mm, vmstat: add infrastructure for per-node vmstats
authorMel Gorman <mgorman@techsingularity.net>
Thu, 28 Jul 2016 22:45:24 +0000 (15:45 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Thu, 28 Jul 2016 23:07:41 +0000 (16:07 -0700)
Patchset: "Move LRU page reclaim from zones to nodes v9"

This series moves LRUs from the zones to the node.  While this is a
current rebase, the test results were based on mmotm as of June 23rd.
Conceptually, this series is simple but there are a lot of details.
Some of the broad motivations for this are;

1. The residency of a page partially depends on what zone the page was
   allocated from.  This is partially combatted by the fair zone allocation
   policy but that is a partial solution that introduces overhead in the
   page allocator paths.

2. Currently, reclaim on node 0 behaves slightly different to node 1. For
   example, direct reclaim scans in zonelist order and reclaims even if
   the zone is over the high watermark regardless of the age of pages
   in that LRU. Kswapd on the other hand starts reclaim on the highest
   unbalanced zone. A difference in distribution of file/anon pages due
   to when they were allocated results can result in a difference in
   again. While the fair zone allocation policy mitigates some of the
   problems here, the page reclaim results on a multi-zone node will
   always be different to a single-zone node.
   it was scheduled on as a result.

3. kswapd and the page allocator scan zones in the opposite order to
   avoid interfering with each other but it's sensitive to timing.  This
   mitigates the page allocator using pages that were allocated very recently
   in the ideal case but it's sensitive to timing. When kswapd is allocating
   from lower zones then it's great but during the rebalancing of the highest
   zone, the page allocator and kswapd interfere with each other. It's worse
   if the highest zone is small and difficult to balance.

4. slab shrinkers are node-based which makes it harder to identify the exact
   relationship between slab reclaim and LRU reclaim.

The reason we have zone-based reclaim is that we used to have
large highmem zones in common configurations and it was necessary
to quickly find ZONE_NORMAL pages for reclaim. Today, this is much
less of a concern as machines with lots of memory will (or should) use
64-bit kernels. Combinations of 32-bit hardware and 64-bit hardware are
rare. Machines that do use highmem should have relatively low highmem:lowmem
ratios than we worried about in the past.

Conceptually, moving to node LRUs should be easier to understand. The
page allocator plays fewer tricks to game reclaim and reclaim behaves
similarly on all nodes.

The series has been tested on a 16 core UMA machine and a 2-socket 48
core NUMA machine. The UMA results are presented in most cases as the NUMA
machine behaved similarly.

pagealloc
---------

This is a microbenchmark that shows the benefit of removing the fair zone
allocation policy. It was tested uip to order-4 but only orders 0 and 1 are
shown as the other orders were comparable.

                                           4.7.0-rc4                  4.7.0-rc4
                                      mmotm-20160623                 nodelru-v9
Min      total-odr0-1               490.00 (  0.00%)           457.00 (  6.73%)
Min      total-odr0-2               347.00 (  0.00%)           329.00 (  5.19%)
Min      total-odr0-4               288.00 (  0.00%)           273.00 (  5.21%)
Min      total-odr0-8               251.00 (  0.00%)           239.00 (  4.78%)
Min      total-odr0-16              234.00 (  0.00%)           222.00 (  5.13%)
Min      total-odr0-32              223.00 (  0.00%)           211.00 (  5.38%)
Min      total-odr0-64              217.00 (  0.00%)           208.00 (  4.15%)
Min      total-odr0-128             214.00 (  0.00%)           204.00 (  4.67%)
Min      total-odr0-256             250.00 (  0.00%)           230.00 (  8.00%)
Min      total-odr0-512             271.00 (  0.00%)           269.00 (  0.74%)
Min      total-odr0-1024            291.00 (  0.00%)           282.00 (  3.09%)
Min      total-odr0-2048            303.00 (  0.00%)           296.00 (  2.31%)
Min      total-odr0-4096            311.00 (  0.00%)           309.00 (  0.64%)
Min      total-odr0-8192            316.00 (  0.00%)           314.00 (  0.63%)
Min      total-odr0-16384           317.00 (  0.00%)           315.00 (  0.63%)
Min      total-odr1-1               742.00 (  0.00%)           712.00 (  4.04%)
Min      total-odr1-2               562.00 (  0.00%)           530.00 (  5.69%)
Min      total-odr1-4               457.00 (  0.00%)           433.00 (  5.25%)
Min      total-odr1-8               411.00 (  0.00%)           381.00 (  7.30%)
Min      total-odr1-16              381.00 (  0.00%)           356.00 (  6.56%)
Min      total-odr1-32              372.00 (  0.00%)           346.00 (  6.99%)
Min      total-odr1-64              372.00 (  0.00%)           343.00 (  7.80%)
Min      total-odr1-128             375.00 (  0.00%)           351.00 (  6.40%)
Min      total-odr1-256             379.00 (  0.00%)           351.00 (  7.39%)
Min      total-odr1-512             385.00 (  0.00%)           355.00 (  7.79%)
Min      total-odr1-1024            386.00 (  0.00%)           358.00 (  7.25%)
Min      total-odr1-2048            390.00 (  0.00%)           362.00 (  7.18%)
Min      total-odr1-4096            390.00 (  0.00%)           362.00 (  7.18%)
Min      total-odr1-8192            388.00 (  0.00%)           363.00 (  6.44%)

This shows a steady improvement throughout. The primary benefit is from
reduced system CPU usage which is obvious from the overall times;

           4.7.0-rc4   4.7.0-rc4
        mmotm-20160623nodelru-v8
User          189.19      191.80
System       2604.45     2533.56
Elapsed      2855.30     2786.39

The vmstats also showed that the fair zone allocation policy was definitely
removed as can be seen here;

                             4.7.0-rc3   4.7.0-rc3
                         mmotm-20160623 nodelru-v8
DMA32 allocs               28794729769           0
Normal allocs              48432501431 77227309877
Movable allocs                       0           0

tiobench on ext4
----------------

tiobench is a benchmark that artifically benefits if old pages remain resident
while new pages get reclaimed. The fair zone allocation policy mitigates this
problem so pages age fairly. While the benchmark has problems, it is important
that tiobench performance remains constant as it implies that page aging
problems that the fair zone allocation policy fixes are not re-introduced.

