Netdev List
 help / color / mirror / Atom feed
From: Pavel Begunkov <asml.silence@gmail.com>
To: io-uring@vger.kernel.org
Cc: asml.silence@gmail.com, netdev@vger.kernel.org
Subject: [PATCH review-only 05/17] io_uring/zcrx: coalesce same-niov RQEs on refill
Date: Sat, 11 Jul 2026 10:11:28 +0100	[thread overview]
Message-ID: <267af1e26c3b17fd9fa9b56e4b2316d7412546db.1783616211.git.asml.silence@gmail.com> (raw)
In-Reply-To: <cover.1783616211.git.asml.silence@gmail.com>

With large rx piages I often see >10 sequential RQEs referring to the
same niov. Instead of putting them one by one, count such RQEs during
parsing and batch refcounting for the niov.

Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
---
 io_uring/zcrx.c | 56 +++++++++++++++++++++++++++++++------------------
 1 file changed, 36 insertions(+), 20 deletions(-)

diff --git a/io_uring/zcrx.c b/io_uring/zcrx.c
index 1b8d748b35e7..cb73dca3c1ee 100644
--- a/io_uring/zcrx.c
+++ b/io_uring/zcrx.c
@@ -359,16 +359,16 @@ static inline atomic_t *io_get_user_counter(struct net_iov *niov)
 	return &area->user_refs[net_iov_idx(niov)];
 }
 
-static bool io_zcrx_put_niov_uref(struct net_iov *niov)
+static bool io_zcrx_put_niov_uref(struct net_iov *niov, unsigned refs)
 {
 	atomic_t *uref = io_get_user_counter(niov);
 	int old;
 
 	old = atomic_read(uref);
 	do {
-		if (unlikely(old == 0))
+		if (unlikely(old < refs))
 			return false;
-	} while (!atomic_try_cmpxchg(uref, &old, old - 1));
+	} while (!atomic_try_cmpxchg(uref, &old, old - refs));
 
 	return true;
 }
@@ -1163,6 +1163,22 @@ static inline bool io_parse_rqe(struct io_uring_zcrx_rqe *rqe,
 	return true;
 }
 
+static bool zcrx_put_refill_niov(struct net_iov *niov, struct page_pool *pp,
+				 unsigned refs)
+{
+	netmem_ref netmem = net_iov_to_netmem(niov);
+
+	if (!io_zcrx_put_niov_uref(niov, refs))
+		return false;
+	if (page_pool_unref_netmem(netmem, refs) != 0)
+		return false;
+	if (unlikely(niov->desc.pp != pp)) {
+		io_zcrx_return_niov(niov);
+		return false;
+	}
+	return true;
+}
+
 static unsigned io_zcrx_ring_refill(struct page_pool *pp,
 				    struct io_zcrx_ifq *ifq,
 				    netmem_ref *netmems, unsigned to_alloc)
@@ -1170,34 +1186,34 @@ static unsigned io_zcrx_ring_refill(struct page_pool *pp,
 	struct zcrx_rq *rq = &ifq->rq;
 	struct io_uring_zcrx_rqe *rqe;
 	struct zcrx_rq_iter it;
+	struct net_iov *niov = NULL;
+	unsigned niov_refs = 0;
 	unsigned allocated = 0;
 
 	guard(spinlock_bh)(&rq->lock);
 
 	zcrx_rq_iter_init(&it, rq);
 
-	while (zcrx_rq_iter_next(&it, rq, &rqe)) {
-		struct net_iov *niov;
-		netmem_ref netmem;
+	while (allocated < to_alloc - 1 && zcrx_rq_iter_next(&it, rq, &rqe)) {
+		struct net_iov *next_niov;
 
-		if (!io_parse_rqe(rqe, ifq, &niov))
-			continue;
-		if (!io_zcrx_put_niov_uref(niov))
+		if (!io_parse_rqe(rqe, ifq, &next_niov))
 			continue;
-
-		netmem = net_iov_to_netmem(niov);
-		if (!page_pool_unref_and_test(netmem))
-			continue;
-
-		if (unlikely(niov->desc.pp != pp)) {
-			io_zcrx_return_niov(niov);
+		if (niov == next_niov) {
+			niov_refs++;
 			continue;
 		}
+		if (niov && zcrx_put_refill_niov(niov, pp, niov_refs)) {
+			netmems[allocated] = net_iov_to_netmem(niov);
+			allocated++;
+		}
+		niov = next_niov;
+		niov_refs = 1;
+	}
 
-		netmems[allocated] = netmem;
+	if (niov && zcrx_put_refill_niov(niov, pp, niov_refs)) {
+		netmems[allocated] = net_iov_to_netmem(niov);
 		allocated++;
-		if (allocated >= to_alloc)
-			break;
 	}
 
 	smp_store_release(&rq->ring->head, rq->cached_head);
@@ -1401,7 +1417,7 @@ static void zcrx_return_buffers(netmem_ref *netmems, unsigned nr)
 		netmem_ref netmem = netmems[i];
 		struct net_iov *niov = netmem_to_net_iov(netmem);
 
