login
Search
When replying, please rewrite the subject suitably and try to Cc: the appropriate developer(s). add-lzo1x-algorithm-to-the-kernel.patch make-common-helpers-for-seq_files-that-work-with-list_head-s.patch lots-of-architectures-enable-arbitary-speed-tty-support.patch Merge serial-assert-dtr-for-serial-console-devices.patch Don't know. I worry about Russell's concern (see the changelog) git-acpi-s390-struct-bin_attribute-changes.patch cpuidle-add-rating-to-the-governors-and-pick-the-one-with-highest-rating-by-default-fix.patch exit-acpi-processor-module-gracefully-if-acpi-is-disabled.patch fix-empty-macros-in-acpi.patch drivers-acpi-sbsc-remove-dead-code.patch acpi-enable-c3-power-state-on-dell-inspiron-8200.patch drivers-acpi-pci_linkc-lower-printk-severity.patch Sent to lenb working-3d-dri-intel-agpko-resume-for-i815-chip.patch Sent to davej cifs-use-simple_prepare_write-to-zero-page-data.patch cifs-zero_user_page-conversion.patch Sent to sfrench bugfix-cpufreq-in-combination-with-performance-governor.patch restore-previously-used-governor-on-a-hot-replugged-cpu.patch Sent to davej kcopyd-use-mutex-instead-of-semaphore.patch Sent to agk powerpc-promc-remove-undef-printk.patch 8xx-mpc885ads-pcmcia-support.patch dts-kill-hardcoded-phandles.patch ppc-remove-dead-code-for-preventing-pread-and-pwrite-calls.patch viotape-use-designated-initializers-for-fops-member.patch make-drivers-char-hvc_consoleckhvcd-static.patch powerpc-enable-arbitary-speed-tty-ioctls-and-split.patch powerpc-tlb_32c-build-fix.patch sky-cpu-and-nexus-code-style-improvement.patch sky-cpu-and-nexus-include-ioh.patch sky-cpu-and-nexus-check-for-platform_get_resource-ret.patch sky-cpu-and-nexus-check-for-create_proc_entry-ret-code.patch sky-cpu-use-c99-style-for-struct-init.patch Sent to ...
Andrew, Linus, These got sent to Andi as well, but the patches are independent of the x86_64 conversion. These are bugfixes (nohz-fix-nohz-x86-dyntick-idle-handling.patch) and general improvements of the core code and the existing i386 code. Can we please merge the above now ? I can resend them or setup a git repo if you want. Andi, any objections against the above i386 fixlets ? Thanks, tglx -
On Thu, 12 Jul 2007 01:03:28 +0200 They all look pretty innocuous to me. Could you please take a second look, decide if any of them should also be in 2.6.22.x and let me know? -
Andrew, i386-hpet-check-if-the-counter-works.patch pcspkr-use-the-global-pit-lock.patch are the only candidates. tglx -
yup, come through -stable a few times, be great to get them upstream, and into .22.y thanks, -chris -
Ok by me, although I suspect a lot of the cases where the hpet one was needed got resolved with the PCI HPET resource fix But it's still safer to check. However I don't think patches should go into stable before they hit Linus' tree. -Andi -
Agreed, we're just waiting ;-) thanks, -chris -
On Thu, 12 Jul 2007 01:59:23 +0200 OK, I queued them up for an akpm->linus transfer. Which will of course be abandoned if an akpm->andi or andi->linus merge happens in the next week or so. -
Well ... a few hours ago I just finished the 'unrelated task' that kept
me from doing much cpuset work the last six months.
So, after a little bit of saved up vacation (SGI sabbatical - yippee!),
I should be able to dig into this, see what Dinikar and Cliff have been
up to here, and make some progress on this. Cliff -- did I see some
work from you go by relating to cpusets and the sched domain hooks in
cpusets?
As before, Andrew, if you're getting bored holding this patch, it's ok
to drop it, even though I still figure you'll see it again, as part of
a patch set to improve this cpuset to sched domain interaction.
--
I won't rest till it's the best ...
Programmer, Linux Scalability
Paul Jackson <pj@sgi.com> 1.925.600.0401
-
> fallocate-implementation-on-i86-x86_64-and-powerpc.patch Still broken: arch/x86_64/ia32/ia32entry.S wants compat_sys_fallocate instead of sys_fallocate. Also compat_sys_fallocate probably should be moved to We reserved a different syscall number than the one that is used right now in the patch. Please drop this patch... Martin or I will wire up the syscall as soon as the x86 variant is merged. Everything else just causes trouble and confusion. -
OK, I dropped all the fallocate patches. That means that a few other syscalls (or at least, revoke) get renumbered. -
Andrew, I want to clarify who is going to push the fallocate patches. I can either push them to Linus as part of the ext4 patch set, or we can wait for you to push them. I thought since you had them in -mm and we were going to wait you to push them (and presume that this was going to happen soon). Alternatively I can push them directly to Linus along with other ext4 patches. We can drop the s390 patch if Martin or Heiko wants to wire it up themselves. As far as I know there hasn't been any real contention on the actual syscall patches, other than the numbering issues, so it seems that pushing them to Linus sooner rather than later is the right thing to do. I don't particularly care who pushes them, just as long as they get pushed. :-) So if you've dropped, shall I push them to Linus as part of the ext4 patches we've been planning on pushing? Regards, - Ted -
How about you send them? The syscall numbers might need to be changed I note that nobody really bothered reviewing all those ext4 patches. Do you feel that they have been adequately reviewed? I don't. I guess I ia64 needs changing too. -
Yes, please drop the s390 patch. In general it seems to be better if only one architecture gets a syscall wired up initially and let other arches follow later. Just wondering if the x86_64 compat syscall gets ever fixed? I think I mentioned already three or four times to Amit that it is broken. Or is it that nobody cares? Dunno.. In addition there used to be a somewhat inofficial rule that new syscalls have to come with a test program, so people can easily test if they wired up the syscall correctly. -
On Tue, 10 Jul 2007 20:05:31 +0200 Yep. otoh, fallocate() was special, because we had so many problems working out how to organise the args so that certain kooky architectures can implement Yes please. I normally just slam the whole .c file into the changelog. I'd support an ununofficial rule that submitters of new syscalls also raise a patch against LTP, come to that... -
s/ununofficial//, please. And extend this to every new kernel interface that's not bound to a specific piece of hardware. -
Agree, and cc manpages maintainer too? (preferably write most of the manpage body as well, IMO). -- SUSE Labs, Novell Inc. -
Yes, please. Docs written by, or with input from, the implementer provide one of the few ways for everyone else to spot differences between implementation and intention. -- Ist Ihr Browser Vista-kompatibel? Jetzt die neuesten Browser-Versionen downloaden: http://www.gmx.net/de/go/browser -
And a patch for the manpages. Definitely in favor. -Andi -
Last time it was brought up was when TAKE5 of the patchset was posted and I had planned to fix this in the TAKE6 - which didn't happen since there was no final descision on the mode flags. Anyhow, the x86_64 compat syscall has already been fixed in the ext4 patch queue. I will repost all the patches rebased on 2.6.22 (as they are in the Ok. Will work on a small testcase and post it soon. -- Regards, Amit Arora -
Where is the latest and greatest version of those patches? Is it still the patch set distributed in 2.6.22-rc6-mm1? I'd mostly like to see the final set of flags we're planning on supporting. But yeah, I second the "sooner rather than later" :) --Mark -- Mark Fasheh Senior Software Developer, Oracle mark.fasheh@oracle.com -
I have posted the latest fallocate patches as part of TAKE6. These patches are exactly same as how they currently look in the ext4 patch queue being maintained by Ted. -- Regards, Amit Arora -
The menuconfig patch looks fine. I looked at the May-2007 discussion of ioctl removal. I don't see that much has changed since then, so it's either be brave/foolish/whatever and see what happens or just wait. I'll gladly send a patch to update the removal date in feature-removal-schedule.txt --- ~Randy *** Remember to use Documentation/SubmitChecklist when testing your code *** -
I have a note here that the ioctl-removal patch needs Dominik consideration. I see no rush on it so I'll just sit on it. -
Paul, Are you working on a new version? I thought it was mostly ready for mainline. -- Regards, vatsa -
in particular CONFIG_FAIR_GROUP_SCHED depends on these APIs. Once the APIs to configure are upstream, the group scheduler can be enabled too. (basically all the group scheduling bits are upstream now as part of CFS.) Ingo -
There are definitely some big changes that I want to make internally to the framework, but I guess they don't have to block pushing the basic framework to mainline. I've got a new patchset that's primarily got all the various -mm fix patches rolled into the appropriate original patches, along with plus some small tweaks - changed the Kconfig files to avoid using "select" - adding the subsystem name as a prefix for each control file to enforce namespace scoping - misc contributions from others Short-term I also want to: - rethink the linked list that runs through each task to its css_group object, since that seemed to hurt performance a bit, but for now that can probably be solved by just ripping it out and going back to scanning the tasklist to enumerate tasks in a container. - extend the options parsing, so we can have more than just a list of subsystems. Probably changing the existing -o<subsys1>,<subsys2>,... to be one of: -osubsys=<subsys1>:<subsys2>:...,<otheropt>=<otherval> -osubsys=<subsys1>,subsys=<subsys2>,subsys=...,<otheropt>=<otherval> (what's the preferred convention for fs mount options with multiple values?) I'd not realised that anything else depending on containers was ready for upstream merge, but if CFS group support is ready then merging a subset of them is probably a good idea, since this is an application that I can see a lot of people wanting to play with. Andrew, how about we merge enough of the container framework to support CFS? Bits we could leave out for now include container_clone() support and the nsproxy subsystem, fork/exit callback hooks, and possibly leave cpusets alone for now (which would also mean we could skip the automatic release-agent stuff). I'm in Tokyo for the Linux Foundation Japan symposium right now, but I should be able to get the new patchset to you for Friday afternoon. Paul -
On Tue, 10 Jul 2007 11:34:38 -0700 mm.. Given that you propose leaving bits out for the 2.6.23 merge, and that changes are still pending and that nothing will _use_ the framework in 2.6.23 I'd be inclined to err on the side of caution and hold it all back from 2.6.23. This has the advantage that the merge will happen after the kernel-summit containers discussion which I suspect will be an important point in the life of this project... -
Andrew, The cpu group scheduler is ready and waiting for the container patches in 2.6.23 :) Here are some options with us: a. (As Paul says) merge enough of container patches to enable its use with cfs group scheduler (and possibly cpusets?) b. Enable group scheduling bits in 2.6.23 using the user-id grouping mechanism (aka fair user scheduler). For 2.6.24, we could remove this interface and use Paul's container patches instead. Since this means change of API interface between 2.6.23 and 2.6.24, I don't prefer this option. c. Enable group scheduling bits only in -mm for now (2.6.23-mmX), using Paul's container patches. I can send you a short patch that hooks up cfs group scheduler with Paul's container infrastructure. If a. is not possible, I would prefer c. Let me know your thoughts .. -- Regards, vatsa -
I'm inclined to take the cautious route here - I don't think people will be dying for the CFS thingy (which I didn't even know about?) in .23, and it's rather a lot of infrastructure to add for a CPU scheduler configurator gadget (what does it do, anyway?) We have plenty of stuff for 2.6.23 already ;) Is this liveable with?? -
Hmm ok, if you think the container patches is too early for 2.6.23, fine. We should definitely target to have it in 2.6.24, by which time I am Fine. I will request you to enable group cpu scheduling in 2.6.23-rcX-mmY atleast, so that it gets some amount of testing. The essential group scheduling bits is already in Linus' tree now (as part of cfs merge), so what you need in -mm is a slim patch to hook it with Paul's container infrastructure (which I trust will continue to be in -mm until it goes mainline). I will send across that slim patch later (to be included in 2.6.23-rc1-mm1 perhaps). -- Regards, vatsa -
another option would be to trivially hook up CONFIG_FAIR_GROUP_SCHED with cpusets, and to offer CONFIG_FAIR_GROUP_SCHED in the Kconfig, dependent on CPUSETS and defaulting to off. That would give it a chance to be tested, benchmarked, etc. Ingo -
ah ... you triggered my procmail filter for 'cpuset' ... ;).
What would it mean to hook up CFS with cpusets? I've a pretty
good idea what a cpuset is, but don't know what kind of purpose
you have in mind for such a hook. Could you say a few words to
that? Thanks.
--
I won't rest till it's the best ...
Programmer, Linux Scalability
Paul Jackson <pj@sgi.com> 1.925.600.0401
-
CFS is the new cpu scheduler in Linus's tree (http://lwn.net/Articles/241085/). It has some group scheduling capabilities added i.e the core scheduler now recognizes the concept of a task-group and providing fair cpu time to each task-group (in addition to providing fair time to each task in a group). The core scheduler however is not concerned with how task groups are formed and/or how tasks migrate between groups. Thats where a patch like Paul Menage's container infrastructure comes in hand - to provide a user-interface for managing task-groups (create/delete task groups, migrate task from one group to another etc). Whatever the chosen user-interface is, cpu scheduler needs to know about such task-group creation/destruction, migration of tasks across groups etc. Unfortunately, the group-scheduler bits will be ready in 2.6.23 while Paul Menage's container patches aren't ready for 2.6.23 yet. So Ingo was proposing we use cpuset as that user interface to manage task-groups. This will be only for 2.6.23. In 2.6.24, when hopefully Paul Menage's container patches will be ready and will be merged, the group cpu scheduler will stop using cpuset as that interface and use the container infrastructure instead. If you recall, I have attempted to use cpuset for such an interface in the past (metered cpusets - see [1]). It brings in some semantic changes for cpusets, most notably: - metered cpusets cannot have grand-children - all cpusets under a metered cpuset need to share the same set of cpus. Is it fine if I introduce these semantic changes, only for 2.6.23 and only when CONFIG_FAIR_GROUP_SCHED is enabled? This will let the group cpu scheduler to receive some amount of testing. The other alternative is to hook up group scheduler with user-id's Reference: 1. http://marc.info/?l=linux-kernel&m=115946525811848&w=2 -- Regards, vatsa -
Good explanation - thanks.
In short, the proposal was to use the task partition defined by cpusets
to define CFS task-groups, until the real process containers are
available.
Or, I see in the next message, Ingo responding favorably to your
alternative, using task uid's to partition the tasks into CFS
task-groups.
Yeah, Ingo's preference for using uid's (or gid's ??) sounds right to
me - a sustainable API.
Wouldn't want to be adding a cpuset API for a single 2.6.N release.
... gid's -- why not?
--
I won't rest till it's the best ...
Programmer, Linux Scalability
Paul Jackson <pj@sgi.com> 1.925.600.0401
-
Or process or process groups, or all of the above :-) One thing to think on though, we cannot have per process,uid,gid,pgrp scheduling for one release only. So we'd have to manage interaction with process containers. It might be that a simple weight multiplication scheme is good enough: weight = uid_weight * pgrp_weight * container_weight Of course, if we'd only have a single level group scheduler (as was proposed IIRC) it'd have to create intersection sets (as there might be non trivial overlaps) based on these various weights and schedule these resulting sets instead of the initial groupings. -
Lets illustrate with some ASCII art: so we have this dual level weight grouping (uid, container) uid: a a a a a b b b b b c c c c c container: A A A A A A A B B B B B B B B set: 1 1 1 1 1 2 2 3 3 3 4 4 4 4 4 resulting in schedule sets 1,2,3,4 so that (for instance) weight_2 = weight_b * weight_A -
We would need something like this to flatten hierarchy, so that for example it is possible to do fair-container scheduling + fair-user/process scheduling inside a container using a hierarchy depth of just 1 (containers) that core scheduler understands. We discussed this a bit at http://marc.info/?l=linux-kernel&m=118054481416140&w=2 and is very much Wouldn't it be simpler if admin created these sets as containers directly? i.e: uid: a a a a a b b b b b c c c c c container: 1 1 1 1 1 2 2 3 3 3 4 4 4 4 4 That way scheduler will not have to "guess" such intersecting schedulable sets/groups. It seems much simpler to me this way. Surely there is some policy which is driving some tasks of userid 'b' to be in container A and some to be in B. It should be trivial enough to hook onto that policy making script and create separate -- Regards, vatsa -
could you just try this and send an as simple patch as possible? This is actually something that non-container people would be interested in as well. (btw., if this goes into 2.6.23 then we cannot possibly turn it off in 2.6.24, so it must be sane - but per UID task groups are certainly sane, the only question is how to configure the per-UID weight after bootup. [the default after-bootup should be something along the lines of 'same weight for all users, a bit more for root'.]) This would make it possible for users to test that thing. (it would also help X-heavy workloads.) Ingo -
Note that interfacing with container infrastructure doesn't preclude the possibility of doing fair-user scheduling (that a normal university server or desktop user would want). All that is needed is a daemon which listens for uid change events from kernel (using process-event connector) and moves the task (whose uid is changing) to an appropriate container for that user. The fact that we will have two interface for group scheduler in 2.6.24 is what worries me a bit (one user-id based and other container based). We would need some mechanism for admin to choose only one interface (and not both together, otherwise the group definitions may conflict), which doesn't sound very clean to me. Ideally I would have liked to hook onto only container infrastructure and let user-space decide grouping policy (whether user-id based or something else). Hmm ..would it help if I maintain a patch outside the mainline which turns on fair-user scheduling in 2.6.23-rcX? Folks will have to apply that patch on top of 2.6.23-rcX to use and test fair-user scheduling. In 2.6.24, when container infrastructure goes in, people can get fair-user scheduling off-the-shelf by simply starting the daemon attached at bootup/initrd time. Or would you rather prefer that I add user-id based interface permanently and in 2.6.24 introduce a compile/run-time switch for admin Yeah ..the container based infrastructure allows for configuring such things very easily using a fs-based interface. In the absence of that, we either provide some /proc interface or settle for the non-configurable -- Regards, vatsa
Yeah.
One -could- take linear combinations, as Peter drew in his ascii art,
but would one -want- to do that?
I imagine some future time, when users of this wonder why the API is
more complicated than seems necessary, with two factors determining
task-groups where one seems sufficient, and the answer is "the other
factor, user-id's, is just there because we needed it as an interim
mechanism, and then had to keep it, to preserve ongoing compatibility.
That's not a very persuasive justification.
--
I won't rest till it's the best ...
Programmer, Linux Scalability
Paul Jackson <pj@sgi.com> 1.925.600.0401
-
I'd very much like to have it, but that is just me. We could take a
weight of 0 to mean disabling of that grouping and default to that. That
way it would not complicate regular behaviour.
It could be implemented with a simple hashing scheme where
sched_group_hash(tsk) and sched_group_cmp(tsk, group->some_task) could
be used to identify a schedule group.
pseudo code:
u64 sched_group_hash(struct task_struct *tsk)
{
u64 hash = 0;
if (tsk->pid->weight)
hash_add(&hash, tsk->pid);
if (tsk->pgrp->weight)
hash_add(&hash, tsk->pgrp);
if (tsk->uid->weight)
hash_add(&hash, tsk->uid);
if (tsk->container->weight)
hash_add(&hash, tsk->container);
...
return hash;
}
s64 sched_group_cmp(struct task_struct *t1, struct task_struct *t2)
{
s64 cmp;
if (t1->pid->weight || t2->pid->weight) {
cmp = t1->pid->weight - t2->pid->weight;
if (cmp)
return cmp;
}
...
return 0;
}
u64 sched_group_weight(struct task_struct *tsk)
{
u64 weight = 1024; /* 1 fixed point 10 bits */
if (tsk->pid->weight) {
weight *= tsk->pid->weight;
weight /= 1024;
}
....
return weight;
}
-
Why? [linear combinations of uid, container, pid, pgrp weighting]
You provide some implementation details and complications, but no
motivation that I noticed.
Well ... a little motivation ... "just me", which would go a long
way of your first name was Linus. For the rest of us ... ;).
--
I won't rest till it's the best ...
Programmer, Linux Scalability
Paul Jackson <pj@sgi.com> 1.925.600.0401
-
Good question, and I really have no other answer than that it seems usefull and not impossible (or even hard) to implement :-/ I'm not even that interested in using it, it just seems like a nice idea. -
I know breaking user-interface is a bad thing across releases. But in this particular case, it's probably ok (since fair-group scheduling is a brand new feature in Linux)? If we have that option of breaking API between 2.6.23 and 2.6.24 for fair-group scheduler, then we are in a much more flexible position. For 2.6.23, I can send a user-id based interface for fair-group scheduler (with some /proc interface to tune group nice value). For 2.6.24, this user-id interface will be removed and we will instead switch to container based interface. Fair-user scheduling will continue to work, its just that users will have to use a daemon (sources sent in previous mail) to enable it on top of container-based interface. Hmm..? -- Regards, vatsa -
Selecting the relevant patches to give enough of the container framework to support a CFS container subsystem (slightly tweaked/updated versions of the base patch, procfs interface patch and tasks file interface patch) is about 1600 lines in kernel/container.c and another 200 in kernel/container.h, which is about 99% of the non-documentation changes. So not tiny, but it's not very intrusive on the rest of the kernel, and would avoid having to introduce a temporary API based on uids. Paul -
Yes that would be good. As long as the user-land interface for process containers doesn't change (much?) between 2.6.23 and later releases this should be a good workaround for us. -- Regards, vatsa -
That's what I was originally thinking too, but since CFS has been merged, CFS group scheduling would use it. Paul -
This is a core PCI change, should probably go through Greg and/or linux-pci instead? Or just send it to Linus directly, iirc everyone was ok with the change for 2.6.23. Also, you can add a Signed-off-by: Jesse Barnes <jesse.barnes@intel.com> to it. Thanks, Jesse -
Andrew, Could you replace the comment/header section of fallocate-implementation-on-i86-x86_64-and-powerpc.patch with the following (attached below) ? This is from the ext4 patches, where Amit had cleaned up description, which will make for a cleaner and easier to understand submission into the git tree. I've reviewed the other fallocate patches, noting the request to drop the s390 patches since Martin has said he will wire up it up after this hits mainline, and the only other change that I've found between what we have in the ext4 tree and -mm is that we have fallocate-on-ia64.patch and fallocate-on-ia64-fix.patch merged into a single patch. It probably would be better to merge them before sending it off to Linus, in the interests of cleanliness and making the tree more git-bisect friendly. Regards, - Ted From: Amit Arora <aarora@in.ibm.com> sys_fallocate() implementation on i386, x86_64 and powerpc fallocate() is a new system call being proposed here which will allow applications to preallocate space to any file(s) in a file system. Each file system implementation that wants to use this feature will need to support an inode operation called ->fallocate(). Applications can use this feature to avoid fragmentation to certain level and thus get faster access speed. With preallocation, applications also get a guarantee of space for particular file(s) - even if later the the system becomes full. Currently, glibc provides an interface called posix_fallocate() which can be used for similar cause. Though this has the advantage of working on all file systems, but it is quite slow (since it writes zeroes to each block that has to be preallocated). Without a doubt, file systems can do this more efficiently within the kernel, by implementing the proposed fallocate() system call. It is expected that posix_fallocate() will be modified to call this new system call first and incase the kernel/filesystem does not implement it, it should fall back to the ...
There were issues with the x86 patch, the s390 patch was wrong and Tony wants the the ia64 patch to use a different syscall number. So I dropped everything. Let's start again from scratch. I'd suggest that for now we go with just an i386/x86_64 implementation, let the arch maintainers wire things up when that has settled down. -
Ok, so no objections if we push the i386/x86_64 implementation (only), plus the ext4 support to Linus? - Ted -
On Tue, 10 Jul 2007 15:07:35 -0400 Sounds like a plan. I haven't seriously looked at ext4 code in many months. When I did I found the changes to be quite incomprehensible, very, very poorly commented and with quite a lot of odd-looking things about which I asked but got, iirc, no useful reply. Hopefully it got better. Which patches are you proposing merging into 2.6.23? -
It's my observation that that plan usually works the best. Arch maintainers come along and wire up batches of syscalls when they have a chance to glance at the ABI, and catch up with x86[-64]. Jeff -
... except when the initial implementer picks an argument order which doesn't work on some archs, as happened with sys_sync_file_range. That is also the case with fallocate IIRC. We did come up with an order that worked for everybody, but that discussion seemed to get totally ignored by the ext4 developers. Paul. -
On Wed, 11 Jul 2007 09:27:40 +1000 It was a long discussion. Can someone please remind us what the signature of the syscall (and the compat handler) should be? -
long sys_fallocate(loff_t offset, loff_t len, int fd, int mode) should work for everybody. The compat handler would be long compat_sys_fallocate(u32 offset_hi, u32 offset_lo, u32 len_hi, u32 len_lo, int fd, int mode) for big-endian, or swap hi/lo for little-endian. (Actually it would be good to have an arch-dependent "stitch two args together" macro and call them offset_0, offset_1 etc.) Paul. -
Well, in the end it was a toss-up between asmlinkage long sys_fallocate(int fd, int mode, loff_t offset, loff_t len) and asmlinkage long sys_fallocate(loff_t offset, loff_t len, int fd, int mode) There were a number of folks who preferred having int fd first, and I *thought* Amit had gotten agreement from either Martin or Heiko that it was ok to do this as an exception, even though it was extra work for that arch. But if not, we can try going back the second alternative, or even the 6 32-bits args (off_high, off_low, len_high, len_low) approach, but I think that drew even more fire. Basically, no one approach made everyone happy, and at the end of the day sometimes you have to choose. I thought we had settled this in May with something that people could live with, but if we need to reopen the discussion, better now than later..... - Ted -
The second approach would work for all architectures.. but some people didn't like (no technical reason) not having fd as first argument. Just go ahead with the current approach. s390 seems to be the only architecture which suffers from this and I wouldn't like to start this discussion again. -
For s390 we would have liked the second approach with the two int's as last arguments since it would avoid the wrapper in the kernel. It does not avoid the wrapper in user space since the call uses 6 register on 31 bit. So the fallocate call need special treatement in glibc so I don't Yes, don't worry about s390 for fallocate, the patch that had been in -mm only had an incorrect system call number. The wrapper is fine. -- blue skies, Martin. "Reality continues to ruin my life." - Calvin. -
Quoting Andrew Morton (akpm@linux-foundation.org): Andrew Morgan has requested a series of changes. Since one of these would involve a change in the on-disk format of file capabilities, I guess these should (sigh) wait another cycle. I will try to get that change out the door next, as soon as possible, so that hopefully there are no more definite blocking requests. thanks, -serge -
On Tuesday, 10 July 2007 10:31, Andrew Morton wrote: Hmm, I'm not sure what that means. Am I supposed to do anything about it? Greetings, Rafael -- "Premature optimization is the root of all evil." - Donald Knuth -
Was this oops silently fixed? http://lkml.org/lkml/2007/3/2/138 2.6.21-rc2-mm1: EIP is at ext2_discard_reservation+0x1c/0x52 I still have that ext2 partition backed up. -
On Tue, Jul 10, 2007 at 01:31:52AM -0700, Andrew Morton wrote: > working-3d-dri-intel-agpko-resume-for-i815-chip.patch > > Sent to davej You managed to sneak this to me just hours before I handed AGP maintainership to Dave Airlie. FWIW, I think this needs to redone in a much more generic manner before it goes mainline. > bugfix-cpufreq-in-combination-with-performance-governor.patch > restore-previously-used-governor-on-a-hot-replugged-cpu.patch > > Sent to davej Will start merging the backlog once someone (ahem) pulls the last lot of stuff. Dave -- http://www.codemonkey.org.uk -
~swap prefetch Nick's only remaining issue which I could remotely identify was to make it cpuset aware: http://marc.info/?l=linux-mm&m=117875557014098&w=2 as discussed with Paul Jackson it was cpuset aware: http://marc.info/?l=linux-mm&m=117895463120843&w=2 I fixed all bugs I could find and improved it as much as I could last kernel cycle. Put me and the users out of our misery and merge it now or delete it forever please. And if the meaningless handwaving that I 100% expect as a response begins again, then that's fine. I'll take that as a no and you can dump it. -- -ck -
The window for 2.6.23 has now closed and your position on this is clear. I've been supporting this code in -mm for 21 months since 16-Oct-2005 without any obvious decision for this code forwards or backwards. I am no longer part of your operating system's kernel's world; thus I cannot support this code any longer. Unless someone takes over the code base for swap prefetch you have to assume it is now unmaintained and should delete it. Please respect my request to not be contacted further regarding this or any other kernel code. -- -ck -
For what it's worth; put me down as supporting the merger of swap prefetch. I've found it useful in the past, Con has maintained it nicely and cleaned up everything that people have pointed out - it's mature, does no harm - let's just get it merged. It's too late for 2.6.23-rc1 now, but let's try and get this in by -rc2 - it's long overdue... -- Jesper Juhl <jesper.juhl@gmail.com> Don't top-post http://www.catb.org/~esr/jargon/html/T/top-post.html Plain text mails only, please http://www.expita.com/nomime.html -
Not talking about swap prefetch itself, but everytime I have asked anyone to instrument or produce some workload where swap prefetch helps, they never do. Fair enough if swap prefetch helps them, but I also want to look at why that is the case and try to improve page reclaim in some of these situations (for example standard overnight cron jobs shouldn't need swap prefetch on a 1 or 2GB system, I would hope). Anyway, back to swap prefetch, I don't know why I've been singled out as the bad guy here. I'm one of the only people who has had a look at the damn thing and tried to point out areas where it could be improved to the point of being included, and outlining things that are needed for it to be merged (ie. numbers). If anyone thinks that makes me the bad guy then they have an utterly inverted understanding of what peer review is for. Finally, everyone who has ever hacked on these heuristicy parts of the VM has heaps of patches that help some workload or some silly test case or (real or percieved) shortfall but have not been merged. It really isn't anything personal. If something really works, then it should be possible to get real numbers in real situations where it helps (OK, swap prefetching won't be as easy as a straight line performance improvement, but still much easier than trying to measure something like scheduler interactivity). Numbers are the best way to add weight to the pro-merge argument, so for all the people who a whining about merging this and don't want to actually work on the code -- post some numbers for where it helps you!! -- SUSE Labs, Novell Inc. -
<Raised eyebrow> You sound frustrated. Perhaps we could be communicating better. I'll start. Unlike others on the cc: line, I don't get paid to hack on the kernel, not even indirectly. So if you find that my lack of providing numbers is giving you heartache, I can only apologize and point at my paying work that requires my attention. That said, I'm willing to run my day to day life through both a swap prefetch kernel and a normal one. *However*, before I go through all the work of instrumenting the damn thing, I'd really like Andrew (or Linus) to lay out his acceptance criteria on the feature. Exactly what *should* I be paying attention to? I've suggested keeping track of process swapin delay total time, and comparing with and without. Is that reasonable? Is it incomplete? Without Andrew's criteria, we're back to where we've been for a long time: lots of work, no forward motion. Perhaps it's a character flaw of mine, but I'd really like to know what would constitute proof here before I invest the effort. Especially given that Con has already written a test case that shows that swap prefetch works, and that I've given you a clear argument for why better (or even perfect) page reclaim can't provide full coverage to all the situations that swap prefetch helps. (Also, it's not like I've got tons free time, y'know? Just like all the rest of you all, I have to pick and choose my battles if I'm going to be effective.) Since this merge period has appeared particularly frazzling for Andrew, I've been keeping silent and waiting for him to get to a point where there's a breather. I didn't feel it would be polite to request yet more work out of him while he had a mess on his hands. But, given this has come to a head, I'm asking now. Andrew? You've always given the impression that you want this run more as an engineering effort than an artistic endeavour, so help us out here. What are your concerns with swap prefetch? What sort of comparative data would you like to see to ...
