Virtualization - Jan 2014 - [PATCH net-next v2 1/4] net: allow > 0 order atomic page alloc in skb_page_frag_refill

skb_page_frag_refill currently permits only order-0 page allocs
unless GFP_WAIT is used. Change skb_page_frag_refill to attempt
higher-order page allocations whether or not GFP_WAIT is used. If
memory cannot be allocated, the allocator will fall back to
successively smaller page allocs (down to order-0 page allocs).

This change brings skb_page_frag_refill in line with the existing
page allocation strategy employed by netdev_alloc_frag, which attempts
higher-order page allocations whether or not GFP_WAIT is set, falling
back to successively lower-order page allocations on failure. Part
of migration of virtio-net to per-receive queue page frag allocators.

Acked-by: Michael S. Tsirkin <mst at redhat.com>
Acked-by: Eric Dumazet <edumazet at google.com>
Signed-off-by: Michael Dalton <mwdalton at google.com>
---
 net/core/sock.c | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)

diff --git a/net/core/sock.c b/net/core/sock.c
index 5393b4b..a0d522a 100644
--- a/net/core/sock.c
+++ b/net/core/sock.c
@@ -1865,9 +1865,7 @@ bool skb_page_frag_refill(unsigned int sz, struct
page_frag *pfrag, gfp_t prio)
 		put_page(pfrag->page);
 	}
 
-	/* We restrict high order allocations to users that can afford to wait */
-	order = (prio & __GFP_WAIT) ? SKB_FRAG_PAGE_ORDER : 0;
-
+	order = SKB_FRAG_PAGE_ORDER;
 	do {
 		gfp_t gfp = prio;
 
-- 
1.8.5.1

The virtio-net driver currently uses netdev_alloc_frag() for GFP_ATOMIC
mergeable rx buffer allocations. This commit migrates virtio-net to use
per-receive queue page frags for GFP_ATOMIC allocation. This change unifies
mergeable rx buffer memory allocation, which now will use skb_refill_frag()
for both atomic and GFP-WAIT buffer allocations.

To address fragmentation concerns, if after buffer allocation there
is too little space left in the page frag to allocate a subsequent
buffer, the remaining space is added to the current allocated buffer
so that the remaining space can be used to store packet data.

Signed-off-by: Michael Dalton <mwdalton at google.com>
---
v2: Use GFP_COLD for RX buffer allocations (as in netdev_alloc_frag()).
    Remove per-netdev GFP_KERNEL page_frag allocator.

 drivers/net/virtio_net.c | 69 ++++++++++++++++++++++++------------------------
 1 file changed, 35 insertions(+), 34 deletions(-)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index c51a988..526dfd8 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c
@@ -78,6 +78,9 @@ struct receive_queue {
 	/* Chain pages by the private ptr. */
 	struct page *pages;
 
+	/* Page frag for packet buffer allocation. */
+	struct page_frag alloc_frag;
+
 	/* RX: fragments + linear part + virtio header */
 	struct scatterlist sg[MAX_SKB_FRAGS + 2];
 
@@ -126,11 +129,6 @@ struct virtnet_info {
 	/* Lock for config space updates */
 	struct mutex config_lock;
 
-	/* Page_frag for GFP_KERNEL packet buffer allocation when we run
-	 * low on memory.
-	 */
-	struct page_frag alloc_frag;
-
 	/* Does the affinity hint is set for virtqueues? */
 	bool affinity_hint_set;
 
@@ -336,8 +334,8 @@ static struct sk_buff *receive_mergeable(struct net_device
*dev,
 	int num_buf = hdr->mhdr.num_buffers;
 	struct page *page = virt_to_head_page(buf);
 	int offset = buf - page_address(page);
-	struct sk_buff *head_skb = page_to_skb(rq, page, offset, len,
-					       MERGE_BUFFER_LEN);
+	unsigned int truesize = max_t(unsigned int, len, MERGE_BUFFER_LEN);
+	struct sk_buff *head_skb = page_to_skb(rq, page, offset, len, truesize);
 	struct sk_buff *curr_skb = head_skb;
 
 	if (unlikely(!curr_skb))
@@ -353,11 +351,6 @@ static struct sk_buff *receive_mergeable(struct net_device
*dev,
 			dev->stats.rx_length_errors++;
 			goto err_buf;
 		}
-		if (unlikely(len > MERGE_BUFFER_LEN)) {
-			pr_debug("%s: rx error: merge buffer too long\n",
-				 dev->name);
-			len = MERGE_BUFFER_LEN;
-		}
 
 		page = virt_to_head_page(buf);
 		--rq->num;
@@ -376,19 +369,20 @@ static struct sk_buff *receive_mergeable(struct net_device
*dev,
 			head_skb->truesize += nskb->truesize;
 			num_skb_frags = 0;
 		}
+		truesize = max_t(unsigned int, len, MERGE_BUFFER_LEN);
 		if (curr_skb != head_skb) {
 			head_skb->data_len += len;
 			head_skb->len += len;
-			head_skb->truesize += MERGE_BUFFER_LEN;
+			head_skb->truesize += truesize;
 		}
 		offset = buf - page_address(page);
 		if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
 			put_page(page);
 			skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
-					     len, MERGE_BUFFER_LEN);
+					     len, truesize);
 		} else {
 			skb_add_rx_frag(curr_skb, num_skb_frags, page,
-					offset, len, MERGE_BUFFER_LEN);
+					offset, len, truesize);
 		}
 	}
 
@@ -578,25 +572,24 @@ static int add_recvbuf_big(struct receive_queue *rq, gfp_t
gfp)
 
 static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
 {
-	struct virtnet_info *vi = rq->vq->vdev->priv;
-	char *buf = NULL;
+	struct page_frag *alloc_frag = &rq->alloc_frag;
+	char *buf;
 	int err;
+	unsigned int len, hole;
 
-	if (gfp & __GFP_WAIT) {
-		if (skb_page_frag_refill(MERGE_BUFFER_LEN, &vi->alloc_frag,
-					 gfp)) {
-			buf = (char *)page_address(vi->alloc_frag.page) +
-			      vi->alloc_frag.offset;
-			get_page(vi->alloc_frag.page);
-			vi->alloc_frag.offset += MERGE_BUFFER_LEN;
-		}
-	} else {
-		buf = netdev_alloc_frag(MERGE_BUFFER_LEN);
-	}
-	if (!buf)
+	if (unlikely(!skb_page_frag_refill(MERGE_BUFFER_LEN, alloc_frag, gfp)))
 		return -ENOMEM;
+	buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
+	get_page(alloc_frag->page);
+	len = MERGE_BUFFER_LEN;
+	alloc_frag->offset += len;
+	hole = alloc_frag->size - alloc_frag->offset;
+	if (hole < MERGE_BUFFER_LEN) {
+		len += hole;
+		alloc_frag->offset += hole;
+	}
 
-	sg_init_one(rq->sg, buf, MERGE_BUFFER_LEN);
+	sg_init_one(rq->sg, buf, len);
 	err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, buf, gfp);
 	if (err < 0)
 		put_page(virt_to_head_page(buf));
@@ -617,6 +610,7 @@ static bool try_fill_recv(struct receive_queue *rq, gfp_t
gfp)
 	int err;
 	bool oom;
 
+	gfp |= __GFP_COLD;
 	do {
 		if (vi->mergeable_rx_bufs)
 			err = add_recvbuf_mergeable(rq, gfp);
@@ -1377,6 +1371,14 @@ static void free_receive_bufs(struct virtnet_info *vi)
 	}
 }
 
+static void free_receive_page_frags(struct virtnet_info *vi)
+{
+	int i;
+	for (i = 0; i < vi->max_queue_pairs; i++)
+		if (vi->rq[i].alloc_frag.page)
+			put_page(vi->rq[i].alloc_frag.page);
+}
+
 static void free_unused_bufs(struct virtnet_info *vi)
 {
 	void *buf;
@@ -1705,9 +1707,8 @@ free_recv_bufs:
 	unregister_netdev(dev);
 free_vqs:
 	cancel_delayed_work_sync(&vi->refill);
+	free_receive_page_frags(vi);
 	virtnet_del_vqs(vi);
-	if (vi->alloc_frag.page)
-		put_page(vi->alloc_frag.page);
 free_stats:
 	free_percpu(vi->stats);
 free:
@@ -1724,6 +1725,8 @@ static void remove_vq_common(struct virtnet_info *vi)
 
 	free_receive_bufs(vi);
 
+	free_receive_page_frags(vi);
+
 	virtnet_del_vqs(vi);
 }
 
@@ -1741,8 +1744,6 @@ static void virtnet_remove(struct virtio_device *vdev)
 	unregister_netdev(vi->dev);
 
 	remove_vq_common(vi);
-	if (vi->alloc_frag.page)
-		put_page(vi->alloc_frag.page);
 
 	flush_work(&vi->config_work);
 
-- 
1.8.5.1

Commit 2613af0ed18a ("virtio_net: migrate mergeable rx buffers to page frag
allocators") changed the mergeable receive buffer size from PAGE_SIZE to
MTU-size, introducing a single-stream regression for benchmarks with large
average packet size. There is no single optimal buffer size for all
workloads.  For workloads with packet size <= MTU bytes, MTU + virtio-net
header-sized buffers are preferred as larger buffers reduce the TCP window
due to SKB truesize. However, single-stream workloads with large average
packet sizes have higher throughput if larger (e.g., PAGE_SIZE) buffers
are used.

