Re: io_uring's openat doesn't work with large (2G+) files

[Jens Axboe <axboe@kernel.dk>]

On 4/8/20 8:30 AM, Dmitry Kadashev wrote:
> On Wed, Apr 8, 2020 at 10:19 PM Jens Axboe <axboe@kernel.dk> wrote:
>>
>> On 4/8/20 7:51 AM, Dmitry Kadashev wrote:
>>> Hi,
>>>
>>> io_uring's openat seems to produce FDs that are incompatible with
>>> large files (>2GB). If a file (smaller than 2GB) is opened using
>>> io_uring's openat then writes -- both using io_uring and just sync
>>> pwrite() -- past that threshold fail with EFBIG. If such a file is
>>> opened with sync openat, then both io_uring's writes and sync writes
>>> succeed. And if the file is larger than 2GB then io_uring's openat
>>> fails right away, while the sync one works.
>>>
>>> Kernel versions: 5.6.0-rc2, 5.6.0.
>>>
>>> A couple of reproducers attached, one demos successful open with
>>> failed writes afterwards, and another failing open (in comparison with
>>> sync  calls).
>>>
>>> The output of the former one for example:
>>>
>>> *** sync openat
>>> openat succeeded
>>> sync write at offset 0
>>> write succeeded
>>> sync write at offset 4294967296
>>> write succeeded
>>>
>>> *** sync openat
>>> openat succeeded
>>> io_uring write at offset 0
>>> write succeeded
>>> io_uring write at offset 4294967296
>>> write succeeded
>>>
>>> *** io_uring openat
>>> openat succeeded
>>> sync write at offset 0
>>> write succeeded
>>> sync write at offset 4294967296
>>> write failed: File too large
>>>
>>> *** io_uring openat
>>> openat succeeded
>>> io_uring write at offset 0
>>> write succeeded
>>> io_uring write at offset 4294967296
>>> write failed: File too large
>>
>> Can you try with this one? Seems like only openat2 gets it set,
>> not openat...
> 
> I've tried specifying O_LARGEFILE explicitly, that did not change the
> behavior. Is this good enough? Much faster for me to check this way
> that rebuilding the kernel. But if necessary I can do that.

Not sure O_LARGEFILE settings is going to do it for x86-64, the patch
should fix it though. Might have worked on 32-bit, though.

> Also, forgot to mention, this is on x86_64, not sure if O_LARGEFILE is
> necessary to do 2G+ files there?

Internally, yes.

-- 
Jens Axboe