Coalesce provided buffer segments
Jens Axboe
maintainerv2 cover Coalesce provided buffer segments
Hi,
When selecting provided buffers for a send/recv for bundles, there's no reason why the number of buffers selected is the same as the mapped segments that will be passed to send/recv. If some (or all) of these buffers are virtually contigious, then they can get collapsed into much fewer segments. Sometimes even just a single segment. This avoids costly iteration on the send/recv processing side.
The return value is the number of bytes sent/received, and the starting buffer ID where the operation begun. This is again identical to how bundles work, from the application point of view this doesn't change anything in terms of how send/recv bundles are handled, hence this is a transparent feature.
Patch 1-2 are just basic prep patches, and patch 3 allows for actual coalescing of segments. This is only enabled for bundles, as those are the types of requests that process multiple buffers in a single operation.
Patches are on top of current -git with the 6.12 changes pulled in, and the partial buffer consumption patchset applied too.
io_uring/kbuf.c | 71 ++++++++++++++++++++++++++++++++++++++++++------- io_uring/kbuf.h | 3 +++ io_uring/net.c | 55 +++++++++++++++++--------------------- io_uring/net.h | 1 + 4 files changed, 89 insertions(+), 41 deletions(-)
Since v1:
- Rebase on current tree(s)
Signature · 1 line
Jens Axboe
Jens Axboe
maintainerPATCH v2 1/3 io_uring/kbuf: have io_provided_buffers_select() take buf_sel_arg
Rather than pass in the iovec in both spots, pass in the buf_sel_arg struct pointer directly. In preparation for needing more of this selection struct off that path.
--- io_uring/kbuf.c | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-)
io_uring/kbuf.c +5 -5
@@ -119,7 +119,7 @@ static void __user *io_provided_buffer_select(struct io_kiocb *req, size_t *len,static int io_provided_buffers_select (struct io_kiocb *req,size_t *len,struct io_buffer_list *bl, -struct iovec *iov) +struct buf_sel_arg *arg) {void __user *buf;
@@ -127,8 +127,8 @@ static int io_provided_buffers_select(struct io_kiocb *req, size_t *len,if (unlikely (!buf))return -ENOBUFS ; - iov[0 ].iov_base = buf; - iov[0 ].iov_len = *len; + arg->iovs[0 ].iov_base = buf; + arg->iovs[0 ].iov_len = *len;return 0 ; }
@@ -310,7 +310,7 @@ int io_buffers_select(struct io_kiocb *req, struct buf_sel_arg *arg,io_kbuf_commit (req, bl, arg->out_len, ret); } }else { - ret =io_provided_buffers_select (req, &arg->out_len, bl, arg->iovs); + ret =io_provided_buffers_select (req, &arg->out_len, bl, arg); } out_unlock:io_ring_submit_unlock (ctx, issue_flags);
@@ -337,7 +337,7 @@ int io_buffers_peek(struct io_kiocb *req, struct buf_sel_arg *arg) }/* don't support multiple buffer selections for legacy */ -return io_provided_buffers_select (req, &arg->max_len, bl, arg->iovs); +return io_provided_buffers_select (req, &arg->max_len, bl, arg); } static int __io_remove_buffers(struct io_ring_ctx *ctx,
Signature · 1 line
2.43.0
Jens Axboe
maintainerPATCH v2 2/3 io_uring/net: pass in io_kiocb to io_bundle_nbufs()
In reply to Jens AxboeIn preparation for needing the io_kiocb in there, and kmsg can always be gotten off that in the first place.
--- io_uring/net.c | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-)
io_uring/net.c +4 -3
@@ -457,8 +457,9 @@ static void io_req_msg_cleanup(struct io_kiocb *req,* data in the iter, then loop the segments to figure out how much we * transferred. */ -static int io_bundle_nbufs (struct io_async_msghdr *kmsg,int ret) +static int io_bundle_nbufs (struct io_kiocb *req,int ret) { +struct io_async_msghdr *kmsg = req->async_data;struct iovec *iov;int nbufs;
@@ -501,7 +502,7 @@ static inline bool io_send_finish(struct io_kiocb *req, int *ret, goto finish; } - cflags =io_put_kbufs (req, *ret,io_bundle_nbufs (kmsg, *ret), issue_flags); + cflags =io_put_kbufs (req, *ret,io_bundle_nbufs (req, *ret), issue_flags);if (bundle_finished || req->flags &REQ_F_BL_EMPTY ) goto finish;
@@ -842,7 +843,7 @@ static inline bool io_recv_finish(struct io_kiocb *req, int *ret, cflags |=IORING_CQE_F_SOCK_NONEMPTY ;if (sr->flags &IORING_RECVSEND_BUNDLE ) { - cflags |=io_put_kbufs (req, *ret,io_bundle_nbufs (kmsg, *ret), + cflags |=io_put_kbufs (req, *ret,io_bundle_nbufs (req, *ret), issue_flags);/* bundle with no more immediate buffers, we're done */ if (req->flags &REQ_F_BL_EMPTY )
Signature · 1 line
2.43.0
Jens Axboe
maintainerPATCH v2 3/3 io_uring/net: allow coalescing of mapped segments
In reply to Jens AxboeFor bundles, when multiple buffers are selected, it's not unlikely that some/all of them will be virtually contigious. If these segments aren't big, then nice wins can be reaped by coalescing them into bigger segments. This makes networking copies more efficient, and reduces the number of iterations that need to be done over an iovec. Ideally, multiple segments that would've been mapped as an ITER_IOVEC before can now be mapped into a single ITER_UBUF iterator.