This commit auto-tunes the mergeable receiver buffer packet size by
choosing the packet buffer size based on an EWMA of the recent packet
sizes for the receive queue. Packet buffer sizes range from MTU_SIZE +
virtio-net header len to PAGE_SIZE. This improves throughput for
large packet workloads, as any workload with average packet size >PAGE_SIZE
will use PAGE_SIZE buffers.

These optimizations interact positively with recent commit
ba275241030c ("virtio-net: coalesce rx frags when possible during
rx"),
which coalesces adjacent RX SKB fragments in virtio_net. The coalescing
optimizations benefit buffers of any size.

Benchmarks taken from an average of 5 netperf 30-second TCP_STREAM runs
between two QEMU VMs on a single physical machine. Each VM has two VCPUs
with all offloads & vhost enabled. All VMs and vhost threads run in a
single 4 CPU cgroup cpuset, using cgroups to ensure that other processes
in the system will not be scheduled on the benchmark CPUs. Trunk includes
SKB rx frag coalescing.

net-next w/ virtio_net before 2613af0ed18a (PAGE_SIZE bufs): 14642.85Gb/s
net-next (MTU-size bufs):  13170.01Gb/s
net-next + auto-tune: 14555.94Gb/s

Jason Wang also reported a throughput increase on mlx4 from 22Gb/s
using MTU-sized buffers to about 26Gb/s using auto-tuning.

Signed-off-by: Michael Dalton <mwdalton at google.com>
---
v2: Add per-receive queue metadata ring to track precise truesize for
    mergeable receive buffers. Remove all truesize approximation. Never
    try to fill a full RX ring (required for metadata ring in v2).

 drivers/net/virtio_net.c | 145 ++++++++++++++++++++++++++++++++++-------------
 1 file changed, 107 insertions(+), 38 deletions(-)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index 526dfd8..f6e1ee0 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c
@@ -26,6 +26,7 @@
 #include <linux/if_vlan.h>
 #include <linux/slab.h>
 #include <linux/cpu.h>
+#include <linux/average.h>
 
 static int napi_weight = NAPI_POLL_WEIGHT;
 module_param(napi_weight, int, 0444);
@@ -36,11 +37,15 @@ module_param(gso, bool, 0444);
 
 /* FIXME: MTU in config. */
 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
-#define MERGE_BUFFER_LEN (ALIGN(GOOD_PACKET_LEN + \
-                                sizeof(struct virtio_net_hdr_mrg_rxbuf), \
-                                L1_CACHE_BYTES))
 #define GOOD_COPY_LEN	128
 
+/* Weight used for the RX packet size EWMA. The average packet size is used to
+ * determine the packet buffer size when refilling RX rings. As the entire RX
+ * ring may be refilled at once, the weight is chosen so that the EWMA will be
+ * insensitive to short-term, transient changes in packet size.
+ */
+#define RECEIVE_AVG_WEIGHT 64
+
 #define VIRTNET_DRIVER_VERSION "1.0.0"
 
 struct virtnet_stats {
@@ -65,11 +70,30 @@ struct send_queue {
 	char name[40];
 };
 
+/* Per-packet buffer context for mergeable receive buffers. */
+struct mergeable_receive_buf_ctx {
+	/* Packet buffer base address. */
+	void *buf;
+
+	/* Original size of the packet buffer for use in SKB truesize. Does not
+	 * include any padding space used to avoid internal fragmentation.
+	 */
+	unsigned int truesize;
+};
+
 /* Internal representation of a receive virtqueue */
 struct receive_queue {
 	/* Virtqueue associated with this receive_queue */
 	struct virtqueue *vq;
 
+	/* Circular buffer of mergeable rxbuf contexts. */
+	struct mergeable_receive_buf_ctx *mrg_buf_ctx;
+
+	/* Number of elements & head index of mrg_buf_ctx. Size must be
+	 * equal to the associated virtqueue's vring size.
+	 */
+	unsigned int mrg_buf_ctx_size, mrg_buf_ctx_head;
+
 	struct napi_struct napi;
 
 	/* Number of input buffers, and max we've ever had. */
@@ -78,6 +102,9 @@ struct receive_queue {
 	/* Chain pages by the private ptr. */
 	struct page *pages;
 
+	/* Average packet length for mergeable receive buffers. */
+	struct ewma mrg_avg_pkt_len;
+
 	/* Page frag for packet buffer allocation. */
 	struct page_frag alloc_frag;
 
@@ -327,32 +354,32 @@ err:
 
 static struct sk_buff *receive_mergeable(struct net_device *dev,
 					 struct receive_queue *rq,
-					 void *buf,
+					 struct mergeable_receive_buf_ctx *ctx,
 					 unsigned int len)
 {
-	struct skb_vnet_hdr *hdr = buf;
+	struct skb_vnet_hdr *hdr = ctx->buf;
 	int num_buf = hdr->mhdr.num_buffers;
-	struct page *page = virt_to_head_page(buf);
-	int offset = buf - page_address(page);
-	unsigned int truesize = max_t(unsigned int, len, MERGE_BUFFER_LEN);
+	struct page *page = virt_to_head_page(ctx->buf);
+	int offset = ctx->buf - page_address(page);
+	unsigned int truesize = max(len, ctx->truesize);
+
 	struct sk_buff *head_skb = page_to_skb(rq, page, offset, len, truesize);
 	struct sk_buff *curr_skb = head_skb;
 
 	if (unlikely(!curr_skb))
 		goto err_skb;
-
 	while (--num_buf) {
 		int num_skb_frags;
 
-		buf = virtqueue_get_buf(rq->vq, &len);
-		if (unlikely(!buf)) {
+		ctx = virtqueue_get_buf(rq->vq, &len);
+		if (unlikely(!ctx)) {
 			pr_debug("%s: rx error: %d buffers out of %d missing\n",
 				 dev->name, num_buf, hdr->mhdr.num_buffers);
 			dev->stats.rx_length_errors++;
 			goto err_buf;
 		}
 
-		page = virt_to_head_page(buf);
+		page = virt_to_head_page(ctx->buf);
 		--rq->num;
 
 		num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
@@ -369,13 +396,13 @@ static struct sk_buff *receive_mergeable(struct net_device
*dev,
 			head_skb->truesize += nskb->truesize;
 			num_skb_frags = 0;
 		}
-		truesize = max_t(unsigned int, len, MERGE_BUFFER_LEN);
+		truesize = max(len, ctx->truesize);
 		if (curr_skb != head_skb) {
 			head_skb->data_len += len;
 			head_skb->len += len;
 			head_skb->truesize += truesize;
 		}
-		offset = buf - page_address(page);
+		offset = ctx->buf - page_address(page);
 		if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
 			put_page(page);
 			skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
@@ -386,19 +413,20 @@ static struct sk_buff *receive_mergeable(struct net_device
*dev,
 		}
 	}
 