                                         4.7.0-rc4             4.7.0-rc4
                                    mmotm-20160623            nodelru-v9
Min      PotentialReadSpeed        89.65 (  0.00%)       90.21 (  0.62%)
Min      SeqRead-MB/sec-1          82.68 (  0.00%)       82.01 ( -0.81%)
Min      SeqRead-MB/sec-2          72.76 (  0.00%)       72.07 ( -0.95%)
Min      SeqRead-MB/sec-4          75.13 (  0.00%)       74.92 ( -0.28%)
Min      SeqRead-MB/sec-8          64.91 (  0.00%)       65.19 (  0.43%)
Min      SeqRead-MB/sec-16         62.24 (  0.00%)       62.22 ( -0.03%)
Min      RandRead-MB/sec-1          0.88 (  0.00%)        0.88 (  0.00%)
Min      RandRead-MB/sec-2          0.95 (  0.00%)        0.92 ( -3.16%)
Min      RandRead-MB/sec-4          1.43 (  0.00%)        1.34 ( -6.29%)
Min      RandRead-MB/sec-8          1.61 (  0.00%)        1.60 ( -0.62%)
Min      RandRead-MB/sec-16         1.80 (  0.00%)        1.90 (  5.56%)
Min      SeqWrite-MB/sec-1         76.41 (  0.00%)       76.85 (  0.58%)
Min      SeqWrite-MB/sec-2         74.11 (  0.00%)       73.54 ( -0.77%)
Min      SeqWrite-MB/sec-4         80.05 (  0.00%)       80.13 (  0.10%)
Min      SeqWrite-MB/sec-8         72.88 (  0.00%)       73.20 (  0.44%)
Min      SeqWrite-MB/sec-16        75.91 (  0.00%)       76.44 (  0.70%)
Min      RandWrite-MB/sec-1         1.18 (  0.00%)        1.14 ( -3.39%)
Min      RandWrite-MB/sec-2         1.02 (  0.00%)        1.03 (  0.98%)
Min      RandWrite-MB/sec-4         1.05 (  0.00%)        0.98 ( -6.67%)
Min      RandWrite-MB/sec-8         0.89 (  0.00%)        0.92 (  3.37%)
Min      RandWrite-MB/sec-16        0.92 (  0.00%)        0.93 (  1.09%)

           4.7.0-rc4   4.7.0-rc4
        mmotm-20160623 approx-v9
User          645.72      525.90
System        403.85      331.75
Elapsed      6795.36     6783.67

This shows that the series has little or not impact on tiobench which is
desirable and a reduction in system CPU usage. It indicates that the fair
zone allocation policy was removed in a manner that didn't reintroduce
one class of page aging bug. There were only minor differences in overall
reclaim activity

                             4.7.0-rc4   4.7.0-rc4
                          mmotm-20160623nodelru-v8
Minor Faults                    645838      647465
Major Faults                       573         640
Swap Ins                             0           0
Swap Outs                            0           0
DMA allocs                           0           0
DMA32 allocs                  46041453    44190646
Normal allocs                 78053072    79887245
Movable allocs                       0           0
Allocation stalls                   24          67
Stall zone DMA                       0           0
Stall zone DMA32                     0           0
Stall zone Normal                    0           2
Stall zone HighMem                   0           0
Stall zone Movable                   0          65
Direct pages scanned             10969       30609
Kswapd pages scanned          93375144    93492094
Kswapd pages reclaimed        93372243    93489370
Direct pages reclaimed           10969       30609
Kswapd efficiency                  99%         99%
Kswapd velocity              13741.015   13781.934
Direct efficiency                 100%        100%
Direct velocity                  1.614       4.512
Percentage direct scans             0%          0%

kswapd activity was roughly comparable. There were differences in direct
reclaim activity but negligible in the context of the overall workload
(velocity of 4 pages per second with the patches applied, 1.6 pages per
second in the baseline kernel).

pgbench read-only large configuration on ext4
---------------------------------------------

pgbench is a database benchmark that can be sensitive to page reclaim
decisions. This also checks if removing the fair zone allocation policy
is safe

pgbench Transactions
                        4.7.0-rc4             4.7.0-rc4
                   mmotm-20160623            nodelru-v8
Hmean    1       188.26 (  0.00%)      189.78 (  0.81%)
Hmean    5       330.66 (  0.00%)      328.69 ( -0.59%)
Hmean    12      370.32 (  0.00%)      380.72 (  2.81%)
Hmean    21      368.89 (  0.00%)      369.00 (  0.03%)
Hmean    30      382.14 (  0.00%)      360.89 ( -5.56%)
Hmean    32      428.87 (  0.00%)      432.96 (  0.95%)

Negligible differences again. As with tiobench, overall reclaim activity
was comparable.

bonnie++ on ext4
----------------

No interesting performance difference, negligible differences on reclaim
stats.

paralleldd on ext4
------------------

This workload uses varying numbers of dd instances to read large amounts of
data from disk.

                               4.7.0-rc3             4.7.0-rc3
                          mmotm-20160623            nodelru-v9
Amean    Elapsd-1       186.04 (  0.00%)      189.41 ( -1.82%)
Amean    Elapsd-3       192.27 (  0.00%)      191.38 (  0.46%)
Amean    Elapsd-5       185.21 (  0.00%)      182.75 (  1.33%)
Amean    Elapsd-7       183.71 (  0.00%)      182.11 (  0.87%)
Amean    Elapsd-12      180.96 (  0.00%)      181.58 ( -0.35%)
Amean    Elapsd-16      181.36 (  0.00%)      183.72 ( -1.30%)

           4.7.0-rc4   4.7.0-rc4
        mmotm-20160623 nodelru-v9
User         1548.01     1552.44
System       8609.71     8515.08
Elapsed      3587.10     3594.54

There is little or no change in performance but some drop in system CPU usage.

                             4.7.0-rc3   4.7.0-rc3
                        mmotm-20160623  nodelru-v9
Minor Faults                    362662      367360
Major Faults                      1204        1143
Swap Ins                            22           0
Swap Outs                         2855        1029
DMA allocs                           0           0
DMA32 allocs                  31409797    28837521
Normal allocs                 46611853    49231282
Movable allocs                       0           0
Direct pages scanned                 0           0
Kswapd pages scanned          40845270    40869088
Kswapd pages reclaimed        40830976    40855294
Direct pages reclaimed               0           0
Kswapd efficiency                  99%         99%
Kswapd velocity              11386.711   11369.769
Direct efficiency                 100%        100%
Direct velocity                  0.000       0.000
Percentage direct scans             0%          0%
Page writes by reclaim            2855        1029
Page writes file                     0           0
Page writes anon                  2855        1029
Page reclaim immediate             771        1628
Sector Reads                 293312636   293536360
Sector Writes                 18213568    18186480
Page rescued immediate               0           0
Slabs scanned                   128257      132747
Direct inode steals                181          56
Kswapd inode steals                 59        1131

