Patch series "add detect count for hung tasks", v2.
This patchset adds a counter, hung_task_detect_count, to track the number
of times hung tasks are detected.
IHMO, hung tasks are a critical metric. Currently, we detect them by
periodically parsing dmesg. However, this method isn't as user-friendly
as using a counter.
Sometimes, a short-lived issue with NIC or hard drive can quickly decrease
the hung_task_warnings to zero. Without warnings, we must directly access
the node to ensure that there are no more hung tasks and that the system
has recovered. After all, load average alone cannot provide a clear
picture.
Once this counter is in place, in a high-density deployment pattern, we
plan to set hung_task_timeout_secs to a lower number to improve stability,
even though this might result in false positives. And then we can set a
time-based threshold: if hung tasks last beyond this duration, we will
automatically migrate containers to other nodes. Based on past
experience, this approach could help avoid many production disruptions.
Moreover, just like other important events such as OOM that already have
counters, having a dedicated counter for hung tasks makes sense ;)
This patch (of 2):
This commit adds a counter, hung_task_detect_count, to track the number of
times hung tasks are detected.
IHMO, hung tasks are a critical metric. Currently, we detect them by
periodically parsing dmesg. However, this method isn't as user-friendly as
using a counter.
Sometimes, a short-lived issue with NIC or hard drive can quickly decrease
the hung_task_warnings to zero. Without warnings, we must directly access
the node to ensure that there are no more hung tasks and that the system
has recovered. After all, load average alone cannot provide a clear
picture.
Once this counter is in place, in a high-density deployment pattern, we
plan to set hung_task_timeout_secs to a lower number to improve stability,
even though this might result in false positives. And then we can set a
time-based threshold: if hung tasks last beyond this duration, we will
automatically migrate containers to other nodes. Based on past experience,
this approach could help avoid many production disruptions.
Moreover, just like other important events such as OOM that already have
counters, having a dedicated counter for hung tasks makes sense.
[ioworker0@gmail.com: proc_doulongvec_minmax instead of proc_dointvec]
Link: https://lkml.kernel.org/r/20241101114833.8377-1-ioworker0@gmail.com
Link: https://lkml.kernel.org/r/20241027120747.42833-1-ioworker0@gmail.com
Link: https://lkml.kernel.org/r/20241027120747.42833-2-ioworker0@gmail.com
Signed-off-by: Mingzhe Yang <mingzhe.yang@ly.com>
Signed-off-by: Lance Yang <ioworker0@gmail.com>
Cc: Bang Li <libang.li@antgroup.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Huang Cun <cunhuang@tencent.com>
Cc: Joel Granados <j.granados@samsung.com>
Cc: Joel Granados <joel.granados@kernel.org>
Cc: John Siddle <jsiddle@redhat.com>
Cc: Kent Overstreet <kent.overstreet@linux.dev>
Cc: Ryan Roberts <ryan.roberts@arm.com>
Cc: Thomas Weißschuh <linux@weissschuh.net>
Cc: Yongliang Gao <leonylgao@tencent.com>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Pull sched_ext fixes from Tejun Heo:
- The fair sched class currently has a bug where its balance() returns
true telling the sched core that it has tasks to run but then NULL
from pick_task(). This makes sched core call sched_ext's pick_task()
without preceding balance() which can lead to stalls in partial mode.
For now, work around by detecting the condition and forcing the CPU
to go through another scheduling cycle.
- Add a missing newline to an error message and fix drgn introspection
tool which went out of sync.
* tag 'sched_ext-for-6.12-rc7-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/sched_ext:
sched_ext: Handle cases where pick_task_scx() is called without preceding balance_scx()
sched_ext: Update scx_show_state.py to match scx_ops_bypass_depth's new type
sched_ext: Add a missing newline at the end of an error message
* kvm-arm64/psci-1.3:
: PSCI v1.3 support, courtesy of David Woodhouse
:
: Bump KVM's PSCI implementation up to v1.3, with the added bonus of
: implementing the SYSTEM_OFF2 call. Like other system-scoped PSCI calls,
: this gets relayed to userspace for further processing with a new
: KVM_SYSTEM_EVENT_SHUTDOWN flag.