Let it just be noted that Con is not the only one who has expended effort on this patch. It's been in -mm for nearly two years and it has meant ongoing effort for me and, to a lesser extent, other MM developers to keep Critera are different for each patch, but it usually comes down to a cost/benefit judgement. Does the benefit of the patch exceed its maintenance cost over the lifetime of the kernel (whatever that is). In this case the answer to that has never been clear to me. The (much older) fs-aio patches were (are) in a similar situation. The other consideration here is, as Nick points out, are the problems which people see this patch solving for them solveable in other, better ways? IOW, is this patch fixing up preexisting deficiencies post-facto? To attack the second question we could start out with bug reports: system A with workload B produces result C. I think result C is wrong for <reasons> Well. The above, plus there's always a lot of stuff happening in MM land, and I haven't seen much in the way of enthusiasm from the usual MM developers. -
<nods> Yes, keeping patches from crufting up and stepping on other patches' toes is hard work; I did it for a bit, and it was one of the more thankless tasks I've tried a hand at. Well, I suspect it's 'lifetime of the feature,' in this case as it's no more user visible than the page replacement algorithm in the first In some cases, it almost certainly is. It also has the troubling aspect of mitigating future regressions without anyone terribly noticing, due to it being able to paper over those hypothetical future I spend a lot of time each day watching my computer fault my workingset back in when I switch contexts. I'd rather I didn't have to do that. Unfortunately, that's a pretty subjective problem report. For whatever it's worth, we have pretty subjective solution reports pointing to swap prefetch as providing a fix for them. My concern is that a subjective problem report may not be good enough. So, what do I measure to make this an objective problem report? And if I do that (and it shows a positive result), will that be good enough to argue for inclusion? -
Ideal would be to find a reproducible-by-others testcase which does what you That depends upon whether there are more suitable ways of fixing "the wrong thing". There may not be - it could well be that present behaviour is correct for the testcase, but it leaves the system in the wrong state for your large workload shift. In that case, prefetching (ie: restoring system state approximately to that which prevailed prior to "testcase") might well be a suitable fix. -
That seems to be the crux of the matter: how to measure subjective usability issues (aka user experience) when simple reports along the lines of "A is much better than B for everyday work" are not enough. The same problem already impaired the "fair scheduler" discussion. It would really help to have a clear direction there. --=20 Tilman Schmidt E-Mail: tilman@imap.cc Bonn, Germany Diese Nachricht besteht zu 100% aus wiederverwerteten Bits. Unge=C3=B6ffnet mindestens haltbar bis: (siehe R=C3=BCckseite)
So let me get this straight - you don't want to merge swap prefetch which exists now and solves issues many people are seeing, and has been tested more than a gazillion other bits & pieces that do get merged - because it could be possible that in the future some other patch, which doesn't yet exist and nobody is working on, may solve the problem better? You know what, just release Linux 0.02 as 2.6.23 because, using your logic, everything that was merged since October 5, 1991 could be replaced by something better. Perhaps. So there's obviously no point having it there in the first place & there'll be untold savings in storage costs and compilation time for the kernel tree, also bandwidth for the mirror sites etc. in the mean time while we wait for the magic pixies to come and deliver the one true piece of code that cannot be I haven't seen much in the way of enthusiasm from developers, period. People are tired of maintaining patches for years that never get merged into mainline because of totally bullshit reasons (usually amounting to NIH syndrome) -- Matt -
From: "Matthew Hawkins" <darthmdh@gmail.com> I have to generally agree that the objections to the swap prefetch patches have been conjecture and in general wasting time and frustrating people. There is a point at which it might be wise to just step back and let the river run it's course and see what happens. Initially, it's good to play games of "what if", but after several months it's not a productive thing and slows down progress for no good reason. If a better mechanism gets implemented, great! We'll can easily replace the swap prefetch stuff at such time. But until then swap prefetch is what we have and it's sat long enough in -mm with no major problems to merge it. -
Um, isn't it up to you? The questions that need to be answered are:
1. What are you trying to achieve? Presumably you have some intended
or desired effect you're trying to get. What's the intended
audience? Who would be expected to see a benefit? Who suffers?
2. How does the code achieve that end? Is it nasty or nice? Has
everyone who's interested in the affected areas at least looked at
the changes, or ideally given them a good review? Does it need
lots of tunables, or is it set-and-forget?
3. Does it achieve the intended end? Numbers are helpful here.
4. Does it make anything worse? A lot or a little? Rare corner
cases, or a real world usage? Again, numbers make the case most
strongly.
I can't say I've been following this particular feature very closely,
but these are the fundamental questions that need to be dealt with in
merging any significant change. And as Nick says, historically point 4
is very important in VM tuning changes, because "obvious" improvements
have often ended up giving pathologically bad results on unexpected
workloads.
J
-
Huh? I'm not Linus or Andrew, with the power to merge a patch to the Dude. My whole question was *what* numbers. Please go back and read it all again. Maybe I was unclear, but I really don't think so. -
I don't feel it is so useful without more context. For example, in most situations where pages get pushed to swap, there will *also* be useful file backed pages being thrown out. Swap prefetch might improve the total swapin delay time very significantly but that may be just a tiny portion of the real problem. Also a random day at the desktop, it is quite a broad scope and pretty well impossible to analyse. If we can first try looking at some specific problems that are easily identified. Looking at your past email, you have a 1GB desktop system and your overnight updatedb run is causing stuff to get swapped out such that swap prefetch makes it significantly better. This is really intriguing to me, and I would hope we can start by making this particular workload "not suck" without swap prefetch (and hopefully make it even better than it currently is with swap prefetch because we'll try not to evict useful file backed pages as well). After that we can look at other problems that swap prefetch helps with, or think of some ways to measure your "whole day" scenario. So when/if you have time, I can cook up a list of things to monitor and possibly a patch to add some instrumentation over this updatedb run. Anyway, I realise swap prefetching has some situations where it will fundamentally outperform even the page replacement oracle. This is why I haven't asked for it to be dropped: it isn't a bad idea at all. However, if we can improve basic page reclaim where it is obviously lacking, that is always preferable. eg: being a highly speculative operation, swap prefetch is not great for power efficiency -- but we still want laptop users to have a good experience as well, right? -- SUSE Labs, Novell Inc. -
Agreed, it's important to make sure we're not being penny-wise and It is pretty broad, but that's also what swap prefetch is targetting. As for hard to analyze, I'm not sure I agree. One can black-box test this stuff with only a few controls. e.g., if I use the same apps each day (mercurial, firefox, xorg, gcc), and the total I/O wait time consistently goes down on a swap prefetch kernel (normalized by some control statistic, such as application CPU time or total I/O, or Always easier, true. Let's start with "My mouse jerks around under memory load." A Google Summer of Code student working on X.Org claims that mlocking the mouse handling routines gives a smooth cursor under load ([1]). It's surprising that the kernel would swap that out in the first place. updatedb is an annoying case, because one would hope that there would be a better way to deal with that highly specific workload. It's also pretty stat dominant, which puts it roughly in the same category as a git diff. (They differ in that updatedb does a lot of open()s and getdents on directories, git merely does a ton of lstat()s instead.) Anyway, my point is that I worry that tuning for an unusual and infrequent workload (which updatedb certainly is), is the wrong way to That would be appreciated. Don't spend huge amounts of time on it, okay? Point me the right direction, and we'll see how far I can run Absolutely. Disk I/O is the enemy, and the best I/O is one you never had to do in the first place. -
Sounds like something that can be altered with a tuneable for workloads where power efficiency is more important than performance. As far as performance goes, empty memory is wasted memory. I think the most important 'measurement' people can make for swap prefetch, if this is even possible to capture, is a positive hit ratio. Under everyman's typical workload, what percentage of pages prefetched end up being used? And what percentage end up discarded? I'm pulling these numbers out of thin air, but I would say, if > 10% is referenced, and < 70% discarded, then that would be significant performance boost well worthwhile. To be clear, I don't know what I'm talking about. It just seems to me however, that debating whether or not to implement a performance boost because we can better tune corner cases is silly. For as long as computers have used swap and unused memory, there will be a performance gain to background prefetching. That doesn't preclude developers from tuning the specific workloads that lead to such. It's not like this were a theoretical discussion of should we develop this or not.. Prefetch is here, now, and working. The only questions I see are: Does the performance gain from Prefetch compensate for the prefetch code memory requirements? Is there someone who's comfortable with lkml politics willing to maintain the thing? -
I'm not saying that we can't try to tackle that problem, but first of all you have a really nice narrow problem where updatedb seems to be causing the kernel to completely do the wrong thing. So we start on OK, I'm not sure what the point is though. Under heavy memory load, things are going to get swapped out... and swap prefetch isn't going to help there (at least, not during the memory load). There are also other issues like whether the CPU scheduler is at fault, etc. Interactive workloads are always the hardest to work out. updatedb Yeah, and I suspect we might be able to do better use-once of inode and dentry caches. It isn't really highly specific: lots of things tend to just scan over a few files once -- updatedb Well it runs every day or so for every desktop Linux user, and it has similarities with other workloads. We don't want to optimise it at the expense of other things, but it _really_ should not be I guess /proc/meminfo, /proc/zoneinfo, /proc/vmstat, /proc/slabinfo before and after the updatedb run with the latest kernel would be a first step. top and vmstat output during the run wouldn't hurt either. Thanks, Nick -- SUSE Labs, Novell Inc. -
Sorry, I headed slightly off-topic. Or perhaps 'up-topic' to the larger issue, which is that the desktop experience has some suckiness to it. My point is that the page replacement algorithm has some choice as to what to evict. The xorg input handler never should have been evicted. It was hopefully a hard example of where the current page replacement policy is falling flat on its face. All that said, this could really easily be handled by xorg mlocking This one is not a scheduler issue, as mlock()ing the mouse handling routines gives a smooth cursor. It's just a pure page replacement problem, as the kernel should never have swapped that out in the first place. <snip things I agreed with> <snip list of things to watch during updatedb run> Ray -
It certainly doesn't run for me ever. Always kind of a "that's not the point" comment but I just keep wondering whenever I see anyone complain about updatedb why the _hell_ they are running it in the first place. If anyone who never uses "locate" for anything simply disable updatedb, the problem will for a large part be solved. This not just meant as a cheap comment; while I can think of a few similar loads even on the desktop (scanning a browser cache, a media player indexing a large amount of media files, ...) I've never heard of problems _other_ than updatedb. So just junk that crap and be happy. Rene. -
- updatedb type of workloads that add tons of inodes and dentries in the slab caches which of course use the pagecache. - streaming large files (read or copying) that fill the pagecache with useless used-once data swap prefetch fix the first case, drop-behind fix the second case. Both have the same symptoms but the cause is different. Personally updatedb doesn't really hurt me. But I don't have that many files on my desktop. I've tried the swap prefetch patch in the past and it was not so noticeable for me. (I don't doubt it's helpful for others) But every time I read or copy a large file around (usually from a server) the slowdown is noticeable for some moments. I just wanted to point this out, if it wasn't clean enough for everyone. I hope both problems get fixed. Best regards, - Eric -
OK, this is where I start to worry. Swap prefetch AFAIKS doesn't fix the updatedb problem very well, because if updatedb has caused swapout then it has filled memory, and swap prefetch doesn't run unless there is free memory (not to mention that updatedb would have paged out other files as well). And drop behind doesn't fix your usual problem where you are downloading from a server, because that is use-once write(2) data which is the problem. And this readahead-based drop behind also doesn't help if data you were reading happened to be a sequence of small files, or otherwise not in good readahead order. Not to say that neither fix some problems, but for such conceptually big changes, it should take a little more effort than a constructed test case and no consideration of the alternatives to get it merged. -- SUSE Labs, Novell Inc. -
Sorry for the confusion. For swap prefetch I should have said "some people claim that it fix their problem". I didn't want to hurt anybody feelings, some people are tired to hear others speak hypothetically about this patch, as it work-for-them (TM). I don't experience the problem. Can't help. For drop behind it fix half the problem. The read case is handled perfectly by Peter's patch. And the copy (read+write) is unchanged. My test case demonstrate it very easily, just look at the numbers. So, I agree with you that drop behind doesn't fix the write() case. Peter has said so himself when I offered to test his patch. As I do experience this problem, I have written a small test program and batch file to help push the patch for acceptance. I'm very willing to help improve the test cases, test patches and write code, time permitting. About this very subject, earlier this year this Andrew suggested me to came up with a test case to demonstrate my problem, well finally I've done so. http://lkml.org/lkml/2007/3/3/164 http://lkml.org/lkml/2007/3/3/166 Lastly, I would go as far to say that the use-once read then copy fix must also work with copies over NFS. I don't know if NFS change the workload on the client station versus the local case, and I don't know if it's still possible to consider data copied this way as use-once. - Eric -
Hey Eric, No, there's a third case which I find the most annoying. I have multiple working sets, the sum of which won't fit into RAM. When I finish one, the kernel had time to preemptively swap back in the other, and yet it didn't. So, I sit around, twiddling my thumbs, waiting for my music player to come back to life, or thunderbird, or... -
On Wed, 25 Jul 2007 09:09:01 -0700 Yes, I'm thinking that's a good problem statement and it isn't something which the kernel even vaguely attempts to address, apart from normal demand paging. We could perhaps improve things with larger and smarter fault readaround, perhaps guided by refault-rate measurement. But that's still demand-paged rather than being proactive/predictive/whatever. None of this is swap-specific though: exactly the same problem would need to be solved for mmapped files and even plain old pagecache. In fact I'd restate the problem as "system is in steady state A, then there is a workload shift causing transition to state B, then the system goes idle. We now wish to reinstate state A in anticipation of a resumption of the original workload". swap-prefetch solves a part of that. A complete solution for anon and file-backed memory could be implemented (ta-da) in userspace using the kernel inspection tools in -mm's maps2-* patches. We would need to add a means by which userspace can repopulate swapcache, but that doesn't sound too hard (especially when you haven't thought about it). And userspace can right now work out which pages from which files are in pagecache so this application can handle pagecache, swap and file-backed memory. (file-backed memory might not even need special treatment, given that it's pagecache anyway). And userspace can do a much better implementation of this how-to-handle-large-load-shifts problem, because it is really quite complex. The system needs to be monitored to determine what is the "usual" state (ie: the thing we wish to reestablish when the transient workload subsides). The system then needs to be monitored to determine when the exceptional workload has started, and when it has subsided, and userspace then needs to decide when to start reestablishing the old working set, at what rate, when to abort doing that, etc. All this would end up needing runtime configurability and tweakability ...
<nod> Could be what I'm noticing, but it's important to note that as others have shown improvement with Con's swap prefetch, it's easily arguable that targeting just swap is good enough for a first Yes, that's a fair transformation / generalization. It's always nice So in your proposed scheme, would userspace be polling, er, <goes and looks through email for maps2 stuff, only finds Rusty's patches to it>, well, /proc/<pids>/something_or_another? A userspace daemon that wakes up regularly to poll a bunch of proc files fills me with glee. Wait, is that glee? I think, no... wait... horror, yes, horror is what I'm feeling. Oy. I mean this in the most respectful way possible, but you're too smart for your own good. I mean, sure, it's possible one could have multiply-chained transient workloads each of which have their optimum workingset, of which there's little overlap with the previous. Mainframes made their names on such loads. Workingset A starts, generates data, finishes and invokes workingset B, of which the only thing they share in common is said data. B finishes and invokes C, etc. So, yeah, that's way too complex to stuff into the kernel. Even if it were possible to do so, I cringe at the thought. And I can't believe that would be a common enough pattern nowadays to justify any hueristics on anyone's part. It's certainly complex enough that I'd like to punt that scenario out of the conversation entirely -- I think it has the potential to give a false impression as to how involved of a process we're talking about here. I'll take an 80% solution for that one problem, and happily declare that the kernel's job is done. In particular, when a resource hog exits (or whatever hueristics prefetch is currently hooking in to), the kernel (or userspace, if that interface could be made sane) could exercise a completely workload agnostic refetch of the last n things evicted, where n is determined by what's suddenly become free (or whatever Con came up ...
Actually, I overdesigned the API, I suspect. What we _could_ do is to provide a way of allowing userspace to say "pretend process A touched page B": adopt its mm and go touch the page. We in fact already have that: PTRACE_PEEKTEXT. So I suspect this could all be done by polling maps2 and using PEEKTEXT. The tricky part would be working out when to poll, and when to reestablish. A neater implementation than PEEKTEXT would be to make the maps2 files writeable(!) so as a party trick you could tar 'em up and then, when you want to reestablish firefox's previous working set, do a untar in /proc/$(pidof firefox)/ -
Sick. But thankfully, unnecessary. The pagemaps give you more than just a present bit, which is all we care about here. We simply need to record which pages are mapped, then reference them all back to life.. -- Mathematics is the supreme nostalgia of our time. -
Welllllll.... there is the taskstats interface. It's not required right now, though, and lacks most of what userspace would need, I think. It does at least currently provide a notification of process exit, which is a clue for when to start reestablishment. Gotta be another way we can get at that... Oh, stat on /proc, does that work? Huh, it does, sort of. It seems to be off by 12 or 13, but hey, that's something. Wish I had the time to look at the maps2 stuff, but regardless, it probably currently provides too much detail for continual polling? I suspect what we'd want to do is to take a detailed snapshot a little after the beginning of a process's lifetime (once the block-in counts subside), then poll aggregate residency or evicition counts to know which processes are suffering the burden of the transient workload. Eh, wait, that doesn't help with inodes. No matter, I guess; I'm the one who said targetting swap-in would be good enough for a first pass. On process exit, if userspace can get a hold of an estimate of the size of what just freed up, it could then spend min(that,evicted_count) on repopulation. That's probably already I'm going to get into trouble if I wake up the other person in the house with my laughter. That's laughter in a positive sense, not a "you're daft" kind of way. Huh. <thinks> So, to go back a little bit, I guess one of my problems with polling is that it means that userspace can only approximate an MRU of what's been evicted. Perhaps an approximation is good enough, I don't know, but that's something to keep in mind. (Hmm, how many pages can an average desktop evict per second? If we poll everything once per second, that's how off we could be.) Another is a more philosophical hangup -- running a process that polls periodically to improve system performance seems backward. Okay, so that's my problem to get over, not yours. Another problem is what poor sod would be willing to write and test this, given that there's already a written ...
The userspace solution has been brought up before. It could be a good way to go. I was thinking about how to do refetching of file backed pages from the kernel, and it isn't impossible, but it it seems like locking would be quite hard and it would be pretty complex and inflexible compared to a userspace solution. Userspace might know what to chuck out, what to keep, what access patterns to use... Not that I want to say anything about swap prefetch getting merged: my inbox is already full of enough "helpful suggestions" about that, so I'll just be happy to have a look at little things like updatedb. -- SUSE Labs, Novell Inc. -
Yes, that is a little thing. I mean, even if the kernel's behaviour during an updatedb run was "perfect" (ie: does what we the designers curently intend it to do (whatever that is)) then the core problem isn't solved: short-term workload evicts your working set and you have to synchronously reestablish it. -
It is a good idea if we can give enough to get started. Then if they run into something they really need to do in the kernel, we can take a look. Page eviction order / prefetch-back-in-order might be tricky to Sure, I know and I was never against swap (and/or file) prefetching to solve this problem. I'm just saying, I'm staying out of that :) -- SUSE Labs, Novell Inc. -
Andrew Morton wrote: But a patch already exist. might be. but merging maps2 have higher risk which should be done in a development branch (er... 2.7, but we don't have it now). -- -
This is off-topic and has been discussed to death, but: Rather than one stable branch and one development branch, we have a few stable branches and a lot of development branches. Some are located at git.kernel.org. Among else, this gives you a predictable release rythm and very timely updated stable branches. -- Stefan Richter -=====-=-=== -=== ===-- http://arcgraph.de/sr/ -
Maybe I'm missing something here, but if the problem is resource allocation when switching from state A to state B, and from B to C, etc.; wouldn't it be a bad thing if state B happened to be (in the future) this state-shifting userspace daemon of which you speak? (Or is that likely to be impossible/unlikely for some other reason which alludes me at the moment?) -- Michael Chang Please avoid sending me Word or PowerPoint attachments. Send me ODT, RTF, or HTML instead. See http://www.gnu.org/philosophy/no-word-attachments.html Thank you. -
Well. I was assuming that the daemon wouldn't be a great memory pig. I suspect it would do practically zero IO and would use little memory. It could even be mlocked, but I doubt if that would be needed. -
I don't really understand the reasons for all those discussions. As long as you have a maintainer, why don't you just put swap prefetch into= =20 the kernel, marked experimental, default deactivated so anyone who just=20 make[s] oldconfig (or defaultconfig) won't get it. If anyone finds a good=20 solution for all those cache 'poisoning' problems and problems concerning=20 swapin on workload changes and such swap prefetch can easily taken out agai= n=20 and no one has to complain and continuing maintaining it. Actually the same goes for plugshed (having it might have kept Con as a=20 valuable developer). I am actually waiting for more than 2 years that reiser4 will make it into = the=20 kernel (sure, marked experimental or even highly experimental) so the=20 patch-journey for every new kernel comes to an end. And most things in-kern= el=20 will surely be tested more intense so bugs will come up much faster.=20 (Constantly running MM kernels is not really an option since many things in= =20 there can't be deactivated if they don't work as expected since lots of=20 patches also concern 'vital' parts of the kernel.) =2E..just 2cents from a happy CK user for it made it possible to watch a mo= vie=20 while compiling the system - which was impossible with mainline kernel, eve= n=20 on dual core 2.2 GHz AMD64 with 4G RAM ! Dirk.
=2E..sorry for the reply to myself. As Gentoo user i'm happy about the freedom of choice in almost every aspect= of=20 the system. But with the kernel this freedom is taken away and i'm left with largely=20 choices other people did. Sure, i can get the sources and patch and change= =20 the kernel myself in every aspect but thats more the possibility given by t= he=20 freedom of the open source of the kernel. I think the kernel should be more open for new ideas, and as long as the=20 maintainer follows the kernel development things can be left in, if the=20 maintainer can't follow anymore they are taken out quite fast again. (This= =20 statement mostly counts for parts of the kernel where a choice is possible = or=20 the coding overhead of making such choice possible is quite low.) A problem of Linux is there are 100s of patches and patchsets out there but= if=20 you want to have more than one (since they offer advantages or functionalit= y=20 in different places) you are mostly left alone, start to integrate all=20 patches by hand and solve so many rejects. =2E.still a happy CK user, but sad that Con left the scene. Dirk.
This is just not good engineering. It is axiomatic that it is easy to add code, but difficult to remove code. It takes -years- to remove code that no one uses. Long after the maintainer disappears, the users (and bug reports!) remain. It is also axiomatic that adding code, particularly core code, often exponentially increases complexity. Jeff -
I'll point out that the code that's so hard to remove is the code that exposes an API to userspace. code that's an internal implementation (like a couple of the things being this is true and may be a valid argument (depending on how large and how intrusive the proposed patch is) David Lang -
Not true. It is highly unlikely that code will get removed if it has active users, even if the maintainer has disappeared. The only things that get removed rapidly are those things mathematically guaranteed to be dead code. _Behavior changes_, driver removals, feature removals happen more frequently than userspace ABI changes -- true -- but the rate of removal is still very, very slow. It is axiomatic that we are automatically burdened with new code for at least 10 years :) That's what you have to assume, when accepting anything. Jeff -
if you propose removing code in such a way that performance suffers then yes, it's hard to remove (and it should be). but if it has no API the code is only visable to the users as a side effect of it's use. if the new code works better then it can be replaced. the scheduler change that we're going through right now is an example, new code came along that was better and the old code went away very quickly. the SLAB/SLOB/SLUB/S**B debate is another example. currently the different versions have different performance advantages and disadvantages, as one drops behind to the point where one of the others is better at all times, a large part of this is that it's so hard to get a replacement that works for userspace API's 10 years is reasonable, for internal features it's not. there is a LOT of internal stuff that was in the kernel 10 (or even 5) years ago that isn't there now. the key is that the behavior as far as users is concerned is better now. David Lang -
Swap Prefetch has existed since September 5, 2005. Please Nick, enlighten us all with your "alternatives" which have been offered (in practical, not theoretical form) in the past 23 months, along with their non-constructed benchmarks proving their case and the hordes of happy users and kernel developers who have tested them out the wazoo and given their backing. Or just take a nice steaming jug of STFU. -- Matt -
Heh. Here we have a VM developer expressing his interest in the problem space, and you offer him a steaming jug of STFU because he doesn't say what you want to hear. I wonder how many killfiles you just entered. -Mike -
Agreed. (a bit OT) People should understand that it's not (I think) about a desktop workload vs enterprise workloads war. I see it mostly as a progression versus regressions trade-off. And adding potentially useless or unmaintained code is a regression from the maintainers POV. The best way to justify a patch and have it integrated is to have a scientific testing method with repeatable numbers. Con has done so for his patch, his benchmark demonstrated good improvements. But I feel some of his supporters have indirectly harmed his cause by their comments. Also, the fact that Con recently stopped maintaining his work out of frustration also don't help having his patch merged. Again I'm not personally pushing this patch, I don't need it. Con has worked for many years on two area that still cause problems for desktop users: scheduler interactivity and pagecache trashing. Now that the scheduler has been fixed, let's have the VM fixed too. Sorry for the slightly OT post, and please don't start a flame war... - Eric -
The alternatives comment was in relation to the readahead based drop behind patch,for which an alternative would be improving use-once, possibly in the way I described. As for swap prefetch, I don't know, I'm not in charge of it being merged or not merged. I do know some people have reported that their updatedb problem gets much better with swap prefetch turned on, and I am trying to work on that too. For you? You also have the alternative to help improve things yourself, and you can modify your own kernel. -- SUSE Labs, Novell Inc. -
It is *not* about updatedb. That is just a trivial case which people notice. Therefore fixing updatedb to be nicer, as was discussed at various points in this thread, is *not* the solution. Most users are also *not*at*all* interested in kernel builds as a metric of system performance. When I'm at work, I run a large, commercial, engineering application. While running, it takes most of the system memory (4GB and up), and it reads and writes very large files. Swap prefetch noticeably helps my desktop too. Can I measure it? Not sure. Can people on lkml fix the application? Certainly not. Frank -
well, there appears to be a fairly large group of people who have subjective opinions that it helps them. but those were dismissed becouse they aren't measurements. so now the measurements of the constructed test case aren't acceptable. what sort of test case would be acceptable? David Lang -
Not at all. But there is also seems to be some people also experiencing problems with basic page reclaim on some of the workloads where these things help. I am not dismissing anybody's claims about anything; I want to try to solve some of these problems. Interestingly, some of the people ranting the most about how the VM sucks Well I never said real world tests aren't acceptable, they are. There is a difference between an "it feels better for me", and some actual real measurement and analysis of said workload. And constructed test cases of course are useful as well, I didn't say they weren't. I don't know what you mean by "acceptable", but you should read my last paragraph again. -- SUSE Labs, Novell Inc. -
On 25/07/07, Nick Piggin <nickpiggin@yahoo.com.au> wrote: Let me tell you about the use-case where swap prefetch helps me. I don't have actual numbers currently, only a subjective "it feels better", but when I get home from work tonight I'll try to collect some actual numbers for you. Anyway, here's a description of the scenario (machine is a AMD Athlon64 X2 4400+, 2GB RAM, 1GB swap, running 32bit kernel & userspace): A KDE desktop with the following running is common for me - A few (konsole) shells open running vim, pine, less, ssh sessions etc. - Eclipse (with CDT) with 20-30 files open in a project. - Firefox with 30+ tabs open. - LyX with a 200+ page document I'm working on open, is running. - Gimp running, usually with at least one or two images open (~1280x1024). - Amarok open and playing my playlist (a few days worth of music). - At least one Konqueror window in filemanager mode running. - More often than not OpenOffice is running with a spreadsheet or text document open. - In the background the machine is running Apache, MySQL, BIND and NFS services for my local LAN, but they see very little actual use. Now, a thing I commonly do is fire up a new shell, pull the latest changes from Linus' git tree and start a script running that builds a allnoconfig kernel, a allmodconfig kernel, a allyesconfig kernel and then 30 randconfig kernels. Obviously that script takes quite a while to run and loads the box quite a bit, so I usually just leave the box alone for a few hours until it is done (sometimes I leave it over night, in which case updatedb also gets added to the mix during the night). This usually pushes the box to use some amount of swap. Without swap prefetch; when I start working with one of the apps I had running before starting the compile job it always feels a little laggy at first. With swap prefetch app response time is not laggy when I come back. The "laggyness" doesn't last too long and is hard to quantify, but I'll try getting some numbers ...
this problem has been around for many years, with many different people working on solutions. it's hardly a case of getting a proposal and trying to get it in without anyone looking at other options. it seems that there are some people (not nessasarily including you) who will oppose this feature until a test is created that shows that it's better. the question is what sort of test will be accepted as valid? I'm not useing this patch, but it sounds as if the people who are useing it are interested in doing whatever testing is required, but so far the situation seems to be a series of "here's a test", "that test isn't valid, try again" loops. which don't seem to be doing anyone any good and are frustrating lots of people, so like several people over the last few days O'm asking the question, "what sort of test would be acceptable as proof that this patch does some good?" David Lang -
What is "this problem"? People have an updatedb problem that is solved by swap prefetching which I want to fix in a different way. There would be a different problem of "run something that uses heaps of memory and swap everything else out, then quit it, wait for a while, and swap prefetching helps". OK, definitely swap prefetching would help there. How much? I don't know. I'd be slightly surprised if it was like an order of magnitude, because not only swap but everything else has been thrown And yet despite my repeated pleas, none of those people has yet spent a I don't think any further proof is needed that the patch does "some" good. Rig up a test case and you could see some seconds shaved off it. Maybe you want to know "how to get this patch merged"? And I don't know that one. I do know that it is fuzzy, and probably doesn't include demanding things of Andrew or Linus. BTW. If you find out the answer to that one, let me know because I have this lockless pagecache patch that has also been around for years, is also just a few hundred lines in the VM, and does do some good too. I'm sure the buffered AIO people and many others would also like to know. -- SUSE Labs, Novell Inc. -
btw., it might help to give specific, precise instructions about what people should do to help you analyze this problem. Ingo -
Ray has been the first one to offer (thank you), and yes I have asked him for precise details of info to collect to hopefully work out what is happening with his first problem. For the general "it feels better for me" it is harder, but not as hard as CPU scheduler. We can measure various types of IO waits, swap in/out events, swap prefetch events and successfulness; see what happens to those as we change swappiness or vfs_cache_pressure etc. -- SUSE Labs, Novell Inc. -
do you mean this paragraph: | I guess /proc/meminfo, /proc/zoneinfo, /proc/vmstat, /proc/slabinfo | before and after the updatedb run with the latest kernel would be a | first step. top and vmstat output during the run wouldn't hurt either. correct? Does "latest kernel" mean v2.6.22.1, or does it have to be v2.6.23-rc1? I guess v2.6.22.1 would be fine as this is a VM problem, not a scheduling problem. the following script will gather all the above information for a 10 seconds interval: http://people.redhat.com/mingo/cfs-scheduler/tools/cfs-debug-info.sh Ray, please run this script before the updatedb run, once during the updatedb run and once after the updatedb run, and send Nick the 3 files it creates. (feel free to Cc: me too) Ingo -
OK fair point, but the counter point that there are real patterns that just use-once a lot of metadata (ls, for example. grep even.) -- SUSE Labs, Novell Inc. -
but if you do use locate then the alturnative becomes sitting around and waiting for find to complete on a regular basis. David Lang -
Yes, but what's locate's usage scenario? I've never, ever wanted to use it. When do you know the name of something but not where it's located, other than situations which "which" wouldn't cover and after just having installed/unpacked something meaning locate doesn't know about it yet either? Rene. -
I use it to find source files and documents all the time. One of my work boxes has <runs a locate work | wc -l> ~38500 files and directories under my source directory. And then there's the "I wrote that tech doc two years ago, where was that. Hmm, what did I name it? Bet it had 323 in the name, and doc in the path." I'd just like updatedb to amortize its work better. If we had some way to track all filesystem events, updatedb could keep a live and accurate index on the filesystem. And this isn't just updatedb that wants that, beagle and tracker et al also want to know filesystem events so that they can index the documents themselves as well as the metadata. And if they do it live, that spreads the cost out, including the VM pressure. Ray -
That would be nice. It'd be great if there was a per-filesystem inotify mode. I can't help but think it'd be more efficient than recursing every directory and adding a watch. Or maybe a netlink thing that could buffer events since filesystem mount until a daemon could get around to starting, so none were lost. --=20 Zan Lynx <zlynx@acm.org>
See "Filesystem Event Reporter" by Yi Yang, that does pretty much exactly that. http://lkml.org/lkml/2006/9/30/98 . Author had things to update, never resubmitted it as far as I can tell. Ray -
We already have this, its called inotify (and if I'm not mistaken, beagle already uses it). Several years ago when it was still a little flakey patch, I built a custom filesystem indexer into an enterprise search engine using it (I needed to pull apart Unix mbox files). The only trouble of course is the action is triggered immediately, which may not always be ideal (but that's a userspace problem) -- Matt -
Yeah, I know about inotify, but it doesn't scale. ray@phoenix:~$ find ~ -type d | wc -l 17933 ray@phoenix:~$ That's not fun with inotify, and that's just my home directory. The vast majority of those are quiet the vast majority of the time, which is the crux of the problem, and why inotify isn't a great fit for on-demand virus scanners or indexers. Ray -
Yeah, the nonrecursive behaviour is a bugger. Also I found it helped to queue operations in userspace and execute periodically rather than trying to execute on every single notification. Worked well for indexing, for virus scanning though you'd want to do some risk analysis. It'd be nice to have a filesystem that handled that sort of thing internally *cough*winfs*cough*. That was my hope for reiserfs a very long time ago with its pluggable fs modules feature. -- Matt -
With all this discussion about updatedb and locate and such, I thought I'd do a Google search, (considering I've never heard of locate before but I've seen updatedb here and there in ps lists) and I found this: http://www.linux.com/articles/114029 That page mentions something called "rlocate", which seems to provide some sort of almost-real-time mechanism, although the method it does so bothers me -- it uses a 2.6 kernel module AND a userspace daemon. And from what I can tell, there's no indication that this almost "real-time" (--I see mentions of a 2 second lag--) system replaces/eliminates updatedb in any way, shape, or form. http://rlocate.sourceforge.net/ - Project "Web Site" http://sourceforge.net/projects/rlocate/ - Source Forge Project Summary The last release also appears a bit dated on sourceforge... release 0.4.0 on 2006-01-15. Just thought I'd mention it. -- Michael Chang Please avoid sending me Word or PowerPoint attachments. Send me ODT, RTF, or HTML instead. See http://www.gnu.org/philosophy/no-word-attachments.html Thank you. -
My favorite use - with 5 Fedora kernels and as many -mm kernels on my laptop, doing a 'locate moby' finds all the moby.c and moby.o and moby.ko for the various releases. For bonus points, something like: ls -lt `locate iwl3945.ko` to find all 19 copies that are on my system, and remind me which ones were compiled when. Or just when you remember the name of some one-off 100-line Perl program that you wrote 6 months ago, but not sure which directory you left it in... ;) You want hard numbers? Here you go - 'locate' versus 'find' (/usr/src/ has about 290K files on it): % strace locate iwl3945.ko >| /tmp/foo3 2>&1 % wc /tmp/foo3 96 592 6252 /tmp/foo3 % strace find /usr/src /lib -name iwl3945.ko >| /tmp/foo4 2>&1 % wc /tmp/foo4 328380 1550032 15708205 /tmp/foo4 # echo 1 > /proc/sys/vm/drop_caches (to empty the caches % time locate iwl3945.ko > /dev/null real 0m0.872s user 0m0.867s sys 0m0.008s % time find /usr/src /lib -name iwl3945.ko > /dev/null find: /usr/src/lost+found: Permission denied real 1m12.241s user 0m1.128s sys 0m3.566s So 96 system calls in 1 second, against 328K calls in a minute. There's your use case, right there. Now if we can just find a way for that find/updatedb to not be as painful to the rest of the system.....
