Patch series "Convert all THP vmstat counters to pages", v6.
This patch series is aimed to convert all THP vmstat counters to pages.
The unit of some vmstat counters are pages, some are bytes, some are
HPAGE_PMD_NR, and some are KiB. When we want to expose these vmstat
counters to the userspace, we have to know the unit of the vmstat counters
is which one. When the unit is bytes or kB, both clearly distinguishable
by the B/KB suffix. But for the THP vmstat counters, we may make mistakes.
For example, the below is some bug fix for the THP vmstat counters:
- 7de2e9f195 ("mm: memcontrol: correct the NR_ANON_THPS counter of hierarchical memcg")
- The first commit in this series ("fix NR_ANON_THPS accounting in charge moving")
This patch series can make the code clear. And make all the unit of the THP
vmstat counters in pages. Finally, the unit of the vmstat counters are
pages, kB and bytes. The B/KB suffix can tell us that the unit is bytes
or kB. The rest which is without suffix are pages.
In this series, I changed the following vmstat counters unit from HPAGE_PMD_NR
to pages. However, there is no change to the print format of output to user
space.
- NR_ANON_THPS
- NR_FILE_THPS
- NR_SHMEM_THPS
- NR_SHMEM_PMDMAPPED
- NR_FILE_PMDMAPPED
Doing this also can make the statistics more accuracy for the THP vmstat
counters. This series is consistent with 8f182270df ("mm/swap.c: flush lru
pvecs on compound page arrival").
Because we use struct per_cpu_nodestat to cache the vmstat counters, which
leads to inaccurate statistics especially THP vmstat counters. In the systems
with hundreds of processors it can be GBs of memory. For example, for a 96
CPUs system, the threshold is the maximum number of 125. And the per cpu
counters can cache 23.4375 GB in total.
The THP page is already a form of batched addition (it will add 512 worth of
memory in one go) so skipping the batching seems like sensible. Although every
THP stats update overflows the per-cpu counter, resorting to atomic global
updates. But it can make the statistics more accuracy for the THP vmstat
counters. From this point of view, I think that do this converting is
reasonable.
Thanks Hugh for mentioning this. This was inspired by Johannes and Roman.
Thanks to them.
This patch (of 7):
The unit of NR_ANON_THPS is HPAGE_PMD_NR already. So it should inc/dec by
one rather than nr_pages.
Link: https://lkml.kernel.org/r/20201228164110.2838-1-songmuchun@bytedance.com
Link: https://lkml.kernel.org/r/20201228164110.2838-2-songmuchun@bytedance.com
Fixes: 468c398233 ("mm: memcontrol: switch to native NR_ANON_THPS counter")
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Pankaj Gupta <pankaj.gupta@cloud.ionos.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Feng Tang <feng.tang@intel.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: NeilBrown <neilb@suse.de>
Cc: Rafael. J. Wysocki <rafael@kernel.org>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Sami Tolvanen <samitolvanen@google.com>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
In general it's unknown in advance if a slab page will contain accounted
objects or not. In order to avoid memory waste, an obj_cgroup vector is
allocated dynamically when a need to account of a new object arises. Such
approach is memory efficient, but requires an expensive cmpxchg() to set
up the memcg/objcgs pointer, because an allocation can race with a
different allocation on another cpu.
But in some common cases it's known for sure that a slab page will contain
accounted objects: if the page belongs to a slab cache with a SLAB_ACCOUNT
flag set. It includes such popular objects like vm_area_struct, anon_vma,
task_struct, etc.
In such cases we can pre-allocate the objcgs vector and simple assign it
to the page without any atomic operations, because at this early stage the
page is not visible to anyone else.
A very simplistic benchmark (allocating 10000000 64-bytes objects in a
row) shows ~15% win. In the real life it seems that most workloads are
not very sensitive to the speed of (accounted) slab allocations.