It basically shows that kswapd was active at roughly the same rate in
both kernels. There was also comparable slab scanning activity and direct
reclaim was avoided in both cases. There appears to be a large difference
in numbers of inodes reclaimed but the workload has few active inodes and
is likely a timing artifact.

stutter
-------

stutter simulates a simple workload. One part uses a lot of anonymous
memory, a second measures mmap latency and a third copies a large file.
The primary metric is checking for mmap latency.

stutter
                             4.7.0-rc4             4.7.0-rc4
                        mmotm-20160623            nodelru-v8
Min         mmap     16.6283 (  0.00%)     13.4258 ( 19.26%)
1st-qrtle   mmap     54.7570 (  0.00%)     34.9121 ( 36.24%)
2nd-qrtle   mmap     57.3163 (  0.00%)     46.1147 ( 19.54%)
3rd-qrtle   mmap     58.9976 (  0.00%)     47.1882 ( 20.02%)
Max-90%     mmap     59.7433 (  0.00%)     47.4453 ( 20.58%)
Max-93%     mmap     60.1298 (  0.00%)     47.6037 ( 20.83%)
Max-95%     mmap     73.4112 (  0.00%)     82.8719 (-12.89%)
Max-99%     mmap     92.8542 (  0.00%)     88.8870 (  4.27%)
Max         mmap   1440.6569 (  0.00%)    121.4201 ( 91.57%)
Mean        mmap     59.3493 (  0.00%)     42.2991 ( 28.73%)
Best99%Mean mmap     57.2121 (  0.00%)     41.8207 ( 26.90%)
Best95%Mean mmap     55.9113 (  0.00%)     39.9620 ( 28.53%)
Best90%Mean mmap     55.6199 (  0.00%)     39.3124 ( 29.32%)
Best50%Mean mmap     53.2183 (  0.00%)     33.1307 ( 37.75%)
Best10%Mean mmap     45.9842 (  0.00%)     20.4040 ( 55.63%)
Best5%Mean  mmap     43.2256 (  0.00%)     17.9654 ( 58.44%)
Best1%Mean  mmap     32.9388 (  0.00%)     16.6875 ( 49.34%)

This shows a number of improvements with the worst-case outlier greatly
improved.

Some of the vmstats are interesting

                             4.7.0-rc4   4.7.0-rc4
                          mmotm-20160623nodelru-v8
Swap Ins                           163         502
Swap Outs                            0           0
DMA allocs                           0           0
DMA32 allocs                 618719206  1381662383
Normal allocs                891235743   564138421
Movable allocs                       0           0
Allocation stalls                 2603           1
Direct pages scanned            216787           2
Kswapd pages scanned          50719775    41778378
Kswapd pages reclaimed        41541765    41777639
Direct pages reclaimed          209159           0
Kswapd efficiency                  81%         99%
Kswapd velocity              16859.554   14329.059
Direct efficiency                  96%          0%
Direct velocity                 72.061       0.001
Percentage direct scans             0%          0%
Page writes by reclaim         6215049           0
Page writes file               6215049           0
Page writes anon                     0           0
Page reclaim immediate           70673          90
Sector Reads                  81940800    81680456
Sector Writes                100158984    98816036
Page rescued immediate               0           0
Slabs scanned                  1366954       22683

While this is not guaranteed in all cases, this particular test showed
a large reduction in direct reclaim activity. It's also worth noting
that no page writes were issued from reclaim context.

This series is not without its hazards. There are at least three areas
that I'm concerned with even though I could not reproduce any problems in
that area.

1. Reclaim/compaction is going to be affected because the amount of reclaim is
   no longer targetted at a specific zone. Compaction works on a per-zone basis
   so there is no guarantee that reclaiming a few THP's worth page pages will
   have a positive impact on compaction success rates.

2. The Slab/LRU reclaim ratio is affected because the frequency the shrinkers
   are called is now different. This may or may not be a problem but if it
   is, it'll be because shrinkers are not called enough and some balancing
   is required.

3. The anon/file reclaim ratio may be affected. Pages about to be dirtied are
   distributed between zones and the fair zone allocation policy used to do
   something very similar for anon. The distribution is now different but not
   necessarily in any way that matters but it's still worth bearing in mind.

VM statistic counters for reclaim decisions are zone-based.  If the kernel
is to reclaim on a per-node basis then we need to track per-node
statistics but there is no infrastructure for that.  The most notable
change is that the old node_page_state is renamed to
sum_zone_node_page_state.  The new node_page_state takes a pglist_data and
uses per-node stats but none exist yet.  There is some renaming such as
vm_stat to vm_zone_stat and the addition of vm_node_stat and the renaming
of mod_state to mod_zone_state.  Otherwise, this is mostly a mechanical
patch with no functional change.  There is a lot of similarity between the
node and zone helpers which is unfortunate but there was no obvious way of
reusing the code and maintaining type safety.

Link: http://lkml.kernel.org/r/1467970510-21195-2-git-send-email-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Rik van Riel <riel@surriel.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
drivers/base/node.c
include/linux/mm.h
include/linux/mmzone.h
include/linux/vmstat.h
mm/page_alloc.c
mm/vmstat.c
mm/workingset.c

index 51c7db2..0a1b643 100644 (file)
@@ -74,16 +74,16 @@ static ssize_t node_read_meminfo(struct device *dev,
                       nid, K(i.totalram),
                       nid, K(i.freeram),
                       nid, K(i.totalram - i.freeram),
-                      nid, K(node_page_state(nid, NR_ACTIVE_ANON) +
-                               node_page_state(nid, NR_ACTIVE_FILE)),
-                      nid, K(node_page_state(nid, NR_INACTIVE_ANON) +
-                               node_page_state(nid, NR_INACTIVE_FILE)),
-                      nid, K(node_page_state(nid, NR_ACTIVE_ANON)),
-                      nid, K(node_page_state(nid, NR_INACTIVE_ANON)),
-                      nid, K(node_page_state(nid, NR_ACTIVE_FILE)),
-                      nid, K(node_page_state(nid, NR_INACTIVE_FILE)),
-                      nid, K(node_page_state(nid, NR_UNEVICTABLE)),
-                      nid, K(node_page_state(nid, NR_MLOCK)));
+                      nid, K(sum_zone_node_page_state(nid, NR_ACTIVE_ANON) +
+                               sum_zone_node_page_state(nid, NR_ACTIVE_FILE)),
+                      nid, K(sum_zone_node_page_state(nid, NR_INACTIVE_ANON) +
+                               sum_zone_node_page_state(nid, NR_INACTIVE_FILE)),
+                      nid, K(sum_zone_node_page_state(nid, NR_ACTIVE_ANON)),
+                      nid, K(sum_zone_node_page_state(nid, NR_INACTIVE_ANON)),
+                      nid, K(sum_zone_node_page_state(nid, NR_ACTIVE_FILE)),
+                      nid, K(sum_zone_node_page_state(nid, NR_INACTIVE_FILE)),
+                      nid, K(sum_zone_node_page_state(nid, NR_UNEVICTABLE)),
+                      nid, K(sum_zone_node_page_state(nid, NR_MLOCK)));
 