Supposing you know the path in one tree, you know the path in all of them, These are ofcourse not necesary. If you discount the time updatedb itself takes it's utterly obvious that _if_ you use it, it's going to be wildly faster than find. Regardless, I'll stand by "[by disabling updatedb] the problem will for a large part be solved" as I expect approximately 94.372 percent of Linux desktop users couldn't care less about locate. Rene. -
i think that approach is illogical: because Linux mis-handled a mixed workload the answer is to ... remove a portion of that workload? To bring your approach to the extreme: what if Linux sucked at running more than two CPU-intense tasks at once. Most desktop users dont do that, so a probably larger than 94.372 percent of Linux desktop users couldn't care less about a proper scheduler. Still, anyone who builds a kernel (the average desktop user wont do that) while using firefox will attest to the fact that it's quite handy that the Linux scheduler can handle mixed workloads pretty well. now, it might be the case that this mixed VM/VFS workload cannot be handled any more intelligently - but that wasnt your argument! The swap-prefetch patch certainly tried to do things more intelligently and the test-case (measurement app) Con provided showed visible improvements in swap-in latency. (and a good number of people posted those results) Ingo -
No. It got snipped but I introduced the comment by saying it was a "that's not the point" kind of thing. Sometimes things that aren't the point are still true though and in the case of Linux desktop users complaining about updatedb runs, a comment that says that for many an obvious solution would be to stop running the damned thing is not in any sense illogical. Also note I'm not against swap prefetch or anything. I don't use it and do not believe I have a pressing need for it, but do suspect it has potential to make quite a bit of difference on some things -- if only to drastically reduce seeks if it means it's swapping in larger chunks than a randomly faulting program would. Rene. -
ok - with that qualification i understand. still, especially for someone like me who frequently deals with source code, 'locate' is indispensible. and the fact is: updatedb discards a considerable portion of the cache completely unnecessarily: on a reasonably complex box no way do all the inodes and dentries fit into all of RAM, so we just trash everything. Maybe the kernel could be extended with a method of opening files in a 'drop from the dcache after use' way. (beagled and backup tools could make use of that facility too.) (Or some other sort of file-cache-invalidation syscall that already exist, which would _also_ result in the immediate zapping of the dentry+inode from the dcache.) Ingo -
I'm wondering how much of this updatedb problem is due to poor layout
of swap and other file systems across disk spindles.
I'll wager that those most impacted by updatedb have just one disk.
I have the following three boxes - three different setups, each with
different updatedb behaviour:
The first box, with 1 GB ram, becomes dog slow as soon as it
breaths on the swap device. Updatedb and backups are painful
intrusions on any interactive work on that system. I sometimes
wait a half minute for a response from an interactive application
anytime it has to go to disk. This box has a single disk spindle,
on an old cheap slow disk, with swap on the opposite end of the
disk from root and the main usr partition. It's a worst case
disk seek test device.
The second box, also with 1 GB ram, has multiple disk spindles,
and swap on its own spindle. I can still notice updatedb and
backup, but it's far far less painful.
The third box has dual CPU cores and 4 GB ram. Updatedb runs
over the entire system in perhaps 30 seconds with no perceptible
impact at all on interactive uses. Everything is still in memory
from the previous updatedb run; the disk is just used to write
out new stuff. Swap is never used on this (sweet) rig.
I'd think that prefetch would help in the single disk spindle
configuration, because it does the swap accesses separately, instead
of intermingling them with root or usr partition accesses, which
would require alot of disk head seeking.
Pretty much anytime that ordinary desktop users complain about
performance as much as they have about this one, it's either disk
head seeks or network delays. Nothing else is -that- slow, to be so
noticeable to so many users just doing ordinary work.
Question:
Could those who have found this prefetch helps them alot say how
many disks they have? In particular, is their swap on the same
disk spindle as their root and user files?
Answer - for ...On Wed, 2007-07-25 at 15:05 -0700, Paul Jackson wrote: A well done swap-to-flash would help here. I sometimes do it anyway to a 4GB CF card but I can tell it's hitting the read/update/write cycles on the flash blocks. The sad thing is that it is still a speed improvement over swapping to laptop disk. --=20 Zan Lynx <zlynx@acm.org>
Swap prefetch helps me. In my case I have a single (10K RPM, Ultra 160 SCSI) disk. # fdisk -l /dev/sda Disk /dev/sda: 36.7 GB, 36703918080 bytes 255 heads, 63 sectors/track, 4462 cylinders Units = cylinders of 16065 * 512 = 8225280 bytes Device Boot Start End Blocks Id System /dev/sda1 1 974 7823623+ 83 Linux /dev/sda2 975 1218 1959930 83 Linux /dev/sda3 1219 1341 987997+ 82 Linux swap /dev/sda4 1342 4462 25069432+ 83 Linux sda1 is "/", sda2 is "/usr/local/" and sda4 is "/home/" But, I don't think updatedb is the problem, at least not just updatedb on its own. My machine has 2GB of RAM, so a single updatedb on its own will not cause it to start swapping, but it does eat up a chunk of mem no doubt about that. The problem with updatedb is simply that it can be a contributing factor to stuff being swapped out, but any memory hungry application can do that - just try building an allyesconfig kernel and see how much the linker eats towards the end. What swap prefetch helps is not updatedb specifically, In my experience it helps any case where you have applications running, then start some memory hungry job that runs for a limited time, push the previously started apps out to swap and then dies (like updatedb or a compile job). Without swap prefetch those apps that were pushed to swap won't be brought back in before they are used (at which time the user is going to have to sit there and wait for them). With swap prefetch, the apps that got swapped out will slowly make their way back once the mem hungry app has died and will then be fully or partly back in memory when the user comes back to them. That's how swap prefetch helps, it's got nothing to do with updatedb as such - at least not as I see it. -- Jesper Juhl <jesper.juhl@gmail.com> Don't top-post http://www.catb.org/~esr/jargon/html/T/top-post.html Plain text mails only, ...
I have found that swap prefetch helped on all of the four machines machine I have, although the effect is more noticeable on machines with slower disks. They all have one hard disk, and root and swap were always on the same disk. I have no idea how to determine how many disk spindles they have, but since the drives are mainly low-end consumer models sold with low-end sub $500 PCs... -- Michael Chang Please avoid sending me Word or PowerPoint attachments. Send me ODT, RTF, or HTML instead. See http://www.gnu.org/philosophy/no-word-attachments.html Thank you. -
On both desktop and laptop. Cheers, -- []s, André Goddard -
The semantic that would benefit my work patterns the most would not be "immediate zapping" - I have 2G of RAM, so often there's no memory pressure, and often a 'find' will be followed by another similar 'find' that will hit a lot of the same dentries and inodes, so may as well save them if we can. Flagging it as "the first to be heaved over the side the instant there *is* pressure" would suit just fine. Or is that the semantic you actually meant?
Okay, but unless I've now managed to really quite horribly confuse myself, that wouldn't have anything to do with _swap_ prefetch would it? Rene. -
it's connected: it would remove updatedb from the VM picture altogether. (updatedb would just cycle through the files with leaving minimal cache disturbance.) hence swap-prefetch could concentrate on the cases where it makes sense to start swap prefetching _without_ destroying other, already cached content: such as when a large app exits and frees gobs of memory back into the buddy allocator. _That_ would be a definitive "no costs and side-effects" point for swap-prefetch to kick in, and it would eliminate this pretty artificial (and unnecessary) 'desktop versus server' controversy and would turn it into a 'helps everyone' feature. Ingo -
which only finds executables that are in the path. I commonly use locate to find config files (or sample config files) for packages that were installed at some point in the past with fairly default configs and now I want to go and tweak them. so I start reading documentation and then need to find out where $disto moved the files to this release (I commonly am working on machines with over a half dozen different distro releases, and none of them RedHat) David Lang -
updatedb isn't the only problem, its just an obvious one. I like the idea of looking into the vfs for this and other one-shot applications (rather than looking at updatedb itself specifically) Many modern applications have a lot of open file handles. For example, I just fired up my usual audio player and sys/fs/file-nr showed another 600 open files (funnily enough, I have roughly that many audio files :) I'm not exactly sure what happens when this one gets swapped out for whatever reason (firefox/java/vmware/etc chews ram, updatedb, whatever) but I'm fairly confident what happens between kswapd and the vfs and whatever else we're caching is not optimal come time for this process to context-switch back in. We're not running a highly-optimised number-crunching scientific app on desktops, we're running a full herd of poorly-coded hogs simultaneously through smaller pens. I don't think anyone is trying to claim that swap prefetch is the be all and end all of this problem's solution, however without it the effects are an order of magnitude worse (I've cited numbers elsewhere, as have several others); its relatively non-intrusive (600+ lines of the 755 changed ones are self-contained), is compile and runtime selectable, and still has a maintainer now that Con has retired. If there was a better solution, it should have been developed sometime in the past 23 months that swap prefetch has addressed it. That's how we got rmap versus aa, and so on. But nobody chose to do so, and continuing to hold out on merging it on the promise of vapourware is ridiculous. That has never been the way linux kernel development has operated. -- Matt -
That's the point, it is an obvious one. So it should be easy to work out why it is going wrong, and fix it. (And hopefully that fixes some And yet nobody wants to take the time to properly analyse why these things are going wrong and reporting their findings? Or if they have, where is that documented? -- SUSE Labs, Novell Inc. -
Hi Nick, I've attached two files with this kind of info. Being up at the cron hours of the morning meant I got a better picture of what my system is doing. Here's a short summary of what I saw in top: beagleindexer used gobs of ram. 600M or so (I have 1G) updatedb didn't use much ram, but while it was running kswapd kept on frequenting the top 10 cpu hogs - it would stick around for 5 seconds or so then disappear for no more than 10 seconds, then come back again. This behaviour persisted during the run. updatedb ran third (beagleindexer was first, then update-dlocatedb) I'm going to do this again, this time under a CFS kernel & use Ingo's sched_debug script to see what the scheduler is doing also. Let me know if there's anything else you wish to see. The running kernel at the time was 2.6.22.1-ck. There's no slabinfo since I'm using slub instead (and I don't have slub debug enabled). Cheers, -- Matt
Hmm OK, beagleindexer. I thought beagle didn't need frequent reindexing Kswapd will use CPU when memory is low, even if there is no swapping. Your "buffers" grew by 600% (from 50MB to 350MB), and slab also grew by a few thousand entries. This is not just a problem when it pushes out swap, it will also harm filebacked working set. This (which Ray's traces also show) is a bit of a problem. As Andrew noticed, use-once isn't working well for buffer cache, and it doesn't really for dentry and inode cache either (although those don't seem to be as much of a problem on your workload). Andrew has done a little test patch for this in -mm, but it probably wants more work and testing. If you can test the -mm kernel and see if things are improved, that would help. -- SUSE Labs, Novell Inc. -
One simple way to fix this would be to implement a fadvise() flag that puts the dentry/inode on a "soon to be expired" list if there are no other references. Then if a dentry allocation needs more memory try to reuse dentries from that list (or better queue) first. Any other access will remove the dentry from the list. Disadvantage would be that the userland would need to be patched, but I guess it's better than adding very dubious heuristics to the kernel. Similar thing could be done for directory buffers although they are probably less of a problem. I expect that C.Lameter's directed dentry/inode freeing in slub will also make a big difference. People who have problems with updatedb should definitely try mm which has it I believe and enable SLUB. -Andi (who always thought swap prefetch was just a workaround, not a real solution) -
Are you going to change every single large memory application in the world? As I wrote before, it is *not* about updatedb, but about all applications that use a lot of memory, and then terminate. Frank -
it is about multiple problems, _one_ problem is updatedb. The _second_ problem is large memory applications. note that updatedb is not a "large memory application". It simply scans through the filesystem and has pretty minimal memory footprint. the _kernel_ ends up blowing up the dentry cache to a rather large size (because it has no idea that updatedb uses every dentry only once). Once we give the kernel the knowledge that the dentry wont be used again by this app, the kernel can do a lot more intelligent decision and not baloon the dentry cache. ( we _do_ want to baloon the dentry cache otherwise - for things like "find" - having a fast VFS is important. But known-use-once things like the daily updatedb job can clearly be annotated properly. ) the 'large memory apps' are a second category of problems. And those are where swap-prefetch could indeed help. (as long as it only 'fills up' the free memory that a large-memory-exit left behind it.) the 'morning after' phenomenon that the majority of testers complained about will likely be resolved by the updatedb change. The second category is likely an improvement too, for swap-happy desktop (and server) workloads. Ingo -
Mutter. /proc/sys/vm/vfs_cache_pressure has been there for what, three years? Are any distros raising it during the updatedb run yet? -
but ... that's system-wide, and the 'dont baloon the dcache' is only a
property of updatedb. Still, it's useful to debug this thing.
below is an updatedb hack that sets vfs_cache_pressure down to 0 during
an updatedb run. Could someone who is affected by the 'morning after'
problem give it a try? If this works then we can think about any other
measures ...
Ingo
--- /etc/cron.daily/mlocate.cron.orig
+++ /etc/cron.daily/mlocate.cron
@@ -1,4 +1,7 @@
#!/bin/sh
nodevs=$(< /proc/filesystems awk '$1 == "nodev" { print $2 }')
renice +19 -p $$ >/dev/null 2>&1
+PREV=`cat /proc/sys/vm/vfs_cache_pressure 2>/dev/null`
+echo 0 > /proc/sys/vm/vfs_cache_pressure 2>/dev/null
/usr/bin/updatedb -f "$nodevs"
+[ "$PREV" != "" ] && echo $PREV > /proc/sys/vm/vfs_cache_pressure 2>/dev/null
-
Sure, but it's practical, isn't it? Who runs (and cares about) vfs-intensive workloads during their wee-small-hours updatedb run? Setting it to zero will maximise the preservation of the vfs caches. You wanted 10000 there. <bets that nobody will test this> -
drops caches prior to both updatedb runs. root@Homer: df -i Filesystem Inodes IUsed IFree IUse% Mounted on /dev/hdc3 12500992 1043544 11457448 9% / udev 129162 1567 127595 2% /dev /dev/hdc1 26104 87 26017 1% /boot /dev/hda1 108144 90676 17468 84% /windows/C /dev/hda5 11136 3389 7747 31% /windows/D /dev/hda6 0 0 0 - /windows/E vfs_cache_pressure=10000, updatedb freshly completed: procs -----------memory---------- ---swap-- -----io---- -system-- ----cpu---- r b swpd free buff cache si so bi bo in cs us sy id wa 1 0 48 76348 420356 104748 0 0 0 0 1137 912 3 1 97 0 ext3_inode_cache 315153 316274 524 7 1 : tunables 54 27 8 : slabdata 45182 45182 0 dentry_cache 224829 281358 136 29 1 : tunables 120 60 8 : slabdata 9702 9702 0 buffer_head 156624 159728 56 67 1 : tunables 120 60 8 : slabdata 2384 2384 0 vfs_cache_pressure=100 (stock), updatedb freshly completed: procs -----------memory---------- ---swap-- -----io---- -system-- ----cpu---- r b swpd free buff cache si so bi bo in cs us sy id wa 1 0 148 83824 270088 116340 0 0 0 0 1095 330 2 1 97 0 ext3_inode_cache 467257 502495 524 7 1 : tunables 54 27 8 : slabdata 71785 71785 0 dentry_cache 292695 408958 136 29 1 : tunables 120 60 8 : slabdata 14102 14102 0 buffer_head 118329 184384 56 67 1 : tunables 120 60 8 : slabdata 2752 2752 1 Note: updatedb doesn't bother my box, not running enough leaky apps I guess. -Mike -
I think that was the wrong thing to do. That will leave gobs of free memory for updatedb to populate with dentries and inodes. Instead, fill all of memory up with pagecache, then do the updatedb. See So you ended up with a couple hundred MB of pagecache preserved. Capturing before-and-after /proc/meminfo would be nice - it's a useful summary. -
Yeah. Before these two runs just to see what difference there was in caches with those two settings, I tried running with a heavier than normal (for me) desktop application mix, to see if it would start swapping, but it didn't. Seems that 1GB ram is enough space for everything I do, and everything updatedb does as well. You need a larger working set to feel the pain I guess. -Mike -
I didn't _fill_ memory, but loaded it up a bit with some real workload data... I tried time sh -c 'git diff v2.6.11 HEAD > /dev/null' to populate the cache, and tried different values for vfs_cache_pressure. Nothing prevented git's data from being trashed by updatedb. Turning the knob downward rapidly became very unpleasant due to swap, (with 0 not surprisingly being a true horror) but turning it up didn't help git one bit. The amount of data that had to be re-read with stock 100 or 10000 was the same, or at least so close that you couldn't see a difference in vmstat and wall-clock. Cache sizes varied, but the bottom line didn't. (wasn't surprised, seems quite reasonable that git's data looks old and useless to the reclaim logic when updatedb runs in between git runs) -Mike -
Did a bit of playing with this with 128MB of memory. - drop caches - read a 1MB file - run slocate.cron With vfs_cache_pressure=100: MemTotal: 116316 kB MemFree: 3196 kB Buffers: 54408 kB Cached: 5128 kB SwapCached: 0 kB Active: 41728 kB Inactive: 27540 kB HighTotal: 0 kB HighFree: 0 kB LowTotal: 116316 kB LowFree: 3196 kB SwapTotal: 1020116 kB SwapFree: 1019496 kB Dirty: 0 kB Writeback: 0 kB AnonPages: 9760 kB Mapped: 3808 kB Slab: 40468 kB SReclaimable: 34824 kB SUnreclaim: 5644 kB PageTables: 720 kB NFS_Unstable: 0 kB Bounce: 0 kB CommitLimit: 1078272 kB Committed_AS: 25988 kB VmallocTotal: 901112 kB VmallocUsed: 656 kB VmallocChunk: 900412 kB HugePages_Total: 0 HugePages_Free: 0 HugePages_Rsvd: 0 Hugepagesize: 4096 kB WIth vfs_cache_pressure=10000: MemTotal: 116316 kB MemFree: 3060 kB Buffers: 80792 kB Cached: 5052 kB SwapCached: 0 kB Active: 59432 kB Inactive: 36140 kB HighTotal: 0 kB HighFree: 0 kB LowTotal: 116316 kB LowFree: 3060 kB SwapTotal: 1020116 kB SwapFree: 1019512 kB Dirty: 0 kB Writeback: 0 kB AnonPages: 9756 kB Mapped: 3832 kB Slab: 14304 kB SReclaimable: 7992 kB SUnreclaim: 6312 kB PageTables: 732 kB NFS_Unstable: 0 kB Bounce: 0 kB CommitLimit: 1078272 kB Committed_AS: 26000 kB VmallocTotal: 901112 kB VmallocUsed: 656 kB VmallocChunk: 900412 kB HugePages_Total: 0 HugePages_Free: 0 HugePages_Rsvd: 0 Hugepagesize: 4096 kB so we reaped quite a lot more slab with the higher vfs_cache_pressure. What I think is killing us here is the blockdev pagecache: the ...
(may tinker a bit, but i'm way rusty. ain't had the urge to mutilate anything down there in quite a while... works just fine for me these Yeah, I mount noatime,nodiratime,data=writeback. ext3's journal with my crusty old disk/fs is painful as heck. -Mike -
So much for that theory. afaict mmapped, active pagecache is immune to updatedb activity. It just sits there while updatedb continues munching away at the slab and blockdev pagecache which it instantiated. I assume we're never getting the VM into enough trouble to tip it over the start-reclaiming-mapped-pages threshold (ie: /proc/sys/vm/swappiness). Start the updatedb on this 128MB machine with 80MB of mapped pagecache, it falls to 55MB fairly soon and then never changes. So hrm. Are we sure that updatedb is the problem? There are quite a few heavyweight things which happen in the wee small hours. -
The balance in _my_ world seems just fine. I don't let any of those system maintenance things run while I'm using the system, and it doesn't bother me if my working set has to be reconstructed after heavy-weight maintenance things are allowed to run. I'm not seeing anything I wouldn't expect to see when running a job the size of updatedb. -Mike -
Do you realize you've totally missed the point? It isn't about what is fine in the Kernel Developers world, but what is fine in the *USERS* world. There are dozens of big businesses pushing Linux for Enterprise performance. Rather than discussing the merit of those patches - some of which just improve the performance of a specific application by 1 or 2 percent - they get a nod and go into the kernel. But when a group of users that don't represent one of those businesses says "Hey, this helps with problems I see on my system" there is a big discussion and ultimately those patches get rejected. Why? Because they'll give an example using a program that they see causing part of the problem and be told "Use program X - it does things differently and shouldn't cause the problem" or "But what causes the problem to happen? The patch treats a symptom of a larger problem". The fucked up part of that is that the (mass of) kernel developers will see a similar report saying "mySQL has a performance problem because of X, this fixes it" and not blink twice - even if it is "treating the symptom and not the cause". It's this attitude more than anything that caused Con to "retire" - at least that is the impression I got from the interviews he's given. (The exact impression was "I'm sick of the kernel developers doing everything they can to help enterprise users and ignoring the home users") So... The problem: Updatedb or another process that uses the FS heavily runs on a users 256MB P3-800 (when it is idle) and the VFS caches grow, causing memory pressure that causes other applications to be swapped to disk. In the morning the user has to wait for the system to swap those applications back in. Questions about it: Q) Does swap-prefetch help with this? A) [From all reports I've seen (*)] Yes, it does. Q) Why does it help? A) Because it pro-actively swaps stuff back-in when the memory pressure that caused it to be swapped out is gone. Q) What causes the ...
Did you notice that I didn't make one disparaging remark about the patch or the concept behind it? Did you notice that I took _my time_ to test, to actually look at the problem? No, you're too busy running your mouth to appreciate the efforts of others. <snips load of useless spleen venting> Do yourself a favor, go dig into the VM source. Read it, understand it (not terribly easy), _then_ come back and preach to me. Have a nice day. -Mike -
If you're done being an ass, take note of the fact that I never even said you were doing that. What I was commenting on was the fact that you (and a lot of the other developers) seem to keep saying "It doesn't happen here, so it I've been trying to do that since the thread started. Note that you snipped where I said (and I'm going to paraphrase myself) "There is another way to fix this, but I don't have the understanding necessary". Now, once more, I'm going to ask: What is so terribly wrong with swap prefetch? Why does it seem that everyone against it says "Its treating a symptom, so it can't go in"? Try coming up with an answer that isn't "I don't see the problem on my $10K system" or similar - try explaining it based on the *technical* merits. Does it cause the processor cache to get thrashed? Does it create locking problems? I stand by my statements, as vitriolic as you and Rene seem to want to get over it. So far in this thread I have not seen one bit of *technical* I am. You being nasty when somebody gets fed up with a line of BS doesn't stop me from having a nice day. Only thing that could make my life any better would be to have the questions I've asked answered, rather than having supposedly intelligent people act like trolls. DRH -- Dialup is like pissing through a pipette. Slow and excruciatingly painful. -
And once again, I personally have nothing against swap-prefetch, or something like it. I can see how it or something like it could be made to improve the lives of people who get up in the morning to find their apps sitting on disk due to memory pressure generated by over-night system maintenance operations. The author himself however, says his implementation can't help with updatedb (though people seem to be saying that it does), or anything else that leaves memory full. That IMHO, makes it of questionable value toward solving what people are saying they want swap-prefetch for in the first place. I personally don't care if swap-prefetch goes in or not. -Mike -
Okay. I have to agree with the author that, in such a situation, it wouldn't help. However there are, without a doubt, other situations where it would help immensely. (memory hogs forcing everything to disk and quitting, one off tasks that don't balloon the cache (kernel compiles, et al) - in those situations swap prefetch would really shine.) DRH -- Dialup is like pissing through a pipette. Slow and excruciatingly painful. -
No it does not. If updatedb filled memory to the point of causing swapping (which noone is reproducing anyway) it HAS FILLED MEMORY and swap-prefetch hasn't any memory to prefetch into -- updatedb itself doesn't use any significant memory. Here's swap-prefetch's author saying the same: http://lkml.org/lkml/2007/2/9/112 | It can't help the updatedb scenario. Updatedb leaves the ram full and | swap prefetch wants to cost as little as possible so it will never | move anything out of ram in preference for the pages it wants to swap | back in. Now please finally either understand this, or tell us how we're wrong. Rene. -
I never said that it was the *program* itself - or *any* specific program (I used "Updatedb" because it has been the big name in the discussion) - doing the filling of memory. I actually said that the problem is that the kernel's caches - VFS and others - will grow *WITHOUT* *LIMIT*, filling all available memory. Swap prefetch on its own will not alleviate *all* of the problem, but it appears to fix enough of it that the problem doesn't seem to bother people anymore. (As I noted later on there are things that can be changes that would also fix things. Those changes, however, are quite tricky and involve changes to the page faulting mechanism, the way the various caches work and a number of other things) In light of the fact that swap prefetch appears to solve the problem for the people that have been vocal about it, and because it is a less intrusive change than the other potential solutions, I'd like to know why all the complaints and arguments against it come down to "Its treating the symptom". I mean it - because I fail to see how it isn't getting at the root of the problem - which is, pretty much, that Swap has classically been and, in the case of most modern systems, still is damned slow. By prefetching those pages that have most recently been evicted the problem of "slow swap" is being directly addressed. You want to know what causes the problem? The current design of the caches. They will extend without much limit, to the point of actually pushing pages I already did. You completely ignored it because I happened to use the magic words "updatedb" and "swap prefetch". Did I ever say it was about "updatedb" in particular? You've got the statement in the part of my post that you quoted. Nope, appears that I used the name as a specific example - and one that has been used previously in the thread. Now drop the damned attitude and start using your brain. Okay? DRH -- Dialup is like pissing through a pipette. Slow and excruciatingly ...
WHICH SWAP-PREFETCH DOES NOT HELP WITH. WHICH SWAP-PREFETCH DOES NOT HELP WITH. WHICH SWAP-PREFETCH DOES NOT HELP WITH. Due to being a generally nice guy, I am going to try _once_ more to try and make you understand. Not twice, once. So pay attention. Right now. Those caches are NOT causing any problem under discussion. If any caches grow to the point of causing swap-out, they have filled memory and swap-prefetch cannot and will not do anything since it needs free (as in not occupied by caches) memory. As such, people maintaining that swap-prefetch helps their situation are not being hit by caches. The only way swap-prefetch can (and will) do anything is when something that by itself takes up lots of memory runs and exits. So can we now please finally drop the fucking red herring and start talking about swap-prefetch? If we accept that some of the people maintaining that swap-prefetch helps them are not in fact deluded -- a bit of a stretch seeing as how not a single one of them is substantiating anything -- we have a number of slightly different possibilities for "something" in the above. -- 1) It could be an inefficient updatedb. Although he isn't experiencing the problem, Bjoern Steinbrink is posting numbers (weeee!) that show that at least the GNU version spawns a large memory "sort" process meaning that on a low-memory box updatedb itself can be what causes the observed problem. While in this situation switching to a different updatedb (slocate, mlocate) obviously makes sense it's the kind of situation where swap-prefetch will help. -- 2) It could be something else entirely such as a backup run. I suppose people would know if they were running anything of the sort though and wouldn't blaim anything on updatedb. Other than that, it's again the situation where swap-prefetch would help. -- 3) The something else entirely can also run _after_ updatedb, kicking out the VFS caches and leaving free memory upon exit. I still suppose the ...
