Re: [PATCH v6.20] io_uring/rsrc: refactor io_import_kbuf() to use single loop

[Caleb Sander Mateos <csander@purestorage.com>]

On Wed, Dec 17, 2025 at 9:28 PM Ming Lei <ming.lei@redhat.com> wrote:
>
> On Thu, Dec 18, 2025 at 09:42:43AM +0800, Ming Lei wrote:
> > On Wed, Dec 17, 2025 at 11:16:47PM +0800, huang-jl wrote:
> > > The code looks correct to me.
> > >
> > > > This simplifies the logic
> > >
> > > I'm not an expert in Linux development, but from my perspective, the
> > > original version seems simpler and more readable. The semantics of
> > > iov_iter_advance() are clear and well-understood.
> > >
> > > That said, I understand the appeal of merging them into a single loop.
> > >
> > > > and avoids the overhead of iov_iter_advance()
> > >
> > > Could you clarify what overhead you mean? If it's the function call
> > > overhead, I think the compiler would inline it anyway. The actual
> > > iteration work seems equivalent between both approaches.
> >
> > iov_iter_advance() is global function, and it can't be inline.
> >
> > Also single loop is more readable, cause ->iov_offset can be ignored easily.
> >
> > In theory, re-calculating nr_segs isn't necessary, it is just for avoiding
> > potential split, however not get idea how it is triggered. Nitesh didn't
> > mention the exact reason:
> >
> > https://lkml.org/lkml/2025/4/16/351
> >
> > I will look at the reason and see if it can be avoided.
>
> The reason is in both bio_iov_bvec_set() and bio_may_need_split().

nr_segs is not just a performance optimization, it's part of the
struct iov_iter API and used by iov_iter_extract_bvec_pages(), as
huang-jl pointed out. I don't think it's a good idea to assume that
nr_segs isn't going to be used and doesn't need to be calculated
correctly.

I think this patch is a definite improvement as it reduces the number
of assumptions about the internal structure of a bvec iov_iter. The
remaining assignment to iter->iov_offset is unfortunate, but I don't
see a great way around it.

My only suggestion would be to separate out a "find first bvec" loop
and a "find last bvec" loop to make the loop body less branchy.
Something like this:
const struct bio_vec *bvec = imu->bvec, bvec_start;
size_t remaining;
unsigned bvec_avail;

while (offset >= bvec->bv_len) {
        offset -= bvec->bv_len;
        bvec++;
}

bvec_avail = bvec->bv_len - offset;
bvec_start = bvec;
remaining = len;
while (remaining > bvec_avail) {
        remaining -= bvec_avail;
        bvec++;
        bvec_avail = bvec->bv_len;
}

iov_iter_bvec(iter, ddir, start_bvec, bvec - start_bvec + 1, len);
iter->iov_offset = offset;

Best,
Caleb

>
> ->bi_vcnt doesn't make sense for cloned bio, and shouldn't be used as multiple
> segment hint.
>
> However, it also shows bio_split_rw() is too heavy.
>
>
> Thanks,
> Ming
>