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 noticable running the NASA Parallel Benchmark when backed by
btrfs, the default root filesystem for some distributions, but also noticable
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.
Signed-off-by: Mel Gorman <mgorman@xxxxxxxxxxxxxxxxxxx>
diff --git a/mm/migrate.c b/mm/migrate.c
index 1e5525a25691..d26832f0723b 100644
--- a/mm/migrate.c
+++ b/mm/migrate.c
@@ -1955,6 +1955,13 @@ int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
(vma->vm_flags & VM_EXEC))
goto out;
+ /*
+ * Also do not migrate dirty pages as not all filesystems can move
+ * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
+ */
+ if (page_is_file_cache(page) && PageDirty(page))
+ goto out;
+
/*
* Rate-limit the amount of data that is being migrated to a node.
* Optimal placement is no good if the memory bus is saturated and
diff --git a/mm/mprotect.c b/mm/mprotect.c
index e3309fcf586b..3cfd095e2bb0 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;