:
: As an added bonus, implement client-side support for hibernation with
: the SYSTEM_OFF2 call.
arm64: Use SYSTEM_OFF2 PSCI call to power off for hibernate
KVM: arm64: nvhe: Pass through PSCI v1.3 SYSTEM_OFF2 call
KVM: selftests: Add test for PSCI SYSTEM_OFF2
KVM: arm64: Add support for PSCI v1.2 and v1.3
KVM: arm64: Add PSCI v1.3 SYSTEM_OFF2 function for hibernation
firmware/psci: Add definitions for PSCI v1.3 specification
Signed-off-by: Oliver Upton <oliver.upton@linux.dev>
In sched_ext API, a repeatedly reported pain point is the overuse of the
verb "dispatch" and confusion around "consume":
- ops.dispatch()
- scx_bpf_dispatch[_vtime]()
- scx_bpf_consume()
- scx_bpf_dispatch[_vtime]_from_dsq*()
This overloading of the term is historical. Originally, there were only
built-in DSQs and moving a task into a DSQ always dispatched it for
execution. Using the verb "dispatch" for the kfuncs to move tasks into these
DSQs made sense.
Later, user DSQs were added and scx_bpf_dispatch[_vtime]() updated to be
able to insert tasks into any DSQ. The only allowed DSQ to DSQ transfer was
from a non-local DSQ to a local DSQ and this operation was named "consume".
This was already confusing as a task could be dispatched to a user DSQ from
ops.enqueue() and then the DSQ would have to be consumed in ops.dispatch().
Later addition of scx_bpf_dispatch_from_dsq*() made the confusion even worse
as "dispatch" in this context meant moving a task to an arbitrary DSQ from a
user DSQ.
Clean up the API with the following renames:
1. scx_bpf_dispatch[_vtime]() -> scx_bpf_dsq_insert[_vtime]()
2. scx_bpf_consume() -> scx_bpf_dsq_move_to_local()
3. scx_bpf_dispatch[_vtime]_from_dsq*() -> scx_bpf_dsq_move[_vtime]*()
This patch performs the third set of renames. Compatibility is maintained
by:
- The previous kfunc names are still provided by the kernel so that old
binaries can run. Kernel generates a warning when the old names are used.
- compat.bpf.h provides wrappers for the new names which automatically fall
back to the old names when running on older kernels. They also trigger
build error if old names are used for new builds.
- scx_bpf_dispatch[_vtime]_from_dsq*() were already wrapped in __COMPAT
macros as they were introduced during v6.12 cycle. Wrap new API in
__COMPAT macros too and trigger build errors on both __COMPAT prefixed and
naked usages of the old names.
The compat features will be dropped after v6.15.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Andrea Righi <arighi@nvidia.com>
Acked-by: Changwoo Min <changwoo@igalia.com>
Acked-by: Johannes Bechberger <me@mostlynerdless.de>
Acked-by: Giovanni Gherdovich <ggherdovich@suse.com>
Cc: Dan Schatzberg <dschatzberg@meta.com>
Cc: Ming Yang <yougmark94@gmail.com>
In sched_ext API, a repeatedly reported pain point is the overuse of the
verb "dispatch" and confusion around "consume":
- ops.dispatch()
- scx_bpf_dispatch[_vtime]()
- scx_bpf_consume()
- scx_bpf_dispatch[_vtime]_from_dsq*()
This overloading of the term is historical. Originally, there were only
built-in DSQs and moving a task into a DSQ always dispatched it for
execution. Using the verb "dispatch" for the kfuncs to move tasks into these
DSQs made sense.
Later, user DSQs were added and scx_bpf_dispatch[_vtime]() updated to be
able to insert tasks into any DSQ. The only allowed DSQ to DSQ transfer was
from a non-local DSQ to a local DSQ and this operation was named "consume".
This was already confusing as a task could be dispatched to a user DSQ from
ops.enqueue() and then the DSQ would have to be consumed in ops.dispatch().
Later addition of scx_bpf_dispatch_from_dsq*() made the confusion even worse
as "dispatch" in this context meant moving a task to an arbitrary DSQ from a
user DSQ.