Con might have been wrong there for boxes with really little memory. My desktop box has not even 300k inodes in use (IIRC someone posted a df -i output showing 1 million inodes in use). Still, the memory footprint of the "sort" process grows up to about 50MB. Assuming that the average filename length stays, that would mean 150MB for the 1 million inode case, just for the "sort" process. Now, sort cannot produce any output before its got all its input, so that RSS usage exists at least as long as the VFS cache is growing due to the ongoing search for files. And then, all that memory that "sort" uses is required, because sort needs to output its results. So if there's memory pressure, the VFS cache is likely to be dropped, because "sort" needs its data, for sorting and producing output. And then sort terminates and leaves that whole lot of memory _unused_. The other actions of updatedb only touch the locate db, which is just a few megs (4.5MB here) big so the cache won't grow that much again. OK, so we got about, say, at least 128MB of totally unused memory, maybe even more. If you look at the vmstat output I sent, you see that I had between 90MB and 128MB free, depending on the swappiness setting, with increased inode usage, that could very well scale up. Conclusion: updatedb does _not_ leave the RAM full. And for a box with little memory (say 256MB) it might even be 50% or more memory that is free after updatedb ran. Might that make swap prefetch kick in? Any faults in that reasoning? Thanks, Björn -
> Any faults in that reasoning? GNU sort uses a merge sort with temporary files on disk. Not sure how much it keeps in memory during that, but it's probably less than 150MB. At some point the dirty limit should kick in and write back the data of the temporary files; so it's not quite the same as anonymous memory. But it's not that different given. It would be better to measure than to guess. At least Andrew's measurements on 128MB actually didn't show updatedb being really that big a problem. Perhaps some people have much more files or simply a less efficient updatedb implementation? I guess the people who complain here that loudly really need to supply some real numbers. -Andi -
memory, only using a single file for some buffering purposes. When they get pushed out by a memory hog (either short term or long term) it takes several seconds for them to be swapped back in. (I'm on a P4-1.3GHz machine with 1G of ram and rarely run more than four programs (Mail Client, XChat, FireFox and a console window) and I've seen this lag in FireFox when switching to it after starting OOo. I've also seen the same sort of lag when exiting OOo. I agree. As I've said previously, it isn't updatedb itself which causes the problem. It's the way the VFS cache seems to just expand and expand - to the point of evicting pages to make room for itself. However, I may be wrong about that - I haven't actually tested it for myself, just looked at the Yes, it could be the proliferation of files. It could also be some other sort of problem that is exposing a corner-case in the VFS cache or the MM. I, personally, am of the opinion that it is likely the aforementioned corner case for people reporting the "updatedb" problem. If it is, then swap-prefetch is just papering over the problem. However I do not have the knowledge and understanding of the subsystems involved to be able to do much I've seen numerous "real numbers" posted about this. As was said earlier in the thread "every time numbers are posted they are claimed to be no good". But hey, nobodies perfect :) Anyway, the discussion seems to be turning to the technical merits of swap-prefetch... Now, a completely different question: During the research (and lots of thinking) I've been doing while this thread has been going on I've often wondered why swap prefetch wasn't already in the kernel. The problem of slow swap-in has long been known, and, given current hardware, the optimal solution would be some sort of data prefetch - similar to what is done to speed up normal disk reads. Swap prefetch looks like it does exactly that. The algo could be argued over and/or improved (to suggest ways to do that ...
Hm, does that change anything? The files need to be read at the end (so
they go into the cache) and are delete afterwards (cache gets freed I
Here's a before/after memory usage for an updatedb run:
root@atjola:~# free -m
total used free shared buffers cached
Mem: 2011 1995 15 0 269 779
-/+ buffers/cache: 946 1064
Swap: 1945 0 1945
root@atjola:~# updatedb
root@atjola:~# free -m
total used free shared buffers cached
Mem: 2011 1914 96 0 209 746
-/+ buffers/cache: 958 1052
Swap: 1945 0 1944
81MB more unused RAM afterwards.
If anyone can make use of that, here's a snippet from /proc/$PID/smaps
of updatedb's sort process, when it was at about its peak memory usage
(according to the RSS column in top), which was about 50MB.
2b90ab3c1000-2b90ae4c3000 rw-p 2b90ab3c1000 00:00 0
Size: 50184 kB
Rss: 50184 kB
Shared_Clean: 0 kB
Shared_Dirty: 0 kB
Private_Clean: 0 kB
Private_Dirty: 50184 kB
sort (GNU coreutils) 5.97
Just to clarify: I'm not complaining either way, neither about not
merging swap prefetch, nor about someone wanting that to be merge. It
was rather the "discussion" that caught my attention... Just in case ;-)
Björn
-
If I'm reading the source code for GNU sort correctly, then the
following snippet of shell code displays how much memory it uses
for its primary buffer on typical GNU/Linux systems:
head -2 /proc/meminfo | awk '
NR == 1 { memtotal = $2 }
NR == 2 { memfree = $2 }
END {
if (memfree > memtotal/8)
m = memfree
else
m = memtotal/8
print "sort size:", m/2, "kB"
}
'
That is, over simplifying, GNU sort looks at the first two entries
in /proc/meminfo, which for example on a machine near me happen to be:
MemTotal: 2336472 kB
MemFree: 110600 kB
and then uses one-half of whichever is -greater- of MemTotal/8 or
MemFree.
... However ... for the typical GNU locate updatedb run, it is sorting
the list of pathnames for almost all files on the system, which is
usually larger than fits in one of these sized buffers. So it ends up
using quite a few of the temporary files you mention, which tends to
chew up easily freed memory.
--
I won't rest till it's the best ...
Programmer, Linux Scalability
Paul Jackson <pj@sgi.com> 1.925.600.0401
-
Note -- with "the updatedb scenario" both he in the above and I are talking about the "VFS caches filling memory cause the problem" not updatedb in Even if it's not 150MB, 50MB is already a lot on a 128 or even a 256MB box. So, yes, we're now at the expected scenario of some hog pushing out things and freeing it upon exit again and it's something swap-prefetch definitely has potential to help with. Said early in the thread it's hard to imagine how it would not help in any such situation so that the discussion may as far as I'm concerned at that point concentrate on whether swap-prefetch hurts anything in others. Some people I believe are not convinced it helps very significantly due to at that point _everything_ having been thrown out but a copy of openoffice with a large spreadsheet open should come back to life much quicker it would No. If the machine goes idle after some memory hog _itself_ pushes things out and then exits, swap-prefetch helps, at the veryvery least potentially. By the way -- I'm unable to make my slocate grow substantial here but I'll try what GNU locate does. If it's really as bad as I hear then regardless of anything else it should really be either fixed or dumped... Rene. -
Yes. GNU locate is broken and nobody should be using it. The updatedb from (my distribution standard) "slocate" uses around 2M allocated total during an entire run while GNU locate allocates some 30M to the sort process alone. GNU locate is also close to 4 times as slow (although that ofcourse only matters on cached runs anyways). So, GNU locate is just a pig pushing things out, with or without any added VFS cache pressure from the things it does by design. As such, we can trust people complaining about it but should first tell them to switch to halfway sane locate implementation. If you run memory hogs on small memory boxes, you're going to suffer. Leaves the fact that swap-prefetch sometimes helps alleviate these and other kinds of memory-hog situations and as such, might not (again) be a bad idea in itself. Rene. -
however there are other programs which are known to take up significant amounts of memory and will cause the issue being described (openoffice for example) please don't get hung up on the text 'updatedb' and accept that there are programs that do run intermittently and do use a significant amount of ram and then free it. -
Different issue. One that's worth pursueing perhaps, but a different issue from the VFS caches issue that people have been trying to track down. Rene. -
people are trying to track down the problem of their machine being slow until enough data is swapped back in to operate normally. in at some situations swap prefetch can help becouse something that used memory freed it so there is free memory that could be filled with data (which is something that Linux does agressivly in most other situations) in some other situations swap prefetch cannot help becouse useless data is getting cached at the expense of useful data. nobody is arguing that swap prefetch helps in the second cast. what people are arguing is that there are situations where it helps for the first case. on some machines and version of updatedb the nighly run of updatedb can cause both sets of problems. but the nightly updatedb run is not the only thing that can cause problems but let's talk about the concept here for a little bit the design is to use CPU and I/O capacity that's otherwise idle to fill free memory with data from swap. pro: more ram has potentially useful data in it con: it takes a little extra effort to give this memory to another app (the page must be removed from the list and zeroed at the time it's needed, I assume that the data is left in swap so that it doesn't have to be written out again) it adds some complexity to the kernel (~500 lines IIRC from this thread) by undoing recent swapouts it can potentially mask problems with swapout it looks to me like unless the code was really bad (and after 23 months in -mm it doesn't sound like it is) that the only significant con left is the potential to mask other problems. however there are many legitimate cases where it is definantly dong the right thing (swapout was correct in pushing out the pages, but now the cause of that preasure is gone). the amount of benifit from this will vary from situation to situation, but it's not reasonable to claim that this provides no benifit (you have benchmark numbers that show it in synthetic benchmarks, and you ...
Actually, I made a mistake when tracking the thread and reading the code for the patch and started to argue just that. But I have to admit I made a mistake - the patches author has stated (as Rene was kind enough to point Solving the cache filling memory case is difficult. There have been a number of discussions about it. The simplest solution, IMHO, would be to place a (configurable) hard limit on the maximum size any of the kernels caches can I'll second this. But with the swap system itself having seen as heavy testing as it has I don't know if it would be masking other problems. That is why I've been asking "What is so wrong with it?" - while it definately doesn't help with programs that cause caches to balloon (that problem does need another solution) it does help to speed things up when a memory hog has exited. (And since its a pretty safe assumption that swap is going to be noticeably slower than RAM this patch seems to me to be a rather visible and Exactly. Though I have seen posts which (to me at least) appear to claim exactly that. It was part of the reason why I got a bit incensed. (The other was that it looked like the kernel devs with the ultra-powerful machines were claiming 'I don't see the problem on my machine, so it doesn't exist'. That sort of attitude is fine, in some cases, but not, IMHO, where performance is Yep. And it's a pretty obvious step forward. The VFS system already does readahead and caching for mounted volumes to improve performance - why not do similar to improve the performance of swap? The only real downside is that swap-prefetch won't be effective in all cases and it will cause some extra power consumption. (drives can't spin-down as soon as the would without it, etc...) While I can only make some suggestions as to how to fix the problem of ballooning caches (I've been wading through the VM code for a few days now and still don't fully understand any of it), the solution to the power consumption seems obvious ...
limiting the size of the cache is also the wrong thing to do in many situations. it's only right if the cache pushes out other data you care about, if you are trying to do one thing as fast as you can you really do want the system to use all the memory it can for the cache. David Lang -
If you had made the statement before I decided to speak up you would have been correct :) Anyway, I try to always admit when I've made a mistake - its part of my philosophy. (There have been times when I haven't done it, but I'm trying to After thinking about this you are partially correct. There are those sorts of situations where you want the system to use all the memory it can for caches. OTOH, if those situations could be described in some sort of simple heuristic, then a soft-limit that uses those heuristics to determine when to let the cache expand could exploit the benefits of having both a limited and unlimited cache. (And, potentially, if the heuristic has allowed a cache to expand beyond the limit then, when the heuristic no longer shows the oversize cache is no longer necessary it could trigger and automatic reclaim of that memory.) (I'm willing to help write and test code to do this exactly. There is no guarantee that I'll be able to help with more than testing - I don't understand the parts of the code involved all that well) DRH -- Dialup is like pissing through a pipette. Slow and excruciatingly painful. -
Oh yes they are. Daniel for example did twice, telling me to turn my brain on in between (if you read it, you may have noticed I got a little annoyed Not to sound pretentious or anything but I assume that Andrew has a fairly good overview of exactly how broken -mm can be at times. How many -mm users use it anyway? He himself said he's not convinced of usefulness having not seen it help for him (and notice that most developers are also users), turned it off due to it annoying him at some point and hasn't seen a serious Which is not a madeup issue, mind you. As an example, I just now tried GNU locate and saw it's a complete pig and specifically unsuitable for the low memory boxes under discussion. Upon completion, it actually frees enough memory that swap-prefetch _could_ help on some boxes, while the real issue I certainly would not want to argue anything of the sort no. As said a few times, I agree that swap-prefetch makes sense and has at least the potential to help some situations that you really wouldnt even want to try and fix any Well, _that_ is what the kernel is already going to great lengths at doing, and it decided that those pages us poor overnight OO.o users want in in the morning weren't reasonable guesses. The kernel also won't any time soon be reading our minds, so any solution would need either user intervention (we could devise a way to tell the kernel "hey ho, I consider these pages to be very important -- try not to swap them out" possible even with a "and if you do, please pull them back in when possible") or we can let swap-prefetch do the "just in case" thing it is doing. While swap-prefetch may not be the be all end all of solutions I agree that having a machine sit around with free memory and applications in swap seems not too useful if (as is the case) fetched pages can be dropped immediately when it turns out swap-prefetch made the wrong decision. So that's for the concept. As to implementation, if I try and look at the ...
> It is. Prefetched pages can be dropped on the floor without additional I/O. Which is essentially free for most cases. In addition your disk access may well have been in idle time (and should be for this sort of stuff) and if it was in the same chunk as something nearby was effectively free anyway. Actual physical disk ops are precious resource and anything that mostly reduces the number will be a win - not to stay swap prefetch is the right answer but accidentally or otherwise there are good reasons it may happen to help. Bigger more linear chunks of writeout/readin is much more important I I've been using it for months with no noticed problem. I turn it on because it might as well get tested. I've not done comparison tests so I can't comment on if its worth it. Lots of -mm testers turn *everything* on because its a test kernel. -
<nod>. The larger the chunks are that we swap out, the less it actually hurts to swap, which might make all this a moot point. Not all I/O is created equal... Ray -
Yes. The swap-prefetch patch ensures that the machine (well, the VM) is very Yes, I believe this might be an important point. Earlier I posted a dumb little VM thrasher: http://lkml.org/lkml/2007/7/25/85 Contrived thing and all, but what it does do is show exactly how bad seeking all over swap-space is. If you push it out before hitting enter, the time it takes easily grows past 10 minutes (with my 768M) versus sub-second (!) when it's all in to start with. What are the tradeoffs here? What wants small chunks? Also, as far as I'm aware Linux does not do things like up the granularity when it notices it's Okay. Rene. -
Small chunks means you get better efficiency of memory use - large chunks mean you may well page in a lot more than you needed to each time (and cause more paging in turn). Your disk would prefer you fed it big linear I/O's - 512KB would probably be my first guess at tuning a large box under load for paging chunk size. More radically if anyone wants to do real researchy type work - how about log structured swap with a cleaner ? Alan -
That probably kills my momentary hope that I was looking at yet another good use of large soft-pages seeing as how 512K would be going overboard a bit Right over my head. Why does log-structure help anything? Rene. -
Log structured disk layouts allow for better placement of writeout, so that you cn eliminate most or all seeks. Seeks are the enemy when trying to get full disk bandwidth. google on log structured disk layout, or somesuch, for details. Ray -
I understand what log structure is generally, but how does it help swapin? Rene. -
Look at the swap out case first. Right now, when swapping out the kernel places whatever it can wherever it can inside the swap space. The closer you are to filling your swap space, the more likely that those swapped out blocks will be all over place, rather than in one nice chunk. Contrast that with a log structured scheme, where the writeout happens to sequential spaces on the drive instead of scattered about. So, at some point when the system needs to fault those blocks that back in, it now has a linear span of sectors to read instead of asking the drive to bounce over twenty tracks for a hundred blocks. So, it eliminates the seeks. My laptop drive can read (huh, how odd, it got slower, need to retest in single user mode), hmm, let's go with about 25 MB/s. If we ask for a single block from each track, though, that'll drop to 4k * (1 second / seek time) which is about a megabyte a second if we're lucky enough to read from consecutive tracks. Even worse if it's not. Seeks are the enemy any time you need to hit the drive for anything, be it swapping or optimizing a database. Ray -
If the problem is reading stuff back in from swap quickly when
needed, then this likely helps, by reducing the seeks needed.
If the problem is reading stuff back in from swap at the *same time*
that the application is reading stuff from some user file system, and if
that user file system is on the same drive as the swap partition
(typical on laptops), then interleaving the user file system accesses
with the swap partition accesses might overwhelm all other performance
problems, due to the frequent long seeks between the two.
In that case, swap layout and swap i/o block size are secondary.
However, pre-fetching, so that swap read back is not interleaved
with application file accesses, could help dramatically.
===
Perhaps we could have a 'wake-up' command, analogous to the various sleep
and hibernate commands. The 'wake-up' command could do whatever of the
following it knew to do, in order to optimize for an anticipated change in
usage patterns:
1) pre-fetch swap
2) clean (write out) dirty pages
3) maximize free memory
4) layout swap nicely
5) pre-fetch a favorite set of apps
Stumble out of bed in the morning, press 'wake-up', start boiling the
water for your coffee, and in another ten minutes, one is ready to rock
and roll.
In case Andrew is so bored he read this far -- yes this wake-up sounds
like user space code, with minimal kernel changes to support any
particular lower level operation that we can't do already.
--
I won't rest till it's the best ...
Programmer, Linux Scalability
Paul Jackson <pj@sgi.com> 1.925.600.0401
-
Ah, so in a normal scenario where a working-set is getting faulted back in, we have the swap storage as well as the file-backed stuff that needs to be read as well. So even if swap is organized perfectly, we're still seeking. Damn. On the other hand, that explains another thing that swap prefetch could be helping with -- if it preemptively faults the swap back in, then the file-backed stuff can be faulted back more quickly, just by the virtue of not needing to seek back and forth to swap for its stuff. Hadn't thought of that. That also implies that people running with swap files rather than swap partitions will see less of an issue. I should dig out my old compact He'd suggested using, uhm, ptrace_peek or somesuch for just such a purpose. The second half of the issue is to know when and what to target. Ray -
Perhaps this applies in some cases ... perhaps.
--
I won't rest till it's the best ...
Programmer, Linux Scalability
Paul Jackson <pj@sgi.com> 1.925.600.0401
-
Yeah, point taken: better data would make this a lot easier to figure out and target fixes. Ray -
That is one reason why I try to have swap on a device dedicated just for it. It helps keep the system from having to seek all over the drive for data. (I remember that this was recommended years ago with Windows - back when you For it to really help swap-prefetch would have to be more aggressive. At the moment (if I'm reading the code correctly) the system has to have close to zero for it to kick in. A tunable knob controlling how much activity is too much for the prefetch to kick in would help with finding a sane default. IMHO it should be the one that provides the most benefit with the least hit to Maybe. It all depends on how much seeking is needed to track down the pages in the swapfile and such. What would really help make the situation even better would be doing the log structured swap + cleaner. The log structured swap + cleaner should provide a performance boost by itself - add in the prefetch mechanism and the benefits are even more visible. Another way to improve performance would require making the page replacement mechanism more intelligent. There are bounds to what can be done in the kernel without negatively impacting performance, but, if I've read the code correctly, there might be a better way to decide which pages to evict. One way to do this would be to implement some mechanism that allows the system to choose a single group of contiguous pages (or, say, a large soft-page) over swapping out a single page at a time. (some form of memory defrag would also be nice, but I can't think of a way to do that without massively breaking everything) The userspace suggestion that was thrown out earlier would have been as error-prone and problematic as FUSE. A solution like you suggest would be workable - its small and does a task that is best done in userspace (IMHO). (IIRC, the original suggestion involved merging maps2 and another patchset into mainline and using that, combined with PEEKTEXT to provide for a userspace swap daemon. Swap, ...
This seems to be now fixing the different problem of swap-space filling up. Moreover though -- what I know about log structure is that generally it optimises for write (swapout) and might make read (swapin) worse due to fragmentation that wouldn't happen with a regular fs structure. I guess that cleaner that Alan mentioned might be involved there -- I don't I am very aware of the costs of seeks (on current magnetic media). Rene. -
I was trying to point out that currently, as an example, memory that is linear in a process' space could be fragmented on disk when swapped out. That's today. Under a log-structured scheme, one could set it up such that something that was linear in RAM could be swapped out linearly on the drive, minimizing seeks on writeout, which will naturally minimize seeks on It looks like I'm not doing a very good job of explaining this, I'm afraid. Suffice it to say that a log structured swap would give optimization Then you should google on `log structured filesystem (primer OR introduction)` and read a few of the links that pop up. You might find Then perhaps you can just take it on faith -- log structured layouts are designed to help minimize seeks, read and write. Ray -
I am particularly bad at faith. Let's take that stupid program that I posted: http://lkml.org/lkml/2007/7/25/85 You push it out before you hit enter, it's written out to swap, at whatever speed. How should it be layed out so that it's swapped in most efficiently after hitting enter? Reading bigger chunks would quite obviously help, but the layout? The program is not a real-world issue and if you do not consider it a useful boundary condition either (okay I guess), how would log structured swap help if I just assume I have plenty of free swap to begin with? Rene. -
You only think you are :-). I'm sure there are lots of things you have Is that generally the case on your systems? Every linux system I've run, regardless of RAM, has always pushed things out to swap. And once there's something already in swap, you now have a packing problem when you want to swap something else out. -
For me, it is generally the case yes. We are still discussing this in the context of desktop machines and their problems with being slow as things have been swapped out and generally I expect a desktop to have plenty of swap which it's not regularly going to fillup significantly since then the Once we're crammed, it gets to be a different situation yes. As far as I'm concerned that's for another thread though. I'm spending too much time on LKML as it is... Rene. -
A simple log optimises writeout (which is latency critical) and can otherwise stall an enitre system. In a log you can also have multiple copies of the same page on disk easily, some stale - so you can write out chunks of data that are not all them removed from memory, just so you get them back more easily if you then do (and I guess you'd mark them accordingly) The second element is a cleaner - something to go around removing stuff from the log that is needed when the disks are idle - and also to repack data in nice linear chunks. So instead of using the empty disk time for page-in you use it for packing data and optimising future paging. -
<Shrug> Well, that doesn't match my systems. My laptop has 400MB in swap:
ray@phoenix:~$ free
total used free shared buffers cached
Mem: 894208 883920 10288 0 3044 163224
-/+ buffers/cache: 717652 176556
No, it's not even when crammed. It's just when there are holes.
mm/swapfile.c does try to cluster things, but doesn't work too hard at
it as we don't want to spend all our time looking for a perfect fit
that may not exist.
Ray
-
Which in your case is slightly more than 1/3 of available swap space. Quite a lot for a desktop indeed. And if it's more than a few percent fragmented, please fix current swapout instead of log structuring it. Rene. -
I'm sure this is true while you are doing the swapout or swapin and the system is waiting for it. but with prefetch you may be able to avoid doing the swapin at a time when the system is waiting for it by doing it at a time when the system is otherwise idle. David Lang -
if that was the case then people should be responding to the request to get it merged with 'but it caused problems for me when I tried it' I see the conclusion as being exactly the opposite. here is a workload with some badly designed userspace software that the kernel can make much more pleasent for users. arguing that users should never use badly designed software in userspace doesn't seem like an argument that will gain much traction. I'm not saying the kernel needs to fix the software itself (ala the sched_yeild issues), but the kernel should try and keep such software from hurting the rest of the system where it can. in this case it can't help it while the bad software is running, but it so there is a legitimate situation where swap-prefetch will help significantly, what is the downside that prevents it from being included? (reading this thread it sometimes seems like the downside is that updatedb shouldn't cause this problem and so if you fixed updatedb there wold be no legitimate benifit, or alturnatly this patch doesn't help updatedb so it's not that they shouldn't have been swapped out (they should have I've seen it mentioned that there is still a maintainer but I missed who it is, but I haven't seen any concerns that can be addressed, they all seem to be 'this is a core concept, people need to think about it' or 'but someone may find a better answer in the future' type of things. it's impossible to address these concerns directly. David Lang -
So you're saying Andrew did not say that? You're jumping to the conclusion And now you do it again :-) There is no conclusion -- just the inescapable observation that swap-prefetch was (or may have been) masking the problem of People being unconvinced it helps all that much, no serious investigation into possible downsides and no consideration of alternatives is three I've personally heard. You don't want to merge a conceptually core VM feature if you're not really convinced. It's not a part of the kernel you can throw a feature into like you could some driver saying "ah, heck, if it makes someone happy" since everything in the VM ends up interacting -- that in fact is actually the hard part of VM as far as I've seen it. And in this situation the proposed feature is something that "papers over a problem" by design -- where it could certainly be that the problem is not solveable in another way simply due to the kernel not growing the possiblity to read user's minds anytime soon (which some might even like to rephrase as "due to no problem existing") but that this gets people a bit anxious is not So do it indirectly. But please don't just say "it help some people (not me mind you!) so merge it and if you don't it's all just politics and we can't do anything about it anyway". Because that's mostly what I've been hearing. And no, I'm not subscribed to any ck mailinglists nor do I hang around its IRC community which will can account for part of that. I expect though that the same holds for the people that actually matter in this, such as Andrew Morton and Nick Piggin. -- 1: people being unconvinced it helps all that much At least partly caused by the updatedb i/dcache red herring that infected this issue. Also, at the point VM pressure has mounted high enough to cause enough to be swapped out to give you a bad experience, a lot of other things have been dropped already as well. It's unsurprising though that it would for example help the issue ...
I don't remember anyone saying that it actually caused problems (including both you and andrew). I (and others) have been trying to learn what problems people believe it has in the hope that they can be addressed one isn't your conclusion then that if people just stopped useing that version of updatedb the problem would be solved and there would be no need for the swap prefetch patch? that seemed to be what you were strongly implying (if people who have lots of memory and so don't use swap will never see the benifit of this patch. over the years many people have investigated the problem and tried to address it in other ways (the better version of updatedb is an attempt to fix it for that program as an example), but there is still a problem. I agree that tinkering with the core VM code should not be done lightly, but this has been put through the proper process and is stalled with no forget the nightly cron jobs for the moment. think of this scenerio. you have your memory fairly full with apps that you have open (including firefox with many tabs), you receive a spreadsheet you need to look at, so you fire up openoffice to look at it. then you exit openoffice and try to go back to firefox (after a pause while you walk to the printer to get the printout of the spreadsheet), only to find that it's going to be sluggish becouse it got swapped out due to the preasure from openoffice. no nightly cron job needed, just enough of a memory hog or a small enough larger swap granularity may help, but waiting for the user to need the ram and have to wait for it to be read back in is always going to be worse for the user then pre-populating the free memory (for the case where the pre-population is right, for other cases it's the same). so I see this as there are fully legitimate situations where this is useful, the 'papering over' effect is not referring to these, it's referring to other possible problems in the future. I see this argument as being in the ...
No. What I said outright, every single time, is that swap-prefetch in itself seems to make sense. And specifically that even if the _direct_ problem is a crummy program, it _still_ makes sense generally. Every single time. But see -- you failed to notice this because you guys are stuck in this dumb adversary "us against them" thing so inherent of (online) communities, where you sit around your own habitats patting each other on the back for extended periods of time and then every once a while go out clinging on to each other vigorously and going "boo! hiss!" at the big bad outside world. I already got overly violent at one point in this thread so I'll leave out any further references to sense-deprived fanboy-culture but please, I said every single time that I'm not against swap-prefetch. I cannot communicate It has not. Concerns that were raised (by specifically Nick Piggin) weren't And swinging a dead rat from its tail facing east-wards while reciting Documentation/CodingStyle. Okay, very very sorry, that was particularly childish, but that "walking to the printer" is ofcourse completely constructed and this _is_ something to take into account. Swap-prefetch wants to be free, which (also again) it is doing a good job at it seems, but this also means that it waits for the VM to be _very_ idle before it does anything and as such, we cannot just forget the "nightly" scenario and pretend it's about something else entirely. As long as the machine's being used, swap-prefetch doesn't kick in. Which is a good feature for swap-prefetch, but also something that needs to weighed alongside its other features in a discussion of alternatives, where for example something like a larger swap granularity would not have anything of the sort to take into account. If it were about walks to the printer, we Arjan van de Ven made another point here about seeking away due to swap-prefetch (just) before the next request comes in, but that's probably a I saw Chris ...
I may have missed them, but what I saw from him weren't specific issues, yes it was contrived for simplicity. the same effect would happen if instead of going back to firefox the user instead went to their e-mail software and read some mail. doing so should how long does the machine need to be idle? if someone spends 30 seconds reading an e-mail that's an incredibly long time for the system and I swapin will always require disk access, and avoiding doing disk access while the user is waiting for it by doing it when the system isn't useing the disk will always be a win (possibly not as large of a win, but still a win) on slow laptop drives where you may only get 20MB/second of reads under optimal situations it doesn't take much reading to be noticed by the and these thing do not conflict with prefetch, they compliment it. improved use-once will avoid pushing things out to swap in the first place. this will help during normal workloads so is valuble in any case. better swapin (I assume you are talking about things like larger swap granularity) will also help during normal workloads when you are thrashing into swap. prefetch will help when you have pushed things out to swap and now have free memory and a momentarily idle system. David Lang -
Something better, ie. the problems with page reclaim being fixed. Why is that nebulous? -- SUSE Labs, Novell Inc. -
becouse that doesn't begin to address all the benifits. the approach of fixing page reclaim and updatedb is pretending that if you only do everything right pages won't get pushed to swap in the first place, and therefor swap prefetch won't be needed. this completely ignores the use case where the swapping was exactly the right thing to do, but memory has been freed up from a program exiting so that you couldnow fill that empty ram with data that was swapped out. David Lang -
What do you mean "address the benefits"? What I want
You should read what I wrote.
Anyway, the fact of the matter is that there are still
fairly significant problems with page reclaim in this
workload which I would like to see fixed.
I personally still think some of the low hanging fruit
*might* be better fixed before swap prefetch gets
merged, but I've repeatedly said I'm sick of getting
dragged back into the whole debate so I'm happy with
whatever Andrew decides to do with it.
I think it is sad to turn it off for laptops, if it
really makes the "desktop" experience so much better.
Surely for _most_ workloads we should be able to
Yeah. However, merging patches (especially when
changing heuristics, especially in page reclaim) is
not about just thinking up a use-case that it works
well for and telling people that they're putting their
heads in the sand if they say anything against it.
Read this thread and you'll find other examples of
patches that have been around for as long or longer
and also have some good use-cases and also have not
been merged.
