Re: [PATCH liburing 1/1] man: add io_uring_register_region.3

[Jens Axboe <axboe@kernel.dk>]

On 1/14/26 11:50 AM, Pavel Begunkov wrote:
> On 1/14/26 17:23, Jens Axboe wrote:
>> On 1/14/26 9:04 AM, Pavel Begunkov wrote:
>>> On 1/14/26 14:54, Pavel Begunkov wrote:
>>>> On 1/14/26 14:42, Pavel Begunkov wrote:
>>>>> On 1/13/26 22:37, Jens Axboe wrote:
>>>>>> On 1/13/26 2:31 PM, Gabriel Krisman Bertazi wrote:
>>>>>>> Pavel Begunkov <asml.silence@gmail.com> writes:
>>>>>>>
>>>>>>>> Describe the region API. As it was created for a bunch of ideas in mind,
>>>>>>>> it doesn't go into details about wait argument passing, which I assume
>>>>>>>> will be a separate page the region description can refer to.
>>>>>>>>
>>>>>>>
>>>>>>> Hey, Pavel.
>>>>>>
>>>>>> I did a bunch of spelling and phrasing fixups when applying, can you
>>>>>> take a look at the repo and send a patch for the others? Thanks!
>>>>>
>>>>> "Upon successful completion, the memory region may then be used, for
>>>>> example, to pass waiting parameters to the io_uring_enter(2) system
>>>>> call in a more efficient manner as it avoids copying wait related data
>>>>> for each wait event."
>>>>>
>>>>> Doesn't matter much, but this change is somewhat misleading. Both copy
>>>>> args same number of times (i.e. unsafe_get_user() instead of
>>>>> copy_from_user()), which is why I was a bit vague with that
>>>>> "in an efficient manner".
>>>>
>>>> Hmm, actually the normal / non-registered way does make an extra
>>>> copy, even though it doesn't have to.
>>>
>>> And the compiler is smart enough to optimise it out since
>>> it's all on stack.
>>
>> Not sure I follow these emails. For the normal case,
>> io_validate_ext_arg() copies in the args via a normal user copy, which
>> depending on options and the arch (or even sub-arch, amd more expensive)
>> is more or less expensive.
> 
> In the end, after prep that is still just a move instruction, e.g.
> for x86. And it loads into a register and stores it into ext_arg,
> just like with registration. User copy needs to prepare page fault
> handling / etc., which could be costly (e.g. I see stac + lfence
> in asm), but that's not exactly about avoiding copies.

Those are implementation details. The user copy is stac/clac, and then
the loads. This is what makes it more expensive. I don't want to be
writing about stac/clac in the man page, that's irrelevant to the user.

-- 
Jens Axboe