+	ewma_add(&rq->mrg_avg_pkt_len, head_skb->len);
 	return head_skb;
 
 err_skb:
 	put_page(page);
 	while (--num_buf) {
-		buf = virtqueue_get_buf(rq->vq, &len);
-		if (unlikely(!buf)) {
+		ctx = virtqueue_get_buf(rq->vq, &len);
+		if (unlikely(!ctx)) {
 			pr_debug("%s: rx error: %d buffers missing\n",
 				 dev->name, num_buf);
 			dev->stats.rx_length_errors++;
 			break;
 		}
-		page = virt_to_head_page(buf);
+		page = virt_to_head_page(ctx->buf);
 		put_page(page);
 		--rq->num;
 	}
@@ -419,12 +447,14 @@ static void receive_buf(struct receive_queue *rq, void
*buf, unsigned int len)
 	if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
 		pr_debug("%s: short packet %i\n", dev->name, len);
 		dev->stats.rx_length_errors++;
-		if (vi->mergeable_rx_bufs)
-			put_page(virt_to_head_page(buf));
-		else if (vi->big_packets)
+		if (vi->mergeable_rx_bufs) {
+			struct mergeable_receive_buf_ctx *ctx = buf;
+			put_page(virt_to_head_page(ctx->buf));
+		} else if (vi->big_packets) {
 			give_pages(rq, buf);
-		else
+		} else {
 			dev_kfree_skb(buf);
+		}
 		return;
 	}
 
@@ -572,29 +602,43 @@ static int add_recvbuf_big(struct receive_queue *rq, gfp_t
gfp)
 
 static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
 {
+	const unsigned int ring_size = rq->mrg_buf_ctx_size;
+	const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 	struct page_frag *alloc_frag = &rq->alloc_frag;
-	char *buf;
+	struct mergeable_receive_buf_ctx *ctx;
 	int err;
 	unsigned int len, hole;
 
-	if (unlikely(!skb_page_frag_refill(MERGE_BUFFER_LEN, alloc_frag, gfp)))
+	len = hdr_len + clamp_t(unsigned int, ewma_read(&rq->mrg_avg_pkt_len),
+				GOOD_PACKET_LEN, PAGE_SIZE - hdr_len);
+	len = ALIGN(len, L1_CACHE_BYTES);
+	if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
 		return -ENOMEM;
-	buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
+
+	ctx = &rq->mrg_buf_ctx[rq->mrg_buf_ctx_head];
+	ctx->buf = (char *)page_address(alloc_frag->page) +
alloc_frag->offset;
+	ctx->truesize = len;
 	get_page(alloc_frag->page);
-	len = MERGE_BUFFER_LEN;
 	alloc_frag->offset += len;
 	hole = alloc_frag->size - alloc_frag->offset;
-	if (hole < MERGE_BUFFER_LEN) {
+	if (hole < len) {
+		/* To avoid internal fragmentation, if there is very likely not
+		 * enough space for another buffer, add the remaining space to
+		 * the current buffer. This extra space is not included in
+		 * ctx->truesize.
+		 */
 		len += hole;
 		alloc_frag->offset += hole;
 	}
 
-	sg_init_one(rq->sg, buf, len);
-	err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, buf, gfp);
-	if (err < 0)
-		put_page(virt_to_head_page(buf));
-
-	return err;
+	sg_init_one(rq->sg, ctx->buf, len);
+	err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, ctx, gfp);
+	if (err < 0) {
+		put_page(virt_to_head_page(ctx->buf));
+		return err;
+	}
+	rq->mrg_buf_ctx_head = (rq->mrg_buf_ctx_head + 1) & (ring_size - 1);
+	return 0;
 }
 
 /*
@@ -610,6 +654,9 @@ static bool try_fill_recv(struct receive_queue *rq, gfp_t
gfp)
 	int err;
 	bool oom;
 
+	/* Do not attempt to add a buffer if the RX ring is full. */
+	if (unlikely(!rq->vq->num_free))
+		return true;
 	gfp |= __GFP_COLD;
 	do {
 		if (vi->mergeable_rx_bufs)
@@ -1354,8 +1401,10 @@ static void virtnet_free_queues(struct virtnet_info *vi)
 {
 	int i;
 
-	for (i = 0; i < vi->max_queue_pairs; i++)
+	for (i = 0; i < vi->max_queue_pairs; i++) {
 		netif_napi_del(&vi->rq[i].napi);
+		kfree(vi->rq[i].mrg_buf_ctx);
+	}
 
 	kfree(vi->rq);
 	kfree(vi->sq);
@@ -1394,12 +1443,14 @@ static void free_unused_bufs(struct virtnet_info *vi)
 		struct virtqueue *vq = vi->rq[i].vq;
 
 		while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
-			if (vi->mergeable_rx_bufs)
-				put_page(virt_to_head_page(buf));
-			else if (vi->big_packets)
+			if (vi->mergeable_rx_bufs) {
+				struct mergeable_receive_buf_ctx *ctx = buf;
+				put_page(virt_to_head_page(ctx->buf));
+			} else if (vi->big_packets) {
 				give_pages(&vi->rq[i], buf);
-			else
+			} else {
 				dev_kfree_skb(buf);
+			}
 			--vi->rq[i].num;
 		}
 		BUG_ON(vi->rq[i].num != 0);
@@ -1509,6 +1560,7 @@ static int virtnet_alloc_queues(struct virtnet_info *vi)
 			       napi_weight);
 
 		sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
+		ewma_init(&vi->rq[i].mrg_avg_pkt_len, 1, RECEIVE_AVG_WEIGHT);
 		sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
 	}
 
@@ -1522,7 +1574,8 @@ err_sq:
 
 static int init_vqs(struct virtnet_info *vi)
 {
-	int ret;
+	struct virtio_device *vdev = vi->vdev;
+	int i, ret;
 
 	/* Allocate send & receive queues */
 	ret = virtnet_alloc_queues(vi);
@@ -1533,12 +1586,28 @@ static int init_vqs(struct virtnet_info *vi)
 	if (ret)
 		goto err_free;
 
+	if (vi->mergeable_rx_bufs) {
+		for (i = 0; i < vi->max_queue_pairs; i++) {
+			struct receive_queue *rq = &vi->rq[i];
+			rq->mrg_buf_ctx_size = virtqueue_get_vring_size(rq->vq);
+			rq->mrg_buf_ctx = kmalloc(sizeof(*rq->mrg_buf_ctx) *
+						  rq->mrg_buf_ctx_size,
+						  GFP_KERNEL);
+			if (!rq->mrg_buf_ctx) {
+				ret = -ENOMEM;
+				goto err_del_vqs;
+			}
+		}
+	}
+
 	get_online_cpus();
 	virtnet_set_affinity(vi);
 	put_online_cpus();
 
 	return 0;
 
+err_del_vqs:
+	vdev->config->del_vqs(vdev);
 err_free:
 	virtnet_free_queues(vi);
 err:
-- 
1.8.5.1

Add initial support for debugfs to virtio-net. Each virtio-net network
device will have a directory under /virtio-net in debugfs. The
per-network device directory will contain one sub-directory per active,
enabled receive queue. If mergeable receive buffers are enabled, each
receive queue directory will contain a read-only file that returns the
current packet buffer size for the receive queue.

Signed-off-by: Michael Dalton <mwdalton at google.com>
---
 drivers/net/virtio_net.c | 314 ++++++++++++++++++++++++++++++++++++++++++++---
 1 file changed, 296 insertions(+), 18 deletions(-)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index f6e1ee0..5da18d6 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c
@@ -27,6 +27,9 @@
 #include <linux/slab.h>
 #include <linux/cpu.h>
 #include <linux/average.h>
+#include <linux/seqlock.h>
+#include <linux/kref.h>
+#include <linux/debugfs.h>
 
 static int napi_weight = NAPI_POLL_WEIGHT;
 module_param(napi_weight, int, 0444);
@@ -35,6 +38,9 @@ static bool csum = true, gso = true;
 module_param(csum, bool, 0444);
 module_param(gso, bool, 0444);
 
+/* Debugfs root directory for all virtio-net devices. */
+static struct dentry *virtnet_debugfs_root;
+
 /* FIXME: MTU in config. */
 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
 #define GOOD_COPY_LEN	128
@@ -102,9 +108,6 @@ struct receive_queue {
 	/* Chain pages by the private ptr. */
 	struct page *pages;
 
-	/* Average packet length for mergeable receive buffers. */
-	struct ewma mrg_avg_pkt_len;
-
 	/* Page frag for packet buffer allocation. */
 	struct page_frag alloc_frag;
 
@@ -115,6 +118,28 @@ struct receive_queue {
 	char name[40];
 };
 
