On Wed, Oct 27, 2021 at 09:06:05AM -0600, Jens Axboe wrote: > On 10/27/21 9:03 AM, Ming Lei wrote: > > On Wed, Oct 27, 2021 at 07:12:31AM -0700, Keith Busch wrote: > >> On Wed, Oct 27, 2021 at 06:16:19AM -0700, Bart Van Assche wrote: > >>> On 10/26/21 10:27 PM, Christoph Hellwig wrote: > >>>> On Tue, Oct 26, 2021 at 01:10:47PM -0700, Bart Van Assche wrote: > >>>>> If blk_insert_cloned_request() is moved into the device mapper then I > >>>>> think that blk_mq_request_issue_directly() will need to be exported. > >>>> > >>>> Which is even worse. > >>>> > >>>>> How > >>>>> about the (totally untested) patch below for removing the > >>>>> blk_insert_cloned_request() call from the UFS-HPB code? > >>>> > >>>> Which again doesn't fix anything. The problem is that it fans out one > >>>> request into two on the same queue, not the specific interface used. > >>> > >>> That patch fixes the reported issue, namely removing the additional accounting > >>> caused by calling blk_insert_cloned_request(). Please explain why it is > >>> considered wrong to fan out one request into two. That code could be reworked > >>> such that the block layer is not involved as Adrian Hunter explained. However, > >>> before someone spends time on making these changes I think that someone should > >>> provide more information about why it is considered wrong to fan out one request > >>> into two. > >> > >> The original request consumes a tag from that queue's tagset. If the > >> lifetime of that tag depends on that same queue having another free tag, > >> you can deadlock. > > > > Just take a quick look at the code, if the spawned request can't be allocated, > > scsi will return BLK_STS_RESOURCE for the original scsi request which will be > > retried later by blk-mq. > > > > So if tag depth is > 1 and max allowed inflight write buffer command is limited > > as 1, there shouldn't be the deadlock. > > > > Or is it possible to reuse the original scsi request's tag for the > > spawned request? Like the trick used in inserting flush request. > > The flush approach did come to mind here as well, but honestly that one is > very ugly and would never have been permitted if it wasn't excluded to be > in the very core code already. But yes, reuse of the existing request is > probably another potentially viable approach. My worry there is that > inevitably you end up needing to stash a lot of data to restore the original, > and we're certainly not adding anything to struct request for that. > > Hence I think being able to find a new request reliably would be better. request with scsi_cmnd may be allocated by the ufshpb driver, even it should be fine to call ufshcd_queuecommand() directly for this driver private IO, if the tag can be reused. One example is scsi_ioctl_reset(). Thanks, Ming