Several people report seeing warnings about inconsistent radix tree
nodes followed by crashes in the workingset code, which all looked like
use-after-free access from the shadow node shrinker.
Dave Jones managed to reproduce the issue with a debug patch applied,
which confirmed that the radix tree shrinking indeed frees shadow nodes
while they are still linked to the shadow LRU:
WARNING: CPU: 2 PID: 53 at lib/radix-tree.c:643 delete_node+0x1e4/0x200
CPU: 2 PID: 53 Comm: kswapd0 Not tainted 4.10.0-rc2-think+ #3
Call Trace:
delete_node+0x1e4/0x200
__radix_tree_delete_node+0xd/0x10
shadow_lru_isolate+0xe6/0x220
__list_lru_walk_one.isra.4+0x9b/0x190
list_lru_walk_one+0x23/0x30
scan_shadow_nodes+0x2e/0x40
shrink_slab.part.44+0x23d/0x5d0
shrink_node+0x22c/0x330
kswapd+0x392/0x8f0
This is the WARN_ON_ONCE(!list_empty(&node->private_list)) placed in the
inlined radix_tree_shrink().
The problem is with 14b468791f ("mm: workingset: move shadow entry
tracking to radix tree exceptional tracking"), which passes an update
callback into the radix tree to link and unlink shadow leaf nodes when
tree entries change, but forgot to pass the callback when reclaiming a
shadow node.
While the reclaimed shadow node itself is unlinked by the shrinker, its
deletion from the tree can cause the left-most leaf node in the tree to
be shrunk. If that happens to be a shadow node as well, we don't unlink
it from the LRU as we should.
Consider this tree, where the s are shadow entries:
root->rnode
|
[0 n]
| |
[s ] [sssss]
Now the shadow node shrinker reclaims the rightmost leaf node through
the shadow node LRU:
root->rnode
|
[0 ]
|
[s ]
Because the parent of the deleted node is the first level below the
root and has only one child in the left-most slot, the intermediate
level is shrunk and the node containing the single shadow is put in
its place:
root->rnode
|
[s ]
The shrinker again sees a single left-most slot in a first level node
and thus decides to store the shadow in root->rnode directly and free
the node - which is a leaf node on the shadow node LRU.
root->rnode
|
s
Without the update callback, the freed node remains on the shadow LRU,
where it causes later shrinker runs to crash.
Pass the node updater callback into __radix_tree_delete_node() in case
the deletion causes the left-most branch in the tree to collapse too.
Also add warnings when linked nodes are freed right away, rather than
wait for the use-after-free when the list is scanned much later.
Fixes: 14b468791f ("mm: workingset: move shadow entry tracking to radix tree exceptional tracking")
Reported-by: Dave Chinner <[email protected]>
Reported-by: Hugh Dickins <[email protected]>
Reported-by: Andrea Arcangeli <[email protected]>
Reported-and-tested-by: Dave Jones <[email protected]>
Signed-off-by: Johannes Weiner <[email protected]>
Cc: Christoph Hellwig <[email protected]>
Cc: Chris Leech <[email protected]>
Cc: Lee Duncan <[email protected]>
Cc: Jan Kara <[email protected]>
Cc: Kirill A. Shutemov <[email protected]>
Cc: Matthew Wilcox <[email protected]>
Cc: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
When FADV_DONTNEED cannot drop all pages in the range, it observes that
some pages might still be on per-cpu LRU caches after recent
instantiation and so initiates remote calls to all CPUs to flush their
local caches. However, in most cases, the fadvise happens from the same
context that instantiated the pages, and any pre-LRU pages in the
specified range are most likely sitting on the local CPU's LRU cache,
and so in many cases this results in unnecessary remote calls, which, in
a loaded system, can hold up the fadvise() call significantly.
[ I didn't record it in the extreme case we observed at Facebook,
unfortunately. We had a slow-to-respond system and noticed it
lru_add_drain_all() leading the profile during fadvise calls. This
patch came out of thinking about the code and how we commonly call
FADV_DONTNEED.
FWIW, I wrote a silly directory tree walker/searcher that recurses
through /usr to read and FADV_DONTNEED each file it finds. On a 2
socket 40 ht machine, over 1% is spent in lru_add_drain_all(). With
the patch, that cost is gone; the local drain cost shows at 0.09%. ]
Try to avoid the remote call by flushing the local LRU cache before even
attempting to invalidate anything. It's a cheap operation, and the
local LRU cache is the most likely to hold any pre-LRU pages in the
specified fadvise range.
Link: http://lkml.kernel.org/r/[email protected]
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Vlastimil Babka <[email protected]>
Acked-by: Mel Gorman <[email protected]>
Acked-by: Hillf Danton <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Since commit 59dc76b0d4 ("mm: vmscan: reduce size of inactive file
list") the size of the active file list is no longer limited to half of
memory. Increase the shadow node limit accordingly to avoid throwing
out shadow entries that might still result in eligible refaults.
The exact size of the active list now depends on the overall size of the
page cache, but converges toward taking up most of the space:
In mm/vmscan.c::inactive_list_is_low(),
* total target max
* memory ratio inactive
* -------------------------------------
* 10MB 1 5MB
* 100MB 1 50MB
* 1GB 3 250MB
* 10GB 10 0.9GB
* 100GB 31 3GB
* 1TB 101 10GB
* 10TB 320 32GB
It would be possible to apply the same precise ratios when determining
the limit for radix tree nodes containing shadow entries, but since it
is merely an approximation of the oldest refault distances in the wild
and the code also makes assumptions about the node population density,
keep it simple and always target the full cache size.
While at it, clarify the comment and the formula for memory footprint.
Link: http://lkml.kernel.org/r/[email protected]
Signed-off-by: Johannes Weiner <[email protected]>
Cc: Rik van Riel <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Shadow entries in the page cache used to be accounted behind the radix
tree implementation's back in the upper bits of node->count, and the
radix tree code extending a single-entry tree with a shadow entry in
root->rnode would corrupt that counter. As a result, we could not put
shadow entries at index 0 if the tree didn't have any other entries, and
that means no refault detection for any single-page file.
Now that the shadow entries are tracked natively in the radix tree's
exceptional counter, this is no longer necessary. Extending and
shrinking the tree from and to single entries in root->rnode now does
the right thing when the entry is exceptional, remove that limitation.
Link: http://lkml.kernel.org/r/[email protected]
Signed-off-by: Johannes Weiner <[email protected]>
Reviewed-by: Jan Kara <[email protected]>
Cc: Kirill A. Shutemov <[email protected]>
Cc: Hugh Dickins <[email protected]>
Cc: Matthew Wilcox <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Currently, we track the shadow entries in the page cache in the upper
bits of the radix_tree_node->count, behind the back of the radix tree
implementation. Because the radix tree code has no awareness of them,
we rely on random subtleties throughout the implementation (such as the
node->count != 1 check in the shrinking code, which is meant to exclude
multi-entry nodes but also happens to skip nodes with only one shadow
entry, as that's accounted in the upper bits). This is error prone and
has, in fact, caused the bug fixed in d3798ae8c6 ("mm: filemap: don't
plant shadow entries without radix tree node").
To remove these subtleties, this patch moves shadow entry tracking from
the upper bits of node->count to the existing counter for exceptional
entries. node->count goes back to being a simple counter of valid
entries in the tree node and can be shrunk to a single byte.
This vastly simplifies the page cache code. All accounting happens
natively inside the radix tree implementation, and maintaining the LRU
linkage of shadow nodes is consolidated into a single function in the
workingset code that is called for leaf nodes affected by a change in
the page cache tree.
This also removes the last user of the __radix_delete_node() return
value. Eliminate it.
Link: http://lkml.kernel.org/r/[email protected]
Signed-off-by: Johannes Weiner <[email protected]>
Reviewed-by: Jan Kara <[email protected]>
Cc: Kirill A. Shutemov <[email protected]>
Cc: Hugh Dickins <[email protected]>
Cc: Matthew Wilcox <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Support handing __radix_tree_replace() a callback that gets invoked for
all leaf nodes that change or get freed as a result of the slot
replacement, to assist users tracking nodes with node->private_list.
This prepares for putting page cache shadow entries into the radix tree
root again and drastically simplifying the shadow tracking.
