On 11/26/21 9:21 AM, Jens Axboe wrote:
> On 11/26/21 2:53 AM, Shinichiro Kawasaki wrote:
>> I ran my test set on v5.16-rc2 and observed a process hang. The test work load
>> repeats file creation on xfs on dm-zoned. This dm-zoned device is on top of 3
>> dm-linear devices. One of them is dm-linear device on non-zoned NVMe device as
>> the cache of the dm-zoned device. The other two are dm-linear devices on zoned
>> SMR HDDs. So far, the hang is recreated 100% with my test system.
>>
>> The kernel message [2] reported hanging tasks. In the call stack, I observe
>> wbt_wait(). Also I observed "inflight 1" value in the "rqos/wbt/inflight"
>> attribute of debug sysfs.
>>
>> # grep -R . /sys/kernel/debug/block/nvme0n1 | grep inflight
>> /sys/kernel/debug/block/nvme0n1/rqos/wbt/inflight:0: inflight 1
>> /sys/kernel/debug/block/nvme0n1/rqos/wbt/inflight:1: inflight 0
>> /sys/kernel/debug/block/nvme0n1/rqos/wbt/inflight:2: inflight 0
>>
>> These symptoms look related to another issue reported to linux-block [1]. As
>> discussed in that thread, I set 0 to /sys/block/nvme0n1/queue/wbt_lat_usec.
>> With this setting, I observed the hang disappeared. Then this hang I observe
>> also related to writeback throttling for the NVMe device.
>>
>> I bisected and found the commit 4f5022453acd ("nvme: wire up completion batching
>> for the IRQ path") is the trigger commit. I reverted this commit from v5.16-rc2,
>> and observed the hang disappeared.
>>
>> Wish this report helps.
>>
>>
>> [1] https://lore.kernel.org/linux-block/b3ba57a7-d363-9c17-c4be-9dbe86875@xxxxxxxxx
>
> Yes looks the same as that one, and that commit was indeed my suspicion
> on what could potentially cause the accounting discrepancy. I'll take a
> look at this.
I sent out a patch in the other thread, please give that a whirl.
--
Jens Axboe