[guro@fb.com: open-code set_page_objcgs() and add some comments, by Johannes]
Link: https://lkml.kernel.org/r/20201113001926.GA2934489@carbon.dhcp.thefacebook.com
[akpm@linux-foundation.org: fix it for mm-slub-call-account_slab_page-after-slab-page-initialization-fix.patch]
Link: https://lkml.kernel.org/r/20201110195753.530157-2-guro@fb.com
Signed-off-by: Roman Gushchin <guro@fb.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Christoph Lameter <cl@linux.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
We are capable of SetPageWorkingset based on refault distances after
commit aae466b005 ("mm/swap: implement workingset detection for
anonymous LRU"). This is done by workingset_refault(), which is right
above the unconditional SetPageWorkingset deleted by this patch.
The unconditional SetPageWorkingset miscategorizes pages that are read
ahead or never belonged to the working set (e.g., tmpfs pages accessed
only once by fd). When those pages are swapped in (after they were
swapped out) for the first time, they skew PSI (when using async swap).
When this happens again, depending on their refault distances, they might
skew workingset_restore_anon counter in addition to PSI because their
shadows indicate they were part of the working set.
Historically, SetPageWorkingset was added as part of the PSI series, and
Johannes said:
"It was meant to mark incoming pages under IO with SetPageWorkingset
when waiting for them constituted a memory stall.
On the page cache side, because we HAVE workingset detection, this was
specific to recently evicted pages that had been active in their
previous life. On the anon side, the aging algorithm had no
distinction between workingset and sporadically used pages. Given the
choice between a) no swapin stalls are pressure and b) all swapin
stalls are pressure, I went with the latter in order to detect swap
storms. The false positive case - high rate of swapin without severe
memory pressure - was relatively unlikely, because we tried to avoid
swapping until everything was completely on fire in the first place."
Link: https://lkml.kernel.org/r/20201209012400.1771150-1-yuzhao@google.com
Link: https://lkml.kernel.org/r/20201214231253.62313-1-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
For page splitting to succeed, the thread asking to split the page has to
be the only one with a reference to the page. Calling
wait_on_page_locked() while holding a reference to the page will
effectively prevent this from happening with sufficient threads waiting on
the same page. Use put_and_wait_on_page_locked() to sleep without holding
a reference to the page, then retry the page lookup after the page is
unlocked.
Since we now get the page lock a little earlier in filemap_update_page(),
we can eliminate a number of duplicate checks. The original intent
(commit ebded02788 ("avoid unnecessary calls to lock_page when waiting
for IO to complete during a read")) behind getting the page lock later was
to avoid re-locking the page after it has been brought uptodate by another
thread. We still avoid that because we go through the normal lookup path
again after the winning thread has brought the page uptodate.
Link: https://lkml.kernel.org/r/20210122160140.223228-7-willy@infradead.org
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: Kent Overstreet <kent.overstreet@gmail.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Patch series "Refactor generic_file_buffered_read", v5.
This is a combination of Christoph's work to refactor
generic_file_buffered_read() and some of my large-page support
which was disrupted by Kent's refactoring of generic_file_buffered_read.
This patch (of 18):
The recent split of generic_file_buffered_read() created some very long
function names which are hard to distinguish from each other. Rename as
follows:
generic_file_buffered_read_readpage -> filemap_read_page
generic_file_buffered_read_pagenotuptodate -> filemap_update_page
generic_file_buffered_read_no_cached_page -> filemap_create_page
generic_file_buffered_read_get_pages -> filemap_get_pages
Link: https://lkml.kernel.org/r/20210122160140.223228-1-willy@infradead.org
Link: https://lkml.kernel.org/r/20210122160140.223228-2-willy@infradead.org
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: Kent Overstreet <kent.overstreet@gmail.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Miaohe Lin <linmiaohe@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Currently the basic tests just validate various page table transformations
after starting with vm_get_page_prot(VM_READ|VM_WRITE|VM_EXEC) protection.