Link: http://lkml.kernel.org/r/[email protected]
Signed-off-by: Johannes Weiner <[email protected]>
Suggested-by: Jan Kara <[email protected]>
Reviewed-by: Jan Kara <[email protected]>
Cc: Kirill A. Shutemov <[email protected]>
Cc: Hugh Dickins <[email protected]>
Cc: Matthew Wilcox <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Page cache shadow entry handling will be a lot simpler when it can use a
single generic replacement function for pages, shadow entries, and
emptying slots.
Make __radix_tree_replace() properly account insertions and deletions in
node->count and garbage collect nodes as they become empty. Then
re-implement radix_tree_delete() on top of it.
Link: http://lkml.kernel.org/r/[email protected]
Signed-off-by: Johannes Weiner <[email protected]>
Reviewed-by: Jan Kara <[email protected]>
Cc: Kirill A. Shutemov <[email protected]>
Cc: Hugh Dickins <[email protected]>
Cc: Matthew Wilcox <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The bug in khugepaged fixed earlier in this series shows that radix tree
slot replacement is fragile; and it will become more so when not only
NULL<->!NULL transitions need to be caught but transitions from and to
exceptional entries as well. We need checks.
Re-implement radix_tree_replace_slot() on top of the sanity-checked
__radix_tree_replace(). This requires existing callers to also pass the
radix tree root, but it'll warn us when somebody replaces slots with
contents that need proper accounting (transitions between NULL entries,
real entries, exceptional entries) and where a replacement through the
slot pointer would corrupt the radix tree node counts.
Link: http://lkml.kernel.org/r/[email protected]
Signed-off-by: Johannes Weiner <[email protected]>
Suggested-by: Jan Kara <[email protected]>
Reviewed-by: Jan Kara <[email protected]>
Cc: Kirill A. Shutemov <[email protected]>
Cc: Hugh Dickins <[email protected]>
Cc: Matthew Wilcox <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The way the page cache is sneaking shadow entries of evicted pages into
the radix tree past the node entry accounting and tracking them manually
in the upper bits of node->count is fraught with problems.
These shadow entries are marked in the tree as exceptional entries,
which are a native concept to the radix tree. Maintain an explicit
counter of exceptional entries in the radix tree node. Subsequent
patches will switch shadow entry tracking over to that counter.
DAX and shmem are the other users of exceptional entries. Since slot
replacements that change the entry type from regular to exceptional must
now be accounted, introduce a __radix_tree_replace() function that does
replacement and accounting, and switch DAX and shmem over.
The increase in radix tree node size is temporary. A followup patch
switches the shadow tracking to this new scheme and we'll no longer need
the upper bits in node->count and shrink that back to one byte.
Link: http://lkml.kernel.org/r/[email protected]
Signed-off-by: Johannes Weiner <[email protected]>
Reviewed-by: Jan Kara <[email protected]>
Cc: Kirill A. Shutemov <[email protected]>
Cc: Hugh Dickins <[email protected]>
Cc: Matthew Wilcox <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
When the shadow page shrinker tries to reclaim a radix tree node but
finds it in an unexpected state - it should contain no pages, and
non-zero shadow entries - there is no need to kill the executing task or
even the entire system. Warn about the invalid state, then leave that
tree node be. Simply don't put it back on the shadow LRU for future
reclaim and move on.
Link: http://lkml.kernel.org/r/[email protected]
Signed-off-by: Johannes Weiner <[email protected]>
Reviewed-by: Jan Kara <[email protected]>
Cc: Kirill A. Shutemov <[email protected]>
Cc: Hugh Dickins <[email protected]>
Cc: Matthew Wilcox <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The radix tree counts valid entries in each tree node. Entries stored
in the tree cannot be removed by simpling storing NULL in the slot or
the internal counters will be off and the node never gets freed again.
When collapsing a shmem page fails, restore the holes that were filled
with radix_tree_insert() with a proper radix tree deletion.
Fixes: f3f0e1d215 ("khugepaged: add support of collapse for tmpfs/shmem pages")
Link: http://lkml.kernel.org/r/[email protected]
Signed-off-by: Johannes Weiner <[email protected]>
Reported-by: Jan Kara <[email protected]>
Acked-by: Kirill A. Shutemov <[email protected]>
Reviewed-by: Jan Kara <[email protected]>
Cc: Hugh Dickins <[email protected]>
Cc: Matthew Wilcox <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Patch series "mm: workingset: radix tree subtleties & single-page file
refaults", v3.
This is another revision of the radix tree / workingset patches based on
feedback from Jan and Kirill.
This is a follow-up to d3798ae8c6 ("mm: filemap: don't plant shadow
entries without radix tree node"). That patch fixed an issue that was
caused mainly by the page cache sneaking special shadow page entries
into the radix tree and relying on subtleties in the radix tree code to
make that work. The fix also had to stop tracking refaults for
single-page files because shadow pages stored as direct pointers in
radix_tree_root->rnode weren't properly handled during tree extension.
These patches make the radix tree code explicitely support and track
such special entries, to eliminate the subtleties and to restore the
thrash detection for single-page files.
This patch (of 9):
When a radix tree iteration drops the tree lock, another thread might
swoop in and free the node holding the current slot. The iteration
needs to do another tree lookup from the current index to continue.
[[email protected]: re-lookup for replacement]
Fixes: f3f0e1d215 ("khugepaged: add support of collapse for tmpfs/shmem pages")
Link: http://lkml.kernel.org/r/[email protected]
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Kirill A. Shutemov <[email protected]>
Reviewed-by: Jan Kara <[email protected]>
Cc: Hugh Dickins <[email protected]>
Cc: Matthew Wilcox <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
On 4.0, we saw a stack corruption from a page fault entering direct
memory cgroup reclaim, calling into btrfs_releasepage(), which then
tried to allocate an extent and recursed back into a kmem charge ad
nauseam:
[...]
btrfs_releasepage+0x2c/0x30
try_to_release_page+0x32/0x50
shrink_page_list+0x6da/0x7a0
shrink_inactive_list+0x1e5/0x510
shrink_lruvec+0x605/0x7f0
shrink_zone+0xee/0x320
do_try_to_free_pages+0x174/0x440
try_to_free_mem_cgroup_pages+0xa7/0x130
try_charge+0x17b/0x830
memcg_charge_kmem+0x40/0x80
new_slab+0x2d9/0x5a0
__slab_alloc+0x2fd/0x44f
kmem_cache_alloc+0x193/0x1e0
alloc_extent_state+0x21/0xc0
__clear_extent_bit+0x2b5/0x400
try_release_extent_mapping+0x1a3/0x220
__btrfs_releasepage+0x31/0x70
btrfs_releasepage+0x2c/0x30
try_to_release_page+0x32/0x50
shrink_page_list+0x6da/0x7a0
shrink_inactive_list+0x1e5/0x510
shrink_lruvec+0x605/0x7f0
shrink_zone+0xee/0x320
do_try_to_free_pages+0x174/0x440
try_to_free_mem_cgroup_pages+0xa7/0x130
try_charge+0x17b/0x830
mem_cgroup_try_charge+0x65/0x1c0
handle_mm_fault+0x117f/0x1510
__do_page_fault+0x177/0x420
do_page_fault+0xc/0x10
page_fault+0x22/0x30
On later kernels, kmem charging is opt-in rather than opt-out, and that
particular kmem allocation in btrfs_releasepage() is no longer being
charged and won't recurse and overrun the stack anymore.
But it's not impossible for an accounted allocation to happen from the
memcg direct reclaim context, and we needed to reproduce this crash many
times before we even got a useful stack trace out of it.
Like other direct reclaimers, mark tasks in memcg reclaim PF_MEMALLOC to
avoid recursing into any other form of direct reclaim. Then let
recursive charges from PF_MEMALLOC contexts bypass the cgroup limit.
Link: http://lkml.kernel.org/r/[email protected]
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Cc: Vladimir Davydov <[email protected]>
Cc: Tejun Heo <[email protected]>
Cc: <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Commit 22f2ac51b6 ("mm: workingset: fix crash in shadow node shrinker
caused by replace_page_cache_page()") switched replace_page_cache() from
raw radix tree operations to page_cache_tree_insert() but didn't take
into account that the latter function, unlike the raw radix tree op,
handles mapping->nrpages. As a result, that counter is bumped for each
page replacement rather than balanced out even.
