os: fix amp hang issue #6623
os: fix amp hang issue #6623
Conversation
|
Hi @gSahitya-samsung since your commit includes my change of including DMB in arm_pause_handler, is #6619 still required? |
@gSahitya-samsung Let's split commits for @lhenry-realtek 's change. Could you cherry pick his commit here? |
gSahitya-samsung
force-pushed
the
smp_hang
branch
from
3c4c758 to
6ee637a
Compare
gSahitya-samsung
changed the title
|
As you know, I think, The rootcause is missed enter_critical_section in up_schedule_sigaction(). #6624 Could you please update if below commit is not needed, Please update it |
Hi @ewoodev, this commit is required, the The |
gSahitya-samsung
force-pushed
the
smp_hang
branch
3 times, most recently
from
7360022 to
be35154
Compare
gSahitya-samsung
changed the title
| @@ -277,7 +278,7 @@ int arm_pause_handler(int irq, void *context, void *arg) | |||
| * interrupt by calling up_cpu_paused(). If the pause event has already | |||
| * been processed then g_cpu_paused[cpu] will not be locked. | |||
| */ | |||
|
|
|||
There was a problem hiding this comment.
Let's remove this unnecessary change.
…spinlock variable is observed in ISR This commit addresses a condition where the spinlock variable becomes un-synced between cores due to out-of-order memory access. DMB is added to ensure the latest copy of the lock is retrieved before operating on it
To pause the cpu, cpu should call `up_cpu_paused`. The call to `up_cpu_paused` will get missed when `g_cpu_irqlock` is unlocked (means none of the cpu having access to irqlock), `arm_pause_handler` rely on `enter_critical_section` to pause the cpu, it must handle `up_cpu_pausereq` without any fail, otherwise other cpu will wait forever in spinlock. Change `while` loop to `do-while` loop in `irq_waitlock`, which will ensure there is no pending `up_cpu_pausereq` before getting lock on `g_cpu_irqlock` otherwise pause request will get missed. The `SP_DMB` is needed to ensure both cpu read correct spinlock value. The `spin_lock(&g_cpu_resumed[cpu])` is moved upward in code to avoid race condition between cpus.
gSahitya-samsung
dismissed stale reviews from sunghan-chang and r-prabu
via
7482d8c
gSahitya-samsung
force-pushed
the
smp_hang
branch
from
be35154 to
7482d8c
Compare
To pause the cpu, cpu should call
up_cpu_paused. The call toup_cpu_pausedwill get missed wheng_cpu_irqlockis unlocked (means none of the cpu having access to irqlock),arm_pause_handlerrely onenter_critical_sectionto pause the cpu, it must handleup_cpu_pausereqwithout any fail, otherwise other cpu will wait forever in spinlock.Change
whileloop todo-whileloop inirq_waitlock, which will ensure there is no pendingup_cpu_pausereqbefore getting lock ong_cpu_irqlockotherwise pause request will get missed.The
SP_DMBis needed to ensure both cpu read correct spinlock value.The
spin_lock(&g_cpu_resumed[cpu])is moved upward in code to avoid race condition between cpus.