+/* Per-receive queue statistics exported via debugfs. */
+struct receive_queue_stats {
+	/* Average packet length of receive queue (for mergeable rx buffers). */
+	struct ewma avg_pkt_len;
+
+	/* Per-receive queue stats debugfs directory. */
+	struct dentry *dbg;
+
+	/* Reference count for the receive queue statistics, needed because
+	 * an open debugfs file may outlive the receive queue and netdevice.
+	 * Open files will remain in-use until all outstanding file descriptors
+	 * are closed, even after the underlying file is unlinked.
+	 */
+	struct kref refcount;
+
+	/* Sequence counter to allow debugfs readers to safely access stats.
+	 * Assumes a single virtio-net writer, which is enforced by virtio-net
+	 * and NAPI.
+	 */
+	seqcount_t dbg_seq;
+};
+
 struct virtnet_info {
 	struct virtio_device *vdev;
 	struct virtqueue *cvq;
@@ -147,6 +172,15 @@ struct virtnet_info {
 	/* Active statistics */
 	struct virtnet_stats __percpu *stats;
 
+	/* Per-receive queue statstics exported via debugfs. Stored in
+	 * virtnet_info to survive freeze/restore -- a task may have a per-rq
+	 * debugfs file open at the time of freeze.
+	 */
+	struct receive_queue_stats **rq_stats;
+
+	/* Per-netdevice debugfs directory. */
+	struct dentry *dbg_dev_root;
+
 	/* Work struct for refilling if we run low on memory. */
 	struct delayed_work refill;
 
@@ -358,6 +392,8 @@ static struct sk_buff *receive_mergeable(struct net_device
*dev,
 					 unsigned int len)
 {
 	struct skb_vnet_hdr *hdr = ctx->buf;
+	struct virtnet_info *vi = netdev_priv(dev);
+	struct receive_queue_stats *rq_stats = vi->rq_stats[vq2rxq(rq->vq)];
 	int num_buf = hdr->mhdr.num_buffers;
 	struct page *page = virt_to_head_page(ctx->buf);
 	int offset = ctx->buf - page_address(page);
@@ -413,7 +449,9 @@ static struct sk_buff *receive_mergeable(struct net_device
*dev,
 		}
 	}
 
-	ewma_add(&rq->mrg_avg_pkt_len, head_skb->len);
+	write_seqcount_begin(&rq_stats->dbg_seq);
+	ewma_add(&rq_stats->avg_pkt_len, head_skb->len);
+	write_seqcount_end(&rq_stats->dbg_seq);
 	return head_skb;
 
 err_skb:
@@ -600,18 +638,30 @@ static int add_recvbuf_big(struct receive_queue *rq, gfp_t
gfp)
 	return err;
 }
 
+static unsigned int get_mergeable_buf_len(struct ewma *avg_pkt_len)
+{
+	const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
+	unsigned int len;
+
+	len = hdr_len + clamp_t(unsigned int, ewma_read(avg_pkt_len),
+				GOOD_PACKET_LEN, PAGE_SIZE - hdr_len);
+	return ALIGN(len, L1_CACHE_BYTES);
+}
+
 static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
 {
 	const unsigned int ring_size = rq->mrg_buf_ctx_size;
-	const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 	struct page_frag *alloc_frag = &rq->alloc_frag;
+	struct virtnet_info *vi = rq->vq->vdev->priv;
 	struct mergeable_receive_buf_ctx *ctx;
 	int err;
 	unsigned int len, hole;
 
-	len = hdr_len + clamp_t(unsigned int, ewma_read(&rq->mrg_avg_pkt_len),
-				GOOD_PACKET_LEN, PAGE_SIZE - hdr_len);
-	len = ALIGN(len, L1_CACHE_BYTES);
+	/* avg_pkt_len is written only in NAPI rx softirq context. We may
+	 * read avg_pkt_len without using the dbg_seq seqcount, as this code
+	 * is called only in NAPI rx softirq context or when NAPI is disabled.
+	 */
+	len =
get_mergeable_buf_len(&vi->rq_stats[vq2rxq(rq->vq)]->avg_pkt_len);
 	if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
 		return -ENOMEM;
 
@@ -1274,13 +1324,101 @@ static void virtnet_get_drvinfo(struct net_device *dev,
 
 }
 
+static ssize_t mergeable_rx_buffer_size_read(struct file *file,
+					     char __user *userbuf,
+					     size_t count,
+					     loff_t *ppos)
+{
+	struct receive_queue_stats *rq_stats = file->private_data;
+	char buf[32];
+	struct ewma avg;
+	unsigned int start, len;
+
+	/* Don't allow partial reads. */
+	if (*ppos)
+		return 0;
+	do {
+		start = read_seqcount_begin(&rq_stats->dbg_seq);
+		avg = rq_stats->avg_pkt_len;
+	} while (read_seqcount_retry(&rq_stats->dbg_seq, start));
+	len = scnprintf(buf, sizeof(buf), "%u\n",
get_mergeable_buf_len(&avg));
+	return simple_read_from_buffer(userbuf, count, ppos, buf, len);
+}
+
+void receive_queue_stats_free(struct kref *ref)
+{
+	struct receive_queue_stats *rq_stats;
+
+	rq_stats = container_of(ref, struct receive_queue_stats, refcount);
+	kfree(rq_stats);
+}
+
+static int receive_queue_stats_debugfs_open(struct inode *inode,
+					    struct file *file)
+{
+	struct receive_queue_stats *rq_stats = inode->i_private;
+	kref_get(&rq_stats->refcount);
+	file->private_data = rq_stats;
+	return 0;
+}
+
+static int receive_queue_stats_debugfs_release(struct inode *inode,
+					       struct file *file)
+{
+	struct receive_queue_stats *rq_stats = inode->i_private;
+	kref_put(&rq_stats->refcount, receive_queue_stats_free);
+	file->private_data = NULL;
+	return 0;
+}
+
+static const struct file_operations mergeable_rx_buffer_size_fops = {
+	.owner = THIS_MODULE,
+	.open = receive_queue_stats_debugfs_open,
+	.read = mergeable_rx_buffer_size_read,
+	.llseek = default_llseek,
+	.release = receive_queue_stats_debugfs_release,
+};
+
+static void receive_queue_debugfs_add(struct receive_queue *rq)
+{
+	struct virtnet_info *vi = rq->vq->vdev->priv;
+	unsigned int rq_index = vq2rxq(rq->vq);
+	struct receive_queue_stats *rq_stats = vi->rq_stats[rq_index];
+	struct dentry *dentry;
+	char name[32];
+
+	if (IS_ERR_OR_NULL(vi->dbg_dev_root))
+		return;
+	scnprintf(name, sizeof(name), "rx-%u", rq_index);
+	dentry = debugfs_create_dir(name, vi->dbg_dev_root);
+	if (IS_ERR_OR_NULL(dentry)) {
+		pr_warn("%s: could not create %s rx queue debugfs dir\n",
+			vi->dev->name, name);
+		return;
+	}
+	rq_stats->dbg = dentry;
+	if (vi->mergeable_rx_bufs)
+		debugfs_create_file("mergeable_rx_buffer_size", S_IRUSR,
+				rq_stats->dbg, rq_stats,
+				&mergeable_rx_buffer_size_fops);
+}
+
+static void receive_queue_debugfs_del(struct receive_queue *rq)
+{
+	struct virtnet_info *vi = rq->vq->vdev->priv;
+	struct receive_queue_stats *rq_stats = vi->rq_stats[vq2rxq(rq->vq)];
+	debugfs_remove_recursive(rq_stats->dbg);
+	rq_stats->dbg = NULL;
+}
+
 /* TODO: Eliminate OOO packets during switching */
 static int virtnet_set_channels(struct net_device *dev,
 				struct ethtool_channels *channels)
 {
 	struct virtnet_info *vi = netdev_priv(dev);
-	u16 queue_pairs = channels->combined_count;
-	int err;
+	u16 new_queue_pairs = channels->combined_count;
+	u16 old_queue_pairs = vi->curr_queue_pairs;
+	int err, i;
 
 	/* We don't support separate rx/tx channels.
 	 * We don't allow setting 'other' channels.
@@ -1288,14 +1426,21 @@ static int virtnet_set_channels(struct net_device *dev,
 	if (channels->rx_count || channels->tx_count ||
channels->other_count)
 		return -EINVAL;
 
