mm, memcg: Try charging a page before setting page up to date
authorMel Gorman <mgorman@suse.de>
Wed, 24 Jun 2015 23:57:27 +0000 (16:57 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Thu, 25 Jun 2015 00:49:43 +0000 (17:49 -0700)
commiteb3c24f305e56caaf5c4bd34d2923839688d470e
tree37766a627064a5b6cce7c0160204b71a5a2e635b
parent4165b9b46181290d7e6ac276080c89b65623c633
mm, memcg: Try charging a page before setting page up to date

Historically memcg overhead was high even if memcg was unused.  This has
improved a lot but it still showed up in a profile summary as being a
problem.

/usr/src/linux-4.0-vanilla/mm/memcontrol.c                           6.6441   395842
  mem_cgroup_try_charge                                                        2.950%   175781
  __mem_cgroup_count_vm_event                                                  1.431%    85239
  mem_cgroup_page_lruvec                                                       0.456%    27156
  mem_cgroup_commit_charge                                                     0.392%    23342
  uncharge_list                                                                0.323%    19256
  mem_cgroup_update_lru_size                                                   0.278%    16538
  memcg_check_events                                                           0.216%    12858
  mem_cgroup_charge_statistics.isra.22                                         0.188%    11172
  try_charge                                                                   0.150%     8928
  commit_charge                                                                0.141%     8388
  get_mem_cgroup_from_mm                                                       0.121%     7184

That is showing that 6.64% of system CPU cycles were in memcontrol.c and
dominated by mem_cgroup_try_charge.  The annotation shows that the bulk
of the cost was checking PageSwapCache which is expected to be cache hot
but is very expensive.  The problem appears to be that __SetPageUptodate
is called just before the check which is a write barrier.  It is
required to make sure struct page and page data is written before the
PTE is updated and the data visible to userspace.  memcg charging does
not require or need the barrier but gets unfairly hit with the cost so
this patch attempts the charging before the barrier.  Aside from the
accidental cost to memcg there is the added benefit that the barrier is
avoided if the page cannot be charged.  When applied the relevant
profile summary is as follows.

/usr/src/linux-4.0-chargefirst-v2r1/mm/memcontrol.c                  3.7907   223277
  __mem_cgroup_count_vm_event                                                  1.143%    67312
  mem_cgroup_page_lruvec                                                       0.465%    27403
  mem_cgroup_commit_charge                                                     0.381%    22452
  uncharge_list                                                                0.332%    19543
  mem_cgroup_update_lru_size                                                   0.284%    16704
  get_mem_cgroup_from_mm                                                       0.271%    15952
  mem_cgroup_try_charge                                                        0.237%    13982
  memcg_check_events                                                           0.222%    13058
  mem_cgroup_charge_statistics.isra.22                                         0.185%    10920
  commit_charge                                                                0.140%     8235
  try_charge                                                                   0.131%     7716

That brings the overhead down to 3.79% and leaves the memcg fault
accounting to the root cgroup but it's an improvement.  The difference
in headline performance of the page fault microbench is marginal as
memcg is such a small component of it.

pft faults
                                       4.0.0                  4.0.0
                                     vanilla            chargefirst
Hmean    faults/cpu-1 1443258.1051 (  0.00%) 1509075.7561 (  4.56%)
Hmean    faults/cpu-3 1340385.9270 (  0.00%) 1339160.7113 ( -0.09%)
Hmean    faults/cpu-5  875599.0222 (  0.00%)  874174.1255 ( -0.16%)
Hmean    faults/cpu-7  601146.6726 (  0.00%)  601370.9977 (  0.04%)
Hmean    faults/cpu-8  510728.2754 (  0.00%)  510598.8214 ( -0.03%)
Hmean    faults/sec-1 1432084.7845 (  0.00%) 1497935.5274 (  4.60%)
Hmean    faults/sec-3 3943818.1437 (  0.00%) 3941920.1520 ( -0.05%)
Hmean    faults/sec-5 3877573.5867 (  0.00%) 3869385.7553 ( -0.21%)
Hmean    faults/sec-7 3991832.0418 (  0.00%) 3992181.4189 (  0.01%)
Hmean    faults/sec-8 3987189.8167 (  0.00%) 3986452.2204 ( -0.02%)

It's only visible at single threaded. The overhead is there for higher
threads but other factors dominate.

Signed-off-by: Mel Gorman <mgorman@suse.de>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/memory.c