Replace the memory controller's custom hierarchical stats code with the
generic rstat infrastructure provided by the cgroup core.
The current implementation does batched upward propagation from the
write side (i.e. as stats change). The per-cpu batches introduce an
error, which is multiplied by the number of subgroups in a tree. In
systems with many CPUs and sizable cgroup trees, the error can be large
enough to confuse users (e.g. 32 batch pages * 32 CPUs * 32 subgroups
results in an error of up to 128M per stat item). This can entirely
swallow allocation bursts inside a workload that the user is expecting
to see reflected in the statistics.
In the past, we've done read-side aggregation, where a memory.stat read
would have to walk the entire subtree and add up per-cpu counts. This
became problematic with lazily-freed cgroups: we could have large
subtrees where most cgroups were entirely idle. Hence the switch to
change-driven upward propagation. Unfortunately, it needed to trade
accuracy for speed due to the write side being so hot.
Rstat combines the best of both worlds: from the write side, it cheaply
maintains a queue of cgroups that have pending changes, so that the read
side can do selective tree aggregation. This way the reported stats
will always be precise and recent as can be, while the aggregation can
skip over potentially large numbers of idle cgroups.
The way rstat works is that it implements a tree for tracking cgroups
with pending local changes, as well as a flush function that walks the
tree upwards. The controller then drives this by 1) telling rstat when
a local cgroup stat changes (e.g. mod_memcg_state) and 2) when a flush
is required to get uptodate hierarchy stats for a given subtree (e.g.
when memory.stat is read). The controller also provides a flush
callback that is called during the rstat flush walk for each cgroup and
aggregates its local per-cpu counters and propagates them upwards.
This adds a second vmstats to struct mem_cgroup (MEMCG_NR_STAT +
NR_VM_EVENT_ITEMS) to track pending subtree deltas during upward
aggregation. It removes 3 words from the per-cpu data. It eliminates
memcg_exact_page_state(), since memcg_page_state() is now exact.
[akpm@linux-foundation.org: merge fix]
[hannes@cmpxchg.org: fix a sleep in atomic section problem]
Link: https://lkml.kernel.org/r/20210315234100.64307-1-hannes@cmpxchg.org
Link: https://lkml.kernel.org/r/20210209163304.77088-7-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Roman Gushchin <guro@fb.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Acked-by: Balbir Singh <bsingharora@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Current users of the rstat code can source root-level statistics from
the native counters of their respective subsystem, allowing them to
forego aggregation at the root level. This optimization is currently
implemented inside the generic rstat code, which doesn't track the root
cgroup and doesn't invoke the subsystem flush callbacks on it.
However, the memory controller cannot do this optimization, because
cgroup1 breaks out memory specifically for the local level, including at
the root level. In preparation for the memory controller switching to
rstat, move the optimization from rstat core to the controllers.
Afterwards, rstat will always track the root cgroup for changes and
invoke the subsystem callbacks on it; and it's up to the subsystem to
special-case and skip aggregation of the root cgroup if it can source
this information through other, cheaper means.
This is the case for the io controller and the cgroup base stats. In
their respective flush callbacks, check whether the parent is the root
cgroup, and if so, skip the unnecessary upward propagation.
The extra cost of tracking the root cgroup is negligible: on stat
changes, we actually remove a branch that checks for the root. The
queueing for a flush touches only per-cpu data, and only the first stat
change since a flush requires a (per-cpu) lock.
Link: https://lkml.kernel.org/r/20210209163304.77088-6-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Tejun Heo <tj@kernel.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Michal Koutný <mkoutny@suse.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Shakeel Butt <shakeelb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Rstat currently only supports the default hierarchy in cgroup2. In
order to replace memcg's private stats infrastructure - used in both
cgroup1 and cgroup2 - with rstat, the latter needs to support cgroup1.
The initialization and destruction callbacks for regular cgroups are
already in place. Remove the cgroup_on_dfl() guards to handle cgroup1.
The initialization of the root cgroup is currently hardcoded to only
handle cgrp_dfl_root.cgrp. Move those callbacks to cgroup_setup_root()
and cgroup_destroy_root() to handle the default root as well as the
various cgroup1 roots we may set up during mounting.
The linking of css to cgroups happens in code shared between cgroup1 and
cgroup2 as well. Simply remove the cgroup_on_dfl() guard.
Linkage of the root css to the root cgroup is a bit trickier: per
default, the root css of a subsystem controller belongs to the default
hierarchy (i.e. the cgroup2 root). When a controller is mounted in its
cgroup1 version, the root css is stolen and moved to the cgroup1 root;
on unmount, the css moves back to the default hierarchy. Annotate
rebind_subsystems() to move the root css linkage along between roots.
