sometimes a rcu callback is just calling kfree() to free a struct's memory (we say this callback is a trivial callback.). this patch introduce kfree_rcu() to do these things directly, easily. There are 4 reasons that we need kfree_rcu(): 1) unloadable modules: a module(rcu callback is defined in this module) using rcu must call rcu_barrier() when unload. rcu_barrier() will increase the system's overhead(the more cpus the worse) and rcu_barrier() is very time-consuming. if all rcu callback defined in this module are trivial callback, we can just call kfree_rcu() instead, save a rcu_barrier() when unload. 2) duplicate code: all trivial callback are duplicate code though the structs to be freed are different. it's just a container_of() and a kfree(). There are about 50% callbacks are trivial callbacks for call_rcu() in current kernel code. 3) cache: the instructions of trivial callback is not in the cache supposedly. calling a trivial callback will let to cache missing very likely. the more trivial callback the more cache missing. OK, this is not a problem now or in a few days: Only less than 1% trivial callback are called in running kernel. 4) future: the number of user of rcu is increasing. new code for rcu is trivial callback very likely. it means more modules using rcu and more duplicate code(may come to 90% of callbacks is trivial callbacks) and more cache missing. Implementation: there were a lot of ideas came out when i implemented kfree_rcu(). I chose the simplest one as this patch shows. but these implementation may cannot be used for to free a struct larger than 16KBytes. kfree_rcu_bh()? kfree_rcu_sched()? these two are not need current. call_rcu_bh() & call_rcu_sched() are hardly be called(and hardly be called for trivial callback). vfree_rcu()? No, vfree() is not atomic function, will not be called in softirq. Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> --- diff ...
All the above could do with some comments explaining what it does. Please use checkpatch. Surely it cannot be resirable to have a "function" which can call kfree either synchronously or asynchronously depending upon the size of the object which was passed to it? If the caller wants rcu treatment of the freeing then that is what the caller must be given. I mean, if the caller can tolerate a synchrnous call to kfree() then the caller should have directly called kfree? I think (and pray) that the above could have been implemented as an How can this work? We take the difference between two pointers, divide that by 4 or 8, then treat the resulting number as the address of an RCU callback function. --
__rcu_reclaim either treats head->func as an offset for kfree or as a
Here it's used:
the softirq that is called after the grace period calls kfree directly
I'd try to make that a compile time error. Is that possible? perhaps
with some __builtin_constant_p (head-ptr) or something like that. Or
__rcu_reclaim() knows that function pointers < 4096 are actually offsets
for kfree.
I like the idea:
- the call to list->func() is probably very difficult to predict for a
branch target predictor.
- it's just a waste not to call kfree directly.
- I'm not sure about the implementation.
--
Manfred
--
__rcu_reclaim(head) is called when @head 's grace period had completed. __rcu_reclaim(head) uses "if (__offset <= __KFREE_RCU_MAX_OFFSET)" to check whether @head is queued by kfree_rcu() or normal call_rcu(). if @head is queued by kfree_rcu(), ((void *)head - sizeof(void *) * __offset) is the pointer that the memory chunk need to be freed. otherwise call head->func(head) as original code. __rcu_reclaim has __prefix, will not be used by user, it is a helper for rcu-machine. it is always called asynchronously. sorry for not comments We will use __rcu_reclaim(head) sort it out when it's grace period has --
You lost me on this one. Suppose that the following sequence of events occurred: a. The module invokes call_rcu() or kfree_rcu(). The callback is queued on CPU 0. b. Perhaps a grace period completes, and the callback is therefore moved to CPU 0's donelist. But CPU 0 is busy, so doesn't get around to invoking the callback. (For example, ksoftirqd.) c. The module is unloaded, and uses kfree_rcu() instead of rcu_barrier(). The callback is queued on CPU 1. d. A grace period completes, and CPU 1 is relatively idle, so invokes its callback quickly. The module is therefore unloaded. e. CPU 0 finally gets around