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>
In order to enable PREEMPT_DYNAMIC for PREEMPT_RT, remove PREEMPT_RT
from the 'Preemption Model' choice. Strictly speaking PREEMPT_RT is
not a change in how preemption works, but rather it makes a ton more
code preemptible.
Notably, take away NONE and VOLUNTARY options for PREEMPT_RT, they make
no sense (but are techincally possible).
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lkml.kernel.org/r/20241007075055.441622332@infradead.org
Change fair to use resched_curr_lazy(), which, when the lazy
preemption model is selected, will set TIF_NEED_RESCHED_LAZY.
This LAZY bit will be promoted to the full NEED_RESCHED bit on tick.
As such, the average delay between setting LAZY and actually
rescheduling will be TICK_NSEC/2.
In short, Lazy preemption will delay preemption for fair class but
will function as Full preemption for all the other classes, most
notably the realtime (RR/FIFO/DEADLINE) classes.
The goal is to bridge the performance gap with Voluntary, such that we
might eventually remove that option entirely.
Suggested-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lkml.kernel.org/r/20241007075055.331243614@infradead.org
Instead of solving the underlying problem of the double invocation of
__sched_fork() for idle tasks, sched-ext decided to hack around the issue
by partially clearing out the entity struct to preserve the already
enqueued node. A provided analysis and solution has been ignored for four
months.
Now that someone else has taken care of cleaning it up, remove the
disgusting hack and clear out the full structure. Remove the comment in the
structure declaration as well, as there is no requirement for @node being
the last element anymore.
Fixes: f0e1a0643a ("sched_ext: Implement BPF extensible scheduler class")
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Tejun Heo <tj@kernel.org>
Link: https://lore.kernel.org/r/87ldy82wkc.ffs@tglx
Idle tasks are initialized via __sched_fork() twice:
fork_idle()
copy_process()
sched_fork()
__sched_fork()
init_idle()
__sched_fork()
Instead of cleaning this up, sched_ext hacked around it. Even when analyis
and solution were provided in a discussion, nobody cared to clean this up.
init_idle() is also invoked from sched_init() to initialize the boot CPU's
idle task, which requires the __sched_fork() invocation. But this can be
trivially solved by invoking __sched_fork() before init_idle() in
sched_init() and removing the __sched_fork() invocation from init_idle().
Do so and clean up the comments explaining this historical leftover.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20241028103142.359584747@linutronix.de
Pull scheduler fixes from Thomas Gleixner:
- Plug a race between pick_next_task_fair() and try_to_wake_up() where
both try to write to the same task, even though both paths hold a
runqueue lock, but obviously from different runqueues.
The problem is that the store to task::on_rq in __block_task() is
visible to try_to_wake_up() which assumes that the task is not
queued. Both sides then operate on the same task.
Cure it by rearranging __block_task() so the the store to task::on_rq
is the last operation on the task.
- Prevent a potential NULL pointer dereference in task_numa_work()
task_numa_work() iterates the VMAs of a process. A concurrent unmap
of the address space can result in a NULL pointer return from
vma_next() which is unchecked.
Add the missing NULL pointer check to prevent this.
- Operate on the correct scheduler policy in task_should_scx()
task_should_scx() returns true when a task should be handled by sched
EXT. It checks the tasks scheduling policy.
This fails when the check is done before a policy has been set.
Cure it by handing the policy into task_should_scx() so it operates
on the requested value.
- Add the missing handling of sched EXT in the delayed dequeue
mechanism. This was simply forgotten.
* tag 'sched-urgent-2024-11-03' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
sched/ext: Fix scx vs sched_delayed
sched: Pass correct scheduling policy to __setscheduler_class
sched/numa: Fix the potential null pointer dereference in task_numa_work()
sched: Fix pick_next_task_fair() vs try_to_wake_up() race
Pull sched_ext fixes from Tejun Heo:
- Instances of scx_ops_bypass() could race each other leading to
misbehavior. Fix by protecting the operation with a spinlock.
- selftest and userspace header fixes
* tag 'sched_ext-for-6.12-rc5-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/sched_ext:
sched_ext: Fix enq_last_no_enq_fails selftest
sched_ext: Make cast_mask() inline
scx: Fix raciness in scx_ops_bypass()
scx: Fix exit selftest to use custom DSQ
sched_ext: Fix function pointer type mismatches in BPF selftests
selftests/sched_ext: add order-only dependency of runner.o on BPFOBJ
Similarly to commit dfa4ed29b1 ("sched_ext: Introduce LLC awareness to
the default idle selection policy"), extend the built-in idle CPU
selection policy to also prioritize CPUs within the same NUMA node.
