Re: [PATCHSET 0/4] Allow relative lookups

[Jens Axboe <axboe@kernel.dk>]

On 2/7/20 3:56 PM, Stefan Metzmacher wrote:
> Hi Jens,
> 
>>> Am 07.02.20 um 16:50 schrieb Jens Axboe:
>>>> Due to an oversight on my part, AT_FDCWD lookups only work when the
>>>> lookup can be done inline, not async. This patchset rectifies that,
>>>> aiming for 5.6 for this one as it would be a shame to have openat etc
>>>> without that.
>>>>
>>>> Just 3 small simple patches - grab the task ->fs, add io-wq suppor for
>>>> passing it in and setting it, and finally add a ->needs_fs to the opcode
>>>> table list of requirements for openat/openat2/statx.
>>>>
>>>> Last patch just ensures we allow AT_FDCWD.
>>>
>>> Thanks! But IOSQE_FIXED_FILE is still not supported and not rejected at
>>> the same time, correct?
>>
>> That's in a separate patch:
>>
>> https://git.kernel.dk/cgit/linux-block/commit/?h=io_uring-5.6&id=5e159663813f0b7837342426cfb68185b6609359
> 
> Do we handle the error path correct?
> As far as I can see io_req_set_file() is called before
> io_{statx,openat,openat2}_prep() and req->file is already filled.
> Maybe a generic way would be better using io_op_defs[op].allow_fixed_file.

Worst case, as far as I can tell, is that we'll think it's a valid
descriptor (because the both the fixed index and fd are valid) and we'll
still error in the prep. Only concern would be that maybe we should make
this an -EBADF return, which would be 100% consistent between them (and
with other cases). I'll make that change.

> In the long run we may want to add support for openat2 with
> IOSQE_FIXED_FILE, then Samba could register the share root directory as
> fixed file and it could be used for all openat2 calls.
> But for now it's fine to just reject it...

It's not impossible to support, it's just that it requires changes
outside of io_uring to do so. So for now it'll just not be possible, I'm
afraid.

-- 
Jens Axboe