Re: [PATCH 1/6] io_uring: expand main struct io_kiocb flags to 64-bits

[Pavel Begunkov <asml.silence@gmail.com>]

On 2/7/24 02:18, Jens Axboe wrote:
> On 2/6/24 5:43 PM, Pavel Begunkov wrote:
>> On 2/6/24 16:22, Jens Axboe wrote:
>>> We're out of space here, and none of the flags are easily reclaimable.
>>> Bump it to 64-bits and re-arrange the struct a bit to avoid gaps.
>>>
>>> Add a specific bitwise type for the request flags, io_request_flags_t.
>>> This will help catch violations of casting this value to a smaller type
>>> on 32-bit archs, like unsigned int.
>>>
>>> No functional changes intended in this patch.
>>>
>>> Signed-off-by: Jens Axboe <axboe@kernel.dk>
>>> ---
...
>>>      typedef void (*io_req_tw_func_t)(struct io_kiocb *req, struct io_tw_state *ts);
>>> @@ -592,15 +595,14 @@ struct io_kiocb {
>>>         * and after selection it points to the buffer ID itself.
>>>         */
>>>        u16                buf_index;
>>> -    unsigned int            flags;
>>>    -    struct io_cqe            cqe;
>>
>> With the current layout the min number of lines we touch per
>> request is 2 (including the op specific 64B), that's includes
>> setting up cqe at init and using it for completing. Moving cqe
>> down makes it 3.
>>
>>> +    atomic_t            refs;
>>
>> We're pulling it refs, which is not touched at all in the hot
>> path. Even if there's a hole I'd argue it's better to leave it
>> at the end.
>>
>>> +
>>> +    io_req_flags_t            flags;
>>>          struct io_ring_ctx        *ctx;
>>>        struct task_struct        *task;
>>>    -    struct io_rsrc_node        *rsrc_node;
>>
>> It's used in hot paths, registered buffers/files, would be
>> unfortunate to move it to the next line.
> 
> Yep I did feel a bit bad about that one... Let me take another stab at
> it.
> 
>>> -
>>>        union {
>>>            /* store used ubuf, so we can prevent reloading */
>>>            struct io_mapped_ubuf    *imu;
>>> @@ -615,18 +617,23 @@ struct io_kiocb {
>>>            struct io_buffer_list    *buf_list;
>>>        };
>>>    +    /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
>>> +    struct hlist_node        hash_node;
>>> +
>>
>> And we're pulling hash_node into the hottest line, which is
>> used only when we arm a poll and remove poll. So, it's mostly
>> for networking, sends wouldn't use it much, and multishots
>> wouldn't normally touch it.
>>
>> As for ideas how to find space:
>> 1) iopoll_completed completed can be converted to flags2
> 
> That's a good idea, but won't immediately find any space as it'd just
> leave a hole anyway. But would be good to note in there perhaps, you
> never know when it needs re-arranging again.

struct io_kiocb {
	unsigned flags;
	...
	u8 	 flags2;
};

I rather proposed to have this, which is definitely borderline
ugly but certainly an option.


>> 2) REQ_F_{SINGLE,DOUBLE}_POLL is a weird duplication. Can
>> probably be combined into one flag, or removed at all.
>> Again, sends are usually not so poll heavy and the hot
>> path for recv is multishot.
> 
> Normal receive is also a hot path, even if multishot should be preferred

The degree of hotness is arguable. It's poll, which takes a
spinlock (and disables irqs), does an indirect call, goes into
he socket internals there touching pretty contended parts
like sock_wq. The relative overhead of looking at f_ops should
be nothing.

But the thought was more about combining them,
REQ_F_POLL_ACTIVE, and clear only if it's not double poll.

> in general. Ditto on non-sockets but still pollable files, doing eg read
> for example.
> 
>> 3) we can probably move req->task down and replace it with
>>
>> get_task() {
>>      if (req->ctx->flags & DEFER_TASKRUN)
>>          task = ctx->submitter_task;
>>      else
>>          task = req->task;
>> }
> 
> Assuming ctx flags is hot, which is would generally be, that's not a bad
> idea at all.

As mentioned, task_work_add would be the main user, and
there is already a different branch for DEFER_TASKRUN,
to it implicitly knows that ctx->submitter_task is correct.

> I'll do another loop over this one.

-- 
Pavel Begunkov