virtio-blk: Fix WARN_ON_ONCE in virtio_queue_rq()
authorSuwan Kim <suwan.kim027@gmail.com>
Tue, 30 Aug 2022 15:01:53 +0000 (00:01 +0900)
committerMichael S. Tsirkin <mst@redhat.com>
Tue, 27 Sep 2022 22:30:49 +0000 (18:30 -0400)
If a request fails at virtio_queue_rqs(), it is inserted to requeue_list
and passed to virtio_queue_rq(). Then blk_mq_start_request() can be called
again at virtio_queue_rq() and trigger WARN_ON_ONCE like below trace because
request state was already set to MQ_RQ_IN_FLIGHT in virtio_queue_rqs()
despite the failure.

[    1.890468] ------------[ cut here ]------------
[    1.890776] WARNING: CPU: 2 PID: 122 at block/blk-mq.c:1143
blk_mq_start_request+0x8a/0xe0
[    1.891045] Modules linked in:
[    1.891250] CPU: 2 PID: 122 Comm: journal-offline Not tainted 5.19.0+ #44
[    1.891504] Hardware name: ChromiumOS crosvm, BIOS 0
[    1.891739] RIP: 0010:blk_mq_start_request+0x8a/0xe0
[    1.891961] Code: 12 80 74 22 48 8b 4b 10 8b 89 64 01 00 00 8b 53
20 83 fa ff 75 08 ba 00 00 00 80 0b 53 24 c1 e1 10 09 d1 89 48 34 5b
41 5e c3 <0f> 0b eb b8 65 8b 05 2b 39 b6 7e 89 c0 48 0f a3 05 39 77 5b
01 0f
[    1.892443] RSP: 0018:ffffc900002777b0 EFLAGS: 00010202
[    1.892673] RAX: 0000000000000000 RBX: ffff888004bc0000 RCX: 0000000000000000
[    1.892952] RDX: 0000000000000000 RSI: ffff888003d7c200 RDI: ffff888004bc0000
[    1.893228] RBP: 0000000000000000 R08: 0000000000000001 R09: ffff888004bc0100
[    1.893506] R10: ffffffffffffffff R11: ffffffff8185ca10 R12: ffff888004bc0000
[    1.893797] R13: ffffc90000277900 R14: ffff888004ab2340 R15: ffff888003d86e00
[    1.894060] FS:  00007ffa143a4640(0000) GS:ffff88807dd00000(0000)
knlGS:0000000000000000
[    1.894412] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[    1.894682] CR2: 00005648577d9088 CR3: 00000000053da004 CR4: 0000000000170ee0
[    1.894953] Call Trace:
[    1.895139]  <TASK>
[    1.895303]  virtblk_prep_rq+0x1e5/0x280
[    1.895509]  virtio_queue_rq+0x5c/0x310
[    1.895710]  ? virtqueue_add_sgs+0x95/0xb0
[    1.895905]  ? _raw_spin_unlock_irqrestore+0x16/0x30
[    1.896133]  ? virtio_queue_rqs+0x340/0x390
[    1.896453]  ? sbitmap_get+0xfa/0x220
[    1.896678]  __blk_mq_issue_directly+0x41/0x180
[    1.896906]  blk_mq_plug_issue_direct+0xd8/0x2c0
[    1.897115]  blk_mq_flush_plug_list+0x115/0x180
[    1.897342]  blk_add_rq_to_plug+0x51/0x130
[    1.897543]  blk_mq_submit_bio+0x3a1/0x570
[    1.897750]  submit_bio_noacct_nocheck+0x418/0x520
[    1.897985]  ? submit_bio_noacct+0x1e/0x260
[    1.897989]  ext4_bio_write_page+0x222/0x420
[    1.898000]  mpage_process_page_bufs+0x178/0x1c0
[    1.899451]  mpage_prepare_extent_to_map+0x2d2/0x440
[    1.899603]  ext4_writepages+0x495/0x1020
[    1.899733]  do_writepages+0xcb/0x220
[    1.899871]  ? __seccomp_filter+0x171/0x7e0
[    1.900006]  file_write_and_wait_range+0xcd/0xf0
[    1.900167]  ext4_sync_file+0x72/0x320
[    1.900308]  __x64_sys_fsync+0x66/0xa0
[    1.900449]  do_syscall_64+0x31/0x50
[    1.900595]  entry_SYSCALL_64_after_hwframe+0x63/0xcd
[    1.900747] RIP: 0033:0x7ffa16ec96ea
[    1.900883] Code: b8 4a 00 00 00 0f 05 48 3d 00 f0 ff ff 77 41 c3
48 83 ec 18 89 7c 24 0c e8 e3 02 f8 ff 8b 7c 24 0c 89 c2 b8 4a 00 00
00 0f 05 <48> 3d 00 f0 ff ff 77 36 89 d7 89 44 24 0c e8 43 03 f8 ff 8b
44 24
[    1.901302] RSP: 002b:00007ffa143a3ac0 EFLAGS: 00000293 ORIG_RAX:
000000000000004a
[    1.901499] RAX: ffffffffffffffda RBX: 0000560277ec6fe0 RCX: 00007ffa16ec96ea
[    1.901696] RDX: 0000000000000000 RSI: 0000000000000000 RDI: 0000000000000016
[    1.901884] RBP: 0000560277ec5910 R08: 0000000000000000 R09: 00007ffa143a4640
[    1.902082] R10: 00007ffa16e4d39e R11: 0000000000000293 R12: 00005602773f59e0
[    1.902459] R13: 0000000000000000 R14: 00007fffbfc007ff R15: 00007ffa13ba4000
[    1.902763]  </TASK>
[    1.902877] ---[ end trace 0000000000000000 ]---

