The nr_deadline_tasks variable in the cpuset structure was introduced by
commit 6c24849f55 ("sched/cpuset: Keep track of SCHED_DEADLINE task
in cpusets"). It is reported by sashiko [1] that nr_deadline_tasks
can currently be modified by inc_dl_tasks_cs() under rq->lock and
by cpuset_attach() under cpuset_mutex. So if both updates happen
simultaneously, the nr_deadline_tasks variable can be corrupted leading
to incorrect operations down the road.
Fix that by changing its type to atomic_t so that nr_deadline_tasks
are always atomically updated. This fix patch is a low hanging fruit.
It can handle some of the races between a concurrent sched_setscheduler()
and cpuset_can_attach()/cpuset_attach() calls, but not all of them like
the other issue raised by sashiko [2]. This will be handled hopefully
in a future follow up patch.
[1] https://sashiko.dev/#/patchset/20260626181923.133658-1-longman%40redhat.com
[2] https://sashiko.dev/#/patchset/20260630033344.352702-1-longman%40redhat.com
Fixes: 6c24849f55 ("sched/cpuset: Keep track of SCHED_DEADLINE task in cpusets")
Reviewed-by: Ridong Chen <ridong.chen@linux.dev>
Signed-off-by: Waiman Long <longman@redhat.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
In order to avoid the average CPU fraction avg(F_g_n) becoming tiny '1/N',
assume each cgroup is maximally concurrent and distrubute 'N*weight', such
that:
F_g_n' = N * F_g_n
Giving:
avg(F_g_n') = N*avg(F_g_n) ~ N * 1/N = 1
And while this sounds like it solves things, remember what that ~ meant. There
is the corner case when a cgroup is minimally loaded, eg a single runnable
task, therefore limit the CPU fraction to that of a nice -20 task to avoid
getting too much load.
This last bit is what makes it different from a previous proposal to allow
raising cpu.weight to '100 * N', that would not limit the mininal concurrency
case and results in a very large F_g_n. And just like F_g_n << 1 is
problematic, so is F_g_n >> 1 for the exact same reasons (it would drown the
kthreads, but it also risks overflowing the load values).
So while this might appear to be a better scheme than the current default
scheme, it doesn't really handle less than maximal concurrency nicely -- it
clips and introduces artificially large weights. So where the traditional SMP
mode works well when nr_tasks << nr_cpus, MAX doesn't work well in that regime
and vice-versa.
The meaning of "cpu.weight" would be: weight per allowed CPU.
Included for completeness (and infrastructure).
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://patch.msgid.link/20260605124051.589618504%40infradead.org
As reported by sashiko [1], cpuset_update_tasks_nodemask() will do
mpol_rebind_mm() and possibly cpuset_migrate_mm() for all threads of
a multithreaded process. Since commit 3df9ca0a2b ("cpuset: migrate
memory only for threadgroup leaders"), cpuset_attach() had been updated
to rebind and migrate memory only for threadgroup leaders to mark the
group leader as the owner of the mm_struct.
To be consistent and avoid unnecessary performance overhead for heavily
multithreaded processes, follow the cpuset_attach() example and perform
memory rebind and migration only for threadgroup leaders.
Also add a paragraph in cgroup-v2.rst under cpuset.mems that the
threadgroup leader is the memory owner of that threadgroup. Therefore
the non-leading threads shouldn't be in other cgroups whose "cpuset.mems"
doesn't fully overlap that of the group leader.
[1] https://sashiko.dev/#/patchset/20260621032816.1806773-1-longman%40redhat.com
Signed-off-by: Waiman Long <longman@redhat.com>
Reviewed-by: Ridong Chen <ridong.chen@linux.dev>
Signed-off-by: Tejun Heo <tj@kernel.org>
As reported by sashiko [1], cpuset_hotplug_update_tasks() may perform
unnecessary task iteration and updating of tasks' CPU and node masks
when mems_allowed and/or cpus_allowed are not set in cpuset v2. It is
due to the fact that the temporary new_cpus and new_mems masks do not
inherit parent's effective_cpus/mems when they are empty which is the
expected behavior for cpuset v2 since commit 4ec22e9c5a ("cpuset:
Enable cpuset controller in default hierarchy").
