summaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorLee Schermerhorn <Lee.Schermerhorn@hp.com>2009-09-21 17:03:40 -0700
committerLinus Torvalds <torvalds@linux-foundation.org>2009-09-22 07:17:41 -0700
commit252c5f94d944487e9f50ece7942b0fbf659c5c31 (patch)
tree0dd21bdb968ce5bd3f30fadce261984d3b1abf98
parent3f96b79ad96263cc0ece7bb340cddf9b2ddfb1b3 (diff)
downloadlinux-stable-252c5f94d944487e9f50ece7942b0fbf659c5c31.tar.gz
linux-stable-252c5f94d944487e9f50ece7942b0fbf659c5c31.tar.bz2
linux-stable-252c5f94d944487e9f50ece7942b0fbf659c5c31.zip
mmap: avoid unnecessary anon_vma lock acquisition in vma_adjust()
We noticed very erratic behavior [throughput] with the AIM7 shared workload running on recent distro [SLES11] and mainline kernels on an 8-socket, 32-core, 256GB x86_64 platform. On the SLES11 kernel [2.6.27.19+] with Barcelona processors, as we increased the load [10s of thousands of tasks], the throughput would vary between two "plateaus"--one at ~65K jobs per minute and one at ~130K jpm. The simple patch below causes the results to smooth out at the ~130k plateau. But wait, there's more: We do not see this behavior on smaller platforms--e.g., 4 socket/8 core. This could be the result of the larger number of cpus on the larger platform--a scalability issue--or it could be the result of the larger number of interconnect "hops" between some nodes in this platform and how the tasks for a given load end up distributed over the nodes' cpus and memories--a stochastic NUMA effect. The variability in the results are less pronounced [on the same platform] with Shanghai processors and with mainline kernels. With 31-rc6 on Shanghai processors and 288 file systems on 288 fibre attached storage volumes, the curves [jpm vs load] are both quite flat with the patched kernel consistently producing ~3.9% better throughput [~80K jpm vs ~77K jpm] than the unpatched kernel. Profiling indicated that the "slow" runs were incurring high[er] contention on an anon_vma lock in vma_adjust(), apparently called from the sbrk() system call. The patch: A comment in mm/mmap.c:vma_adjust() suggests that we don't really need the anon_vma lock when we're only adjusting the end of a vma, as is the case for brk(). The comment questions whether it's worth while to optimize for this case. Apparently, on the newer, larger x86_64 platforms, with interesting NUMA topologies, it is worth while--especially considering that the patch [if correct!] is quite simple. We can detect this condition--no overlap with next vma--by noting a NULL "importer". The anon_vma pointer will also be NULL in this case, so simply avoid loading vma->anon_vma to avoid the lock. However, we DO need to take the anon_vma lock when we're inserting a vma ['insert' non-NULL] even when we have no overlap [NULL "importer"], so we need to check for 'insert', as well. And Hugh points out that we should also take it when adjusting vm_start (so that rmap.c can rely upon vma_address() while it holds the anon_vma lock). akpm: Zhang Yanmin reprts a 150% throughput improvement with aim7, so it might be -stable material even though thiss isn't a regression: "this issue is not clear on dual socket Nehalem machine (2*4*2 cpu), but is severe on large machine (4*8*2 cpu)" [hugh.dickins@tiscali.co.uk: test vma start too] Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: Nick Piggin <npiggin@suse.de> Cc: Eric Whitney <eric.whitney@hp.com> Tested-by: "Zhang, Yanmin" <yanmin_zhang@linux.intel.com> Cc: <stable@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
-rw-r--r--mm/mmap.c4
1 files changed, 2 insertions, 2 deletions
diff --git a/mm/mmap.c b/mm/mmap.c
index 56eb871c32d0..b6d74b3dbbc2 100644
--- a/mm/mmap.c
+++ b/mm/mmap.c
@@ -570,9 +570,9 @@ again: remove_next = 1 + (end > next->vm_end);
/*
* When changing only vma->vm_end, we don't really need
- * anon_vma lock: but is that case worth optimizing out?
+ * anon_vma lock.
*/
- if (vma->anon_vma)
+ if (vma->anon_vma && (insert || importer || start != vma->vm_start))
anon_vma = vma->anon_vma;
if (anon_vma) {
spin_lock(&anon_vma->lock);