On 8/15/21 9:42 PM, Olivier Langlois wrote:
[...]

When I have first encountered the issue, the very first thing that I
did try was to create a simple test program that would synthetize the
problem.

After few time consumming failed attempts, I just gave up the idea and
simply settle to my prod program that showcase systematically the
problem every time that I kill the process with a SEGV signal.

In a nutshell, all the program does is to issue read operations with
io_uring on a TCP socket on which there is a constant data stream.

Now that I have a better understanding of what is going on, I think
that one way that could reproduce the problem consistently could be
along those lines:

1. Create a pipe
2. fork a child
3. Initiate a read operation on the pipe with io_uring from the child
4. Let the parent kill its child with a core dump generating signal.
5. Write something in the pipe from the parent so that the io_uring
read operation completes while the core dump is generated.

I guess that I'll end up doing that if I cannot fix the issue with my
current setup but here is what I have attempted so far:

1. Call io_uring_files_cancel from do_coredump
2. Same as #1 but also make sure that TIF_NOTIFY_SIGNAL is cleared on
returning from io_uring_files_cancel

Those attempts didn't work but lurking in the io_uring dev mailing list
is starting to pay off. I thought that I did reach the bottom of the
rabbit hole in my journey of understanding io_uring but the recent
patch set sent by Hao Xu

https://lore.kernel.org/io-uring/90fce498-968e-6812-7b6a-fdf8520ea8d9@kernel.dk/T/#t (local)

made me realize that I still haven't assimilated all the small io_uring
nuances...

Here is my feedback. From my casual io_uring code reader point of view,
it is not 100% obvious what the difference is between
io_uring_files_cancel and io_uring_task_cancel

As you mentioned, io_uring_task_cancel() cancels and waits for all
requests submitted by current task, used in exec() and SQPOLL because
of potential races.

io_uring_task_cancel() cancels only selected ones and


io_uring_files_cancel()
cancels and waits only some specific requests that we absolutely have
to, e.g. in 5.15 it'll be only requests referencing the ring itself.
It's used on normal task exit.

io_uring_task_cancel() cancels and waits all requests submitted by
current task, used on exec() because of races.



As you mentioned 

It seems like io_uring_files_cancel is cancelling polls only if they
have the REQ_F_INFLIGHT flag set.

I have no idea what an inflight request means and why someone would
want to call io_uring_files_cancel over io_uring_task_cancel.

I guess that if I was to meditate on the question for few hours, I
would at some point get some illumination strike me but I believe that
it could be a good idea to document in the code those concepts for
helping casual readers...

Bottomline, I now understand that io_uring_files_cancel does not cancel
all the requests. Therefore, without fully understanding what I am
doing, I am going to replace my call to io_uring_files_cancel from
do_coredump with io_uring_task_cancel and see if this finally fix the
issue for good.

What I am trying to do is to cancel pending io_uring requests to make
sure that TIF_NOTIFY_SIGNAL isn't set while core dump is generated.

Maybe another solution would simply be to modify __dump_emit to make it
resilient to TIF_NOTIFY_SIGNAL as Eric W. Biederman originally
suggested.

or maybe do both...

Not sure which approach is best. If someone has an opinion, I would be
curious to hear it.

Greetings,

-- 
Pavel Begunkov