io_uring/net: allow coalescing of mapped segments

For 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.

Signed-off-by: Jens Axboe <axboe@kernel.dk>

diff --git a/io_uring/kbuf.c b/io_uring/kbuf.c
index 2ccf33b8daa49216b570c2c1f6a1880b5abec87b..49251eddfc52ee8c8a1b7ab5f9928e3154e595fd 100644 (file)
--- a/io_uring/kbuf.c
+++ b/io_uring/kbuf.c
@@ -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 1;
 }
 
@@ -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 <= MAX_RW_COUNT) {
+                       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)
 {

diff --git a/io_uring/kbuf.h b/io_uring/kbuf.h
index 36aadfe5ac0027f432227d7874d59f08115a5bf0..37a35ce31c0c65acbaf69161fafddd7f80384a83 100644 (file)
--- a/io_uring/kbuf.h
+++ b/io_uring/kbuf.h
@@ -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,

diff --git a/io_uring/net.c b/io_uring/net.c
index 14dd60bed676aef13e1bc3e956a0bf3bb77ff19e..4553465805f727ebe7ee786781b937ce5f775ba6 100644 (file)
--- a/io_uring/net.c
+++ b/io_uring/net.c
@@ -460,33 +460,16 @@ static void io_req_msg_cleanup(struct io_kiocb *req,
 static int io_bundle_nbufs(struct io_kiocb *req, int ret)
 {
        struct io_async_msghdr *kmsg = req->async_data;
-       struct iovec *iov;
-       int nbufs;
 
-       /* no data is always zero segments, and a ubuf is always 1 segment */
+       /* no data is always zero segments */
        if (ret <= 0)
                return 0;
-       if (iter_is_ubuf(&kmsg->msg.msg_iter))
-               return 1;
-
-       iov = kmsg->free_iov;
-       if (!iov)
-               iov = &kmsg->fast_iov;
-
-       /* if all data was transferred, it's basic pointer math */
+       /* if all data was transferred, we already know the number of buffers */
        if (!iov_iter_count(&kmsg->msg.msg_iter))
-               return iter_iov(&kmsg->msg.msg_iter) - iov;
-
-       /* short transfer, count segments */
-       nbufs = 0;
-       do {
-               int this_len = min_t(int, iov[nbufs].iov_len, ret);
-
-               nbufs++;
-               ret -= this_len;
-       } while (ret);
+               return kmsg->nbufs;
 
-       return nbufs;
+       /* short transfer, iterate buffers to find number of segments */
+       return io_buffer_segments(req, ret);
 }
 
 static inline bool io_send_finish(struct io_kiocb *req, int *ret,
@@ -600,6 +583,7 @@ retry_bundle:
                        .iovs = &kmsg->fast_iov,
                        .max_len = min_not_zero(sr->len, INT_MAX),
                        .nr_iovs = 1,
+                       .coalesce = !(issue_flags & IO_URING_F_UNLOCKED),
                };
 
                if (kmsg->free_iov) {
@@ -623,6 +607,7 @@ retry_bundle:
                        req->flags |= REQ_F_NEED_CLEANUP;
                }
                sr->len = arg.out_len;
+               kmsg->nbufs = arg.nsegs;
 
                if (ret == 1) {
                        sr->buf = arg.iovs[0].iov_base;
@@ -1078,6 +1063,7 @@ static int io_recv_buf_select(struct io_kiocb *req, struct io_async_msghdr *kmsg
                        .iovs = &kmsg->fast_iov,
                        .nr_iovs = 1,
                        .mode = KBUF_MODE_EXPAND,
+                       .coalesce = true,
                };
 
                if (kmsg->free_iov) {
@@ -1093,7 +1079,18 @@ static int io_recv_buf_select(struct io_kiocb *req, struct io_async_msghdr *kmsg
                if (unlikely(ret < 0))
                        return ret;
 
-               /* special case 1 vec, can be a fast path */
+               if (arg.iovs != &kmsg->fast_iov && arg.iovs != kmsg->free_iov) {
+                       kmsg->free_iov_nr = arg.nsegs;
+                       kmsg->free_iov = arg.iovs;
+                       req->flags |= REQ_F_NEED_CLEANUP;
+               }
+               kmsg->nbufs = arg.nsegs;
+
+               /*
+                * Special case 1 vec, can be a fast path. Note that multiple
+                * contig buffers may get mapped to a single vec, but we can
+                * still use ITER_UBUF for those.
+                */
                if (ret == 1) {
                        sr->buf = arg.iovs[0].iov_base;
                        sr->len = arg.iovs[0].iov_len;
@@ -1101,11 +1098,6 @@ static int io_recv_buf_select(struct io_kiocb *req, struct io_async_msghdr *kmsg
                }
                iov_iter_init(&kmsg->msg.msg_iter, ITER_DEST, arg.iovs, ret,
                                arg.out_len);
-               if (arg.iovs != &kmsg->fast_iov && arg.iovs != kmsg->free_iov) {
-                       kmsg->free_iov_nr = ret;
-                       kmsg->free_iov = arg.iovs;
-                       req->flags |= REQ_F_NEED_CLEANUP;
-               }
        } else {
                void __user *buf;
 

diff --git a/io_uring/net.h b/io_uring/net.h
index 52bfee05f06a14ea3df09b94ab0b26f2afad4e20..b9a453da4a0ff44a0b0c15557df04051ef2838d0 100644 (file)
--- a/io_uring/net.h
+++ b/io_uring/net.h
@@ -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;