With this change applied, the built-in CPU idle selection policy follows
this logic:
- always prioritize CPUs from fully idle SMT cores,
- select the same CPU if possible,
- select a CPU within the same LLC domain,
- select a CPU within the same NUMA node.
Both NUMA and LLC awareness features are enabled only when the system
has multiple NUMA nodes or multiple LLC domains.
In the future, we may want to improve the NUMA node selection to account
the node distance from prev_cpu. Currently, the logic only tries to keep
tasks running on the same NUMA node. If all CPUs within a node are busy,
the next NUMA node is chosen randomly.
Signed-off-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Commit 98442f0ccd ("sched: Fix delayed_dequeue vs
switched_from_fair()") overlooked that __setscheduler_prio(), now
__setscheduler_class() relies on p->policy for task_should_scx(), and
moved the call before __setscheduler_params() updates it, causing it
to be using the old p->policy value.
Resolve this by changing task_should_scx() to take the policy itself
instead of a task pointer, such that __sched_setscheduler() can pass
in the updated policy.
Fixes: 98442f0ccd ("sched: Fix delayed_dequeue vs switched_from_fair()")
Signed-off-by: Aboorva Devarajan <aboorvad@linux.ibm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Tejun Heo <tj@kernel.org>
What psi needs to do on each enqueue and dequeue has gotten more
subtle, and the generic sched code trying to distill this into a bool
for the callbacks is awkward.
Pass the flags directly and let psi parse them. For that to work, the
#include "stats.h" (which has the psi callback implementations) needs
to be below the flag definitions in "sched.h". Move that section
further down, next to some of the other accounting stuff.
This also puts the ENQUEUE_SAVE/RESTORE branch behind the psi jump
label, slightly reducing overhead when PSI=y but runtime disabled.
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20241014144358.GB1021@cmpxchg.org
When running stress-ng-vm-segv test, we found a null pointer dereference
error in task_numa_work(). Here is the backtrace:
[323676.066985] Unable to handle kernel NULL pointer dereference at virtual address 0000000000000020
......
[323676.067108] CPU: 35 PID: 2694524 Comm: stress-ng-vm-se
......
[323676.067113] pstate: 23401009 (nzCv daif +PAN -UAO +TCO +DIT +SSBS BTYPE=--)
[323676.067115] pc : vma_migratable+0x1c/0xd0
[323676.067122] lr : task_numa_work+0x1ec/0x4e0
[323676.067127] sp : ffff8000ada73d20
[323676.067128] x29: ffff8000ada73d20 x28: 0000000000000000 x27: 000000003e89f010
[323676.067130] x26: 0000000000080000 x25: ffff800081b5c0d8 x24: ffff800081b27000
[323676.067133] x23: 0000000000010000 x22: 0000000104d18cc0 x21: ffff0009f7158000
[323676.067135] x20: 0000000000000000 x19: 0000000000000000 x18: ffff8000ada73db8
[323676.067138] x17: 0001400000000000 x16: ffff800080df40b0 x15: 0000000000000035
[323676.067140] x14: ffff8000ada73cc8 x13: 1fffe0017cc72001 x12: ffff8000ada73cc8
[323676.067142] x11: ffff80008001160c x10: ffff000be639000c x9 : ffff8000800f4ba4
[323676.067145] x8 : ffff000810375000 x7 : ffff8000ada73974 x6 : 0000000000000001
[323676.067147] x5 : 0068000b33e26707 x4 : 0000000000000001 x3 : ffff0009f7158000
[323676.067149] x2 : 0000000000000041 x1 : 0000000000004400 x0 : 0000000000000000
[323676.067152] Call trace:
[323676.067153] vma_migratable+0x1c/0xd0
[323676.067155] task_numa_work+0x1ec/0x4e0
[323676.067157] task_work_run+0x78/0xd8
[323676.067161] do_notify_resume+0x1ec/0x290
[323676.067163] el0_svc+0x150/0x160
[323676.067167] el0t_64_sync_handler+0xf8/0x128
[323676.067170] el0t_64_sync+0x17c/0x180
[323676.067173] Code: d2888001 910003fd f9000bf3 aa0003f3 (f9401000)
[323676.067177] SMP: stopping secondary CPUs
[323676.070184] Starting crashdump kernel...
stress-ng-vm-segv in stress-ng is used to stress test the SIGSEGV error
handling function of the system, which tries to cause a SIGSEGV error on
return from unmapping the whole address space of the child process.
