Re: PROBLEM: double fault in md_end_io

[Guoqing Jiang <jgq516@gmail.com>]

On 2021/4/10 上午5:40, Paweł Wiejacha wrote:
> Hello,
>
> Two of my machines constantly crash with a double fault like this:
>
> 1146  <0>[33685.629591] traps: PANIC: double fault, error_code: 0x0
> 1147  <4>[33685.629593] double fault: 0000 [#1] SMP NOPTI
> 1148  <4>[33685.629594] CPU: 10 PID: 2118287 Comm: kworker/10:0
> Tainted: P           OE     5.11.8-051108-generic #202103200636
> 1149  <4>[33685.629595] Hardware name: ASUSTeK COMPUTER INC. KRPG-U8
> Series/KRPG-U8 Series, BIOS 4201 09/25/2020
> 1150  <4>[33685.629595] Workqueue: xfs-conv/md12 xfs_end_io [xfs]
> 1151  <4>[33685.629596] RIP: 0010:__slab_free+0x23/0x340
> 1152  <4>[33685.629597] Code: 4c fe ff ff 0f 1f 00 0f 1f 44 00 00 55
> 48 89 e5 41 57 49 89 cf 41 56 49 89 fe 41 55 41 54 49 89 f4 53 48 83
> e4 f0 48 83 ec 70 <48> 89 54 24 28 0f 1f 44 00 00 41 8b 46 28 4d 8b 6c
> 24 20 49 8b 5c
> 1153  <4>[33685.629598] RSP: 0018:ffffa9bc00848fa0 EFLAGS: 00010086
> 1154  <4>[33685.629599] RAX: ffff94c04d8b10a0 RBX: ffff94437a34a880
> RCX: ffff94437a34a880
> 1155  <4>[33685.629599] RDX: ffff94437a34a880 RSI: ffffcec745e8d280
> RDI: ffff944300043b00
> 1156  <4>[33685.629599] RBP: ffffa9bc00849040 R08: 0000000000000001
> R09: ffffffff82a5d6de
> 1157  <4>[33685.629600] R10: 0000000000000001 R11: 000000009c109000
> R12: ffffcec745e8d280
> 1158  <4>[33685.629600] R13: ffff944300043b00 R14: ffff944300043b00
> R15: ffff94437a34a880
> 1159  <4>[33685.629601] FS:  0000000000000000(0000)
> GS:ffff94c04d880000(0000) knlGS:0000000000000000
> 1160  <4>[33685.629601] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> 1161  <4>[33685.629602] CR2: ffffa9bc00848f98 CR3: 000000014d04e000
> CR4: 0000000000350ee0
> 1162  <4>[33685.629602] Call Trace:
> 1163  <4>[33685.629603]  <IRQ>
> 1164  <4>[33685.629603]  ? kfree+0x3bc/0x3e0
> 1165  <4>[33685.629603]  ? mempool_kfree+0xe/0x10
> 1166  <4>[33685.629603]  ? mempool_kfree+0xe/0x10
> 1167  <4>[33685.629604]  ? mempool_free+0x2f/0x80
> 1168  <4>[33685.629604]  ? md_end_io+0x4a/0x70
> 1169  <4>[33685.629604]  ? bio_endio+0xdc/0x130
> 1170  <4>[33685.629605]  ? bio_chain_endio+0x2d/0x40
> 1171  <4>[33685.629605]  ? md_end_io+0x5c/0x70
> 1172  <4>[33685.629605]  ? bio_endio+0xdc/0x130
> 1173  <4>[33685.629605]  ? bio_chain_endio+0x2d/0x40
> 1174  <4>[33685.629606]  ? md_end_io+0x5c/0x70
> 1175  <4>[33685.629606]  ? bio_endio+0xdc/0x130
> 1176  <4>[33685.629606]  ? bio_chain_endio+0x2d/0x40
> 1177  <4>[33685.629607]  ? md_end_io+0x5c/0x70
> ... repeated ...
> 1436  <4>[33685.629677]  ? bio_endio+0xdc/0x130
> 1437  <4>[33685.629677]  ? bio_chain_endio+0x2d/0x40
> 1438  <4>[33685.629677]  ? md_end_io+0x5c/0x70
> 1439  <4>[33685.629677]  ? bio_endio+0xdc/0x130
> 1440  <4>[33685.629678]  ? bio_chain_endio+0x2d/0x40
> 1441  <4>[33685.629678]  ? md_
> 1442  <4>[33685.629679] Lost 357 message(s)!
>
> This happens on:
> 5.11.8-051108-generic #202103200636 SMP Sat Mar 20 11:17:32 UTC 2021
> and on 5.8.0-44-generic #50~20.04.1-Ubuntu
> (https://changelogs.ubuntu.com/changelogs/pool/main/l/linux/linux_5.8.0-44.50/changelog)
> which contains backported
> https://github.com/torvalds/linux/commit/41d2d848e5c09209bdb57ff9c0ca34075e22783d
> ("md: improve io stats accounting").
> The 5.8.18-050818-generic #202011011237 SMP Sun Nov 1 12:40:15 UTC
> 2020 which does not contain above suspected change does not crash.
>
> If there's a better way/place to report this bug just let me know. If
> not, here are steps to reproduce:
>
> 1. Create a RAID 0 device using three Micron_9300_MTFDHAL7T6TDP disks.
> mdadm --create --verbose /dev/md12 --level=stripe --raid-devices=3
> /dev/nvme0n1p1 /dev/nvme1n1p1 /dev/nvme2n1p1
>
> 2. Setup xfs on it:
> mkfs.xfs /dev/md12 and mount it
>
> 3. Write to a file on this filesystem:
> while true; do rm -rf /mnt/md12/crash* ; for i in `seq 8`; do dd
> if=/dev/zero of=/mnt/md12/crash$i bs=32K count=50000000 & done; wait;
> done
> Wait for a crash (usually less than 20 min).
>
> I couldn't reproduce it with a single dd process (maybe I have to wait
> a little longer), but a single cat
> /very/large/file/on/cephfs/over100GbE > /mnt/md12/crash is enough for
> this double fault to occur.

I guess it is related with bio split, if raid0_make_request calls 
bio_chain for the split bio, then
it's bi_end_io is changed to bio_chain_endio. Could you try this?

--- a/drivers/md/md.c
+++ b/drivers/md/md.c
@@ -489,7 +489,7 @@ static blk_qc_t md_submit_bio(struct bio *bio)
                 return BLK_QC_T_NONE;
         }

-       if (bio->bi_end_io != md_end_io) {
+       if (bio->bi_end_io != md_end_io && bio->bi_end_io != 
bio_chain_endio) {

If the above works, then we could miss the statistics of split bio from 
blk_queue_split which is
called before hijack bi_end_io, so we may need to move blk_queue_split 
after hijack bi_end_io.

Thanks,
Guoqing