to executing its callback, but the module has been unloaded, so there is nothingness where the callback function used to be. We get an oops. Indeed! There was something similar to kfree_rcu() proposed some years back, but it was rejected because it contained more code than did the trivial callbacks. :-/ But there are more such callbacks these days, so might be worth Reducing code footprint would be a good thing. Do you have stats on OK, so the idea is that structures whose rcu_head is near the front of the structure have the offset of the rcu_head put into the ->func field instead of a pointer to the callback function? Of course, it doesn't need to be too near the beginning of the function... All arches are guaranteed not to have kernel text in the low 16K OK, so we pass in the pointer to the rcu_head structure, followed by the offset in pointer-sized units, but with the latter cast to a pointer to a callback function? Hmmm.... Kinky.... Then after the grace period completes, the __rcu_reclaim() sorts --
How to usage kfree_rcu:
struct my_struct {
int data;
struct rcu_head rcu;
};
----------------original code:--------------------------
void my_struct_release_rcu(struct rcu_head *rcu)
{
struct my_struct *p;
item = container_of(rcu, struct my_struct, rcu);
kfree(p);
}
void some_fuction()
{
struct my_struct *p;
.....;
call_rcu(&p->rcu, my_struct_release_rcu);
.....;
}
---end---
-----------------after use kfree_rcu:--------------------
/* my_struct_release_rcu() was removed */
void some_fuction()
{
struct my_struct *p;
.....;
kfree_rcu(p, &p->rcu);
.....;
}
---end---
1) unloadable modules:
A) use my_struct_release_rcu():
when we unload this modules, we need call rcu_barrier() to wait
all my_struct_release_rcu() had called.
B) use kfree_rcu():
if all trivial callback are removed and kfree_rcu() are used instead,
we do not need to wait anything. just quick finish unloading.
2) duplicate code:
A) use my_struct_release_rcu():
All trivial callback are very like my_struct_release_rcu(),
all are duplicate code.
B) use kfree_rcu():
all trivial callback are removed, not duplicate code like
my_struct_release_rcu().
3) cache:
A) use my_struct_release_rcu():
my_struct_release_rcu() is called rarely, when my_struct_release_rcu()
is being called, cache missing will occur.
B) use kfree_rcu():
my_struct_release_rcu() is removed, not such cache missing.
4) future:
A) use my_struct_release_rcu():
when new user use rcu, the most callback is trivial callback
like my_struct_release_rcu(). this is the common of using rcu.
so the problems of above are more and more heavy.
B) use kfree_rcu():
fix these problems for ever.
--
On Thu, Sep 18, 2008 at 04:59:33PM +0800, Lai Jiangshan wrote: OK, so the trick is that the module -never- uses call_rcu() directly, instead using -only- kfree_rcu(), along perhaps also with synchronize_rcu(). Because kfree_rcu() does not reference module text, you then don't need to wait at all. Good point! Thanx, Paul --
Hmm: why is rcu_barrier() sufficient to prevent races?
Offlining a cpu reorders rcu callbacks - rcu_barrier() can return before
all previous call_rcu() callbacks were called.
--
Manfred
--
The rcu_barrier() family of functions registers a callback on each CPU, and waits until all these callbacks have been invoked. The CPU offlining process preserves the order of the callbacks that were registered on a given CPU. Thus, when rcu_barrier() returns, all RCU callbacks previously registered are guaranteed to have already been invoked, regardless of what CPUs might have been offlined and onlined in the meantime. Thanx, Paul --
You are right: I mixed up rcu_barrier() and synchronize_rcu().
--
Manfred
--
We save a rcu_barrier() only when all rcu callback defined in this module are trivial callback and we use kfree_rcu to instead them. trivial callbacks are the most common callbacks, so some module may used I did not have stats on the kernel text size, I think these cache missing are caused by lots of different trivial callbacks in everywhere, (unsigned long)head->func is always <= 4095, not 14K or 32K. we just guaranteed not to have kernel text in the low 4k of memory. the real offset is (sizeof(void *) * (unsigned long)head->func), Yes, kernel pointers have redundant information, we use the low 4k as offset. when ->func < 4k, it stand for offset, when ->func >= 4k, --
The Tiny Linux guys might be interested in even a small reduction in Good point! --