-	if (queue_pairs > vi->max_queue_pairs)
+	if (new_queue_pairs > vi->max_queue_pairs)
 		return -EINVAL;
 
 	get_online_cpus();
-	err = virtnet_set_queues(vi, queue_pairs);
+	err = virtnet_set_queues(vi, new_queue_pairs);
 	if (!err) {
-		netif_set_real_num_tx_queues(dev, queue_pairs);
-		netif_set_real_num_rx_queues(dev, queue_pairs);
+		if (new_queue_pairs < old_queue_pairs) {
+			for (i = new_queue_pairs; i < old_queue_pairs; i++)
+				receive_queue_debugfs_del(&vi->rq[i]);
+		} else {
+			for (i = old_queue_pairs; i < new_queue_pairs; i++)
+				receive_queue_debugfs_add(&vi->rq[i]);
+		}
+		netif_set_real_num_tx_queues(dev, new_queue_pairs);
+		netif_set_real_num_rx_queues(dev, new_queue_pairs);
 
 		virtnet_set_affinity(vi);
 	}
@@ -1336,7 +1481,44 @@ static int virtnet_change_mtu(struct net_device *dev, int
new_mtu)
 	return 0;
 }
 
+/* Must be called only after the net_device name has been expanded. */
+static void virtnet_debugfs_init(struct virtnet_info *vi)
+{
+	int i;
+
+	if (IS_ERR_OR_NULL(virtnet_debugfs_root))
+		return;
+	vi->dbg_dev_root = debugfs_create_dir(vi->dev->name,
+					      virtnet_debugfs_root);
+	if (IS_ERR_OR_NULL(vi->dbg_dev_root)) {
+		pr_warn("%s: could not create netdevice debugfs dir\n",
+			vi->dev->name);
+		return;
+	}
+	for (i = 0; i < vi->curr_queue_pairs; i++)
+		receive_queue_debugfs_add(&vi->rq[i]);
+}
+
+static void virtnet_debugfs_cleanup(struct virtnet_info *vi)
+{
+	int i;
+
+	for (i = 0; i < vi->max_queue_pairs; i++)
+		receive_queue_debugfs_del(&vi->rq[i]);
+	debugfs_remove_recursive(vi->dbg_dev_root);
+	vi->dbg_dev_root = NULL;
+}
+
+static int virtnet_init(struct net_device *dev)
+{
+	struct virtnet_info *vi = netdev_priv(dev);
+
+	virtnet_debugfs_init(vi);
+	return 0;
+}
+
 static const struct net_device_ops virtnet_netdev = {
+	.ndo_init	     = virtnet_init,
 	.ndo_open            = virtnet_open,
 	.ndo_stop   	     = virtnet_close,
 	.ndo_start_xmit      = start_xmit,
@@ -1560,7 +1742,6 @@ static int virtnet_alloc_queues(struct virtnet_info *vi)
 			       napi_weight);
 
 		sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
-		ewma_init(&vi->rq[i].mrg_avg_pkt_len, 1, RECEIVE_AVG_WEIGHT);
 		sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
 	}
 
@@ -1614,6 +1795,39 @@ err:
 	return ret;
 }
 
+static int virtnet_rename(struct notifier_block *this,
+			  unsigned long event, void *ptr)
+{
+	struct net_device *dev = netdev_notifier_info_to_dev(ptr);
+	struct virtnet_info *vi;
+
+	if (event != NETDEV_CHANGENAME || dev->netdev_ops != &virtnet_netdev)
+		return NOTIFY_DONE;
+	vi = netdev_priv(dev);
+	if (IS_ERR_OR_NULL(vi->dbg_dev_root))
+		return NOTIFY_DONE;
+	if (IS_ERR_OR_NULL(debugfs_rename(virtnet_debugfs_root,
+					  vi->dbg_dev_root,
+					  virtnet_debugfs_root, dev->name))) {
+		pr_warn("%s: failed debugfs rename, removing old debugfs dir\n",
+			dev->name);
+		virtnet_debugfs_cleanup(vi);
+	}
+	return NOTIFY_DONE;
+}
+
+static void virtnet_release_receive_queue_stats(struct virtnet_info *vi)
+{
+	int i;
+
+	for (i = 0; i < vi->max_queue_pairs; i++) {
+		struct receive_queue_stats *rq_stats = vi->rq_stats[i];
+		if (rq_stats)
+			kref_put(&rq_stats->refcount, receive_queue_stats_free);
+	}
+	kfree(vi->rq_stats);
+}
+
 static int virtnet_probe(struct virtio_device *vdev)
 {
 	int i, err;
@@ -1723,10 +1937,24 @@ static int virtnet_probe(struct virtio_device *vdev)
 	vi->curr_queue_pairs = 1;
 	vi->max_queue_pairs = max_queue_pairs;
 
+	vi->rq_stats = kzalloc(sizeof(vi->rq_stats[0]) *
+			       vi->max_queue_pairs, GFP_KERNEL);
+	if (!vi->rq_stats)
+		goto free_dev_stats;
+	for (i = 0; i < vi->max_queue_pairs; i++) {
+		vi->rq_stats[i] = kzalloc(sizeof(*vi->rq_stats[0]), GFP_KERNEL);
+		if (!vi->rq_stats[i])
+			goto free_rq_stats;
+		seqcount_init(&vi->rq_stats[i]->dbg_seq);
+		kref_init(&vi->rq_stats[i]->refcount);
+		ewma_init(&vi->rq_stats[i]->avg_pkt_len, 1,
+			  RECEIVE_AVG_WEIGHT);
+	}
+
 	/* Allocate/initialize the rx/tx queues, and invoke find_vqs */
 	err = init_vqs(vi);
 	if (err)
-		goto free_stats;
+		goto free_rq_stats;
 
 	netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
 	netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
@@ -1777,8 +2005,11 @@ free_recv_bufs:
 free_vqs:
 	cancel_delayed_work_sync(&vi->refill);
 	free_receive_page_frags(vi);
+	virtnet_debugfs_cleanup(vi);
 	virtnet_del_vqs(vi);
-free_stats:
+free_rq_stats:
+	virtnet_release_receive_queue_stats(vi);
+free_dev_stats:
 	free_percpu(vi->stats);
 free:
 	free_netdev(dev);
@@ -1812,10 +2043,12 @@ static void virtnet_remove(struct virtio_device *vdev)
 
 	unregister_netdev(vi->dev);
 
+	virtnet_debugfs_cleanup(vi);
 	remove_vq_common(vi);
 
 	flush_work(&vi->config_work);
 
+	virtnet_release_receive_queue_stats(vi);
 	free_percpu(vi->stats);
 	free_netdev(vi->dev);
 }
@@ -1884,6 +2117,19 @@ static int virtnet_restore(struct virtio_device *vdev)
 }
 #endif
 
+static void virtnet_register_debugfs(void)
+{
+	virtnet_debugfs_root = debugfs_create_dir("virtio-net", NULL);
+	if (IS_ERR_OR_NULL(virtnet_debugfs_root))
+		pr_warn("Could not create virtio-net debugfs dir\n");
+}
+
+static void virtnet_unregister_debugfs(void)
+{
+	debugfs_remove_recursive(virtnet_debugfs_root);
+	virtnet_debugfs_root = NULL;
+}
+
 static struct virtio_device_id id_table[] = {
 	{ VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
 	{ 0 },
@@ -1917,7 +2163,39 @@ static struct virtio_driver virtio_net_driver = {
 #endif
 };
 
-module_virtio_driver(virtio_net_driver);
+static struct notifier_block virtnet_rename_notifier = {
+	.notifier_call = virtnet_rename,
+};
+
+static int __init init(void)
+{
+	int err;
+
+	virtnet_register_debugfs();
+	err = register_netdevice_notifier(&virtnet_rename_notifier);
+	if (err)
+		goto free_debugfs;
+	err = register_virtio_driver(&virtio_net_driver);
+	if (err)
+		goto free_notifier;
+	return 0;
+
+free_notifier:
+	unregister_netdevice_notifier(&virtnet_rename_notifier);
+free_debugfs:
+	virtnet_unregister_debugfs();
+	return err;
+}
+
+static void __exit cleanup(void)
+{
+	unregister_virtio_driver(&virtio_net_driver);
+	unregister_netdevice_notifier(&virtnet_rename_notifier);
+	virtnet_unregister_debugfs();
+}
+
+module_init(init);
+module_exit(cleanup);
 
