Skip to contentSkip to conversation

sched_ext: Reset dsq_vtime and slice when a task leaves SCX

PATCH Superseded Andrea Righi 2 participants 3 messages Jun 8

Subsystems (2)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore

Andrea Righi

author reviewer
#0 lore raw

PATCH 1/1 sched_ext: Reset dsq_vtime and slice when a task leaves SCX

When a task switches out of the sched_ext class, p->scx.dsq_vtime and p->scx.slice keep whatever values they last held. On enable, slice is reset to slice_dfl, but dsq_vtime is owned by the BPF scheduler and is never cleared by the core, so a task that leaves SCX and later returns carries a stale dsq_vtime across the round-trip.

The stale values are also visible to other SCX schedulers if they inspects the scx fields of non-SCX tasks.

Zero dsq_vtime and slice in switched_from_scx() after scx_disable_task() so the fields are reset to the same baseline a freshly forked task has. The reset is done after the disable callback so the BPF scheduler can still observe the task's final values.

---
 kernel/sched/ext.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)
kernel/sched/ext.c +4 -3
index 8e88a25bc602f..915d5ece5d277 100644
@@ -3942,10 +3942,11 @@ static void switched_from_scx(struct rq *rq, struct task_struct *p)
 	 * scx_disable_task() would WARN on the non-%ENABLED state and trigger a
 	 * NONE -> READY validation failure.
 	 */
-	if (scx_get_task_state(p) == SCX_TASK_NONE)
-		return;
+	if (scx_get_task_state(p) != SCX_TASK_NONE)
+		scx_disable_task(scx_task_sched(p), p);
 
-	scx_disable_task(scx_task_sched(p), p);
+	p->scx.dsq_vtime = 0;
+	p->scx.slice = 0;
 }
 
 static void switched_to_scx(struct rq *rq, struct task_struct *p) {}
Signature · 1 line
2.54.0
#1 lore raw
Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

New issues:
- [High] Tasks returning to the SCX scheduling class receive a zero slice, violating the intended baseline and causing immediate slice expiration.

Pre-existing issues:
- [High] Tasks with an explicit SCHED_EXT policy retain a stale dsq_vtime across global BPF scheduler unloads and reloads, leading to corrupted task prioritization. --

commit 7f8eecaba1a8e23922f489ed563df593e55f3478
Author: Andrea Righi <[email protected]>

sched_ext: Reset dsq_vtime and slice when a task leaves SCX

When a task switches out of the sched_ext class, p->scx.dsq_vtime and p->scx.slice keep whatever values they last held. This patch zeroes dsq_vtime and slice in switched_from_scx() after scx_disable_task() so the fields are reset to the same baseline a freshly forked task has.

Quoted 4 lines
kernel/sched/ext.c +0 -0
index bd8a3332999f4..38157af00fb77 100644

[ ... ]

kernel/sched/ext.c static void switched_from_scx(struct rq *rq, struct task_struct *p) · line 4104

Quoted 12 lines
kernel/sched/ext.c +4 -3
@@ -4097,10 +4097,11 @@ static void switched_from_scx(struct rq *rq, struct task_struct *p)
 	 * scx_disable_task() would WARN on the non-%ENABLED state and trigger a
 	 * NONE -> READY validation failure.
 	 */
-	if (scx_get_task_state(p) == SCX_TASK_NONE)
-		return;
+	if (scx_get_task_state(p) != SCX_TASK_NONE)
+		scx_disable_task(scx_task_sched(p), p);
 
-	scx_disable_task(scx_task_sched(p), p);
+	p->scx.dsq_vtime = 0;
+	p->scx.slice = 0;

[Severity: High]
Does zeroing the slice here mean a task returning to the SCX scheduling class will receive a zero slice?

If neither switched_to_scx() nor __scx_enable_task() resets it to SCX_SLICE_DFL, it seems this violates the commit's intent of restoring the baseline of a freshly forked task.

Will entering the BPF scheduler with a 0 slice cause the core to emit a warned_zero_slice warning and trigger immediate slice exhaustion upon dispatch (falling back to a 1 nanosecond slice)?

