blk-mq: fix discard merge with scheduler attached
authorJens Axboe <axboe@kernel.dk>
Thu, 1 Feb 2018 21:01:02 +0000 (14:01 -0700)
committerJens Axboe <axboe@kernel.dk>
Thu, 1 Feb 2018 21:01:02 +0000 (14:01 -0700)
I ran into an issue on my laptop that triggered a bug on the
discard path:

WARNING: CPU: 2 PID: 207 at drivers/nvme/host/core.c:527 nvme_setup_cmd+0x3d3/0x430
 Modules linked in: rfcomm fuse ctr ccm bnep arc4 binfmt_misc snd_hda_codec_hdmi nls_iso8859_1 nls_cp437 vfat snd_hda_codec_conexant fat snd_hda_codec_generic iwlmvm snd_hda_intel snd_hda_codec snd_hwdep mac80211 snd_hda_core snd_pcm snd_seq_midi snd_seq_midi_event snd_rawmidi snd_seq x86_pkg_temp_thermal intel_powerclamp kvm_intel uvcvideo iwlwifi btusb snd_seq_device videobuf2_vmalloc btintel videobuf2_memops kvm snd_timer videobuf2_v4l2 bluetooth irqbypass videobuf2_core aesni_intel aes_x86_64 crypto_simd cryptd snd glue_helper videodev cfg80211 ecdh_generic soundcore hid_generic usbhid hid i915 psmouse e1000e ptp pps_core xhci_pci xhci_hcd intel_gtt
 CPU: 2 PID: 207 Comm: jbd2/nvme0n1p7- Tainted: G     U           4.15.0+ #176
 Hardware name: LENOVO 20FBCTO1WW/20FBCTO1WW, BIOS N1FET59W (1.33 ) 12/19/2017
 RIP: 0010:nvme_setup_cmd+0x3d3/0x430
 RSP: 0018:ffff880423e9f838 EFLAGS: 00010217
 RAX: 0000000000000000 RBX: ffff880423e9f8c8 RCX: 0000000000010000
 RDX: ffff88022b200010 RSI: 0000000000000002 RDI: 00000000327f0000
 RBP: ffff880421251400 R08: ffff88022b200000 R09: 0000000000000009
 R10: 0000000000000000 R11: 0000000000000000 R12: 000000000000ffff
 R13: ffff88042341e280 R14: 000000000000ffff R15: ffff880421251440
 FS:  0000000000000000(0000) GS:ffff880441500000(0000) knlGS:0000000000000000
 CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
 CR2: 000055b684795030 CR3: 0000000002e09006 CR4: 00000000001606e0
 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
 Call Trace:
  nvme_queue_rq+0x40/0xa00
  ? __sbitmap_queue_get+0x24/0x90
  ? blk_mq_get_tag+0xa3/0x250
  ? wait_woken+0x80/0x80
  ? blk_mq_get_driver_tag+0x97/0xf0
  blk_mq_dispatch_rq_list+0x7b/0x4a0
  ? deadline_remove_request+0x49/0xb0
  blk_mq_do_dispatch_sched+0x4f/0xc0
  blk_mq_sched_dispatch_requests+0x106/0x170
  __blk_mq_run_hw_queue+0x53/0xa0
  __blk_mq_delay_run_hw_queue+0x83/0xa0
  blk_mq_run_hw_queue+0x6c/0xd0
  blk_mq_sched_insert_request+0x96/0x140
  __blk_mq_try_issue_directly+0x3d/0x190
  blk_mq_try_issue_directly+0x30/0x70
  blk_mq_make_request+0x1a4/0x6a0
  generic_make_request+0xfd/0x2f0
  ? submit_bio+0x5c/0x110
  submit_bio+0x5c/0x110
  ? __blkdev_issue_discard+0x152/0x200
  submit_bio_wait+0x43/0x60
  ext4_process_freed_data+0x1cd/0x440
  ? account_page_dirtied+0xe2/0x1a0
  ext4_journal_commit_callback+0x4a/0xc0
  jbd2_journal_commit_transaction+0x17e2/0x19e0
  ? kjournald2+0xb0/0x250
  kjournald2+0xb0/0x250
  ? wait_woken+0x80/0x80
  ? commit_timeout+0x10/0x10
  kthread+0x111/0x130
  ? kthread_create_worker_on_cpu+0x50/0x50
  ? do_group_exit+0x3a/0xa0
  ret_from_fork+0x1f/0x30
 Code: 73 89 c1 83 ce 10 c1 e1 10 09 ca 83 f8 04 0f 87 0f ff ff ff 8b 4d 20 48 8b 7d 00 c1 e9 09 48 01 8c c7 00 08 00 00 e9 f8 fe ff ff <0f> ff 4c 89 c7 41 bc 0a 00 00 00 e8 0d 78 d6 ff e9 a1 fc ff ff
 ---[ end trace 50d361cc444506c8 ]---
 print_req_error: I/O error, dev nvme0n1, sector 847167488

