arch/arm64: Fix topology initialization for core scheduling
PATCH v2 1/1 arch/arm64: Fix topology initialization for core scheduling
Arm64 systems rely on store_cpu_topology() to call update_siblings_masks() to transfer the toplogy to the various cpu masks. This needs to be done before the call to notify_cpu_starting() which tells the scheduler about each cpu found, otherwise the core scheduling data structures are setup in a way that does not match the actual topology.
Without this change stress-ng (which enables core scheduling in its prctl tests) causes a warning and then a crash (trimmed for legibility):
[ 1853.805168] ------------[ cut here ]------------ [ 1853.809784] task_rq(b)->core != rq->core [ 1853.809792] WARNING: CPU: 117 PID: 0 at kernel/sched/fair.c:11102 cfs_prio_less+0x1b4/0x1c4 ... [ 1854.015210] Unable to handle kernel NULL pointer dereference at virtual address 0000000000000010 ... [ 1854.231256] Call trace: [ 1854.233689] pick_next_task+0x3dc/0x81c [ 1854.237512] __schedule+0x10c/0x4cc [ 1854.240988] schedule_idle+0x34/0x54
---
This is a similar issue to
f2703def339c ("MIPS: smp: fill in sibling and core maps earlier")
which fixed it for MIPS.
v2: Fixed the commit message. No code change.
arch/arm64/kernel/smp.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
arch/arm64/kernel/smp.c +1 -1
@@ -234,6 +234,7 @@ asmlinkage notrace void secondary_start_kernel(void)* Log the CPU info before it is marked online and might get read. */ cpuinfo_store_cpu (); +store_cpu_topology (cpu);/* * Enable GIC and timers.
@@ -242,7 +243,6 @@ asmlinkage notrace void secondary_start_kernel(void)ipi_setup (cpu); -store_cpu_topology (cpu);numa_add_cpu (cpu);/*
Signature · 1 line
2.18.0
Quoted 20 lines · Phil Auld
Arm64 systems rely on
store_cpu_topology()to callupdate_siblings_masks()to transfer the toplogy to the various cpu masks. This needs to be done before the call tonotify_cpu_starting()which tells the scheduler about each cpu found, otherwise the core scheduling data structures are setup in a way that does not match the actual topology.Without this change stress-ng (which enables core scheduling in its prctl tests) causes a warning and then a crash (trimmed for legibility):
[ 1853.805168] ------------[ cut here ]------------ [ 1853.809784] task_rq(b)->core != rq->core [ 1853.809792] WARNING: CPU: 117 PID: 0 at kernel/sched/fair.c:11102 cfs_prio_less+0x1b4/0x1c4 ... [ 1854.015210] Unable to handle kernel NULL pointer dereference at virtual address 0000000000000010 ... [ 1854.231256] Call trace: [ 1854.233689] pick_next_task+0x3dc/0x81c [ 1854.237512] __schedule+0x10c/0x4cc [ 1854.240988] schedule_idle+0x34/0x54
Quoted 8 lines
Fixes 9edeaea1bc45 ("sched: Core-wide rq->lock")Signed-off-by Phil Auld <[email protected]>---
This is a similar issue to
f2703def339c("MIPS: smp: fill in sibling and core maps earlier")
which fixed it for MIPS.v2: Fixed the commit message. No code change.
Ah, the reason is that smt_mask is not correctly setup, so we bail on cpumask_weight(smt_mask) == 1 for !leaders in:
notify_cpu_starting()
cpuhp_invoke_callback_range()
sched_cpu_starting()
sched_core_cpu_starting()
which leads to rq->core not being correctly set for !leader-rq's.
LGTM. Tested on: HPE Apollo 70 X1
[...]
Quoted 33 lines · On Thu, Mar 31, 2022 at 11:04:31AM +0200 Dietmar Eggemann wrote:
Quoted 20 lines · Phil Auld
Arm64 systems rely on
store_cpu_topology()to callupdate_siblings_masks()to transfer the toplogy to the various cpu masks. This needs to be done before the call tonotify_cpu_starting()which tells the scheduler about each cpu found, otherwise the core scheduling data structures are setup in a way that does not match the actual topology.Without this change stress-ng (which enables core scheduling in its prctl tests) causes a warning and then a crash (trimmed for legibility):
[ 1853.805168] ------------[ cut here ]------------ [ 1853.809784] task_rq(b)->core != rq->core [ 1853.809792] WARNING: CPU: 117 PID: 0 at kernel/sched/fair.c:11102 cfs_prio_less+0x1b4/0x1c4 ... [ 1854.015210] Unable to handle kernel NULL pointer dereference at virtual address 0000000000000010 ... [ 1854.231256] Call trace: [ 1854.233689] pick_next_task+0x3dc/0x81c [ 1854.237512] __schedule+0x10c/0x4cc [ 1854.240988] schedule_idle+0x34/0x54Quoted 8 lines
Fixes 9edeaea1bc45 ("sched: Core-wide rq->lock")Signed-off-by Phil Auld <[email protected]>---
This is a similar issue to
f2703def339c("MIPS: smp: fill in sibling and core maps earlier")
which fixed it for MIPS.v2: Fixed the commit message. No code change.
