On 5/4/21 1:57 AM, Martin Wilck wrote:
> No, I can't. I haven't reviewed every caller. I thought about the
> function's documentation, which simply says "find the tagged command by
> host". Maybe I got this wrong.
>
> You're right that returning non-in-flight requests makes no sense for
> drivers calling this function in the request completion code path.
>
> The situation was a bit different until recently, when drivers used
> scsi_host_find_tag() also for iterating over commands. The commit
> message of e73a5e8e8003 stated that it avoids 'random errors' in this
> case; I am not sure if that's actually the case. I can at least imagine
> that a driver would want to abort or otherwise invalidate requests even
> if they're not in flight yet (e.g. when shutting down a controller).
> As the drivers have now been fixed to use blk_mq_tagset_busy_iter(),
> there's no need to discuss this further.
Unless if there is a strong argument I'd like to keep commit
e73a5e8e8003 ("scsi: core: Only return started requests from
scsi_host_find_tag()"). I think that commit implements the behavior that
is needed in the completion path of SCSI LLDs.
Thanks,
Bart.