-		if (!io_zcrx_put_niov_uref(niov))
+		if (!io_zcrx_put_niov_uref(niov, 1))
 			continue;
 		if (!page_pool_unref_and_test(netmem))
 			continue;
-- 
2.54.0


WARNING: multiple messages have this Message-ID (diff)
From: Pavel Begunkov <asml.silence@gmail.com>
To: io-uring@vger.kernel.org
Cc: asml.silence@gmail.com, netdev@vger.kernel.org
Subject: [PATCH review-only 05/17] io_uring/zcrx: coalesce same-niov RQEs on refill
Date: Sat, 11 Jul 2026 11:39:58 +0100	[thread overview]
Message-ID: <267af1e26c3b17fd9fa9b56e4b2316d7412546db.1783616211.git.asml.silence@gmail.com> (raw)
Message-ID: <20260711103958.hlViraouRVnJ-N4kcI0nW7GaFlVKC8n2OfnFCWmmM4s@z> (raw)
In-Reply-To: <cover.1783616211.git.asml.silence@gmail.com>

With large rx piages I often see >10 sequential RQEs referring to the
same niov. Instead of putting them one by one, count such RQEs during
parsing and batch refcounting for the niov.

Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
---
 io_uring/zcrx.c | 56 +++++++++++++++++++++++++++++++------------------
 1 file changed, 36 insertions(+), 20 deletions(-)

diff --git a/io_uring/zcrx.c b/io_uring/zcrx.c
index 1b8d748b35e7..cb73dca3c1ee 100644
--- a/io_uring/zcrx.c
+++ b/io_uring/zcrx.c
@@ -359,16 +359,16 @@ static inline atomic_t *io_get_user_counter(struct net_iov *niov)
 	return &area->user_refs[net_iov_idx(niov)];
 }
 
-static bool io_zcrx_put_niov_uref(struct net_iov *niov)
+static bool io_zcrx_put_niov_uref(struct net_iov *niov, unsigned refs)
 {
 	atomic_t *uref = io_get_user_counter(niov);
 	int old;
 
 	old = atomic_read(uref);
 	do {
-		if (unlikely(old == 0))
+		if (unlikely(old < refs))
 			return false;
-	} while (!atomic_try_cmpxchg(uref, &old, old - 1));
+	} while (!atomic_try_cmpxchg(uref, &old, old - refs));
 
 	return true;
 }
@@ -1163,6 +1163,22 @@ static inline bool io_parse_rqe(struct io_uring_zcrx_rqe *rqe,
 	return true;
 }
 
+static bool zcrx_put_refill_niov(struct net_iov *niov, struct page_pool *pp,
+				 unsigned refs)
+{
+	netmem_ref netmem = net_iov_to_netmem(niov);
+
+	if (!io_zcrx_put_niov_uref(niov, refs))
+		return false;
+	if (page_pool_unref_netmem(netmem, refs) != 0)
+		return false;
+	if (unlikely(niov->desc.pp != pp)) {
+		io_zcrx_return_niov(niov);
+		return false;
+	}
+	return true;
+}
+
 static unsigned io_zcrx_ring_refill(struct page_pool *pp,
 				    struct io_zcrx_ifq *ifq,
 				    netmem_ref *netmems, unsigned to_alloc)
@@ -1170,34 +1186,34 @@ static unsigned io_zcrx_ring_refill(struct page_pool *pp,
 	struct zcrx_rq *rq = &ifq->rq;
 	struct io_uring_zcrx_rqe *rqe;
 	struct zcrx_rq_iter it;
+	struct net_iov *niov = NULL;
+	unsigned niov_refs = 0;
 	unsigned allocated = 0;
 
 	guard(spinlock_bh)(&rq->lock);
 
 	zcrx_rq_iter_init(&it, rq);
 
-	while (zcrx_rq_iter_next(&it, rq, &rqe)) {
-		struct net_iov *niov;
-		netmem_ref netmem;
+	while (allocated < to_alloc - 1 && zcrx_rq_iter_next(&it, rq, &rqe)) {
+		struct net_iov *next_niov;
 
-		if (!io_parse_rqe(rqe, ifq, &niov))
-			continue;
-		if (!io_zcrx_put_niov_uref(niov))
+		if (!io_parse_rqe(rqe, ifq, &next_niov))
 			continue;
-
-		netmem = net_iov_to_netmem(niov);
-		if (!page_pool_unref_and_test(netmem))
-			continue;
-
-		if (unlikely(niov->desc.pp != pp)) {
-			io_zcrx_return_niov(niov);
+		if (niov == next_niov) {
+			niov_refs++;
 			continue;
 		}
+		if (niov && zcrx_put_refill_niov(niov, pp, niov_refs)) {
+			netmems[allocated] = net_iov_to_netmem(niov);
+			allocated++;
+		}
+		niov = next_niov;
+		niov_refs = 1;
+	}
 
