On 2024/04/22 16:25, Christoph Hellwig wrote: > On Sat, Apr 20, 2024 at 04:58:11PM +0900, Damien Le Moal wrote: >> A zone write plug BIO work function blk_zone_wplug_bio_work() calls >> submit_bio_noacct_nocheck() to execute the next unplugged BIO. This >> function may block. So executing zone plugs BIO works using the block >> layer global kblockd workqueue can potentially lead to preformance or >> latency issues as the number of concurrent work for a workqueue is >> limited to WQ_DFL_ACTIVE (256). >> 1) For a system with a large number of zoned disks, issuing write >> requests to otherwise unused zones may be delayed wiating for a work >> thread to become available. >> 2) Requeue operations which use kblockd but are independent of zone >> write plugging may alsoi end up being delayed. >> >> To avoid these potential performance issues, create a workqueue per >> zoned device to execute zone plugs BIO work. The workqueue max active >> parameter is set to the maximum number of zone write plugs allocated >> with the zone write plug mempool. This limit is equal to the maximum >> number of open zones of the disk and defaults to 128 for disks that do >> not have a limit on the number of open zones. > > Looks good: > > Reviewed-by: Christoph Hellwig <hch@xxxxxx> > > Should the zone write plug submission do non-blocking submissions as well > to avoid stalling in the workqueue thread all the time? I do not think that the stalling actually happens that often. The 2 main cases I see are: 1) Out of tag so we block on tag allocation when preparing the request in submit_bio_noacct_nocheck(), or 2) The device has BLK_MQ_F_BLOCKING set for its tag set (e.g. nullblk with memory backing). For (1), we could use RQF_NOWAIT to prevent blocking, but then we would need to retry later on with a timer to make forward progress for the plug. And I do not think we can actually avoid (2). So in the end, I do not see a clean way to completely avoid blocking in all cases. -- Damien Le Moal Western Digital Research