Fix that and avoid unnecessary work by enhancing
compute_effective_cpumask() to add the empty cpumask check
and inheriting the parent's versions if empty when in v2. A new
compute_effective_nodemask() helper is also added to perform a similar
function for new effective_mems.
Add new test_cpuset_prs.sh test cases to confirm that effective_cpus
will inherit the parent's version if cpuset.cpus is empty.
[1] https://sashiko.dev/#/patchset/20260621032816.1806773-1-longman%40redhat.com
Suggested-by: Ridong Chen <ridong.chen@linux.dev>
Fixes: 4ec22e9c5a ("cpuset: Enable cpuset controller in default hierarchy")
Signed-off-by: Waiman Long <longman@redhat.com>
Reviewed-by: Ridong Chen <ridong.chen@linux.dev>
Signed-off-by: Tejun Heo <tj@kernel.org>
Pull cgroup updates from Tejun Heo:
- Last cycle deferred css teardown on cgroup removal until the cgroup
depopulated, so a css is not taken offline while tasks can still
reference it. Disabling a controller through cgroup.subtree_control
still had the same problem. This reworks the deferral from per-cgroup
to per-css so that path is covered too.
- New RDMA controller monitoring files: rdma.peak for per-device peak
usage and rdma.events / rdma.events.local for resource-limit
exhaustion. The max-limit parser was rewritten, fixing two input
parsing bugs.
- cpuset: fix a sched-domain leak on the domain-rebuild failure path
and skip a redundant hardwall ancestor scan on v2.
- Misc: pair the remaining lockless cgroup.max.* reads with WRITE_ONCE,
assorted selftest robustness fixes, and doc path corrections.
* tag 'cgroup-for-7.2' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/cgroup: (22 commits)
cgroup: Migrate tasks to the root css when a controller is rebound
docs: cgroup: Fix stale source file paths
cgroup/cpuset: Free sched domains on rebuild guard failure
cgroup: pair max limit READ_ONCE() with WRITE_ONCE()
selftests/cgroup: enable memory controller in hugetlb memcg test
cgroup/rdma: Drop unnecessary READ_ONCE() on event counters
cgroup: Defer kill_css_finish() in cgroup_apply_control_disable()
cgroup: Add per-subsys-css kill_css_finish deferral
cgroup: Move populated counters to cgroup_subsys_state
cgroup: Annotate unlocked nr_populated_* accesses with READ_ONCE/WRITE_ONCE
cgroup: Inline cgroup_has_tasks() in cgroup.h
cgroup/rdma: document rdma.peak, rdma.events and rdma.events.local
cgroup/rdma: add rdma.events.local for per-cgroup allocation failure attribution
cgroup/rdma: add rdma.events to track resource limit exhaustion
cgroup/rdma: add rdma.peak for per-device peak usage tracking
selftests/cgroup: check malloc return value in alloc_anon functions
cgroup/cpuset: Skip hardwall ancestor scan in cpuset v2 in cpuset_current_node_allowed()
selftests/cgroup: fix misleading debug message in test_cgfreezer_time_child
selftests/cgroup: fix child process escaping to parent cleanup in test_cpucg_nice
selftests/cgroup: Add NULL check after malloc in cgroup_util.c
...
Just like we have memory.peak, introduce a dmem.peak, which uses the
page_counter support for that.
For now, make it read-only.
This allows for memory usage monitoring without polling dmem.current when
the information needed is the maximum device memory used. That can be used
for capacity planning, such that dmem.max can be properly setup for a given
workload. It can also be used for debugging to determine whether a given
workload would have caused eviction or system memory use.