____________________________________________________________________________________
Yahoo!7 Mail has just got even bigger and better with unlimited storage on all webmail accounts.
http://au.docs.yahoo.com/mail/unlimitedstorage.html
-
On 8/6/07, Nick Piggin <nickpiggin@yahoo.com.au> wrote: What do you think Andrew? Swap prefetch is not the panacea, it's not going to solve all the problems but it seems to improve the "desktop experience" and it has been discussed and reviewed a lot (it's has even been discussed more than it should have be). Are you going to push upstream the patch? Ciao, -- Paolo http://paolo.ciarrocchi.googlepages.com/ -
I wonder what happens if you try that on ext2. There we'd get directory contents in per-directory page cache, so the picture might change... -
afacit ext2 just forgets to run mark_page_accessed for directory pages altogether, so it'll be equivalent to ext3 with that one-liner, I expect. The directory pagecache on ext2 might get reclaimed faster because those pages are eligible for reclaiming via the reclaim of their inodes, whereas ext3's directories are in blockdev pagecache, for which the reclaim-via-inode mechanism cannot happen. I should do some testing with mmapped files. -
Good idea for updatedb. However, it may be a bad idea for files that are often written to. Turning an inode write into a read plus a write does not sound like such a hot idea, we really want to keep those in the cache. I think what you need is to ignore multiple references to the same page when they all happen in one time interval, counting them only if they happen in multiple time intervals. The use-once cleanup (which takes a page flag for PG_new, I know...) would solve that problem. However, it would introduce the problem of having to scan all the pages on the list before a page becomes freeable. We would have to add some background scanning (or a separate list for PG_new pages) to make the initial pageout run use an acceptable amount of CPU time. Not sure that complexity will be worth it... -- Politics is the struggle between those who want to make their country the best in the world, and those who believe it already is. Each group calls the other unpatriotic. -
Remember that this problem applies to both inode blocks and to directory blocks. Yes, it might be useful to hold onto an inode block for a future Yes, the sudden burst of accesses for adjacent inode/dirents will be a common pattern, and it'd make heaps of sense to treat that as a single touch. It'd have to be done in the fs I guess, and it might be a bit hard to do. And it turns out that embedding the touch_buffer() all the way down in __find_get_block() was convenient, but it's going to be tricky to change. For now I'm fairly inclined to just nuke the touch_buffer() on the read side and maybe add one on the modification codepaths and see what happens. I suspect that the situation we have now is so bad that pretty much anything we do will be an improvement. I've always wondered "ytf is there so much blockdev pagecache?" This machine I'm typing at: MemTotal: 3975080 kB MemFree: 750400 kB Buffers: 547736 kB Cached: 1299532 kB SwapCached: 12772 kB Active: 1789864 kB Inactive: 861420 kB HighTotal: 0 kB HighFree: 0 kB LowTotal: 3975080 kB LowFree: 750400 kB SwapTotal: 4875716 kB SwapFree: 4715660 kB Dirty: 76 kB Writeback: 0 kB Mapped: 638036 kB Slab: 522724 kB CommitLimit: 6863256 kB Committed_AS: 1115632 kB PageTables: 14452 kB VmallocTotal: 34359738367 kB VmallocUsed: 36432 kB VmallocChunk: 34359696379 kB HugePages_Total: 0 HugePages_Free: 0 HugePages_Rsvd: 0 Hugepagesize: 2048 kB More that a quarter of my RAM in fs metadata! Most of it I'll bet is on the active list. And the fs on which I do most of the work is mounted noatime.. -
wrong, it's active on three of my boxes already :) But then again, i
never had these hangover problems. (not really expected with gigs of RAM
anyway)
Ingo
--- /etc/cron.daily/mlocate.cron.orig
+++ /etc/cron.daily/mlocate.cron
@@ -1,4 +1,7 @@
#!/bin/sh
nodevs=$(< /proc/filesystems awk '$1 == "nodev" { print $2 }')
renice +19 -p $$ >/dev/null 2>&1
+PREV=`cat /proc/sys/vm/vfs_cache_pressure 2>/dev/null`
+echo 10000 > /proc/sys/vm/vfs_cache_pressure 2>/dev/null
/usr/bin/updatedb -f "$nodevs"
+[ "$PREV" != "" ] && echo $PREV > /proc/sys/vm/vfs_cache_pressure 2>/dev/null
-
[...] mlocate by design doesn't thrash the cache as much. People using slocate (distros other than Redhat ;) are going to be hit worse. See http://carolina.mff.cuni.cz/~trmac/blog/mlocate/ updatedb by itself doesn't really bug me, its just that on occasion its still running at 7am which then doesn't assist my single spindle come swapin of the other apps! I'm considering getting one of the old ide drives out in the garage and shifting swap onto it. The swap prefetch patch has mainly assisted me in the "state A -> B -> A" scenario. A lot. -- Matt -
You should start it earlier then - assuming it doesn't already start at the earliest opportunity? Helge Hafting -
there's another side-effect: it likely results in the zapping of thousands of dentries that were cached nicely before. So we might exchange 'all my apps are swapped out' experience with 'all file access is slow'. The latter is _probably_ still an improvement over the balooning, but i'm not sure. What we _really_ want is an updatedb that does not disturb the dcache. Ingo -
Yup. Nobody has begun to think about preserving dcache/icache across load Well. Hopefully this time next year you can prep a 16MB container and toss your updatedb inside that. Maybe set its peak disk bandwidth utilisation too. However that won't work ;) because I don't think anyone is looking at containerisation of vfs cache memory yet. Perhaps full-on openvz has it, dunno. But updatedb is a special case, because it is so vfs-intensive. For lots of other workloads (those which use heaps of pagecache), resource management via containerisation will work nicely. -
BTW, I really wonder how much pain could be avoided if updatedb recorded mtime of directories and checked it. I.e. instead of just doing blind find(1), walk the stored directory tree comparing timestamps with those in filesystem. If directory mtime has not changed, don't bother rereading it and just go for (stored) subdirectories. If it has changed - reread the sucker. If we have a match for stored subdirectory of changed directory, check inumber; if it doesn't match, consider the entire subtree as new one. AFAICS, that could eliminate quite a bit of IO... -
That would just save reading the directories. Not sure it helps that much. Much better would be actually if it didn't stat the individual files (and force their dentries/inodes in). I bet it does that to find out if they are directories or not. But in a modern system it could just check the type in the dirent on file systems that support that and not do a stat. Then you would get much less dentries/inodes. Also I expect in general the new slub dcache freeing that is pending will improve things a lot. But even if updatedb was fixed to be more efficient we probably still need a general solution for other tree walking programs that cannot be optimized this way. -Andi -
FWIW, find(1) does *not* stat non-directories (and neither would this approach). So it's just dentries for directories and you can't realistically skip those. OK, you could - if you had banned cross-directory rename for directories and propagated "dirty since last look" towards root (note that it would be a boolean, not a timestamp). Then we could skip unchanged subtrees completely... -
Could we help it a little from kernel and set 'dirty since last look' on directory renames? I mean, this is not only updatedb. KDE startup is limited by this, too. It would be nice to have effective 'what change in tree' operation. Pavel -- (english) http://www.livejournal.com/~pavelmachek (cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html -
Someone mentioned a variant of slocate above that they called mlocate,
and that Red Hat ships, that seems to do this (if I understand you and
what mlocate does correctly.)
--
I won't rest till it's the best ...
Programmer, Linux Scalability
Paul Jackson <pj@sgi.com> 1.925.600.0401
-
updatedb pushing out program data may be able to be improved on with drop behind or similar. however another scenerio that causes a similar problem is when a user is busy useing one of the big memory hogs and then switches to another (think you could make a synthetic test by writing a memory hog that allocates 3/4 of your ram then pauses waiting for input and then randomly accesses the memory for a while (say randomly accessing 2x # of pages allocated) and then pausing again before repeating run two of these, alternating which one is running at any one time. time how long it takes to do the random accesses. the difference in this time should be a fair example of how much it would impact the user. by the way, I've also seen comments on the Postgres performance mailing list about how slow linux is compared to other OS's in pulling data back in that's been pushed out to swap (not a factor on dedicated database almost always true, however there is some amount of I/O that is free with todays drives (remember, they read the entire track into ram and then give you the sectors on the track that you asked for). and if you have a raid array this is even more true. if you read one sector in from a raid5 array you have done all the same I/O that you would have to do to read in the entire stripe, but I don't believe that the current system will keep it all around if it exceeds the readahead limit. so in many cases readahead may end up being significantly cheaper then you expect. David Lang -
Hoo boy, lots of messages this morning. (Al? I've added you to the CC: because of your swap-in vs swap-out speed report from January. See below -- half-way down or so -- for more detals.) Yes, and that was the core of my original report months ago. I'm working for a while on one task, go to openoffice to view a report, or gimp to tweak the colors on a photo before uploading it, and then go back to my email and... and... and... there we go. The faults that Con wrote a benchmark much like that. It showed measurable improvement Yeah, akpm and... one of the usual suspects, had mentioned something such as 2.6 is half the speed of 2.4 for swapin. (Let's see if I can find a reference for that, it's been a year or more...) Okay, misremembered. Swap in is half the speed of swap out ( http://lkml.org/lkml/2007/1/22/173 ). Al Boldi (added to the CC:, poor sod), is the one who knows how to measure that, I'm guessing. Al? How are you coming up with those figures? I'm interested in reproducing it. It could be due to something stupid, such as the VM Yeah, I knew I'd get called on that one :-). It's the seeks that'll really kill you, and as you say once you're on the track the rest is practically free (which is why the VM should prefer to evict larger chunks at a time rather than lots of small things, see http://lkml.org/lkml/2007/7/23/214 for something that's heading the Fengguang Wu is doing lots of active work on making the readahead suck less. Ping him and he'll likely take an active interest in the RAID stuff. Ray -
Thanks for asking. I'm rather surprised why nobody's noticing any of this slowdown. To be fair, it's not really a regression, on the contrary, 2.4 is lot worse wrt swapin and swapout, and Rik van Riel even considers a 50% swapin slowdown wrt swapout something like better than expected (see thread '[RFC] kswapd: Kernel Swapper performance'). He probably meant random swapin, which seems to offer a 4x slowdown. There are two ways to reproduce this: 1. swsusp to disk reports ~44mb/s swapout, and ~25mb/s swapin during resume 2. tmpfs swapout is superfast, whereas swapin is really slow (see thread '[PATCH] free swap space when (re)activating page') Here is an excerpt from that thread (note machine config in first line): ============================================ RAM 512mb , SWAP 1G #mount -t tmpfs -o size=1G none /dev/shm #time cat /dev/full > /dev/shm/x.dmp 15sec #time cat /dev/shm/x.dmp > /dev/null 58sec #time cat /dev/shm/x.dmp > /dev/null 72sec #time cat /dev/shm/x.dmp > /dev/null 85sec #time cat /dev/shm/x.dmp > /dev/null 93sec #time cat /dev/shm/x.dmp > /dev/null 99sec ============================================ As you can see, swapout is running full wirespeed, whereas swapin not only is 4x slower, but increasingly gets the VM tangled up to end at a ~6x slowdown. So again, I'm really surprised people haven't noticed. Thanks! -- Al -
Sorry for the late reply. Well I think I reported this or another swap/tmpfs performance issue earlier ( http://marc.info/?t=116542915700004&r=1&w=2 ), we got the suggestion to increase /proc/sys/vm/page-cluster to 5, but we never came around to try it. Maybe this was the reason for my report to be almost entirely ignored, sorry for that. Regards, Magnus -
Notenotenote, not sure what you're going to show with it (times are simply as horrendous as I'd expect) but thought I'd try to inject something other than steaming cups of 4-letter beverages. Rene.
when the swap readahead is enabled does it make a significant difference in the time to do the random access? if it does that should show a direct benifit of the patch in a simulation of a relativly common workflow (startup a memory hog like openoffice then try and go back to your prior work) David Lang -
I don't use swap prefetch (nor -ck or -mm). If someone who has the patch applied waits to hit enter until swap prefetch has prefetched it all back in again, it certainly will. Swap prefetch's potential to do larger reads back from swapspace than a random segfaulting app could well be very significant. Reads are dwarved by seeks. If this program does what you wanted, please use it to show us. Rene. -
I haven't used swap prefetch either, the call was put out for what could be used to test the performance, and I was suggesting a test. if nobody else follows up on this I'll try to get some time to test it myself in a day or two. David Lang -
this assumes that this isn't ruled an invalid test in the meantime. in any case thanks for codeing this up so quickly. David Lang -
Let's save a little time and guess. While two instances of the hog are running no physical memory is free (as together they take up 1.5x physical) meaning that swap-prefetch wouldn't get a change to do anything and wouldn't make a difference. As such, the two instances test as you suggested would in fact not be testing anything it seems. However, if you quit one, and idle long enough to continue with the other one until swap-prefetch prefetched all its memory back in, it should be a difference on the order of minutes, even total if swap prefetch fetched it back in without seeking al over swap-space, and "total" isn't applicable if the idle time really is free. A program randomly touching single pages all over memory is a contrived worst case scenario and not a real-world issue. It is a boundary condition though, and it's simply quite impossible to think of any example where swap-prefetch would _not_ give you a snappier feeling machine after you've been idling. So really the only question would seem to be -- does it hurt any if you have _not_ been? Rene. -
NACK on this one. This bloats romfs by almost half of it's previous size to add mtd support to it. Given that romfs is a compltely trivial filesystem it's much better to have a separate filesystem driver handling the format on mtd instead of adding all these indirections. In addition to that argument the switch on the underlying subsystem is done horrible. There's lots of ifdefs instead of proper functions pointers, there's one file containing both block and mtd code instead of seaparate files, etc. And the get_unmapped_area method in a bare filesystem needs a _lot_ of explanation. -
The rest of it is nacked anyway, until we unify the point and get_unmapped_area methods of the MTD API. -- dwmw2 -
Methinks you meant nommu-make-it-possible-for-romfs-to-use-mtd-devices.patch, not romfs-printk-format-warnings.patch. I'll drop nommu-make-it-possible-for-romfs-to-use-mtd-devices.patch, thamks. -
Thanks. I was certainly getting confused. --- ~Randy *** Remember to use Documentation/SubmitChecklist when testing your code *** -
Care to drop the patches James NACKed every single time? -
I'm not aware of any which fit that description. There may be a couple in there which fix real bugs in an unapproved way. But I keep such patches as a matter of policy, so people keep on getting pestered about their bugs. -
> pagefault-in-write deadlock fixes. Will hold for 2.6.24. Why that? This stuff has been in forever and is needed at various levels. We need this in for anything to move forward on the buffered write front. -
At Nick's request. More work is needed and the code hasn't had a lot of -
Umm, Andrew - mixing new userspace interface, compltely rewritten drivers and simple fixes in a simple misc category doesn't exactly help reading this list :) -
Hopefull this will be done during the 2.6.23 merge window, but right now it's not (yet). -
> lguest-export-symbols-for-lguest-as-a-module.patch __put_task_struct is one of those no way in hell should this be exported things because we don't want modules messing with task lifetimes. Fortunately I can't find anything actually using this in lguest, so it looks the issue has been solved in the meantime. I also have a rather bad feeling about exporting access_process_vm. This is the proverbial sledge hammer for access to user vm addresses and I'd rather keep it away from module programmers with "if all you have is a hammer ..." in mind. In lguest this is used by send_dma which from my short reading of the code seems to be the central IPC mechanism. The double copy here doesn't look very efficient to me either. Maybe some VM folks could look into a better way to archive this that might be both more Just started to reading this (again) so no useful comment here, but it would be nice if the code could follow CodingStyle and place the || and && at the end of the line in multiline conditionals instead of at the beginning of the new one. -
To do inter-guest (ie. inter-process) I/O you really have to make sure It's not a double copy: it's a map & copy. If KVM develops inter-guest I/O then this could all be extracted into a Surprisingly, you have a point here. Since the key purpose of lguest is as demonstration code, it meticulously match kernel style. I shall immediately prepare a patch to convert the rest of the kernel to the correct "&& at beginning of line" style. Rusty. -
From: Rusty Russell <rusty@rustcorp.com.au> You should just let it exit and when it does you receive some kind of exit notification that resets your virtual device channel. I think the reference counting approach is error and deadlock prone. Be more loose and let the events reset the virtual devices when guests go splat. -
There are two places where we grab task refcnt. One might be avoidable
(will test and get back) but the deferred wakeup isn't really:
/* We cache one process to wakeup: helps for batching & wakes outside locks. */
void set_wakeup_process(struct lguest *lg, struct task_struct *p)
{
if (p == lg->wake)
return;
if (lg->wake) {
wake_up_process(lg->wake);
put_task_struct(lg->wake);
}
lg->wake = p;
if (lg->wake)
get_task_struct(lg->wake);
}
We drop the lock after I/O, and then do this wakeup. Meanwhile the
other task might have exited.
I could get rid of it, but I don't think there's anything wrong with the
code...
Cheers,
Rusty.
-
From: Rusty Russell <rusty@rustcorp.com.au> I already understand what you're doing. Is it possible to use exit notifiers to handle this case? That's what I'm trying to suggest. :) -
Sure, the process has /dev/lguest open, so I can do something in the close routine. Instead of keeping a reference to the tsk, I can keep a reference to the struct lguest (currently it doesn't have or need a refcnt). Then I need another lock, to protect lg->tsk. This seems like a lot of dancing to avoid one export. If it's that important I'd far rather drop the code and do a normal wakeup under the big lguest lock for 2.6.23. Cheers, Rusty. -
From: Rusty Russell <rusty@rustcorp.com.au> I'm not against the export, so use if it really helps. Ref-counting just seems clumsy to me given how the hw assisted -
<handwaving> We seem to be taking the reference against the wrong thing here. It should be against the mm, not against a task_struct? -
This is solely for the wakeup: you don't wake an mm 8) The mm reference is held as well under the big lguest_mutex (mm gets destroyed before files get closed, so we definitely do need to hold a reference). I just completed benchmarking: the cached wakeup with the current naive drivers makes no difference (at one stage I was playing with batched hypercalls, where it seemed to help). Thanks Christoph, DaveM! === Remove export of __put_task_struct, and usage in lguest lguest takes a reference count of tasks for two reasons. The first is bogus: the /dev/lguest close callback will be called before the task is destroyed anyway, so no need to take a reference on open. The second is code to defer waking up tasks for inter-guest I/O, but the current lguest drivers are too simplistic to benefit (only batched hypercalls will see an effect, and it's likely that lguests' entire I/O model will be replaced with virtio and ringbuffers anyway). Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> --- drivers/lguest/hypercalls.c | 1 - drivers/lguest/io.c | 18 +----------------- drivers/lguest/lg.h | 1 - drivers/lguest/lguest_user.c | 2 -- kernel/fork.c | 1 - 5 files changed, 1 insertion(+), 22 deletions(-) =================================================================== --- a/drivers/lguest/hypercalls.c +++ b/drivers/lguest/hypercalls.c @@ -189,5 +189,4 @@ void do_hypercalls(struct lguest *lg) do_hcall(lg, lg->regs); clear_hcall(lg); } - set_wakeup_process(lg, NULL); } =================================================================== --- a/drivers/lguest/io.c +++ b/drivers/lguest/io.c @@ -296,7 +296,7 @@ static int dma_transfer(struct lguest *s /* Do this last so dst doesn't simply sleep on lock. */ set_bit(dst->interrupt, dstlg->irqs_pending); - set_wakeup_process(srclg, dstlg->tsk); + wake_up_process(dstlg->tsk); return i == dst->num_dmas; fail: @@ -333,7 +333,6 @@ again: /* Give ...
The version that just got into mainline still has the __put_task_struct export despite not needing it anymore. Care to fix this up? -
No, it got patched in then immediately patched out again. Andrew mis-mixed my patches, but there have been so many of them I find it hard to blame him. Rusty. -
Indeed, the export is gone in last mainline gone. -
What about Open /dev/lguest transfer fd using SCM_RIGHTS (or clone()?) close fd in original task exit() ? My feeling is that if you want to be bound to a task, not a file, you need to use syscalls, not ioctls. -- error compiling committee.c: too many arguments to function -
"Don't do that". You'll lose the ability to access the operations on the fd once you are no longer the original task (explicit check)., It's not an exact match, but a file is a remarkably convenient abstraction for a non-ABI such as lguest. Of course, Carsten was talking about unifying the lguest & kvm userspace interface, so this could well change anyway. Cheers, Rusty. -
On Wed, 11 Jul 2007 14:23:24 +0200 Christoph Hellwig wrote: I prefer them at the ends of lines also, but that's not in CodingStyle, it's just how we do it most of the time (so "coding style", without caps). --- ~Randy *** Remember to use Documentation/SubmitChecklist when testing your code *** -
On Wed, 11 Jul 2007 14:23:24 +0200 Ther are a couple of calls to put_task_struct() in there, and that needs hm, well, access_process_vm() is a convenience wrapper around -
Andrew Morton wrote: The lumpy reclaim patches originally came out of work to support Mel's anti-fragmentation work. As such I think they have become somewhat attached to those patches. Whilst lumpy is most effective where placement controls are in place as offered by Mel's work, we see benefit from reduction in the "blunderbuss" effect when we reclaim at higher orders. While placement control is pretty much required for the very highest orders such as huge page size, lower order allocations are benefited in terms of lower collateral damage. There are now a few areas other than huge page allocations which can benefit. Stacks are still order 1. Jumbo frames want higher order contiguous pages for there incoming hardware buffers. SLUB is showing performance benefits from moving to a higher allocation order. All of these should benefit from more aggressive targeted reclaim, indeed I have been surprised just how often my test workloads trigger lumpy at order 1 to get new stacks. Truly representative work loads are hard to generate for some of these. Though we have heard some encouraging noises from those who can reproduce these problems. [...] -apw -
[Seems a PEBKAC occured on the subject line, resending lest it become a victim of "oh thats spam".] -
I'd expect that the main application for lumpy-reclaim is in keeping a pool of order-2 (say) pages in reserve for GFP_ATOMIC allocators. ie: jumbo frames. At present this relies upon the wakeup_kswapd(..., order) mechanism. How effective is this at solving the jumbo frame problem? (And do we still have a jumbo frame problem? Reports seems to have subsided) -
The patches have an application with hugepage pool resizing. When lumpy-reclaim is used used with ZONE_MOVABLE, the hugepages pool can be resized with greater reliability. Testing on a desktop machine with 2GB of RAM showed that growing the hugepage pool with ZONE_MOVABLE on it's own was very slow as the success rate was quite low. Without lumpy-reclaim, each attempt to grow the pool by 100 pages would yield 1 or 2 hugepages. With lumpy-reclaim, getting 40 to 70 hugepages on each attempt was typical. -- Mel Gorman Part-time Phd Student Linux Technology Center University of Limerick IBM Dublin Software Lab -
The tie in between allocator and kswapd is essentially unchanged, so if allocators are dropping below the watermarks at the specified order, reclaim will be triggered at that order. Reclaim continues until we return above the high watermarks, at the order at which we are reclaiming. What lumpy brings is a greater targetting of effort to get the pages. kswapd now uses the desired allocator order when applying reclaim. This leads to pressure being applied to contigious areas at the required order, and so a higher chance of that order becoming available. Traditional reclaim could end up applying pressure to a number of pages, but not all pages in any area at the required order, leading to a very low chance of success. By targetting areas at the required order we significantly increase the chances of success for any given amount of reclaim. As we will reclaim until we have the desired number of free pages, we will have to reclaim less to achieve this compared to random reclaim. This certainly is appealing intuitivly, and our testing at higher orders shows that the cost of each reclaimed page is lower and more importantly the time to reclaim each page is reduced. So for a 'continuing' consumer like an incoming packet stream, we should have to do much less work and thus disrupt the system as a whole much less to get its pages. Where demand for atomic higher order pages is not heavy we would expect kswapd to maintain free levels pages more readily and so under higher demand. Though it should be stressed without placement control success rates drop off significantly at higher orders as the probabality of reclaim succeeding on all pages in the area subsided) It is not in the least bit clear if the problem is resolved or if the reporters have simply gone quiet. Overall the approach taken in lumpy reclaim seems to be a logical extension of the regular reclaim algorithm, leading to more efficient reclaim. -apw -
Chris is a moving target. Thankfully we have Shannon Nelson taking over Chris' duties. Shannon, can you take a look at these and see what needs to happen to it ? Most likely these just need to be pushed to the right person. Cheers, Auke PS: I think we should add an I/OAT / DMA engine section in the MAINTAINERS... -
Auke: Thanks for the introduction :-). Andrew: All three of these patches are reasonable and can be pushed on up. You can add my sign-off to all three: I'll be posting a MAINTAINERS patch Real Soon Now with my name on IOAT/DMA. sln ====================================================================== Mr. Shannon Nelson LAN Access Division, Intel Corp. Shannon.Nelson@intel.com I don't speak for Intel (503) 712-7659 Parents can't afford to be squeamish. -
On Tue, 10 Jul 2007 11:05:45 -0700 OK, the way it works is that I send these patches at the git tree maintainer, then the git tree maintainer merges them (this step is unreliable) and then when I repull that git tree maintainer's tree I see that they got merged so I drop them from -mm. The git tree maintainer decides when to send them to Linus. I am presently pulling git://lost.foo-projects.org/~cleech/linux-2.6#master into -mm. Will you be taking over the IOAT git tree? If so, please send me a suitable git URL when it's ready. The above tree has several changes in it from January (see ftp://ftp.kernel.org/pub/linux/kernel/people/akpm/patches/2.6/2.6.22-rc6/2.6.22-rc6-mm...). Please take a look at those, work out what we should do with it all. -
I'll be getting there Real Soon Now. The transition seems to be a Will do. Thanks for your patience. sln ====================================================================== Mr. Shannon Nelson LAN Access Division, Intel Corp. Shannon.Nelson@intel.com I don't speak for Intel (503) 712-7659 Parents can't afford to be squeamish. -
(just to provide my indicator of status) see Alan's comments. I've been ignoring pata_acpi for a while, because are other pata_platform people happy with this? I don't know embedded should be combined, really. will merge eventually. basic concept OK, Needs a bug fix, so that the newly modified loop doesn't scan the final Any of the above worth 2.6.23? Just wondering if they were useful Just the general march of progress on new hardware :) I would like to see this support merged in /some/ form. We've been telling Intel for years they were sillyheads for not bothering with an IOMMU. Now that they have, we should give them a cookie and support good technology. Jeff -
On Tue, 10 Jul 2007 13:42:16 -0400 This is just a silly remove-unneeded-cast-of-void* cleanup. I wrote this as a fixup against libata-add-irq_flags-to-struct-pata_platform_info.patch with the intention of folding it into that base patch, but you went and merged the submitter's original patch so this trivial fixup got stranded in -mm. Feel free to give Oh, I thought these were the patches which affected scsi and which James 3x59x-fix-pci-resource-management.patch: you wrote it ;) I have a comment here: - I don't remember the story with cardbus either. Presumably once upon a time the cardbus layer was claiming IO regions on behalf of cardbus devices (?) Need to think about that. update-smc91x-driver-with-arm-versatile-board-info.patch: See comment from rmk in changelog: ftp://ftp.kernel.org/pub/linux/kernel/people/akpm/patches/2.6/2.6.22-rc6/2.6.22-rc6-mm... Deepak, can we move this along a bit please? drivers-net-ns83820c-add-paramter-to-disable-auto.patch: See comments in changelog: ftp://ftp.kernel.org/pub/linux/kernel/people/akpm/patches/2.6/2.6.22-rc6/2.6.22-rc6-mm... The first few patches will a) fix up our writev performance regression and b) reintroduce the writev() deadlock which the writev()-regresion-adding patch fixed. OK, thanks. -
Well ... my concern was really how to make them more generic ... ahci isn't the only controller that can do phy power management, and it also seemed to me that the most generic entity for power management was the transport rather than the SCSI mid-layer, but that debate is still ongoing. James -
I'm sorry, I didn't look closely enough. I was referring to the hrm. ISTR James wanted some cleanups, Kristen did some cleanups, then looking at the cleanups decided they were needed / appropriate at this time. Anyway, these are in my mbox queue and the libata portions (of which the code is the majority) seem OK. Need to give them a final review. Jeff -
On Wed, 11 Jul 2007 00:31:23 +0400 yup, that's what "I have a comment" meant ;) The comment seems rather bogus actually. Let's just merge it. -
Mmm.. Ben had 2 comments last year: I know very little about hardware and only own the fiber version of this card. Even if i tried to make code for the copper version, it would probably blow it up the phy and set the switches on fire ;). This is pretty much Russian to me. I wouldnt know where to find the "link-autonegotiation-state-machine-for-fibre-cards" or know what to do with it anyway :). The "disable_autoneg" is a convenient feature (for me and the other guy who made the same patch last year) and i consider it a harmless feature in every way. It is simply an 'if'-statement, that skips the "start autoneg" function upon load. We can simply remove the feature entirely if it is deemed undesirable. So in conclusion: - I vote "use the patch as-is", but im fine with it being changed. - If it needs support for copper, someone else has to code it. Regards - Dan -
Hello.
Now that the fix for CONFIG_PCI=n has been merged, what's left is to test
You should have, I was sending it to you.
WBR, Sergei
-
The more work may turn out being too much for you (although it is nothing exactly tricky that would introduce subtle bugs, it is a fair amont of churn). However, in that case we can still merge these two: mm-fix-fault-vs-invalidate-race-for-linear-mappings.patch mm-fix-clear_page_dirty_for_io-vs-fault-race.patch Which fix real bugs that need fixing (and will at least help to get some of my patches off your hands). -- SUSE Labs, Novell Inc. -
OK, so does that mean we can finally get the block_page_mkwrite patches merged? i.e.: http://marc.info/?l=linux-kernel&m=117426058311032&w=2 http://marc.info/?l=linux-kernel&m=117426070111136&w=2 I've got up-to-date versions of them ready to go and they've been consistently tested thanks to the XFSQA test I wrote for the bug that it fixes. I've been holding them out-of-tree for months now because ->fault was supposed to supercede this interface..... Cheers, Dave. -- Dave Chinner Principal Engineer SGI Australian Software Group -
Yeah, as I've said, don't hold them back because of me. They are relatively simple enough that I don't see why they couldn't be merged in this window. -- SUSE Labs, Novell Inc. -
Xen is probably going to be merged. I'm still not fully happy about the review status of the drivers and xenbus, but there doesn't seem to be much value in delaying it further. I'll consolidate the fixes and fixes-to-fixes. It's still not clear to me this is any useful. The current code can run a program on MCE which should be really fast enough Might need more testing? I'm sceptical about the dynticks code. It just rips out the x86-64 timing code completely, which needs a lot more review and testing. Probably not .23 -Andi -
For the mtrr trim patch at least, I think the coverage we've received in -mm is probably sufficient (the failure mode would be fairly obvious). The only thing I'm nervous about is adding AMD support for the quirk, since I don't have any way of testing it. We can easily add that later though, if a tester steps forward or we see demand for it (should just be an extra conditional in the trim code). Thanks, Jesse -
What you just did here is a slap in the face to a lot of contributors
who worked hard on this code :(
Let me tell you about the history of this project first. Arjan wrote the
first version of it a year ago, and it was added to -rt and tested there
by many people and went through many iterations and fixes. Chris Wright
then created a x86_64 clockevents cleanup and dynticks enabling patchset
from it this spring and sent it to lkml three and a half months ago, on
March 31:
http://lwn.net/Articles/229094/
Thomas, the high resolution timers and clockevents maintainer,
immediately picked up Chris' splitup/splitout/cleanup work and fixed and
extended it, and sent a first cut to lkml on May 6th:
http://lwn.net/Articles/233226/
Thomas then sent an updated version of the x86_64 clockevents cleanup
and dynticks code to lkml (on June 10th), for a second round of review:
http://lwn.net/Articles/237687/
As Thomas stated it in his submission:
" The patch set has been tested in the -hrt and -rt trees for quite a
while and the initial problems have been sorted out. Thanks to the
folks from the PowerTop project for testing and feedback. "
Then on June 16th Thomas sent the third series:
http://lwn.net/Articles/238834/
(which too was in -rt and was tested there on numerous machines. It was
also added to -mm.)
