Re: [PATCH 0/6] Allow signals for IO threads

From: Jens Axboe
Date: Fri Mar 26 2021 - 08:57:22 EST


On 3/26/21 5:48 AM, Stefan Metzmacher wrote:
>
> Am 26.03.21 um 01:39 schrieb Jens Axboe:
>> Hi,
>>
>> As discussed in a previous thread today, the seemingly much saner approach
>> is just to allow signals (including SIGSTOP) for the PF_IO_WORKER IO
>> threads. If we just have the threads call get_signal() for
>> signal_pending(), then everything just falls out naturally with how
>> we receive and handle signals.
>>
>> Patch 1 adds support for checking and calling get_signal() from the
>> regular IO workers, the manager, and the SQPOLL thread. Patch 2 unblocks
>> SIGSTOP from the default IO thread blocked mask, and the rest just revert
>> special cases that were put in place for PF_IO_WORKER threads.
>>
>> With this done, only two special cases remain for PF_IO_WORKER, and they
>> aren't related to signals so not part of this patchset. But both of them
>> can go away as well now that we have "real" threads as IO workers, and
>> then we'll have zero special cases for PF_IO_WORKER.
>>
>> This passes the usual regression testing, my other usual 24h run has been
>> kicked off. But I wanted to send this out early.
>>
>> Thanks to Linus for the suggestion. As with most other good ideas, it's
>> obvious once you hear it. The fact that we end up with _zero_ special
>> cases with this is a clear sign that this is the right way to do it
>> indeed. The fact that this series is 2/3rds revert further drives that
>> point home. Also thanks to Eric for diligent review on the signal side
>> of things for the past changes (and hopefully ditto on this series :-))
>
> Ok, I'm testing a8ff6a3b20bd16d071ef66824ae4428529d114f9 from
> your io_uring-5.12 branch.
>
> And using this patch:
> diff --git a/examples/io_uring-cp.c b/examples/io_uring-cp.c
> index cc7a227a5ec7..6e26a4214015 100644
> --- a/examples/io_uring-cp.c
> +++ b/examples/io_uring-cp.c
> @@ -116,13 +116,16 @@ static void queue_write(struct io_uring *ring, struct io_data *data)
> io_uring_submit(ring);
> }
>
> -static int copy_file(struct io_uring *ring, off_t insize)
> +static int copy_file(struct io_uring *ring, off_t _insize)
> {
> + off_t insize = _insize;
> unsigned long reads, writes;
> struct io_uring_cqe *cqe;
> off_t write_left, offset;
> int ret;
>
> +again:
> + insize = _insize;
> write_left = insize;
> writes = reads = offset = 0;
>
> @@ -221,6 +224,12 @@ static int copy_file(struct io_uring *ring, off_t insize)
> }
> }
>
> + {
> + struct timespec ts = { .tv_nsec = 999999, };
> + nanosleep(&ts, NULL);
> + goto again;
> + }
> +
> return 0;
> }
>
> Running ./io_uring-cp ~/linux-image-5.12.0-rc2+-dbg_5.12.0-rc2+-5_amd64.deb file
> What I see is this:
>
> kill -SIGSTOP to any thread I used a worker with pid 2061 here, results in
>
> root@ub1704-166:~# head /proc/2061/status
> Name: iou-wrk-2041
> Umask: 0022
> State: R (running)
> Tgid: 2041
> Ngid: 0
> Pid: 2061
> PPid: 1857
> TracerPid: 0
> Uid: 0 0 0 0
> Gid: 0 0 0 0
> root@ub1704-166:~# head /proc/2041/status
> Name: io_uring-cp
> Umask: 0022
> State: T (stopped)
> Tgid: 2041
> Ngid: 0
> Pid: 2041
> PPid: 1857
> TracerPid: 0
> Uid: 0 0 0 0
> Gid: 0 0 0 0
> root@ub1704-166:~# head /proc/2042/status
> Name: iou-mgr-2041
> Umask: 0022
> State: T (stopped)
> Tgid: 2041
> Ngid: 0
> Pid: 2042
> PPid: 1857
> TracerPid: 0
> Uid: 0 0 0 0
> Gid: 0 0 0 0
>
> So userspace and iou-mgr-2041 stop, but the workers don't.
> 49 workers burn cpu as much as possible.
>
> kill -KILL 2061
> results in this:
> - all workers are gone
> - iou-mgr-2041 is gone
> - io_uring-cp waits in status D forever
>
> root@ub1704-166:~# head /proc/2041/status
> Name: io_uring-cp
> Umask: 0022
> State: D (disk sleep)
> Tgid: 2041
> Ngid: 0
> Pid: 2041
> PPid: 1857
> TracerPid: 0
> Uid: 0 0 0 0
> Gid: 0 0 0 0
> root@ub1704-166:~# cat /proc/2041/stack
> [<0>] io_wq_destroy_manager+0x36/0xa0
> [<0>] io_wq_put_and_exit+0x2b/0x40
> [<0>] io_uring_clean_tctx+0xc5/0x110
> [<0>] __io_uring_files_cancel+0x336/0x4e0
> [<0>] do_exit+0x16b/0x13b0
> [<0>] do_group_exit+0x8b/0x140
> [<0>] get_signal+0x219/0xc90
> [<0>] arch_do_signal_or_restart+0x1eb/0xeb0