Signed-off-by: Thadeu Lima de Souza Cascardo <cascardo@igalia.com>
Reviewed-by: Michal Koutný <mkoutny@suse.com>
Reviewed-by: Maarten Lankhorst <dev@lankhorst.se>
Link: https://patch.msgid.link/20260514-dmem_peak-v3-1-b64ce5d3ac38@igalia.com
Signed-off-by: Maarten Lankhorst <dev@lankhorst.se>
Pull cgroup fixes from Tejun Heo:
"One cpuset fix and a maintenance update, both low-risk:
- Fix cpuset partition CPU accounting under sibling CPU exclusion
that could produce wrong CPU assignments and trigger
scheduling-domain warnings. Includes selftests.
- Update an email address in MAINTAINERS"
* tag 'cgroup-for-7.1-rc6-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/cgroup:
cgroup/cpuset: Change Ridong's email
cgroup/cpuset: Add test cases for sibling CPU exclusion on partition update
cgroup/cpuset: Use effective_xcpus in partcmd_update add/del mask calculation
cgroup_apply_control_disable() defers kill_css_finish() while a css is
still populated, relying on css_update_populated() to fire the deferred
kill once the populated count reaches zero.
This deadlocks when a controller is rebound out of a hierarchy. Mounting
an implicit_on_dfl controller such as perf_event as a v1 hierarchy steals
it off the default hierarchy, and rebind_subsystems() kills its
per-cgroup csses while they are still populated. The migration run in the
same step keeps the old css for a controller no longer in the hierarchy's
mask, so no task is migrated off the dying csses. Their populated count
never reaches zero, the deferred kill_css_finish() never fires, and the
next cgroup_lock_and_drain_offline() hangs forever under cgroup_mutex.
That migration is already a no-op pass over the rebound subtree. Add
cgroup_rebind_ss_mask so find_existing_css_set() resolves the leaving
controllers to the root css. Their tasks are migrated there, the
per-cgroup csses depopulate, and cgroup_apply_control_disable() kills
them synchronously. The deferral stays correct for the rmdir and
controller-disable paths it was meant for.
Fixes: 1dffd95575 ("cgroup: Defer kill_css_finish() in cgroup_apply_control_disable()")
Reported-by: Mark Brown <broonie@kernel.org>
Closes: https://lore.kernel.org/all/41cd159c-54e5-45e0-81df-eaf36a6c028e@sirena.org.uk/
Reported-by: Bert Karwatzki <spasswolf@web.de>
Closes: https://lore.kernel.org/all/4e986b4ed7e16547805d54b6e67d09120bc4d2f2.camel@web.de/
Tested-by: Mark Brown <broonie@kernel.org>
Tested-by: Bert Karwatzki <spasswolf@web.de>
Signed-off-by: Tejun Heo <tj@kernel.org>
generate_sched_domains() returns sched-domain masks and optional
attributes that are normally handed to partition_sched_domains(), which
takes ownership of them.
rebuild_sched_domains_locked() has a WARN guard after
generate_sched_domains() and before partition_sched_domains() to avoid
passing offline CPUs into the scheduler domain rebuild path. If that
guard fires, the function currently returns directly without freeing
the generated doms and attr.
Free the generated sched-domain masks and attributes before returning
from the guard failure path.
Signed-off-by: Guopeng Zhang <zhangguopeng@kylinos.cn>
Reviewed-by: Waiman Long <longman@redhat.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
cgroup.max.descendants and cgroup.max.depth are shown through seq_file.
Their show callbacks read cgrp->max_descendants and cgrp->max_depth with
READ_ONCE(), respectively.
The corresponding write callbacks update the same scalar fields while
holding the cgroup lock, but the seq_file show path does not serialize
against those stores. This leaves the lockless show-side loads annotated
with READ_ONCE(), while the corresponding stores remain plain stores.
Use WRITE_ONCE() for the updates so the intended lockless access is marked
consistently on both sides. This does not change locking, ordering, or
user-visible semantics.
Assisted-by: OpenAI-Codex:gpt-5.5
Signed-off-by: Ren Tamura <ren.tamura.oss@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
When sibling CPU exclusion occurs, a partition's user_xcpus may contain
CPUs that were never actually granted to it. These CPUs are present in
user_xcpus(cs) but not in cs->effective_xcpus.
