The root cause:
After just cached_dev_free do cancel_writeback_rate_update_dwork
without bch_register_lock
.
at the same time. Wirting the writeback_percent by sysfs witch
bch_register_lock will insert a writeback_rate_update work.
cached_dev_free with bch_register_lock to do bcache_device_free.
(it is introduce by patch 80265d8dfd77792e133793cef44a21323aac2908)
pls：
1： run the shell script
#!/bin/bash
while [ true ]
do
echo 0 > /sys/block/bcache0/bcache/writeback_percent
done
2: hotplug the cache disk
On 12/3/20, Coly Li <colyli@xxxxxxx> wrote:
> On 12/3/20 2:25 PM, Yi Li wrote:
>>> On 12/1/20 12:35 PM, Yi Li wrote:
>>>> sorry, This patch will cause deadlock, i will check and redo it.
>>>
>>> Can you try latest upstream kernel firstly ? Before spending more time
>>> on the fix.
>>>
>>
>> This issue just happened three times （xenserver7.5 dom0 kernel） on the
>> same machine and cannot reproduce it now. and have not reproduce it
>> using the lastest uptream kernel.
>>
>
> Hmm, this is something very probably that I am not able to help. It
> seems the kernel is a third-part maintained Linux v4.4 based kernel +
> bcache backport, which is out of my view.
>
> If similar problem happens on latest upstream kernel, or at least v5.8+
> kernel, I can help to take a look.
>
>
>>> If I remember correctly, when cancel_writeback_rate_update_dwork() is
>>> not timed out, the cache set memory won't be freed before the
>>> writeback_rate_update worker terminates. It is possible that I miss
>>> something in the code, but I suggest to test with a kernel after v5.3,
>>> and better a v5.8+ kernel.
>>>
>>> Coly Li
>>>
>> Thanks.
>>
>> it is confused that why writeback_rate_update worker run again after
>> cancel_delayed_work_sync( kernel log telled).
>>
>
> [snipped]
>
> Coly Li
>