> [<0>] exit_to_user_mode_prepare+0x115/0x1a0
> [<0>] syscall_exit_to_user_mode+0x27/0x50
> [<0>] do_syscall_64+0x45/0x90
> [<0>] entry_SYSCALL_64_after_hwframe+0x44/0xae
>
> The 3rd problem is that gdb in a ubuntu 20.04 userspace vm hangs forever:
>
> root@ub1704-166:~/samba.git# LANG=C strace -o /dev/shm/strace.txt -f -ttT gdb --pid 2417
> GNU gdb (Ubuntu 9.2-0ubuntu1~20.04) 9.2
> Copyright (C) 2020 Free Software Foundation, Inc.
> License GPLv3+: GNU GPL version 3 or later <http://gnu.org/licenses/gpl.html>
> This is free software: you are free to change and redistribute it.
> There is NO WARRANTY, to the extent permitted by law.
> Type "show copying" and "show warranty" for details.
> This GDB was configured as "x86_64-linux-gnu".
> Type "show configuration" for configuration details.
> For bug reporting instructions, please see:
> <http://www.gnu.org/software/gdb/bugs/>.
> Find the GDB manual and other documentation resources online at:
> <http://www.gnu.org/software/gdb/documentation/>.
>
> For help, type "help".
> Type "apropos word" to search for commands related to "word".
> Attaching to process 2417
> [New LWP 2418]
> [New LWP 2419]
>
> <here it hangs forever>
>
> The related parts of 'pstree -a -t -p':
>
> ├─bash,2048
> │ └─io_uring-cp,2417 /root/kernel/sn-devel-184-builds/linux-image-5.12.0-rc2+-dbg_5.12.0-rc2+-5_amd64.deb file
> │ ├─{iou-mgr-2417},2418
> │ └─{iou-wrk-2417},2419
> ├─bash,2167
> │ └─strace,2489 -o /dev/shm/strace.txt -f -ttT gdb --pid 2417
> │ └─gdb,2492 --pid 2417
> │ └─gdb,2494 --pid 2417
>
> root@ub1704-166:~# cat /proc/sys/kernel/yama/ptrace_scope
> 0
>
> root@ub1704-166:~# head /proc/2417/status
> Name: io_uring-cp
> Umask: 0022
> State: t (tracing stop)
> Tgid: 2417
> Ngid: 0
> Pid: 2417
> PPid: 2048
> TracerPid: 2492
> Uid: 0 0 0 0
> Gid: 0 0 0 0
> root@ub1704-166:~# head /proc/2418/status
> Name: iou-mgr-2417
> Umask: 0022
> State: t (tracing stop)
> Tgid: 2417
> Ngid: 0
> Pid: 2418
> PPid: 2048
> TracerPid: 2492
> Uid: 0 0 0 0
> Gid: 0 0 0 0
> root@ub1704-166:~# head /proc/2419/status
> Name: iou-wrk-2417
> Umask: 0022
> State: R (running)
> Tgid: 2417
> Ngid: 0
> Pid: 2419
> PPid: 2048
> TracerPid: 2492
> Uid: 0 0 0 0
> Gid: 0 0 0 0
> root@ub1704-166:~# head /proc/2492/status
> Name: gdb
> Umask: 0022
> State: S (sleeping)
> Tgid: 2492
> Ngid: 0
> Pid: 2492
> PPid: 2489
> TracerPid: 2489
> Uid: 0 0 0 0
> Gid: 0 0 0 0
> root@ub1704-166:~# head /proc/2494/status
> Name: gdb
> Umask: 0022
> State: t (tracing stop)
> Tgid: 2494
> Ngid: 0
> Pid: 2494
> PPid: 2492
> TracerPid: 2489
> Uid: 0 0 0 0
> Gid: 0 0 0 0
>
>
> Maybe these are related and 2494 gets the SIGSTOP that was supposed to
> be handled by 2419.
>
> strace.txt is attached.
>
> Just a wild guess (I don't have time to test this), but maybe this
> will fix it:
>
> diff --git a/fs/io-wq.c b/fs/io-wq.c
> index 07e7d61524c7..ee5a402450db 100644
> --- a/fs/io-wq.c
> +++ b/fs/io-wq.c
> @@ -503,8 +503,7 @@ static int io_wqe_worker(void *data)
> if (io_flush_signals())
> continue;
> ret = schedule_timeout(WORKER_IDLE_TIMEOUT);
> - if (try_to_freeze() || ret)
> - continue;
> + try_to_freeze();
> if (signal_pending(current)) {
> struct ksignal ksig;
>
> @@ -514,8 +513,7 @@ static int io_wqe_worker(void *data)
> continue;
> }
> /* timed out, exit unless we're the fixed worker */
> - if (test_bit(IO_WQ_BIT_EXIT, &wq->state) ||
> - !(worker->flags & IO_WORKER_F_FIXED))
> + if (ret == 0 && !(worker->flags & IO_WORKER_F_FIXED))
> break;
> }
>
> When the worker got a signal I guess ret is not 0 and we'll
> never hit the if (signal_pending()) statement...

Right, the logic was a bit wrong there, and we can also just drop
try_to_freeze() from all of them now, we don't have to special
case that anymore.

Can you try the current branch? I folded in fixes for that.
That will definitely fix case 1+3, the #2 with kill -KILL is kind
of puzzling. I'll try and reproduce that with the current tree and see
what happens. But that feels like it's either not a new thing, or it's
the same core issue as 1+3 (though I don't quite see how, unless the
failure to catch the signal will elude get_signal() forever in the
worker, I guess that's possible).

--
Jens Axboe