The partcmd_update path in update_parent_effective_cpumask() uses
user_xcpus(cs) (via the local variable xcpus) to compute the addmask
(CPUs to return to parent) and delmask (CPUs to request from parent).
This is incorrect:
1) When newmask removes a CPU that was previously excluded by a
sibling, addmask incorrectly includes that CPU and tries to return
it to the parent even though the partition never actually owned it,
causing CPU overlap with sibling partitions and triggering warnings
in generate_sched_domains().
2) When newmask adds a previously excluded CPU that is now available,
delmask fails to request it from the parent because user_xcpus(cs)
already includes it.
Fix this by using cs->effective_xcpus instead of user_xcpus(cs) in all
partcmd_update paths that calculate addmask or delmask, including the
PERR_NOCPUS error handling paths.
Reproducers:
Example 1 - Removing a sibling-excluded CPU incorrectly returns it:
# cd /sys/fs/cgroup
# echo "0-1" > a1/cpuset.cpus
# echo "root" > a1/cpuset.cpus.partition
# echo "0-2" > b1/cpuset.cpus
# echo "root" > b1/cpuset.cpus.partition
# echo "2" > b1/cpuset.cpus
# cat cpuset.cpus.effective
# Actual: 0-1,3 Expected: 3
Example 2 - Expanding to a previously excluded CPU fails to request it:
# cd /sys/fs/cgroup
# echo "0-1" > a1/cpuset.cpus
# echo "root" > a1/cpuset.cpus.partition
# echo "0-2" > b1/cpuset.cpus
# echo "root" > b1/cpuset.cpus.partition
# echo "member" > a1/cpuset.cpus.partition
# echo "1-2" > b1/cpuset.cpus
# cat cpuset.cpus.effective
# Actual: 0-1,3 Expected: 0,3
Fixes: 2a3602030d ("cgroup/cpuset: Don't invalidate sibling partitions on cpuset.cpus conflict")
Cc: stable@vger.kernel.org # v7.0+
Suggested-by: Zhang Guopeng <zhangguopeng@kylinos.cn>
Signed-off-by: Sun Shaojie <sunshaojie@kylinos.cn>
Reviewed-by: Waiman Long <longman@redhat.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Pull cgroup fixes from Tejun Heo:
"Two rstat fixes:
- Out-of-bounds access in the css_rstat_updated() BPF kfunc when
called with an unchecked user-supplied cpu
- Over-strict NMI guard after the recent switch to try_cmpxchg left
sparc and ppc64 unable to queue rstat updates from NMI"
* tag 'cgroup-for-7.1-rc4-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/cgroup:
cgroup: rstat: relax NMI guard after switch to try_cmpxchg
cgroup/rstat: validate cpu before css_rstat_cpu() access
Commit 36df6e3dbd ("cgroup: make css_rstat_updated nmi safe") used
this_cpu_cmpxchg() for the lockless insertion, and therefore required
both ARCH_HAVE_NMI_SAFE_CMPXCHG and ARCH_HAS_NMI_SAFE_THIS_CPU_OPS in
the NMI guard: on archs without the latter, this_cpu_cmpxchg() falls
back to "local_irq_save() + plain cmpxchg", and local_irq_save()
cannot mask NMIs.
Commit 3309b63a22 ("cgroup: rstat: use LOCK CMPXCHG in
css_rstat_updated") later replaced this_cpu_cmpxchg() with plain
try_cmpxchg() to fix cross-CPU lockless-list corruption, but left the
NMI guard untouched. After that switch, css_rstat_updated() no longer
performs any this_cpu_*() RMW operations and only relies on the arch
having NMI-safe cmpxchg, so ARCH_HAS_NMI_SAFE_THIS_CPU_OPS is no
longer required in the guard.
Relax the guard accordingly so that archs which have HAVE_NMI and
ARCH_HAVE_NMI_SAFE_CMPXCHG but not ARCH_HAS_NMI_SAFE_THIS_CPU_OPS
(e.g. sparc, powerpc on PPC64/BOOK3S) can benefit from the existing
CONFIG_MEMCG_NMI_SAFETY_REQUIRES_ATOMIC path. Without this, the css
is never queued in NMI on those archs, and the atomics staged by
account_{slab,kmem}_nmi_safe() are not drained by flush_nmi_stats().
