Hi guys,

While investigating the performance issue reported by Ming [0], I am
seeing
this hang in certain scenarios:

tivated0KB /s] [0/0/0 iops] [eta 1158048815d:13h:31m:49s] [ 740.499917]
rcu: INFO: rcu_preempt detected stalls on CPUs/tasks:ops] [eta
34722d:05h:17m:25s] [ 740.505994] rcu: Tasks blocked on level-1 rcu_node
(CPUs 0-15):
[ 740.511982] (detected by 64, t=5255 jiffies, g=6105, q=6697) [
740.517703]
rcu: All QSes seen, last rcu_preempt kthread activity 0 (4295075897-
4295075897), jiffies_till_next_fqs=1, root ->qsmask 0x1 [ 740.723625] BUG:
scheduling while atomic: swapper/64/0/0x00000008 [ 740.729692] Modules
linked in:
[ 740.732737] CPU: 64 PID: 0 Comm: swapper/64 Tainted: G W 5.12.0-rc7-
g7589ed97c1da-dirty #322 [ 740.742432] Hardware name: Huawei TaiShan
2280 V2/BC82AMDC, BIOS
2280-V2 CS V5.B133.01 03/25/2021
[ 740.751264] Call trace:
[ 740.753699] dump_backtrace+0x0/0x1b0
[ 740.757353] show_stack+0x18/0x68
[ 740.760654] dump_stack+0xd8/0x134
[ 740.764046] __schedule_bug+0x60/0x78
[ 740.767694] __schedule+0x620/0x6d8
[ 740.771168] schedule_idle+0x20/0x40
[ 740.774730] do_idle+0x19c/0x278
[ 740.777945] cpu_startup_entry+0x24/0x68 [ 740.781850]
secondary_start_kernel+0x178/0x188
[ 740.786362] 0x0
^Cbs: 12 (f=12): [r(12)] [0.0% done] [1626MB/0KB/0KB /s] [416K/0/0 iops]
[eta
34722d:05h:16m:28s]
fio: terminating on signal 2

I thought it merited a separate thread.

[ 740.723625] BUG: scheduling while atomic: swapper/64/0/0x00000008
Looks bad ...

The scenario to create seems to be running fio with rw=randread and mq-
deadline IO scheduler. And heavily loading the system - running fio on a
subset of available CPUs seems to help (recreate).

When it occurs, the system becomes totally unresponsive.

It could be a LLDD bug, but I am doubtful.

Has anyone else seen this or help try to recreate?

John - I have not seen such issue on megaraid_sas driver. Is this something
to do with CPU lock up ?
Can you try your test with "rq_affinity=2" ? megaraid_sas driver detect CPU
lockup (flood of completion on single CPU) and it use irq_poll interface to
avoid such loop.
Since you mentioned you noticed issue with hisi_sas v2 without hostwide tag
I can think of similar stuffs in this case.

How cpus to irq affinity settled in your case. ? Is it 1-1 mapping ?

Kashyap

scsi debug or null_blk don't seem to load the system heavily enough to
recreate.

I have seen it on 5.11 also. I see it on hisi_sas v2 and v3 hw drivers,
And I don't
think it's related to hostwide tags, as for hisi_sas v2 hw driver, I unset
that flag
and can still see it.

Thanks,
John

[0]
https://lore.kernel.org/linux-scsi/89ebc37c-21d6-c57e-4267-
cac49a3e5953@huawei.com/T/#t