Instead scan over the entire protection_map[] for better coverage. It
also makes sure that all these basic page table tranformations checks hold
true irrespective of the starting protection value for the page table
entry. There is also a slight change in the debug print format for basic
tests to capture the protection value it is being tested with. The
modified output looks something like
[pte_basic_tests ]: Validating PTE basic ()
[pte_basic_tests ]: Validating PTE basic (read)
[pte_basic_tests ]: Validating PTE basic (write)
[pte_basic_tests ]: Validating PTE basic (read|write)
[pte_basic_tests ]: Validating PTE basic (exec)
[pte_basic_tests ]: Validating PTE basic (read|exec)
[pte_basic_tests ]: Validating PTE basic (write|exec)
[pte_basic_tests ]: Validating PTE basic (read|write|exec)
[pte_basic_tests ]: Validating PTE basic (shared)
[pte_basic_tests ]: Validating PTE basic (read|shared)
[pte_basic_tests ]: Validating PTE basic (write|shared)
[pte_basic_tests ]: Validating PTE basic (read|write|shared)
[pte_basic_tests ]: Validating PTE basic (exec|shared)
[pte_basic_tests ]: Validating PTE basic (read|exec|shared)
[pte_basic_tests ]: Validating PTE basic (write|exec|shared)
[pte_basic_tests ]: Validating PTE basic (read|write|exec|shared)
This adds a missing argument 'struct mm_struct *' in pud_basic_tests()
test . This never got exposed before as PUD based THP is available only
on X86 platform where mm_pmd_folded(mm) call gets macro replaced without
requiring the mm_struct i.e __is_defined(__PAGETABLE_PMD_FOLDED).
Link: https://lkml.kernel.org/r/1611137241-26220-3-git-send-email-anshuman.khandual@arm.com
Signed-off-by: Anshuman Khandual <anshuman.khandual@arm.com>
Tested-by: Gerald Schaefer <gerald.schaefer@de.ibm.com> [s390]
Reviewed-by: Steven Price <steven.price@arm.com>
Suggested-by: Catalin Marinas <catalin.marinas@arm.com>
Cc: Christophe Leroy <christophe.leroy@csgroup.eu>
Cc: Gerald Schaefer <gerald.schaefer@linux.ibm.com>
Cc: Paul Walmsley <paul.walmsley@sifive.com>
Cc: Vineet Gupta <vgupta@synopsys.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
In deactivate_slab() we currently move all but one objects on the cpu
freelist to the page freelist one by one using the costly cmpxchg_double()
operation. Then we unfreeze the page while moving the last object on page
freelist, with a final cmpxchg_double().
This can be optimized to avoid the cmpxchg_double() per object. Just
count the objects on cpu freelist (to adjust page->inuse properly) and
also remember the last object in the chain. Then splice page->freelist to
the last object and effectively add the whole cpu freelist to
page->freelist while unfreezing the page, with a single cmpxchg_double().
Link: https://lkml.kernel.org/r/20210115183543.15097-1-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: Jann Horn <jannh@google.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
SLAB has been using get/put_online_cpus() around creating, destroying and
shrinking kmem caches since 95402b3829 ("cpu-hotplug: replace
per-subsystem mutexes with get_online_cpus()") in 2008, which is supposed
to be replacing a private mutex (cache_chain_mutex, called slab_mutex
today) with system-wide mechanism, but in case of SLAB it's in fact used
in addition to the existing mutex, without explanation why.
SLUB appears to have avoided the cpu hotplug lock initially, but gained it
due to common code unification, such as 20cea9683e ("mm, sl[aou]b: Move
kmem_cache_create mutex handling to common code").
Regardless of the history, checking if the hotplug lock is actually needed
today suggests that it's not, and therefore it's better to avoid this
system-wide lock and the ordering this imposes wrt other locks (such as
slab_mutex).
Specifically, in SLAB we have for_each_online_cpu() in do_tune_cpucache()
protected by slab_mutex, and cpu hotplug callbacks that also take the
slab_mutex, which is also taken by the common slab function that currently
also take the hotplug lock. Thus the slab_mutex protection should be
sufficient. Also per-cpu array caches are allocated for each possible
cpu, so not affected by their online/offline state.
In SLUB we have for_each_online_cpu() in functions that show statistics
and are already unprotected today, as racing with hotplug is not harmful.
Otherwise SLUB relies on percpu allocator. The slub_cpu_dead() hotplug
callback takes the slab_mutex.
To sum up, this patch removes get/put_online_cpus() calls from slab as it
should be safe without further adjustments.
Link: https://lkml.kernel.org/r/20210113131634.3671-4-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Qian Cai <cai@redhat.com>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Since commit 03afc0e25f ("slab: get_online_mems for
kmem_cache_{create,destroy,shrink}") we are taking memory hotplug lock for
SLAB and SLUB when creating, destroying or shrinking a cache. It is quite
a heavy lock and it's best to avoid it if possible, as we had several
issues with lockdep complaining about ordering in the past, see e.g.
e4f8e513c3 ("mm/slub: fix a deadlock in show_slab_objects()").
The problem scenario in 03afc0e25f (solved by the memory hotplug lock)
can be summarized as follows: while there's slab_mutex synchronizing new
kmem cache creation and SLUB's MEM_GOING_ONLINE callback
slab_mem_going_online_callback(), we may miss creation of kmem_cache_node
for the hotplugged node in the new kmem cache, because the hotplug
callback doesn't yet see the new cache, and cache creation in
init_kmem_cache_nodes() only inits kmem_cache_node for nodes in the
N_NORMAL_MEMORY nodemask, which however may not yet include the new node,
as that happens only later after the MEM_GOING_ONLINE callback.
Instead of using get/put_online_mems(), the problem can be solved by SLUB
maintaining its own nodemask of nodes for which it has allocated the
per-node kmem_cache_node structures. This nodemask would generally mirror
the N_NORMAL_MEMORY nodemask, but would be updated only in under SLUB's
control in its memory hotplug callbacks under the slab_mutex. This patch
adds such nodemask and its handling.
Commit 03afc0e25f mentiones "issues like [the one above]", but there
don't appear to be further issues. All the paths (shared for SLAB and
SLUB) taking the memory hotplug locks are also taking the slab_mutex,
except kmem_cache_shrink() where 03afc0e25f replaced slab_mutex with
get/put_online_mems().
We however cannot simply restore slab_mutex in kmem_cache_shrink(), as
SLUB can enters the function from a write to sysfs 'shrink' file, thus
holding kernfs lock, and in kmem_cache_create() the kernfs lock is nested
within slab_mutex. But on closer inspection we don't actually need to
protect kmem_cache_shrink() from hotplug callbacks: While SLUB's
__kmem_cache_shrink() does for_each_kmem_cache_node(), missing a new node
added in parallel hotplug is not fatal, and parallel hotremove does not
free kmem_cache_node's anymore after the previous patch, so use-after free
cannot happen. The per-node shrinking itself is protected by
n->list_lock. Same is true for SLAB, and SLOB is no-op.
SLAB also doesn't need the memory hotplug locking, which it only gained by
03afc0e25f through the shared paths in slab_common.c. Its memory
hotplug callbacks are also protected by slab_mutex against races with
these paths. The problem of SLUB relying on N_NORMAL_MEMORY doesn't apply
to SLAB, as its setup_kmem_cache_nodes relies on N_ONLINE, and the new
node is already set there during the MEM_GOING_ONLINE callback, so no
special care is needed for SLAB.
As such, this patch removes all get/put_online_mems() usage by the slab
subsystem.
Link: https://lkml.kernel.org/r/20210113131634.3671-3-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Qian Cai <cai@redhat.com>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Patch series "mm, slab, slub: remove cpu and memory hotplug locks".
