On Fri, 2018-07-20 at 10:23 -0600, Keith Busch wrote: +AD4- On Fri, Jul 20, 2018 at 04:20:01PM +-0000, Bart Van Assche wrote: +AD4- +AD4- On Fri, 2018-07-20 at 10:12 -0600, Keith Busch wrote: +AD4- +AD4- +AD4- On Fri, Jul 20, 2018 at 04:03:18PM +-0000, Bart Van Assche wrote: +AD4- +AD4- +AD4- +AD4- On Fri, 2018-07-20 at 09:56 -0600, Keith Busch wrote: +AD4- +AD4- +AD4- +AD4- +AD4- SCSI is the only block driver that wants this behavior. Moving it back +AD4- +AD4- +AD4- +AD4- +AD4- to generic where it used to be breaks other block drivers. +AD4- +AD4- +AD4- +AD4- +AD4- +AD4- +AD4- +AD4- That's new to me. What would break in the NVMe driver if the above change would be +AD4- +AD4- +AD4- +AD4- present in the block layer? +AD4- +AD4- +AD4- +AD4- +AD4- +AD4- This is what causes the block layer to lose completions, and most drivers +AD4- +AD4- +AD4- don't want the kernel to lose their completions. +AD4- +AD4- +AD4- +AD4- Hello Keith, +AD4- +AD4- +AD4- +AD4- Have you considered to introduce a fourth state for block layer requests to +AD4- +AD4- avoid that completions that occur while a timeout handler is in progress get +AD4- +AD4- lost? That would avoid that completions get lost not only for the NVMe driver +AD4- +AD4- but also for SCSI drivers. See e.g. the MQ+AF8-RQ+AF8-TIMED+AF8-OUT state in +AD4- +AD4- https://www.mail-archive.com/linux-block+AEA-vger.kernel.org/msg22196.html +AD4- +AD4- Yes, I've considered that, and I really want to use it, but scsi may +AD4- still reference a freed request in scmd+AF8-eh+AF8-abort+AF8-handler that way. I think that's a misunderstanding. If scsi+AF8-times+AF8-out() queues an abort asynchronously then it tells the block layer through its return value that the SCSI core still owns the request and hence that the block layer should ignore any completions that occur until the SCSI core calls scsi+AF8-finish+AF8-command(). That scsi+AF8-finish+AF8-command() will trigger a call to +AF8AXw-blk+AF8-mq+AF8-end+AF8-request(). The scsi+AF8-times+AF8-out() return value I was referring to is called BLK+AF8-EH+AF8-DONE today and was called BLK+AF8-EH+AF8-NOT+AF8-HANDLED in kernel version v4.17. This also means that I got the BLK+AF8-EH+AF8-NOT+AF8-HANDLED case wrong in +ACI-blk-mq: Rework blk-mq timeout handling again+ACI-: in that case concurrent a blk+AF8-mq+AF8-complete+AF8-request() call should be ignored instead of triggering request completion. Bart.