The mapping->nrpages counter is used to skip needless radix tree walks
when invalidating, truncating, syncing inodes without pages, as well as
statistics for userspace. Since the error is positive, we'll do more
page cache tree walks than necessary; we won't miss a necessary one.
And we'll report more buffer pages to userspace than there are. The
error is limited to fuse inodes.
Fixes: 22f2ac51b6 ("mm: workingset: fix crash in shadow node shrinker caused by replace_page_cache_page()")
Signed-off-by: Johannes Weiner <[email protected]>
Cc: Andrew Morton <[email protected]>
Cc: Miklos Szeredi <[email protected]>
Cc: [email protected]
Signed-off-by: Linus Torvalds <[email protected]>
When the underflow checks were added to workingset_node_shadow_dec(),
they triggered immediately:
kernel BUG at ./include/linux/swap.h:276!
invalid opcode: 0000 [#1] SMP
Modules linked in: isofs usb_storage fuse xt_CHECKSUM ipt_MASQUERADE nf_nat_masquerade_ipv4 tun nf_conntrack_netbios_ns nf_conntrack_broadcast ip6t_REJECT nf_reject_ipv6
soundcore wmi acpi_als pinctrl_sunrisepoint kfifo_buf tpm_tis industrialio acpi_pad pinctrl_intel tpm_tis_core tpm nfsd auth_rpcgss nfs_acl lockd grace sunrpc dm_crypt
CPU: 0 PID: 20929 Comm: blkid Not tainted 4.8.0-rc8-00087-gbe67d60ba944 #1
Hardware name: System manufacturer System Product Name/Z170-K, BIOS 1803 05/06/2016
task: ffff8faa93ecd940 task.stack: ffff8faa7f478000
RIP: page_cache_tree_insert+0xf1/0x100
Call Trace:
__add_to_page_cache_locked+0x12e/0x270
add_to_page_cache_lru+0x4e/0xe0
mpage_readpages+0x112/0x1d0
blkdev_readpages+0x1d/0x20
__do_page_cache_readahead+0x1ad/0x290
force_page_cache_readahead+0xaa/0x100
page_cache_sync_readahead+0x3f/0x50
generic_file_read_iter+0x5af/0x740
blkdev_read_iter+0x35/0x40
__vfs_read+0xe1/0x130
vfs_read+0x96/0x130
SyS_read+0x55/0xc0
entry_SYSCALL_64_fastpath+0x13/0x8f
Code: 03 00 48 8b 5d d8 65 48 33 1c 25 28 00 00 00 44 89 e8 75 19 48 83 c4 18 5b 41 5c 41 5d 41 5e 5d c3 0f 0b 41 bd ef ff ff ff eb d7 <0f> 0b e8 88 68 ef ff 0f 1f 84 00
RIP page_cache_tree_insert+0xf1/0x100
This is a long-standing bug in the way shadow entries are accounted in
the radix tree nodes. The shrinker needs to know when radix tree nodes
contain only shadow entries, no pages, so node->count is split in half
to count shadows in the upper bits and pages in the lower bits.
Unfortunately, the radix tree implementation doesn't know of this and
assumes all entries are in node->count. When there is a shadow entry
directly in root->rnode and the tree is later extended, the radix tree
implementation will copy that entry into the new node and and bump its
node->count, i.e. increases the page count bits. Once the shadow gets
removed and we subtract from the upper counter, node->count underflows
and triggers the warning. Afterwards, without node->count reaching 0
again, the radix tree node is leaked.
Limit shadow entries to when we have actual radix tree nodes and can
count them properly. That means we lose the ability to detect refaults
from files that had only the first page faulted in at eviction time.
Fixes: 449dd6984d ("mm: keep page cache radix tree nodes in check")
Signed-off-by: Johannes Weiner <[email protected]>
Reported-and-tested-by: Linus Torvalds <[email protected]>
Reviewed-by: Jan Kara <[email protected]>
Cc: Andrew Morton <[email protected]>
Cc: [email protected]
Signed-off-by: Linus Torvalds <[email protected]>
Antonio reports the following crash when using fuse under memory pressure:
kernel BUG at /build/linux-a2WvEb/linux-4.4.0/mm/workingset.c:346!
invalid opcode: 0000 [#1] SMP
Modules linked in: all of them
CPU: 2 PID: 63 Comm: kswapd0 Not tainted 4.4.0-36-generic #55-Ubuntu
Hardware name: System manufacturer System Product Name/P8H67-M PRO, BIOS 3904 04/27/2013
task: ffff88040cae6040 ti: ffff880407488000 task.ti: ffff880407488000
RIP: shadow_lru_isolate+0x181/0x190
Call Trace:
__list_lru_walk_one.isra.3+0x8f/0x130
list_lru_walk_one+0x23/0x30
scan_shadow_nodes+0x34/0x50
shrink_slab.part.40+0x1ed/0x3d0
shrink_zone+0x2ca/0x2e0
kswapd+0x51e/0x990
kthread+0xd8/0xf0
ret_from_fork+0x3f/0x70
which corresponds to the following sanity check in the shadow node
tracking:
BUG_ON(node->count & RADIX_TREE_COUNT_MASK);
The workingset code tracks radix tree nodes that exclusively contain
shadow entries of evicted pages in them, and this (somewhat obscure)
line checks whether there are real pages left that would interfere with
reclaim of the radix tree node under memory pressure.
While discussing ways how fuse might sneak pages into the radix tree
past the workingset code, Miklos pointed to replace_page_cache_page(),
and indeed there is a problem there: it properly accounts for the old
page being removed - __delete_from_page_cache() does that - but then
does a raw raw radix_tree_insert(), not accounting for the replacement
page. Eventually the page count bits in node->count underflow while
leaving the node incorrectly linked to the shadow node LRU.
To address this, make sure replace_page_cache_page() uses the tracked
page insertion code, page_cache_tree_insert(). This fixes the page
accounting and makes sure page-containing nodes are properly unlinked
from the shadow node LRU again.
Also, make the sanity checks a bit less obscure by using the helpers for
checking the number of pages and shadows in a radix tree node.
Fixes: 449dd6984d ("mm: keep page cache radix tree nodes in check")
Link: http://lkml.kernel.org/r/[email protected]
Signed-off-by: Johannes Weiner <[email protected]>
Reported-by: Antonio SJ Musumeci <[email protected]>
Debugged-by: Miklos Szeredi <[email protected]>
Cc: <[email protected]> [3.15+]
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Commit 23047a96d7 ("mm: workingset: per-cgroup cache thrash
detection") added a page->mem_cgroup lookup to the cache eviction,
refault, and activation paths, as well as locking to the activation
path, and the vm-scalability tests showed a regression of -23%.
While the test in question is an artificial worst-case scenario that
doesn't occur in real workloads - reading two sparse files in parallel
at full CPU speed just to hammer the LRU paths - there is still some
optimizations that can be done in those paths.
Inline the lookup functions to eliminate calls. Also, page->mem_cgroup
doesn't need to be stabilized when counting an activation; we merely
need to hold the RCU lock to prevent the memcg from being freed.
This cuts down on overhead quite a bit:
23047a96d7 063f6715e77a7be5770d6081fe
---------------- --------------------------
%stddev %change %stddev
\ | \
21621405 +- 0% +11.3% 24069657 +- 2% vm-scalability.throughput
[[email protected]: drop unnecessary include file]
[[email protected]: add WARN_ON_ONCE()s]
Link: http://lkml.kernel.org/r/[email protected]
Link: http://lkml.kernel.org/r/[email protected]
Reported-by: Ye Xiaolong <[email protected]>
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Cc: Vladimir Davydov <[email protected]>
Signed-off-by: Guenter Roeck <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The valid cgroup hierarchy ID range includes 0, so we can't filter for
positive numbers when freeing it, or it'll leak the first ID. No big
deal, just disruptive when reading the code.
The ID is freed during error handling and when the reference count hits
zero, so the double-free test is not necessary; remove it.
Link: http://lkml.kernel.org/r/[email protected]
Signed-off-by: Johannes Weiner <[email protected]>
Cc: Vladimir Davydov <[email protected]>
Cc: Tejun Heo <[email protected]>
Cc: Nikolay Borisov <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The memory controller has quite a bit of state that usually outlives the
cgroup and pins its CSS until said state disappears. At the same time
it imposes a 16-bit limit on the CSS ID space to economically store IDs
in the wild. Consequently, when we use cgroups to contain frequent but
small and short-lived jobs that leave behind some page cache, we quickly
run into the 64k limitations of outstanding CSSs. Creating a new cgroup
fails with -ENOSPC while there are only a few, or even no user-visible
cgroups in existence.
Although pinning CSSs past cgroup removal is common, there are only two
instances that actually need an ID after a cgroup is deleted: cache
shadow entries and swapout records.
Cache shadow entries reference the ID weakly and can deal with the CSS
having disappeared when it's looked up later. They pose no hurdle.
Swap-out records do need to pin the css to hierarchically attribute
swapins after the cgroup has been deleted; though the only pages that
remain swapped out after offlining are tmpfs/shmem pages. And those
references are under the user's control, so they are manageable.
This patch introduces a private 16-bit memcg ID and switches swap and
cache shadow entries over to using that. This ID can then be recycled
after offlining when the CSS remains pinned only by objects that don't
specifically need it.
This script demonstrates the problem by faulting one cache page in a new
cgroup and deleting it again:
set -e
mkdir -p pages
for x in `seq 128000`; do
[ $((x % 1000)) -eq 0 ] && echo $x
mkdir /cgroup/foo
echo $$ >/cgroup/foo/cgroup.procs
echo trex >pages/$x
echo $$ >/cgroup/cgroup.procs
rmdir /cgroup/foo
done
When run on an unpatched kernel, we eventually run out of possible IDs
even though there are no visible cgroups:
[root@ham ~]# ./cssidstress.sh
[...]
65000
mkdir: cannot create directory '/cgroup/foo': No space left on device
After this patch, the IDs get released upon cgroup destruction and the
cache and css objects get released once memory reclaim kicks in.
[[email protected]: init the IDR]
Link: http://lkml.kernel.org/r/[email protected]
Fixes: b2052564e6 ("mm: memcontrol: continue cache reclaim from offlined groups")
Link: http://lkml.kernel.org/r/[email protected]
Signed-off-by: Johannes Weiner <[email protected]>
Reported-by: John Garcia <[email protected]>
Reviewed-by: Vladimir Davydov <[email protected]>
Acked-by: Tejun Heo <[email protected]>
Cc: Nikolay Borisov <[email protected]>
Cc: <[email protected]> [3.19+]
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
css_idr allocation starts at 1, so index 0 will never point to an
item. css_from_id() currently filters that before asking idr_find(),
but idr_find() would also just return NULL, so this is not needed.
Signed-off-by: Johannes Weiner <[email protected]>
Signed-off-by: Tejun Heo <[email protected]>
The valid cgroup hierarchy ID range includes 0, so we can't filter for
positive numbers when freeing it, or it'll leak the first ID. No big
deal, just disruptive when reading the code.
The ID is freed during error handling and when the reference count
hits zero, so the double-free test is not necessary; remove it.
Signed-off-by: Johannes Weiner <[email protected]>
Signed-off-by: Tejun Heo <[email protected]>
Andres observed that his database workload is struggling with the
transaction journal creating pressure on frequently read pages.
Access patterns like transaction journals frequently write the same
pages over and over, but in the majority of cases those pages are never
read back. There are no caching benefits to be had for those pages, so
activating them and having them put pressure on pages that do benefit
from caching is a bad choice.
Leave page activations to read accesses and don't promote pages based on
writes alone.
It could be said that partially written pages do contain cache-worthy
data, because even if *userspace* does not access the unwritten part,
the kernel still has to read it from the filesystem for correctness.
However, a counter argument is that these pages enjoy at least *some*
protection over other inactive file pages through the writeback cache,
in the sense that dirty pages are written back with a delay and cache
reclaim leaves them alone until they have been written back to disk.
Should that turn out to be insufficient and we see increased read IO
from partial writes under memory pressure, we can always go back and
update grab_cache_page_write_begin() to take (pos, len) so that it can
tell partial writes from pages that don't need partial reads. But for
now, keep it simple.
Signed-off-by: Johannes Weiner <[email protected]>
Reported-by: Andres Freund <[email protected]>
Cc: Rik van Riel <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Cgroup2 currently doesn't have a per-cgroup swappiness setting. We
might want to add one later - that's a different discussion - but until
we do, the cgroups should always follow the system setting. Otherwise
it will be unchangeably set to whatever the ancestor inherited from the
system setting at the time of cgroup creation.
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Acked-by: Vladimir Davydov <[email protected]>
Cc: <[email protected]> [4.5]
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
uncharge_list() does an unusual list walk because the function can take
regular lists with dedicated list_heads as well as singleton lists where
a single page is passed via the page->lru list node.
This can sometimes lead to confusion as well as suggestions to replace
the loop with a list_for_each_entry(), which wouldn't work.
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Cc: Vladimir Davydov <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Setting the original memory.limit_in_bytes hardlimit is subject to a
race condition when the desired value is below the current usage. The
code tries a few times to first reclaim and then see if the usage has
dropped to where we would like it to be, but there is no locking, and
the workload is free to continue making new charges up to the old limit.
Thus, attempting to shrink a workload relies on pure luck and hope that
the workload happens to cooperate.
To fix this in the cgroup2 memory.max knob, do it the other way round:
set the limit first, then try enforcement. And if reclaim is not able
to succeed, trigger OOM kills in the group. Keep going until the new
limit is met, we run out of OOM victims and there's only unreclaimable
memory left, or the task writing to memory.max is killed. This allows
users to shrink groups reliably, and the behavior is consistent with
what happens when new charges are attempted in excess of memory.max.
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Cc: Vladimir Davydov <[email protected]>
Cc: <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
When setting memory.high below usage, nothing happens until the next
charge comes along, and then it will only reclaim its own charge and not
the now potentially huge excess of the new memory.high. This can cause
groups to stay in excess of their memory.high indefinitely.
To fix that, when shrinking memory.high, kick off a reclaim cycle that
goes after the delta.
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Cc: Vladimir Davydov <[email protected]>
Cc: <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
In machines with 140G of memory and enterprise flash storage, we have
seen read and write bursts routinely exceed the kswapd watermarks and
cause thundering herds in direct reclaim. Unfortunately, the only way
to tune kswapd aggressiveness is through adjusting min_free_kbytes - the
system's emergency reserves - which is entirely unrelated to the
system's latency requirements. In order to get kswapd to maintain a
250M buffer of free memory, the emergency reserves need to be set to 1G.
That is a lot of memory wasted for no good reason.
On the other hand, it's reasonable to assume that allocation bursts and
overall allocation concurrency scale with memory capacity, so it makes
sense to make kswapd aggressiveness a function of that as well.
Change the kswapd watermark scale factor from the currently fixed 25% of
the tunable emergency reserve to a tunable 0.1% of memory.
Beyond 1G of memory, this will produce bigger watermark steps than the
current formula in default settings. Ensure that the new formula never
chooses steps smaller than that, i.e. 25% of the emergency reserve.
On a 140G machine, this raises the default watermark steps - the
distance between min and low, and low and high - from 16M to 143M.
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Mel Gorman <[email protected]>
Acked-by: Rik van Riel <[email protected]>
Acked-by: David Rientjes <[email protected]>
Cc: Joonsoo Kim <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
When the OOM killer scans tasks and encounters a PF_EXITING one, it
force-selects that task regardless of the score. The problem is that if
that task got stuck waiting for some state the allocation site is
holding, the OOM reaper can not move on to the next best victim.
Frankly, I don't even know why we check for exiting tasks in the OOM
killer. We've tried direct reclaim at least 15 times by the time we
decide the system is OOM, there was plenty of time to exit and free
memory; and a task might exit voluntarily right after we issue a kill.
This is testing pure noise. Remove it.
Signed-off-by: Tetsuo Handa <[email protected]>
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Cc: Kirill A. Shutemov <[email protected]>
Cc: Mel Gorman <[email protected]>
Cc: David Rientjes <[email protected]>
Cc: Oleg Nesterov <[email protected]>
Cc: Hugh Dickins <[email protected]>
Cc: Andrea Argangeli <[email protected]>
Cc: Rik van Riel <[email protected]>
Cc: Sasha Levin <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Migration accounting in the memory controller used to have to handle
both oldpage and newpage being on the LRU already; fuse's page cache
replacement used to pass a recycled newpage that had been uncharged but
not freed and removed from the LRU, and the memcg migration code used to
uncharge oldpage to "pass on" the existing charge to newpage.
Nowadays, pages are no longer uncharged when truncated from the page
cache, but rather only at free time, so if a LRU page is recycled in
page cache replacement it'll also still be charged. And we bail out of
the charge transfer altogether in that case. Tell commit_charge() that
we know newpage is not on the LRU, to avoid taking the zone->lru_lock
unnecessarily from the migration path.
But also, oldpage is no longer uncharged inside migration. We only use
oldpage for its page->mem_cgroup and page size, so we don't care about
its LRU state anymore either. Remove any mention from the kernel doc.
Signed-off-by: Johannes Weiner <[email protected]>
Suggested-by: Hugh Dickins <[email protected]>
Acked-by: Vladimir Davydov <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Cc: Mateusz Guzik <[email protected]>
Cc: Sergey Senozhatsky <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Rather than scattering mem_cgroup_migrate() calls all over the place,
have a single call from a safe place where every migration operation
eventually ends up in - migrate_page_copy().
Signed-off-by: Johannes Weiner <[email protected]>
Suggested-by: Hugh Dickins <[email protected]>
Acked-by: Vladimir Davydov <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Cc: Mateusz Guzik <[email protected]>
Cc: Sergey Senozhatsky <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
There are several users that nest lock_page_memcg() inside lock_page()
to prevent page->mem_cgroup from changing. But the page lock prevents
pages from moving between cgroups, so that is unnecessary overhead.
Remove lock_page_memcg() in contexts with locked contexts and fix the
debug code in the page stat functions to be okay with the page lock.
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Vladimir Davydov <[email protected]>
Cc: Michal Hocko <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Now that migration doesn't clear page->mem_cgroup of live pages anymore,
it's safe to make lock_page_memcg() and the memcg stat functions take
pages, and spare the callers from memcg objects.
[[email protected]: fix warnings]
Signed-off-by: Johannes Weiner <[email protected]>
Suggested-by: Vladimir Davydov <[email protected]>
Acked-by: Vladimir Davydov <[email protected]>
Cc: Michal Hocko <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Changing a page's memcg association complicates dealing with the page,
so we want to limit this as much as possible. Page migration e.g. does
not have to do that. Just like page cache replacement, it can forcibly
charge a replacement page, and then uncharge the old page when it gets
freed. Temporarily overcharging the cgroup by a single page is not an
issue in practice, and charging is so cheap nowadays that this is much
preferrable to the headache of messing with live pages.
The only place that still changes the page->mem_cgroup binding of live
pages is when pages move along with a task to another cgroup. But that
path isolates the page from the LRU, takes the page lock, and the move
lock (lock_page_memcg()). That means page->mem_cgroup is always stable
in callers that have the page isolated from the LRU or locked. Lighter
unlocked paths, like writeback accounting, can use lock_page_memcg().
[[email protected]: fix build]
[[email protected]: fix lockdep splat]
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Vladimir Davydov <[email protected]>
Cc: Johannes Weiner <[email protected]>
Cc: Michal Hocko <[email protected]>
Cc: Greg Thelen <[email protected]>
Cc: Hugh Dickins <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Cache thrash detection (see a528910e12 "mm: thrash detection-based
file cache sizing" for details) currently only works on the system
level, not inside cgroups. Worse, as the refaults are compared to the
global number of active cache, cgroups might wrongfully get all their
refaults activated when their pages are hotter than those of others.
Move the refault machinery from the zone to the lruvec, and then tag
eviction entries with the memcg ID. This makes the thrash detection
work correctly inside cgroups.
[[email protected]: do not return from workingset_activation() with locked rcu and page]
Signed-off-by: Johannes Weiner <[email protected]>
Signed-off-by: Sergey Senozhatsky <[email protected]>
Reviewed-by: Vladimir Davydov <[email protected]>
Cc: Michal Hocko <[email protected]>
Cc: David Rientjes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
For per-cgroup thrash detection, we need to store the memcg ID inside
the radix tree cookie as well. However, on 32 bit that doesn't leave
enough bits for the eviction timestamp to cover the necessary range of
recently evicted pages. The radix tree entry would look like this:
[ RADIX_TREE_EXCEPTIONAL(2) | ZONEID(2) | MEMCGID(16) | EVICTION(12) ]
12 bits means 4096 pages, means 16M worth of recently evicted pages.
But refaults are actionable up to distances covering half of memory. To
not miss refaults, we have to stretch out the range at the cost of how
precisely we can tell when a page was evicted. This way we can shave
off lower bits from the eviction timestamp until the necessary range is
covered. E.g. grouping evictions into 1M buckets (256 pages) will
stretch the longest representable refault distance to 4G.
This patch implements eviction buckets that are automatically sized
according to the available bits and the necessary refault range, in
preparation for per-cgroup thrash detection.
The maximum actionable distance is currently half of memory, but to
support memory hotplug of up to 200% of boot-time memory, we size the
buckets to cover double the distance. Beyond that, thrashing won't be
detectable anymore.
During boot, the kernel will print out the exact parameters, like so:
[ 0.113929] workingset: timestamp_bits=12 max_order=18 bucket_order=6
In this example, there are 12 radix entry bits available for the
eviction timestamp, to cover a maximum distance of 2^18 pages (this is a
1G machine). Consequently, evictions must be grouped into buckets of
2^6 pages, or 256K.
Signed-off-by: Johannes Weiner <[email protected]>
Reviewed-by: Vladimir Davydov <[email protected]>
Cc: Michal Hocko <[email protected]>
Cc: David Rientjes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Per-cgroup thrash detection will need to derive a live memcg from the
eviction cookie, and doing that inside unpack_shadow() will get nasty
with the reference handling spread over two functions.
In preparation, make unpack_shadow() clearly about extracting static
data, and let workingset_refault() do all the higher-level handling.
Signed-off-by: Johannes Weiner <[email protected]>
Reviewed-by: Vladimir Davydov <[email protected]>
Cc: Michal Hocko <[email protected]>
Cc: David Rientjes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
These patches tag the page cache radix tree eviction entries with the
memcg an evicted page belonged to, thus making per-cgroup LRU reclaim
work properly and be as adaptive to new cache workingsets as global
reclaim already is.
This should have been part of the original thrash detection patch
series, but was deferred due to the complexity of those patches.
This patch (of 5):
So far the only sites that needed to exclude charge migration to
stabilize page->mem_cgroup have been per-cgroup page statistics, hence
the name mem_cgroup_begin_page_stat(). But per-cgroup thrash detection
will add another site that needs to ensure page->mem_cgroup lifetime.
Rename these locking functions to the more generic lock_page_memcg() and
unlock_page_memcg(). Since charge migration is a cgroup1 feature only,
we might be able to delete it at some point, and these now easy to
identify locking sites along with it.
Signed-off-by: Johannes Weiner <[email protected]>
Suggested-by: Vladimir Davydov <[email protected]>
Acked-by: Vladimir Davydov <[email protected]>
Cc: Michal Hocko <[email protected]>
Cc: David Rientjes <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Add cgroup_no_v1= to kernel-parameters.txt, and a small blurb to
cgroup-v2.txt section about transitioning from cgroup to cgroup2.
Signed-off-by: Johannes Weiner <[email protected]>
Signed-off-by: Tejun Heo <[email protected]>
Testing cgroup2 can be painful with system software automatically
mounting and populating all cgroup controllers in v1 mode. Sometimes
they can be unmounted from rc.local, sometimes even that is too late.
Provide a commandline option to disable certain controllers in v1
mounts, so that they remain available for cgroup2 mounts.
Example use:
cgroup_no_v1=memory,cpu
cgroup_no_v1=all
Disabling will be confirmed at boot-time as such:
[ 0.013770] Disabling cpu control group subsystem in v1 mounts
[ 0.016004] Disabling memory control group subsystem in v1 mounts
Signed-off-by: Johannes Weiner <[email protected]>
Signed-off-by: Tejun Heo <[email protected]>
MEM_CGROUP_STAT_NSTATS is just a delimiter for cgroup1 statistics, not
an actual array entry. Reuse it for the first cgroup2 stat entry, like
in the event array.