Some related work caused me to look at how we use get/put_mems_online()
and get/put_online_cpus() during kmem cache
creation/descruction/shrinking, and realize that it should be actually
safe to remove all of that with rather small effort (as e.g. Michal Hocko
suspected in some of the past discussions already). This has the benefit
to avoid rather heavy locks that have caused locking order issues already
in the past. So this is the result, Patches 2 and 3 remove memory hotplug
and cpu hotplug locking, respectively. Patch 1 is due to realization that
in fact some races exist despite the locks (even if not removed), but the
most sane solution is not to introduce more of them, but rather accept
some wasted memory in scenarios that should be rare anyway (full memory
hot remove), as we do the same in other contexts already.
This patch (of 3):
Commit e4f8e513c3 ("mm/slub: fix a deadlock in show_slab_objects()") has
fixed a problematic locking order by removing the memory hotplug lock
get/put_online_mems() from show_slab_objects(). During the discussion, it
was argued [1] that this is OK, because existing slabs on the node would
prevent a hotremove to proceed.
That's true, but per-node kmem_cache_node structures are not necessarily
allocated on the same node and may exist even without actual slab pages on
the same node. Any path that uses get_node() directly or via
for_each_kmem_cache_node() (such as show_slab_objects()) can race with
freeing of kmem_cache_node even with the !NULL check, resulting in
use-after-free.
To that end, commit e4f8e513c3 argues in a comment that:
* We don't really need mem_hotplug_lock (to hold off
* slab_mem_going_offline_callback) here because slab's memory hot
* unplug code doesn't destroy the kmem_cache->node[] data.
While it's true that slab_mem_going_offline_callback() doesn't free the
kmem_cache_node, the later callback slab_mem_offline_callback() actually
does, so the race and use-after-free exists. Not just for
show_slab_objects() after commit e4f8e513c3, but also many other places
that are not under slab_mutex. And adding slab_mutex locking or other
synchronization to SLUB paths such as get_any_partial() would be bad for
performance and error-prone.
The easiest solution is therefore to make the abovementioned comment true
and stop freeing the kmem_cache_node structures, accepting some wasted
memory in the full memory node removal scenario. Analogically we also
don't free hotremoved pgdat as mentioned in [1], nor the similar per-node
structures in SLAB. Importantly this approach will not block the
hotremove, as generally such nodes should be movable in order to succeed
hotremove in the first place, and thus the GFP_KERNEL allocated
kmem_cache_node will come from elsewhere.
[1] https://lore.kernel.org/linux-mm/20190924151147.GB23050@dhcp22.suse.cz/
Link: https://lkml.kernel.org/r/20210113131634.3671-1-vbabka@suse.cz
Link: https://lkml.kernel.org/r/20210113131634.3671-2-vbabka@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Qian Cai <cai@redhat.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Currently, a trace record generated by the RCU core is as below.
... kmem_cache_free: call_site=rcu_core+0x1fd/0x610 ptr=00000000f3b49a66
It doesn't tell us what the RCU core has freed.
This patch adds the slab name to trace_kmem_cache_free().
The new format is as follows.
... kmem_cache_free: call_site=rcu_core+0x1fd/0x610 ptr=0000000037f79c8d name=dentry
... kmem_cache_free: call_site=rcu_core+0x1fd/0x610 ptr=00000000f78cb7b5 name=sock_inode_cache
... kmem_cache_free: call_site=rcu_core+0x1fd/0x610 ptr=0000000018768985 name=pool_workqueue
... kmem_cache_free: call_site=rcu_core+0x1fd/0x610 ptr=000000006a6cb484 name=radix_tree_node
We can use it to understand what the RCU core is going to free. For
example, some users maybe interested in when the RCU core starts
freeing reclaimable slabs like dentry to reduce memory pressure.
Link: https://lkml.kernel.org/r/20201216072804.8838-1-jian.w.wen@oracle.com
Signed-off-by: Jacob Wen <jian.w.wen@oracle.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