Normally this program will not cause kernel crashes. But before the
munmap system call returns to user mode, a potential task_numa_work()
for numa balancing could be added and executed. In this scenario, since the
child process has no vma after munmap, the vma_next() in task_numa_work()
will return a null pointer even if the vma iterator restarts from 0.
Recheck the vma pointer before dereferencing it in task_numa_work().
Fixes: 214dbc4281 ("sched: convert to vma iterator")
Signed-off-by: Shawn Wang <shawnwang@linux.alibaba.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: stable@vger.kernel.org # v6.2+
Link: https://lkml.kernel.org/r/20241025022208.125527-1-shawnwang@linux.alibaba.com
ops.dispatch() and ops.yield() may be fed a NULL task_struct pointer.
set_arg_maybe_null() is used to tell the verifier that they should be NULL
checked before being dereferenced. BPF now has an a lot prettier way to
express this - tagging arguments in CFI stubs with __nullable. Replace
set_arg_maybe_null() with __nullable CFI stub tags.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: David Vernet <void@manifault.com>
Acked-by: Martin KaFai Lau <martin.lau@kernel.org>
Acked-by: Alexei Starovoitov <ast@kernel.org>
CFI stubs can be used to tag arguments with __nullable (and possibly other
tags in the future) but for that to work the CFI stubs must have names that
are recognized by BPF. Rename them.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: David Vernet <void@manifault.com>
Acked-by: Martin KaFai Lau <martin.lau@kernel.org>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Rely on the scheduler topology information to implement basic LLC
awareness in the sched_ext build-in idle selection policy.
This allows schedulers using the built-in policy to make more informed
decisions when selecting an idle CPU in systems with multiple LLCs, such
as NUMA systems or chiplet-based architectures, and it helps keep tasks
within the same LLC domain, thereby improving cache locality.
For efficiency, LLC awareness is applied only to tasks that can run on
all the CPUs in the system for now. If a task's affinity is modified
from user space, it's the responsibility of user space to choose the
appropriate optimized scheduling domain.
Signed-off-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Update ops.select_cpu() documentation to clarify that this method is not
called for tasks that are restricted to run on a single CPU, as these
tasks do not have the option to select a different CPU.
Signed-off-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Syzkaller robot reported KCSAN tripping over the
ASSERT_EXCLUSIVE_WRITER(p->on_rq) in __block_task().
The report noted that both pick_next_task_fair() and try_to_wake_up()
were concurrently trying to write to the same p->on_rq, violating the
assertion -- even though both paths hold rq->__lock.
The logical consequence is that both code paths end up holding a
different rq->__lock. And looking through ttwu(), this is possible
when the __block_task() 'p->on_rq = 0' store is visible to the ttwu()
'p->on_rq' load, which then assumes the task is not queued and
continues to migrate it.
Rearrange things such that __block_task() releases @p with the store
and no code thereafter will use @p again.
Fixes: 152e11f6df ("sched/fair: Implement delayed dequeue")
Reported-by: syzbot+0ec1e96c2cdf5c0e512a@syzkaller.appspotmail.com
Reported-by: Kent Overstreet <kent.overstreet@linux.dev>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Marco Elver <elver@google.com>
Link: https://lkml.kernel.org/r/20241023093641.GE16066@noisy.programming.kicks-ass.net
Overlapping fixes solving the same bug slightly differently:
7266f0a6d3 fs/bcachefs: Fix __wait_on_freeing_inode() definition of waitqueue entry
3b80552e70 bcachefs: __wait_for_freeing_inode: Switch to wait_bit_queue_entry
Use the upstream version.
Signed-off-by: Ingo Molnar <mingo@kernel.org>
In the sched_ext built-in idle CPU selection logic, when handling a
WF_SYNC wakeup, we always attempt to migrate the task to the waker's
CPU, as the waker is expected to yield the CPU after waking the task.
However, it may be preferable to keep the task on its previous CPU if
the waker's CPU is cache-affine.
The same approach is also used by the fair class and in other scx
schedulers, like scx_rusty and scx_bpfland.
Therefore, apply the same logic to the built-in idle CPU selection
policy as well.