Example from an io_uring network backend receiving data, with various transfer sizes, over a 100G network link.
recv size coalesce threads bw cpu usage bw diff
=====================================================================
64 0 1 23GB/sec 100% 64 1 1 46GB/sec 79% +100% 64 0 4 81GB/sec 370% 64 1 4 96GB/sec 160% + 20% 256 0 1 44GB/sec 90% 256 1 1 47GB/sec 48% + 7% 256 0 4 90GB/sec 190% 256 1 4 96GB/sec 120% + 7% 1024 0 1 49GB/sec 60% 1024 1 1 50GB/sec 53% + 2% 1024 0 4 94GB/sec 140% 1024 1 4 96GB/sec 120% + 2%
where obviously small buffer sizes benefit the most, but where an efficiency gain is seen even at higher buffer sizes as well.
--- io_uring/kbuf.c | 61 +++++++++++++++++++++++++++++++++++++++++++++---- io_uring/kbuf.h | 3 +++ io_uring/net.c | 48 ++++++++++++++++---------------------- io_uring/net.h | 1 + 4 files changed, 80 insertions(+), 33 deletions(-)
io_uring/kbuf.c +56 -5
@@ -129,6 +129,7 @@ static int io_provided_buffers_select(struct io_kiocb *req, size_t *len, arg->iovs[0 ].iov_base = buf; arg->iovs[0 ].iov_len = *len; + arg->nsegs =1 ;return 0 ; }
@@ -194,11 +195,16 @@ void __user *io_buffer_select(struct io_kiocb *req, size_t *len,/* cap it at a reasonable 256, will be one page even for 4K */ #define PEEK_MAX_IMPORT 256 +/* +* Returns how many iovecs were used to fill the range. arg->nsegs contains +* the number of buffers mapped, which may be less than the return value if +* segments were coalesced. +*/ static int io_ring_buffers_peek (struct io_kiocb *req,struct buf_sel_arg *arg,struct io_buffer_list *bl) {struct io_uring_buf_ring *br = bl->buf_ring; -struct iovec *iov = arg->iovs; +struct iovec *prev_iov, *iov = arg->iovs;int nr_iovs = arg->nr_iovs;__u16 nr_avail, tail, head;struct io_uring_buf *buf;
@@ -253,9 +259,11 @@ static int io_ring_buffers_peek(struct io_kiocb *req, struct buf_sel_arg *arg,if (!arg->max_len) arg->max_len =INT_MAX ; + prev_iov =NULL ; req->buf_index = buf->bid; do { u32 len = buf->len; + void __user *ubuf; /* truncate end piece, if needed, for non partial buffers */ if (len > arg->max_len) {
@@ -264,10 +272,20 @@ static int io_ring_buffers_peek(struct io_kiocb *req, struct buf_sel_arg *arg, buf->len = len; } - iov->iov_base =u64_to_user_ptr (buf->addr); - iov->iov_len = len; - iov++; + ubuf =u64_to_user_ptr (buf->addr); +if (prev_iov && + prev_iov->iov_base + prev_iov->iov_len == ubuf && + prev_iov->iov_len + len <=INT_MAX ) { + prev_iov->iov_len += len; + }else { + iov->iov_base = ubuf; + iov->iov_len = len; +if (arg->coalesce) + prev_iov = iov; + iov++; + } + arg->nsegs++; arg->out_len += len; arg->max_len -= len;if (!arg->max_len)
@@ -280,7 +298,8 @@ static int io_ring_buffers_peek(struct io_kiocb *req, struct buf_sel_arg *arg, req->flags |=REQ_F_BL_EMPTY ; req->flags |=REQ_F_BUFFER_RING ; - req->buf_list = bl; +if (arg->coalesce) + req->buf_list = bl;return iov - arg->iovs; }
@@ -340,6 +359,38 @@ int io_buffers_peek(struct io_kiocb *req, struct buf_sel_arg *arg)return io_provided_buffers_select (req, &arg->max_len, bl, arg); } +int io_buffer_segments (struct io_kiocb *req,int nbytes) +{ +struct io_uring_buf_ring *br; +struct io_buffer_list *bl; +int nbufs =0 ; +unsigned bid; + +/* +* Safe to use ->buf_list here, as coalescing can only have happened +* if we remained lock throughout the operation. Unlocked usage must +* not have buf_sel_arg->coalesce set to true +*/ + bl = req->buf_list; +if (unlikely (!bl || !(bl->flags &IOBL_BUF_RING ))) +return 1 ; + + bid = req->buf_index; + br = bl->buf_ring; +do { +struct io_uring_buf *buf; +int this_len; + + buf =io_ring_head_to_buf (br, bid, bl->mask); + this_len =min_t (int, buf->len, nbytes); + nbufs++; + bid++; + nbytes -= this_len; + }while (nbytes); + +return nbufs; +} +static int __io_remove_buffers (struct io_ring_ctx *ctx,struct io_buffer_list *bl,unsigned nbufs) {
io_uring/kbuf.h +3 -0
@@ -61,8 +61,11 @@ struct buf_sel_arg {size_t max_len;unsigned short nr_iovs;unsigned short mode; +unsigned short nsegs; +bool coalesce; }; +int io_buffer_segments (struct io_kiocb *req,int nbytes);void __user *io_buffer_select (struct io_kiocb *req,size_t *len,unsigned int issue_flags); int io_buffers_select(struct io_kiocb *req, struct buf_sel_arg *arg,
io_uring/net.c +20 -28
89 lines not on this page · raw message
io_uring/net.h +1 -0
@@ -9,6 +9,7 @@ struct io_async_msghdr {/* points to an allocated iov, if NULL we use fast_iov instead */ struct iovec *free_iov;int free_iov_nr; +int nbufs;int namelen;__kernel_size_t controllen;__kernel_size_t payloadlen;
Signature · 1 line
2.43.0