Link: https://lkml.kernel.org/r/20210209163304.77088-5-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Roman Gushchin <guro@fb.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Patch series "mm: memcontrol: switch to rstat", v3.
This series converts memcg stats tracking to the streamlined rstat
infrastructure provided by the cgroup core code. rstat is already used by
the CPU controller and the IO controller. This change is motivated by
recent accuracy problems in memcg's custom stats code, as well as the
benefits of sharing common infra with other controllers.
The current memcg implementation does batched tree aggregation on the
write side: local stat changes are cached in per-cpu counters, which are
then propagated upward in batches when a threshold (32 pages) is exceeded.
This is cheap, but the error introduced by the lazy upward propagation
adds up: 32 pages times CPUs times cgroups in the subtree. We've had
complaints from service owners that the stats do not reliably track and
react to allocation behavior as expected, sometimes swallowing the results
of entire test applications.
The original memcg stat implementation used to do tree aggregation
exclusively on the read side: local stats would only ever be tracked in
per-cpu counters, and a memory.stat read would iterate the entire subtree
and sum those counters up. This didn't keep up with the times:
- Cgroup trees are much bigger now. We switched to lazily-freed
cgroups, where deleted groups would hang around until their remaining
page cache has been reclaimed. This can result in large subtrees that
are expensive to walk, while most of the groups are idle and their
statistics don't change much anymore.
- Automated monitoring increased. With the proliferation of userspace
oom killing, proactive reclaim, and higher-resolution logging of
workload trends in general, top-level stat files are polled at least
once a second in many deployments.
- The lifetime of cgroups got shorter. Where most cgroup setups in the
past would have a few large policy-oriented cgroups for everything
running on the system, newer cgroup deployments tend to create one
group per application - which gets deleted again as the processes
exit. An aggregation scheme that doesn't retain child data inside the
parents loses event history of the subtree.
Rstat addresses all three of those concerns through intelligent,
persistent read-side aggregation. As statistics change at the local
level, rstat tracks - on a per-cpu basis - only those parts of a subtree
that have changes pending and require aggregation. The actual
aggregation occurs on the colder read side - which can now skip over
(potentially large) numbers of recently idle cgroups.
===
The test_kmem cgroup selftest is currently failing due to excessive
cumulative vmstat drift from 100 subgroups:
ok 1 test_kmem_basic
memory.current = 8810496
slab + anon + file + kernel_stack = 17074568
slab = 6101384
anon = 946176
file = 0
kernel_stack = 10027008
not ok 2 test_kmem_memcg_deletion
ok 3 test_kmem_proc_kpagecgroup
ok 4 test_kmem_kernel_stacks
ok 5 test_kmem_dead_cgroups
ok 6 test_percpu_basic
As you can see, memory.stat items far exceed memory.current. The kernel
stack alone is bigger than all of charged memory. That's because the
memory of the test has been uncharged from memory.current, but the
negative vmstat deltas are still sitting in the percpu caches.
The test at this time isn't even counting percpu, pagetables etc. yet,
which would further contribute to the error. The last patch in the series
updates the test to include them - as well as reduces the vmstat
tolerances in general to only expect page_counter batching.
With all patches applied, the (now more stringent) test succeeds:
ok 1 test_kmem_basic
ok 2 test_kmem_memcg_deletion
ok 3 test_kmem_proc_kpagecgroup
ok 4 test_kmem_kernel_stacks
ok 5 test_kmem_dead_cgroups
ok 6 test_percpu_basic
===
A kernel build test confirms that overhead is comparable. Two kernels are
built simultaneously in a nested tree with several idle siblings:
root - kernelbuild - one - two - three - four - build-a (defconfig, make -j16)
`- build-b (defconfig, make -j16)
`- idle-1
`- ...
`- idle-9
During the builds, kernelbuild/memory.stat is read once a second.
A perf diff shows that the changes in cycle distribution is
minimal. Top 10 kernel symbols:
0.09% +0.08% [kernel.kallsyms] [k] __mod_memcg_lruvec_state
0.00% +0.06% [kernel.kallsyms] [k] cgroup_rstat_updated
0.08% -0.05% [kernel.kallsyms] [k] __mod_memcg_state.part.0
0.16% -0.04% [kernel.kallsyms] [k] release_pages
0.00% +0.03% [kernel.kallsyms] [k] __count_memcg_events
0.01% +0.03% [kernel.kallsyms] [k] mem_cgroup_charge_statistics.constprop.0
0.10% -0.02% [kernel.kallsyms] [k] get_mem_cgroup_from_mm
0.05% -0.02% [kernel.kallsyms] [k] mem_cgroup_update_lru_size
0.57% +0.01% [kernel.kallsyms] [k] asm_exc_page_fault
===
The on-demand aggregated stats are now fully accurate:
$ grep -e nr_inactive_file /proc/vmstat | awk '{print($1,$2*4096)}'; \
grep -e inactive_file /sys/fs/cgroup/memory.stat
vanilla: patched:
nr_inactive_file 1574105088 nr_inactive_file 1027801088
inactive_file 1577410560 inactive_file 1027801088
===
This patch (of 8):
The memcg hotunplug callback erroneously flushes counts on the local CPU,
not the counts of the CPU going away; those counts will be lost.