[Severity: High]
This isn't a bug introduced by this patch, but does the approach of clearing dsq_vtime here miss tasks during global BPF scheduler unloads and reloads?

If tasks have an explicit SCHED_EXT policy, they remain in the ext_sched_class when a custom BPF scheduler is unloaded. During scx_root_disable() and scx_root_enable(), old_class == new_class evaluates to true, so switched_from_scx() is never invoked and dsq_vtime is never cleared.

When the new BPF scheduler is loaded, will these tasks retain the dsq_vtime from the previous scheduler, causing incorrect sorting and prioritization in vtime-sorted dispatch queues?

Quoted 3 lines
kernel/sched/ext.c +0 -0
 }
 
 static void switched_to_scx(struct rq *rq, struct task_struct *p) {}
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
author reviewer
#2 lore raw
Quoted 41 lines · On Mon, Jun 08, 2026 at 02:11:11PM +0000, [email protected] wrote:

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

New issues:
- [High] Tasks returning to the SCX scheduling class receive a zero slice, violating the intended baseline and causing immediate slice expiration.

Pre-existing issues:
- [High] Tasks with an explicit SCHED_EXT policy retain a stale dsq_vtime across global BPF scheduler unloads and reloads, leading to corrupted task prioritization. --

commit 7f8eecaba1a8e23922f489ed563df593e55f3478
Author: Andrea Righi <[email protected]>

sched_ext: Reset dsq_vtime and slice when a task leaves SCX

When a task switches out of the sched_ext class, p->scx.dsq_vtime and p->scx.slice keep whatever values they last held. This patch zeroes dsq_vtime and slice in switched_from_scx() after scx_disable_task() so the fields are reset to the same baseline a freshly forked task has.

Quoted 4 lines
kernel/sched/ext.c +0 -0
index bd8a3332999f4..38157af00fb77 100644

[ ... ]

kernel/sched/ext.c static void switched_from_scx(struct rq *rq, struct task_struct *p) · line 4104

Quoted 12 lines
kernel/sched/ext.c +4 -3
@@ -4097,10 +4097,11 @@ static void switched_from_scx(struct rq *rq, struct task_struct *p)
 	 * scx_disable_task() would WARN on the non-%ENABLED state and trigger a
 	 * NONE -> READY validation failure.
 	 */
-	if (scx_get_task_state(p) == SCX_TASK_NONE)
-		return;
+	if (scx_get_task_state(p) != SCX_TASK_NONE)
+		scx_disable_task(scx_task_sched(p), p);
 
-	scx_disable_task(scx_task_sched(p), p);
+	p->scx.dsq_vtime = 0;
+	p->scx.slice = 0;

[Severity: High]
Does zeroing the slice here mean a task returning to the SCX scheduling class will receive a zero slice?

Quoted 7 lines

If neither switched_to_scx() nor __scx_enable_task() resets it to SCX_SLICE_DFL, it seems this violates the commit's intent of restoring the baseline of a freshly forked task.

Will entering the BPF scheduler with a 0 slice cause the core to emit a warned_zero_slice warning and trigger immediate slice exhaustion upon dispatch (falling back to a 1 nanosecond slice)?

This looks like a false positive. And I think it's semantically correct to reset the slice to zero when a task exits the SCHED_EXT class. I'll rephrase the "freshly forked task" part, which might be confusing.

Quoted 5 lines

[Severity: High]
This isn't a bug introduced by this patch, but does the approach of clearing dsq_vtime here miss tasks during global BPF scheduler unloads and reloads?

Quoted 8 lines

If tasks have an explicit SCHED_EXT policy, they remain in the ext_sched_class when a custom BPF scheduler is unloaded. During scx_root_disable() and scx_root_enable(), old_class == new_class evaluates to true, so switched_from_scx() is never invoked and dsq_vtime is never cleared.

When the new BPF scheduler is loaded, will these tasks retain the dsq_vtime from the previous scheduler, causing incorrect sorting and prioritization in vtime-sorted dispatch queues?

This one instead looks legit, we should probably reset dsq_vtime and slice in scx_disable_task() after ops.disable() has observed their final values. I'll fix this in v2.

-Andrea