Clean up the API with the following renames:
1. scx_bpf_dispatch[_vtime]() -> scx_bpf_dsq_insert[_vtime]()
2. scx_bpf_consume() -> scx_bpf_dsq_move_to_local()
3. scx_bpf_dispatch[_vtime]_from_dsq*() -> scx_bpf_dsq_move[_vtime]*()
This patch performs the second rename. Compatibility is maintained by:
- The previous kfunc names are still provided by the kernel so that old
binaries can run. Kernel generates a warning when the old names are used.
- compat.bpf.h provides wrappers for the new names which automatically fall
back to the old names when running on older kernels. They also trigger
build error if old names are used for new builds.
The compat features will be dropped after v6.15.
v2: Comment and documentation updates.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Andrea Righi <arighi@nvidia.com>
Acked-by: Changwoo Min <changwoo@igalia.com>
Acked-by: Johannes Bechberger <me@mostlynerdless.de>
Acked-by: Giovanni Gherdovich <ggherdovich@suse.com>
Cc: Dan Schatzberg <dschatzberg@meta.com>
Cc: Ming Yang <yougmark94@gmail.com>
In sched_ext API, a repeatedly reported pain point is the overuse of the
verb "dispatch" and confusion around "consume":
- ops.dispatch()
- scx_bpf_dispatch[_vtime]()
- scx_bpf_consume()
- scx_bpf_dispatch[_vtime]_from_dsq*()
This overloading of the term is historical. Originally, there were only
built-in DSQs and moving a task into a DSQ always dispatched it for
execution. Using the verb "dispatch" for the kfuncs to move tasks into these
DSQs made sense.
Later, user DSQs were added and scx_bpf_dispatch[_vtime]() updated to be
able to insert tasks into any DSQ. The only allowed DSQ to DSQ transfer was
from a non-local DSQ to a local DSQ and this operation was named "consume".
This was already confusing as a task could be dispatched to a user DSQ from
ops.enqueue() and then the DSQ would have to be consumed in ops.dispatch().
Later addition of scx_bpf_dispatch_from_dsq*() made the confusion even worse
as "dispatch" in this context meant moving a task to an arbitrary DSQ from a
user DSQ.
Clean up the API with the following renames:
1. scx_bpf_dispatch[_vtime]() -> scx_bpf_dsq_insert[_vtime]()
2. scx_bpf_consume() -> scx_bpf_dsq_move_to_local()
3. scx_bpf_dispatch[_vtime]_from_dsq*() -> scx_bpf_dsq_move[_vtime]*()
This patch performs the first set of renames. Compatibility is maintained
by:
- The previous kfunc names are still provided by the kernel so that old
binaries can run. Kernel generates a warning when the old names are used.
- compat.bpf.h provides wrappers for the new names which automatically fall
back to the old names when running on older kernels. They also trigger
build error if old names are used for new builds.
The compat features will be dropped after v6.15.
v2: Documentation updates.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Andrea Righi <arighi@nvidia.com>
Acked-by: Changwoo Min <changwoo@igalia.com>
Acked-by: Johannes Bechberger <me@mostlynerdless.de>
Acked-by: Giovanni Gherdovich <ggherdovich@suse.com>
Cc: Dan Schatzberg <dschatzberg@meta.com>
Cc: Ming Yang <yougmark94@gmail.com>
Logic to prevent callbacks from acquiring new references for the program
(i.e. leaving acquired references), and releasing caller references
(i.e. those acquired in parent frames) was introduced in commit
9d9d00ac29 ("bpf: Fix reference state management for synchronous callbacks").
This was necessary because back then, the verifier simulated each
callback once (that could potentially be executed N times, where N can
be zero). This meant that callbacks that left lingering resources or
cleared caller resources could do it more than once, operating on
undefined state or leaking memory.
With the fixes to callback verification in commit
ab5cfac139 ("bpf: verify callbacks as if they are called unknown number of times"),
all of this extra logic is no longer necessary. Hence, drop it as part
of this commit.
Cc: Eduard Zingerman <eddyz87@gmail.com>
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Link: https://lore.kernel.org/r/20241109231430.2475236-3-memxor@gmail.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
When bpf_spin_lock was introduced originally, there was deliberation on
whether to use an array of lock IDs, but since bpf_spin_lock is limited
to holding a single lock at any given time, we've been using a single ID
to identify the held lock.