 MODULE_DEVICE_TABLE(virtio, id_table);
 MODULE_DESCRIPTION("Virtio network driver");
-- 
1.8.5.1

On 01/07/2014 01:25 PM, Michael Dalton wrote:
> Commit 2613af0ed18a ("virtio_net: migrate mergeable rx buffers to page
frag
> allocators") changed the mergeable receive buffer size from PAGE_SIZE
to
> MTU-size, introducing a single-stream regression for benchmarks with large
> average packet size. There is no single optimal buffer size for all
> workloads.  For workloads with packet size <= MTU bytes, MTU +
virtio-net
> header-sized buffers are preferred as larger buffers reduce the TCP window
> due to SKB truesize. However, single-stream workloads with large average
> packet sizes have higher throughput if larger (e.g., PAGE_SIZE) buffers
> are used.
>
> This commit auto-tunes the mergeable receiver buffer packet size by
> choosing the packet buffer size based on an EWMA of the recent packet
> sizes for the receive queue. Packet buffer sizes range from MTU_SIZE +
> virtio-net header len to PAGE_SIZE. This improves throughput for
> large packet workloads, as any workload with average packet size >>
PAGE_SIZE will use PAGE_SIZE buffers.
>
> These optimizations interact positively with recent commit
> ba275241030c ("virtio-net: coalesce rx frags when possible during
rx"),
> which coalesces adjacent RX SKB fragments in virtio_net. The coalescing
> optimizations benefit buffers of any size.
>
> Benchmarks taken from an average of 5 netperf 30-second TCP_STREAM runs
> between two QEMU VMs on a single physical machine. Each VM has two VCPUs
> with all offloads & vhost enabled. All VMs and vhost threads run in a
> single 4 CPU cgroup cpuset, using cgroups to ensure that other processes
> in the system will not be scheduled on the benchmark CPUs. Trunk includes
> SKB rx frag coalescing.
>
> net-next w/ virtio_net before 2613af0ed18a (PAGE_SIZE bufs): 14642.85Gb/s
> net-next (MTU-size bufs):  13170.01Gb/s
> net-next + auto-tune: 14555.94Gb/s
>
> Jason Wang also reported a throughput increase on mlx4 from 22Gb/s
> using MTU-sized buffers to about 26Gb/s using auto-tuning.
>
> Signed-off-by: Michael Dalton <mwdalton at google.com>
> ---
> v2: Add per-receive queue metadata ring to track precise truesize for
>     mergeable receive buffers. Remove all truesize approximation. Never
>     try to fill a full RX ring (required for metadata ring in v2).
>
>  drivers/net/virtio_net.c | 145
++++++++++++++++++++++++++++++++++-------------
>  1 file changed, 107 insertions(+), 38 deletions(-)
>
> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> index 526dfd8..f6e1ee0 100644
> --- a/drivers/net/virtio_net.c
> +++ b/drivers/net/virtio_net.c
> @@ -26,6 +26,7 @@
>  #include <linux/if_vlan.h>
>  #include <linux/slab.h>
>  #include <linux/cpu.h>
> +#include <linux/average.h>
>  
>  static int napi_weight = NAPI_POLL_WEIGHT;
>  module_param(napi_weight, int, 0444);
> @@ -36,11 +37,15 @@ module_param(gso, bool, 0444);
>  
>  /* FIXME: MTU in config. */
>  #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
> -#define MERGE_BUFFER_LEN (ALIGN(GOOD_PACKET_LEN + \
> -                                sizeof(struct virtio_net_hdr_mrg_rxbuf), \
> -                                L1_CACHE_BYTES))
>  #define GOOD_COPY_LEN	128
>  
> +/* Weight used for the RX packet size EWMA. The average packet size is
used to
> + * determine the packet buffer size when refilling RX rings. As the entire
RX
> + * ring may be refilled at once, the weight is chosen so that the EWMA
will be
> + * insensitive to short-term, transient changes in packet size.
> + */
> +#define RECEIVE_AVG_WEIGHT 64
> +
>  #define VIRTNET_DRIVER_VERSION "1.0.0"
>  
>  struct virtnet_stats {
> @@ -65,11 +70,30 @@ struct send_queue {
>  	char name[40];
>  };
>  
> +/* Per-packet buffer context for mergeable receive buffers. */
> +struct mergeable_receive_buf_ctx {
> +	/* Packet buffer base address. */
> +	void *buf;
> +
> +	/* Original size of the packet buffer for use in SKB truesize. Does not
> +	 * include any padding space used to avoid internal fragmentation.
> +	 */
> +	unsigned int truesize;
> +};
> +
>  /* Internal representation of a receive virtqueue */
>  struct receive_queue {
>  	/* Virtqueue associated with this receive_queue */
>  	struct virtqueue *vq;
>  
> +	/* Circular buffer of mergeable rxbuf contexts. */
> +	struct mergeable_receive_buf_ctx *mrg_buf_ctx;
> +
> +	/* Number of elements & head index of mrg_buf_ctx. Size must be
> +	 * equal to the associated virtqueue's vring size.
> +	 */
> +	unsigned int mrg_buf_ctx_size, mrg_buf_ctx_head;
> +
>  	struct napi_struct napi;
>  
>  	/* Number of input buffers, and max we've ever had. */
> @@ -78,6 +102,9 @@ struct receive_queue {
>  	/* Chain pages by the private ptr. */
>  	struct page *pages;
>  
> +	/* Average packet length for mergeable receive buffers. */
> +	struct ewma mrg_avg_pkt_len;
> +
>  	/* Page frag for packet buffer allocation. */
>  	struct page_frag alloc_frag;
>  
> @@ -327,32 +354,32 @@ err:
>  
>  static struct sk_buff *receive_mergeable(struct net_device *dev,
>  					 struct receive_queue *rq,
> -					 void *buf,
> +					 struct mergeable_receive_buf_ctx *ctx,
>  					 unsigned int len)
>  {
> -	struct skb_vnet_hdr *hdr = buf;
> +	struct skb_vnet_hdr *hdr = ctx->buf;
>  	int num_buf = hdr->mhdr.num_buffers;
> -	struct page *page = virt_to_head_page(buf);
> -	int offset = buf - page_address(page);
> -	unsigned int truesize = max_t(unsigned int, len, MERGE_BUFFER_LEN);
> +	struct page *page = virt_to_head_page(ctx->buf);
> +	int offset = ctx->buf - page_address(page);
> +	unsigned int truesize = max(len, ctx->truesize);
This looks unnecessary, we've already had precise truesize for this
buffer. ctx->trusize should be always greater or equal to len
here.
> +
>  	struct sk_buff *head_skb = page_to_skb(rq, page, offset, len, truesize);
>  	struct sk_buff *curr_skb = head_skb;
>  
>  	if (unlikely(!curr_skb))
>  		goto err_skb;
> -
>  	while (--num_buf) {
>  		int num_skb_frags;
>  
> -		buf = virtqueue_get_buf(rq->vq, &len);
> -		if (unlikely(!buf)) {
> +		ctx = virtqueue_get_buf(rq->vq, &len);
> +		if (unlikely(!ctx)) {
>  			pr_debug("%s: rx error: %d buffers out of %d missing\n",
>  				 dev->name, num_buf, hdr->mhdr.num_buffers);
>  			dev->stats.rx_length_errors++;
>  			goto err_buf;
>  		}
>  
> -		page = virt_to_head_page(buf);
> +		page = virt_to_head_page(ctx->buf);
>  		--rq->num;
>  
>  		num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
> @@ -369,13 +396,13 @@ static struct sk_buff *receive_mergeable(struct
