On Thu, Apr 17, 2014 at 11:48:21AM -0700, Dan Williams wrote: > The AHCI spec allows implementations to issue commands in tag order > rather than FIFO order: > > 5.3.2.12 P:SelectCmd > HBA sets pSlotLoc = (pSlotLoc + 1) mod (CAP.NCS + 1) > or HBA selects the command to issue that has had the > PxCI bit set to '1' longer than any other command > pending to be issued. > > The result is that commands posted sequentially (time-wise) may play out > of sequence when issued by hardware. > > This behavior has likely been hidden by drives that arrange for commands > to complete in issue order. However, it appears recent drives (two from > different vendors that we have found so far) inflict out-of-order > completions as a matter of course. So, we need to take care to maintain > ordered submission, otherwise we risk triggering a drive to fall out of > sequential-io automation and back to random-io processing, which incurs > large latency and degrades throughput. > > This issue was found in simple benchmarks where QD=2 seq-write > performance was 30-50% *greater* than QD=32 seq-write performance. > > Tagging for -stable and making the change globally since it has a low > risk-to-reward ratio. Also, word is that recent versions of an unnamed > OS also does it this way now. So, drives in the field are already > experienced with this tag ordering scheme. > > Cc: <stable@xxxxxxxxxxxxxxx> > Cc: Dave Jiang <dave.jiang@xxxxxxxxx> > Cc: Ed Ciechanowski <ed.ciechanowski@xxxxxxxxx> > Reviewed-by: Matthew Wilcox <matthew.r.wilcox@xxxxxxxxx> > Signed-off-by: Dan Williams <dan.j.williams@xxxxxxxxx> Applied to libata/for-3.15-fixes. Thanks. -- tejun -- To unsubscribe from this list: send the line "unsubscribe stable" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html