Flush the CPU that is actually going away.
Also simplify the code a bit by using mod_memcg_state() and
count_memcg_events() instead of open-coding the upward flush - this is
comparable to how vmstat.c handles hotunplug flushing.
Link: https://lkml.kernel.org/r/20210209163304.77088-1-hannes@cmpxchg.org
Link: https://lkml.kernel.org/r/20210209163304.77088-2-hannes@cmpxchg.org
Fixes: a983b5ebee ("mm: memcontrol: fix excessive complexity in memory.stat reporting")
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Roman Gushchin <guro@fb.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
In the era of async memcg oom-killer, the commit a0d8b00a33 ("mm: memcg:
do not declare OOM from __GFP_NOFAIL allocations") added the code to skip
memcg oom-killer for __GFP_NOFAIL allocations. The reason was that the
__GFP_NOFAIL callers will not enter aync oom synchronization path and will
keep the task marked as in memcg oom. At that time the tasks marked in
memcg oom can bypass the memcg limits and the oom synchronization would
have happened later in the later userspace triggered page fault. Thus
letting the task marked as under memcg oom bypass the memcg limit for
arbitrary time.
With the synchronous memcg oom-killer (commit 29ef680ae7 ("memcg, oom:
move out_of_memory back to the charge path")) and not letting the task
marked under memcg oom to bypass the memcg limits (commit 1f14c1ac19
("mm: memcg: do not allow task about to OOM kill to bypass the limit")),
we can again allow __GFP_NOFAIL allocations to trigger memcg oom-kill.
This will make memcg oom behavior closer to page allocator oom behavior.
Link: https://lkml.kernel.org/r/20210223204337.2785120-1-shakeelb@google.com
Signed-off-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
For simplification commit 991e767385 ("mm: memcontrol: account kernel
stack per node") changed the per zone vmalloc backed stack pages
accounting to per node.
By doing that we have lost a certain precision because those pages might
live in different NUMA nodes. In the end NR_KERNEL_STACK_KB exported to
the userspace might be over estimated on some nodes while underestimated
on others. But this is not a real world problem, just a problem found
by reading the code. So there is no actual data to showing how much
impact it has on users.
This doesn't impose any real problem to correctnes of the kernel
behavior as the counter is not used for any internal processing but it
can cause some confusion to the userspace.
Address the problem by accounting each vmalloc backing page to its own
node.
Link: https://lkml.kernel.org/r/20210303151843.81156-1-songmuchun@bytedance.com
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Roman Gushchin <guro@fb.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Use the newly added unpin_user_page_range_dirty_lock() for more quickly
unpinning a consecutive range of pages represented as compound pages.
This will also calculate number of pages to unpin (for the tail pages
which matching head page) and thus batch the refcount update.
Running a test program which calls memory range reg/unreg on a region 1G
in size and measures cost of both operations together (in a guest using
rxe) with THP and hugetlbfs:
Before:
590 rounds in 5.003 sec: 8480.335 usec / round
6898 rounds in 60.001 sec: 8698.367 usec / round
After:
2688 rounds in 5.002 sec: 1860.786 usec / round
32517 rounds in 60.001 sec: 1845.225 usec / round
Link: https://lkml.kernel.org/r/20210212130843.13865-5-joao.m.martins@oracle.com
Signed-off-by: Joao Martins <joao.m.martins@oracle.com>
Acked-by: Jason Gunthorpe <jgg@nvidia.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Doug Ledford <dledford@redhat.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Matthew Wilcox <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Rather than decrementing the head page refcount one by one, we walk the
page array and checking which belong to the same compound_head. Later on
we decrement the calculated amount of references in a single write to the
head page. To that end switch to for_each_compound_head() does most of
the work.
set_page_dirty() needs no adjustment as it's a nop for non-dirty head
pages and it doesn't operate on tail pages.
This considerably improves unpinning of pages with THP and hugetlbfs:
- THP
gup_test -t -m 16384 -r 10 [-L|-a] -S -n 512 -w
PIN_LONGTERM_BENCHMARK (put values): ~87.6k us -> ~23.2k us
- 16G with 1G huge page size
gup_test -f /mnt/huge/file -m 16384 -r 10 [-L|-a] -S -n 512 -w
PIN_LONGTERM_BENCHMARK: (put values): ~87.6k us -> ~27.5k us
Link: https://lkml.kernel.org/r/20210212130843.13865-3-joao.m.martins@oracle.com
Signed-off-by: Joao Martins <joao.m.martins@oracle.com>
Reviewed-by: John Hubbard <jhubbard@nvidia.com>
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Doug Ledford <dledford@redhat.com>
Cc: Matthew Wilcox <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Patch series "mm/gup: page unpining improvements", v4.
