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>

diff --git a/mm/mmap.c b/mm/mmap.c
index 56eb871c32d0884b1c4996d79d8691f5698e6d29..b6d74b3dbbc25eb4ab73edc18ff46b5ba1fa5b96 100644 (file)
--- 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);