Fixes: b2807f07f4 ("mm: memcontrol: add "sock" to cgroup2 memory.stat")
Signed-off-by: Johannes Weiner <[email protected]>
Cc: Vladimir Davydov <[email protected]>
Cc: Michal Hocko <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Commit b76437579d ("procfs: mark thread stack correctly in
proc/<pid>/maps") added [stack:TID] annotation to /proc/<pid>/maps.
Finding the task of a stack VMA requires walking the entire thread list,
turning this into quadratic behavior: a thousand threads means a
thousand stacks, so the rendering of /proc/<pid>/maps needs to look at a
million combinations.
The cost is not in proportion to the usefulness as described in the
patch.
Drop the [stack:TID] annotation to make /proc/<pid>/maps (and
/proc/<pid>/numa_maps) usable again for higher thread counts.
The [stack] annotation inside /proc/<pid>/task/<tid>/maps is retained, as
identifying the stack VMA there is an O(1) operation.
Siddesh said:
"The end users needed a way to identify thread stacks programmatically and
there wasn't a way to do that. I'm afraid I no longer remember (or have
access to the resources that would aid my memory since I changed
employers) the details of their requirement. However, I did do this on my
own time because I thought it was an interesting project for me and nobody
really gave any feedback then as to its utility, so as far as I am
concerned you could roll back the main thread maps information since the
information is available in the thread-specific files"
Signed-off-by: Johannes Weiner <[email protected]>
Cc: "Kirill A. Shutemov" <[email protected]>
Cc: Siddhesh Poyarekar <[email protected]>
Cc: Shaohua Li <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The cgroup methods are no longer used after baac50bbc3 ("net:
tcp_memcontrol: simplify linkage between socket and page counter").
The hunk to delete them was included in the original patch but must
have gotten lost during conflict resolution on the way upstream.
Fixes: baac50bbc3 ("net: tcp_memcontrol: simplify linkage between socket and page counter")
Signed-off-by: Johannes Weiner <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Changing page->mem_cgroup of a live page is tricky and fragile. In
particular, the memcg writeback code relies on that mapping being stable
and users of mem_cgroup_replace_page() not overlapping with dirtyable
inodes.
Page cache replacement doesn't have to do that, though. Instead of being
clever and transferring the charge from the old page to the new,
force-charge the new page and leave the old page alone. A temporary
overcharge won't matter in practice, and the old page is going to be freed
shortly after this anyway. And this is not performance critical.
Signed-off-by: Johannes Weiner <[email protected]>
Cc: Michal Hocko <[email protected]>
Cc: Vladimir Davydov <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The creation and teardown of struct mem_cgroup is fairly messy and
that has attracted mistakes and subtle bugs before.
The main cause for this is that there is no clear model about what
needs to happen when, and that attracts more chaos. So create one:
1. mem_cgroup_alloc() should allocate struct mem_cgroup and its
auxiliary members and initialize work items, locks etc. so that the
object it returns is fully initialized and in a neutral state.
2. mem_cgroup_css_alloc() will use mem_cgroup_alloc() to obtain a new
memcg object and configure it and the system according to the role
of the new memory-controlled cgroup in the hierarchy.
3. mem_cgroup_css_online() is no longer needed to synchronize with
iterators, but it verifies css->id which isn't available earlier.
4. mem_cgroup_css_offline() implements stuff that needs to happen upon
the user-visible destruction of a cgroup, which includes stopping
all user interfacing as well as releasing certain structures when
continued memory consumption would be unexpected at that point.
5. mem_cgroup_css_free() prepares the system and the memcg object for
the object's disappearance, neutralizes its state, and then gives
it back to mem_cgroup_free().
6. mem_cgroup_free() releases struct mem_cgroup and auxiliary memory.
[[email protected]: fix SLOB build regression]
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Vladimir Davydov <[email protected]>
Cc: Michal Hocko <[email protected]>
Signed-off-by: Arnd Bergmann <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
There are no more external users of struct cg_proto, flatten the
structure into struct mem_cgroup.
Since using those struct members doesn't stand out as much anymore,
add cgroup2 static branches to make it clearer which code is legacy.
Suggested-by: Vladimir Davydov <[email protected]>
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Vladimir Davydov <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
What CONFIG_INET and CONFIG_LEGACY_KMEM guard inside the memory
controller code is insignificant, having these conditionals is not
worth the complication and fragility that comes with them.
[[email protected]: rework mem_cgroup_css_free() statement ordering]
Signed-off-by: Johannes Weiner <[email protected]>
Cc: Michal Hocko <[email protected]>
Acked-by: Vladimir Davydov <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The original cgroup memory controller has an extension to account slab
memory (and other "kernel memory" consumers) in a separate "kmem"
counter, once the user set an explicit limit on that "kmem" pool.
However, this includes various consumers whose sizes are directly linked
to userspace activity. Accounting them as an optional "kmem" extension
is problematic for several reasons:
1. It leaves the main memory interface with incomplete semantics. A
user who puts their workload into a cgroup and configures a memory
limit does not expect us to leave holes in the containment as big
as the dentry and inode cache, or the kernel stack pages.
2. If the limit set on this random historical subgroup of consumers is
reached, subsequent allocations will fail even when the main memory
pool available to the cgroup is not yet exhausted and/or has
reclaimable memory in it.
3. Calling it 'kernel memory' is misleading. The dentry and inode
caches are no more 'kernel' (or no less 'user') memory than the
page cache itself. Treating these consumers as different classes is
a historical implementation detail that should not leak to users.
So, in addition to page cache, anonymous memory, and network socket
memory, account the following memory consumers per default in the
cgroup2 memory controller:
- threadinfo
- task_struct
- task_delay_info
- pid
- cred
- mm_struct
- vm_area_struct and vm_region (nommu)
- anon_vma and anon_vma_chain
- signal_struct
- sighand_struct
- fs_struct
- files_struct
- fdtable and fdtable->full_fds_bits
- dentry and external_name
- inode for all filesystems.
This should give us reasonable memory isolation for most common
workloads out of the box.
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Cc: Tejun Heo <[email protected]>
Acked-by: Vladimir Davydov <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The cgroup2 memory controller will account important in-kernel memory
consumers per default. Move all necessary components to CONFIG_MEMCG.
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Vladimir Davydov <[email protected]>
Cc: Michal Hocko <[email protected]>
Cc: Arnd Bergmann <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The cgroup2 memory controller will include important in-kernel memory
consumers per default, including socket memory, but it will no longer
carry the historic tcp control interface.
Separate the kmem state init from the tcp control interface init in
preparation for that.
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Cc: Tejun Heo <[email protected]>
Acked-by: Vladimir Davydov <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
On any given memcg, the kmem accounting feature has three separate
states: not initialized, structures allocated, and actively accounting
slab memory. These are represented through a combination of the
kmem_acct_activated and kmem_acct_active flags, which is confusing.
Convert to a kmem_state enum with the states NONE, ALLOCATED, and
ONLINE. Then rename the functions to modify the state accordingly.
This follows the nomenclature of css object states more closely.
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Cc: Tejun Heo <[email protected]>
Acked-by: Vladimir Davydov <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The kmem page_counter's limit is initialized to PAGE_COUNTER_MAX inside
mem_cgroup_css_online(). There is no need to repeat this from
memcg_propagate_kmem().
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Cc: Tejun Heo <[email protected]>
Acked-by: Vladimir Davydov <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
This series adds accounting of the historical "kmem" memory consumers to
the cgroup2 memory controller.
These consumers include the dentry cache, the inode cache, kernel stack
pages, and a few others that are pointed out in patch 7/8. The
footprint of these consumers is directly tied to userspace activity in
common workloads, and so they have to be part of the minimally viable
configuration in order to present a complete feature to our users.
The cgroup2 interface of the memory controller is far from complete, but
this series, along with the socket memory accounting series, provides
the final semantic changes for the existing memory knobs in the cgroup2
interface, which is scheduled for initial release in the next merge
window.
This patch (of 8):
Remove unused css argument frmo memcg_init_kmem()
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Cc: Tejun Heo <[email protected]>
Acked-by: Vladimir Davydov <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
According to <linux/jump_label.h> the direct use of struct static_key is
deprecated. Update the socket and slab accounting code accordingly.
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: David S. Miller <[email protected]>
Reported-by: Jason Baron <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Let the networking stack know when a memcg is under reclaim pressure so
that it can clamp its transmit windows accordingly.