 #ifdef CONFIG_HIGHMEM
        n += sprintf(buf + n,
@@ -117,31 +117,31 @@ static ssize_t node_read_meminfo(struct device *dev,
                       "Node %d ShmemPmdMapped: %8lu kB\n"
 #endif
                        ,
-                      nid, K(node_page_state(nid, NR_FILE_DIRTY)),
-                      nid, K(node_page_state(nid, NR_WRITEBACK)),
-                      nid, K(node_page_state(nid, NR_FILE_PAGES)),
-                      nid, K(node_page_state(nid, NR_FILE_MAPPED)),
-                      nid, K(node_page_state(nid, NR_ANON_PAGES)),
+                      nid, K(sum_zone_node_page_state(nid, NR_FILE_DIRTY)),
+                      nid, K(sum_zone_node_page_state(nid, NR_WRITEBACK)),
+                      nid, K(sum_zone_node_page_state(nid, NR_FILE_PAGES)),
+                      nid, K(sum_zone_node_page_state(nid, NR_FILE_MAPPED)),
+                      nid, K(sum_zone_node_page_state(nid, NR_ANON_PAGES)),
                       nid, K(i.sharedram),
-                      nid, node_page_state(nid, NR_KERNEL_STACK) *
+                      nid, sum_zone_node_page_state(nid, NR_KERNEL_STACK) *
                                THREAD_SIZE / 1024,
-                      nid, K(node_page_state(nid, NR_PAGETABLE)),
-                      nid, K(node_page_state(nid, NR_UNSTABLE_NFS)),
-                      nid, K(node_page_state(nid, NR_BOUNCE)),
-                      nid, K(node_page_state(nid, NR_WRITEBACK_TEMP)),
-                      nid, K(node_page_state(nid, NR_SLAB_RECLAIMABLE) +
-                               node_page_state(nid, NR_SLAB_UNRECLAIMABLE)),
-                      nid, K(node_page_state(nid, NR_SLAB_RECLAIMABLE)),
+                      nid, K(sum_zone_node_page_state(nid, NR_PAGETABLE)),
+                      nid, K(sum_zone_node_page_state(nid, NR_UNSTABLE_NFS)),
+                      nid, K(sum_zone_node_page_state(nid, NR_BOUNCE)),
+                      nid, K(sum_zone_node_page_state(nid, NR_WRITEBACK_TEMP)),
+                      nid, K(sum_zone_node_page_state(nid, NR_SLAB_RECLAIMABLE) +
+                               sum_zone_node_page_state(nid, NR_SLAB_UNRECLAIMABLE)),
+                      nid, K(sum_zone_node_page_state(nid, NR_SLAB_RECLAIMABLE)),
 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
-                      nid, K(node_page_state(nid, NR_SLAB_UNRECLAIMABLE)),
-                      nid, K(node_page_state(nid, NR_ANON_THPS) *
+                      nid, K(sum_zone_node_page_state(nid, NR_SLAB_UNRECLAIMABLE)),
+                      nid, K(sum_zone_node_page_state(nid, NR_ANON_THPS) *
                                       HPAGE_PMD_NR),
-                      nid, K(node_page_state(nid, NR_SHMEM_THPS) *
+                      nid, K(sum_zone_node_page_state(nid, NR_SHMEM_THPS) *
                                       HPAGE_PMD_NR),
-                      nid, K(node_page_state(nid, NR_SHMEM_PMDMAPPED) *
+                      nid, K(sum_zone_node_page_state(nid, NR_SHMEM_PMDMAPPED) *
                                       HPAGE_PMD_NR));
 #else
-                      nid, K(node_page_state(nid, NR_SLAB_UNRECLAIMABLE)));
+                      nid, K(sum_zone_node_page_state(nid, NR_SLAB_UNRECLAIMABLE)));
 #endif
        n += hugetlb_report_node_meminfo(nid, buf + n);
        return n;
@@ -160,12 +160,12 @@ static ssize_t node_read_numastat(struct device *dev,
                       "interleave_hit %lu\n"
                       "local_node %lu\n"
                       "other_node %lu\n",
-                      node_page_state(dev->id, NUMA_HIT),
-                      node_page_state(dev->id, NUMA_MISS),
-                      node_page_state(dev->id, NUMA_FOREIGN),
-                      node_page_state(dev->id, NUMA_INTERLEAVE_HIT),
-                      node_page_state(dev->id, NUMA_LOCAL),
-                      node_page_state(dev->id, NUMA_OTHER));
+                      sum_zone_node_page_state(dev->id, NUMA_HIT),
+                      sum_zone_node_page_state(dev->id, NUMA_MISS),
+                      sum_zone_node_page_state(dev->id, NUMA_FOREIGN),
+                      sum_zone_node_page_state(dev->id, NUMA_INTERLEAVE_HIT),
+                      sum_zone_node_page_state(dev->id, NUMA_LOCAL),
+                      sum_zone_node_page_state(dev->id, NUMA_OTHER));
 }
 static DEVICE_ATTR(numastat, S_IRUGO, node_read_numastat, NULL);
 
@@ -173,12 +173,18 @@ static ssize_t node_read_vmstat(struct device *dev,
                                struct device_attribute *attr, char *buf)
 {
        int nid = dev->id;
+       struct pglist_data *pgdat = NODE_DATA(nid);
        int i;
        int n = 0;
 
        for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
                n += sprintf(buf+n, "%s %lu\n", vmstat_text[i],
-                            node_page_state(nid, i));
+                            sum_zone_node_page_state(nid, i));
+
+       for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
+               n += sprintf(buf+n, "%s %lu\n",
+                            vmstat_text[i + NR_VM_ZONE_STAT_ITEMS],
+                            node_page_state(pgdat, i));
 
        return n;
 }
index 97065e1..08ed53e 100644 (file)
@@ -933,6 +933,11 @@ static inline struct zone *page_zone(const struct page *page)
        return &NODE_DATA(page_to_nid(page))->node_zones[page_zonenum(page)];
 }
 
+static inline pg_data_t *page_pgdat(const struct page *page)
+{
+       return NODE_DATA(page_to_nid(page));
+}
+
 #ifdef SECTION_IN_PAGE_FLAGS
 static inline void set_page_section(struct page *page, unsigned long section)
 {
index 19425e9..078ecb8 100644 (file)
@@ -160,6 +160,10 @@ enum zone_stat_item {
        NR_FREE_CMA_PAGES,
        NR_VM_ZONE_STAT_ITEMS };
 