Then on June 23rd Thomas sent the fourth series of the x86_64
clockevents and dynticks code:
http://lwn.net/Articles/239620/
We finally have someone (Thomas) with core kernel clue who actually
_cares_ about the x86 time code and does not see it as an ugly chore,
one who collects the right patches and maintains the -hrt tree and
co-maintains the -rt tree and interacts with other contributors. What he
did was _hard_ to do but we are making really good progress:
http://lkml.org/lkml/2007/7/5/242
" All in all, personally I'm very happy to see Linux making such a
huge step forward with tickless and ...I can understand being disappointed, but not quite as upset as you appear to be. so have you (Ingo) reviewed the ext4 patches? or reiser4 patches? or lumpy reclaim? or anti-fragmentation? I certainly haven't. I can barely keep up with reading about 1/2 of lkml emails. And in my non-scientific method, I think that we are suffering from both (a) more patch submittals and (b) fewer qualified reviewers (per kernel KLOC) than we had 3-5 years ago. I don't see how you can expect Andrew to review these or any other specific patchset. Do you have some suggestions on how to clone Andrew? --- ~Randy *** Remember to use Documentation/SubmitChecklist when testing your code *** -
Randy, Ingo was talking to Andi, the x86_64 maintainer, not to Andrew. And I share his opinion that the maintainer of the subsystem, which is affected by such a fundamental patch, could have at least shown any public sign of interest, disgust, comment or what ever in a 3+ month time frame. Especially about a patch, which is a logical consequence of an almost two years public and transparent effort to consolidate the time code in the kernel. I for my part have no problem maintaining the set for another round out of tree and weed out eventually problems in -mm, but my expectation for qualified response of the responsible maintainer is exactly zero right now. Thanks, tglx -
Yep, I see that when I re-read it. I apologize. --- ~Randy *** Remember to use Documentation/SubmitChecklist when testing your code *** -
Well I spent a lot of time making the x86-64 timing code work well on a variety of machines; working around a wide variety of hardware and platform bugs. I obviously don't agree on your description I told him my objections privately earlier. Basically i would like to see an actually debuggable step-by-step change, not a rip everything out. If that isn't possible it needs very careful review which just hasn't happened yet. But I'm not convinced even step by step is not possible here. I thought it was clear that rip everything out is rarely a good idea in Linux land? That's really not something I should need to harp on repeatedly. -Andi -
Andi, You promised privately to do a thorough review as well, which I'm still There is no step by step thing. You convert an arch to clock events or If you have technical objections, put them on the table. Point by point. All I heard so far from you are platitudes, which are not worth the electrons to transport them. tglx -
Indeed, about the only thing can be done is to take a slower approach to converging the arch specific implementations (hpet, pit, etc). thanks, -chris -
I did some reviewing, but never the big write up and feedback. That was my fault, sorry. -Andi -
Hi Andi, I'm going to change topic big time because your sentence above perfectly applies to the O(1) scheduler too. It's not like process schedulers are sacred and there shall be only one, while I/O schedulers and packet schedulers are profane and there can be many of them. FWIW IMHO the right way would have been to make the new scheduler pluggable and switchable at runtime, too bad it was ripped off instead. The difficulty of making the scheduler pluggable isn't really enormous, there have been patches floating around to achieve it, some I even deal with them myself once. The only positive side of being forced to CFS I can imagine, is that more testing will make it more stable and more tuned more quickly. But I'm fairly certain Ingo's good enough to achieve without it, perhaps with a few more weeks. Personally I very much like the unfariness of O(1), I'm afraid CFS will overschedule under a certain number of workloads in its attempt to provide a complete fair queieing at all costs, and it won't deal with the X server as nicely as O(1), but I may as well be wrong. The only thing I'm more sure about is that the computational complexity is higher, and that reason alone is a good technical reason to provide both and let the java folks stick with O(1) if they want. -
I disagree to a large degree. We almost never have problems with code you can "think about". Sure, bugs happen, but code that everybody runs the same generally doesn't break. So a CPU scheduler doesn't worry me all that much. CPU schedulers are "easy". What worries me is interfaces to hardware that we know looks different for different people. That means that any testing that one person has done doesn't necessarily translate to anything at *all* on another persons machine. The timer problems we had when merging the stuff in 2.6.21 just scarred me. I'd _really_ hate to have to go through that again. And no, the "gradual" thing where the patch that actually *enables* something isn't very gradual at all, so that's the absolutely worst kind of thing, because then people can "git bisect" to the point where it got enabled and tell us that's where things broke, but that doesn't actually say anything at all about the patch that actually implements the new behaviour. So the "enable" kind of patch is actually the worst of the lot, when it comes to hardware. When it comes to pure software algorithms, and things like schedulers, you'll still obviously have timing issues and tuning, but generally things *work*, which makes it a lot easier to debug and describe. Linus -
Hi, A little more advance warning wouldn't have hurt though. The new scheduler does _a_lot_ of heavy 64 bit calculations without any attempt to scale that down a little... One can blame me now for not having it brought up earlier, but discussions with Ingo are not something I'm looking forward to. :( bye, Roman -
See prio_to_weight[], prio_to_wmult[] and sysctl_sched_stat_granularity. Perhaps more can be done, but "without any attempt..." isn't accurate. -Mike -
Hi, Calculating these values at runtime would have been completely insane, the alternative would be a crummy approximation, so using a lookup table is actually a good thing. That's not the problem. BTW could someone please verify the prio_to_wmult table, especially [16] and [21] look a little off, like a digit was cut off. While I'm at this, the 10% scaling there looks a little much (unless there are other changes I haven't looked at yet), the old code used more like 5%. This would mean a prio -20 task would get 98.86% cpu time compared to a prio 0 task, that was previously about the difference between -20 and 19 (and it would have previously gotten only 88.89%), now a prio -20 task would get 99.98% cpu time compared to a prio 19 task. The individual levels are unfortunately not that easily comparable, but at the overall scale the change looks IMHO a little drastic. bye, Roman -
Roman Zippel noticed inconsistency of the wmult table.
wmult[16] has a missing digit.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
diff --git a/kernel/sched.c b/kernel/sched.c
index 0559665..3332bbb 100644
--- a/kernel/sched.c
+++ b/kernel/sched.c
@@ -750,7 +750,7 @@ static const u32 prio_to_wmult[40] = {
48356, 60446, 75558, 94446, 118058, 147573,
184467, 230589, 288233, 360285, 450347,
562979, 703746, 879575, 1099582, 1374389,
- 717986, 2147483, 2684354, 3355443, 4194304,
+ 1717986, 2147483, 2684354, 3355443, 4194304,
5244160, 6557201, 8196502, 10250518, 12782640,
16025997, 19976592, 24970740, 31350126, 39045157,
49367440, 61356675, 76695844, 95443717, 119304647,
-
[snip] While we're at it, isn't the comment above the wmult table incorrect? The multiplier is 1.25, meaning a 25% change per nice level, not 10%. - Jim -
yes, the weight multiplier 1.25, but the actual difference in CPU utilization, when running two CPU intense tasks, is ~10%: PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 8246 mingo 20 0 1576 244 196 R 55 0.0 0:11.96 loop 8247 mingo 21 1 1576 244 196 R 45 0.0 0:10.52 loop so the first task 'wins' +10% CPU utilization (relative to the 50% it had before), the second task 'loses' -10% CPU utilization (relative to the 50% it had before). so what the comment says is true: * The "10% effect" is relative and cumulative: from _any_ nice level, * if you go up 1 level, it's -10% CPU usage, if you go down 1 level * it's +10% CPU usage. for there to be a ~+10% change in CPU utilization for a task that races against another CPU-intense task there needs to be a ~25% change in the weight. Ingo -
in any case more documentation is justified, so i've added some
clarification to the comments - see the patch below.
Ingo
------------------------>
Subject: sched: improve weight-array comments
From: Ingo Molnar <mingo@elte.hu>
improve the comments around the wmult array (which controls the weight
of niced tasks). Clarify that to achieve a 10% difference in CPU
utilization, a weight multiplier of 1.25 has to be used.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
---
kernel/sched.c | 4 +++-
1 file changed, 3 insertions(+), 1 deletion(-)
Index: linux/kernel/sched.c
===================================================================
--- linux.orig/kernel/sched.c
+++ linux/kernel/sched.c
@@ -736,7 +736,9 @@ static void update_curr_load(struct rq *
*
* The "10% effect" is relative and cumulative: from _any_ nice level,
* if you go up 1 level, it's -10% CPU usage, if you go down 1 level
- * it's +10% CPU usage.
+ * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
+ * If a task goes up by ~10% and another task goes down by ~10% then
+ * the relative distance between them is ~25%.)
*/
static const int prio_to_weight[40] = {
/* -20 */ 88818, 71054, 56843, 45475, 36380, 29104, 23283, 18626, 14901, 11921,
-
Ah ok so it's 10% of the original CPU usage, not relative to a tasks share from before. While I guess I still think in terms of relative CPU share, your comments now make sense to me. Thanks for the clarification. -
Hi, As soon as you add another loop the difference changes again, while it's always correct to say it gets 25% more cpu time (which I still think is a little too much). bye, Roman -
yep, and i'll add the relative effect to the comment too. Ingo -
Hi, Why did you cut off the rest of the sentence? To illustrate the problem a little different: a task with a nice level -20 got around 700% more cpu time (or 8 times more), now it gets 8500% more cpu time (or 86.7 times more). You don't think that change to the nice levels is a little drastic? bye, Roman -
(no need to become hostile, i answered to that portion of your sentence separately, which was logically detached from the other portion of your This was discussed on lkml in detail, see the CFS threads. It has been a common request for nice levels to be more logical (i.e. to make them universal and to detach them from HZ) and for them to be more effective as well. Ingo -
Hi, Could you please stop with these accusations? Which are quite big, so I skipped most of it, a more precise pointer would Huh? What has this to do with HZ? The scheduler used ticks internally, but it's irrelevant to what the user sees via the nice levels. So the question still stands that this change may be a little drastic, as you changed the nice levels of _all_ users, not just of those who were previously interested in CFS. bye, Roman -
unfortunately you are wrong again - there are various HZ related artifacts in the nice level support code of the old scheduler. v2.6.22, CONFIG_HZ=100, nice +19 task against a nice-0 CPU-intense task: PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 2446 mingo 25 0 1576 244 196 R 90.9 0.0 0:32.79 loop 2448 mingo 39 19 1580 248 196 R 9.1 0.0 0:02.94 loop v2.6.22, CONFIG_HZ=250, nice +19 task against a nice-0 CPU-intense task: PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 2358 mingo 25 0 1576 248 196 R 96.1 0.0 0:31.97 loop_silent 2363 mingo 39 19 1576 244 196 R 3.9 0.0 0:01.24 loop_silent v2.6.22, CONFIG_HZ=300, nice +19 task against a nice-0 CPU-intense task: PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 2332 mingo 25 0 1580 248 196 R 95.1 0.0 0:11.84 loop_silent 2335 mingo 39 19 1576 244 196 R 3.1 0.0 0:00.39 loop_silent to sum it up: a nice +19 task (the most commonly used nice level in practice) gets 9.1%, 3.9%, 3.1% of CPU time on the old scheduler, depending on the value of HZ. This is quite inconsistent and illogical. this HZ dependency of nice levels existed for many years, and the new scheduler solves that inconsistency - every nice level will get the same amount of time, regardless of HZ. Ingo -
Hi, You're correct that you can find artifacts in the extreme cases, it's subjective whether this is a serious problem. It's nice that these artifacts are gone, but that still doesn't explain why this ratio had to be increase that much from around 1:10 to 1:69. bye, Roman -
More dynamic range is better? If you actually want a task to get 20x the CPU time of another, the older scheduler doesn't really allow it. Getting 1/69th of a modern CPU is still a fair number of cycles. Nevermind 1/69th of a machine with > 64 cores. -- Mathematics is the supreme nostalgia of our time. -
yeah. furthermore, nice -20 is only admin-selectable. Here are the current CPU-use values for positive nice levels: nice 0: 100.00% nice 1: 80.00% nice 2: 64.10% nice 3: 51.28% nice 4: 40.98% nice 5: 32.78% nice 6: 26.24% nice 7: 21.00% nice 8: 16.77% nice 9: 13.42% nice 10: 10.74% nice 11: 8.59% nice 12: 6.87% nice 13: 5.50% nice 14: 4.39% nice 15: 3.51% nice 16: 2.81% nice 17: 2.25% nice 18: 1.80% nice 19: 1.44% here's the CPU utilization table for negative nice levels (relative to a nice -20 task): nice 0: 1.15% nice -1: 1.44% nice -2: 1.80% nice -3: 2.25% nice -4: 2.81% nice -5: 3.51% nice -6: 4.39% nice -7: 5.50% nice -8: 6.87% nice -9: 8.59% nice -10: 10.74% nice -11: 13.42% nice -12: 16.77% nice -13: 21.00% nice -14: 26.24% nice -15: 32.78% nice -16: 40.98% nice -17: 51.28% nice -18: 64.10% nice -19: 80.00% nice -20: 100.00% these are pretty sane, and symmetric across the origo. Nice -20 is the odd one out, because there is no nice +20. But its value is still logical, it's the mirror image of an imaginery nice +20. and note that even on the old scheduler, nice-0 was "3200% more powerful" than nice +19 (with CONFIG_HZ=300), and nice -19 was only 700% more powerful than nice-0. So not only was it inconsistent (and i can create scary numbers too ;), it gave the admin-controlled negative nice levels less of a punch than to user-controlled nice +19. A number of people complainted about that, and CFS addresses this. in fact i like it that nice -20 has a slightly bigger punch than it used to have before: it might remove the need to run audio apps (and other multimedia apps) under SCHED_FIFO. (SCHED_FIFO is unprotected against lockups, while under CFS a nice 0 task is still starvation protected against a nice -20 task.) furthermore, there is a quality of implementation issue as well, look at the definition of the nice system call: asmlinkage long sys_nice(int ...
Hi, How did you get that value? At any HZ the ratio should be around 1:10 "Slightly bigger"??? You're joking, right? Especially the user levels are doing something completely different now, which may break user expectation. While the user couldn't expect anything precise, it's still a big difference whether a process at nice 5 gets 75% of the time or only 30%. bye, Roman -
you are wrong again. I sent you the numbers earlier today already: | PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND | 2332 mingo 25 0 1580 248 196 R 95.1 0.0 0:11.84 loop | 2335 mingo 39 19 1576 244 196 R 3.1 0.0 0:00.39 loop 3.1% is 3067% more than 95.1%, and the ratio is 1:30.67. You again deny above that this is the case, and there's nothing i can do about your denial of facts - that is your own private problem. Ingo -
Hi, Ingo, how am I supposed to react to this? I'm asking a simple question and I get this? I'm at serious loss how to deal with you. :-( Above is based on theoritical values, for a 300HZ kernel these two processes should get 30 and 3 ticks. Should there be any rounding error or off by one error so that the processes get one tick less than they should get or one tick is accounted to the wrong process, my theoritical value is still within the possible error range and doesn't contradict your practical values. Playing around with some other nice levels, confirms the theory that something is a little off, so I'm quite correct at saying that the ratio _should_ be 1:10. OTOH you are the one who is wrong about me (again). :-( bye, Roman -
Hi, Rechecking everything there was actually a small error in my test program, so the ratio should be at 1:20. Sorry about that mistake. Nice level 19 shows the largest artifacts, as that level only gets a single tick, so the ratio is often 1:HZ/10 (except for 1000HZ where it's 5:100). Nevertheless it's still true that in general nice levels were independent of HZ (that's all I wanted to say a couple of mails ago). Ingo, you can start now gloating, but contrary to you I have no problems with admitting mistakes and apologizing for them. The point is just that I'm reacting better to factual arguments instead of flames (and I think bye, Roman -
Roman, please do me a favor, and ask me the following question: " Ingo, you've been maintaining the scheduler for years. In fact you wrote the old nice code we are talking about here. You changed it a number of times since then. So you really know what's going on here. Why does the old nice code behave like that for nice +19 levels? " I've been waiting for that obvious question, and i _might_ be able to answer it, but somehow it never occured to you ;-) Thanks, Ingo -
Hi, Do you have any idea how insulting and arrogant this is? Let me translate for you, how this arrived: "O Ingo, who art our god of the scheduler. You have blessed the paths I walked in. You kept me from sinning numerous times. Your wisdom is infinite. Guide me on the journey that layeth ahead of me into this world knowledge of Your truth." (I apologize already in advance, if I should have hurt anyones religious feelings.) It's obvious that you have more experience with the scheduler code, but does that make you unfailable? Does that give you the right to act like a jerk? I do make mistakes, I try to learn from them and life goes on, I have no problem with that, but what I have a problem with is if someone is abusing this to his own advantage. I have to be extremely carful what I say to you, because you jump on the first small mistake and I have to bear your insults like "there's nothing i can do about your denial of facts - that is your own private problem." I have no problems with facts, I'm only trying very hard to ignore your arrogant behaviour... If you have something to contribute to this discussion which might clear things up, then just say it, but I'm not going to beg for it. bye, Roman -
Roman, it is really not about 'experience', and yes, we all make frequent mistakes. it's about the plain fact that i happened to write _both_ the old and the new code you were talking about all along. In this discussion about nice levels you were (very) agressively asserting things that were untrue, you were suggesting that i dont understand the code, instead of simply asking me why the code was written in such a way and what the motivation behind it was. I'd be glad to attempt to answer such a friendly question, if you are interested in asking it and if you are interested in my answer. Thanks, Ingo -
Hi, Instead of simply asserting things, how about you provide some examples? I made so far a single mistake of mixing up nice levels 18 and 19. If you would point me to such examples, I could learn how to tone it down a little, since the nice levels are not the only issue I have with the new scheduler, the heavy stuff is still about to come. The problem here is there is too much burnt ground so I can't just present raw ideas, which get flamed by you, I have to be sufficiently confident they are valid, Again, please point me to examples, so I at least have a chance to clear things up, since it was never my intention to make such a suggestion, but this gives me no chance to defend myself. OTOH I can tell you exactly how you continuously insult me, e.g. by suggesting I ask "stupid questions" or that I'm in "denial of facts". Don't make such suggestions if you have no idea how insulting they are. Especially the one deleted insult above where you have the impertinence to quote it, such tone is more appropriate between lord and inferior, where the latter have to make a request and the former "might" grant it. _Never_ make me beg. :-( bye, Roman -
uhm, [and the uninterested reader might want to skip to the next mail the ";-)" emoticon (and its contents) clearly signals this as a sarcastic, tongue-in-cheek remark. To make it even clearer, please ok? (If you didnt see/read it as sarcastic straight away then my apologies for insulting you!) The "_might_ be able to answer" bit is of course sarcastic too, and contrary to your (i have to say, pretty absurd) suggestion i did not suggest that i "might be _willing_ to answer" - which would be quite arrogant indeed and which i never said or suggested. To make it even clearer: i'm definitely able to answer questions about code i wrote originally and which i just changed, were you to show genuine interest in hearing my opinion :-) Ingo -
Hi, To take another example why is this still insulting and inappropriate, this is a behaviour I would characterize as school bullying: A bully attacks someone obviously weaker than himself and for example takes something away and than continues like "If you ask nicely I'll give it back to you.", this often accompied by laughter to signal he's enjoying himself and the power he has, but for the other person it's everything but funny. Maybe you don't know what it feels like, but I do and I can't find anything funny, sarcastic or whatever about this, no matter how many smileys or other tags you add there. If the communication is already that troubled as this, such "humor" is really the worst thing you can do and I Sorry, that is too little too late. You've apologized before and you continued to make fun of me personally to the point of spreading wrong information about me, which you could have very easily verified yourself, if you only wanted. What I want from you is that you treat me with respect and to keep your "sarcasm" to yourself. I told you very clearly how I think about you requoting this crap and yet you repeat it again _twice_, so on the one hand I get this apology attempt and on the other hand you continue to kick me in the crotch? How do you think am I supposed to feel about this? It's also always interesting what you don't respond to. I asked you for examples which would prove the (rather strong) assertions you made about me, what does it tell me now if you can't back up your statements? bye, Roman -
Hi, You can already have that, the complete range level from 19 to -20 was about 1:80. There is also something like too much range, I tried it with top at 19 and as soon as something runs at -20 it's practically dead, because it gets now only 1/5900 of cpu time. bye, Roman -
But that is irrelevant: all tasks start out at nice 0, and what matters is the dynamic range around 0. So the dynamic range has been made uniform in the positive from 1:10...1:20...1:30 to 1:69 for nice +19, and from 1:8 to 1:69 in the minus. (with 1:86 nice -20) If you look at the negative nice levels alone it's a substantial increase but if you compare it with positive nice levels you'll similar kinds of dynamic ranges were already present in the old scheduler and you'll see why we've done it. Negative nice levels are admin-controlled, the increase in the negative levels is is not a big issue and people actually like the increased dynamic range and the consistency. The positive range _might_ be a bigger issue but there we were largely inconsistent anyway, and again, people like the increased dynamic range. Ingo -
Hi, So let's look at them: for (i=0;i<20;i++) print i, " : ", (20-i)*5, " : ", 100*1.25^-i, " : ", e(l(2)*(-i/5))*100, "\n"; 0 : 100 : 100 : 100.00000000000000000000 1 : 95 : 80.00000000000000000000 : 87.05505632961241391300 2 : 90 : 64.00000000000000000000 : 75.78582832551990411700 3 : 85 : 51.20000000000000000000 : 65.97539553864471296900 4 : 80 : 40.96000000000000000000 : 57.43491774985175034000 5 : 75 : 32.76800000000000000000 : 50.00000000000000000000 6 : 70 : 26.21440000000000000000 : 43.52752816480620695700 7 : 65 : 20.97152000000000000000 : 37.89291416275995205900 8 : 60 : 16.77721600000000000000 : 32.98769776932235648400 9 : 55 : 13.42177280000000000000 : 28.71745887492587517000 10 : 50 : 10.73741824000000000000 : 25.00000000000000000000 11 : 45 : 8.58993459200000000000 : 21.76376408240310347800 12 : 40 : 6.87194767360000000000 : 18.94645708137997602900 13 : 35 : 5.49755813888000000000 : 16.49384888466117824200 14 : 30 : 4.39804651110400000000 : 14.35872943746293758500 15 : 25 : 3.51843720888320000000 : 12.50000000000000000000 16 : 20 : 2.81474976710656000000 : 10.88188204120155173900 17 : 15 : 2.25179981368524800000 : 9.47322854068998801400 18 : 10 : 1.80143985094819840000 : 8.24692444233058912100 19 : 5 : 1.44115188075855872000 : 7.17936471873146879200 (nice level : old % : new % : my suggested %) Your levels divert very quickly from what they used to be (upto a factor of 7), it's also not really easy to remember what the individual levels mean. I at least try to keep them somewhat in the range they used to be (and the difference is limited to a factor of about 2), also every 5 levels the amount of cpu time is halved, which is very easy to remember. If you need more dynamic range, is there a law that prevents us from going beyond 19? For example: for (i=20;i<=30;i++) print i, " : ", (20-i)*5, " : ", 100*1.25^-i, " : ", e(l(2)*(-i/5))*100, "\n"; 20 : 0 : 1.15292150460684697600 : 6.25000000000000000000 21 : -5 : .92233720368547758000 : ...
Ingo, that _does_ sound excessive. How about trying a much less aggressive nice-level (and preferably linear, not exponential)? Linus -
I actually like the extra range, it allows for a much softer punch of background tasks even on somewhat slower boxen. I've been testing CFS on my 1200 MHz lappy for some time and a strongly niced kbuild leaves a very usable system. The old scheduler would leave the thing rather jumpy. And while CFS fully fixes the jumpyness, I just did a nice +13 (which should be equivalent to the old schedulers nice +19 for my HZ) and did a nice +19 kbuild and I can definitely feel the difference between them. Early CFS versions had an pretty aggressive nice range (0.1% for +19), and that has been toned down based on feedback. The current levels seem to work well, at least on my boxen. - Peter -
Hi, The extra range is not really a problem, in http://www.ussg.iu.edu/hypermail/linux/kernel/0707.2/0850.html I suggested how we can have both. bye, Roman -
By breaking the UNIX model of nice levels. Not an option in my book. -
yeah, that's pretty much out of question. Ingo -
Hi, Breaking user expectations of nice levels is? bye, Roman -
http://www.opengroup.org/onlinepubs/009695399/basedefs/xbd_chap03.html specifically: "3.239 Nice Value A number used as advice to the system to alter process scheduling. Numerically smaller values give a process additional preference when scheduling a process to run. Numerically larger values reduce the preference and make a process less likely to run. Typically, a process with a smaller nice value runs to completion more quickly than an equivalent process with a higher nice value. The symbol {NZERO} specifies the default nice value of the system." The only expectation is that a process with a lower nice level gets more time. Any other expectation is a bug. -
Hi, Yes, users are buggy, they expect a lot of stupid things... Is this really reason enough to break this? What exactly is the damage if setpriority() accepts a few more levels? bye, Roman -
_changing_ it is an option within reason, and we've done it a couple of times already in the past, and even within CFS (as Peter correctly observed) we've been through a couple of iterations already. And as i mentioned it before, the outer edge of nice levels (+19, by far the most commonly used nice level) was inconsistent to begin with: 3%, 5%, 9% of nice-0, depending on HZ. So changing that to a consistent (and user-requested) 1.5% is a much smaller change than you seem to make it out to be. CFS itself is a far larger "change of expectations" than this tweak to nice levels. So by your standard we could never change the scheduler. (which your ultimate argument might be after all =B-) Ingo -
Hi, Why do you constantly stress level 19? Yes, that one is special, all other How old is CFS and how many users did it have so far? How many users has The percentage levels are off by a factor of upto _seven_, sorry I fail Careful, you make assertion about me, for which you have absolutely no base, adding a smiley doesn't make this any funnier. bye, Roman -
i constantly stress it for the reason i mentioned a good number of
times: because it's by far the most commonly used (and complained about)
nice level. =B-)
but because you are asking, i'm glad to give you some first-hand
historic background about Linux nice levels (in case you are interested)
and the motivations behind their old and new implementations:
nice levels were always so weak under Linux (just read Peter's report)
that people continuously bugged me about making nice +19 tasks use up
much less CPU time. Unfortunately that was not that easy to implement
(otherwise we'd have done it long ago) because nice level support was
historically coupled to timeslice length, and timeslice units were
driven by the HZ tick, so the smallest timeslice was 1/HZ.
In the O(1) scheduler (about 4 years ago) i changed negative nice levels
to be much stronger than they were before in 2.4 (and people were happy
about that change), and i also intentionally calibrated the linear
timeslice rule so that nice +19 level would be _exactly_ 1 jiffy. To
better understand it, the timeslice graph went like this (cheesy ASCII
art alert!):
A
\ | [timeslice length]
\ |
\ |
\ |
\ |
\|___100msecs
|^ . _
| ^ . _
| ^ . _
-*----------------------------------*-----> [nice level]
-20 | +19
|
|
so that if someone wants to really renice tasks, +19 would give a much
bigger hit than the normal linear rule would do. (The solution of
changing the ABI to extend priorities was discarded early on.)
This approach worked to some degree for some time, but later on with
HZ=1000 it caused 1 jiffy to be 1 msec, which meant 0.1% CPU usage which
we felt to be a bit excessive. Excessive _not_ because it's too small of
a CPU ...Hi, I guess I should be thankful now? I'm curious why you post this now, after I "asked" about this. Most of the information is either rather generic or not specific enough for the problem at hand. If you had posted this information earlier, it had been far more valueable as it could have been a nice base for a discussion. But posting it this late I can't lose the feeling you're more interested Not completely. For negative nice levels you mentioned audio apps, but these aren't really interested in a fair share, they would use the higher percentage only to guarantee they get the amount of time they need independent of the current load. I think they would be better served with e.g. a deadline scheduler, which guarantees them an absolute time share not a relative one. On the other end with positive levels I more remember requests for something closer to idle scheduling, where a process only runs when nothing else is running. So assuming we had scheduling classes for the above use cases, what other reasons are left for such extreme nice levels? My proposed nice levels have otherwise the same properties as yours (e.g. being consistent). There is one propery you haven't commented on at all yet. My proposed levels give the average use a far better idea what they actually mean, i.e. that every 5 levels the process gets double/halve the cpu time. This is IMO a considerable advantage. bye, Roman -
Hi, BTW what is the "UNIX model of nice levels"? SUS specifies the limit via NZERO, which is defined as "Minimum Acceptable Value: 20", I can't find any information that it must be 20. bye, Roman -
I have never encountered a UNIX where it is anything other than 20. Convention (alas not specification) does dictate 20. -
Hi, I think the exponential increase isn't the problem. The old code did approximate something like this rather crudely with the result that there was a big gap between level 0 and -1. Something like this: echo 'for (i=-20;i<=20;i++) print i, " : ", 1024*e(l(2)*(-i/20*3)), "\n";' | bc -l would produce a range similiar to the old code. Replacing the factor 3 with 4 would be IMO a more reasonable increase and had the advantage for the user that it's easier to understand that every 5 levels the time a process gets is doubled. bye, Roman -
Hi, I more meant serious attempts. At this point I'm not that much interested in a few localized optimizations, what I'm interested in is how can this optimized at the design level (e.g. how can arch information be used to simplify things). So I spent quite a bit of time looking through cfs and experimenting with some ideas. I want to put the main focus on the performance aspect, but there are a few other issues as well. But first something else (especially for Ingo): I tried to be very careful with any claims made in this mail, but this of course doesn't exclude the possibility of errors, in which case I'd appreciate any corrections. Any explanations done in this mail don't imply that anyone needs any such explanations, they're done to keep things in context, so that interested readers have a chance to follow even if they don't have the complete background information. Any suggestions made don't imply that they have to be implemented like this, there are more an incentive for further discussion and I'm always interested in better solutions. A first indication that something may not be quite right is the increase in code size: 2.6.22: text data bss dec hex filename 10150 24 3344 13518 34ce kernel/sched.o recent git: text data bss dec hex filename 14724 228 2020 16972 424c kernel/sched.o That's i386 without stats/debug. A lot of the new code is in regularly executed regions and it's often not exactly trivial code as cfs added lots of heavy 64bit calculations. With the increased text comes increased runtime memory usage, e.g. task_struct increased so that only 5 of them instead 6 fit now into 8KB. Since sched-design-CFS.txt doesn't really go into any serious detail, so the EEVDF paper was more helpful and after playing with the ideas a little I noticed that the whole idea of fair scheduling can be explained somewhat simpler and I'm a little surprised not finding it ...
Hi Roman,
Took me most of today trying to figure out WTH you did in fs2.c, more
math and fundamental explanations would have been good. So please bear
with me as I try to recap this thing. (No, your code was very much _not_
obvious, a few comments and broken out functions would have made a world
of a difference)
So, for each task we keep normalised time
normalised time := time/weight
using Bresenham's algorithm we can do this prefectly (up until a renice
- where you'd get errors)
avg_frac += weight_inv
weight_inv = X / weight
avg = avg_frac / weight0_inv
weight0_inv = X / weight0
avg = avg_frac / (X / weight0)
= (X / weight) / (X / weight0)
= X / weight * weight0 / X
= weight0 / weight
So avg ends up being in units of [weight0/weight].
Then, in order to allow sleeping, we need to have a global clock to sync
with. Its this global clock that gave me headaches to reconstruct.
We're looking for a time like this:
rq_time := sum(time)/sum(weight)
And you commented that the /sum(weight) part is where CFS obtained its
accumulating rounding error? (I'm inclined to believe the error will
statistically be 0, but I'll readily accept otherwise if you can show a
practical 'exploit')
Its not obvious how to do this using modulo logic like Bresenham because
that would involve using a gcm of all possible weights.
What you ended up with is quite interesting if correct.
sum_avg_frac += weight_inv_{i}
however by virtue of the scheduler minimising:
avg_{i} - avg_{j} | i != j
this gets a factor of:
weight_{i}/sum_{j}^{N}(weight_{j})
( seems correct, needs more analysis though, this is very much a
statistical step based on the previous constraint. this might
very well introduce some errors )
resulting in:
sum_avg_frac += sum_{i}^{N}(weight_inv_{i} *
weight_{i}/sum_{j}^{N}(weight_{j}))
...Hi, Thanks for the effort though. :) I know I'm not the best explaining these things, so I really appreciate I think I've sent you off into the wrong direction somehow. Sorry. :) Let's ignore the average for a second, normalized time is maintained as: normalized time := time * (2^16 / weight) The important point is that I keep the value in full resolution of 2^-16 vsec units (vsec for virtual second or sec/weight, where every tasks gets weight seconds for every virtual second, to keep things simpler I also omit the nano prefix from the units for a moment). Compared to that CFS maintains a global normalized value in 1 vsec units. Since I don't round the value down I avoid the accumulating error, this means that time_norm += time_delta1 * (2^16 / weight) time_norm += time_delta2 * (2^16 / weight) is the same as time_norm += (time_delta1 + time_delta2) * (2^16 / weight) CFS for example does this delta_mine = calc_delta_mine(delta_exec, curr->load.weight, lw); in above terms this means time = time_delta * weight * (2^16 / weight_sum) / 2^16 The last shift now rounds the value down and if one does that 1000 times per second, the resolution of the value that is finally accounted to wait_runtime is also reduced appropriately. The other rounding problem is based on that this term x * prio_to_weight[i] * prio_to_wmult[i] / 2^32 doesn't produce x for most values in that tables (the same applies to the weight sum), so if we have chains, where the values are converted from one scale to the other, a rounding error is produced. In CFS this happens now because wait_runtime is maintained in nanoseconds and fair_clock is a normalized value. The problem here isn't that these errors might have a statistical relevance, as they are usually completely overshadowed by measurement errors anyway. The problem is that these errors exist at all, this means they have to be compensated somehow, so that they don't accumulate over time and then ...