Fixes: 3309b63a22 ("cgroup: rstat: use LOCK CMPXCHG in css_rstat_updated")
Signed-off-by: Cunlong Li <shenxiaogll@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
css_rstat_updated() is exposed as a BPF kfunc and accepts a
caller-provided cpu argument. The function uses cpu for per-cpu rstat
lookups without checking whether it refers to a valid possible CPU.
A BPF iter/cgroup program with CAP_BPF and CAP_PERFMON can pass an
invalid cpu value. On an unfixed UBSCAN_BOUNDS test kernel, cpu ==
0x7fffffff triggers:
UBSAN: array-index-out-of-bounds in kernel/cgroup/rstat.c:31:9
index 2147483647 is out of range for type 'long unsigned int [64]'
Call Trace:
css_rstat_updated
bpf_iter_run_prog
cgroup_iter_seq_show
bpf_seq_read
Add cpu validation to the BPF-facing css_rstat_updated() kfunc and
move the common implementation to __css_rstat_updated() for in-kernel
callers.
Fixes: a319185be9 ("cgroup: bpf: enable bpf programs to integrate with rstat")
Signed-off-by: Qing Ming <a0yami@mailbox.org>
Signed-off-by: Tejun Heo <tj@kernel.org>
All accesses to the event counters are serialized by rdmacg_mutex,
making the READ_ONCE() annotations unnecessary. Remove them.
Signed-off-by: Tao Cui <cuitao@kylinos.cn>
Signed-off-by: Tejun Heo <tj@kernel.org>
Same race shape as the rmdir path that 93618edf75 ("cgroup: Defer css
percpu_ref kill on rmdir until cgroup is depopulated") fixed: a task past
exit_signals() whose cset subsys[ssid] still pins the disabled controller's
css can be touching subsys state while ->css_offline() runs. The earlier
patches in this series built up the per-subsys-css deferral machinery and
routed cgroup_destroy_locked() through it. Apply the same shape to
cgroup_apply_control_disable():
kill_css_sync(css);
if (!css_is_populated(css))
kill_css_finish(css);
When the dying css is still populated, kill_css_finish() is deferred. The
walker in css_update_populated() fires kill_finish_work once the css's
hierarchical populated count drops to zero.
cgroup_lock_and_drain_offline()'s wait predicate switches from
percpu_ref_is_dying() to css_is_dying(). CSS_DYING is set by kill_css_sync()
and is a strict superset of percpu_ref_is_dying. Without this change, a +cpu
re-enable after a deferred -cpu disable would skip the drain (percpu_ref
isn't killed yet) and observe the still-CSS_DYING css through cgroup_css(),
treating it as live.
Signed-off-by: Tejun Heo <tj@kernel.org>
93618edf75 ("cgroup: Defer css percpu_ref kill on rmdir until cgroup is
depopulated") deferred kill_css_finish() at the cgroup level: rmdir waits
for the entire cgroup's populated count to drop to zero, then fires
kill_css_finish() on every subsystem css at once. Replace that with
per-subsys-css deferral. Each subsystem css now tracks its own hierarchical
populated count and independently defers its kill_css_finish() until its own
subtree drains.
The rmdir-race fix carries through unchanged in shape. The dying css's
->css_offline() still waits until no PF_EXITING task references it, and v2's
cgroup-level machinery goes away.
cgroup_apply_control_disable() has the same race shape (PF_EXITING tasks
pinning a css whose ->css_offline() is about to run) and stays synchronous
here. This patch lays the groundwork for fixing it - per-cgroup waiting
can't gate one subsys css being killed while the rest of the cgroup stays
live, but per-css can.
Subtree-wide invariant preserved: a dying ancestor css stays populated
through nr_populated_children until every dying descendant's task drains, so
the walker fires the ancestor's kill_finish_work only after all descendants
have drained.
