Bart,
I would say it is very possible. There was a period when the pool
locks were switched to _bh spinlocks but that was later reversed back
to _irqsave locks which cleaned up some failures.
I don't know which version Yi Zhang was using. The root cause of this
bug was caused by
librdmacm making verbs API calls while holding _irqsave locks which I
didn't figure out until later.
Bob
On Thu, Apr 7, 2022 at 1:51 PM Bart Van Assche <bvanassche@xxxxxxx> wrote:
>
> On 4/5/22 20:08, Yi Zhang wrote:
> > [ 296.616660] ================================
> > [ 296.620931] WARNING: inconsistent lock state
> > [ 296.625207] 5.18.0-rc1 #1 Tainted: G S I
> > [ 296.630259] --------------------------------
> > [ 296.634531] inconsistent {SOFTIRQ-ON-W} -> {IN-SOFTIRQ-W} usage.
> > [ 296.640535] ksoftirqd/30/188 [HC0[0]:SC1[1]:HE0:SE0] takes:
> > [ 296.646106] ffff888151491468 (&xa->xa_lock#15){+.?.}-{2:2}, at:
> > rxe_pool_get_index+0x72/0x1d0 [rdma_rxe]
>
> Hi Bob,
>
> Do you agree that the root cause of this issue is in the rdma_rxe driver?
>
> Thanks,
>
> Bart.
