On 10/24/24 15:48, Jens Axboe wrote:
On 10/24/24 8:44 AM, Pavel Begunkov wrote:
On 10/23/24 17:07, Jens Axboe wrote:
Provided buffers inform the kernel which buffer group ID to pick a
buffer from for transfer. Normally that buffer contains the usual
addr + length information, as well as a buffer ID that is passed back
at completion time to inform the application of which buffer was used
for the transfer.
However, if registered and provided buffers are combined, then the
provided buffer must instead tell the kernel which registered buffer
index should be used, and the length/offset within that buffer. Rather
than store the addr + length, the application must instead store this
information instead.
If provided buffers are used with send zc, then those buffers must be
an index into a registered buffer. Change the mapping type to use
KBUF_MODE_BVEC, which tells the kbuf handlers to turn the mappings
into bio_vecs rather than iovecs. Then all that is needed is to
setup our iov_iterator to use iov_iter_bvec().
Signed-off-by: Jens Axboe <axboe@xxxxxxxxx>
---
...
diff --git a/io_uring/net.h b/io_uring/net.h
index 52bfee05f06a..e052762cf85d 100644
--- a/io_uring/net.h
+++ b/io_uring/net.h
@@ -5,9 +5,15 @@
struct io_async_msghdr {
#if defined(CONFIG_NET)
- struct iovec fast_iov;
+ union {
+ struct iovec fast_iov;
+ struct bio_vec fast_bvec;
+ };
/* points to an allocated iov, if NULL we use fast_iov instead */
- struct iovec *free_iov;
+ union {
+ struct iovec *free_iov;
+ struct bio_vec *free_bvec;
I'd rather not do it like that, aliasing with reusing memory and
counting the number is a recipe for disaster when scattered across
code. E.g. seems you change all(?) iovec allocations to allocate
based on the size of the larger structure.
Counting bytes as in my series is less fragile, otherwise it needs
a new structure and a set of helpers that can be kept together.
I have been pondering this, because I'm not a huge fan either. But
outside of the space side, it does come out pretty nicely/clean. This
series is really just a WIP posting as per the RFC, mostly just so we
can come up with something that's clean enough and works for both cases,
as it does have the caching that your series does not. And to
facilitate some more efficient TX/RX zero copy testing.
The thing is, I know how to implement caching on top of my series,
but as I commented in the other thread, I don't think it can reuse
the incremental helper well. At least it can try to unify the imu
checking / parsing, bvec copy-setup, but that's minor.
--
Pavel Begunkov
