Skip to contentSkip to conversation

io_uring fixes for 5.12-rc3

GIT PULL Merged Jens Axboe 3 participants 9 messages Mar 12, 2021 io_uring

lore
Share & tools
lore

Jens Axboe

author maintainer
#0 lore raw

Hi Linus,

Not quite as small this week as I had hoped, but at least this should be the end of it. All the little known issues have been ironed out - most of it little stuff, but cancelations being the bigger part. Only minor tweaks and/or regular fixes expected beyond this point.

- Fix the creds tracking for async (io-wq and SQPOLL)

- Various SQPOLL fixes related to parking, sharing, forking, IOPOLL, completions, and life times. Much simpler now.

- Make IO threads unfreezable by default, on account of a bug report that had them spinning on resume. Honestly not quite sure why thawing leaves us with a perpetual signal pending (causing the spin), but for now make them unfreezable like there were in 5.11 and prior.

- Move personality_idr to xarray, solving a use-after-free related to removing an entry from the iterator callback. Buffer idr needs the same treatment.

- Re-org around and task vs context tracking, enabling the fixing of cancelations, and then cancelation fixes on top.

- Various little bits of cleanups and hardening, and removal of now dead parts.

Please pull!

The following changes since commit a38fd8748464831584a19438cbb3082b5a2dab15:

Linux 5.12-rc2 (2021-03-05 17:33:41 -0800)

are available in the Git repository at:

git://git.kernel.dk/linux-block.git tags/io_uring-5.12-2021-03-12

for you to fetch changes up to 58f99373834151e1ca7edc49bc5578d9d40db099:

io_uring: fix OP_ASYNC_CANCEL across tasks (2021-03-12 09:42:56 -0700)

----------------------------------------------------------------
io_uring-5.12-2021-03-12

----------------------------------------------------------------
Jens Axboe (9):
io-wq: fix race in freeing 'wq' and worker access
io-wq: always track creds for async issue
io_uring: SQPOLL parking fixes
io-wq: remove unused 'user' member of io_wq
io_uring: move all io_kiocb init early in io_init_req()
io_uring: always wait for sqd exited when stopping SQPOLL thread
kernel: make IO threads unfreezable by default
io_uring: force creation of separate context for ATTACH_WQ and non-threads
io_uring: perform IOPOLL reaping if canceler is thread itself

Matthew Wilcox (Oracle) (1):
io_uring: Convert personality_idr to XArray

Pavel Begunkov (20):
io_uring: make del_task_file more forgiving
io_uring: introduce ctx to tctx back map
io_uring: do ctx initiated file note removal
io_uring: don't take task ring-file notes
io_uring: index io_uring->xa by ctx not file
io_uring: warn when ring exit takes too long
io_uring: cancel reqs of all iowq's on ring exit
io-wq: warn on creating manager while exiting
io_uring: fix unrelated ctx reqs cancellation
io_uring: clean R_DISABLED startup mess
io_uring: fix io_sq_offload_create error handling
io_uring: add io_disarm_next() helper
io_uring: fix complete_post races for linked req
io_uring: fix invalid ctx->sq_thread_idle
io_uring: remove indirect ctx into sqo injection
io_uring: cancel deferred requests in try_cancel
io_uring: remove useless ->startup completion
io_uring: prevent racy sqd->thread checks
io_uring: cancel sqpoll via task_work
io_uring: fix OP_ASYNC_CANCEL across tasks

Stefan Metzmacher (2):
io_uring: run __io_sq_thread() with the initial creds from io_uring_setup()
io_uring: kill io_sq_thread_fork() and return -EOWNERDEAD if the sq_thread is gone

Yang Li (1):
io_uring: remove unneeded variable 'ret'

yangerkun (1):
io-wq: fix ref leak for req in case of exit cancelations

 fs/io-wq.c               |  25 +-
 fs/io-wq.h               |   2 +-
 fs/io_uring.c            | 837 ++++++++++++++++++++++++++---------------------
 include/linux/io_uring.h |   2 +-
 kernel/fork.c            |   1 +
 5 files changed, 470 insertions(+), 397 deletions(-)
Signature · 1 line
Jens Axboe
#1 lore raw
Quoted 5 lines · Jens Axboe

- Make IO threads unfreezable by default, on account of a bug report that had them spinning on resume. Honestly not quite sure why thawing leaves us with a perpetual signal pending (causing the spin), but for now make them unfreezable like there were in 5.11 and prior.

That "not quite sure" doesn't exactly give my the warm and fuzzies, but not being new is ok, I guess. But I'd really like you to try to figure out what is actually going on.

I'm _guessing_ it's just that now those threads are user threads, and then the freezing logic expects them to freeze/thaw using a signal machinery or something like that. And that doesn't work when there is no signal handling for those threads.

But it would be good to _know_.

Linus

#2 lore raw
Quoted 5 lines · Linus Torvalds

I'm _guessing_ it's just that now those threads are user threads, and then the freezing logic expects them to freeze/thaw using a signal machinery or something like that. And that doesn't work when there is no signal handling for those threads.

IOW, I think it's this logic in freeze_task():

        if (!(p->flags & PF_KTHREAD))
                fake_signal_wake_up(p);
        else
                wake_up_state(p, TASK_INTERRUPTIBLE);

where that "not a PF_KTHREAD" test will trigger for the io_uring threads, and it does that fake_signal_wake_up(), and then signal_wake_up() does that

        set_tsk_thread_flag(t, TIF_SIGPENDING);

but the io_uring thread has no way to react to it.