Signed-off-by: Andrea Righi <andrea.righi@linux.dev>
Signed-off-by: Tejun Heo <tj@kernel.org>
Conflicts:
kernel/sched/ext.c
There's a context conflict between this upstream commit:
3fdb9ebcec sched_ext: Start schedulers with consistent p->scx.slice values
... and this fix in sched/urgent:
98442f0ccd sched: Fix delayed_dequeue vs switched_from_fair()
Resolve it.
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Pull sched_ext fixes from Tejun Heo:
- More issues reported in the enable/disable paths on large machines
with many tasks due to scx_tasks_lock being held too long. Break up
the task iterations
- Remove ops.select_cpu() dependency in bypass mode so that a
misbehaving implementation can't live-lock the machine by pushing all
tasks to few CPUs in bypass mode
- Other misc fixes
* tag 'sched_ext-for-6.12-rc3-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/sched_ext:
sched_ext: Remove unnecessary cpu_relax()
sched_ext: Don't hold scx_tasks_lock for too long
sched_ext: Move scx_tasks_lock handling into scx_task_iter helpers
sched_ext: bypass mode shouldn't depend on ops.select_cpu()
sched_ext: Move scx_buildin_idle_enabled check to scx_bpf_select_cpu_dfl()
sched_ext: Start schedulers with consistent p->scx.slice values
Revert "sched_ext: Use shorter slice while bypassing"
sched_ext: use correct function name in pick_task_scx() warning message
selftests: sched_ext: Add sched_ext as proper selftest target
As described in commit b07996c7ab ("sched_ext: Don't hold
scx_tasks_lock for too long"), we're doing a cond_resched() every 32
calls to scx_task_iter_next() to avoid RCU and other stalls. That commit
also added a cpu_relax() to the codepath where we drop and reacquire the
lock, but as Waiman described in [0], cpu_relax() should only be
necessary in busy loops to avoid pounding on a cacheline (or to allow a
hypertwin to more fully utilize a core).
Let's remove the unnecessary cpu_relax().
[0]: https://lore.kernel.org/all/35b3889b-904a-4d26-981f-c8aa1557a7c7@redhat.com/
Cc: Waiman Long <llong@redhat.com>
Signed-off-by: David Vernet <void@manifault.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Let's define the "scheduling context" as all the scheduler state
in task_struct for the task chosen to run, which we'll call the
donor task, and the "execution context" as all state required to
actually run the task.
Currently both are intertwined in task_struct. We want to
logically split these such that we can use the scheduling
context of the donor task selected to be scheduled, but use
the execution context of a different task to actually be run.
To this purpose, introduce rq->donor field to point to the
task_struct chosen from the runqueue by the scheduler, and will
be used for scheduler state, and preserve rq->curr to indicate
the execution context of the task that will actually be run.
This patch introduces the donor field as a union with curr, so it
doesn't cause the contexts to be split yet, but adds the logic to
handle everything separately.
[add additional comments and update more sched_class code to use
rq::proxy]
[jstultz: Rebased and resolved minor collisions, reworked to use
accessors, tweaked update_curr_common to use rq_proxy fixing rt
scheduling issues]
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Juri Lelli <juri.lelli@redhat.com>
Signed-off-by: Connor O'Brien <connoro@google.com>
Signed-off-by: John Stultz <jstultz@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Metin Kaya <metin.kaya@arm.com>
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Tested-by: Metin Kaya <metin.kaya@arm.com>
Link: https://lore.kernel.org/r/20241009235352.1614323-8-jstultz@google.com
commit 223baf9d17 ("sched: Fix performance regression introduced by mm_cid")
introduced a per-mm/cpu current concurrency id (mm_cid), which keeps
a reference to the concurrency id allocated for each CPU. This reference
expires shortly after a 100ms delay.
These per-CPU references keep the per-mm-cid data cache-local in
situations where threads are running at least once on each CPU within
each 100ms window, thus keeping the per-cpu reference alive.
However, intermittent workloads behaving in bursts spaced by more than
100ms on each CPU exhibit bad cache locality and degraded performance
compared to purely per-cpu data indexing, because concurrency IDs are
allocated over various CPUs and cores, therefore losing cache locality
of the associated data.