+enum node_stat_item {
+       NR_VM_NODE_STAT_ITEMS
+};
+
 /*
  * We do arithmetic on the LRU lists in various places in the code,
  * so it is important to keep the active lists LRU_ACTIVE higher in
@@ -267,6 +271,11 @@ struct per_cpu_pageset {
 #endif
 };
 
+struct per_cpu_nodestat {
+       s8 stat_threshold;
+       s8 vm_node_stat_diff[NR_VM_NODE_STAT_ITEMS];
+};
+
 #endif /* !__GENERATING_BOUNDS.H */
 
 enum zone_type {
@@ -695,6 +704,10 @@ typedef struct pglist_data {
        struct list_head split_queue;
        unsigned long split_queue_len;
 #endif
+
+       /* Per-node vmstats */
+       struct per_cpu_nodestat __percpu *per_cpu_nodestats;
+       atomic_long_t           vm_stat[NR_VM_NODE_STAT_ITEMS];
 } pg_data_t;
 
 #define node_present_pages(nid)        (NODE_DATA(nid)->node_present_pages)
index d2da8e0..d1744aa 100644 (file)
@@ -106,20 +106,38 @@ static inline void vm_events_fold_cpu(int cpu)
                zone_idx(zone), delta)
 
 /*
- * Zone based page accounting with per cpu differentials.
+ * Zone and node-based page accounting with per cpu differentials.
  */
-extern atomic_long_t vm_stat[NR_VM_ZONE_STAT_ITEMS];
+extern atomic_long_t vm_zone_stat[NR_VM_ZONE_STAT_ITEMS];
+extern atomic_long_t vm_node_stat[NR_VM_NODE_STAT_ITEMS];
 
 static inline void zone_page_state_add(long x, struct zone *zone,
                                 enum zone_stat_item item)
 {
        atomic_long_add(x, &zone->vm_stat[item]);
-       atomic_long_add(x, &vm_stat[item]);
+       atomic_long_add(x, &vm_zone_stat[item]);
+}
+
+static inline void node_page_state_add(long x, struct pglist_data *pgdat,
+                                enum node_stat_item item)
+{
+       atomic_long_add(x, &pgdat->vm_stat[item]);
+       atomic_long_add(x, &vm_node_stat[item]);
 }
 
 static inline unsigned long global_page_state(enum zone_stat_item item)
 {
-       long x = atomic_long_read(&vm_stat[item]);
+       long x = atomic_long_read(&vm_zone_stat[item]);
+#ifdef CONFIG_SMP
+       if (x < 0)
+               x = 0;
+#endif
+       return x;
+}
+
+static inline unsigned long global_node_page_state(enum node_stat_item item)
+{
+       long x = atomic_long_read(&vm_node_stat[item]);
 #ifdef CONFIG_SMP
        if (x < 0)
                x = 0;
@@ -161,31 +179,44 @@ static inline unsigned long zone_page_state_snapshot(struct zone *zone,
 }
 
 #ifdef CONFIG_NUMA
-
-extern unsigned long node_page_state(int node, enum zone_stat_item item);
-
+extern unsigned long sum_zone_node_page_state(int node,
+                                               enum zone_stat_item item);
+extern unsigned long node_page_state(struct pglist_data *pgdat,
+                                               enum node_stat_item item);
 #else
-
-#define node_page_state(node, item) global_page_state(item)
-
+#define sum_zone_node_page_state(node, item) global_page_state(item)
+#define node_page_state(node, item) global_node_page_state(item)
 #endif /* CONFIG_NUMA */
 
 #define add_zone_page_state(__z, __i, __d) mod_zone_page_state(__z, __i, __d)
 #define sub_zone_page_state(__z, __i, __d) mod_zone_page_state(__z, __i, -(__d))
+#define add_node_page_state(__p, __i, __d) mod_node_page_state(__p, __i, __d)
+#define sub_node_page_state(__p, __i, __d) mod_node_page_state(__p, __i, -(__d))
 
 #ifdef CONFIG_SMP
 void __mod_zone_page_state(struct zone *, enum zone_stat_item item, long);
 void __inc_zone_page_state(struct page *, enum zone_stat_item);
 void __dec_zone_page_state(struct page *, enum zone_stat_item);
 
+void __mod_node_page_state(struct pglist_data *, enum node_stat_item item, long);
+void __inc_node_page_state(struct page *, enum node_stat_item);
+void __dec_node_page_state(struct page *, enum node_stat_item);
+
 void mod_zone_page_state(struct zone *, enum zone_stat_item, long);
 void inc_zone_page_state(struct page *, enum zone_stat_item);
 void dec_zone_page_state(struct page *, enum zone_stat_item);
 
+void mod_node_page_state(struct pglist_data *, enum node_stat_item, long);
+void inc_node_page_state(struct page *, enum node_stat_item);
+void dec_node_page_state(struct page *, enum node_stat_item);
+
 extern void inc_zone_state(struct zone *, enum zone_stat_item);
+extern void inc_node_state(struct pglist_data *, enum node_stat_item);
 extern void __inc_zone_state(struct zone *, enum zone_stat_item);
+extern void __inc_node_state(struct pglist_data *, enum node_stat_item);
 extern void dec_zone_state(struct zone *, enum zone_stat_item);
 extern void __dec_zone_state(struct zone *, enum zone_stat_item);
+extern void __dec_node_state(struct pglist_data *, enum node_stat_item);
 
 void quiet_vmstat(void);
 void cpu_vm_stats_fold(int cpu);
@@ -213,16 +244,34 @@ static inline void __mod_zone_page_state(struct zone *zone,
        zone_page_state_add(delta, zone, item);
 }
 
+static inline void __mod_node_page_state(struct pglist_data *pgdat,
+                       enum node_stat_item item, int delta)
+{
+       node_page_state_add(delta, pgdat, item);
+}
+
 static inline void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
 {
        atomic_long_inc(&zone->vm_stat[item]);
-       atomic_long_inc(&vm_stat[item]);
+       atomic_long_inc(&vm_zone_stat[item]);
+}
+
+static inline void __inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
+{
+       atomic_long_inc(&pgdat->vm_stat[item]);
+       atomic_long_inc(&vm_node_stat[item]);
 }
 
 static inline void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
 {
        atomic_long_dec(&zone->vm_stat[item]);
-       atomic_long_dec(&vm_stat[item]);
+       atomic_long_dec(&vm_zone_stat[item]);
+}
+
+static inline void __dec_node_state(struct pglist_data *pgdat, enum node_stat_item item)
+{
+       atomic_long_dec(&pgdat->vm_stat[item]);
+       atomic_long_dec(&vm_node_stat[item]);
 }
 
 static inline void __inc_zone_page_state(struct page *page,
@@ -231,12 +280,26 @@ static inline void __inc_zone_page_state(struct page *page,
        __inc_zone_state(page_zone(page), item);
 }
 