In preparation for introducing spin locks that can be taken multiple
times, introduce support for acquiring multiple lock IDs. For this
purpose, reuse the acquired_refs array and store both lock and pointer
references. We tag the entry with REF_TYPE_PTR or REF_TYPE_LOCK to
disambiguate and find the relevant entry. The ptr field is used to track
the map_ptr or btf (for bpf_obj_new allocations) to ensure locks can be
matched with protected fields within the same "allocation", i.e.
bpf_obj_new object or map value.
The struct active_lock is changed to an int as the state is part of the
acquired_refs array, and we only need active_lock as a cheap way of
detecting lock presence.
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Link: https://lore.kernel.org/r/20241109231430.2475236-2-memxor@gmail.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
For htab of maps, when the map is removed from the htab, it may hold the
last reference of the map. bpf_map_fd_put_ptr() will invoke
bpf_map_free_id() to free the id of the removed map element. However,
bpf_map_fd_put_ptr() is invoked while holding a bucket lock
(raw_spin_lock_t), and bpf_map_free_id() attempts to acquire map_idr_lock
(spinlock_t), triggering the following lockdep warning:
=============================
[ BUG: Invalid wait context ]
6.11.0-rc4+ #49 Not tainted
-----------------------------
test_maps/4881 is trying to lock:
ffffffff84884578 (map_idr_lock){+...}-{3:3}, at: bpf_map_free_id.part.0+0x21/0x70
other info that might help us debug this:
context-{5:5}
2 locks held by test_maps/4881:
#0: ffffffff846caf60 (rcu_read_lock){....}-{1:3}, at: bpf_fd_htab_map_update_elem+0xf9/0x270
#1: ffff888149ced148 (&htab->lockdep_key#2){....}-{2:2}, at: htab_map_update_elem+0x178/0xa80
stack backtrace:
CPU: 0 UID: 0 PID: 4881 Comm: test_maps Not tainted 6.11.0-rc4+ #49
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), ...
Call Trace:
<TASK>
dump_stack_lvl+0x6e/0xb0
dump_stack+0x10/0x20
__lock_acquire+0x73e/0x36c0
lock_acquire+0x182/0x450
_raw_spin_lock_irqsave+0x43/0x70
bpf_map_free_id.part.0+0x21/0x70
bpf_map_put+0xcf/0x110
bpf_map_fd_put_ptr+0x9a/0xb0
free_htab_elem+0x69/0xe0
htab_map_update_elem+0x50f/0xa80
bpf_fd_htab_map_update_elem+0x131/0x270
htab_map_update_elem+0x50f/0xa80
bpf_fd_htab_map_update_elem+0x131/0x270
bpf_map_update_value+0x266/0x380
__sys_bpf+0x21bb/0x36b0
__x64_sys_bpf+0x45/0x60
x64_sys_call+0x1b2a/0x20d0
do_syscall_64+0x5d/0x100
entry_SYSCALL_64_after_hwframe+0x76/0x7e
One way to fix the lockdep warning is using raw_spinlock_t for
map_idr_lock as well. However, bpf_map_alloc_id() invokes
idr_alloc_cyclic() after acquiring map_idr_lock, it will trigger a
similar lockdep warning because the slab's lock (s->cpu_slab->lock) is
still a spinlock.
Instead of changing map_idr_lock's type, fix the issue by invoking
htab_put_fd_value() after htab_unlock_bucket(). However, only deferring
the invocation of htab_put_fd_value() is not enough, because the old map
pointers in htab of maps can not be saved during batched deletion.
Therefore, also defer the invocation of free_htab_elem(), so these
to-be-freed elements could be linked together similar to lru map.
There are four callers for ->map_fd_put_ptr:
(1) alloc_htab_elem() (through htab_put_fd_value())
It invokes ->map_fd_put_ptr() under a raw_spinlock_t. The invocation of
htab_put_fd_value() can not simply move after htab_unlock_bucket(),
because the old element has already been stashed in htab->extra_elems.