Whenever the reclaim efficiency of a cgroup's LRU lists drops low enough
for a MEDIUM or HIGH vmpressure event to occur, assert a pressure state
in the socket and tcp memory code that tells it to curb consumption
growth from sockets associated with said control group.
Traditionally, vmpressure reports for the entire subtree of a memcg
under pressure, which drops useful information on the individual groups
reclaimed. However, it's too late to change the userinterface, so add a
second reporting mode that reports on the level of reclaim instead of at
the level of pressure, and use that report for sockets.
vmpressure events are naturally edge triggered, so for hysteresis assert
socket pressure for a second to allow for subsequent vmpressure events
to occur before letting the socket code return to normal.
This will likely need finetuning for a wider variety of workloads, but
for now stick to the vmpressure presets and keep hysteresis simple.
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: David S. Miller <[email protected]>
Reviewed-by: Vladimir Davydov <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Socket memory can be a significant share of overall memory consumed by
common workloads. In order to provide reasonable resource isolation in
the unified hierarchy, this type of memory needs to be included in the
tracking/accounting of a cgroup under active memory resource control.
Overhead is only incurred when a non-root control group is created AND
the memory controller is instructed to track and account the memory
footprint of that group. cgroup.memory=nosocket can be specified on the
boot commandline to override any runtime configuration and forcibly
exclude socket memory from active memory resource control.
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: David S. Miller <[email protected]>
Reviewed-by: Vladimir Davydov <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The unified hierarchy memory controller doesn't expose the memory+swap
counter to userspace, but its accounting is hardcoded in all charge
paths right now, including the per-cpu charge cache ("the stock").
To avoid adding yet more pointless memory+swap accounting with the
socket memory support in unified hierarchy, disable the counter
altogether when in unified hierarchy mode.
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Reviewed-by: Vladimir Davydov <[email protected]>
Acked-by: David S. Miller <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The unified hierarchy memory controller is going to use this jump label
as well to control the networking callbacks. Move it to the memory
controller code and give it a more generic name.
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Reviewed-by: Vladimir Davydov <[email protected]>
Acked-by: David S. Miller <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
There won't be any separate counters for socket memory consumed by
protocols other than TCP in the future. Remove the indirection and link
sockets directly to their owning memory cgroup.
Signed-off-by: Johannes Weiner <[email protected]>
Reviewed-by: Vladimir Davydov <[email protected]>
Acked-by: David S. Miller <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
There won't be a tcp control soft limit, so integrating the memcg code
into the global skmem limiting scheme complicates things unnecessarily.
Replace this with simple and clear charge and uncharge calls--hidden
behind a jump label--to account skb memory.
Note that this is not purely aesthetic: as a result of shoehorning the
per-memcg code into the same memory accounting functions that handle the
global level, the old code would compare the per-memcg consumption
against the smaller of the per-memcg limit and the global limit. This
allowed the total consumption of multiple sockets to exceed the global
limit, as long as the individual sockets stayed within bounds. After
this change, the code will always compare the per-memcg consumption to
the per-memcg limit, and the global consumption to the global limit, and
thus close this loophole.
Without a soft limit, the per-memcg memory pressure state in sockets is
generally questionable. However, we did it until now, so we continue to
enter it when the hard limit is hit, and packets are dropped, to let
other sockets in the cgroup know that they shouldn't grow their transmit
windows, either. However, keep it simple in the new callback model and
leave memory pressure lazily when the next packet is accepted (as
opposed to doing it synchroneously when packets are processed). When
packets are dropped, network performance will already be in the toilet,
so that should be a reasonable trade-off.
As described above, consumption is now checked on the per-memcg level
and the global level separately. Likewise, memory pressure states are
maintained on both the per-memcg level and the global level, and a
socket is considered under pressure when either level asserts as much.
Signed-off-by: Johannes Weiner <[email protected]>
Reviewed-by: Vladimir Davydov <[email protected]>
Acked-by: David S. Miller <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
tcp_memcontrol replicates the global sysctl_mem limit array per cgroup,
but it only ever sets these entries to the value of the memory_allocated
page_counter limit. Use the latter directly.
Signed-off-by: Johannes Weiner <[email protected]>
Reviewed-by: Vladimir Davydov <[email protected]>
Acked-by: David S. Miller <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The number of allocated sockets is used for calculations in the soft
limit phase, where packets are accepted but the socket is under memory
pressure.
Since there is no soft limit phase in tcp_memcontrol, and memory
pressure is only entered when packets are already dropped, this is
actually dead code. Remove it.
As this is the last user of parent_cg_proto(), remove that too.
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: David S. Miller <[email protected]>
Reviewed-by: Vladimir Davydov <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Move the jump-label from sock_update_memcg() and sock_release_memcg() to
the callsite, and so eliminate those function calls when socket
accounting is not enabled.
This also eliminates the need for dummy functions because the calls will
be optimized away if the Kconfig options are not enabled.
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: David S. Miller <[email protected]>
Reviewed-by: Vladimir Davydov <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
When a cgroup currently breaches its socket memory limit, it enters
memory pressure mode for itself and its *ancestors*. This throttles
transmission in unrelated sibling and cousin subtrees that have nothing
to do with the breached limit.
On the contrary, breaching a limit should make that group and its
*children* enter memory pressure mode. But this happens already, albeit
lazily: if an ancestor limit is breached, siblings will enter memory
pressure on their own once the next packet arrives for them.
So no additional hierarchy code is needed. Remove the bogus stuff.
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: David S. Miller <[email protected]>
Reviewed-by: Vladimir Davydov <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
When charging socket memory, the code currently checks only the local
page counter for excess to determine whether the memcg is under socket
pressure. But even if the local counter is fine, one of the ancestors
could have breached its limit, which should also force this child to
enter socket pressure. This currently doesn't happen.
Fix this by using page_counter_try_charge() first. If that fails, it
means that either the local counter or one of the ancestors are in
excess of their limit, and the child should enter socket pressure.
Fixes: 3e32cb2e0a ("mm: memcontrol: lockless page counters")
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: David S. Miller <[email protected]>
Reviewed-by: Vladimir Davydov <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
A later patch will need this symbol in files other than memcontrol.c, so
export it now and replace mem_cgroup_root_css at the same time.
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Acked-by: David S. Miller <[email protected]>
Reviewed-by: Vladimir Davydov <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The MemAvailable item in /proc/meminfo is to give users a hint of how
much memory is allocatable without causing swapping, so it excludes the
zones' low watermarks as unavailable to userspace.
However, for a userspace allocation, kswapd will actually reclaim until
the free pages hit a combination of the high watermark and the page
allocator's lowmem protection that keeps a certain amount of DMA and
DMA32 memory from userspace as well.
Subtract the full amount we know to be unavailable to userspace from the
number of free pages when calculating MemAvailable.
Signed-off-by: Johannes Weiner <[email protected]>
Cc: Rik van Riel <[email protected]>
Cc: Mel Gorman <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The dirty balance reserve that dirty throttling has to consider is
merely memory not available to userspace allocations. There is nothing
writeback-specific about it. Generalize the name so that it's reusable
outside of that context.
Signed-off-by: Johannes Weiner <[email protected]>
Cc: Rik van Riel <[email protected]>
Cc: Mel Gorman <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
To make it easier to quickly find what's needed list the basic
resource controllers of cgroup2 first - io, memory, cpu - while
pushing the more exotic and/or legacy controllers to the bottom.
tj: Removed spurious "&& CGROUPS" from CGROUP_PERF as suggested by Li.
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Zefan Li <[email protected]>
Signed-off-by: Tejun Heo <[email protected]>
The config options for the different cgroup controllers use various
terms: resource controller, cgroup subsystem, etc. Simplify this to
"controller", which is clear enough in the cgroup context.
Signed-off-by: Johannes Weiner <[email protected]>
Signed-off-by: Tejun Heo <[email protected]>
Vladimir architected and authored much of the current state of the
memcg's slab memory accounting and tracking. Make sure he gets CC'd on
bug reports ;-)
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Vladimir Davydov <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
page_counter_try_charge() currently returns 0 on success and -ENOMEM on
failure, which is surprising behavior given the function name.
Make it follow the expected pattern of try_stuff() functions that return a
boolean true to indicate success, or false for failure.