So now the io_uring thread will see "I have pending signals" (even if there is no actual pending signal - it's just a way to get normal processes to get out of TASK_INTERRUPTIBLE and return to user space handling).

And the pending fake signal will never go away, because there is no "return to user space" handling.

So I think the fix is to simply make freeze_task() not do that fake signal thing for IO-uring threads either.

Linus

#3 lore raw
Quoted 3 lines · Linus Torvalds

So I think the fix is to simply make freeze_task() not do that fake signal thing for IO-uring threads either.

Note: I've pulled your tree, I just want you to dig deeper, verify, and check this for perhaps re-instating freezability.

Linus

#4 lore raw
Quoted 3 lines · Linus Torvalds

Note: I've pulled your tree, I just want you to dig deeper, verify, and check this for perhaps re-instating freezability.

Also note that I don't think it's _just_ that freeze_task() thing that is needed for the fix.

I think the io_uring threads then need to do the whole "try_to_freeze()" dance in the main loop or something, so that it does actually freeze.

Linus

author maintainer
#5 lore raw
Quoted 6 lines · Linus Torvalds
Quoted 3 lines · Linus Torvalds

Note: I've pulled your tree, I just want you to dig deeper, verify, and check this for perhaps re-instating freezability.

Quoted 6 lines

Also note that I don't think it's _just_ that freeze_task() thing that is needed for the fix.

I think the io_uring threads then need to do the whole "try_to_freeze()" dance in the main loop or something, so that it does actually freeze.

Right, and they did before the change. The freezing did work, the thaw just left us with that sigpending and caused the never-scheduler busy behavior. But I'll unify it all when the thawing is sorted (which I think it is), and then enable it for all three types.

Signature · 1 line
Jens Axboe
author maintainer
#6 lore raw
In reply to Linus Torvalds
Quoted 28 lines · Linus Torvalds
Quoted 5 lines · Linus Torvalds

I'm _guessing_ it's just that now those threads are user threads, and then the freezing logic expects them to freeze/thaw using a signal machinery or something like that. And that doesn't work when there is no signal handling for those threads.

IOW, I think it's this logic in freeze_task():

        if (!(p->flags & PF_KTHREAD))
                fake_signal_wake_up(p);
        else
                wake_up_state(p, TASK_INTERRUPTIBLE);

where that "not a PF_KTHREAD" test will trigger for the io_uring threads, and it does that fake_signal_wake_up(), and then signal_wake_up() does that

        set_tsk_thread_flag(t, TIF_SIGPENDING);

but the io_uring thread has no way to react to it.

So now the io_uring thread will see "I have pending signals" (even if there is no actual pending signal - it's just a way to get normal processes to get out of TASK_INTERRUPTIBLE and return to user space handling).

Quoted 5 lines

And the pending fake signal will never go away, because there is no "return to user space" handling.

So I think the fix is to simply make freeze_task() not do that fake signal thing for IO-uring threads either.

Ah good catch, I missed it. Yes I bet this is exactly what it is, and it just needs the one liner doing:

         if (!(p->flags & (PF_IO_WORKER | PF_KTHREAD)))
                 fake_signal_wake_up(p);
         else
                 wake_up_state(p, TASK_INTERRUPTIBLE);

instead. I'll try this out and test it, then we can drop the PF_NOFREEZE business going forward. Thanks!

Signature · 1 line
Jens Axboe
author maintainer
#7 lore raw
In reply to Linus Torvalds
Quoted 10 lines · Linus Torvalds
Quoted 5 lines · Jens Axboe

- Make IO threads unfreezable by default, on account of a bug report that had them spinning on resume. Honestly not quite sure why thawing leaves us with a perpetual signal pending (causing the spin), but for now make them unfreezable like there were in 5.11 and prior.

That "not quite sure" doesn't exactly give my the warm and fuzzies, but not being new is ok, I guess. But I'd really like you to try to figure out what is actually going on.

Oh I agree, which is why I wanted to bring it up as I do feel like we're just working around this issue.

Quoted 6 lines

I'm _guessing_ it's just that now those threads are user threads, and then the freezing logic expects them to freeze/thaw using a signal machinery or something like that. And that doesn't work when there is no signal handling for those threads.

But it would be good to _know_.

That's very much my guess too, since the symptom is that we loop all the time because we think there's a signal pending. I am investigating it and reached out to Rafael to try and figure this out, but at the same time I did not want to have an -rc3 that caused resume issues for users. Hence the "let's just disable freezing for now so that things work, get it fixed for real when we can" kind of solution.

Signature · 1 line
Jens Axboe
merge bot Merged
#8 lore raw
In reply to Jens Axboe
The pull request you sent on Fri, 12 Mar 2021 12:48:29 -0700:

The pull request you sent on Fri, 12 Mar 2021 12:48:29 -0700:

Quoted 1 line

git://git.kernel.dk/linux-block.git tags/io_uring-5.12-2021-03-12

has been merged into torvalds/linux.git:
https://git.kernel.org/torvalds/c/9278be92f22979a026a68206e226722138c9443d

Thank you!

Signature · 2 lines
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/prtracker.html