To avoid calling blk_mq_start_request() twice, This patch moves the
execution of blk_mq_start_request() to the end of virtblk_prep_rq().
And instead of requeuing failed request to plug list in the error path of
virtblk_add_req_batch(), it uses blk_mq_requeue_request() to change failed
request state to MQ_RQ_IDLE. Then virtblk can safely handle the request
on the next trial.

Fixes: 0e9911fa768f ("virtio-blk: support mq_ops->queue_rqs()")
Reported-by: Alexandre Courbot <acourbot@chromium.org>
Tested-by: Alexandre Courbot <acourbot@chromium.org>
Signed-off-by: Suwan Kim <suwan.kim027@gmail.com>
Message-Id: <20220830150153.12627-1-suwan.kim027@gmail.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Acked-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Pankaj Raghav <p.raghav@samsung.com>
drivers/block/virtio_blk.c

index 30255fcaf18121c5051ba156e821ee0fd33ba4aa..dd9a05174726bceae82339d63f9cd4496057b8cb 100644 (file)
@@ -322,14 +322,14 @@ static blk_status_t virtblk_prep_rq(struct blk_mq_hw_ctx *hctx,
        if (unlikely(status))
                return status;
 
-       blk_mq_start_request(req);
-
        vbr->sg_table.nents = virtblk_map_data(hctx, req, vbr);
        if (unlikely(vbr->sg_table.nents < 0)) {
                virtblk_cleanup_cmd(req);
                return BLK_STS_RESOURCE;
        }
 
+       blk_mq_start_request(req);
+
        return BLK_STS_OK;
 }
 
@@ -391,8 +391,7 @@ static bool virtblk_prep_rq_batch(struct request *req)
 }
 
 static bool virtblk_add_req_batch(struct virtio_blk_vq *vq,
-                                       struct request **rqlist,
-                                       struct request **requeue_list)
+                                       struct request **rqlist)
 {
        unsigned long flags;
        int err;
@@ -408,7 +407,7 @@ static bool virtblk_add_req_batch(struct virtio_blk_vq *vq,
                if (err) {
                        virtblk_unmap_data(req, vbr);
                        virtblk_cleanup_cmd(req);
-                       rq_list_add(requeue_list, req);
+                       blk_mq_requeue_request(req, true);
                }
        }
 
@@ -436,7 +435,7 @@ static void virtio_queue_rqs(struct request **rqlist)
 
                if (!next || req->mq_hctx != next->mq_hctx) {
                        req->rq_next = NULL;
-                       kick = virtblk_add_req_batch(vq, rqlist, &requeue_list);
+                       kick = virtblk_add_req_batch(vq, rqlist);
                        if (kick)
                                virtqueue_notify(vq->vq);