Re: Consistently reproducible deadlock with simple io_uring program

[Jens Axboe <axboe@kernel.dk>]

On 8/15/20 7:12 PM, Drew DeVault wrote:
> Kernel 5.7.12-arch1-1 x86_64, Arch Linux.
> 
> I'm working on a new implementation of the userspace end of io_uring for
> a new programming language under design. The program is pretty simple:
> it sets up an io_uring with flags set to zero, obtains an SQE, prepares
> it to read 1024 bytes from stdin, and then calls io_uring_enter with
> submit and min_complete both set to 1.
> 
> The code is set up to grab the CQE and interpret the data in the buffer
> after this, but I'm not sure if this side ever gets run in userspace,
> because my kernel immediately locks up after I press enter on the
> controlling terminal to submit some data to stdin.
> 
> I have uploaded a binary which reproduces the problem here:
> 
> https://yukari.sr.ht/io_uring-crashme
> 
> If you want to reproduce it from source, reach out to me out of band; it
> will be a challenge. This new programming langauge is technically GPL'd,
> but it hasn't been released to the public, YMMV getting it set up on
> localhost, and it doesn't generate DWARF symbols so you're going to be
> debugging assembly no matter what.

This should be fixed by this one:

https://git.kernel.dk/cgit/linux-block/commit/?h=io_uring-5.9&id=d4e7cd36a90e38e0276d6ce0c20f5ccef17ec38c

which will go into 5.7/5.8 stable shortly. Affected files are the ones
that use double waitqueues at the same time for polling, which are
mostly just tty...

-- 
Jens Axboe