+static inline void __inc_node_page_state(struct page *page,
+                       enum node_stat_item item)
+{
+       __inc_node_state(page_pgdat(page), item);
+}
+
+
 static inline void __dec_zone_page_state(struct page *page,
                        enum zone_stat_item item)
 {
        __dec_zone_state(page_zone(page), item);
 }
 
+static inline void __dec_node_page_state(struct page *page,
+                       enum node_stat_item item)
+{
+       __dec_node_state(page_pgdat(page), item);
+}
+
+
 /*
  * We only use atomic operations to update counters. So there is no need to
  * disable interrupts.
@@ -245,7 +308,12 @@ static inline void __dec_zone_page_state(struct page *page,
 #define dec_zone_page_state __dec_zone_page_state
 #define mod_zone_page_state __mod_zone_page_state
 
+#define inc_node_page_state __inc_node_page_state
+#define dec_node_page_state __dec_node_page_state
+#define mod_node_page_state __mod_node_page_state
+
 #define inc_zone_state __inc_zone_state
+#define inc_node_state __inc_node_state
 #define dec_zone_state __dec_zone_state
 
 #define set_pgdat_percpu_threshold(pgdat, callback) { }
index d7a7970..7d4ff81 100644 (file)
@@ -4204,8 +4204,8 @@ void si_meminfo_node(struct sysinfo *val, int nid)
        for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
                managed_pages += pgdat->node_zones[zone_type].managed_pages;
        val->totalram = managed_pages;
-       val->sharedram = node_page_state(nid, NR_SHMEM);
-       val->freeram = node_page_state(nid, NR_FREE_PAGES);
+       val->sharedram = sum_zone_node_page_state(nid, NR_SHMEM);
+       val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
 #ifdef CONFIG_HIGHMEM
        for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
                struct zone *zone = &pgdat->node_zones[zone_type];
@@ -5330,6 +5330,11 @@ static void __meminit setup_zone_pageset(struct zone *zone)
        zone->pageset = alloc_percpu(struct per_cpu_pageset);
        for_each_possible_cpu(cpu)
                zone_pageset_init(zone, cpu);
+
+       if (!zone->zone_pgdat->per_cpu_nodestats) {
+               zone->zone_pgdat->per_cpu_nodestats =
+                       alloc_percpu(struct per_cpu_nodestat);
+       }
 }
 
 /*
@@ -6033,6 +6038,7 @@ void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
        reset_deferred_meminit(pgdat);
        pgdat->node_id = nid;
        pgdat->node_start_pfn = node_start_pfn;
+       pgdat->per_cpu_nodestats = NULL;
 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
        get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
        pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
index 7997f52..3345d39 100644 (file)
@@ -86,8 +86,10 @@ void vm_events_fold_cpu(int cpu)
  *
  * vm_stat contains the global counters
  */
-atomic_long_t vm_stat[NR_VM_ZONE_STAT_ITEMS] __cacheline_aligned_in_smp;
-EXPORT_SYMBOL(vm_stat);
+atomic_long_t vm_zone_stat[NR_VM_ZONE_STAT_ITEMS] __cacheline_aligned_in_smp;
+atomic_long_t vm_node_stat[NR_VM_NODE_STAT_ITEMS] __cacheline_aligned_in_smp;
+EXPORT_SYMBOL(vm_zone_stat);
+EXPORT_SYMBOL(vm_node_stat);
 
 #ifdef CONFIG_SMP
 
@@ -167,19 +169,36 @@ int calculate_normal_threshold(struct zone *zone)
  */
 void refresh_zone_stat_thresholds(void)
 {
+       struct pglist_data *pgdat;
        struct zone *zone;
        int cpu;
        int threshold;
 
+       /* Zero current pgdat thresholds */
+       for_each_online_pgdat(pgdat) {
+               for_each_online_cpu(cpu) {
+                       per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold = 0;
+               }
+       }
+
        for_each_populated_zone(zone) {
+               struct pglist_data *pgdat = zone->zone_pgdat;
                unsigned long max_drift, tolerate_drift;
 
                threshold = calculate_normal_threshold(zone);
 
-               for_each_online_cpu(cpu)
+               for_each_online_cpu(cpu) {
+                       int pgdat_threshold;
+
                        per_cpu_ptr(zone->pageset, cpu)->stat_threshold
                                                        = threshold;
 
+                       /* Base nodestat threshold on the largest populated zone. */
+                       pgdat_threshold = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold;
+                       per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold
+                               = max(threshold, pgdat_threshold);
+               }
+
                /*
                 * Only set percpu_drift_mark if there is a danger that
                 * NR_FREE_PAGES reports the low watermark is ok when in fact
@@ -238,6 +257,26 @@ void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 }
 EXPORT_SYMBOL(__mod_zone_page_state);
 
+void __mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
+                               long delta)
+{
+       struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
+       s8 __percpu *p = pcp->vm_node_stat_diff + item;
+       long x;
+       long t;
+
+       x = delta + __this_cpu_read(*p);
+
+       t = __this_cpu_read(pcp->stat_threshold);
+
+       if (unlikely(x > t || x < -t)) {
+               node_page_state_add(x, pgdat, item);
+               x = 0;
+       }
+       __this_cpu_write(*p, x);
+}
+EXPORT_SYMBOL(__mod_node_page_state);
+
 /*
  * Optimized increment and decrement functions.
  *
@@ -277,12 +316,34 @@ void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
        }
 }
 
+void __inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
+{
+       struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
+       s8 __percpu *p = pcp->vm_node_stat_diff + item;
+       s8 v, t;
+
+       v = __this_cpu_inc_return(*p);
+       t = __this_cpu_read(pcp->stat_threshold);
+       if (unlikely(v > t)) {
+               s8 overstep = t >> 1;
+
+               node_page_state_add(v + overstep, pgdat, item);
+               __this_cpu_write(*p, -overstep);
+       }
+}
+
 void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
 {
        __inc_zone_state(page_zone(page), item);
 }
 EXPORT_SYMBOL(__inc_zone_page_state);
 
