Re: [BUG? liburing] io_uring_register_files_update with liburing 2.0 on 5.13.17

[Jens Axboe <axboe@kernel.dk>]

On Sat, Sep 18, 2021 at 2:26 PM Jens Axboe <axboe@kernel.dk> wrote:
>
> On 9/18/21 2:13 PM, Victor Stewart wrote:
> > On Sat, Sep 18, 2021 at 3:41 PM Jens Axboe <axboe@kernel.dk> wrote:
> >>
> >> On 9/18/21 7:41 AM, Victor Stewart wrote:
> >>> just auto updated from 5.13.16 to 5.13.17, and suddenly my fixed
> >>> file registrations fail with EOPNOTSUPP using liburing 2.0.
> >>>
> >>> static inline struct io_uring ring;
> >>> static inline int *socketfds;
> >>>
> >>> // ...
> >>>
> >>> void enableFD(int fd)
> >>> {
> >>>    int result = io_uring_register_files_update(&ring, fd,
> >>>                       &(socketfds[fd] = fd), 1);
> >>>    printf("enableFD, result = %d\n", result);
> >>> }
> >>>
> >>> maybe this is due to the below and related work that
> >>> occurred at the end of 5.13 and liburing got out of sync?
> >>>
> >>> https://github.com/torvalds/linux/commit/992da01aa932b432ef8dc3885fa76415b5dbe43f#diff-79ffab63f24ef28eec3badbc8769e2a23e0475ab1fbe390207269ece944a0824
> >>>
> >>> and can't use liburing 2.1 because of the api changes since 5.13.
> >>
> >> That's very strange, the -EOPNOTSUPP should only be possible if you
> >> are not passing in the ring fd for the register syscall. You should
> >> be able to mix and match liburing versions just fine, the only exception
> >> is sometimes between releases (of both liburing and the kernel) where we
> >> have the liberty to change the API of something that was added before
> >> release.
> >>
> >> Can you do an strace of it and attach?
> >
> > oh ya the EOPNOTSUPP was my bug introduced trying to debug.
> >
> > here's the real bug...
> >
> > io_uring_register(13, IORING_REGISTER_FILES, [-1, -1, -1, 3, 4, 5, 6, 7,
> > 8, 9, 10, 11, 12, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1,
> > -1, -1, -1, -1, -1,
> > -1, ...], 32768) = -1 EMFILE (Too many open files)
> >
> > 32,768 is 1U << 15 aka IORING_MAX_FIXED_FILES, but i tried
> > 16,000 just to try and same issue.
> >
> > maybe you're not allowed to have pre-filled (aka non negative 1)
> > entries upon the initial io_uring_register_files call anymore?
> >
> > this was working until the 5.13.16 -> 5.13.17 transition.
>
> Ah yes that makes more sense. You need to up RLIMIT_NOFILE, the
> registered files are under that protection now too. This is also why it
> was brought back to stable. A bit annoying, but it was needed for the
> direct file support to have some sanity there.
>
> So use rlimit(RLIMIT_NOFILE,...) from the app or ulimit -n to bump the
> limit.

BTW, this could be incorporated into io_uring_register_files and
io_uring_register_files_tags(), might not be a bad idea in general. Just
have it check rlim.rlim_cur for RLIMIT_NOFILE, and if it's smaller than
'nr_files', then bump it. That'd hide it nicely, instead of throwing a
failure.

-- 
Jens Axboe