在 2020/12/1 20:34, Ming Lei 写道:
On Mon, Nov 30, 2020 at 11:22:33AM +0000, John Garry wrote:
Hi all,
Some guys internally upgraded to v5.10-rcX and start to see a hang after dd
+ sync for a large file:
- mount /dev/sda1 (ext4 filesystem) to directory /mnt;
- run "if=/dev/zero of=test1 bs=1M count=2000" on directory /mnt;
- run "sync"
and get:
[ 367.912761] INFO: task jbd2/sdb1-8:3602 blocked for more than 120
seconds.
[ 367.919618] Not tainted 5.10.0-rc1-109488-g32ded76956b6 #948
[ 367.925776] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs"
disables this message.
[ 367.933579] task:jbd2/sdb1-8 state:D stack: 0 pid: 3602
ppid: 2 flags:0x00000028
[ 367.941901] Call trace:
[ 367.944351] __switch_to+0xb8/0x168
[ 367.947840] __schedule+0x30c/0x670
[ 367.951326] schedule+0x70/0x108
[ 367.954550] io_schedule+0x1c/0xe8
[ 367.957948] bit_wait_io+0x18/0x68
[ 367.961346] __wait_on_bit+0x78/0xf0
[ 367.964919] out_of_line_wait_on_bit+0x8c/0xb0
[ 367.969356] __wait_on_buffer+0x30/0x40
[ 367.973188] jbd2_journal_commit_transaction+0x1370/0x1958
[ 367.978661] kjournald2+0xcc/0x260
[ 367.982061] kthread+0x150/0x158
[ 367.985288] ret_from_fork+0x10/0x34
[ 367.988860] INFO: task sync:3823 blocked for more than 120 seconds.
[ 367.995102] Not tainted 5.10.0-rc1-109488-g32ded76956b6 #948
[ 368.001265] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs"
disables this message.
[ 368.009067] task:sync state:D stack: 0 pid: 3823 ppid:
3450 flags:0x00000009
[ 368.017397] Call trace:
[ 368.019841] __switch_to+0xb8/0x168
[ 368.023320] __schedule+0x30c/0x670
[ 368.026804] schedule+0x70/0x108
[ 368.030025] jbd2_log_wait_commit+0xbc/0x158
[ 368.034290] ext4_sync_fs+0x188/0x1c8
[ 368.037947] sync_fs_one_sb+0x30/0x40
[ 368.041606] iterate_supers+0x9c/0x138
[ 368.045350] ksys_sync+0x64/0xc0
[ 368.048569] __arm64_sys_sync+0x10/0x20
[ 368.052398] el0_svc_common.constprop.3+0x68/0x170
[ 368.057177] do_el0_svc+0x24/0x90
[ 368.060482] el0_sync_handler+0x118/0x168
[ 368.064478] el0_sync+0x158/0x180
The issue was reported here originally:
https://lore.kernel.org/linux-ext4/4d18326e-9ca2-d0cb-7cb8-cb56981280da@xxxxxxxxxxxxx/
But it looks like issue related to recent work for SCSI MQ.
They can only create with hisi_sas v3 hw. I could not create with megaraid
sas on the same dev platform or hisi_sas on a similar dev board.
Reverting "scsi: core: Only re-run queue in scsi_end_request() if device
queue is busy" seems solve the issue. Also, checking out to patch prior to
"scsi: hisi_sas: Switch v3 hw to MQ" seems to not have the issue.
If the issue can be reproduced, you may try the following patch:
I tried the change, and the issue is still.
We find that the number of completed IO is less than dispatched, but
from sysfs of block device (such as
/sys/devices/pci0000:74/0000:74:02.0/host0/port-0:0/end_device-0:0/target0:0:0/0:0:0:0/block/sda/sda1/inflight),
the number of inflight is 0.
diff --git a/drivers/scsi/scsi_lib.c b/drivers/scsi/scsi_lib.c
index 60c7a7d74852..f95bd0e5006e 100644
--- a/drivers/scsi/scsi_lib.c
+++ b/drivers/scsi/scsi_lib.c
@@ -602,6 +602,9 @@ static bool scsi_end_request(struct request *req, blk_status_t error,
__blk_mq_end_request(req, error);
+ if (unlikely(req->end_io))
+ smp_mb();
+
scsi_run_queue_async(sdev);
percpu_ref_put(&q->q_usage_counter);
Thanks,
Ming
.
