All of lore.kernel.org
 help / color / mirror / Atom feed
From: Ed Czeck <ed.czeck@atomicrules.com>
To: dev@dpdk.org, stephen@networkplumber.org
Cc: Shepard Siegel <shepard.siegel@atomicrules.com>,
	John Miller <john.miller@atomicrules.com>
Subject: [PATCH v3 4/4] net/ark: improve Rx queue recovery after mbuf exhaustion
Date: Mon, 15 Sep 2025 11:02:04 -0400	[thread overview]
Message-ID: <20250915150204.1719628-4-ed.czeck@atomicrules.com> (raw)
In-Reply-To: <20250915150204.1719628-1-ed.czeck@atomicrules.com>

use 4-K aligned buffers to reduce PCIe requests
reduce message spew
attempt to allocate smaller chunks of buffers during starvation

Signed-off-by: Ed Czeck <ed.czeck@atomicrules.com>
---
v3:
- Use #define for pcie boundary

The PCIE spec states "requests must not specify an Address/Length
combination that causes a Memory Space access to cross a 4-KB
boundary."  Buffer alignment to 4KB boundaries will minimize the
number of 4-KB boundary in a given buffer and then can reduce
TLP read requests from FPGA to host.

v2:
- reduced message to single line.
- Added comments on buffer alignment.  PCIe devices deal with page
size of 4096 bytes.  By aligning this buffer to a 4K boundary, we
will reduce the number of PCIe read requests.
---
 drivers/net/ark/ark_ethdev_rx.c | 35 ++++++++++++++++++++++-----------
 1 file changed, 24 insertions(+), 11 deletions(-)

diff --git a/drivers/net/ark/ark_ethdev_rx.c b/drivers/net/ark/ark_ethdev_rx.c
index 1b5c4b64a4..98fd52949b 100644
--- a/drivers/net/ark/ark_ethdev_rx.c
+++ b/drivers/net/ark/ark_ethdev_rx.c
@@ -14,6 +14,7 @@
 #define ARK_RX_META_SIZE 32
 #define ARK_RX_META_OFFSET (RTE_PKTMBUF_HEADROOM - ARK_RX_META_SIZE)
 #define ARK_RX_MPU_CHUNK (64U)
+#define ARK_PCIE_BOUNDARY 4096
 
 /* Forward declarations */
 struct ark_rx_queue;
@@ -42,6 +43,7 @@ struct __rte_cache_aligned ark_rx_queue {
 	rx_user_meta_hook_fn rx_user_meta_hook;
 	void *ext_user_data;
 
+	uint32_t starvation;
 	uint32_t dataroom;
 	uint32_t headroom;
 
@@ -57,8 +59,6 @@ struct __rte_cache_aligned ark_rx_queue {
 	/* The queue Index is used within the dpdk device structures */
 	uint16_t queue_index;
 
-	uint32_t unused;
-
 	/* next cache line - fields written by device */
 	alignas(RTE_CACHE_LINE_MIN_SIZE) RTE_MARKER cacheline1;
 
@@ -187,10 +187,11 @@ eth_ark_dev_rx_queue_setup(struct rte_eth_dev *dev,
 				   nb_desc * sizeof(struct rte_mbuf *),
 				   512,
 				   socket_id);
+	/* Align buffer to PCIe's boundary size to reduce upstream read request TLPs from FPGA */
 	queue->paddress_q =
 		rte_zmalloc_socket("Ark_rx_queue paddr",
 				   nb_desc * sizeof(rte_iova_t),
-				   512,
+				   ARK_PCIE_BOUNDARY,
 				   socket_id);
 
 	if (queue->reserve_q == 0 || queue->paddress_q == 0) {
@@ -265,6 +266,9 @@ eth_ark_recv_pkts(void *rx_queue,
 		return 0;
 	if (unlikely(nb_pkts == 0))
 		return 0;
+	if (unlikely(queue->starvation))
+		eth_ark_rx_seed_mbufs(queue);
+
 	prod_index = queue->prod_index;
 	cons_index = queue->cons_index;
 	if (prod_index == cons_index)
@@ -453,7 +457,7 @@ eth_ark_rx_stop_queue(struct rte_eth_dev *dev, uint16_t queue_id)
 static inline int
 eth_ark_rx_seed_mbufs(struct ark_rx_queue *queue)
 {
-	uint32_t limit = (queue->cons_index & ~(ARK_RX_MPU_CHUNK - 1)) +
+	uint32_t limit = RTE_ALIGN_FLOOR(queue->cons_index, ARK_RX_MPU_CHUNK) +
 		queue->queue_size;
 	uint32_t seed_index = queue->seed_index;
 