+void __inc_node_page_state(struct page *page, enum node_stat_item item)
+{
+       __inc_node_state(page_pgdat(page), item);
+}
+EXPORT_SYMBOL(__inc_node_page_state);
+
 void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
 {
        struct per_cpu_pageset __percpu *pcp = zone->pageset;
@@ -299,12 +360,34 @@ void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
        }
 }
 
+void __dec_node_state(struct pglist_data *pgdat, enum node_stat_item item)
+{
+       struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
+       s8 __percpu *p = pcp->vm_node_stat_diff + item;
+       s8 v, t;
+
+       v = __this_cpu_dec_return(*p);
+       t = __this_cpu_read(pcp->stat_threshold);
+       if (unlikely(v < - t)) {
+               s8 overstep = t >> 1;
+
+               node_page_state_add(v - overstep, pgdat, item);
+               __this_cpu_write(*p, overstep);
+       }
+}
+
 void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
 {
        __dec_zone_state(page_zone(page), item);
 }
 EXPORT_SYMBOL(__dec_zone_page_state);
 
+void __dec_node_page_state(struct page *page, enum node_stat_item item)
+{
+       __dec_node_state(page_pgdat(page), item);
+}
+EXPORT_SYMBOL(__dec_node_page_state);
+
 #ifdef CONFIG_HAVE_CMPXCHG_LOCAL
 /*
  * If we have cmpxchg_local support then we do not need to incur the overhead
@@ -318,8 +401,8 @@ EXPORT_SYMBOL(__dec_zone_page_state);
  *     1       Overstepping half of threshold
  *     -1      Overstepping minus half of threshold
 */
-static inline void mod_state(struct zone *zone, enum zone_stat_item item,
-                            long delta, int overstep_mode)
+static inline void mod_zone_state(struct zone *zone,
+       enum zone_stat_item item, long delta, int overstep_mode)
 {
        struct per_cpu_pageset __percpu *pcp = zone->pageset;
        s8 __percpu *p = pcp->vm_stat_diff + item;
@@ -359,26 +442,88 @@ static inline void mod_state(struct zone *zone, enum zone_stat_item item,
 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
                         long delta)
 {
-       mod_state(zone, item, delta, 0);
+       mod_zone_state(zone, item, delta, 0);
 }
 EXPORT_SYMBOL(mod_zone_page_state);
 
 void inc_zone_state(struct zone *zone, enum zone_stat_item item)
 {
-       mod_state(zone, item, 1, 1);
+       mod_zone_state(zone, item, 1, 1);
 }
 
 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
 {
-       mod_state(page_zone(page), item, 1, 1);
+       mod_zone_state(page_zone(page), item, 1, 1);
 }
 EXPORT_SYMBOL(inc_zone_page_state);
 
 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
 {
-       mod_state(page_zone(page), item, -1, -1);
+       mod_zone_state(page_zone(page), item, -1, -1);
 }
 EXPORT_SYMBOL(dec_zone_page_state);
+
+static inline void mod_node_state(struct pglist_data *pgdat,
+       enum node_stat_item item, int delta, int overstep_mode)
+{
+       struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
+       s8 __percpu *p = pcp->vm_node_stat_diff + item;
+       long o, n, t, z;
+
+       do {
+               z = 0;  /* overflow to node counters */
+
+               /*
+                * The fetching of the stat_threshold is racy. We may apply
+                * a counter threshold to the wrong the cpu if we get
+                * rescheduled while executing here. However, the next
+                * counter update will apply the threshold again and
+                * therefore bring the counter under the threshold again.
+                *
+                * Most of the time the thresholds are the same anyways
+                * for all cpus in a node.
+                */
+               t = this_cpu_read(pcp->stat_threshold);
+
+               o = this_cpu_read(*p);
+               n = delta + o;
+
+               if (n > t || n < -t) {
+                       int os = overstep_mode * (t >> 1) ;
+
+                       /* Overflow must be added to node counters */
+                       z = n + os;
+                       n = -os;
+               }
+       } while (this_cpu_cmpxchg(*p, o, n) != o);
+
+       if (z)
+               node_page_state_add(z, pgdat, item);
+}
+
+void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
+                                       long delta)
+{
+       mod_node_state(pgdat, item, delta, 0);
+}
+EXPORT_SYMBOL(mod_node_page_state);
+
+void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
+{
+       mod_node_state(pgdat, item, 1, 1);
+}
+
+void inc_node_page_state(struct page *page, enum node_stat_item item)
+{
+       mod_node_state(page_pgdat(page), item, 1, 1);
+}
+EXPORT_SYMBOL(inc_node_page_state);
+
+void dec_node_page_state(struct page *page, enum node_stat_item item)
+{
+       mod_node_state(page_pgdat(page), item, -1, -1);
+}
+EXPORT_SYMBOL(dec_node_page_state);
 #else
 /*
  * Use interrupt disable to serialize counter updates
@@ -424,21 +569,69 @@ void dec_zone_page_state(struct page *page, enum zone_stat_item item)
        local_irq_restore(flags);
 }
 EXPORT_SYMBOL(dec_zone_page_state);
-#endif
 
+void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
+{
+       unsigned long flags;
+
+       local_irq_save(flags);
+       __inc_node_state(pgdat, item);
+       local_irq_restore(flags);
+}
+EXPORT_SYMBOL(inc_node_state);
+
+void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
+                                       long delta)
+{
+       unsigned long flags;
+
+       local_irq_save(flags);
+       __mod_node_page_state(pgdat, item, delta);
+       local_irq_restore(flags);
+}
+EXPORT_SYMBOL(mod_node_page_state);
+
+void inc_node_page_state(struct page *page, enum node_stat_item item)
+{
+       unsigned long flags;
+       struct pglist_data *pgdat;
+
+       pgdat = page_pgdat(page);
+       local_irq_save(flags);
+       __inc_node_state(pgdat, item);
+       local_irq_restore(flags);
+}
+EXPORT_SYMBOL(inc_node_page_state);
+
+void dec_node_page_state(struct page *page, enum node_stat_item item)
+{
+       unsigned long flags;
+
+       local_irq_save(flags);
+       __dec_node_page_state(page, item);
+       local_irq_restore(flags);
+}
+EXPORT_SYMBOL(dec_node_page_state);
+#endif
 