Roman, Thanks for the testing and the feedback, it's much appreciated! :-) On what platform did you do your tests, and what .config did you use (and could you please send me your .config)? Please also send me the output of this script: http://people.redhat.com/mingo/cfs-scheduler/tools/cfs-debug-info.sh (if the output is too large send it to me privately, or bzip2 -9 it.) Could you also please send the source code for the "l.c" and "lt.c" apps you used for your testing so i can have a look. Thanks! Ingo -
I haven't been able to reproduce this with any combination of features, and massive_intr tweaked to his work/sleep cycle. I notice he's collecting stats though, and they look funky. Recompiling. -Mike -
yeah, the posted numbers look most weird, but there's a complete lack of any identification of test environment - so we'll need some more word from Roman. Perhaps this was run on some really old box that does not have a high-accuracy sched_clock()? The patch below should simulate that scenario on 32-bit x86. Ingo Index: linux/arch/i386/kernel/tsc.c =================================================================== --- linux.orig/arch/i386/kernel/tsc.c +++ linux/arch/i386/kernel/tsc.c @@ -110,7 +110,7 @@ unsigned long long native_sched_clock(vo * very important for it to be as fast as the platform * can achive it. ) */ - if (unlikely(!tsc_enabled && !tsc_unstable)) +// if (unlikely(!tsc_enabled && !tsc_unstable)) /* No locking but a rare wrong value is not a big deal: */ return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ); -
Ah, thanks. I noticed that clocksource= went away. I'll test with stats, with and without jiffies resolution. -Mike -
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ P COMMAND 6465 root 20 0 1432 356 296 R 30 0.0 1:02.55 1 chew 6462 root 20 0 1576 216 140 R 23 0.0 0:50.29 1 massive_intr_x 6463 root 20 0 1576 216 140 R 23 0.0 0:50.23 1 massive_intr_x 6464 root 20 0 1576 216 140 R 23 0.0 0:50.28 1 massive_intr_x Well, jiffies resolution clock did upset fairness a bit with a right at jiffies resolution burn time, but not nearly as bad as on Roman's box, and not in favor of the sleepers. With the longer burn time of stock massive_intr.c (8ms burn, 1ms sleep), lower resolution clock didn't upset it. PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ P COMMAND 6511 root 20 0 1572 220 140 R 25 0.0 1:00.11 1 massive_intr 6512 root 20 0 1572 220 140 R 25 0.0 1:00.14 1 massive_intr 6514 root 20 0 1432 356 296 R 25 0.0 1:00.31 1 chew 6513 root 20 0 1572 220 140 R 24 0.0 1:00.14 1 massive_intr -Mike -
Hi,
l.c is a simple busy loop (well, with the option to start many of them).
This is lt.c, what it does is to run a bit less than a jiffie, so it
needs a low resolution clock to trigger the problem:
#include <stdio.h>
#include <signal.h>
#include <time.h>
#include <sys/time.h>
#define NSEC 1000000000
#define USEC 1000000
#define PERIOD (NSEC/1000)
int i;
void worker(int sig)
{
struct timeval tv;
long long t0, t;
gettimeofday(&tv, 0);
//printf("%u,%lu\n", i, tv.tv_usec);
t0 = (long long)tv.tv_sec * 1000000 + tv.tv_usec + PERIOD / 1000 - 50;
do {
gettimeofday(&tv, 0);
t = (long long)tv.tv_sec * 1000000 + tv.tv_usec;
} while (t < t0);
}
int main(int ac, char **av)
{
int cnt;
timer_t timer;
struct itimerspec its;
struct sigaction sa;
cnt = i = atoi(av[1]);
sa.sa_handler = worker;
sa.sa_flags = 0;
sigemptyset(&sa.sa_mask);
sigaction(SIGALRM, &sa, 0);
clock_gettime(CLOCK_MONOTONIC, &its.it_value);
its.it_interval.tv_sec = 0;
its.it_interval.tv_nsec = PERIOD * cnt;
while (--i > 0 && fork() > 0)
;
its.it_value.tv_nsec += i * PERIOD;
if (its.it_value.tv_nsec > NSEC) {
its.it_value.tv_sec++;
its.it_value.tv_nsec -= NSEC;
}
timer_create(CLOCK_MONOTONIC, 0, &timer);
timer_settime(timer, TIMER_ABSTIME, &its, 0);
printf("%u,%lu\n", i, its.it_interval.tv_nsec);
while (1)
pause();
return 0;
}
-
thanks. Just to make sure, while you said that your TSC was off on that laptop, the bootup log of yours suggests a working TSC: Time: tsc clocksource has been installed. and still your fl.c testcases produces the top output that you've reported in your first mail? If so then this could be a regression. Or did you turn off the tsc manually via notsc? (or was it with a different .config or on a different machine)? Please help us figure this out exactly, we dont want a real regression go unnoticed. If you can reproduce that problem with a working TSC then please generate a second cfs-debug-info.sh snapshot _while_ your fl+l workload is running and send that to me (i'll reply back to it publicly). Thanks, Ingo -
Standard kernels often disable the TSC later after running a bit with it (e.g. on any cpufreq change without p state invariant TSC) -Andi -
I assume that what Roman hit was that he had explicitly disabled the TSC because of TSC instability with the "notsc" kernel command line. Which disabled is *entirely*. That *used* to be the right thing to do, since the gettimeofday() logic originally didn't know about TSC instability, and it just resulted in somewhat flaky timekeeping. These days, of course, we should notice it on our own, and just switch away from the TSC as a reliable clock-source, but still allow it to be used for the cases where absolute accuracy is not a big issue. So I suspect that Roman - by virtue of being an old-timer - ends up having a workaround for an old problem that isn't needed, and that in turn ends up meaning that his scheduler clock also ends up using the really not very good timer tick.. Linus -
It might just have been cpufreq. That nearly hits everybody with cpufreq unless you have a pstate invariant TSC; and that's pretty much always the case on older laptops. It used to not be that drastic, but since i386 switched to the The rewritten sched_clock() i still have queued does just that. I planned to submit it for .23, but then during later in deepth testing on my machine park I found a show stopper that I couldn't fix on time. Hopefully for .24 -Andi -
but that does not appear to be the case, the debug info i got from Roman includes the following boot options: Kernel command line: auto BOOT_IMAGE=2.6.23-rc1-git9 ro root=306 there's no "notsc" option there. Andi's theory cannot be true either, Roman's debug info also shows this /proc/<PID>/sched data: clock-delta : 95 that means that sched_clock() is in high-res mode, the TSC is alive and kicking and a sched_clock() call took 95 nanoseconds. Roman, could you please help us with this mystery? Ingo -
Hi, Actually, Andi is right. What I sent you was generated directly after boot, as I had to reboot for the right kernel, so a little later appeared this: Aug 1 14:54:30 spit kernel: eth0: link up, 100Mbps, full-duplex, lpa 0x45E1 Aug 1 15:09:56 spit kernel: Clocksource tsc unstable (delta = 656747233 ns) Aug 1 15:09:56 spit kernel: Time: pit clocksource has been installed. bye, Roman -
just to make sure, how does 'top' output of the l + "lt 3" testcase look like now on your laptop? Yesterday it was this: 4544 roman 20 0 1796 520 432 S 32.1 0.4 0:21.08 lt 4545 roman 20 0 1796 344 256 R 32.1 0.3 0:21.07 lt 4546 roman 20 0 1796 344 256 R 31.7 0.3 0:21.07 lt 4547 roman 20 0 1532 272 216 R 3.3 0.2 0:01.94 l and i'm still wondering how that output was possible. Ingo -
Hi, I disabled the jiffies logic and the result is still the same, so this problem isn't related to resolution at all. I traced it a little and what's happing is that the busy loop really only gets little time, it only runs inbetween the timer tasks. When the timer task is woken up __enqueue_sleeper() updates sleeper_bonus and a little later when the busy loop is preempted __update_curr() is called a last time and it's fully hit by the sleeper_bonus. So the timer tasks use less time than they actually get and thus produce overflows, the busy loop OTOH is punished and underflows. So it seems my initial suspicion was right and this logic is dodgy, what is it actually supposed to do? Why is some random task accounted with the sleeper_bonus? bye, Roman PS: Can I still expect answer about all the other stuff? -
how did you disable the jiffies logic? Also, could you please send me the cfs-debug-info.sh: http://people.redhat.com/mingo/cfs-scheduler/tools/cfs-debug-info.sh captured _while_ the above workload is running. This is the third time i've asked for that :-) to establish that the basic sched_clock() behavior is sound on that box, could you please also run this tool: http://people.redhat.com/mingo/cfs-scheduler/tools/tsc-dump.c please run it both while the system is idle, and while there's a CPU hog running: while :; do :; done & and send me that output too? (it's 2x 60 lines only) Thanks! Ingo -
Hi,
Is there any reason to believe my analysis is wrong?
So far you haven't answered a single question about the CFS design...
Anyway, I give you something better - the raw trace data for 2ms:
1186747669.274790012: update_curr 0xc7fb06f0,479587,319708,21288884188,159880,7360532
1186747669.274790375: dequeue_entity 0xc7fb06f0,21280402988,159880
1186747669.274792580: sched 2848,2846,0xc7432cb0,-7520413
1186747669.274820987: update_curr 0xc7432ce0,29302,-130577,21288913490,1,-7680293
1186747669.274821269: dequeue_entity 0xc7432ce0,21296077409,1
1186747669.274821930: enqueue_entity 0xc7432ce0,21296593783,1
1186747669.274826979: update_curr 0xc7432ce0,5707,5707,21288919197,1,-7680294
1186747669.274827724: enqueue_entity 0xc7432180,21280919197,639451
1186747669.274829948: update_curr 0xc7432ce0,1553,-318172,21288920750,319726,-8000000
1186747669.274831878: sched 2846,2847,0xc7432150,8000000
1186747669.275789883: update_curr 0xc7432180,479797,319935,21289400547,159864,7360339
1186747669.275790295: dequeue_entity 0xc7432180,21280919197,159864
1186747669.275792439: sched 2847,2846,0xc7432cb0,-7520203
1186747669.275820819: update_curr 0xc7432ce0,29238,-130625,21289429785,1,-7680067
1186747669.275821109: dequeue_entity 0xc7432ce0,21296593783,1
1186747669.275821763: enqueue_entity 0xc7432ce0,21297109852,1
1186747669.275826887: update_curr 0xc7432ce0,5772,5772,21289435557,1,-7680068
1186747669.275827652: enqueue_entity 0xc7fb0ca0,21281435557,639881
1186747669.275829826: update_curr 0xc7432ce0,1549,-318391,21289437106,319941,-8000000
1186747669.275831584: sched 2846,2849,0xc7fb0c70,8000000
About the values:
update_curr: sched_entity, delta_fair, delta_mine, fair_clock, sleeper_bonus, wait_runtime
(final values at the end of __update_curr)
{en,de}queue_entity: sched_entity, fair_key, sleeper_bonus
(at the start of __enqueue_entity/__dequeue_entity)
sched: prev_pid,pid,current,wait_runtime
(at the end of scheduling, note that current has a small structure
offset to ...please first give me the debug data captured with the script above (while the workload is running) - so that i can see the full picture about what's happening. Thanks, Ingo -
I guess I'm going to have to give up on trying to reproduce this... my 3GHz P4 is just not getting there from here. Last attempt, compiled UP, HZ=1000 dynticks, full preempt and highres timers fwiw. 6392 root 20 0 1696 332 248 R 25.5 0.0 3:00.14 0 lt 6393 root 20 0 1696 332 248 R 24.9 0.0 3:00.15 0 lt 6391 root 20 0 1696 488 404 R 24.7 0.0 3:00.20 0 lt 6394 root 20 0 2888 1232 1028 R 24.5 0.1 2:58.58 0 sh -Mike -
Hi, Except for UP and HZ=1000, everything else is pretty much turned off. If you use a very recent kernel, the problem may not be visible like this anymore. It may be a bit easier to reproduce, if you change the end time t0 in lt.c a little. Also try to start the busy loop first. bye, Roman -
Not yet, but if you give Ingo what he wants (as opposed to what you're giving him) it'll be easier for him to answer what's going wrong, and perhaps "fix" the problem to boot. (The scripts gives info about CPU characteristics, interrupts, modules, etc. -- you know, all those "unknown" variables.) And perhaps a patch to show what parts you commented out, too, so one can tell if anything got broken (unintentionally). -- Michael Chang Please avoid sending me Word or PowerPoint attachments. Send me ODT, RTF, or HTML instead. See http://www.gnu.org/philosophy/no-word-attachments.html Thank you. -
Hi, He already has most of this information and the trace shows _exactly_ what's going on. All this information should be more than enough to allow an initial judgement whether my analysis is correct. Also none of this information is needed to explain the CFS logic a little more, which I'm still waiting for... bye, Roman -
i'll need the other bits of information too to have a complete picture about what's going on while your test is running - to maximize the chances of me being able to fix it. I'm a bit perplexed (and a bit worried) about this - you've spent _far_ more effort to _not send_ that script output (captured while the workload is running) than it would have taken to do it :-/ If you'd like me to fix bugs then please just send it (in private mail if you want) - or give me an ssh login to that box - whichever variant you prefer. Thanks, Ingo -
Roman, fortunately all bug reporters are not like you. It's amazing how long you can resist sending a simple bug report to a developer! Maybe you consider that you need to fix the bug by yourself after you understand the code, but if you systematically refuse to return the small information Ingo asks you, we will have to wait for some more cooperative users to be hit by the same bug when 2.6.23 is released, which is stupid. I thought you could at least understand that one developer who is used to read traces from the same tool every day will be far faster at decoding a trace from the same tool than trying to figure out what your self-maid dump means. It's the exact same reason I ask for pcap files when people send me outputs of tcpdumps without the information I *need*. I you definitely do not want to cooperate, stop asking for a personal explanation, and go figure by yourself how the code works. BTW, in the trace you "kindly offered" in exchange for the cfs-debug-info dump, you show several useful variables, but nothing says where they are captured. And as you can see, they're changing. That's a fantastic trace for a developer, really... Please try to be a little bit more transparent if you really want the bugs fixed, and don't behave as if you wanted this bug to survive till -final. Thanks, Willy -
Hi, I'm more amazed how long Ingo can resist providing some explanations (not just about this problem). It's not like I haven't given him anything, he already has the test programs, he already knows the system configuration. Could you please ask Ingo the same? I'm simply trying to get some transparancy into the CFS design. Without further information it's difficult to tell, whether something is supposed to work this way or it's a bug. In this case it's quite possible that due to a recent change my testcase doesn't work anymore. Should I consider the problem fixed or did it just go into hiding? Without more information it's difficult to verify this independently. bye, Roman -
one more small thing: could you please send your exact .config (Mike asked for that too, and i too on two prior occasions). Sometimes unexpected little details in the .config make a difference, we are not asking you that because we are second-guessing you in any way, the reason is simple: i frequently boot _the very .config that others use_, and see surprising reproducability of bugs that i couldnt trigger before. It's standard procedure to just pick up the .config of others to eliminate a whole bunch of degrees of freedom for a bug to hide behind - and your "it's a pretty standard config" description doesnt really achieve that. It probably wont make a real difference, but it's really easy for you to send and it's still very useful when one tries to eliminate possibilities and when one wants to concentrate on the remaining possibilities alone. Thanks again, Ingo -
Hi, The thing I'm afraid about CFS is its possible unpredictability, which would make it hard to reproduce problems and we may end up with users with unexplainable weird problems. That's the main reason I'm trying so hard to push for a design discussion. Just to give an idea here are two more examples of irregular behaviour, which are hopefully easier to reproduce. 1. Two simple busy loops, one of them is reniced to 15, according to my calculations the reniced task should get about 3.4% (1/(1.25^15+1)), but I get this: PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 4433 roman 20 0 1532 300 244 R 99.2 0.2 5:05.51 l 4434 roman 35 15 1532 72 16 R 0.7 0.1 0:10.62 l OTOH upto nice level 12 I get what I expect. 2. If I start 20 busy loops, initially I see in top that every task gets 5% and time increments equally (as it should): PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 4492 roman 20 0 1532 68 16 R 5.0 0.1 0:02.86 l 4491 roman 20 0 1532 68 16 R 5.0 0.1 0:02.86 l 4490 roman 20 0 1532 68 16 R 5.0 0.1 0:02.86 l 4489 roman 20 0 1532 68 16 R 5.0 0.1 0:02.86 l 4488 roman 20 0 1532 68 16 R 5.0 0.1 0:02.86 l 4487 roman 20 0 1532 68 16 R 5.0 0.1 0:02.86 l 4486 roman 20 0 1532 68 16 R 5.0 0.1 0:02.86 l 4485 roman 20 0 1532 68 16 R 5.0 0.1 0:02.86 l 4484 roman 20 0 1532 68 16 R 5.0 0.1 0:02.86 l 4483 roman 20 0 1532 68 16 R 5.0 0.1 0:02.86 l 4482 roman 20 0 1532 68 16 R 5.0 0.1 0:02.86 l 4481 roman 20 0 1532 68 16 R 5.0 0.1 0:02.86 l 4480 roman 20 0 1532 68 16 R 5.0 0.1 0:02.86 l 4479 roman 20 0 1532 68 16 R 5.0 0.1 0:02.86 l 4478 roman 20 0 1532 68 16 R 5.0 0.1 0:02.86 l 4477 roman 20 0 1532 68 16 R 5.0 0.1 0:02.86 l 4476 roman 20 0 ...
You may be interested by looking at the very early CFS versions. The design was much more naive and understandable. After that, a lot of tricks have been added to take into account a lot of uses and corner cases, which may Do you see this only at -15, or starting with -15 and below ? Willy -
note that the typo was not in the weight table but in the inverse weight table which didnt really affect CPU utilization (that's why we didnt notice the typo sooner). Regarding the above problem with nice +15 being beefier than intended i'd suggest to re-test with a doubled /proc/sys/kernel/sched_runtime_limit value, or with: echo 30 > /proc/sys/kernel/sched_features i think this was scheduling jitter caused by the larger granularity of negatively reniced tasks. This got improved recently, with latest -git i get: PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 3108 root 5 -15 1576 248 196 R 5.0 0.0 0:07.26 loop_silent 3109 root 5 -15 1576 248 196 R 5.0 0.0 0:07.26 loop_silent 3110 root 5 -15 1576 248 196 R 5.0 0.0 0:07.26 loop_silent 3111 root 5 -15 1576 244 196 R 5.0 0.0 0:07.26 loop_silent 3112 root 5 -15 1576 248 196 R 5.0 0.0 0:07.26 loop_silent 3113 root 5 -15 1576 248 196 R 5.0 0.0 0:07.26 loop_silent that's picture-perfect CPU time distribution. But, and that's fair to say, i never ran such an artificial workload of 20x nice -15 infinite loops (!) before, and boy does interactivity suck (as expected) ;) Ingo -
everything looks good in your debug output and the TSC dump data, except
for the wait_runtime values, they are quite out of balance - and that
balance cannot be explained with jiffies granularity or with any sort of
sched_clock() artifact. So this clearly looks like a CFS regression that
should be fixed.
the only relevant thing that comes to mind at the moment is that last
week Peter noticed a buggy aspect of sleeper bonuses (in that we do not
rate-limit their output, hence we 'waste' them instead of redistributing
them), and i've got the small patch below in my queue to fix that -
could you give it a try?
this is just a blind stab into the dark - i couldnt see any real impact
from that patch in various workloads (and it's not upstream yet), so it
might not make a big difference. The trace you did (could you send the
source for that?) seems to implicate sleeper bonuses though.
if this patch doesnt help, could you check the general theory whether
it's related to sleeper-fairness, via turning it off:
echo 30 > /proc/sys/kernel/sched_features
does the bug go away if you do that? If sleeper bonuses are showing too
many artifacts then we could turn it off for final .23.
Ingo
--------------------->
Subject: sched: fix sleeper bonus
From: Ingo Molnar <mingo@elte.hu>
Peter Ziljstra noticed that the sleeper bonus deduction code was not
properly rate-limited: a task that scheduled more frequently would get a
disproportionately large deduction. So limit the deduction to delta_exec
and limit production to runtime_limit.
Not-Yet-Signed-off-by: Ingo Molnar <mingo@elte.hu>
---
kernel/sched_fair.c | 12 ++++++------
1 file changed, 6 insertions(+), 6 deletions(-)
Index: linux/kernel/sched_fair.c
===================================================================
--- linux.orig/kernel/sched_fair.c
+++ linux/kernel/sched_fair.c
@@ -75,7 +75,7 @@ enum {
unsigned int sysctl_sched_features __read_mostly =
SCHED_FEAT_FAIR_SLEEPERS *1 ...Hi, It doesn't make much of a difference. OTOH if I disabled the sleeper code completely in __update_curr(), I get this: PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 3139 roman 20 0 1796 344 256 R 21.7 0.3 0:02.68 lt 3138 roman 20 0 1796 344 256 R 21.7 0.3 0:02.68 lt 3137 roman 20 0 1796 520 432 R 21.7 0.4 0:02.68 lt 3136 roman 20 0 1532 268 216 R 34.5 0.2 0:06.82 l Disabling this code completely via sched_features makes only a minor difference: PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 3139 roman 20 0 1796 344 256 R 20.4 0.3 0:09.94 lt 3138 roman 20 0 1796 344 256 R 20.4 0.3 0:09.94 lt 3137 roman 20 0 1796 520 432 R 20.4 0.4 0:09.94 lt Can we please skip to the point, where you try to explain the intention a little more? If I had to guess that this is supposed to keep the runtime balance, then it would be better to use wait_runtime to adjust fair_clock, from where it would be evenly distributed to all tasks (but this had to be done during enqueue and dequeue). OTOH this also had then a consequence for the wait queue, as fair_clock is used to calculate fair_key. IMHO current wait_runtime should have some influence in calculating the sleep bonus, so that wait_runtime doesn't constantly overflow for tasks which only run occasionally. bye, Roman -
I thought this was history. With your config, I was finally able to reproduce the anomaly (only with your proggy though), and Ingo's patch does indeed fix it here. Freshly reproduced anomaly and patch verification, running 2.6.23-rc3 with your config, both with and without Ingo's patch reverted: 6561 root 20 0 1696 492 404 S 32.0 0.0 0:30.83 0 lt 6562 root 20 0 1696 336 248 R 32.0 0.0 0:30.79 0 lt 6563 root 20 0 1696 336 248 R 32.0 0.0 0:30.80 0 lt 6564 root 20 0 2888 1236 1028 R 4.6 0.1 0:05.26 0 sh 6507 root 20 0 2888 1236 1028 R 25.8 0.1 0:30.75 0 sh 6504 root 20 0 1696 492 404 R 24.4 0.0 0:29.26 0 lt 6505 root 20 0 1696 336 248 R 24.4 0.0 0:29.26 0 lt 6506 root 20 0 1696 336 248 R 24.4 0.0 0:29.25 0 lt -Mike -
oh, great! I'm glad we didnt discard this as a pure sched_clock resolution artifact. Roman, a quick & easy request: please send the usual cfs-debug-info.sh output captured while your testcase is running. (Preferably try .23-rc3 or later as Mike did, which has the most recent scheduler code, it includes the patch i sent to you already.) I'll reply to your sleeper-fairness questions separately, but in any case we need to figure out what's happening on your box - if you can still reproduce it with .23-rc3. Thanks, Ingo -
Hi, I did update to 2.6.23-rc3-git1 first, but I ended up reverting the patch, as I didn't notice it had been applied already. Sorry about that. With this patch the underflows are gone, but there are still the overflows, so the questions from the last mail still remain. bye, Roman -
It's a matter of time balance. It takes a short time to send the output of a script, and it takes a very long time to explain how things work. I often encounter the same situation with haproxy. People ask me to explain them in detail how this or that would apply to their context, and it's often easier for me to provide them with a 5-lines patch to add the feature they need, than to spend half an hour explaining why and how it I know that Ingo tends to reply to a question with another question. But as I said, imagine if he has to explain the same things to each person who asks him for it. I think that a more constructive approach would be to point what is missing/unclear/inexact in the doc so that he adds some paragraphs for you and everyone else. If you need this information to debug, generally, problems that appear only on one person's side and which suddenly disappear are either caused by some random buggy patch left in the tree (not your case it seems), or by an obscure bug of the feature being tested which will resurface from time to time as long as it's not identified. Willy -
I still can't reproduce this here. Can you please send your .config, so I can try again with a config as close to yours as possible? -Mike -
Mike and me have managed to reproduce similarly looking 'top' output, but it takes some effort: we had to deliberately run a non-TSC sched_clock(), CONFIG_HZ=100, !CONFIG_NO_HZ and !CONFIG_HIGH_RES_TIMERS. in that case 'top' accounting symptoms similar to the above are not due to the scheduler starvation you suspected, but due the effect of a low-resolution scheduler clock and a tightly coupled timer/scheduler tick to it. I tried the very same workload on 2.6.22 (with the same .config) and i saw similarly anomalous 'top' output. (Not only can one create really anomalous CPU usage, one can completely hide tasks from 'top' output.) if your test-box has a high-resolution sched_clock() [easily possible] then please send us the lt.c and l.c code so that we can have a look. Ingo -
Hi, I used my old laptop for these tests, where tsc is indeed disabled due to Well, it magnifies the rounding problems in CFS. I mainly wanted to test a little the behaviour of CFS and I thought a saw patch which enabled the use of TSC in these cases, so I didn't check sched_clock(). Anyway, I want to point out that this wasn't the main focus of what I wrote. bye, Roman -
why do you say that? 2.6.22 behaves similarly with a low-res
sched_clock(). This has nothing to do with 'rounding problems'!
i tried your fl.c and if sched_clock() is high-resolution it's scheduled
_perfectly_ by CFS:
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
5906 mingo 20 0 1576 244 196 R 71.2 0.0 0:30.11 l
5909 mingo 20 0 1844 344 260 S 9.6 0.0 0:04.02 lt
5907 mingo 20 0 1844 508 424 S 9.5 0.0 0:04.01 lt
5908 mingo 20 0 1844 344 260 S 9.5 0.0 0:04.02 lt
if sched_clock() is low-resolution then indeed the 'lt' tasks will
"hide":
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
2366 mingo 20 0 1576 248 196 R 99.9 0.0 0:07.95 loop_silent
1 root 20 0 2132 636 548 S 0.0 0.0 0:04.64 init
but that's nothing new. CFS cannot conjure up time measurement methods
that do not exist. If you have a low-res clock and if you create an app
that syncs precisely to the tick of that clock via timers that run off
that exact tick then there's nothing the scheduler can do about it. It
is false to charachterise this as 'sleeper starvation' or 'rounding
error' like you did. No amount of rounding logic can create a
high-resolution clock out of thin air.
Ingo
-
Hi, Please calm down. You apparantly already get worked up about one of the secondary problems. I didn't say 'sleeper starvation' or 'rounding error', these are your words and it's your perception of what I said. sched_clock() can have a low resolution, which can be a problem for the scheduler. This is all this program demonstrates. If and how this problem should be solved is a completely different issue, about which I haven't said anything yet and since it's not that important right now I'll leave it at that for now. bye, Roman -
Oh dear :-) It was indeed my preception that yesterday you said: | A problem here is that this can be exploited, if a job is spread over | ^^^^^^^^^^^^^^^^^^^^^ | a few threads, they can get more time relativ to other tasks, e.g. in | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ | this example there are three tasks that run only for about 1ms every | 3ms, but they get far more time than should have gotten fairly: | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ | | 4544 roman 20 0 1796 520 432 S 32.1 0.4 0:21.08 lt | 4545 roman 20 0 1796 344 256 R 32.1 0.3 0:21.07 lt | 4546 roman 20 0 1796 344 256 R 31.7 0.3 0:21.07 lt | 4547 roman 20 0 1532 272 216 R 3.3 0.2 0:01.94 l [ http://lkml.org/lkml/2007/7/31/668 ] ( the underlined portion, in other words, is called 'starvation'.) And again today, i clearly perceived you to say: | > in that case 'top' accounting symptoms similar to the above are not | > due to the scheduler starvation you suspected, but due the effect of | > a low-resolution scheduler clock and a tightly coupled | > timer/scheduler tick to it. | | Well, it magnifies the rounding problems in CFS. [ http://lkml.org/lkml/2007/8/1/153 ] But you are right, that must be my perception alone, you couldnt possibly have said any of that =B-) Or are you perhaps one of those who claims that saying something analogous to sleeper starvation does not equal to talking about 'sleeper starvation' and saying something about 'rounding problems in CFS' does in no way mean you were talking about rounding errors? :-) Ingo -
Hi, *sigh* and here you go off again nitpicking on a minor issue just to prove your point... When I wrote the earlier stuff I hadn't realized it was resolution related, so things have to be put into proper context and you make it yourself a little easy by equating them. Yippi, you found another small error I made, can we drop this now? Please? bye, Roman -
..which is pretty much the state of play for lots of non-x86 hardware. -- Mathematics is the supreme nostalgia of our time. -
question is if it's significantly worse than before. With a 100 or 1000Hz timer, you can't expect perfect fairness just due to the extremely rough measurement of time spent... -
Well, at least we're able to *measure* that task 'l' used 3.3% and that tasks 'lt' used 32%. If we're able to measure it, then that's already fine enough to be able to adjust future timeslices credits. Granted it may be rough for small periods (a few jiffies), but it should be fair for larger periods. Or at least it should *report* some fair distribution. Willy -
but the testcase here uses a LOT shorter time than jiffies... not "a few jiffies". -- if you want to mail me at work (you don't), use arjan (at) linux.intel.com Test the interaction between Linux and your BIOS via http://www.linuxfirmwarekit.org -
But if we rely on the same sampling method, at least we will report something consistent with what happens. And sampling is often the correct method to get finer resolution on a macroscopic scale. I mean, we're telling users that we include the "completely fair scheduler" in 2.6.23, a scheduler which will ensure that all tasks get a fair share of CPU time. A user starts top and sees 33%+32%+32+3% for 4 tasks while he would have expected to see 25%+25%+25%+25%. You can try to explain users that it's the fairest distribution, but they will have a hard time believing it, especially when they measure the time spent on CPU with the "time" command. OK this is all sampling, but we should try to avoid relying on different sources of data for computation and reporting. Time and Top should report something close to 4*25% for comparable tasks. And if not, because of some sampling problem, maybe the scheduler cannot be that fair in some situations, but either it should make use of the sampling time and top use, or top and time should rely on the view of the scheduler. I'll try to quickly hack up a program which makes use of rdtsc from userspace to precisely measure user-space time, and disable TSC use from the kernel to see how the values diverge. Regards, Willy -
Indeed. I'm just pointing out that not having TSC, fast HZ, no-HZ mode, or high-res timers should not be treated as an unusual circumstance. That's a PC-centric view. -- Mathematics is the supreme nostalgia of our time. -
actually, you dont need high-res or fast HZ or TSC to reduce those timer artifacts: all you need is _two_ (low-res, slow) hw clocks. Most platforms do have that (even the really really cheap ones), but arches do not set up the scheduler tick one of them and the timer tick to the other, and to skew the periodic-timer programming setup a bit (by nature of physics they are usually already skewed a bit) so that the scheduler tick and timer tick are not coupled. This whole thing is not a big deal on embedded anyway. (you dont get students log in to the toaster or to the fridge to run timer exploits, do you? :-) Ingo -
The question is if it would be that hard to add TSC equivalent sched_clock() support to more systems. At least a lot of CPUs I have ever looked at had some kind of fast clock available. Perhaps it's more laziness of the developers or cut'n'paste that these are not as widely used as they should be? -Andi -
that's without CONFIG_SMP, right? :-) On SMP they are about net break
even:
text data bss dec hex filename
26535 4173 24 30732 780c kernel/sched.o-2.6.22
28378 2574 16 30968 78f8 kernel/sched.o-2.6.23-git
(plus a further ~1.5K per CPU data reduction which is not visible here)
btw., here's the general change in size of a generic vmlinux from .22 to
.23-git, using the same .config:
text data bss dec hex filename
5256628 520760 1331200 7108588 6c77ec vmlinux.22
5306918 535844 1327104 7169866 6d674a vmlinux.23-git
+50K. (this was on UP)
In any case, there's still some debugging code in the scheduler (beyond
SCHED_DEBUG), i'll work some more on reducing it.