Introduce the following changes to improve per-mm-cid cache locality:
- Add a "recent_cid" field to the per-mm/cpu mm_cid structure to keep
track of which mm_cid value was last used, and use it as a hint to
attempt re-allocating the same concurrency ID the next time this
mm/cpu needs to allocate a concurrency ID,
- Add a per-mm CPUs allowed mask, which keeps track of the union of
CPUs allowed for all threads belonging to this mm. This cpumask is
only set during the lifetime of the mm, never cleared, so it
represents the union of all the CPUs allowed since the beginning of
the mm lifetime (note that the mm_cpumask() is really arch-specific
and tailored to the TLB flush needs, and is thus _not_ a viable
approach for this),
- Add a per-mm nr_cpus_allowed to keep track of the weight of the
per-mm CPUs allowed mask (for fast access),
- Add a per-mm max_nr_cid to keep track of the highest number of
concurrency IDs allocated for the mm. This is used for expanding the
concurrency ID allocation within the upper bound defined by:
min(mm->nr_cpus_allowed, mm->mm_users)
When the next unused CID value reaches this threshold, stop trying
to expand the cid allocation and use the first available cid value
instead.
Spreading allocation to use all the cid values within the range
[ 0, min(mm->nr_cpus_allowed, mm->mm_users) - 1 ]
improves cache locality while preserving mm_cid compactness within the
expected user limits,
- In __mm_cid_try_get, only return cid values within the range
[ 0, mm->nr_cpus_allowed ] rather than [ 0, nr_cpu_ids ]. This
prevents allocating cids above the number of allowed cpus in
rare scenarios where cid allocation races with a concurrent
remote-clear of the per-mm/cpu cid. This improvement is made
possible by the addition of the per-mm CPUs allowed mask,
- In sched_mm_cid_migrate_to, use mm->nr_cpus_allowed rather than
t->nr_cpus_allowed. This criterion was really meant to compare
the number of mm->mm_users to the number of CPUs allowed for the
entire mm. Therefore, the prior comparison worked fine when all
threads shared the same CPUs allowed mask, but not so much in
scenarios where those threads have different masks (e.g. each
thread pinned to a single CPU). This improvement is made
possible by the addition of the per-mm CPUs allowed mask.
* Benchmarks
Each thread increments 16kB worth of 8-bit integers in bursts, with
a configurable delay between each thread's execution. Each thread run
one after the other (no threads run concurrently). The order of
thread execution in the sequence is random. The thread execution
sequence begins again after all threads have executed. The 16kB areas
are allocated with rseq_mempool and indexed by either cpu_id, mm_cid
(not cache-local), or cache-local mm_cid. Each thread is pinned to its
own core.
Testing configurations:
8-core/1-L3: Use 8 cores within a single L3
24-core/24-L3: Use 24 cores, 1 core per L3
192-core/24-L3: Use 192 cores (all cores in the system)
384-thread/24-L3: Use 384 HW threads (all HW threads in the system)
Intermittent workload delays between threads: 200ms, 10ms.
Hardware:
CPU(s): 384
On-line CPU(s) list: 0-383
Vendor ID: AuthenticAMD
Model name: AMD EPYC 9654 96-Core Processor
Thread(s) per core: 2
Core(s) per socket: 96
Socket(s): 2
Caches (sum of all):
L1d: 6 MiB (192 instances)
L1i: 6 MiB (192 instances)
L2: 192 MiB (192 instances)
L3: 768 MiB (24 instances)
Each result is an average of 5 test runs. The cache-local speedup
is calculated as: (cache-local mm_cid) / (mm_cid).
Intermittent workload delay: 200ms
per-cpu mm_cid cache-local mm_cid cache-local speedup
(ns) (ns) (ns)
8-core/1-L3 1374 19289 1336 14.4x
24-core/24-L3 2423 26721 1594 16.7x
192-core/24-L3 2291 15826 2153 7.3x
384-thread/24-L3 1874 13234 1907 6.9x
Intermittent workload delay: 10ms
per-cpu mm_cid cache-local mm_cid cache-local speedup
(ns) (ns) (ns)
8-core/1-L3 662 756 686 1.1x
24-core/24-L3 1378 3648 1035 3.5x
192-core/24-L3 1439 10833 1482 7.3x
384-thread/24-L3 1503 10570 1556 6.8x
[ This deprecates the prior "sched: NUMA-aware per-memory-map concurrency IDs"
patch series with a simpler and more general approach. ]
[ This patch applies on top of v6.12-rc1. ]
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Marco Elver <elver@google.com>
Link: https://lore.kernel.org/lkml/20240823185946.418340-1-mathieu.desnoyers@efficios.com/
The memory barrier rmb() in generic idle loop do_idle() function is not
needed, it doesn't order any load instruction, just remove it as needless
rmb() can cause performance impact.
