regression: fixed file hang
Jens Axboe
maintainerHi Xiaoguang,
Was doing some other testing today, and noticed a hang with fixed files. I did a bit of poor mans bisecting, and came up with this one:
commit 0558955373023b08f638c9ede36741b0e4200f58
Author: Xiaoguang Wang <[email protected]> Date: Tue Mar 31 14:05:18 2020 +0800
io_uring: refactor file register/unregister/update handling
If I revert this one, the test completes fine.
The case case is pretty simple, just run t/io_uring from the fio repo, default settings:
[ fio] # t/io_uring /dev/nvme0n1p2
Added file /dev/nvme0n1p2sq_ring ptr = 0x0x7fe1cb81f000
sqes ptr = 0x0x7fe1cb81d000cq_ring ptr = 0x0x7fe1cb81b000
polled=1, fixedbufs=1, buffered=0 QD=128, sq_ring=128, cq_ring=256 submitter=345
IOPS=240096, IOS/call=32/31, inflight=91 (91) IOPS=249696, IOS/call=32/31, inflight=99 (99)
^CExiting on signal 2
and ctrl-c it after a second or so. You'll then notice a kworker that is stuck in io_sqe_files_unregister(), here:
/* wait for all refs nodes to complete */ wait_for_completion(&data->done);
I'll try and debug this a bit, and for some reason it doens't trigger with the liburing fixed file setup. Just wanted to throw this out there, so if you have cycles, please do take a look at it.
Signature · 1 line
Jens Axboe
Jens Axboe
maintainerQuoted 30 lines · Jens Axboe
Hi Xiaoguang,
Was doing some other testing today, and noticed a hang with fixed files. I did a bit of poor mans bisecting, and came up with this one:
commit
0558955373023b08f638c9ede36741b0e4200f58Author: Xiaoguang Wang <[email protected]> Date: Tue Mar 31 14:05:18 2020 +0800
io_uring: refactor file register/unregister/update handlingIf I revert this one, the test completes fine.
The case case is pretty simple, just run t/io_uring from the fio repo, default settings:
[ fio] # t/io_uring
/dev/nvme0n1p2
Added file/dev/nvme0n1p2sq_ringptr = 0x0x7fe1cb81f000
sqes ptr = 0x0x7fe1cb81d000cq_ringptr = 0x0x7fe1cb81b000
polled=1, fixedbufs=1, buffered=0 QD=128, sq_ring=128, cq_ring=256 submitter=345
IOPS=240096, IOS/call=32/31, inflight=91 (91) IOPS=249696, IOS/call=32/31, inflight=99 (99)
^CExiting on signal 2and ctrl-c it after a second or so. You'll then notice a kworker that is stuck in
io_sqe_files_unregister(), here:
Quoted 6 lines
/* wait for all refs nodes to complete */ wait_for_completion(&data->done);I'll try and debug this a bit, and for some reason it doens't trigger with the liburing fixed file setup. Just wanted to throw this out there, so if you have cycles, please do take a look at it.
https://lore.kernel.org/io-uring/[email protected]/T/#u
Signature · 1 line
Jens Axboe
hi,
Quoted 31 lines · Jens Axboe
Quoted 30 lines
Hi Xiaoguang,
Was doing some other testing today, and noticed a hang with fixed files. I did a bit of poor mans bisecting, and came up with this one:
commit
0558955373023b08f638c9ede36741b0e4200f58Author: Xiaoguang Wang <[email protected]> Date: Tue Mar 31 14:05:18 2020 +0800
io_uring: refactor file register/unregister/update handlingIf I revert this one, the test completes fine.
The case case is pretty simple, just run t/io_uring from the fio repo, default settings:
[ fio] # t/io_uring
/dev/nvme0n1p2
Added file/dev/nvme0n1p2sq_ringptr = 0x0x7fe1cb81f000
sqes ptr = 0x0x7fe1cb81d000cq_ringptr = 0x0x7fe1cb81b000
polled=1, fixedbufs=1, buffered=0 QD=128, sq_ring=128, cq_ring=256 submitter=345
IOPS=240096, IOS/call=32/31, inflight=91 (91) IOPS=249696, IOS/call=32/31, inflight=99 (99)
^CExiting on signal 2and ctrl-c it after a second or so. You'll then notice a kworker that is stuck in
io_sqe_files_unregister(), here:
Quoted 8 lines
Quoted 6 lines
/* wait for all refs nodes to complete */ wait_for_completion(&data->done);I'll try and debug this a bit, and for some reason it doens't trigger with the liburing fixed file setup. Just wanted to throw this out there, so if you have cycles, please do take a look at it.
Thanks for this fix, and sorry, it's my bad, I didn't cover this case when sending patches. Can you share your test cases or test method when developing io_uring? Usually I just run test cases under liburing/test, seems it's not enough.
Regards,
Xiaoguang Wang
Quoted 2 lines
Jens Axboe
maintainerQuoted 40 lines · Xiaoguang Wang
hi,
Quoted 38 lines · Jens Axboe
Quoted 36 lines
Hi Xiaoguang,
Was doing some other testing today, and noticed a hang with fixed files. I did a bit of poor mans bisecting, and came up with this one:
commit
0558955373023b08f638c9ede36741b0e4200f58Author: Xiaoguang Wang <[email protected]> Date: Tue Mar 31 14:05:18 2020 +0800
io_uring: refactor file register/unregister/update handlingIf I revert this one, the test completes fine.
The case case is pretty simple, just run t/io_uring from the fio repo, default settings:
[ fio] # t/io_uring
/dev/nvme0n1p2
Added file/dev/nvme0n1p2sq_ringptr = 0x0x7fe1cb81f000
sqes ptr = 0x0x7fe1cb81d000cq_ringptr = 0x0x7fe1cb81b000
polled=1, fixedbufs=1, buffered=0 QD=128, sq_ring=128, cq_ring=256 submitter=345
IOPS=240096, IOS/call=32/31, inflight=91 (91) IOPS=249696, IOS/call=32/31, inflight=99 (99)
^CExiting on signal 2and ctrl-c it after a second or so. You'll then notice a kworker that is stuck in
io_sqe_files_unregister(), here:/* wait for all refs nodes to complete */ wait_for_completion(&data->done);I'll try and debug this a bit, and for some reason it doens't trigger with the liburing fixed file setup. Just wanted to throw this out there, so if you have cycles, please do take a look at it.
Quoted 5 lines
Thanks for this fix, and sorry, it's my bad, I didn't cover this case when sending patches. Can you share your test cases or test method when developing
io_uring? Usually I just run test cases under liburing/test, seems it's not enough.
It really should be enough, the case that triggered this issue is the combination of fixed files and polled IO. I'll need to add that to eg test/read-write.c, which does a lot of combinations already.
The only issue is that polled IO only works on some files / devices. I've been meaning to add a config file to the liburing regression tests, so you can configure a device to use for testing. With an NVMe device in there, we should be able to have full coverage.
Signature · 1 line
Jens Axboe