Re: [PATCH] io_uring/zcrx: reorder fd allocation and disclosure in zcrx_export()

[Jens Axboe <axboe@kernel.dk>]

On 4/5/26 5:53 PM, Bertie Tryner wrote:
> Currently, zcrx_export() allocates and discloses a file descriptor to
> userspace before the backing file is successfully created. If file
> creation fails, the fd is released back to the pool, but the number
> has already been written to the user-provided control structure.
> 
> While this requires a misbehaving or racing userspace to trigger,
> it is better practice to ensure the file descriptor is only
> disclosed once the operation is guaranteed to succeed. This aligns
> the ZCRX export logic with the standard patterns used in the VFS
> layer and other fd-publishing paths.

Like I explained earlier, there's no "race" here at all. The file is
never visible until fd_install() has been done. Any attempt to use the
fd before that happens will get a NULL file in the kernel, and the IO
operation failed.

The operation clearly fails, and the error is returned to the
application. If the application is so buggy that it ignores that and
wants to use the 'fd' value, then it's just buggy. Simple as that, do
stupid things and win stupid prizes.

As a cleanup, this is fine. But the commit message is horribly
(deliberately?) misleading and that should get fixed. I'll let Pavel
decide what to do with this change.

-- 
Jens Axboe