Decoding the assembly, the request claims to have 0xffff segments,
while nvme counts two. This turns out to be because we don't check
for a data carrying request on the mq scheduler path, and since
blk_phys_contig_segment() returns true for a non-data request,
we decrement the initial segment count of 0 and end up with
0xffff in the unsigned short.

There are a few issues here:

1) We should initialize the segment count for a discard to 1.
2) The discard merging is currently using the data limits for
   segments and sectors.

Fix this up by having attempt_merge() correctly identify the
request, and by initializing the segment count correctly
for discards.

This can only be triggered with mq-deadline on discard capable
devices right now, which isn't a common configuration.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
block/blk-core.c
block/blk-merge.c

index 134fd34..d0d1042 100644 (file)
@@ -3283,6 +3283,8 @@ void blk_rq_bio_prep(struct request_queue *q, struct request *rq,
 {
        if (bio_has_data(bio))
                rq->nr_phys_segments = bio_phys_segments(q, bio);
+       else if (bio_op(bio) == REQ_OP_DISCARD)
+               rq->nr_phys_segments = 1;
 
        rq->__data_len = bio->bi_iter.bi_size;
        rq->bio = rq->biotail = bio;
index 8452fc7..782940c 100644 (file)
@@ -550,6 +550,24 @@ static bool req_no_special_merge(struct request *req)
        return !q->mq_ops && req->special;
 }
 
+static bool req_attempt_discard_merge(struct request_queue *q, struct request *req,
+               struct request *next)
+{
+       unsigned short segments = blk_rq_nr_discard_segments(req);
+
+       if (segments >= queue_max_discard_segments(q))
+               goto no_merge;
+       if (blk_rq_sectors(req) + bio_sectors(next->bio) >
+           blk_rq_get_max_sectors(req, blk_rq_pos(req)))
+               goto no_merge;
+
+       req->nr_phys_segments = segments + blk_rq_nr_discard_segments(next);
+       return true;
+no_merge:
+       req_set_nomerge(q, req);
+       return false;
+}
+
 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
                                struct request *next)
 {
@@ -683,9 +701,13 @@ static struct request *attempt_merge(struct request_queue *q,
         * If we are allowed to merge, then append bio list
         * from next to rq and release next. merge_requests_fn
         * will have updated segment counts, update sector
-        * counts here.
+        * counts here. Handle DISCARDs separately, as they
+        * have separate settings.
         */
-       if (!ll_merge_requests_fn(q, req, next))
+       if (req_op(req) == REQ_OP_DISCARD) {
+               if (!req_attempt_discard_merge(q, req, next))
+                       return NULL;
+       } else if (!ll_merge_requests_fn(q, req, next))
                return NULL;
 
        /*
@@ -715,7 +737,8 @@ static struct request *attempt_merge(struct request_queue *q,
 
        req->__data_len += blk_rq_bytes(next);
 
-       elv_merge_requests(q, req, next);
+       if (req_op(req) != REQ_OP_DISCARD)
+               elv_merge_requests(q, req, next);
 
        /*
         * 'next' is going away, so update stats accordingly