The changes from V4 are minimal. Closing of a race window and a fix
to NR_ISOLATED_* are the big changes. Mostly, this is adding a lot of
Reviewed-by's. Thanks a million to the people that reviewed this. It caught
a lot of issues and was a big help.
Kosaki-san raised concerns about the direct compact patch (patch 10/11) where
he'd prefer to see it directly integrated with lumpy reclaim. I responded
with a number of points but heard nothing further. To follow the suggestion,
the complexity of the algorithm would need to increase significantly and
the exit conditions become trickier to use the same type of logic as lumpy
reclaim. Conceivably when there is better data available, that approach
will be taken but overall I don't think it's the best starting point.
Are there any further obstacles to merging this?
Changelog since V4
o Remove unnecessary check for PageLRU and PageUnevictable
o Fix isolated accounting
o Close race window between page_mapcount and rcu_read_lock
o Added a lot more Reviewed-by tags
Changelog since V3
o Document sysfs entries (subseqently, merged independently)
o COMPACTION should depend on MMU
o Comment updates
o Ensure proc/sysfs triggering of compaction fully completes
o Rename anon_vma refcount to external_refcount
o Rebase to mmotm on top of 2.6.34-rc1
Changelog since V2
o Move unusable and fragmentation indices to separate proc files
o Express indices as being between 0 and 1
o Update copyright notice for compaction.c
o Avoid infinite loop when split free page fails
o Init compact_resume at least once (impacted x86 testing)
o Fewer pages are isolated during compaction.
o LRU lists are no longer rotated when page is busy
o NR_ISOLATED_* is updated to avoid isolating too many pages
o Update zone LRU stats correctly when isolating pages
o Reference count anon_vma instead of insufficient locking with
use-after-free races in memory compaction
o Watch for unmapped anon pages during migration
o Remove unnecessary parameters on a few functions
o Add Reviewed-by's. Note that I didn't add the Acks and Reviewed
for the proc patches as they have been split out into separate
files and I don't know if the Acks are still valid.
Changelog since V1
o Update help blurb on CONFIG_MIGRATION
o Max unusable free space index is 100, not 1000
o Move blockpfn forward properly during compaction
o Cleanup CONFIG_COMPACTION vs CONFIG_MIGRATION confusion
o Permissions on /proc and /sys files should be 0200
o Reduce verbosity
o Compact all nodes when triggered via /proc
o Add per-node compaction via sysfs
o Move defer_compaction out-of-line
o Fix lock oddities in rmap_walk_anon
o Add documentation
This patchset is a memory compaction mechanism that reduces external
fragmentation memory by moving GFP_MOVABLE pages to a fewer number of
pageblocks. The term "compaction" was chosen as there are is a number of
mechanisms that are not mutually exclusive that can be used to defragment
memory. For example, lumpy reclaim is a form of defragmentation as was slub
"defragmentation" (really a form of targeted reclaim). Hence, this is called
"compaction" to distinguish it from other forms of defragmentation.
In this implementation, a full compaction run involves two scanners operating
within a zone - a migration and a free scanner. The migration scanner
starts at the beginning of a zone and finds all movable pages within one
pageblock_nr_pages-sized area and isolates them on a migratepages list. The
free scanner begins at the end of the zone and searches on a per-area
basis for enough free pages to migrate all the pages on the migratepages
list. As each area is respectively migrated or exhausted of free pages,
the scanners are advanced one area. A compaction run completes within a
zone when the two scanners meet.
This method is a bit primitive but is easy to understand and greater
sophistication would require maintenance of counters on a per-pageblock
basis. This would have a big impact on allocator fast-paths to improve
compaction which is a poor trade-off.
It also does not try relocate virtually contiguous pages to be physically
contiguous. However, assuming transparent hugepages were in use, a
hypothetical khugepaged might reuse compaction code to isolate free pages,
split them and relocate userspace pages for promotion.
Memory compaction can be triggered in one of three ways. It may be triggered
explicitly by writing any value to /proc/sys/vm/compact_memory and compacting
all of memory. It can be triggered on a per-node basis by writing any
value to /sys/devices/system/node/nodeN/compact where N is the node ID to
be compacted. When a process fails to allocate a high-order page, it may
compact memory in an attempt to satisfy the allocation instead of entering
direct reclaim. Explicit compaction does not finish until the two scanners
meet and direct compaction ends if a suitable page becomes available that
would meet watermarks.
The series is in 11 patches. The first three are not "core" to the series
but are important pre-requisites.
Patch 1 reference counts anon_vma for rmap_walk_anon(). Without this
patch, it's possible to use anon_vma after free if the caller is
not holding a VMA or mmap_sem for the pages in question. While
there should be no existing user that causes this problem,
it's a requirement for memory compaction to be stable. The patch
is at the start of the series for bisection reasons.
Patch 2 skips over anon pages during migration that are no longer mapped
because there still appeared to be a small window between when
a page was isolated and migration started during which anon_vma
could disappear.
Patch 3 merges the KSM and migrate counts. It could be merged with patch 1
but would be slightly harder to review.
Patch 4 allows CONFIG_MIGRATION to be set without CONFIG_NUMA
Patch 5 exports a "unusable free space index" via /proc/pagetypeinfo. It's
a measure of external fragmentation that takes the size of the
allocation request into account. It can also be calculated from
userspace so can be dropped if requested
Patch 6 exports a "fragmentation index" which only has meaning when an
allocation request fails. It determines if an allocation failure
would be due to a lack of memory or external fragmentation.