 /*
  * Fold a differential into the global counters.
  * Returns the number of counters updated.
  */
-static int fold_diff(int *diff)
+static int fold_diff(int *zone_diff, int *node_diff)
 {
        int i;
        int changes = 0;
 
        for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
-               if (diff[i]) {
-                       atomic_long_add(diff[i], &vm_stat[i]);
+               if (zone_diff[i]) {
+                       atomic_long_add(zone_diff[i], &vm_zone_stat[i]);
+                       changes++;
+       }
+
+       for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
+               if (node_diff[i]) {
+                       atomic_long_add(node_diff[i], &vm_node_stat[i]);
                        changes++;
        }
        return changes;
@@ -462,9 +655,11 @@ static int fold_diff(int *diff)
  */
 static int refresh_cpu_vm_stats(bool do_pagesets)
 {
+       struct pglist_data *pgdat;
        struct zone *zone;
        int i;
-       int global_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
+       int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
+       int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
        int changes = 0;
 
        for_each_populated_zone(zone) {
@@ -477,7 +672,7 @@ static int refresh_cpu_vm_stats(bool do_pagesets)
                        if (v) {
 
                                atomic_long_add(v, &zone->vm_stat[i]);
-                               global_diff[i] += v;
+                               global_zone_diff[i] += v;
 #ifdef CONFIG_NUMA
                                /* 3 seconds idle till flush */
                                __this_cpu_write(p->expire, 3);
@@ -516,7 +711,22 @@ static int refresh_cpu_vm_stats(bool do_pagesets)
                }
 #endif
        }
-       changes += fold_diff(global_diff);
+
+       for_each_online_pgdat(pgdat) {
+               struct per_cpu_nodestat __percpu *p = pgdat->per_cpu_nodestats;
+
+               for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
+                       int v;
+
+                       v = this_cpu_xchg(p->vm_node_stat_diff[i], 0);
+                       if (v) {
+                               atomic_long_add(v, &pgdat->vm_stat[i]);
+                               global_node_diff[i] += v;
+                       }
+               }
+       }
+
+       changes += fold_diff(global_zone_diff, global_node_diff);
        return changes;
 }
 
@@ -527,9 +737,11 @@ static int refresh_cpu_vm_stats(bool do_pagesets)
  */
 void cpu_vm_stats_fold(int cpu)
 {
+       struct pglist_data *pgdat;
        struct zone *zone;
        int i;
-       int global_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
+       int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
+       int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
 
        for_each_populated_zone(zone) {
                struct per_cpu_pageset *p;
@@ -543,11 +755,27 @@ void cpu_vm_stats_fold(int cpu)
                                v = p->vm_stat_diff[i];
                                p->vm_stat_diff[i] = 0;
                                atomic_long_add(v, &zone->vm_stat[i]);
-                               global_diff[i] += v;
+                               global_zone_diff[i] += v;
                        }
        }
 
-       fold_diff(global_diff);
+       for_each_online_pgdat(pgdat) {
+               struct per_cpu_nodestat *p;
+
+               p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
+
+               for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
+                       if (p->vm_node_stat_diff[i]) {
+                               int v;
+
+                               v = p->vm_node_stat_diff[i];
+                               p->vm_node_stat_diff[i] = 0;
+                               atomic_long_add(v, &pgdat->vm_stat[i]);
+                               global_node_diff[i] += v;
+                       }
+       }
+
+       fold_diff(global_zone_diff, global_node_diff);
 }
 
 /*
@@ -563,16 +791,19 @@ void drain_zonestat(struct zone *zone, struct per_cpu_pageset *pset)
                        int v = pset->vm_stat_diff[i];
                        pset->vm_stat_diff[i] = 0;
                        atomic_long_add(v, &zone->vm_stat[i]);
-                       atomic_long_add(v, &vm_stat[i]);
+                       atomic_long_add(v, &vm_zone_stat[i]);
                }
 }
 #endif
 
 #ifdef CONFIG_NUMA
 /*
- * Determine the per node value of a stat item.
+ * Determine the per node value of a stat item. This function
+ * is called frequently in a NUMA machine, so try to be as
+ * frugal as possible.
  */
-unsigned long node_page_state(int node, enum zone_stat_item item)
+unsigned long sum_zone_node_page_state(int node,
+                                enum zone_stat_item item)
 {
        struct zone *zones = NODE_DATA(node)->node_zones;
        int i;
@@ -584,6 +815,19 @@ unsigned long node_page_state(int node, enum zone_stat_item item)
        return count;
 }
 
+/*
+ * Determine the per node value of a stat item.
+ */
+unsigned long node_page_state(struct pglist_data *pgdat,
+                               enum node_stat_item item)
+{
+       long x = atomic_long_read(&pgdat->vm_stat[item]);
+#ifdef CONFIG_SMP
+       if (x < 0)
+               x = 0;
+#endif
+       return x;
+}
 #endif
 
 #ifdef CONFIG_COMPACTION
@@ -1287,6 +1531,7 @@ static void *vmstat_start(struct seq_file *m, loff_t *pos)
        if (*pos >= ARRAY_SIZE(vmstat_text))
                return NULL;
        stat_items_size = NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long) +
+                         NR_VM_NODE_STAT_ITEMS * sizeof(unsigned long) +
                          NR_VM_WRITEBACK_STAT_ITEMS * sizeof(unsigned long);
 
 #ifdef CONFIG_VM_EVENT_COUNTERS
@@ -1301,6 +1546,10 @@ static void *vmstat_start(struct seq_file *m, loff_t *pos)
                v[i] = global_page_state(i);
        v += NR_VM_ZONE_STAT_ITEMS;
 
+       for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
+               v[i] = global_node_page_state(i);
+       v += NR_VM_NODE_STAT_ITEMS;
+
        global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
                            v + NR_DIRTY_THRESHOLD);
        v += NR_VM_WRITEBACK_STAT_ITEMS;
@@ -1390,7 +1639,7 @@ int vmstat_refresh(struct ctl_table *table, int write,
        if (err)
                return err;
        for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
-               val = atomic_long_read(&vm_stat[i]);
+               val = atomic_long_read(&vm_zone_stat[i]);
                if (val < 0) {
                        switch (i) {
                        case NR_ALLOC_BATCH:
index d7cc4bb..5ffba0c 100644 (file)
@@ -351,12 +351,13 @@ static unsigned long count_shadow_nodes(struct shrinker *shrinker,
        shadow_nodes = list_lru_shrink_count(&workingset_shadow_nodes, sc);
        local_irq_enable();
 
-       if (memcg_kmem_enabled())
+       if (memcg_kmem_enabled()) {
                pages = mem_cgroup_node_nr_lru_pages(sc->memcg, sc->nid,
                                                     LRU_ALL_FILE);
-       else
-               pages = node_page_state(sc->nid, NR_ACTIVE_FILE) +
-                       node_page_state(sc->nid, NR_INACTIVE_FILE);
+       } else {
+               pages = sum_zone_node_page_state(sc->nid, NR_ACTIVE_FILE) +
+                       sum_zone_node_page_state(sc->nid, NR_INACTIVE_FILE);
+       }
 
        /*
         * Active cache pages are limited to 50% of memory, and shadow