The rmb() was introduced by the tglx/history.git commit f2f1b44c75c4
("[PATCH] Remove RCU abuse in cpu_idle()") to order the loads between
cpu_idle_map and pm_idle. It pairs with wmb() in function cpu_idle_wait().
And then with the removal of cpu_idle_state in function cpu_idle() and
wmb() in function cpu_idle_wait() in commit 783e391b7b ("x86: Simplify
cpu_idle_wait"), rmb() no longer has a reason to exist.
After that, commit d166991234 ("idle: Implement generic idle function")
implemented a generic idle function cpu_idle_loop() which resembles the
functionality found in arch/. And it retained the rmb() in generic idle
loop in file kernel/cpu/idle.c.
And at last, commit cf37b6b484 ("sched/idle: Move cpu/idle.c to
sched/idle.c") moved cpu/idle.c to sched/idle.c. And commit c1de45ca83
("sched/idle: Add support for tasks that inject idle") renamed function
cpu_idle_loop() to do_idle().
History Tree: https://git.kernel.org/pub/scm/linux/kernel/git/tglx/history.git
Signed-off-by: Zhongqiu Han <quic_zhonhan@quicinc.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20241009093745.9504-1-quic_zhonhan@quicinc.com
Sean noted that ever since commit 152e11f6df ("sched/fair: Implement
delayed dequeue") KVM's preemption notifiers have started
mis-classifying preemption vs blocking.
Notably p->on_rq is no longer sufficient to determine if a task is
runnable or blocked -- the aforementioned commit introduces tasks that
remain on the runqueue even through they will not run again, and
should be considered blocked for many cases.
Add the task_is_runnable() helper to classify things and audit all
external users of the p->on_rq state. Also add a few comments.
Fixes: 152e11f6df ("sched/fair: Implement delayed dequeue")
Reported-by: Sean Christopherson <seanjc@google.com>
Tested-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lkml.kernel.org/r/20241010091843.GK33184@noisy.programming.kicks-ass.net
Since sched_delayed tasks remain queued even after blocking, the load
balancer can migrate them between runqueues while PSI considers them
to be asleep. As a result, it misreads the migration requeue followed
by a wakeup as a double queue:
psi: inconsistent task state! task=... cpu=... psi_flags=4 clear=. set=4
First, call psi_enqueue() after p->sched_class->enqueue_task(). A
wakeup will clear p->se.sched_delayed while a migration will not, so
psi can use that flag to tell them apart.
Then teach psi to migrate any "sleep" state when delayed-dequeue tasks
are being migrated.
Delayed-dequeue tasks can be revived by ttwu_runnable(), which will
call down with a new ENQUEUE_DELAYED. Instead of further complicating
the wakeup conditional in enqueue_task(), identify migration contexts
instead and default to wakeup handling for all other cases.
It's not just the warning in dmesg, the task state corruption causes a
permanent CPU pressure indication, which messes with workload/machine
health monitoring.
Debugged-by-and-original-fix-by: K Prateek Nayak <kprateek.nayak@amd.com>
Fixes: 152e11f6df ("sched/fair: Implement delayed dequeue")
Closes: https://lore.kernel.org/lkml/20240830123458.3557-1-spasswolf@web.de/
Closes: https://lore.kernel.org/all/cd67fbcd-d659-4822-bb90-7e8fbb40a856@molgen.mpg.de/
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Link: https://lkml.kernel.org/r/20241010193712.GC181795@cmpxchg.org
Commit 2e0199df25 ("sched/fair: Prepare exit/cleanup paths for delayed_dequeue")
and its follow up fixes try to deal with a rather unfortunate
situation where is task is enqueued in a new class, even though it
shouldn't have been. Mostly because the existing ->switched_to/from()
hooks are in the wrong place for this case.
This all led to Paul being able to trigger failures at something like
once per 10k CPU hours of RCU torture.
For now, do the ugly thing and move the code to the right place by
ignoring the switch hooks.
Note: Clean up the whole sched_class::switch*_{to,from}() thing.