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Cc: Vladimir Davydov <[email protected]
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
memory.current on the root level doesn't add anything that wouldn't be
more accurate and detailed using system statistics. It already doesn't
include slabs, and it'll be a pain to keep in sync when further memory
types are accounted in the memory controller. Remove it.
Note that this applies to the new unified hierarchy interface only.
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Cc: Tejun Heo <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The should_alloc_retry() function was meant to encapsulate retry
conditions of the allocator slowpath, but there are still checks
remaining in the main function, and much of how the retrying is
performed also depends on the OOM killer progress. The physical
separation of those conditions make the code hard to follow.
Inline the should_alloc_retry() checks. Notes:
- The __GFP_NOFAIL check is already done in __alloc_pages_may_oom(),
replace it with looping on OOM killer progress
- The pm_suspended_storage() check is meant to skip the OOM killer
when reclaim has no IO available, move to __alloc_pages_may_oom()
- The order <= PAGE_ALLOC_COSTLY order is re-united with its original
counterpart of checking whether reclaim actually made any progress
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Cc: David Rientjes <[email protected]>
Cc: Tetsuo Handa <[email protected]>
Cc: Andrea Arcangeli <[email protected]>
Cc: Dave Chinner <[email protected]>
Cc: Vlastimil Babka <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The zonelist locking and the oom_sem are two overlapping locks that are
used to serialize global OOM killing against different things.
The historical zonelist locking serializes OOM kills from allocations with
overlapping zonelists against each other to prevent killing more tasks
than necessary in the same memory domain. Only when neither tasklists nor
zonelists from two concurrent OOM kills overlap (tasks in separate memcgs
bound to separate nodes) are OOM kills allowed to execute in parallel.
The younger oom_sem is a read-write lock to serialize OOM killing against
the PM code trying to disable the OOM killer altogether.
However, the OOM killer is a fairly cold error path, there is really no
reason to optimize for highly performant and concurrent OOM kills. And
the oom_sem is just flat-out redundant.
Replace both locking schemes with a single global mutex serializing OOM
kills regardless of context.
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Acked-by: David Rientjes <[email protected]>
Cc: Tetsuo Handa <[email protected]>
Cc: Andrea Arcangeli <[email protected]>
Cc: Dave Chinner <[email protected]>
Cc: Vlastimil Babka <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Disabling the OOM killer needs to exclude allocators from entering, not
existing victims from exiting.
Right now the only waiter is suspend code, which achieves quiescence by
disabling the OOM killer. But later on we want to add waits that hold
the lock instead to stop new victims from showing up.
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Acked-by: David Rientjes <[email protected]>
Cc: Tetsuo Handa <[email protected]>
Cc: Andrea Arcangeli <[email protected]>
Cc: Dave Chinner <[email protected]>
Cc: Vlastimil Babka <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
It turns out that the mechanism to wait for exiting OOM victims is less
generic than it looks: it won't issue wakeups unless the OOM killer is
disabled.
The reason this check was added was the thought that, since only the OOM
disabling code would wait on this queue, wakeup operations could be
saved when that specific consumer is known to be absent.
However, this is quite the handgrenade. Later attempts to reuse the
waitqueue for other purposes will lead to completely unexpected bugs and
the failure mode will appear seemingly illogical. Generally, providers
shouldn't make unnecessary assumptions about consumers.
This could have been replaced with waitqueue_active(), but it only saves
a few instructions in one of the coldest paths in the kernel. Simply
remove it.
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Acked-by: David Rientjes <[email protected]>
Cc: Tetsuo Handa <[email protected]>
Cc: Andrea Arcangeli <[email protected]>
Cc: Dave Chinner <[email protected]>
Cc: Vlastimil Babka <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Rename unmark_oom_victim() to exit_oom_victim(). Marking and unmarking
are related in functionality, but the interface is not symmetrical at
all: one is an internal OOM killer function used during the killing, the
other is for an OOM victim to signal its own death on exit later on.
This has locking implications, see follow-up changes.
While at it, rename mark_tsk_oom_victim() to mark_oom_victim(), which
is easier on the eye.
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: David Rientjes <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Cc: Tetsuo Handa <[email protected]>
Cc: Andrea Arcangeli <[email protected]>
Cc: Dave Chinner <[email protected]>
Cc: Vlastimil Babka <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
On -rt, the VM_BUG_ON(!irqs_disabled()) triggers inside the memcg
swapout path because the spin_lock_irq(&mapping->tree_lock) in the
caller doesn't actually disable the hardware interrupts - which is fine,
because on -rt the tophalves run in process context and so we are still
safe from preemption while updating the statistics.
Remove the VM_BUG_ON() but keep the comment of what we rely on.
Signed-off-by: Johannes Weiner <[email protected]>
Reported-by: Clark Williams <[email protected]>
Cc: Fernando Lopez-Lezcano <[email protected]>
Cc: Steven Rostedt <[email protected]>
Cc: Thomas Gleixner <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Historically, !__GFP_FS allocations were not allowed to invoke the OOM
killer once reclaim had failed, but nevertheless kept looping in the
allocator.
Commit 9879de7373 ("mm: page_alloc: embed OOM killing naturally into
allocation slowpath"), which should have been a simple cleanup patch,
accidentally changed the behavior to aborting the allocation at that
point. This creates problems with filesystem callers (?) that currently
rely on the allocator waiting for other tasks to intervene.
Revert the behavior as it shouldn't have been changed as part of a
cleanup patch.
Fixes: 9879de7373 ("mm: page_alloc: embed OOM killing naturally into allocation slowpath")
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Reported-by: Tetsuo Handa <[email protected]>
Cc: Theodore Ts'o <[email protected]>
Cc: Dave Chinner <[email protected]>
Acked-by: David Rientjes <[email protected]>
Cc: Oleg Nesterov <[email protected]>
Cc: Mel Gorman <[email protected]>
Cc: <[email protected]> [3.19.x]
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The memcg control knobs indicate the highest possible value using the
symbolic name "infinity", which is long and awkward to type.
Switch to the string "max", which is just as descriptive but shorter and
sweeter.
This changes a user interface, so do it before the release and before
the development flag is dropped from the default hierarchy.
Signed-off-by: Johannes Weiner <[email protected]>
Cc: Michal Hocko <[email protected]>
Cc: Tejun Heo <[email protected]>
Cc: Vladimir Davydov <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The swap controller code is scattered all over the file. Gather all
the code that isn't directly needed by the memory controller at the
end of the file in its own CONFIG_MEMCG_SWAP section.
Signed-off-by: Johannes Weiner <[email protected]>
Cc: Michal Hocko <[email protected]>
Reviewed-by: Vladimir Davydov <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The initialization code for the per-cpu charge stock and the soft
limit tree is compact enough to inline it into mem_cgroup_init().
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Reviewed-by: Vladimir Davydov <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
- No need to test the node for N_MEMORY. node_online() is enough for
node fallback to work in slab, use NUMA_NO_NODE for everything else.
- Remove the BUG_ON() for allocation failure. A NULL pointer crash is
just as descriptive, and the absent return value check is obvious.
- Move local variables to the inner-most blocks.
- Point to the tree structure after its initialized, not before, it's
just more logical that way.
Signed-off-by: Johannes Weiner <[email protected]>
Cc: Michal Hocko <[email protected]>
Cc: Vladimir Davydov <[email protected]>
Cc: Guenter Roeck <[email protected]>
Cc: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Turn the move type enum into flags and give the flags field a shorter
name. Once that is done, move_anon() and move_file() are simple enough to
just fold them into the callsites.
[[email protected]: tweak MOVE_MASK definition, per Michal]
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Reviewed-by: Vladimir Davydov <[email protected]>
Cc: Greg Thelen <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[[email protected]: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Cc: Vladimir Davydov <[email protected]>
Cc: Greg Thelen <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The unified hierarchy interface for memory cgroups will no longer use "-1"
to mean maximum possible resource value. In preparation for this, make
the string an argument and let the caller supply it.
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Cc: Vladimir Davydov <[email protected]>
Cc: Greg Thelen <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The complexity of memcg page stat synchronization is currently leaking
into the callsites, forcing them to keep track of the move_lock state and
the IRQ flags. Simplify the API by tracking it in the memcg.
Signed-off-by: Johannes Weiner <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Reviewed-by: Vladimir Davydov <[email protected]>
Cc: Wu Fengguang <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Johannes Weiner