On Thu, Dec 13, 2018 at 12:48:11PM -0700, Jens Axboe wrote: > On 12/11/18 4:01 PM, Dennis Zhou wrote: > > The blk-iolatency controller measures the time from rq_qos_throttle() to > > rq_qos_done_bio() and attributes this time to the first bio that needs > > to create the request. This means if a bio is plug-mergeable or > > bio-mergeable, it gets to bypass the blk-iolatency controller. > > > > The recent series, to tag all bios w/ blkgs in [1] changed the timing > > incorrectly as well. First, the iolatency controller was tagging bios > > and using that information if it should process it in rq_qos_done_bio(). > > However, now that all bios are tagged, this caused the atomic_t for the > > struct rq_wait inflight count to underflow resulting in a stall. Second, > > now the timing was using the duration a bio from generic_make_request() > > rather than the timing mentioned above. > > > > This patch fixes these issues by reusing the BLK_QUEUE_ENTERED flag to > > determine if a bio has entered the request layer and is responsible for > > starting a request. Stacked drivers don't recurse through > > blk_mq_make_request(), so the overhead of using time between > > generic_make_request() and the blk_mq_get_request() should be minimal. > > blk-iolatency now checks if this flag is set to determine if it should > > process the bio in rq_qos_done_bio(). > > I'm having a hard time convincing myself that this is correct... Maybe > we should just add a new flag for this specific use case? Or feel free > to convince me otherwise. > I mean it'll work for now, but then when somebody else wants to do something similar *cough*io.weight*cough* it'll need a new flag. I kind of hate adding a new flag for every controller, but then again it's not like there's thousands of these things. I'm having a hard time coming up with a solution other than a per-tracker flag. As for this specific version, I still think it needs to be in iolatency itself, trying to make it generic just means it'll get fucked up again later down the line. Thanks, Josef