Ingo
-
Hi, That's still quite an increase in some rather important code paths and it's not just the code size, but also code complexity which is important That's why I mentioned the increased runtime memory usage... bye, Roman -
sched_yield() is being reworked at the moment. But in general we want apps to move away to sane locking constructs ASAP. There's some movement in the 3D space at least. Ingo -
jiffies based sched_clock should be soon very rare. It's probably not worth optimizing for it. -Andi -
Hi, I'm not so sure about that. sched_clock() has to be fast, so many archs may want to continue to use jiffies. As soon as one does that one can also save a lot of computational overhead by using 32bit instead of 64bit. The question is then how easy that is possible. bye, Roman -
i think Andi was talking about the vast majority of the systems out there. For example, check out the arch demography of current Fedora installs (according to the Smolt opt-in UUID based user metrics): http://smolt.fedoraproject.org/ i686: 74743 x86_64: 18599 i386: 1208 ppc: 527 ppc64: 396 sparc64: 14 --------------- Total: 95488 even pure i386 (kernels, not systems) is a only 1.2% of all installs. By the time the CFS kernel gets into a distro (a few months at minimum, typically a year) this percentage will go down further. And embedded doesnt really care about task-statistics corner cases [ (it likely doesnt have 'top' installed - likely doesnt even have /proc mounted or even built in ;-) ]. of course CFS should not do _worse_ stats than what we had before, and should not break or massively misbehave. Also, anything sane we can do for low-resolution arches we should do (and we already do quite a bit - the while wmult stuff is to avoid expensive divisions) - and i regularly booted CFS with a low-resolution clock to make sure it works. So i'm not trying to duck anything, we've just got to keep our design priorities right :-) Ingo -
I meant that in many cases where the TSC is considered unreliable today it'll be possible to use it anyways at least for sched_clock() (and possibly even gtod()) The exception would be system which really have none, but there should be very few of those. -Andi -
I have to say, it would be interesting to try to use 32-bit arithmetic. I also think it's likely a mistake to do a nanosecond resolution. That's part of what forces us to 64 bits, and it's just not even an *interesting* resolution. It would be better, I suspect, to make the scheduler clock totally distinct from the other clock sources (many architectures have per-cpu cycle counters), and *not* try to even necessarily force it to be a "time-based" one. So I think it would be entirely appropriate to - do something that *approximates* microseconds. Using microseconds instead of nanoseconds would likely allow us to do 32-bit arithmetic in more areas, without any real overflow. And quite frankly, even on fast CPU's, the scheduler is almost certainly not going to be able to take any advantage of the nanosecond resolution. Just about anything takes a microsecond - including IO. I don't think nanoseconds are worth the ten extra bits they need, if we could do microseconds in 32 bits. And the "approximates" thing would be about the fact that we don't actually care about "absolute" microseconds as much as something that is in the "roughly a microsecond" area. So if we say "it doesn't have to be microseconds, but it should be within a factor of two of a ms", we could avoid all the expensive divisions (even if they turn into multiplications with reciprocals), and just let people *shift* the CPU counter instead. In fact, we could just say that we don't even care about CPU counters that shift frequency - so what? It gets a bit further off the "ideal microsecond", but the scheduler just cares about _relative_ times between tasks (and that the total latency is within some reasonable value), it doesn't really care about absolute time. Hmm? It would still be true that something that is purely based on timer ticks will always be liable to have rounding errors that will inevitably mean that you ...
I would add that I have been bothered by the 64-bit arithmetics when trying to see what could be improved in the code. In fact, it's very hard to optimize anything when you have arithmetics on integers larger than the CPU's, and gcc is known not to emit very good code in this situation (I remember it could not play with registers renaming, etc...). However, I undertand why Ingo chose to use 64 bits. It has the advantage that the numbers never wrap within 584 years. I'm well aware that it's very difficult to keep tasks ordered according to a key which can wrap. But if we consider that we don't need to be more precise than the return value from gettimeofday() that all applications use, we see that a bunch of microseconds is enough. 32 bits at the microsecond level wraps around every hour. We may accept to recompute all keys every hour. It's not that dramatic. The problem is how to detect that we will need to. I remember a trick used by Tim Schmielau in his jiffies64 patch for 2.4. He kept a copy of the highest bit of the lower word in the lowest bit of the higher word, and considered that the lower one could not wrap before we could check it. I liked this approach, which could be translated here in something like the following : Have all keys use 32-bit resolution, and monitor the 32nd bit. All tasks must have the same value in this bit, otherwise we consider that their keys have wrapped. The "current" value of this bit is copied somewhere. When we walk the tree and find a task with a key which does not have its 32nd bit equal to the current value, it means that this key has wrapped, so we have to use this information in our arithmetics. When all keys have their 32nd bit different from the "current" value, then we switch this value to reflect the new 32nd bit, and everything is in sync again. The only requirement is that no key wraps around before the "current" value is switched. This implies that no couple of tasks could have their keys distant by more than 31 bits (35 ...
This has changed in recent gccs. It doesn't force register pairs anymore. If you define an appropiate window and use some macros for the comparisons gettimeofday() has too strict requirements, that make it unnecessarily slow You don't need to recompute keys; just use careful comparisons using If you're worried about wrapping in one hour why is wrapping in two hours not a problem? I have one request though. If anybody adds anything complicated for this please make it optional so that 64bit platforms are not burdened by it. -Andi -
yeah. Note that i largely detached sched_clock() from the GTOD clocksources already in CFS, so part of this is already implemented and the intention is clear. For example, when the following happens: Marking TSC unstable due to: possible TSC halt in C2. Clocksource tsc unstable (delta = -71630388 ns) sched_clock() does _not_ stop using the TSC. It is very careful with the TSC value, it checks against wraps, jumping, etc. (the whole rq_clock() wrapper around sched_clock()), but still tries to use the highest resolution time source possible, even if that time source is not good enough for GTOD's purposes anymore. So the scheduler clock is already Note that there is a relatively easy way of reducing the effects of such intentional coupling: turn on CONFIG_HIGH_RES_TIMERS. That decouples the scheduler tick from the jiffy tick and works against such 'exploits' - _even_ if the scheduler clock is otherwise low resolution. Also enable CONFIG_NO_HZ and the whole thing (of when the scheduler tick kicks in) becomes very hard to predict. [ So while in a low-res clock situation scheduling will always be less precise, with hres-timers and dynticks we have a natural 'random sampler' mechanism so that no task can couple to the scheduler tick - accidentally or even intentionally. The only 'unavoidable coupling' scenario is when the hardware has only a single, low-resolution time sampling method. (that is pretty rare though, even in the ultra-embedded space. If a box has two independent hw clocks, even if they are low resolution, the timer tick can be yeah. We tried to do as much of that as possible, please read on below for (many) more details. There's no short summary i'm afraid :-/ Most importantly, CFS _already_ includes a number of measures that act against too frequent math. So even though you can see 64-bit math code in it, it's only rarely called if your clock has a low resolution - and that happens all automatically! (see below ...
Hi, You're comparing apples with oranges, I explicitely said: "At this point I'm not that much interested in a few localized optimizations, what I'm interested in is how can this optimized at the design level" IMO it's very important to keep computational and algorithmic complexity separately, I want to concentrate on the latter, so unless you can _prove_ that a similiar set of optimizations is impossible within my example, I'm going to ignore them for now. CFS has already gone through several versions of optimization and tuning, expecting the same from my design prototype is a little confusing... I want to analyze the foundation CFS is based on, in the review I mentioned a number of other issues and design related questions. If you need more time, that's fine, but I'd appreciate more background information related to that and not that you only jump on the more trivial Come on, Ingo, you can do better than that, I did mention in my review some of the requirements for the data types. I'm amazed how you can get to that judgement so quickly, could you please substantiate that a little more? I admit that the lack of source comments is an open invitation for further questions and Peter did exactly this and his comments were great - I'm hoping for more like that. You OTOH jump to conclusions based on a partial understanding what I'm actually trying to do. Ingo, how about you provide some of the mathematical prove CFS is based on? Can you prove that the rounding errors are irrelevant? Can you prove that all the limit checks can have no adverse effect? I tried that and I'm not entirely convinced of that, but maybe it's just me, so I'd love to see someone else's attempt at this. A major goal of my design is it to be able to define the limits within the scheduler is working correctly, so I know which information is relevant and what can be approximated. bye, Roman -
Hi Linus, On that theme, expressing the subsecond part of high precision time in decimal instead of left-aligned binary always was an insane idea. Applications end up with silly numbers of multiplies and divides (likely as not incorrect) whereas they would often just need a simple shift as you say, if the tv struct had been defined sanely from the start. As a bonus, whenever precision gets bumped up, the new bits appear on the right in formerly zero locations on the right, meaning little if any code needs to change. What we have in the incumbent libc timeofday scheme is the moral equivalent of BCD. Of course libc is unlikely ever to repent, but we can at least put off converting into the awkward decimal format until the last possible instant. In other words, I do not see why xtime is expressed as a tv instead of simple 32.32 fixed point. Perhaps somebody can elucidate me? Regards, Daniel -
Hi, The basic problem is that one needs a number of bits (at least 16) for normalization, which limits the time range one can work with. This means that 32 bit leaves only room for 1 millisecond resolution, the remainder could maybe saved and reused later. So AFAICT using micro- or nanosecond resolution doesn't make much computational difference. bye, Roman -
yeah, thanks for the reminder, this is on my todo list. As i suspect you
noticed it too, much of the task_struct size increase is not fundamental
and not related to 64-bit math at all - it's simply debug and
instrumentation overhead.
Look at the following table (i386, nodebug):
size
----
pre-CFS 1328
CFS 1472
CFS+patch 1376
the very small patch below gets rid of 96 bytes. And that's only the
beginning.
Ingo
-------------------------------------------------->
---
include/linux/sched.h | 21 +++++++++++++--------
1 file changed, 13 insertions(+), 8 deletions(-)
Index: linux/include/linux/sched.h
===================================================================
--- linux.orig/include/linux/sched.h
+++ linux/include/linux/sched.h
@@ -905,23 +905,28 @@ struct sched_entity {
struct rb_node run_node;
unsigned int on_rq;
+ u64 exec_start;
+ u64 sum_exec_runtime;
u64 wait_start_fair;
+ u64 sleep_start_fair;
+
+#ifdef CONFIG_SCHEDSTATS
u64 wait_start;
- u64 exec_start;
+ u64 wait_max;
+ s64 sum_wait_runtime;
+
u64 sleep_start;
- u64 sleep_start_fair;
- u64 block_start;
u64 sleep_max;
+ s64 sum_sleep_runtime;
+
+ u64 block_start;
u64 block_max;
u64 exec_max;
- u64 wait_max;
- u64 last_ran;
- u64 sum_exec_runtime;
- s64 sum_wait_runtime;
- s64 sum_sleep_runtime;
unsigned long wait_runtime_overruns;
unsigned long wait_runtime_underruns;
+#endif
+
#ifdef CONFIG_FAIR_GROUP_SCHED
struct sched_entity *parent;
/* rq on which this entity is (to be) queued: */
-
I brought that up a couple of weeks ago, got handwaved at and gave up. It still isn't obvious to me that all that arith needs to be 64-bit on 32-bit machines, or even on 64-bit. 4e9 is a big number. -
On Wed, 11 Jul 2007 23:16:38 +0200, Andi Kleen said: I'm seeing a bit of a disconnect here. If you spent all that time making it work, how come the guys who developed the patch are saying you didn't provide Odd, I looked at the patchset fairly closely a number of times, as I was hand-retrofitting the -rc[1-4] versions onto -rc[1-4]-mm kernels, and it looked to *me* like it was a nice set of 20 or so step-by-step changes (bisectable and everything - I got to do that once trying to figure out which one I botched). Was there something in there that I missed?
I think Andi's referring to the existing x86_64 code, which gets replaced by the patchset in question. -
<Takes a closer look at the patches> D'Oh! :) Yeah, the -rc4 version I'm looking at is like a dozen 1-3K patches setting up and cleaning up, and then one monster 65K patch doing the clockevents conversion, then another 6 or 8 small ones. Yeah, that one big patch really doesn't look separable to me. But as I said, I'm just a crash test dummy here. :) Andrew - how do you feel about keeping this in the -mm tree until Linus, Andi, Ingo, and Thomas get on the same page (which may be around the 2.6.24 merge window, by my guesstimate)?
Well, that's supposed to be Andi's tree and aggregated by Andrew into -mm. But keeping it in -mm isn't the hard part. It's getting enough testing to convince Linus it's safe, since there's no simple way to enable clockevents in a slow manner. IOW, keeping it in -mm just postpones the issue. -
I think it should be. That big patch really does do a *lot* more than just the "clockevents conversion". It does all the hpet clock setup changes etc that are about the hardware, and have *nothing* to do with actually changing the interfaces. For example, look at the hpet.c part of that patch. Totally independent cleanups of everything else. Or look at the changes to __setup_APIC_LVTT(). Same thing. All the actual hardware interface changes are *totally* independent of the software interface changes, and a lot of them are just cleanups. But those hardware interface changes are easily the things that can break, where some cleanup results in register writes being done in a different order or something, and so if there's a bug there (and it's not visible on most setups), now you cannot tell where the bug is. Another example: setup_APIC_timer() used to wait for a timer interrupt trigger to happen on the i8259 timer (or HPET). That code just got removed (or maybe it got moved so subtly that I just don't see it). What has that got to do with switching from the old timer interface to the new one? NOTHING. So those kinds of changes that change hardware access functions should have been done separately. Maybe there's a machine where that early synchronization was necessary for some subtle timing reason. If so, removing it sounds like a bug, no? Wouldn't it have been nice to see that removal as a separate patch that was independent of the interface switch- over? I'd be a *lot* happier with switching over interfaces if I thought that the low-level hardware drivers didn't change at the same time. But they *do* change, afaik. Linus -
The patch-set itself actually looks fine, as far as I'm concerned. But it does seem to have that "enable everything in one go" problem. I'd much rather see one time source at a time being converted, and enabled then and there, so that when people report problems and do a bisection, if it was HPET that broke, you get the commit that changed HPET. As it is, looking at that set, it *looks* like you'd get the "ok, now enable it all" as the commit that breaks, which tells you hardly anything, since the commit that _shows_ the behaviour has absolutely nothing to do with the code that actually causes it. But yeah, the patch series per se doesn't look bad. If it wasn't for me being burnt by the last big switch-over for timers, I probably wouldn't mind it at all, personally. Linus -
Ingo, I'm sorry to say so, but your answer just convinced me that you're wrong, and we MUST NOT take that code. That was *exactly* the same thing you talked about when I refused to take the original timer changes into 2.6.20. You were talking about how lots of people had worked really hard, and how it was really tested. And it damn well was NOT really tested, and 2.6.21 ended up being a horribly painful experience (one of the more painful kernel releases in recent times), and we ended up havign to fix a *lot* of stuff. And you admitted you were wrong at the time. Now you do the *exact* same thing. Here's a big clue: it doesn't matter one _whit_ how much face-slapping you get, or how much effort some programmers have put into the code. It's untested. And no, we are *not* going to do another "rip everything out, and replace it with new code" again. Over my dead body. We're going to do this thing gradually, or not at all. And if somebody feels slighted by the face-slap, and thinks he has already done enough, and isn't interested in doing it gradually, then good riddance. The "not at all" seems like a good idea, and maybe we can re-visit this in a year or two. I'm not going to have another 2.6.21 on my hands. Linus -
Linus, Can you please shed some light on me, how exactly you switch an architecture gradually to clock events. You simply can not convert PIT today and the HPET next week followed by the local APIC in three month. I have no problem to brew this for some more time. I got not repulsed by the 2.6.20 decision, but I have no clue how to communicate with a black hole. tglx -
For example, we can make sure that the code in question that actually touches the hardware stays exactly the same, and then just move the interfaces around - and basically guarantee that _zero_ hardware-specific issues pop up when you switch over, for example. That way there is a gradual change-over. The other approach (which would be nice _too_) is to actually try to convert one clock source at a time. Why is that not an option? Linus -
Linus, We need to give control to the clock events core code once we convert one clock event device. Having two competing subsystems controlling different devices (e.g. PIT and APIC) is not really desirable. The HPET change, which is the larger part of the conversion set simply because we now share the code with i386, might be split out by disabling HPET in the first step, doing the PIT / APIC conversion and then the HPET one in a separate step. Thanks, tglx -
The timer specific changes (i.e. the merges between arches) can be done more slowly, but the setup above is basically where I started, and it was already broken on one of my test boxes. Anyway, I'll help you however I can, because it's important to me to get this merged. thanks, -chris -
Can't you take the entire legacy clock system and wrap it as a single legacy clock source? Then you take bits out of the old system and put them as independent sources in the new system? When the legacy clock system is empty, you remove the legacy clock source. -- Mathematics is the supreme nostalgia of our time. -
But that misses the point. It means that the commit that actually *changes* the code never actually gets tested on its own Why not just fix up the HPET code so that it can be shared *first*. Without the other conversion? Really - What's so wrong with the hpet.c changes in the *absense* of conversion to clockevents? Those changes seem to be totally independent - just abstracting ou tthe "hpet_get_virt_address()" stuff etc. None of that has anything to do with clockevents, as far as I can see. In other words, you now change a i386-only file, and maybe it breaks subtly on i386 as a result. Wouldn't it be nicer to see that breakage as a separate event? Then, the x86-64 clockevents code will switch over entirely, but now it switches over to something we can say has gotten testing, and we know the switch-over won't break any 32-bit code, because the switch-over literally didn't change anything at all for that case. See? THAT is what I mean by "gradual". Bugs happen, but if we can make _independent_ bugs show up in _independent_ commits, that will make it much easier to figure out what happened. The same is true of a lot of the APIC timer code. Sure, that patch has the actual conversion in it, and you don't have the cross-architecture issues, but more than 50% of the patch seems to be just cleanup that is independent of the actual switch-over, no? Again, if it was done as a "one patch for cleanup, and another patch that actually switches the higher-level interfaces around", then the two mostly independent issues (of "hardware access/initialization" vs "higher-level changes in how it got called") get done as two independent commits. And no, I really probably wouldn't ask for this, but 2.6.21 showed *exactly* this problem. Trivial debugging helps like "git bisect" didn't help at all, because all the problems started when the new code was "activated", not when it was actually brought in. Linus -
Linus, Sure, I meant to do the HPET changes to i386 separate as a preparatory patch. Sharing HPET before the conversion is nasty at best (it involves a ton Well, we know that it works on i386, but once we turn on the x64 switch we have not tested the shared code for x64 yet. I try to find some practicable compromise between the big bang patch and I said before, that I'm going to split them further. tglx -
I don't think it's that much cleanup. One of my goals for x86-64 was always to have it support modern x86 only; this means in particularly most of the old bug workaround removed. With the APIC timer merging a lot of that crap gets back in. I would prefer to keep APIC code separate. -Andi -
Andi, Care to look at the patch ? It _IS_ seperate. Only HPET and PIT got shared. tglx -
i dont think "clean, modern x86 code" will ever happen - x86_64 has and is going to have the exact same type of crap. And i'll say a weird thing now: that is a _blessing_. Why? Because this crap in question originates from the _diversity_ of the platform, and that is a much larger asset than the cost of the quirks can ever be! What you suggest does not end up in "clean 64-bit code", it ends up in "a bit less crappy 64-bit code", plus a lot of unnecessary duplication of effort and duplication of code - which easily introduces more crap total than it gets rid of ... The x86 architecture isnt fully analogous to a random piece of device hardware that evolves. It is more of a collector of random pieces of hardware that evolve independently, and as such it will always be exposed to human messups in a factorized way. "The pristine, clean architecture" is an utopia and it will never come until humans design hardware. Under your scheme we'll end up with is two sets of code which share some of the workarounds and dont share some others. No, in fact we _already_ ended up with two sets of code that is crappy in different ways. We had countless cases of bugs fixed in i386 but not fixed in x86_64. (and vice versa) Sharing code for similar hardware is almost always good. I think the PowerPC experience (although it is not a fully equivalent case) about them merging their 32-bit and 64-bit architectures was an overwhelmingly positive move, and x86 could learn a thing or two from that. The only way to fight crappy hardware is to map it, to understand it and to design as cleanly in the presence of it as possible. Having two sets of code for the same thing hardly serves that purpose. In fact, having _more_ crappy hardware _forces_ us to do a cleaner design (up to a pain threshold). Ingo -
Yes, but it will be new crap, but no old crap anymore. If you always pile the new crap on the old crap at some point the whole thing might fall over. 64bit was intended as a fresh start. Admittedly we're getting more and more workarounds too and sometimes when I want to remove cruft i find out it is still needed on some 64bit boxes (e.g. see my repeated attempts to clean up the irq 0 The equivalent to the powerpc way would be essentially to report i386 into the x86-64 code base and leave the really old hardware only in arch/i386. I've considered doing it, but it would be an awful lot of work and to tempt distributions to actually use the new port would require going back quite a long time. And at least immediately it would end up with three cases to do things instead of two like currently. -Andi -
Well that's just silly. The right way will never create 3 ways, but always keep the limit to the existing 2 where the differences aren't worth reconciling, and 1 for anything that is common. It will be a fair amount of work, so any constructive input you have upfront would be helpful. thanks, -chris -
I think there's no such thing as a fresh start for a diverse architecture - the ia64 failure has proven that. x86_64 CPUs still do A20 emulation today (!). We still have people running industrial boards on real i386 DX CPUs, with the latest upstream kernel. 15 years ago an i386 DX was already quite obsolete. 32-bit is not going to go away in our lifetime, and we'll want to support it in a first-grade way. We better realize that prospect and have it right before our eyes in a single tree wherever it makes sense to share code - i'm certainly not talking about sharing mtrr/centaur.c or k8.c. (and i'm not necessarily suggesting to share io_apic.c either - although it's certainly borderline.) Ingo -
x86-64 doesn't care about a lot of x86 baggage and a lot of things have been even obsoleted in the platform. In practice the backwards compatibility on x86 isn't that great either. For example a significant number of new systems don't Yes, but those for example would be perfectly happy with an arch/i386 with all APIC and SMP code stripped out. Only the few people who still run dual P5s might not, but those could continue using old kernels. But eventually I think that would be the right clean way: arch/i386 stripped down port for truly old systems like the embedded 386 upto 586 or early 686. No SMP or APIC. arch/x86 supporting 32bit and 64bit for reasonably modern systems. NUMAQ/Voyager/P5-SMP/visual workstation gone [frankly the user base of those is too small to justify the code impact] It's just quite ugly to get there and when you think through it the actual advantages of such a setup it is likely not enough to justify the significant work to make it work. Also I wouldn't have any idea how to regression test significant changes to arch/i386 aimed at old systems. e.g. I don't think the powerpc people actually tried to still support really old systems where it is hard to do regression tests anymore, only really supported platforms. So while such a setup would be quite nice the practical problems of getting there are nasty. Also I must admit I prefer hacking on new code instead. -Andi -
That's not quite right. Leaving the code unchanged caused breakage already. The PIT is damn stupid and can be sensitive to how quickly it's programmed. So code that enable/disable didn't change, but frequency It was that way for x86_64, that's the first thing I fixed (since it was done by fully disabling all other timers but the one coverted ;-) -
Sure. We cannot avoid *all* problems. Bugs happen. But at least we could try to make sure that there aren't totally unnecessary changes in that switch-over patch. Which there definitely were, as far as I can tell. Linus -
one note is that the "talk differently to hardware" thing is in part already tested with the 32 bit tickless code; a lot of people (80% ?) are still using the 32 bit OS on their 64 bit machines, and the 32 bit code already talks in the "new way" to this hardware.... (and since Fedora 7 already ships tickless for 32 bit there are quite a lot of people using that in practice, in addition to the kernel.org kernel users) I would expect just about all the hardware interaction issues to have popped up already because of this "run 32 bit on 64 bit hardware" thing. -- if you want to mail me at work (you don't), use arjan (at) linux.intel.com Test the interaction between Linux and your BIOS via http://www.linuxfirmwarekit.org -
yes - i was (way too!) upset about it, and your reasoning for the
rejection was hard (on us) but fair: you wanted a quiet 2.6.20, and you
yes. We had 12 -hrt/dynticks merge related regressions between
2.6.21-rc1 and -final, and 4 after final. Here's a quick post-mortem:
12 fixes after -rc1:
[PATCH] i386: Fix bogus return value in hpet_next_event()
[PATCH] clockevents: remove bad designed sysfs support for now
[PATCH] clocksource: Fix thinko in watchdog selection
[PATCH] dynticks: fix hrtimer rounding error in next_timer_interrupt
[PATCH] i386: add command line option "local_apic_timer_c2_ok"
[PATCH] i386: disable local apic timer via command line or dmi quirk
[PATCH] i386: clockevents fix breakage on Geode/Cyrix PIT
[PATCH] i386: trust the PM-Timer calibration of the local APIC timer
[PATCH] clockevents: Fix suspend/resume to disk hangs
[PATCH] highres: do not run the TIMER_SOFTIRQ after switching to highres mode
[PATCH] hrtimer: prevent overrun DoS in hrtimer_forward()
[PATCH] Save/restore periodic tick information over suspend/resume implementations
4 fixes after -final:
2.6.21.1: -
2.6.21.2:
[PATCH] clocksource: fix resume logic
2.6.21.3: -
2.6.21.4: -
2.6.21.5:
[PATCH] NOHZ: Rate limit the local softirq pending warning output
[PATCH] Ignore bogus ACPI info for offline CPUs
[PATCH] i386: HPET, check if the counter works
2.6.21.6: -
it's all pretty quiet today on the dynticks regressions front. (there
are no open regressions in either the upstream i386 code or in the devel
patches we are aware of. Forced-HPET in -mm, which is not part of this
queue in question [but which is done for dynticks], has one open
regression.)
The majority of the above bugs were in the infrastructure code. (the
worst was the generic resume/suspend one fixed in 2.6.21.2) And sadly, a
fair number of the infrastructure bugs we introduced during the frentic
clockevents/dynticks rewrites/redesigns we ...One thing I'll happily talk about is that while 2.6.21 was painful, you and Thomas in particular were both very responsible about the thing, so no, I'm not at all complaining or worried about it in that sense! I just really _really_ wish we could have two fairly stable releases in a row. I think 2.6.22 has the potential to be a pretty good setup, and I'd really like to avoid having another 2.6.21 immediately afterwards. So I'm not worried about integration and getting fixes when things break per se, but I *am* worried that this is an area where we've traditionally had lots of unexpected problems. And hey, maybe this time there will be none. I just still smart from the last time, so I'd prefer it to go more smoothly this time around. Linus -
The first and third of these are just simple Kconfig updates, and the
middle one just updates the list of symbols which shouldn't be warned
about in CONFIG_RELOCATABLE's absolute symbol check. They're completely
This appears to fix a real bug; the only question is whether x86-64
needs the same treatment. I'm not sure if the original bug reporter
(dave young) has confirmed it fixed his problem.
J
-
^^^ That patch was supposed to be merged for 2.6.22 (you told me you forgot to merge it) and has been for a long time in mm. Does it now need to be rereviewed for 2.6.23? The other pieces of the quicklist patch for core and other arches were merged for 2.6.22. -
It's just on the normal re-review list. But it'll likely go in. -Andi -
Here's some fix: Signed-off-by: Jan Engelhardt <jengelh@gmx.de> --- arch/x86_64/Kconfig | 20 ++++++++++---------- 1 file changed, 10 insertions(+), 10 deletions(-) Index: linux-2.6.22-rc6/arch/x86_64/Kconfig =================================================================== --- linux-2.6.22-rc6.orig/arch/x86_64/Kconfig +++ linux-2.6.22-rc6/arch/x86_64/Kconfig @@ -753,11 +753,11 @@ config DMAR depends on PCI_MSI && ACPI && EXPERIMENTAL default y help - DMA remapping(DMAR) devices support enables independent address - translations for Direct Memory Access(DMA) from Devices. + DMA remapping (DMAR) devices support enables independent address + translations for Direct Memory Access (DMA) from devices. These DMA remapping devices are reported via ACPI tables - and includes pci device scope covered by these DMA - remapping device. + and include PCI device scope covered by these DMA + remapping devices. config DMAR_GFX_WA bool "Support for Graphics workaround" @@ -765,9 +765,9 @@ config DMAR_GFX_WA default y help Current Graphics drivers tend to use physical address - for DMA and avoid using DMA api's. Setting this config + for DMA and avoid using DMA APIs. Setting this config option permits the IOMMU driver to set a unity map for - all the OS visible memory. Hence the driver can continue + all the OS-visible memory. Hence the driver can continue to use physical addresses for DMA. config DMAR_FLOPPY_WA @@ -775,10 +775,10 @@ config DMAR_FLOPPY_WA depends on DMAR default y help - Floppy disk drivers are know to by pass dma api calls - their by failing to work when IOMMU is enabled. This - work around will setup a 1 to 1 mappings for the first - 16M to make floppy(isa device) work. + Floppy disk drivers are know to bypass DMA API calls + thereby failing to work when IOMMU is enabled. This + workaround will setup a 1:1 mapping for the first + 16M to make floppy (an ISA device) work. ...
) That posting was just a change proposal for the drivers/isdn/Kconfig part, not a complete replacement for the entire patch. If you'd care to reissue that patch with the modification I proposed, I'll gladly ack it. Alternatively I can also send a full replacement patch if you prefer. Regards, Tilman --=20 Tilman Schmidt E-Mail: tilman@imap.cc Bonn, Germany Diese Nachricht besteht zu 100% aus wiederverwerteten Bits. Unge=F6ffnet mindestens haltbar bis: (siehe R=FCckseite)
Since I did not really see much of a difference between our two approaches, I'd be grateful if you could send a full replacement in the hopes that I see the global picture. Thanks, Jan -- -
As far as I can tell the antifrag patches are stable and are significantly enhancing various aspects of the VM and also make it more reliable. SLUB can use it to increase scalability. MM has been using order 3 allocs via SLUB for months now without a problem. Without the antifrag patches order 1 allocs could cause OOMs. It opens the door for functionality that we wanted for a long time such a memory unplug etc. -
SLUB using high orders without page allocation failures do depend on two very questionable patches that I brought to attention in my inital merge mail. If grouping pages by mobility goes through, I'll be revisiting that properly to make sure it can work without deadlocking ever under any circumstances. Right now, it theoritically could livelock although I've never been able to reproduce it. The patches as they are will work for high-order allocations if you are willing to wait and reclaim memory. The more stressful users need more effort but it's already been shown that it can be made work with one And I want to avoid a catch-22 here where the features that depend on grouping pages by mobility have to exist before grouping pages by mobility is pushed through. I would like the patches to go through on the grounds that higher order allocations can succeed. However, I am also happy to say that order-0 pages should be used as much as possible, that case should always be made as fast as possible and the world must not end if a high-order allocation fails. -- -- Mel Gorman Part-time Phd Student Linux Technology Center University of Limerick IBM Dublin Software Lab -
SLUB can easily be made to not use higher order pages. If the SLUB mobility patches are not merged then higher order page use can be explicitly enabled via passing the following to the kernel on boot slub_max_order=<desired max order> If they are merged then the higher order page use can be disabled in case of