[PATCH 2/2] nvme-fc: Wait with a timeout for queue to freeze

[Ming Lei]

On Mon, Jul 05, 2021 at 06:34:00PM +0200, Daniel Wagner wrote:
> On Tue, Jun 29, 2021 at 09:39:30AM +0800, Ming Lei wrote:
> > Can you investigate a bit on why there is the hang? FC shouldn't use
> > managed IRQ, so the interrupt won't be shutdown.
> 
> So far, I was not able to figure out why this hangs. In my test setup I
> don't have to do any I/O, I just toggle the remote port.
> 
>   grep busy /sys/kernel/debug/block/*/hctx*/tags | grep -v busy=0
> 
> and this seems to confirm, no I/O in flight.

What is the output of the following command after the hang is triggered?

(cd /sys/kernel/debug/block/nvme0n1 && find . -type f -exec grep -aH . {} \;)

Suppose the hang disk is nvme0n1.

> 
> So I started to look at the q_usage_counter. The obvious observational
> is that counter is not 0. The least bit is set, thus we are in atomic
> mode. 
> 
> (gdb) p/x *((struct request_queue*)0xffff8ac992fbef20)->q_usage_counter->data
> $10 = {
>   count = {
>     counter = 0x8000000000000001
>   }, 
>   release = 0xffffffffa02e78b0, 
>   confirm_switch = 0x0, 
>   force_atomic = 0x0, 
>   allow_reinit = 0x1, 
>   rcu = {
>     next = 0x0, 
>     func = 0x0
>   }, 
>   ref = 0xffff8ac992fbef30
> }
> 
> I am a bit confused about the percpu-refcount API. My naive
> interpretation is that when we are in atomic mode percpu_ref_is_zero()
> can't be used. But this seems rather strange. I must miss something.

No, percpu_ref_is_zero() is fine to be called in atomic mode.


Thanks,
Ming