On Wed, Nov 27, 2024 at 02:18:34PM +0900, Damien Le Moal wrote: > After some debugging, I understood the issue. It is not a problem with the > queue usage counter but rather an issue with REQ_NOWAIT BIOs that may be failed > with bio_wouldblock_error() *after* having been processed by > blk_zone_plug_bio(). E.g. blk_mq_get_new_requests() may fail to get a request > due to REQ_NOWAIT being set and fail the BIO using bio_wouldblock_error(). This > in turn will lead to a call to disk_zone_wplug_set_error() which will mark the > zone write plug with the ERROR flag. However, since the BIO failure is not from > a failed request, we are not calling disk_zone_wplug_unplug_bio(), which if > called would run the error recovery for the zone. That is, the zone write plug > of the BLK_STS_AGAIN failed BIO is left "busy", which result in the BIO to be > added to the plug BIO list when it is re-submitted again. But we donot have any > write BIO on-going for this zone, the BIO ends up stuck in the zone write plug > list holding a queue reference count, which causes the freeze to never terminate. Did you trace where the bio_wouldblock_error is coming from? Probably a failing request allocation? Can we call the guts of blk_zone_plug_bio after allocating the request to avoid this?
