On 05/02/2021 14:45, Jens Axboe wrote:
> On 2/5/21 3:06 AM, Pavel Begunkov wrote:
>> On 05/02/2021 09:58, Stefan Metzmacher wrote:
>>> Hi Pavel,
>>>
>>>>>> static int io_sendmsg_copy_hdr(struct io_kiocb *req,
>>>>>> struct io_async_msghdr *iomsg)
>>>>>> {
>>>>>> - iomsg->iov = iomsg->fast_iov;
>>>>>> iomsg->msg.msg_name = &iomsg->addr;
>>>>>> + iomsg->free_iov = iomsg->fast_iov;
>>>>>
>>>>> Why this? Isn't the idea of this patch that free_iov is never == fast_iov?
>>>>
>>>> That's a part of __import_iovec() and sendmsg_copy_msghdr() API, you pass
>>>> fast_iov as such and get back NULL or a newly allocated one in it.
>>> I think that should at least get a comment to make this clear and
>>> maybe a temporary variable like this:
>>>
>>> tmp_iov = iomsg->fast_iov;
>>> __import_iovec(..., &tmp_iov, ...);
>>> iomsg->free_iov = tmp_iov;
>>
>> I'd rather disagree. It's a well known (ish) API, and I deliberately
>> placed such assignments right before import_iovec/etc.
>
> Agree on that, it's kind of a weird idiom, but it's used throughout the
> kernel. However:
Maybe someday someone will refactor it and make it accepting inline_vecs
directly...
>
>>>>> kfree() handles NULL, or is this a hot path and we want to avoid a function call?
>>>>
>>>> Yes, the hot path here is not having iov allocated, and Jens told before
>>>> that he had observed overhead for a similar place in io_[read,write].
>>>
>>> Ok, a comment would also help here...
>
> I do agree on that one, since otherwise we get patches for it as has
> been proven by the few other spots... I'll add then when queueing this
> up.
I didn't disagree with that, just forgot to throw one to protect ourselves
from static analysis tools. Thanks for fixing up
--
Pavel Begunkov