Patch 7 is the compaction mechanism although it's unreachable at this point
Patch 8 adds a means of compacting all of memory with a proc trgger
Patch 9 adds a means of compacting a specific node with a sysfs trigger
Patch 10 adds "direct compaction" before "direct reclaim" if it is
determined there is a good chance of success.
Patch 11 temporarily disables compaction if an allocation failure occurs
after compaction.
Testing of compaction was in three stages. For the test, debugging, preempt,
the sleep watchdog and lockdep were all enabled but nothing nasty popped
out. min_free_kbytes was tuned as recommended by hugeadm to help fragmentation
avoidance and high-order allocations. It was tested on X86, X86-64 and PPC64.
Ths first test represents one of the easiest cases that can be faced for
lumpy reclaim or memory compaction.
1. Machine freshly booted and configured for hugepage usage with
a) hugeadm --create-global-mounts
b) hugeadm --pool-pages-max DEFAULT:8G
c) hugeadm --set-recommended-min_free_kbytes
d) hugeadm --set-recommended-shmmax
The min_free_kbytes here is important. Anti-fragmentation works best
when pageblocks don't mix. hugeadm knows how to calculate a value that
will significantly reduce the worst of external-fragmentation-related
events as reported by the mm_page_alloc_extfrag tracepoint.
2. Load up memory
a) Start updatedb
b) Create in parallel a X files of pagesize*128 in size. Wait
until files are created. By parallel, I mean that 4096 instances
of dd were launched, one after the other using &. The crude
objective being to mix filesystem metadata allocations with
the buffer cache.
c) Delete every second file so that pageblocks are likely to
have holes
d) kill updatedb if it's still running
At this point, the system is quiet, memory is full but it's full with
clean filesystem metadata and clean buffer cache that is unmapped.
This is readily migrated or discarded so you'd expect lumpy reclaim
to have no significant advantage over compaction but this is at
the POC stage.
3. In increments, attempt to allocate 5% of memory as hugepages.
Measure how long it took, how successful it was, how many
direct reclaims took place and how how many compactions. Note
the compaction figures might not fully add up as compactions
can take place for orders other than the hugepage size
X86 vanilla compaction
Final page count 930 941 (attempted 1002)
pages reclaimed 74630 3861
X86-64 vanilla compaction
Final page count: 916 916 (attempted 1002)
Total pages reclaimed: 122076 49800
PPC64 vanilla compaction
Final page count: 91 94 (attempted 110)
Total pages reclaimed: 80252 96299
There was not a dramatic improvement in success rates but it wouldn't be
expected in this case either. What was important is that significantly fewer
pages were reclaimed in all cases reducing the amount of IO required to
satisfy a huge page allocation.
The second tests were all performance related - kernbench, netperf, iozone
and sysbench. None showed anything too remarkable.
The last test was a high-order allocation stress test. Many kernel compiles
are started to fill memory with a pressured mix of kernel and movable
allocations. During this, an attempt is made to allocate 90% of memory
as huge pages - one at a time with small delays between attempts to avoid
flooding the IO queue.
vanilla compaction
Percentage of request allocated X86 98 99
Percentage of request allocated X86-64 93 99
Percentage of request allocated PPC64 59 76
Success rates are a little higher, particularly on PPC64 with the larger
huge pages. What is most interesting is the latency when allocating huge
pages.
X86: http://www.csn.ul.ie/~mel/postings/compaction-20100312/highalloc-interlatency-arnold-compaction-stress-v4r3-mean.ps
X86_64: http://www.csn.ul.ie/~mel/postings/compaction-20100312/highalloc-interlatency-hydra-compaction-stress-v4r3-mean.ps
PPC64: http://www.csn.ul.ie/~mel/postings/compaction-20100312/highalloc-interlatency-powyah-compaction-stress-v4r3-mean.ps
X86 latency is reduced the least but it is depending heavily on the HIGHMEM
zone to allocate many of its huge pages which is a relatively
straight-forward job. X86-64 and PPC64 both show very significant reductions
in average time taken to allocate huge pages. It is not reduced to zero
because the system is under enough memory pressure that reclaim is still
required for some of the allocations.
What is also enlightening in the same directory is the "stddev" files. Each
of them show that the variance between allocation times is drastically
reduced.
Andrew, assuming no major complaints, how do you feel about picking these up?
Documentation/ABI/stable/sysfs-devices-node | 7 +
Documentation/ABI/testing/sysfs-devices-node | 7 +
Documentation/filesystems/proc.txt | 68 +++-
Documentation/sysctl/vm.txt | 11 +
drivers/base/node.c | 3 +
include/linux/compaction.h | 76 ++++
include/linux/mm.h | 1 +
include/linux/mmzone.h | 7 +
include/linux/rmap.h | 27 +-
include/linux/swap.h | 6 +
include/linux/vmstat.h | 2 +
kernel/sysctl.c | 11 +
mm/Kconfig | 20 +-
mm/Makefile | 1 +
mm/compaction.c | 555 ++++++++++++++++++++++++++
mm/ksm.c | 4 +-
mm/migrate.c | 22 +
mm/page_alloc.c | 68 ++++
mm/rmap.c | 10 +-
mm/vmscan.c | 5 -
mm/vmstat.c | 217 ++++++++++
21 files changed, 1101 insertions(+), 27 deletions(-)
create mode 100644 Documentation/ABI/stable/sysfs-devices-node
create mode 100644 Documentation/ABI/testing/sysfs-devices-node
create mode 100644 include/linux/compaction.h
create mode 100644 mm/compaction.c
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxxx For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>