-		netmems[allocated] = netmem;
+	if (niov && zcrx_put_refill_niov(niov, pp, niov_refs)) {
+		netmems[allocated] = net_iov_to_netmem(niov);
 		allocated++;
-		if (allocated >= to_alloc)
-			break;
 	}
 
 	smp_store_release(&rq->ring->head, rq->cached_head);
@@ -1401,7 +1417,7 @@ static void zcrx_return_buffers(netmem_ref *netmems, unsigned nr)
 		netmem_ref netmem = netmems[i];
 		struct net_iov *niov = netmem_to_net_iov(netmem);
 
-		if (!io_zcrx_put_niov_uref(niov))
+		if (!io_zcrx_put_niov_uref(niov, 1))
 			continue;
 		if (!page_pool_unref_and_test(netmem))
 			continue;
-- 
2.54.0


  parent reply	other threads:[~2026-07-11  9:12 UTC|newest]

Thread overview: 36+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2026-07-11  9:11 [PATCH review-only 00/17] zcrx RQ improvements and dynamic memory provisioning Pavel Begunkov
2026-07-11  9:11 ` [PATCH review-only 01/17] io_uring/zcrx: scale refilling with large pages Pavel Begunkov
2026-07-11 10:39   ` Pavel Begunkov
2026-07-11  9:11 ` [PATCH review-only 02/17] io_uring/zcrx: move RQ head/tail to separate cache lines Pavel Begunkov
2026-07-11 10:39   ` Pavel Begunkov
2026-07-11  9:11 ` [PATCH review-only 03/17] io_uring/zcrx: add RQ iterator Pavel Begunkov
2026-07-11 10:39   ` Pavel Begunkov
2026-07-11  9:11 ` [PATCH review-only 04/17] io_uring/zcrx: cache RQ tail Pavel Begunkov
2026-07-11 10:39   ` Pavel Begunkov
2026-07-11  9:11 ` Pavel Begunkov [this message]
2026-07-11 10:39   ` [PATCH review-only 05/17] io_uring/zcrx: coalesce same-niov RQEs on refill Pavel Begunkov
2026-07-11  9:11 ` [PATCH review-only 06/17] io_uring/zcrx: constify area_reg on import Pavel Begunkov
2026-07-11 10:39   ` Pavel Begunkov
2026-07-11  9:11 ` [PATCH review-only 07/17] io_uring/zcrx: add helper for deriving area token Pavel Begunkov
2026-07-11 10:40   ` Pavel Begunkov
2026-07-11  9:11 ` [PATCH review-only 08/17] io_uring/zcrx: don't pass ifq_reg to area creation Pavel Begunkov
2026-07-11 10:40   ` Pavel Begunkov
2026-07-11  9:11 ` [PATCH review-only 09/17] io_uring/zcrx: split dmabuf unmap and release Pavel Begunkov
2026-07-11 10:40   ` Pavel Begunkov
2026-07-11  9:11 ` [PATCH review-only 10/17] io_uring/zcrx: unmap under netdev lock Pavel Begunkov
2026-07-11 10:40   ` Pavel Begunkov
2026-07-11  9:11 ` [PATCH review-only 11/17] io_uring/zcrx: split append out of area creation Pavel Begunkov
2026-07-11 10:40   ` Pavel Begunkov
2026-07-11  9:11 ` [PATCH review-only 12/17] io_uring/zcrx: move freelist lock to struct zcrx Pavel Begunkov
2026-07-11 10:40   ` Pavel Begunkov
2026-07-11  9:11 ` [PATCH review-only 13/17] io_uring/zcrx: array of areas Pavel Begunkov
2026-07-11 10:40   ` Pavel Begunkov
2026-07-11  9:11 ` [PATCH review-only 14/17] io_uring/zcrx: pass area_id to __zcrx_create_area() Pavel Begunkov
2026-07-11 10:40   ` Pavel Begunkov
2026-07-11  9:11 ` [PATCH review-only 15/17] io_uring/zcrx: add dynamic area creation Pavel Begunkov
2026-07-11 10:40   ` Pavel Begunkov
2026-07-11  9:11 ` [PATCH review-only 16/17] io_urint/zcrx: narrow var scope in io_zcrx_recv_skb() Pavel Begunkov
2026-07-11 10:40   ` Pavel Begunkov
2026-07-11 10:39 ` [PATCH review-only 00/17] zcrx RQ improvements and dynamic memory provisioning Pavel Begunkov
2026-07-11 10:39 ` [PATCH RESEND " Pavel Begunkov
2026-07-11 10:40 ` [PATCH review-only 17/17] io_uring/zcrx: don't reload skb_shinfo Pavel Begunkov

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=267af1e26c3b17fd9fa9b56e4b2316d7412546db.1783616211.git.asml.silence@gmail.com \
    --to=asml.silence@gmail.com \
    --cc=io-uring@vger.kernel.org \
    --cc=netdev@vger.kernel.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox