a11d6784d7 ("sched_ext: Fix missing rq lock in scx_bpf_cpuperf_set()")
added a call to scx_ops_error() which was renamed to scx_error() in
for-6.16. Fix it up.
scx_bpf_cpuperf_set() can be used to set a performance target level on
any CPU. However, it doesn't correctly acquire the corresponding rq
lock, which may lead to unsafe behavior and trigger the following
warning, due to the lockdep_assert_rq_held() check:
[ 51.713737] WARNING: CPU: 3 PID: 3899 at kernel/sched/sched.h:1512 scx_bpf_cpuperf_set+0x1a0/0x1e0
...
[ 51.713836] Call trace:
[ 51.713837] scx_bpf_cpuperf_set+0x1a0/0x1e0 (P)
[ 51.713839] bpf_prog_62d35beb9301601f_bpfland_init+0x168/0x440
[ 51.713841] bpf__sched_ext_ops_init+0x54/0x8c
[ 51.713843] scx_ops_enable.constprop.0+0x2c0/0x10f0
[ 51.713845] bpf_scx_reg+0x18/0x30
[ 51.713847] bpf_struct_ops_link_create+0x154/0x1b0
[ 51.713849] __sys_bpf+0x1934/0x22a0
Fix by properly acquiring the rq lock when possible or raising an error
if we try to operate on a CPU that is not the one currently locked.
Fixes: d86adb4fc0 ("sched_ext: Add cpuperf support")
Signed-off-by: Andrea Righi <arighi@nvidia.com>
Acked-by: Changwoo Min <changwoo@igalia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Some kfuncs provided by sched_ext may need to operate on a struct rq,
but they can be invoked from various contexts, specifically, different
scx callbacks.
While some of these callbacks are invoked with a particular rq already
locked, others are not. This makes it impossible for a kfunc to reliably
determine whether it's safe to access a given rq, triggering potential
bugs or unsafe behaviors, see for example [1].
To address this, track the currently locked rq whenever a sched_ext
callback is invoked via SCX_CALL_OP*().
This allows kfuncs that need to operate on an arbitrary rq to retrieve
the currently locked one and apply the appropriate action as needed.
[1] https://lore.kernel.org/lkml/20250325140021.73570-1-arighi@nvidia.com/
Suggested-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrea Righi <arighi@nvidia.com>
Acked-by: Changwoo Min <changwoo@igalia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
SCX_EV_ENQ_SLICE_DFL gives the impression that the event only occurs
when the tasks were enqueued, which seems not accurate. What it actually
means is the refilling with defalt slice, and this can occur either when
enqueue or pick_task. Let's change the variable to
SCX_EV_REFILL_SLICE_DFL.
Signed-off-by: Honglei Wang <jameshongleiwang@126.com>
Acked-by: Changwoo Min <changwoo@igalia.com>
Acked-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
scx_has_op is used to encode which ops are implemented by the BPF scheduler
into an array of static_keys. While this saves a bit of branching overhead,
that is unlikely to be noticeable compared to the overall cost. As the
global static_keys can't work with the planned hierarchical multiple
scheduler support, replace the static_key array with a bitmap.
In repeated hackbench runs before and after static_keys removal on an AMD
Ryzen 3900X, I couldn't tell any measurable performance difference.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Changwoo Min <changwoo@igalia.com>
Acked-by: Andrea Righi <arighi@nvidia.com>
scx_ops_allow_queued_wakeup is used to encode SCX_OPS_ALLOW_QUEUED_WAKEUP
into a static_key. The test is gated behind scx_enabled(), and, even when
sched_ext is enabled, is unlikely for the static_key usage to make any
meaningful difference. It is made to use a static_key mostly because there
was no reason not to. However, global static_keys can't work with the
planned hierarchical multiple scheduler support. Remove the static_key and
instead test SCX_OPS_ALLOW_QUEUED_WAKEUP directly.
In repeated hackbench runs before and after static_keys removal on an AMD
Ryzen 3900X, I couldn't tell any measurable performance difference.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Changwoo Min <changwoo@igalia.com>
Acked-by: Andrea Righi <arighi@nvidia.com>
scx_ops_cpu_preempt is used to encode whether ops.cpu_acquire/release() are
implemented into a static_key. These tests aren't hot enough for static_key
usage to make any meaningful difference and are made to use a static_key
mostly because there was no reason not to. However, global static_keys can't
work with the planned hierarchical multiple scheduler support. Remove the
static_key and instead use an internal ops flag SCX_OPS_HAS_CPU_PREEMPT to
record and test whether ops.cpu_acquire/release() are implemented.
In repeated hackbench runs before and after static_keys removal on an AMD
Ryzen 3900X, I couldn't tell any measurable performance difference.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Changwoo Min <changwoo@igalia.com>
Acked-by: Andrea Righi <arighi@nvidia.com>
scx_ops_enq_last/exiting/migration_disabled are used to encode the
corresponding SCX_OPS_ flags into static_keys. These flags aren't hot enough
for static_key usage to make any meaningful difference and are made
static_keys mostly because there was no reason not to. However, global
static_keys can't work with the planned hierarchical multiple scheduler
support. Remove the static_keys and test the ops flags directly.
In repeated hackbench runs before and after static_keys removal on an AMD
Ryzen 3900X, I couldn't tell any measurable performance difference.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Changwoo Min <changwoo@igalia.com>
Acked-by: Andrea Righi <arighi@nvidia.com>
Pull for-6.15-fixes to receive:
e776b26e37 ("sched_ext: Remove cpu.weight / cpu.idle unimplemented warnings")
which conflicts with:
1a7ff7216c ("sched_ext: Drop "ops" from scx_ops_enable_state and friends")
The former removes code updated by the latter. Resolved by removing the
updated section.
Signed-off-by: Tejun Heo <tj@kernel.org>
SCX_OPS_HAS_CGROUP_WEIGHT was only used to suppress the missing cgroup
weight support warnings. Now that the warnings are removed, the flag doesn't
do anything. Mark it for deprecation and remove its usage from scx_flatcg.
v2: Actually include the scx_flatcg update.
Signed-off-by: Tejun Heo <tj@kernel.org>
Suggested-and-reviewed-by: Andrea Righi <arighi@nvidia.com>
sched_ext generates warnings when cpu.weight / cpu.idle are set to
non-default values if the BPF scheduler doesn't implement weight support.
These warnings don't provide much value while adding constant annoyance. A
BPF scheduler may not implement any particular behavior and there's nothing
particularly special about missing cgroup weight support. Drop the warnings.
Signed-off-by: Tejun Heo <tj@kernel.org>
Replace kzalloc with kvzalloc for the exit_dump buffer allocation, which
can require large contiguous memory depending on the implementation.
This change prevents allocation failures by allowing the system to fall
back to vmalloc when contiguous memory allocation fails.
Since this buffer is only used for debugging purposes, physical memory
contiguity is not required, making vmalloc a suitable alternative.
Cc: stable@vger.kernel.org
Fixes: 07814a9439 ("sched_ext: Print debug dump after an error exit")
Suggested-by: Rik van Riel <riel@surriel.com>
Signed-off-by: Breno Leitao <leitao@debian.org>
Acked-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Provide a new kfunc, scx_bpf_select_cpu_and(), that can be used to apply
the built-in idle CPU selection policy to a subset of allowed CPU.
This new helper is basically an extension of scx_bpf_select_cpu_dfl().
However, when an idle CPU can't be found, it returns a negative value
instead of @prev_cpu, aligning its behavior more closely with
scx_bpf_pick_idle_cpu().
It also accepts %SCX_PICK_IDLE_* flags, which can be used to enforce
strict selection to @prev_cpu's node (%SCX_PICK_IDLE_IN_NODE), or to
request only a full-idle SMT core (%SCX_PICK_IDLE_CORE), while applying
the built-in selection logic.
With this helper, BPF schedulers can apply the built-in idle CPU
selection policy restricted to any arbitrary subset of CPUs.
Example usage
=============
Possible usage in ops.select_cpu():
s32 BPF_STRUCT_OPS(foo_select_cpu, struct task_struct *p,
s32 prev_cpu, u64 wake_flags)
{
const struct cpumask *cpus = task_allowed_cpus(p) ?: p->cpus_ptr;
s32 cpu;
cpu = scx_bpf_select_cpu_and(p, prev_cpu, wake_flags, cpus, 0);
if (cpu >= 0) {
scx_bpf_dsq_insert(p, SCX_DSQ_LOCAL, SCX_SLICE_DFL, 0);
return cpu;
}
return prev_cpu;
}
Results
=======
Load distribution on a 4 sockets, 4 cores per socket system, simulated
using virtme-ng, running a modified version of scx_bpfland that uses
scx_bpf_select_cpu_and() with 0xff00 as the allowed subset of CPUs:
$ vng --cpu 16,sockets=4,cores=4,threads=1
...
$ stress-ng -c 16
...
$ htop
...
0[ 0.0%] 8[||||||||||||||||||||||||100.0%]
1[ 0.0%] 9[||||||||||||||||||||||||100.0%]
2[ 0.0%] 10[||||||||||||||||||||||||100.0%]
3[ 0.0%] 11[||||||||||||||||||||||||100.0%]
4[ 0.0%] 12[||||||||||||||||||||||||100.0%]
5[ 0.0%] 13[||||||||||||||||||||||||100.0%]
6[ 0.0%] 14[||||||||||||||||||||||||100.0%]
7[ 0.0%] 15[||||||||||||||||||||||||100.0%]
With scx_bpf_select_cpu_dfl() tasks would be distributed evenly across
all the available CPUs.
Signed-off-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Modify scx_select_cpu_dfl() to take the allowed cpumask as an explicit
argument, instead of implicitly using @p->cpus_ptr.
This prepares for future changes where arbitrary cpumasks may be passed
to the built-in idle CPU selection policy.
This is a pure refactoring with no functional changes.
Signed-off-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
The tag "ops" is used for two different purposes. First, to indicate that
the entity is directly related to the operations such as flags carried in
sched_ext_ops. Second, to indicate that the entity applies to something
global such as enable or bypass states. The second usage is historical and
causes confusion rather than clarifying anything. For example,
scx_ops_enable_state enums are named SCX_OPS_* and thus conflict with
scx_ops_flags. Let's drop the second usages.
Drop "ops" from SCX_OPS_TASK_ITER_BATCH.
Signed-off-by: Tejun Heo <tj@kernel.org>
Suggested-and-acked-by: Andrea Righi <arighi@nvidia.com>
The tag "ops" is used for two different purposes. First, to indicate that
the entity is directly related to the operations such as flags carried in
sched_ext_ops. Second, to indicate that the entity applies to something
global such as enable or bypass states. The second usage is historical and
causes confusion rather than clarifying anything. For example,
scx_ops_enable_state enums are named SCX_OPS_* and thus conflict with
scx_ops_flags. Let's drop the second usages.
Drop "ops" from scx_ops_{init|exit|enable|disable}[_task]() and friends.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Andrea Righi <arighi@nvidia.com>
The tag "ops" is used for two different purposes. First, to indicate that
the entity is directly related to the operations such as flags carried in
sched_ext_ops. Second, to indicate that the entity applies to something
global such as enable or bypass states. The second usage is historical and
causes confusion rather than clarifying anything. For example,
scx_ops_enable_state enums are named SCX_OPS_* and thus conflict with
scx_ops_flags. Let's drop the second usages.
Drop "ops" from scx_ops_exit(), scx_ops_error() and friends.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Andrea Righi <arighi@nvidia.com>
The tag "ops" is used for two different purposes. First, to indicate that
the entity is directly related to the operations such as flags carried in
sched_ext_ops. Second, to indicate that the entity applies to something
global such as enable or bypass states. The second usage is historical and
causes confusion rather than clarifying anything. For example,
scx_ops_enable_state enums are named SCX_OPS_* and thus conflict with
scx_ops_flags. Let's drop the second usages.
Drop "ops" from scx_ops_bypass(), scx_ops_breather() and friends. Update
scx_show_state.py accordingly.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Andrea Righi <arighi@nvidia.com>
The tag "ops" is used for two different purposes. First, to indicate that
the entity is directly related to the operations such as flags carried in
sched_ext_ops. Second, to indicate that the entity applies to something
global such as enable or bypass states. The second usage is historical and
causes confusion rather than clarifying anything. For example,
scx_ops_enable_state enums are named SCX_OPS_* and thus conflict with
scx_ops_flags. Let's drop the second usages.
Drop "ops" from scx_ops_helper, scx_ops_enable_mutex and __scx_ops_enabled.
Update scx_show_state.py accordingly.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Andrea Righi <arighi@nvidia.com>
The tag "ops" is used for two different purposes. First, to indicate that
the entity is directly related to the operations such as flags carried in
sched_ext_ops. Second, to indicate that the entity applies to something
global such as enable or bypass states. The second usage is historical and
causes confusion rather than clarifying anything. For example,
scx_ops_enable_state enums are named SCX_OPS_* and thus conflict with
scx_ops_flags. Let's drop the second usages.
Drop "ops" from scx_ops_enable_state and friends. Update scx_show_state.py
accordingly.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Andrea Righi <arighi@nvidia.com>
A BPF scheduler may want to use the built-in idle cpumasks in ops.init()
before the scheduler is fully initialized, either directly or through a
BPF timer for example.
However, this would result in an error, since the idle state has not
been properly initialized yet.
This can be easily verified by modifying scx_simple to call
scx_bpf_get_idle_cpumask() in ops.init():
$ sudo scx_simple
DEBUG DUMP
===========================================================================
scx_simple[121] triggered exit kind 1024:
runtime error (built-in idle tracking is disabled)
...
Fix this by properly initializing the idle state before ops.init() is
called. With this change applied:
$ sudo scx_simple
local=2 global=0
local=19 global=11
local=23 global=11
...
Fixes: d73249f887 ("sched_ext: idle: Make idle static keys private")
Signed-off-by: Andrea Righi <arighi@nvidia.com>
Reviewed-by: Joel Fernandes <joelagnelf@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
create_dsq and therefore the scx_bpf_create_dsq kfunc currently silently
ignore duplicate entries. As a sched_ext scheduler is creating each DSQ
for a different purpose this is surprising behaviour.
Replace rhashtable_insert_fast which ignores duplicates with
rhashtable_lookup_insert_fast that reports duplicates (though doesn't
return their value). The rest of the code is structured correctly and
this now returns -EEXIST.
Tested by adding an extra scx_bpf_create_dsq to scx_simple. Previously
this was ignored, now init fails with a -17 code. Also ran scx_lavd
which continued to work well.
Signed-off-by: Jake Hillion <jake@hillion.co.uk>
Acked-by: Andrea Righi <arighi@nvidia.com>
Fixes: f0e1a0643a ("sched_ext: Implement BPF extensible scheduler class")
Cc: stable@vger.kernel.org # v6.12+
Signed-off-by: Tejun Heo <tj@kernel.org>
Pull scheduler updates from Ingo Molnar:
"Core & fair scheduler changes:
- Cancel the slice protection of the idle entity (Zihan Zhou)
- Reduce the default slice to avoid tasks getting an extra tick
(Zihan Zhou)
- Force propagating min_slice of cfs_rq when {en,de}queue tasks
(Tianchen Ding)
- Refactor can_migrate_task() to elimate looping (I Hsin Cheng)
- Add unlikey branch hints to several system calls (Colin Ian King)
- Optimize current_clr_polling() on certain architectures (Yujun
Dong)
Deadline scheduler: (Juri Lelli)
- Remove redundant dl_clear_root_domain call
- Move dl_rebuild_rd_accounting to cpuset.h
Uclamp:
- Use the uclamp_is_used() helper instead of open-coding it (Xuewen
Yan)
- Optimize sched_uclamp_used static key enabling (Xuewen Yan)
Scheduler topology support: (Juri Lelli)
- Ignore special tasks when rebuilding domains
- Add wrappers for sched_domains_mutex
- Generalize unique visiting of root domains
- Rebuild root domain accounting after every update
- Remove partition_and_rebuild_sched_domains
- Stop exposing partition_sched_domains_locked
RSEQ: (Michael Jeanson)
- Update kernel fields in lockstep with CONFIG_DEBUG_RSEQ=y
- Fix segfault on registration when rseq_cs is non-zero
- selftests: Add rseq syscall errors test
- selftests: Ensure the rseq ABI TLS is actually 1024 bytes
Membarriers:
- Fix redundant load of membarrier_state (Nysal Jan K.A.)
Scheduler debugging:
- Introduce and use preempt_model_str() (Sebastian Andrzej Siewior)
- Make CONFIG_SCHED_DEBUG unconditional (Ingo Molnar)
Fixes and cleanups:
- Always save/restore x86 TSC sched_clock() on suspend/resume
(Guilherme G. Piccoli)
- Misc fixes and cleanups (Thorsten Blum, Juri Lelli, Sebastian
Andrzej Siewior)"
* tag 'sched-core-2025-03-22' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (40 commits)
cpuidle, sched: Use smp_mb__after_atomic() in current_clr_polling()
sched/debug: Remove CONFIG_SCHED_DEBUG
sched/debug: Remove CONFIG_SCHED_DEBUG from self-test config files
sched/debug, Documentation: Remove (most) CONFIG_SCHED_DEBUG references from documentation
sched/debug: Make CONFIG_SCHED_DEBUG functionality unconditional
sched/debug: Make 'const_debug' tunables unconditional __read_mostly
sched/debug: Change SCHED_WARN_ON() to WARN_ON_ONCE()
rseq/selftests: Fix namespace collision with rseq UAPI header
include/{topology,cpuset}: Move dl_rebuild_rd_accounting to cpuset.h
sched/topology: Stop exposing partition_sched_domains_locked
cgroup/cpuset: Remove partition_and_rebuild_sched_domains
sched/topology: Remove redundant dl_clear_root_domain call
sched/deadline: Rebuild root domain accounting after every update
sched/deadline: Generalize unique visiting of root domains
sched/topology: Wrappers for sched_domains_mutex
sched/deadline: Ignore special tasks when rebuilding domains
tracing: Use preempt_model_str()
xtensa: Rely on generic printing of preemption model
x86: Rely on generic printing of preemption model
s390: Rely on generic printing of preemption model
...
Make scx_select_cpu_dfl() more consistent with the other idle-related
APIs by returning a negative value when an idle CPU isn't found.
No functional changes, this is purely a refactoring.
Signed-off-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Enable passing idle flags (%SCX_PICK_IDLE_*) to scx_select_cpu_dfl(),
to enforce strict selection criteria, such as selecting an idle CPU
strictly within @prev_cpu's node or choosing only a fully idle SMT core.
This functionality will be exposed through a dedicated kfunc in a
separate patch.
Signed-off-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
scx_bpf_reenqueue_local() can be invoked from ops.cpu_release() to give
tasks that are queued to the local DSQ a chance to migrate to other
CPUs, when a CPU is taken by a higher scheduling class.
However, there is no point re-enqueuing tasks that can only run on that
particular CPU, as they would simply be re-added to the same local DSQ
without any benefit.
Therefore, skip per-CPU tasks in scx_bpf_reenqueue_local().
Signed-off-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Add tracing support to track sched_ext core events
(/sched_ext/sched_ext_event). This may be useful for debugging sched_ext
schedulers that trigger a particular event.
The trace point can be used as other trace points, so it can be used in,
for example, `perf trace` and BPF programs, as follows:
======
$> sudo perf trace -e sched_ext:sched_ext_event --filter 'name == "SCX_EV_ENQ_SLICE_DFL"'
======
======
struct tp_sched_ext_event {
struct trace_entry ent;
u32 __data_loc_name;
s64 delta;
};
SEC("tracepoint/sched_ext/sched_ext_event")
int rtp_add_event(struct tp_sched_ext_event *ctx)
{
char event_name[128];
unsigned short offset = ctx->__data_loc_name & 0xFFFF;
bpf_probe_read_str((void *)event_name, 128, (char *)ctx + offset);
bpf_printk("name %s delta %lld", event_name, ctx->delta);
return 0;
}
======
Signed-off-by: Changwoo Min <changwoo@igalia.com>
Acked-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
The event count could be negative in the future,
so change the event type from u64 to s64.
Signed-off-by: Changwoo Min <changwoo@igalia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Pull for-6.14-fixes to receive:
9360dfe4cb ("sched_ext: Validate prev_cpu in scx_bpf_select_cpu_dfl()")
which conflicts with:
337d1b354a ("sched_ext: Move built-in idle CPU selection policy to a separate file")
Signed-off-by: Tejun Heo <tj@kernel.org>
If a BPF scheduler provides an invalid CPU (outside the nr_cpu_ids
range) as prev_cpu to scx_bpf_select_cpu_dfl() it can cause a kernel
crash.
To prevent this, validate prev_cpu in scx_bpf_select_cpu_dfl() and
trigger an scx error if an invalid CPU is specified.
Fixes: f0e1a0643a ("sched_ext: Implement BPF extensible scheduler class")
Cc: stable@vger.kernel.org # v6.12+
Signed-off-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
a6250aa251 ("sched_ext: Handle cases where pick_task_scx() is called
without preceding balance_scx()") added a workaround to handle the cases
where pick_task_scx() is called without prececing balance_scx() which is due
to a fair class bug where pick_taks_fair() may return NULL after a true
return from balance_fair().
The workaround detects when pick_task_scx() is called without preceding
balance_scx() and emulates SCX_RQ_BAL_KEEP and triggers kicking to avoid
stalling. Unfortunately, the workaround code was testing whether @prev was
on SCX to decide whether to keep the task running. This is incorrect as the
task may be on SCX but no longer runnable.
This could lead to a non-runnable task to be returned from pick_task_scx()
which cause interesting confusions and failures. e.g. A common failure mode
is the task ending up with (!on_rq && on_cpu) state which can cause
potential wakers to busy loop, which can easily lead to deadlocks.
Fix it by testing whether @prev has SCX_TASK_QUEUED set. This makes
@prev_on_scx only used in one place. Open code the usage and improve the
comment while at it.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reported-by: Pat Cody <patcody@meta.com>
Fixes: a6250aa251 ("sched_ext: Handle cases where pick_task_scx() is called without preceding balance_scx()")
Cc: stable@vger.kernel.org # v6.12+
Acked-by: Andrea Righi <arighi@nvidia.com>
Similarly to scx_bpf_nr_cpu_ids(), introduce a new kfunc
scx_bpf_nr_node_ids() to expose the maximum number of NUMA nodes in the
system.
BPF schedulers can use this information together with the new node-aware
kfuncs, for example to create per-node DSQs, validate node IDs, etc.
Signed-off-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Using a single global idle mask can lead to inefficiencies and a lot of
stress on the cache coherency protocol on large systems with multiple
NUMA nodes, since all the CPUs can create a really intense read/write
activity on the single global cpumask.
Therefore, split the global cpumask into multiple per-NUMA node cpumasks
to improve scalability and performance on large systems.
The concept is that each cpumask will track only the idle CPUs within
its corresponding NUMA node, treating CPUs in other NUMA nodes as busy.
In this way concurrent access to the idle cpumask will be restricted
within each NUMA node.
The split of multiple per-node idle cpumasks can be controlled using the
SCX_OPS_BUILTIN_IDLE_PER_NODE flag.
By default SCX_OPS_BUILTIN_IDLE_PER_NODE is not enabled and a global
host-wide idle cpumask is used, maintaining the previous behavior.
NOTE: if a scheduler explicitly enables the per-node idle cpumasks (via
SCX_OPS_BUILTIN_IDLE_PER_NODE), scx_bpf_get_idle_cpu/smtmask() will
trigger an scx error, since there are no system-wide cpumasks.
= Test =
Hardware:
- System: DGX B200
- CPUs: 224 SMT threads (112 physical cores)
- Processor: INTEL(R) XEON(R) PLATINUM 8570
- 2 NUMA nodes
Scheduler:
- scx_simple [1] (so that we can focus at the built-in idle selection
policy and not at the scheduling policy itself)
Test:
- Run a parallel kernel build `make -j $(nproc)` and measure the average
elapsed time over 10 runs:
avg time | stdev
---------+------
before: 52.431s | 2.895
after: 50.342s | 2.895
= Conclusion =
Splitting the global cpumask into multiple per-NUMA cpumasks helped to
achieve a speedup of approximately +4% with this particular architecture
and test case.
The same test on a DGX-1 (40 physical cores, Intel Xeon E5-2698 v4 @
2.20GHz, 2 NUMA nodes) shows a speedup of around 1.5-3%.
On smaller systems, I haven't noticed any measurable regressions or
improvements with the same test (parallel kernel build) and scheduler
(scx_simple).
Moreover, with a modified scx_bpfland that uses the new NUMA-aware APIs
I observed an additional +2-2.5% performance improvement with the same
test.
[1] https://github.com/sched-ext/scx/blob/main/scheds/c/scx_simple.bpf.c
Cc: Yury Norov [NVIDIA] <yury.norov@gmail.com>
Signed-off-by: Andrea Righi <arighi@nvidia.com>
Reviewed-by: Yury Norov [NVIDIA] <yury.norov@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Add the new scheduler flag SCX_OPS_BUILTIN_IDLE_PER_NODE, which allows
BPF schedulers to select between using a global flat idle cpumask or
multiple per-node cpumasks.
This only introduces the flag and the mechanism to enable/disable this
feature without affecting any scheduling behavior.
Cc: Yury Norov [NVIDIA] <yury.norov@gmail.com>
Signed-off-by: Andrea Righi <arighi@nvidia.com>
Reviewed-by: Yury Norov [NVIDIA] <yury.norov@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Make all the static keys used by the idle CPU selection policy private
to ext_idle.c. This avoids unnecessary exposure in headers and improves
code encapsulation.
Cc: Yury Norov <yury.norov@gmail.com>
Signed-off-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Add a sysfs entry at /sys/kernel/sched_ext/root/events to expose core
event counters through the files system interface. Each line of the file
shows the event name and its counter value.
In addition, the format of scx_dump_event() is adjusted as the event name
gets longer.
Signed-off-by: Changwoo Min <changwoo@igalia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
A task wakeup can be either processed on the waker's CPU or bounced to the
wakee's previous CPU using an IPI (ttwu_queue). Bouncing to the wakee's CPU
avoids the waker's CPU locking and accessing the wakee's rq which can be
expensive across cache and node boundaries.
When ttwu_queue path is taken, select_task_rq() and thus ops.select_cpu()
may be skipped in some cases (racing against the wakee switching out). As
this confused some BPF schedulers, there wasn't a good way for a BPF
scheduler to tell whether idle CPU selection has been skipped, ops.enqueue()
couldn't insert tasks into foreign local DSQs, and the performance
difference on machines with simple toplogies were minimal, sched_ext
disabled ttwu_queue.
However, this optimization makes noticeable difference on more complex
topologies and a BPF scheduler now has an easy way tell whether
ops.select_cpu() was skipped since 9b671793c7 ("sched_ext, scx_qmap: Add
and use SCX_ENQ_CPU_SELECTED") and can insert tasks into foreign local DSQs
since 5b26f7b920 ("sched_ext: Allow SCX_DSQ_LOCAL_ON for direct
dispatches").
Implement SCX_OPS_ALLOW_QUEUED_WAKEUP which allows BPF schedulers to choose
to enable ttwu_queue optimization.
v2: Update the patch description and comment re. ops.select_cpu() being
skipped in some cases as opposed to always as per Neel.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reported-by: Neel Natu <neelnatu@google.com>
Reported-by: Barret Rhoden <brho@google.com>
Cc: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Andrea Righi <arighi@nvidia.com>
Now when we use scx_bpf_task_cgroup() in ops.tick() to get the cgroup of
the current task, the following error will occur:
scx_foo[3795244] triggered exit kind 1024:
runtime error (called on a task not being operated on)
The reason is that we are using SCX_CALL_OP() instead of SCX_CALL_OP_TASK()
when calling ops.tick(), which triggers the error during the subsequent
scx_kf_allowed_on_arg_tasks() check.
SCX_CALL_OP_TASK() was first introduced in commit 36454023f5 ("sched_ext:
Track tasks that are subjects of the in-flight SCX operation") to ensure
task's rq lock is held when accessing task's sched_group. Since ops.tick()
is marked as SCX_KF_TERMINAL and task_tick_scx() is protected by the rq
lock, we can use SCX_CALL_OP_TASK() to avoid the above issue. Similarly,
the same changes should be made for ops.disable() and ops.exit_task(), as
they are also protected by task_rq_lock() and it's safe to access the
task's task_group.
Fixes: 36454023f5 ("sched_ext: Track tasks that are subjects of the in-flight SCX operation")
Signed-off-by: Chuyi Zhou <zhouchuyi@bytedance.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Pull to receive f3f08c3acf ("sched_ext: Fix incorrect assumption about
migration disabled tasks in task_can_run_on_remote_rq()") which conflicts
with 26176116d9 ("sched_ext: Count SCX_EV_DISPATCH_LOCAL_DSQ_OFFLINE in
the right spot") in for-6.15.
While fixing migration disabled task handling, 3296682157 ("sched_ext: Fix
migration disabled handling in targeted dispatches") assumed that a
migration disabled task's ->cpus_ptr would only have the pinned CPU. While
this is eventually true for migration disabled tasks that are switched out,
->cpus_ptr update is performed by migrate_disable_switch() which is called
right before context_switch() in __scheduler(). However, the task is
enqueued earlier during pick_next_task() via put_prev_task_scx(), so there
is a race window where another CPU can see the task on a DSQ.
If the CPU tries to dispatch the migration disabled task while in that
window, task_allowed_on_cpu() will succeed and task_can_run_on_remote_rq()
will subsequently trigger SCHED_WARN(is_migration_disabled()).
WARNING: CPU: 8 PID: 1837 at kernel/sched/ext.c:2466 task_can_run_on_remote_rq+0x12e/0x140
Sched_ext: layered (enabled+all), task: runnable_at=-10ms
RIP: 0010:task_can_run_on_remote_rq+0x12e/0x140
...
<TASK>
consume_dispatch_q+0xab/0x220
scx_bpf_dsq_move_to_local+0x58/0xd0
bpf_prog_84dd17b0654b6cf0_layered_dispatch+0x290/0x1cfa
bpf__sched_ext_ops_dispatch+0x4b/0xab
balance_one+0x1fe/0x3b0
balance_scx+0x61/0x1d0
prev_balance+0x46/0xc0
__pick_next_task+0x73/0x1c0
__schedule+0x206/0x1730
schedule+0x3a/0x160
__do_sys_sched_yield+0xe/0x20
do_syscall_64+0xbb/0x1e0
entry_SYSCALL_64_after_hwframe+0x77/0x7f
Fix it by converting the SCHED_WARN() back to a regular failure path. Also,
perform the migration disabled test before task_allowed_on_cpu() test so
that BPF schedulers which fail to handle migration disabled tasks can be
noticed easily.
While at it, adjust scx_ops_error() message for !task_allowed_on_cpu() case
for brevity and consistency.
Signed-off-by: Tejun Heo <tj@kernel.org>
Fixes: 3296682157 ("sched_ext: Fix migration disabled handling in targeted dispatches")
Acked-by: Andrea Righi <arighi@nvidia.com>
Reported-by: Jake Hillion <jakehillion@meta.com>
per-CPU cpumasks are dominantly accessed from their own local CPUs,
so allocate them node-local to improve performance.
Signed-off-by: Li RongQing <lirongqing@baidu.com>
Acked-by: Changwoo Min <changwoo@igalia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Count the number of times a migration disabled task is automatically
dispatched to its local DSQ.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Changwoo Min <changwoo@igalia.com>
SCX_EV_DISPATCH_LOCAL_DSQ_OFFLINE wasn't quite right in two aspects:
- It counted both migration disabled and offline events.
- It didn't count events from scx_bpf_dsq_move() path.
Fix it by moving the counting into task_can_run_on_remote_rq() which is
shared by both paths and can distinguish the different rejection conditions.
The argument @trigger_error is renamed to @enforce as it now does more than
just triggering error.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Changwoo Min <changwoo@igalia.com>
Pull to receive:
- 2fa0fbeb69 ("sched_ext: Implement auto local dispatching of migration disabled tasks")
- 3296682157 ("sched_ext: Fix migration disabled handling in targeted dispatches")
as planned for-6.15 changes depend on them (e.g. adding event counter for
implicit migration disabled task handling).
A dispatch operation that can target a specific local DSQ -
scx_bpf_dsq_move_to_local() or scx_bpf_dsq_move() - checks whether the task
can be migrated to the target CPU using task_can_run_on_remote_rq(). If the
task can't be migrated to the targeted CPU, it is bounced through a global
DSQ.
task_can_run_on_remote_rq() assumes that the task is on a CPU that's
different from the targeted CPU but the callers doesn't uphold the
assumption and may call the function when the task is already on the target
CPU. When such task has migration disabled, task_can_run_on_remote_rq() ends
up returning %false incorrectly unnecessarily bouncing the task to a global
DSQ.
Fix it by updating the callers to only call task_can_run_on_remote_rq() when
the task is on a different CPU than the target CPU. As this is a bit subtle,
for clarity and documentation:
- Make task_can_run_on_remote_rq() trigger SCHED_WARN_ON() if the task is on
the same CPU as the target CPU.
- is_migration_disabled() test in task_can_run_on_remote_rq() cannot trigger
if the task is on a different CPU than the target CPU as the preceding
task_allowed_on_cpu() test should fail beforehand. Convert the test into
SCHED_WARN_ON().
Signed-off-by: Tejun Heo <tj@kernel.org>
Fixes: 4c30f5ce4f ("sched_ext: Implement scx_bpf_dispatch[_vtime]_from_dsq()")
Fixes: 0366017e09 ("sched_ext: Use task_can_run_on_remote_rq() test in dispatch_to_local_dsq()")
Cc: stable@vger.kernel.org # v6.12+
Migration disabled tasks are special and pinned to their previous CPUs. They
tripped up some unsuspecting BPF schedulers as their ->nr_cpus_allowed may
not agree with the bits set in ->cpus_ptr. Make it easier for BPF schedulers
by automatically dispatching them to the pinned local DSQs by default. If a
BPF scheduler wants to handle migration disabled tasks explicitly, it can
set SCX_OPS_ENQ_MIGRATION_DISABLED.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Andrea Righi <arighi@nvidia.com>
Add a core event, SCX_EV_ENQ_SLICE_DFL, which represents how many
tasks have been enqueued (or pick_task-ed or select_cpu-ed) with
a default time slice (SCX_SLICE_DFL).
Scheduling a task with SCX_SLICE_DFL unintentionally would be a source
of latency spikes because SCX_SLICE_DFL is relatively long (20 msec).
Thus, soaring the SCX_EV_ENQ_SLICE_DFL value would be a sign of BPF
scheduler bugs, causing latency spikes, especially when ops.select_cpu()
is provided.
__scx_add_event() is used since the caller holds an rq lock or p->pi_lock,
so the preemption has already been disabled.
Signed-off-by: Changwoo Min <changwoo@igalia.com>
Acked-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
