Re: [PATCHSET 0/2] Add support for IORING_RECVSEND_POLL_FIRST

[Jens Axboe <axboe@kernel.dk>]

On 4/29/22 12:31 PM, Hao Xu wrote:
> On 4/27/22 09:54, Jens Axboe wrote:
>> Hi,
>>
>> I had a re-think on the flags2 addition [1] that was posted earlier
>> today, and I don't really like the fact that flags2 then can't work
>> with ioprio for read/write etc. We might also want to extend the
>> ioprio field for other types of IO in the future.
>>
>> So rather than do that, do a simpler approach and just add an io_uring
>> specific flag set for send/recv and friends. This then allow setting
>> IORING_RECVSEND_POLL_FIRST in sqe->addr2 for those, and if set, io_uring
>> will arm poll first rather than attempt a send/recv operation.
>>
>> [1] https://lore.kernel.org/io-uring/20220426183343.150273-1-axboe@kernel.dk/
>>
> 
> Hi Jens,
> Could we use something like the high bits of sqe->fd to store general
> flags2 since I saw the number of open FDs can be about (1<<20) at
> most. Though I'm not sure if we can assume the limitation of fd won't
> change in the future..

I think that's a bit iffy, it's pretty universally true (at least on
*NIX) that and fd is a signed int. So I'd be hesitant to do that. I
don't mind the flags here, generally we have potentially 3 per request:

- SQE specific ones, these are the IOSQE_* flags and are meant to be
  generally applicable to all/most commands. io_uring internal, have no
  meaning outside of io_uring.

- Some requests are layered on top of existing functionality, the
  recv/recvmsg is a good example. They have their own sets of flags. We
  generally use the sqe->rw_flags space for those.

- Lastly, io_uring modifiers to specific requests. That's what this
  patch adds. They don't make sense to the lower layers, but they are
  specific to this request type for io_uring.

The 3rd type is put in sqe->rw_flags for io_uring specific opcodes, but
for commands that already have flags in the 2nd category, we have to put
them somewhere else. Not a big deal imho, at least as long as the
request type has space in the sqe for it. They generally do, they did in
this case.

-- 
Jens Axboe