Ah, the reason is that
smt_maskis not correctly setup, so we bail oncpumask_weight(smt_mask) == 1for !leaders in:
Quoted 7 lines
notify_cpu_starting()
cpuhp_invoke_callback_range()
sched_cpu_starting()
sched_core_cpu_starting()which leads to
rq->corenot being correctly set for !leader-rq's.
Exactly, sorry I was not clearer. smt_mask must be setup correctly by the time sched_core_cpu_starting() is called. (Maybe I should crib some of the above lines into the commit message?)
Quoted 5 lines
LGTM. Tested on: HPE Apollo 70 X1
Reviewed-by Dietmar Eggemann <[email protected]>Tested-by Dietmar Eggemann <[email protected]>
Thanks!
Cheers,
Phil
Quoted 2 lines
[...]
Quoted 2 lines · Phil Auld
Quoted 1 line · On Thu, Mar 31, 2022 at 11:04:31AM +0200 Dietmar Eggemann wrote:
On 30/03/2022 17:56, Phil Auld wrote:
[...]
Quoted 8 lines
Quoted 8 lines
Ah, the reason is that
smt_maskis not correctly setup, so we bail oncpumask_weight(smt_mask) == 1for !leaders in:
notify_cpu_starting()
cpuhp_invoke_callback_range()
sched_cpu_starting()
sched_core_cpu_starting()
Quoted 6 lines
Quoted 2 lines
which leads to
rq->corenot being correctly set for !leader-rq's.Exactly, sorry I was not clearer.
smt_maskmust be setup correctly by the timesched_core_cpu_starting()is called. (Maybe I should crib some of the above lines into the commit message?)
Yeah, maybe, it wouldn't hurt I guess. IMHO mentioning stress-ng's prctl needs PR_SCHED_CORE support could also be handy since today's stress-ng packages don't seem to have this yet.
Quoted 17 lines · On Thu, Mar 31, 2022 at 04:37:50PM +0200 Dietmar Eggemann wrote:
Quoted 2 lines · Phil Auld
Quoted 1 line · On Thu, Mar 31, 2022 at 11:04:31AM +0200 Dietmar Eggemann wrote:
On 30/03/2022 17:56, Phil Auld wrote:
[...]
Quoted 11 lines
Quoted 10 lines
Ah, the reason is that
smt_maskis not correctly setup, so we bail oncpumask_weight(smt_mask) == 1for !leaders in:
notify_cpu_starting()
cpuhp_invoke_callback_range()
sched_cpu_starting()
sched_core_cpu_starting()which leads to
rq->corenot being correctly set for !leader-rq's.
Quoted 8 lines
Quoted 3 lines
Exactly, sorry I was not clearer.
smt_maskmust be setup correctly by the timesched_core_cpu_starting()is called. (Maybe I should crib some of the above lines into the commit message?)Yeah, maybe, it wouldn't hurt I guess. IMHO mentioning stress-ng's prctl needs
PR_SCHED_COREsupport could also be handy since today's stress-ng packages don't seem to have this yet.
My scripts clone it so I did not realize that was not in prepackaged versions yet. But that said, that's really just a way to tickle the problem. Anyone using core scheduling on such a system will hit this (at least the WARN part, the actual crash was harder to create w/o all the threads and tasks stress-ng uses).
I can send a v3 with a further commit message update.
Cheers,
Phil
Quoted 4 lines · Phil Auld
Quoted 3 lines · On Thu, Mar 31, 2022 at 04:37:50PM +0200 Dietmar Eggemann wrote:
Quoted 2 lines · Phil Auld
Quoted 1 line · On Thu, Mar 31, 2022 at 11:04:31AM +0200 Dietmar Eggemann wrote:
On 30/03/2022 17:56, Phil Auld wrote:
[...]
Quoted 10 lines
Quoted 4 lines
Yeah, maybe, it wouldn't hurt I guess. IMHO mentioning stress-ng's prctl needs
PR_SCHED_COREsupport could also be handy since today's stress-ng packages don't seem to have this yet.My scripts clone it so I did not realize that was not in prepackaged versions yet. But that said, that's really just a way to tickle the problem. Anyone using core scheduling on such a system will hit this (at least the WARN part, the actual crash was harder to create w/o all the threads and tasks stress-ng uses).
OK.
I just saw that there is even a kselftest for this:
9f2699007493 - kselftest: Add test for core sched prctl interface (2021-05-12 Chris Hyser)
But it doesn't trigger the issue.
Quoted 1 line
I can send a v3 with a further commit message update.
Sounds good.