@@ -461,23 +465,32 @@ eth_ark_rx_seed_mbufs(struct ark_rx_queue *queue)
 	uint32_t seed_m = queue->seed_index & queue->queue_mask;
 
 	uint32_t nb = limit - seed_index;
+	int status;
 
 	/* Handle wrap around -- remainder is filled on the next call */
 	if (unlikely(seed_m + nb > queue->queue_size))
 		nb = queue->queue_size - seed_m;
 
 	struct rte_mbuf **mbufs = &queue->reserve_q[seed_m];
-	int status = rte_pktmbuf_alloc_bulk(queue->mb_pool, mbufs, nb);
+	do {
+		status = rte_pktmbuf_alloc_bulk(queue->mb_pool, mbufs, nb);
+		if (status == 0)
+			break;
+		/* Try again with a smaller request, keeping aligned with chunk size */
+		nb = RTE_ALIGN_FLOOR(nb / 2, ARK_RX_MPU_CHUNK);
+	} while (nb >= ARK_RX_MPU_CHUNK);
 
 	if (unlikely(status != 0)) {
-		ARK_PMD_LOG(NOTICE,
-			    "Could not allocate %u mbufs from pool"
-			    " for RX queue %u;"
-			    " %u free buffers remaining in queue\n",
-			    nb, queue->queue_index,
-			    queue->seed_index - queue->cons_index);
+		if (queue->starvation == 0) {
+			ARK_PMD_LOG(NOTICE,
+				    "Could not allocate %u mbufs from pool for RX queue %u; %u free buffers remaining\n",
+				    ARK_RX_MPU_CHUNK, queue->queue_index,
+				    queue->seed_index - queue->cons_index);
+			queue->starvation = 1;
+		}
 		return -1;
 	}
+	queue->starvation = 0;
 
 	if (ARK_DEBUG_CORE) {		/* DEBUG */
 		while (count != nb) {
-- 
2.34.1


  parent reply	other threads:[~2025-09-15 15:02 UTC|newest]

Thread overview: 16+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2025-09-03 21:28 [PATCH 1/4] net/ark: add PCIe IDS for newly supported devices Ed Czeck
2025-09-03 21:28 ` [PATCH 2/4] net/ark: remove double mbuf frees Ed Czeck
2025-09-03 21:28 ` [PATCH 3/4] net/ark: improve ring handling for segmented packets Ed Czeck
2025-09-03 21:28 ` [PATCH 4/4] net/ark: improve Rx queue recovery after mbuf exhaustion Ed Czeck
2025-09-06  4:51   ` Stephen Hemminger
2025-09-10  0:33   ` Stephen Hemminger
2025-09-10  0:33   ` Stephen Hemminger
2025-09-10 19:04 ` [PATCH v2 1/4] net/ark: add PCIe IDS for newly supported devices Ed Czeck
2025-09-10 19:04   ` [PATCH v2 2/4] net/ark: remove double mbuf frees Ed Czeck
2025-09-10 19:04   ` [PATCH v2 3/4] net/ark: improve ring handling for segmented packets Ed Czeck
2025-09-10 19:04   ` [PATCH v2 4/4] net/ark: improve Rx queue recovery after mbuf exhaustion Ed Czeck
2025-09-15 15:02 ` [PATCH v3 1/4] net/ark: add PCIe IDS for newly supported devices Ed Czeck
2025-09-15 15:02   ` [PATCH v3 2/4] net/ark: remove double mbuf frees Ed Czeck
2025-09-15 15:02   ` [PATCH v3 3/4] net/ark: improve ring handling for segmented packets Ed Czeck
2025-09-15 15:02   ` Ed Czeck [this message]
2025-09-15 15:36   ` [PATCH v3 1/4] net/ark: add PCIe IDS for newly supported devices Stephen Hemminger

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=20250915150204.1719628-4-ed.czeck@atomicrules.com \
    --to=ed.czeck@atomicrules.com \
    --cc=dev@dpdk.org \
    --cc=john.miller@atomicrules.com \
    --cc=shepard.siegel@atomicrules.com \
    --cc=stephen@networkplumber.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 an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.