Fixes: 2e0199df25 ("sched/fair: Prepare exit/cleanup paths for delayed_dequeue")
Reported-by: Paul E. McKenney <paulmck@kernel.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20241003185037.GA5594@noisy.programming.kicks-ass.net
With KASAN and PREEMPT_RT enabled, calling task_work_add() in
task_tick_mm_cid() may cause the following splat.
[ 63.696416] BUG: sleeping function called from invalid context at kernel/locking/spinlock_rt.c:48
[ 63.696416] in_atomic(): 1, irqs_disabled(): 1, non_block: 0, pid: 610, name: modprobe
[ 63.696416] preempt_count: 10001, expected: 0
[ 63.696416] RCU nest depth: 1, expected: 1
This problem is caused by the following call trace.
sched_tick() [ acquire rq->__lock ]
-> task_tick_mm_cid()
-> task_work_add()
-> __kasan_record_aux_stack()
-> kasan_save_stack()
-> stack_depot_save_flags()
-> alloc_pages_mpol_noprof()
-> __alloc_pages_noprof()
-> get_page_from_freelist()
-> rmqueue()
-> rmqueue_pcplist()
-> __rmqueue_pcplist()
-> rmqueue_bulk()
-> rt_spin_lock()
The rq lock is a raw_spinlock_t. We can't sleep while holding
it. IOW, we can't call alloc_pages() in stack_depot_save_flags().
The task_tick_mm_cid() function with its task_work_add() call was
introduced by commit 223baf9d17 ("sched: Fix performance regression
introduced by mm_cid") in v6.4 kernel.
Fortunately, there is a kasan_record_aux_stack_noalloc() variant that
calls stack_depot_save_flags() while not allowing it to allocate
new pages. To allow task_tick_mm_cid() to use task_work without
page allocation, a new TWAF_NO_ALLOC flag is added to enable calling
kasan_record_aux_stack_noalloc() instead of kasan_record_aux_stack()
if set. The task_tick_mm_cid() function is modified to add this new flag.
The possible downside is the missing stack trace in a KASAN report due
to new page allocation required when task_work_add_noallloc() is called
which should be rare.
Fixes: 223baf9d17 ("sched: Fix performance regression introduced by mm_cid")
Signed-off-by: Waiman Long <longman@redhat.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20241010014432.194742-1-longman@redhat.com
Iterating with scx_task_iter involves scx_tasks_lock and optionally the rq
lock of the task being iterated. Both locks can be released during iteration
and the iteration can be continued after re-grabbing scx_tasks_lock.
Currently, all lock handling is pushed to the caller which is a bit
cumbersome and makes it difficult to add lock-aware behaviors. Make the
scx_task_iter helpers handle scx_tasks_lock.
- scx_task_iter_init/scx_taks_iter_exit() now grabs and releases
scx_task_lock, respectively. Renamed to
scx_task_iter_start/scx_task_iter_stop() to more clearly indicate that
there are non-trivial side-effects.
- Add __ prefix to scx_task_iter_rq_unlock() to indicate that the function
is internal.
- Add scx_task_iter_unlock/relock(). The former drops both rq lock (if held)
and scx_tasks_lock and the latter re-locks only scx_tasks_lock.
This doesn't cause behavior changes and will be used to implement stall
avoidance.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: David Vernet <void@manifault.com>
Bypass mode was depending on ops.select_cpu() which can't be trusted as with
the rest of the BPF scheduler. Always enable and use scx_select_cpu_dfl() in
bypass mode.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: David Vernet <void@manifault.com>
Move the sanity check from the inner function scx_select_cpu_dfl() to the
exported kfunc scx_bpf_select_cpu_dfl(). This doesn't cause behavior
differences and will allow using scx_select_cpu_dfl() in bypass mode
regardless of scx_builtin_idle_enabled.
Signed-off-by: Tejun Heo <tj@kernel.org>
The disable path caps p->scx.slice to SCX_SLICE_DFL. As the field is already
being ignored at this stage during disable, the only effect this has is that
when the next BPF scheduler is loaded, it won't see unreasonable left-over
slices. Ultimately, this shouldn't matter but it's better to start in a
known state. Drop p->scx.slice capping from the disable path and instead
reset it to SCX_SLICE_DFL in the enable path.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: David Vernet <void@manifault.com>
This reverts commit 6f34d8d382.
Slice length is ignored while bypassing and tasks are switched on every tick
and thus the patch does not make any difference. The perceived difference
was from test noise.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: David Vernet <void@manifault.com>
