[lkml]   [2008]   [Nov]   [27]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [RFC v1][PATCH]page_fault retry with NOPAGE_RETRY
Nick Piggin wrote:
> On Wed, Nov 26, 2008 at 11:57:24AM -0800, Mike Waychison wrote:
>> Nick Piggin wrote:
>>> On Tue, Nov 25, 2008 at 10:42:47AM -0800, Ying Han wrote:
>>>>>> The patch flags current->flags to PF_FAULT_MAYRETRY as identify that
>>>>>> the caller can tolerate the retry in the filemap_fault call patch.
>>>>>> Benchmark is done by mmap in huge file and spaw 64 thread each
>>>>>> faulting in pages in reverse order, the the result shows 8%
>>>>>> porformance hit with the patch.
>>>>> I suspect we also want to see the cases where this change helps?
>>>> i am working on more benchmark to show performance improvement.
>>> Can't you share the actual improvement you see inside Google?
>>> Google must be doing something funky with threads, because both
>>> this patch and their new malloc allocator apparently were due to
>>> mmap_sem contention problems, right?
>> One of the big improvements we see with this patch is the ability to
>> read out files in /proc/pid much faster. Consider the following events:
>> - an application has a high count of threads sleeping with
>> read_lock(mmap_sem) held in the fault path (on the order of hundreds).
>> - one of the threads in the application then blocks in
>> write_lock(mmap_sem) in the mmap()/munmap() paths
>> - now our monitoring software tries to read some of the /proc/pid files
>> and blocks behind the waiting writer due to the fairness of the rwsems.
>> This basically has to wait for all faults ahead of the reader to
>> terminate (and let go of the reader lock) and then the writer to have a
>> go at mmap_sem. This can take an extremely long time.
>> This patch helps a lot in this case as it keeps the writer from waiting
>> behind all the waiting readers, so it executes much faster.
> Hmm. How quantifiable is the benefit? Does it actually matter that you
> can read the proc file much faster? (this is for some automated workload
> management daemon or something, right?)

Correct. I don't recall the numbers from the pathelogical cases we were
seeing, but iirc, it was on the order of 10s of seconds, likely
exascerbated by slower than usual disks. I've been digging through my
inbox to find numbers without much success -- we've been using a variant
of this patch since 2.6.11.

Török however identified mmap taking on the order of several
milliseconds due to this exact problem:

> Would it be possible to reduce mmap()/munmap() activity? eg. if it is
> due to a heap memory allocator, then perhaps do more batching or set
> some hysteresis.

I know our tcmalloc team had made great strides to reduce mmap_sem
contention for the heap, but there are various other bits of the stack
that really want to mmap files..

We generally try to avoid such things, but sometimes it a) can't be
easily avoided (third party libraries for instance) and b) when it hits
us, it affects the overall health of the machine/cluster (the monitoring
daemons get blocked, which isn't very healthy).

>>> That was before the kernel and glibc got together to fix the stupid
>>> mmap_sem problem in malloc (shown up in that FreeBSD MySQL thread);
>>> and before private futexes. I would be interested to know if Google
>>> still has problems that require this patch...
>> I'm not very familiar with the 'malloc' problem in glibc. Was this just
>> overhead in heap growth/shrinkage causing problems?
> As far as I understand, glibc does actually seperate notions of allocated
> virtual memory and allocated pages. By that I mean that it is very careful to
> return pages back to the system when they are unused, but it does try to
> batch up changes to the virtual address space. Unfortunately, it used to
> return pages by doing a mmap call with PROT_NONE, then to start using
> that virtual memory again, it would mmap with PROT_READ|PROT_WRITE.
> This meant that a malloc/touch/free would look like this:
> mmap <- down_write(mmap_sem)
> page fault <- down_read(mmap_sem)
> mmap <- down_write(mmap_sem)
> What we did was to make it use madvise(MADV_DONTNEED) to throw the pages
> away. Then the kernel was changed to implement MADV_DONTNEED using only
> down_read. Then the same sequence is just this:
> page fault <- down_read(mmap_sem)
> madvise <- down_read(mmap_sem)
> (because changes to virtual memory allocation are batched).
> I thought google's malloc was primarily written to fix this bad behaviour,
> because with the new behaviour, glibc's malloc seems to beat google's
> malloc on the performance and scalability tests I was running at the time,
> as well as being more memory footprint friendly.
> OTOH, it would be possible that hysteresis watermarks in glibc are not big
> enough for a given application, which could introduce virtual address
> activity back into the workload. These can be tuned, however.

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to
More majordomo info at
Please read the FAQ at

 \ /
  Last update: 2008-11-27 10:33    [W:0.124 / U:0.992 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site