It may be reused immediately after htab_unlock_bucket() and the
invocation of htab_put_fd_value() after htab_unlock_bucket() may release
the newly-added element incorrectly. Therefore, saving the map pointer
of the old element for htab of maps before unlocking the bucket and
releasing the map_ptr after unlock. Beside the map pointer in the old
element, should do the same thing for the special fields in the old
element as well.
(2) free_htab_elem() (through htab_put_fd_value())
Its caller includes __htab_map_lookup_and_delete_elem(),
htab_map_delete_elem() and __htab_map_lookup_and_delete_batch().
For htab_map_delete_elem(), simply invoke free_htab_elem() after
htab_unlock_bucket(). For __htab_map_lookup_and_delete_batch(), just
like lru map, linking the to-be-freed element into node_to_free list
and invoking free_htab_elem() for these element after unlock. It is safe
to reuse batch_flink as the link for node_to_free, because these
elements have been removed from the hash llist.
Because htab of maps doesn't support lookup_and_delete operation,
__htab_map_lookup_and_delete_elem() doesn't have the problem, so kept
it as is.
(3) fd_htab_map_free()
It invokes ->map_fd_put_ptr without raw_spinlock_t.
(4) bpf_fd_htab_map_update_elem()
It invokes ->map_fd_put_ptr without raw_spinlock_t.
After moving free_htab_elem() outside htab bucket lock scope, using
pcpu_freelist_push() instead of __pcpu_freelist_push() to disable
the irq before freeing elements, and protecting the invocations of
bpf_mem_cache_free() with migrate_{disable|enable} pair.
Signed-off-by: Hou Tao <houtao1@huawei.com>
Link: https://lore.kernel.org/r/20241106063542.357743-2-houtao@huaweicloud.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Adding support to attach BPF program for entry and return probe
of the same function. This is common use case which at the moment
requires to create two uprobe multi links.
Adding new BPF_TRACE_UPROBE_SESSION attach type that instructs
kernel to attach single link program to both entry and exit probe.
It's possible to control execution of the BPF program on return
probe simply by returning zero or non zero from the entry BPF
program execution to execute or not the BPF program on return
probe respectively.
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Acked-by: Andrii Nakryiko <andrii@kernel.org>
Link: https://lore.kernel.org/bpf/20241108134544.480660-4-jolsa@kernel.org
Introduce FORCE_CON flag to printk. The new flag will make it possible to
create a context where printk messages will never be suppressed.
This mechanism will be used in the next patch to create a force_con
context on sysrq handling, removing an existing workaround on the
loglevel global variable. The workaround existed to make sure that sysrq
header messages were sent to all consoles, but this doesn't work with
deferred messages because the loglevel might be restored to its original
value before a console flushes the messages.
Signed-off-by: Marcos Paulo de Souza <mpdesouza@suse.com>
Reviewed-by: John Ogness <john.ogness@linutronix.de>
Reviewed-by: Petr Mladek <pmladek@suse.com>
Acked-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Link: https://lore.kernel.org/r/20241105-printk-loud-con-v2-1-bd3ecdf7b0e4@suse.com
Signed-off-by: Petr Mladek <pmladek@suse.com>
Add a light version of override/revert_creds(), this should only be
used when the credentials in question will outlive the critical
section and the critical section doesn't change the ->usage of the
credentials.
Suggested-by: Christian Brauner <brauner@kernel.org>
Signed-off-by: Vinicius Costa Gomes <vinicius.gomes@intel.com>
Signed-off-by: Amir Goldstein <amir73il@gmail.com>
Pick up e7ac4daeed ("mm: count zeromap read and set for swapout and
swapin") in order to move
mm: define obj_cgroup_get() if CONFIG_MEMCG is not defined
mm: zswap: modify zswap_compress() to accept a page instead of a folio
mm: zswap: rename zswap_pool_get() to zswap_pool_tryget()
mm: zswap: modify zswap_stored_pages to be atomic_long_t
mm: zswap: support large folios in zswap_store()
mm: swap: count successful large folio zswap stores in hugepage zswpout stats
mm: zswap: zswap_store_page() will initialize entry after adding to xarray.
mm: add per-order mTHP swpin counters
from mm-unstable into mm-stable.
