Re: [PATCH v2] io-wq: handle hashed writes in chains

[Pavel Begunkov <asml.silence@gmail.com>]

[-- Attachment #1.1: Type: text/plain, Size: 1245 bytes --]

On 23/03/2020 00:16, Pavel Begunkov wrote:
> On 22/03/2020 23:20, Pavel Begunkov wrote:
>> On 22/03/2020 23:15, Jens Axboe wrote:
>>> On 3/22/20 2:05 PM, Jens Axboe wrote:
>>>> On 3/22/20 1:51 PM, Jens Axboe wrote:
>>>>>> Please, tell if you see a hole in the concept. And as said, there is
>>>>>> still a bug somewhere.
>>>>
>>>> One quick guess would be that you're wanting to use both work->list and
>>>> work->data, the latter is used by links which is where you are crashing.
>>>> Didn't check your list management yet so don't know if that's the case,
>>>> but if you're still on the list (or manipulating it after), then that
>>>> would be a bug.
>>>
>>> IOW, by the time you do work->func(&work), the item must be off the
>>> list. Does indeed look like that's exactly the bug you have.
>>>
>>
>> Good guess. I made sure to grab next before ->func(), see next_hashed. And it's
>> not in @work_list, because io_get_next_work() removes it. However, somebody may
>> expect it to be NULL or something. Thanks! I'll check it
> 
> You're right. Re-enqueue corrupts ->data, so should be fixed for previous
> patches as well. I'll send it
> 

And the diff works fine if put on top of the fix!

-- 
Pavel Begunkov


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 833 bytes --]