>>>> Not single thread -- but a "global OF lock" yes.  Not that
quoted text
>>>> it matters too much, (almost) all property accesses are init
>>>> time anyway (which is effectively single threaded).
>>>
>>> Not that true anymore. A lot of driver probe is being threaded
>>> nowadays,
>>> either bcs of the new multithread probing bits, or because they get
>>> loaded by userland from some initramfs etc..
>>
>> The kernel doesn't care if one CPU is in OF land while the others
>> are doing other stuff -- well you have to make sure the OF won't
>> try to use a hardware device at the same time as the kernel, true.
>
> True, but at the very least you have to prevent multiple cpus
> from enterring OFW.  In fact this is very important.

Yes.  "Global OF lock".

quoted text
> OFW is not multi-threaded

You are not _guaranteed_ it is multithreaded, and you don't
know it's threading model (or how to do thread synchronisation).

quoted text
> therefore you can't let multiple CPUs call
> into OFW at one time.  You must use some kind of locking mechanism,
> and that locking mechanism is not simple because it has to not just
> stop the other cpus, it has to be able to stop the other cpus yet
> still allow them to receive SMP cross-calls from the firmware if the
> OFW call is 'stop' or similar.

YOu don't need to *stop* the other CPUs, you just need to
prevent them from entering the client interface.  Put a lock
in front.

quoted text
>> I'm a bit concerned about the 100kB or so of data duplication
>> (on a *quite big* device tree), and the extra code you need
>> (all changes have to be done to both tree copies).  Maybe
>> I shouldn't be worried; still, it's obviously not a great
>> idea to *require* any arch to get and keep a full copy of
>> the tree -- it's wasteful and unnecessary.
>
> The largest amount of memory I've ever seen consumed on sparc64
> was 76K and this is 1) 64-bit and 2) an ENORMOUS machine with
> lots of cpus and devices.  And I know because sparc64 prints
> a kernel message at boot which states how much memory was
> consumed by the in-kernel device tree copy.

The in-OF tree uses a bit more memory, depending on implementation.
It's hard to tell though, it contains so much more than the
properties-only tree, perhaps you're right.

quoted text
> Please let's get over this memory consumption non-issue and move
> on to more productive talk.

Okay -- so answer the second part of my concern please: if you keep
a copy, you need to keep both in sync -- that means every change
by the kernel has to be done twice, and you won't ever be told about
changes by the OF, so you have to get a full fresh copy every single
time you return from an OF client call that could have changed a
property.


Segher

-
unsubscribe notice
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/