On Wed, Sep 01, 2021 at 05:36:55PM +0800, Ming Lei wrote: > On Tue, Aug 31, 2021 at 10:27:28AM +0800, luojiaxing wrote: > > Hi, Ming > > > > > > Sorry to reply so late, This issue occur in low probability, > > > > so it take some time to confirm. > > > > > > On 2021/8/26 15:29, Ming Lei wrote: > > > On Thu, Aug 26, 2021 at 11:00:34AM +0800, luojiaxing wrote: > > > > Dear all: > > > > > > > > > > > > I meet some problem when test hisi_sas driver(under SCSI) based on 5.14-rc4 > > > > kernel, it's found that error handle can not be triggered after > > > > > > > > abnormal IO occur in some test with a low probability. For example, > > > > circularly run disk hardreset or disable all local phy of expander when > > > > running fio. > > > > > > > > > > > > We add some tracepoint and print to see what happen, and we got the > > > > following information: > > > > > > > > (1).print rq and rq_state at bt_tags_iter() to confirm how many IOs is > > > > running now. > > > > > > > > <4>[ 897.431182] bt_tags_iter: rqs[2808]: 0xffff202007bd3000; rq_state: 1 > > > > <4>[ 897.437514] bt_tags_iter: rqs[3185]: 0xffff0020c5261e00; rq_state: 1 > > > > <4>[ 897.443841] bt_tags_iter: rqs[3612]: 0xffff00212f242a00; rq_state: 1 > > > > <4>[ 897.450167] bt_tags_iter: rqs[2808]: 0xffff00211d208100; rq_state: 1 > > > > <4>[ 897.456492] bt_tags_iter: rqs[2921]: 0xffff00211d208100; rq_state: 1 > > > > <4>[ 897.462818] bt_tags_iter: rqs[1214]: 0xffff002151d21b00; rq_state: 1 > > > > <4>[ 897.469143] bt_tags_iter: rqs[2648]: 0xffff0020c4bfa200; rq_state: 1 > > > > > > > > The preceding information show that rq with tag[2808] is found in different > > > > hctx by bt_tags_iter() and with different pointer saved in tags->rqs[]. > > > > > > > > And tag[2808] own the same pointer value saved in rqs[] with tag[2921]. It's > > > > wrong because our driver share tag between all hctx, so it's not possible > > > What is your io scheduler? I guess it is deadline, > > > > > > yes > > > > > > > and can you observe > > > such issue by switching to none? > > > > > > Yes, it happen when switched to none > > > > > > > > > > The tricky thing is that one request dumped may be re-allocated to other tag > > > after returning from bt_tags_iter(). > > > > > > > to allocate one tag to different rq. > > > > > > > > > > > > (2).check tracepoints(temporarily add) in blk_mq_get_driver_tag() and > > > > blk_mq_put_tag() to see where this tag is come from. > > > > > > > > Line 1322969: <...>-20189 [013] .... 893.427707: > > > > blk_mq_get_driver_tag: rqs[2808]: 0xffff00211d208100 > > > > Line 1322997: irq/1161-hisi_s-7602 [012] d..1 893.427814: > > > > blk_mq_put_tag_in_free_request: rqs[2808]: 0xffff00211d208100 > > > > Line 1331257: <...>-20189 [013] .... 893.462663: > > > > blk_mq_get_driver_tag: rqs[2860]: 0xffff00211d208100 > > > > Line 1331289: irq/1161-hisi_s-7602 [012] d..1 893.462785: > > > > blk_mq_put_tag_in_free_request: rqs[2860]: 0xffff00211d208100 > > > > Line 1338493: <...>-20189 [013] .... 893.493519: > > > > blk_mq_get_driver_tag: rqs[2921]: 0xffff00211d208100 > > > > > > > > As we can see this rq is allocated to tag[2808] once, and finially come to > > > > tag[2921], but rqs[2808] still save the pointer. > > > Yeah, we know this kind of handling, but not see it as issue. > > > > > > > There will be no problem until we encounter a rare situation. > > > > > > > > For example, tag[2808] is reassigned to another hctx for execution, then > > > > some IO meet some error. > > > I guess the race is triggered when 2808 is just assigned, meantime > > > ->rqs[] isn't updated. > > > > > > As we shared tag between hctx, so if 2808 was assinged to other hctx. > > > > So previous hctx's rqs will not updated。 > > > > > > > > Before waking up the error handle thread, SCSI compares the values of > > > > scsi_host_busy() and shost->host_failed. > > > > > > > > If the values are different, SCSI waits for the completion of some I/Os. > > > > According to the print provided by (1), the return value of scsi_host_busy() > > > > should be 7 for tag [2808] is calculated twice, > > > > > > > > and the value of shost->host_failed is 6. As a result, this two values are > > > > never equal, and error handle cannot be triggered. > > > > > > > > > > > > A temporary workaround is provided and can solve the problem as: > > > > > > > > diff --git a/block/blk-mq-tag.c b/block/blk-mq-tag.c > > > > index 2a37731..e3dc773 100644 > > > > --- a/block/blk-mq-tag.c > > > > +++ b/block/blk-mq-tag.c > > > > @@ -190,6 +190,7 @@ void blk_mq_put_tag(struct blk_mq_tags *tags, struct > > > > blk_mq_ctx *ctx, > > > > BUG_ON(tag >= tags->nr_reserved_tags); > > > > sbitmap_queue_clear(tags->breserved_tags, tag, ctx->cpu); > > > > } > > > > + tags->rqs[tag] = NULL; > > > > } > > > > > > > > > > > > Since we did not encounter this problem in some previous kernel versions, we > > > > wondered if the community already knew about the problem or could provide > > > > some solutions. > > > Can you try the following patch? > > > > > > I tested it. it can fix the bug. > > > > > > However, if there is still a problem in the following scenario? For example, > > driver tag 0 is assigned > > > > to rq0 in hctx0, and reclaimed after rq completed. Next time driver tag 0 is > > still assigned to rq0 but > > > > in hctx1. So at this time, bt_tags_iter will still got two rqs. > > Each hctx has its own rq pool so far, so no such issue you worried. > > John's patch works towards sharing rq pool among hctxs in case of > shared sbitmap, not merged yet, but ->rqs[] should be shared too, still > no such issue. > > Follows the revised patch for handling the stale request in ->rqs[] issue: > > > diff --git a/block/blk-mq-tag.c b/block/blk-mq-tag.c > index 86f87346232a..ff5caeb82542 100644 > --- a/block/blk-mq-tag.c > +++ b/block/blk-mq-tag.c > @@ -208,7 +208,7 @@ static struct request *blk_mq_find_and_get_req(struct blk_mq_tags *tags, > > spin_lock_irqsave(&tags->lock, flags); > rq = tags->rqs[bitnr]; > - if (!rq || !refcount_inc_not_zero(&rq->ref)) > + if (!rq || rq->tag != bitnr || !refcount_inc_not_zero(&rq->ref)) > rq = NULL; > spin_unlock_irqrestore(&tags->lock, flags); > return rq; Explain the issue a bit in detail: scsi_host_check_in_flight() is used to counting scsi in-flight requests after scsi host is blocked, so no new scsi command can be marked as SCMD_STATE_INFLIGHT. However, driver tag allocation still can be run at that time by blk-mq core. The issue is in blk_mq_tagset_busy_iter(). One request is in-flight, but this request may be kept in another slot of ->rqs[], meantime the slot can be allocated out but ->rqs[] isn't updated yet. Then the in-flight request is counted twice as SCMD_STATE_INFLIGHT. Thanks, Ming