summaryrefslogtreecommitdiff
path: root/mm/mprotect.c
diff options
context:
space:
mode:
authorMel Gorman <mgorman@techsingularity.net>2018-04-10 16:29:20 -0700
committerLinus Torvalds <torvalds@linux-foundation.org>2018-04-11 10:28:31 -0700
commit09a913a7a947fb6f624379e9da22670994942b85 (patch)
tree5ab8bb2f6376fc679d1e9e1a1776682974df4792 /mm/mprotect.c
parente8eddfd2d9cc00a65147353afed8398651a54736 (diff)
sched/numa: avoid trapping faults and attempting migration of file-backed dirty pages
change_pte_range is called from task work context to mark PTEs for receiving NUMA faulting hints. If the marked pages are dirty then migration may fail. Some filesystems cannot migrate dirty pages without blocking so are skipped in MIGRATE_ASYNC mode which just wastes CPU. Even when they can, it can be a waste of cycles when the pages are shared forcing higher scan rates. This patch avoids marking shared dirty pages for hinting faults but also will skip a migration if the page was dirtied after the scanner updated a clean page. This is most noticeable running the NASA Parallel Benchmark when backed by btrfs, the default root filesystem for some distributions, but also noticeable when using XFS. The following are results from a 4-socket machine running a 4.16-rc4 kernel with some scheduler patches that are pending for the next merge window. 4.16.0-rc4 4.16.0-rc4 schedtip-20180309 nodirty-v1 Time cg.D 459.07 ( 0.00%) 444.21 ( 3.24%) Time ep.D 76.96 ( 0.00%) 77.69 ( -0.95%) Time is.D 25.55 ( 0.00%) 27.85 ( -9.00%) Time lu.D 601.58 ( 0.00%) 596.87 ( 0.78%) Time mg.D 107.73 ( 0.00%) 108.22 ( -0.45%) is.D regresses slightly in terms of absolute time but note that that particular load varies quite a bit from run to run. The more relevant observation is the total system CPU usage. 4.16.0-rc4 4.16.0-rc4 schedtip-20180309 nodirty-v1 User 71471.91 70627.04 System 11078.96 8256.13 Elapsed 661.66 632.74 That is a substantial drop in system CPU usage and overall the workload completes faster. The NUMA balancing statistics are also interesting NUMA base PTE updates 111407972 139848884 NUMA huge PMD updates 206506 264869 NUMA page range updates 217139044 275461812 NUMA hint faults 4300924 3719784 NUMA hint local faults 3012539 3416618 NUMA hint local percent 70 91 NUMA pages migrated 1517487 1358420 While more PTEs are scanned due to changes in what faults are gathered, it's clear that a far higher percentage of faults are local as the bulk of the remote hits were dirty pages that, in this case with btrfs, had no chance of migrating. The following is a comparison when using XFS as that is a more realistic filesystem choice for a data partition 4.16.0-rc4 4.16.0-rc4 schedtip-20180309 nodirty-v1r47 Time cg.D 485.28 ( 0.00%) 442.62 ( 8.79%) Time ep.D 77.68 ( 0.00%) 77.54 ( 0.18%) Time is.D 26.44 ( 0.00%) 24.79 ( 6.24%) Time lu.D 597.46 ( 0.00%) 597.11 ( 0.06%) Time mg.D 142.65 ( 0.00%) 105.83 ( 25.81%) That is a reasonable gain on two relatively long-lived workloads. While not presented, there is also a substantial drop in system CPu usage and the NUMA balancing stats show similar improvements in locality as btrfs did. Link: http://lkml.kernel.org/r/20180326094334.zserdec62gwmmfqf@techsingularity.net Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Reviewed-by: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Diffstat (limited to 'mm/mprotect.c')
-rw-r--r--mm/mprotect.c9
1 files changed, 9 insertions, 0 deletions
diff --git a/mm/mprotect.c b/mm/mprotect.c
index c1d6af7455da..625608bc8962 100644
--- a/mm/mprotect.c
+++ b/mm/mprotect.c
@@ -27,6 +27,7 @@
#include <linux/pkeys.h>
#include <linux/ksm.h>
#include <linux/uaccess.h>
+#include <linux/mm_inline.h>
#include <asm/pgtable.h>
#include <asm/cacheflush.h>
#include <asm/mmu_context.h>
@@ -89,6 +90,14 @@ static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
page_mapcount(page) != 1)
continue;
+ /*
+ * While migration can move some dirty pages,
+ * it cannot move them all from MIGRATE_ASYNC
+ * context.
+ */
+ if (page_is_file_cache(page) && PageDirty(page))
+ continue;
+
/* Avoid TLB flush if possible */
if (pte_protnone(oldpte))
continue;