Re: [PATCH v3 0/3] Support userspace-selected fds

[Jens Axboe <axboe@kernel.dk>]

On 4/3/20 10:57 PM, Josh Triplett wrote:
> (Note: numbering this updated version v3, to avoid confusion with Jens'
> v2 that built on my v1. Jens, if you like this approach, please feel
> free to stack your additional patches from the io_uring-fd-select branch
> atop this series. 5.8 material, not intended for the current merge window.)
> 
> This new version has an API to atomically increase the minimum fd and
> return the previous minimum, rather than just getting and setting the
> minimum; this makes it easier to allocate a range. (A library that might
> initialize after the program has already opened other file descriptors may need
> to check for existing open fds in the range after reserving it, and
> reserve more fds if needed; this can be done entirely in userspace, and
> we can't really do anything simpler in the kernel due to limitations on
> file-descriptor semantics, so this patch series avoids introducing any
> extra complexity in the kernel.)
> 
> This new version also supports a __get_specific_unused_fd_flags call
> which accepts the limit for RLIMIT_NOFILE as an argument, analogous to
> __get_unused_fd_flags, since io_uring needs that to correctly handle
> RLIMIT_NOFILE.
> 
> 
> Inspired by the X protocol's handling of XIDs, allow userspace to select
> the file descriptor opened by a call like openat2, so that it can use
> the resulting file descriptor in subsequent system calls without waiting
> for the response the initial openat2 syscall.
> 
> The first patch is independent of the other two; it allows reserving
> file descriptors below a certain minimum for userspace-selected fd
> allocation only.
> 
> The second patch implements userspace-selected fd allocation for
> openat2, introducing a new O_SPECIFIC_FD flag and an fd field in struct
> open_how. In io_uring, this allows sequences like openat2/read/close
> without waiting for the openat2 to complete. Multiple such sequences can
> overlap, as long as each uses a distinct file descriptor.
> 
> The third patch adds userspace-selected fd allocation to pipe2 as well.
> I did this partly as a demonstration of how simple it is to wire up
> O_SPECIFIC_FD support for a new fd-allocating system call, and partly in
> the hopes that this may make it more useful to wire up io_uring support
> for pipe2 in the future.

This looks pretty clean and flexible to me, and it'll work fine for
io_uring.

Care to post this a bit wider (and CC Al)?

-- 
Jens Axboe