Pull misc fixes from Andrew Morton:
"20 hotfixes, 14 of which are cc:stable.
Three affect DAMON. Lorenzo's five-patch series to address the
mmap_region error handling is here also.
Apart from that, various singletons"
* tag 'mm-hotfixes-stable-2024-11-09-22-40' of git://git.kernel.org/pub/scm/linux/kernel/git/akpm/mm:
mailmap: add entry for Thorsten Blum
ocfs2: remove entry once instead of null-ptr-dereference in ocfs2_xa_remove()
signal: restore the override_rlimit logic
fs/proc: fix compile warning about variable 'vmcore_mmap_ops'
ucounts: fix counter leak in inc_rlimit_get_ucounts()
selftests: hugetlb_dio: check for initial conditions to skip in the start
mm: fix docs for the kernel parameter ``thp_anon=``
mm/damon/core: avoid overflow in damon_feed_loop_next_input()
mm/damon/core: handle zero schemes apply interval
mm/damon/core: handle zero {aggregation,ops_update} intervals
mm/mlock: set the correct prev on failure
objpool: fix to make percpu slot allocation more robust
mm/page_alloc: keep track of free highatomic
mm: resolve faulty mmap_region() error path behaviour
mm: refactor arch_calc_vm_flag_bits() and arm64 MTE handling
mm: refactor map_deny_write_exec()
mm: unconditionally close VMAs on error
mm: avoid unsafe VMA hook invocation when error arises on mmap hook
mm/thp: fix deferred split unqueue naming and locking
mm/thp: fix deferred split queue not partially_mapped
In commit 24cc57d8fa ("padata: Honor the caller's alignment in case of
chunk_size 0"), the line 'ps.chunk_size = max(ps.chunk_size, 1ul)' was
added, making 'ps.chunk_size = 1U' redundant and never executed.
Signed-off-by: Zicheng Qu <quzicheng@huawei.com>
Acked-by: Daniel Jordan <daniel.m.jordan@oracle.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
sched_ext dispatches tasks from the BPF scheduler from balance_scx() and
thus every pick_task_scx() call must be preceded by balance_scx(). While
this usually holds, due to a bug, there are cases where the fair class's
balance() returns true indicating that it has tasks to run on the CPU and
thus terminating balance() calls but fails to actually find the next task to
run when pick_task() is called. In such cases, pick_task_scx() can be called
without preceding balance_scx().
Detect this condition using SCX_RQ_BAL_PENDING flags. If detected, keep
running the previous task if possible and avoid stalling from entering idle
without balancing.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/Ztj_h5c2LYsdXYbA@slm.duckdns.org
4c30f5ce4f ("sched_ext: Implement scx_bpf_dispatch[_vtime]_from_dsq()")
added four kfuncs for dispatching while iterating. They are allowed from the
dispatch and unlocked contexts but two of the kfuncs were only added in the
dispatch section. Add missing declarations in the unlocked section.
Signed-off-by: Tejun Heo <tj@kernel.org>
Fixes: 4c30f5ce4f ("sched_ext: Implement scx_bpf_dispatch[_vtime]_from_dsq()")
scf_handler() is used as a SMP function call. This function is always
invoked in IRQ-context even with forced-threading enabled. This function
frees memory which not allowed on PREEMPT_RT because the locking
underneath is using sleeping locks.
Add a per-CPU scf_free_pool where each SMP functions adds its memory to
be freed. This memory is then freed by scftorture_invoker() on each
iteration. On the majority of invocations the number of items is less
than five. If the thread sleeps/ gets delayed the number exceed 350 but
did not reach 400 in testing. These were the spikes during testing.
The bulk free of 64 pointers at once should improve the give-back if the
list grows. The list size is ~1.3 items per invocations.
Having one global scf_free_pool with one cleaning thread let the list
grow to over 10.000 items with 32 CPUs (again, spikes not the average)
especially if the CPU went to sleep. The per-CPU part looks like a good
compromise.
