Hi,
在 2025/09/23 16:12, Shinichiro Kawasaki 写道:
On Sep 18, 2025 / 16:53, Yu Kuai wrote:
From: Yu Kuai <yukuai3@xxxxxxxxxx>
While I'm looking at another deadlock issue for blk-throtl, it's found
during test that lockdep is reporting another issue quite eazy, I'm
adding regerssion test for now, if anyone is interested. Otherwise, I'll
go back for this after I finish the problem at hand later.
BTW, maybe we can support to test for scsi_debug instead of null_blk for
all the throtl tests.
This is a good chance to use blktests' feature to repeat test cases by different
conditions [1]. Test cases can implement set_conditions() hooks for that
purpose.
[1] https://github.com/linux-blktests/blktests/blob/a997ab0ca9b77ddf31b8c632bfe57f7bd75e305e/new#L183
Yu, FYI, I quickly implemented the set_conditions() hooks for the throtl group,
and attach it as a patch. When you have time, please try it out. Each test case
runs twice for null_blk and scsi_debug [2].
- With this patch, the lockdep splat is recreated at throtl/004. It is
observed at throtl/001 also as show in [2].
- throtl/002 fails for scsi_debug. I guess performance difference between
null_blk and scsi_debug is the cause, but not so sure.
- Results are placed in results/nodev_nullb and results/nodev_sdebug.
If you think this approach is good, I will prepare a formal patch series.
I do not look into the patch in detail, however, I think this is
perfect.
[2]
# ./check throtl
throtl/001 (nullb) (basic functionality) [passed]
runtime 4.546s ... 4.753s
throtl/001 (sdebug) (basic functionality) [failed]
runtime 5.386s ... 5.842s
something found in dmesg:
[ 79.444897] [ T1156] run blktests throtl/001 at 2025-09-23 17:05:22
[ 79.684859] [ T1260] sd 9:0:0:0: [sdd] Synchronizing SCSI cache
[ 80.002576] [ T1262] scsi_debug:sdebug_driver_probe: scsi_debug: trim poll_queues to 0. poll_q/nr_hw = (0/1)
[ 80.003950] [ T1262] scsi host9: scsi_debug: version 0191 [20210520]
dev_size_mb=1024, opts=0x0, submit_queues=1, statistics=0
[ 80.009307] [ T1262] scsi 9:0:0:0: Direct-Access Linux scsi_debug 0191 PQ: 0 ANSI: 7
[ 80.011952] [ C2] scsi 9:0:0:0: Power-on or device reset occurred
[ 80.018583] [ T1262] sd 9:0:0:0: Attached scsi generic sg3 type 0
[ 80.020594] [ T103] sd 9:0:0:0: [sdd] 2097152 512-byte logical blocks: (1.07 GB/1.00 GiB)
[ 80.022826] [ T103] sd 9:0:0:0: [sdd] Write Protect is off
...
(See '/home/shin/Blktests/blktests/results/nodev_sdebug/throtl/001.dmesg' for the entire message)
throtl/002 (nullb) (iops limit over IO split) [passed]
runtime 2.586s ... 2.486s
throtl/002 (sdebug) (iops limit over IO split) [failed]
runtime 6.137s ... 6.467s
--- tests/throtl/002.out 2025-09-23 14:05:35.011885439 +0900
+++ /home/shin/Blktests/blktests/results/nodev_sdebug/throtl/002.out.bad 2025-09-23 17:05:37.209794003 +0900
@@ -1,4 +1,4 @@
Running throtl/002
-1
-1
+3
+3
I don't feel this is related to performance, we'll have to check later.
Thanks,
Kuai
Test complete
throtl/003 (nullb) (bps limit over IO split) [passed]
runtime 2.554s ... 2.424s
throtl/003 (sdebug) (bps limit over IO split) [passed]
runtime 3.075s ... 2.924s
throtl/004 (nullb) (delete disk while IO is throttled) [passed]
runtime 1.269s ... 1.071s
throtl/004 (sdebug) (delete disk while IO is throttled) [passed]
runtime 2.142s ... 1.471s
throtl/005 (nullb) (change config with throttled IO) [passed]
runtime 3.345s ... 3.420s
throtl/005 (sdebug) (change config with throttled IO) [passed]
runtime 3.905s ... 3.984s
throtl/006 (nullb) (test if meta IO has higher priority than data IO) [passed]
runtime 13.014s ... 13.086s
throtl/006 (sdebug) (test if meta IO has higher priority than data IO) [passed]
runtime 5.715s ... 5.631s
throtl/007 (nullb) (bps limit with iops limit over io split) [passed]
runtime 4.527s ... 4.476s
throtl/007 (sdebug) (bps limit with iops limit over io split) [passed]
runtime 5.065s ... 5.043s