Add paired smp_mb()s in kill_css_sync() and css_update_populated() to fence
the StoreLoad on (CSS_DYING, populated counter), guaranteeing that either
the walker queues kill_finish_work or the caller fires synchronously.
cgroup_destroy_locked() was implicitly fenced by an unrelated css_set_lock
pair; cgroup_apply_control_disable() in the next patch is not.
Signed-off-by: Tejun Heo <tj@kernel.org>
Later patches replace the cgroup-level finish_destroy_work deferral added
by 93618edf75 ("cgroup: Defer css percpu_ref kill on rmdir until cgroup
is depopulated") with a per-subsys-css deferral. That needs each subsystem
css to track its own populated count. Move the populated counters from
cgroup onto cgroup_subsys_state. cgroup->self is itself a
cgroup_subsys_state and self.parent walks the same chain as cgroup_parent(),
so cgroup_update_populated() generalizes to a single css_update_populated()
taking a css. The cgroup-side bookkeeping runs only when the walk started
from a self css.
Keep nr_populated_{domain,threaded}_children on cgroup. Both sum to
self.nr_populated_children, but staying as dedicated fields to allow readers
like cgroup_can_be_thread_root() unlocked access.
css_set_update_populated() also walks the per-subsys-css chain so each
subsystem css's hierarchical populated count is maintained. No reader
consumes those counts yet.
Signed-off-by: Tejun Heo <tj@kernel.org>
cgroup_update_populated() updates nr_populated_csets,
nr_populated_domain_children, and nr_populated_threaded_children under
css_set_lock, but cgroup_has_tasks(), cgroup_is_populated(), and
cgroup_can_be_thread_root() read them without holding it. Use
READ_ONCE/WRITE_ONCE.
Signed-off-by: Tejun Heo <tj@kernel.org>
cpuset reads cs->css.cgroup->nr_populated_csets directly in two places to
test whether a cgroup has tasks. cgroup.c already has a matching helper,
cgroup_has_tasks(). Move it to cgroup.h as static inline and use that
instead. This is to prepare for relocation of cgroup->nr_populated_csets. No
semantic change.
Signed-off-by: Tejun Heo <tj@kernel.org>
Add per-cgroup local event counters to track RDMA resource limit
exhaustion from the perspective of individual cgroups. The
rdma.events.local file reports two per-resource counters:
- max: number of times this cgroup's limit was the one that blocked
an allocation in the subtree
- alloc_fail: number of allocation attempts originating from this
cgroup that failed due to an ancestor's limit
This mirrors the design of pids.events.local, where events are
attributed to the cgroup that imposed the limit, not necessarily the
cgroup where the allocation was attempted.
Also extend rdma.events with a hierarchical alloc_fail counter that
tracks allocation failures propagating upward from the requesting
cgroup, complementing the existing max counter, so that rdma.events
and rdma.events.local share the same output format.
Signed-off-by: Tao Cui <cuitao@kylinos.cn>
Signed-off-by: Tejun Heo <tj@kernel.org>
Add per-device hierarchical event counters to track when RDMA resource
limits are exceeded. The rdma.events file reports max event counts
propagated upward from the cgroup whose limit was hit to all ancestors.
This mirrors the design of pids.events, where events are attributed to
the cgroup that imposed the limit, not necessarily the cgroup where the
allocation was attempted. Userspace can monitor this file via
poll/epoll for real-time notification of resource exhaustion.
Signed-off-by: Tao Cui <cuitao@kylinos.cn>
Signed-off-by: Tejun Heo <tj@kernel.org>
Waiman LongandTejun Heo
Farhad AlemiandTejun Heo
Peter Zijlstra
Thomas Zimmermann
Linus Torvalds
Thadeu Lima de Souza CascardoandMaarten Lankhorst
Guopeng ZhangandTejun Heo
Ren TamuraandTejun Heo
Sun ShaojieandTejun Heo
Cunlong LiandTejun Heo
Qing MingandTejun Heo
Tao CuiandTejun Heo