Reported-by: "Paul E. McKenney" <paulmck@kernel.org>
Closes: https://lore.kernel.org/lkml/41619255-cdc2-4573-a360-7794fc3614f7@paulmck-laptop/
Tested-by: Paul E. McKenney <paulmck@kernel.org>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Reviewed-by: Boqun Feng <boqun.feng@gmail.com>
Tested-by: Boqun Feng <boqun.feng@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Memory allocations can not happen within regions with explicit disabled
preemption PREEMPT_RT. The problem is that the locking structures
underneath are sleeping locks.
Move the memory allocation outside of the preempt-disabled section. Keep
the GFP_ATOMIC for the allocation to behave like a "ememergncy
allocation".
Tested-by: Paul E. McKenney <paulmck@kernel.org>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Reviewed-by: Boqun Feng <boqun.feng@gmail.com>
Tested-by: Boqun Feng <boqun.feng@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
The smp_call_function() needs to be invoked with the wait flag set to
wait until scf_cleanup_handler() is done. This ensures that all SMP
function calls, that have been queued earlier, complete at this point.
Tested-by: Paul E. McKenney <paulmck@kernel.org>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Reviewed-by: Boqun Feng <boqun.feng@gmail.com>
Tested-by: Boqun Feng <boqun.feng@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Update the comments in sched_ext_ops to clarify this table is for
a BPF scheduler and a userland scheduler should also rely on the
sched_ext_ops table through the BPF scheduler.
Signed-off-by: Changwoo Min <changwoo@igalia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
On 2 x Intel Sapphire Rapids machines with 224 logical CPUs, a poorly
behaving BPF scheduler can live-lock the system by making multiple CPUs bang
on the same DSQ to the point where soft-lockup detection triggers before
SCX's own watchdog can take action. It also seems possible that the machine
can be live-locked enough to prevent scx_ops_helper, which is an RT task,
from running in a timely manner.
Implement scx_softlockup() which is called when three quarters of
soft-lockup threshold has passed. The function immediately enables the ops
breather and triggers an ops error to initiate ejection of the BPF
scheduler.
The previous and this patch combined enable the kernel to reliably recover
the system from live-lock conditions that can be triggered by a poorly
behaving BPF scheduler on Intel dual socket systems.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Douglas Anderson <dianders@chromium.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
A poorly behaving BPF scheduler can live-lock the system by e.g. incessantly
banging on the same DSQ on a large NUMA system to the point where switching
to the bypass mode can take a long time. Turning on the bypass mode requires
dequeueing and re-enqueueing currently runnable tasks, if the DSQs that they
are on are live-locked, this can take tens of seconds cascading into other
failures. This was observed on 2 x Intel Sapphire Rapids machines with 224
logical CPUs.
Inject artifical delays while the bypass mode is switching to guarantee
timely completion.
While at it, move __scx_ops_bypass_lock into scx_ops_bypass() and rename it
to bypass_lock.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reported-by: Valentin Andrei <vandrei@meta.com>
Reported-by: Patrick Lu <patlu@meta.com>
Pull sched_ext/for-6.12-fixes to receive 0e7ffff1b8 ("scx: Fix raciness in
scx_ops_bypass()"). Planned updates for scx_ops_bypass() depends on it.
Signed-off-by: Tejun Heo <tj@kernel.org>
When the LLC and NUMA domains fully overlap, enabling both optimizations
in the built-in idle CPU selection policy is redundant, as it leads to
searching for an idle CPU within the same domain twice.
Likewise, if all online CPUs are within a single LLC domain, LLC
optimization is unnecessary.
Therefore, detect overlapping domains and enable topology optimizations
only when necessary.
Moreover, rely on the online CPUs for this detection logic, instead of
using the possible CPUs.
Fixes: 860a45219b ("sched_ext: Introduce NUMA awareness to the default idle selection policy")
Signed-off-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Prior to commit d646969055 ("Reimplement RLIMIT_SIGPENDING on top of
ucounts") UCOUNT_RLIMIT_SIGPENDING rlimit was not enforced for a class of
signals. However now it's enforced unconditionally, even if
override_rlimit is set. This behavior change caused production issues.
For example, if the limit is reached and a process receives a SIGSEGV
signal, sigqueue_alloc fails to allocate the necessary resources for the
signal delivery, preventing the signal from being delivered with siginfo.