This series improves page unpinning, with an eye on improving MR
deregistration for big swaths of memory (which is bound by the page
unpining), particularly:
1) Decrement the head page by @ntails and thus reducing a lot the
number of atomic operations per compound page. This is done by
comparing individual tail pages heads, and counting number of
consecutive tails on which they match heads and based on that update
head page refcount. Should have a visible improvement in all page
(un)pinners which use compound pages
2) Introducing a new API for unpinning page ranges (to avoid the trick
in the previous item and be based on math), and use that in RDMA
ib_mem_release (used for mr deregistration).
Performance improvements: unpin_user_pages() for hugetlbfs and THP
improves ~3x (through gup_test) and RDMA MR dereg improves ~4.5x with the
new API. See patches 2 and 4 for those.
This patch (of 4):
Add a helper that iterates over head pages in a list of pages. It
essentially counts the tails until the next page to process has a
different head that the current. This is going to be used by
unpin_user_pages() family of functions, to batch the head page refcount
updates once for all passed consecutive tail pages.
Link: https://lkml.kernel.org/r/20210212130843.13865-1-joao.m.martins@oracle.com
Link: https://lkml.kernel.org/r/20210212130843.13865-2-joao.m.martins@oracle.com
Signed-off-by: Joao Martins <joao.m.martins@oracle.com>
Suggested-by: Jason Gunthorpe <jgg@nvidia.com>
Reviewed-by: John Hubbard <jhubbard@nvidia.com>
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Cc: Doug Ledford <dledford@redhat.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Page writeback doesn't hold a page reference, which allows truncate to
free a page the second PageWriteback is cleared. This used to require
special attention in test_clear_page_writeback(), where we had to be
careful not to rely on the unstable page->memcg binding and look up all
the necessary information before clearing the writeback flag.
Since commit 073861ed77 ("mm: fix VM_BUG_ON(PageTail) and
BUG_ON(PageWriteback)") test_clear_page_writeback() is called with an
explicit reference on the page, and this dance is no longer needed.
Use unlock_page_memcg() and dec_lruvec_page_state() directly.
This removes the last user of the lock_page_memcg() return value, change
it to void. Touch up the comments in there as well. This also removes
the last extern user of __unlock_page_memcg(), make it static. Further,
it removes the last user of dec_lruvec_state(), delete it, along with a
few other unused helpers.
Link: https://lkml.kernel.org/r/YCQbYAWg4nvBFL6h@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Roman Gushchin <guro@fb.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
If the I/O completed successfully, the page will remain Uptodate, even
if it is subsequently truncated. If the I/O completed with an error,
this check would cause us to retry the I/O if the page were truncated
before we woke up. There is no need to retry the I/O; the I/O to fill
the page failed, so we can legitimately just return -EIO.
This code was originally added by commit 56f0d5fe6851 ("[PATCH]
readpage-vs-invalidate fix") in 2005 (this commit ID is from the
linux-fullhistory tree; it is also commit ba1f08f14b52 in tglx-history).
At the time, truncate_complete_page() called ClearPageUptodate(), and so
this was fixing a real bug. In 2008, commit 84209e02de ("mm: dont clear
PG_uptodate on truncate/invalidate") removed the call to
ClearPageUptodate, and this check has been unnecessary ever since.
It doesn't do any real harm, but there's no need to keep it.
Link: https://lkml.kernel.org/r/20210303222547.1056428-1-willy@infradead.org
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Acked-by: William Kucharski <william.kucharski@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
After splitting generic_file_buffered_read() into smaller parts, it turns
out we can reuse one of the parts in filemap_fault(). This fixes an
oversight -- waiting for the I/O to complete is now interruptible by a
fatal signal. And it saves us a few bytes of text in an unlikely path.
$ ./scripts/bloat-o-meter before.o after.o
add/remove: 0/0 grow/shrink: 0/1 up/down: 0/-207 (-207)
Function old new delta
filemap_fault 2187 1980 -207
Total: Before=37491, After=37284, chg -0.55%
Link: https://lkml.kernel.org/r/20210226140011.2883498-1-willy@infradead.org
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Kent Overstreet <kent.overstreet@gmail.com>
Cc: Christoph Hellwig <hch@lst.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Johannes Weiner
Shakeel Butt
Muchun Song
Zhiyuan Dai
Yang Shi
Joao Martins
Nikita Ermakov
Rui Sun
Matthew Wilcox (Oracle)
Jens Axboe