This prevents the process from correctly identifying the fault address and
handling the error. From the user-space perspective, applications are
unaware that the limit has been reached and that the siginfo is
effectively 'corrupted'. This can lead to unpredictable behavior and
crashes, as we observed with java applications.
Fix this by passing override_rlimit into inc_rlimit_get_ucounts() and skip
the comparison to max there if override_rlimit is set. This effectively
restores the old behavior.
Link: https://lkml.kernel.org/r/20241104195419.3962584-1-roman.gushchin@linux.dev
Fixes: d646969055 ("Reimplement RLIMIT_SIGPENDING on top of ucounts")
Signed-off-by: Roman Gushchin <roman.gushchin@linux.dev>
Co-developed-by: Andrei Vagin <avagin@google.com>
Signed-off-by: Andrei Vagin <avagin@google.com>
Acked-by: Oleg Nesterov <oleg@redhat.com>
Acked-by: Alexey Gladkov <legion@kernel.org>
Cc: Kees Cook <kees@kernel.org>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
While PREEMPT_RT is undoubtedly totally awesome, it does not, at this
time, make sense to have all{yes,mod}config select it.
Reported-by: Stephen Rothwell <sfr@canb.auug.org.au>
Fixes: 35772d627b ("sched: Enable PREEMPT_DYNAMIC for PREEMPT_RT")
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
For clarity. It's increasingly hard to reason about the code, when KASLR
is moving around the boundaries. In this case where KASLR is randomizing
the location of the kernel image within physical memory, the maximum
number of address bits for physical memory has not changed.
What has changed is the ending address of memory that is allowed to be
directly mapped by the kernel.
Let's name the variable, and the associated macro accordingly.
Also, enhance the comment above the direct_map_physmem_end definition,
to further clarify how this all works.
Link: https://lkml.kernel.org/r/20241009025024.89813-1-jhubbard@nvidia.com
Signed-off-by: John Hubbard <jhubbard@nvidia.com>
Reviewed-by: Pankaj Gupta <pankaj.gupta@amd.com>
Acked-by: David Hildenbrand <david@redhat.com>
Acked-by: Will Deacon <will@kernel.org>
Reviewed-by: Mike Rapoport (Microsoft) <rppt@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Alistair Popple <apopple@nvidia.com>
Cc: Jordan Niethe <jniethe@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
The hrtimer_init*() API is replaced by hrtimer_setup*() variants to
initialize the timer including the callback function at once.
hrtimer_init_sleeper_on_stack() does not need user to setup the callback
function separately, so a new variant would not be strictly necessary.
Nonetheless, to keep the naming convention consistent, introduce
hrtimer_setup_sleeper_on_stack(). hrtimer_init_on_stack() will be removed
once all users are converted.
Signed-off-by: Nam Cao <namcao@linutronix.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Link: https://lore.kernel.org/all/7b5e18e6dd0ace9eaa211201528cb9dc23752454.1730386209.git.namcao@linutronix.de
To initialize hrtimer on stack, hrtimer_init_on_stack() needs to be called
and also hrtimer::function must be set. This is error-prone and awkward to
use.
Introduce hrtimer_setup_on_stack() which does both of these things, so that
users of hrtimer can be simplified.
The new setup function also has a sanity check for the provided function
pointer. If NULL, a warning is emitted and a dummy callback installed.
hrtimer_init_on_stack() will be removed as soon as all of its users have
been converted to the new function.
Signed-off-by: Nam Cao <namcao@linutronix.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Link: https://lore.kernel.org/all/4b05e2ab3a82c517adf67fabc0f0cd8fe118b97c.1730386209.git.namcao@linutronix.de
To initialize hrtimer, hrtimer_init() needs to be called and also
hrtimer::function must be set. This is error-prone and awkward to use.
Introduce hrtimer_setup() which does both of these things, so that users of
hrtimer can be simplified.
The new setup function also has a sanity check for the provided function
pointer. If NULL, a warning is emitted and a dummy callback installed.
hrtimer_init() will be removed as soon as all of its users have been
converted to the new function.
Signed-off-by: Nam Cao <namcao@linutronix.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Link: https://lore.kernel.org/all/5057c1ddbfd4b92033cd93d37fe38e6b069d5ba6.1730386209.git.namcao@linutronix.de