net_device *dev,
>  			head_skb->truesize += nskb->truesize;
>  			num_skb_frags = 0;
>  		}
> -		truesize = max_t(unsigned int, len, MERGE_BUFFER_LEN);
> +		truesize = max(len, ctx->truesize);
And this.
>  		if (curr_skb != head_skb) {
>  			head_skb->data_len += len;
>  			head_skb->len += len;
>  			head_skb->truesize += truesize;
>  		}
> -		offset = buf - page_address(page);
> +		offset = ctx->buf - page_address(page);
>  		if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
>  			put_page(page);
>  			skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
> @@ -386,19 +413,20 @@ static struct sk_buff *receive_mergeable(struct
net_device *dev,
>  		}
>  	}
>  
> +	ewma_add(&rq->mrg_avg_pkt_len, head_skb->len);
>  	return head_skb;
>  
>  err_skb:
>  	put_page(page);
>  	while (--num_buf) {
> -		buf = virtqueue_get_buf(rq->vq, &len);
> -		if (unlikely(!buf)) {
> +		ctx = virtqueue_get_buf(rq->vq, &len);
> +		if (unlikely(!ctx)) {
>  			pr_debug("%s: rx error: %d buffers missing\n",
>  				 dev->name, num_buf);
>  			dev->stats.rx_length_errors++;
>  			break;
>  		}
> -		page = virt_to_head_page(buf);
> +		page = virt_to_head_page(ctx->buf);
>  		put_page(page);
>  		--rq->num;
>  	}
> @@ -419,12 +447,14 @@ static void receive_buf(struct receive_queue *rq,
void *buf, unsigned int len)
>  	if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
>  		pr_debug("%s: short packet %i\n", dev->name, len);
>  		dev->stats.rx_length_errors++;
> -		if (vi->mergeable_rx_bufs)
> -			put_page(virt_to_head_page(buf));
> -		else if (vi->big_packets)
> +		if (vi->mergeable_rx_bufs) {
> +			struct mergeable_receive_buf_ctx *ctx = buf;
> +			put_page(virt_to_head_page(ctx->buf));
> +		} else if (vi->big_packets) {
>  			give_pages(rq, buf);
> -		else
> +		} else {
>  			dev_kfree_skb(buf);
> +		}
>  		return;
>  	}
>  
> @@ -572,29 +602,43 @@ static int add_recvbuf_big(struct receive_queue *rq,
gfp_t gfp)
>  
>  static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
>  {
> +	const unsigned int ring_size = rq->mrg_buf_ctx_size;
> +	const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
>  	struct page_frag *alloc_frag = &rq->alloc_frag;
> -	char *buf;
> +	struct mergeable_receive_buf_ctx *ctx;
>  	int err;
>  	unsigned int len, hole;
>  
> -	if (unlikely(!skb_page_frag_refill(MERGE_BUFFER_LEN, alloc_frag, gfp)))
> +	len = hdr_len + clamp_t(unsigned int,
ewma_read(&rq->mrg_avg_pkt_len),
> +				GOOD_PACKET_LEN, PAGE_SIZE - hdr_len);
> +	len = ALIGN(len, L1_CACHE_BYTES);
> +	if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
>  		return -ENOMEM;
> -	buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
> +
> +	ctx = &rq->mrg_buf_ctx[rq->mrg_buf_ctx_head];
> +	ctx->buf = (char *)page_address(alloc_frag->page) +
alloc_frag->offset;
> +	ctx->truesize = len;
>  	get_page(alloc_frag->page);
> -	len = MERGE_BUFFER_LEN;
>  	alloc_frag->offset += len;
>  	hole = alloc_frag->size - alloc_frag->offset;
> -	if (hole < MERGE_BUFFER_LEN) {
> +	if (hole < len) {
> +		/* To avoid internal fragmentation, if there is very likely not
> +		 * enough space for another buffer, add the remaining space to
> +		 * the current buffer. This extra space is not included in
> +		 * ctx->truesize.
> +		 */
What's the reason that this extra space is not accounted for
truesize?
>  		len += hole;
>  		alloc_frag->offset += hole;
>  	}
>  
> -	sg_init_one(rq->sg, buf, len);
> -	err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, buf, gfp);
> -	if (err < 0)
> -		put_page(virt_to_head_page(buf));
> -
> -	return err;
> +	sg_init_one(rq->sg, ctx->buf, len);
> +	err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, ctx, gfp);
> +	if (err < 0) {
> +		put_page(virt_to_head_page(ctx->buf));
> +		return err;
Should we also roll back the frag offset added above to avoid leaking
frags?
> +	}
> +	rq->mrg_buf_ctx_head = (rq->mrg_buf_ctx_head + 1) & (ring_size
- 1);
> +	return 0;
>  }
>  
>  /*
> @@ -610,6 +654,9 @@ static bool try_fill_recv(struct receive_queue *rq,
gfp_t gfp)
>  	int err;
>  	bool oom;
>  
> +	/* Do not attempt to add a buffer if the RX ring is full. */
> +	if (unlikely(!rq->vq->num_free))
> +		return true;
I haven't figured out why this is needed. It seems safe for
virtqueue_add_inbuf() just fail in add_recv_xx()?
>  	gfp |= __GFP_COLD;
>  	do {
>  		if (vi->mergeable_rx_bufs)
> @@ -1354,8 +1401,10 @@ static void virtnet_free_queues(struct virtnet_info
*vi)
>  {
>  	int i;
>  
> -	for (i = 0; i < vi->max_queue_pairs; i++)
> +	for (i = 0; i < vi->max_queue_pairs; i++) {
>  		netif_napi_del(&vi->rq[i].napi);
> +		kfree(vi->rq[i].mrg_buf_ctx);
> +	}
>  
>  	kfree(vi->rq);
>  	kfree(vi->sq);
> @@ -1394,12 +1443,14 @@ static void free_unused_bufs(struct virtnet_info
*vi)
>  		struct virtqueue *vq = vi->rq[i].vq;
>  
>  		while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
> -			if (vi->mergeable_rx_bufs)
> -				put_page(virt_to_head_page(buf));
> -			else if (vi->big_packets)
> +			if (vi->mergeable_rx_bufs) {
> +				struct mergeable_receive_buf_ctx *ctx = buf;
> +				put_page(virt_to_head_page(ctx->buf));
> +			} else if (vi->big_packets) {
>  				give_pages(&vi->rq[i], buf);
> -			else
> +			} else {
>  				dev_kfree_skb(buf);
> +			}
>  			--vi->rq[i].num;
>  		}
>  		BUG_ON(vi->rq[i].num != 0);
> @@ -1509,6 +1560,7 @@ static int virtnet_alloc_queues(struct virtnet_info
*vi)
>  			       napi_weight);
>  
>  		sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
> +		ewma_init(&vi->rq[i].mrg_avg_pkt_len, 1, RECEIVE_AVG_WEIGHT);
>  		sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
>  	}
>  
> @@ -1522,7 +1574,8 @@ err_sq:
>  
>  static int init_vqs(struct virtnet_info *vi)
>  {
> -	int ret;
> +	struct virtio_device *vdev = vi->vdev;
> +	int i, ret;
>  
>  	/* Allocate send & receive queues */
>  	ret = virtnet_alloc_queues(vi);
> @@ -1533,12 +1586,28 @@ static int init_vqs(struct virtnet_info *vi)
>  	if (ret)
>  		goto err_free;
>  
> +	if (vi->mergeable_rx_bufs) {
> +		for (i = 0; i < vi->max_queue_pairs; i++) {
> +			struct receive_queue *rq = &vi->rq[i];
> +			rq->mrg_buf_ctx_size = virtqueue_get_vring_size(rq->vq);
> +			rq->mrg_buf_ctx = kmalloc(sizeof(*rq->mrg_buf_ctx) *
> +						  rq->mrg_buf_ctx_size,
> +						  GFP_KERNEL);
> +			if (!rq->mrg_buf_ctx) {
> +				ret = -ENOMEM;
> +				goto err_del_vqs;
> +			}
> +		}
> +	}
> +
>  	get_online_cpus();
>  	virtnet_set_affinity(vi);
>  	put_online_cpus();
>  
>  	return 0;
>  
> +err_del_vqs:
> +	vdev->config->del_vqs(vdev);
>  err_free:
>  	virtnet_free_queues(vi);
>  err:

On 01/07/2014 01:25 PM, Michael Dalton wrote:
> Add initial support for debugfs to virtio-net. Each virtio-net network
> device will have a directory under /virtio-net in debugfs. The
> per-network device directory will contain one sub-directory per active,
> enabled receive queue. If mergeable receive buffers are enabled, each
> receive queue directory will contain a read-only file that returns the
> current packet buffer size for the receive queue.
>
> Signed-off-by: Michael Dalton <mwdalton at google.com>
This looks more complicated than expected. How about just adding an
entry in sysfs onto the existed network class device which looks more
simpler?

On Mon, Jan 06, 2014 at 09:25:52PM -0800, Michael Dalton
wrote:
> skb_page_frag_refill currently permits only order-0 page allocs
> unless GFP_WAIT is used. Change skb_page_frag_refill to attempt
> higher-order page allocations whether or not GFP_WAIT is used. If
> memory cannot be allocated, the allocator will fall back to
> successively smaller page allocs (down to order-0 page allocs).
> 
> This change brings skb_page_frag_refill in line with the existing
> page allocation strategy employed by netdev_alloc_frag, which attempts
> higher-order page allocations whether or not GFP_WAIT is set, falling
> back to successively lower-order page allocations on failure. Part
> of migration of virtio-net to per-receive queue page frag allocators.
> 
> Acked-by: Michael S. Tsirkin <mst at redhat.com>
> Acked-by: Eric Dumazet <edumazet at google.com>
> Signed-off-by: Michael Dalton <mwdalton at google.com>
> ---
>  net/core/sock.c | 4 +---
>  1 file changed, 1 insertion(+), 3 deletions(-)
> 
> diff --git a/net/core/sock.c b/net/core/sock.c
> index 5393b4b..a0d522a 100644
> --- a/net/core/sock.c
> +++ b/net/core/sock.c
> @@ -1865,9 +1865,7 @@ bool skb_page_frag_refill(unsigned int sz, struct
page_frag *pfrag, gfp_t prio)
>  		put_page(pfrag->page);
>  	}
>  
> -	/* We restrict high order allocations to users that can afford to wait */
> -	order = (prio & __GFP_WAIT) ? SKB_FRAG_PAGE_ORDER : 0;
> -
> +	order = SKB_FRAG_PAGE_ORDER;
>  	do {
>  		gfp_t gfp = prio;

Eric said we also need a patch to add __GFP_NORETRY, right?
Probably before this one in series.
> -- 
> 1.8.5.1

On Mon, Jan 06, 2014 at 09:25:55PM -0800, Michael Dalton
wrote:
> Add initial support for debugfs to virtio-net. Each virtio-net network
> device will have a directory under /virtio-net in debugfs. The
> per-network device directory will contain one sub-directory per active,
> enabled receive queue. If mergeable receive buffers are enabled, each
> receive queue directory will contain a read-only file that returns the
> current packet buffer size for the receive queue.
> 
> Signed-off-by: Michael Dalton <mwdalton at google.com>
thanks, I'll play with it.
Could you tell us meanwhile, what's the typical size that you see?
> ---
>  drivers/net/virtio_net.c | 314
++++++++++++++++++++++++++++++++++++++++++++---
>  1 file changed, 296 insertions(+), 18 deletions(-)
> 
> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> index f6e1ee0..5da18d6 100644
> --- a/drivers/net/virtio_net.c
> +++ b/drivers/net/virtio_net.c
> @@ -27,6 +27,9 @@
>  #include <linux/slab.h>
>  #include <linux/cpu.h>
>  #include <linux/average.h>
> +#include <linux/seqlock.h>
> +#include <linux/kref.h>
> +#include <linux/debugfs.h>
>  
>  static int napi_weight = NAPI_POLL_WEIGHT;
>  module_param(napi_weight, int, 0444);
> @@ -35,6 +38,9 @@ static bool csum = true, gso = true;
>  module_param(csum, bool, 0444);
>  module_param(gso, bool, 0444);
>  
> +/* Debugfs root directory for all virtio-net devices. */
> +static struct dentry *virtnet_debugfs_root;
> +
>  /* FIXME: MTU in config. */
>  #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
>  #define GOOD_COPY_LEN	128
> @@ -102,9 +108,6 @@ struct receive_queue {
>  	/* Chain pages by the private ptr. */
>  	struct page *pages;
>  
> -	/* Average packet length for mergeable receive buffers. */
> -	struct ewma mrg_avg_pkt_len;
> -
>  	/* Page frag for packet buffer allocation. */
>  	struct page_frag alloc_frag;
>  
> @@ -115,6 +118,28 @@ struct receive_queue {
>  	char name[40];
>  };
>  
> +/* Per-receive queue statistics exported via debugfs. */
> +struct receive_queue_stats {
> +	/* Average packet length of receive queue (for mergeable rx buffers). */
> +	struct ewma avg_pkt_len;
> +
> +	/* Per-receive queue stats debugfs directory. */
> +	struct dentry *dbg;
> +
> +	/* Reference count for the receive queue statistics, needed because
> +	 * an open debugfs file may outlive the receive queue and netdevice.
> +	 * Open files will remain in-use until all outstanding file descriptors
> +	 * are closed, even after the underlying file is unlinked.
> +	 */
> +	struct kref refcount;
> +
> +	/* Sequence counter to allow debugfs readers to safely access stats.
> +	 * Assumes a single virtio-net writer, which is enforced by virtio-net
> +	 * and NAPI.
> +	 */
> +	seqcount_t dbg_seq;
> +};
> +
>  struct virtnet_info {
>  	struct virtio_device *vdev;
>  	struct virtqueue *cvq;
> @@ -147,6 +172,15 @@ struct virtnet_info {
>  	/* Active statistics */
>  	struct virtnet_stats __percpu *stats;
>  
> +	/* Per-receive queue statstics exported via debugfs. Stored in
> +	 * virtnet_info to survive freeze/restore -- a task may have a per-rq
> +	 * debugfs file open at the time of freeze.
> +	 */
> +	struct receive_queue_stats **rq_stats;
> +
> +	/* Per-netdevice debugfs directory. */
> +	struct dentry *dbg_dev_root;
> +
>  	/* Work struct for refilling if we run low on memory. */
>  	struct delayed_work refill;
>  
> @@ -358,6 +392,8 @@ static struct sk_buff *receive_mergeable(struct
net_device *dev,
>  					 unsigned int len)
>  {
>  	struct skb_vnet_hdr *hdr = ctx->buf;
> +	struct virtnet_info *vi = netdev_priv(dev);
> +	struct receive_queue_stats *rq_stats =
vi->rq_stats[vq2rxq(rq->vq)];
>  	int num_buf = hdr->mhdr.num_buffers;
>  	struct page *page = virt_to_head_page(ctx->buf);
>  	int offset = ctx->buf - page_address(page);
> @@ -413,7 +449,9 @@ static struct sk_buff *receive_mergeable(struct
net_device *dev,
>  		}
>  	}
>  
> -	ewma_add(&rq->mrg_avg_pkt_len, head_skb->len);
> +	write_seqcount_begin(&rq_stats->dbg_seq);
> +	ewma_add(&rq_stats->avg_pkt_len, head_skb->len);
> +	write_seqcount_end(&rq_stats->dbg_seq);
>  	return head_skb;
>  
>  err_skb:
> @@ -600,18 +638,30 @@ static int add_recvbuf_big(struct receive_queue *rq,
gfp_t gfp)
>  	return err;
>  }
>  
> +static unsigned int get_mergeable_buf_len(struct ewma *avg_pkt_len)
> +{
> +	const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
> +	unsigned int len;
> +
> +	len = hdr_len + clamp_t(unsigned int, ewma_read(avg_pkt_len),
> +				GOOD_PACKET_LEN, PAGE_SIZE - hdr_len);
> +	return ALIGN(len, L1_CACHE_BYTES);
> +}
> +
>  static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
>  {
>  	const unsigned int ring_size = rq->mrg_buf_ctx_size;
> -	const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
>  	struct page_frag *alloc_frag = &rq->alloc_frag;
> +	struct virtnet_info *vi = rq->vq->vdev->priv;
>  	struct mergeable_receive_buf_ctx *ctx;
>  	int err;
>  	unsigned int len, hole;
>  
> -	len = hdr_len + clamp_t(unsigned int,
ewma_read(&rq->mrg_avg_pkt_len),
> -				GOOD_PACKET_LEN, PAGE_SIZE - hdr_len);
> -	len = ALIGN(len, L1_CACHE_BYTES);
> +	/* avg_pkt_len is written only in NAPI rx softirq context. We may
> +	 * read avg_pkt_len without using the dbg_seq seqcount, as this code
> +	 * is called only in NAPI rx softirq context or when NAPI is disabled.
> +	 */
> +	len =
get_mergeable_buf_len(&vi->rq_stats[vq2rxq(rq->vq)]->avg_pkt_len);
>  	if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
>  		return -ENOMEM;
>  
> @@ -1274,13 +1324,101 @@ static void virtnet_get_drvinfo(struct net_device
*dev,
>  
>  }
>  
> +static ssize_t mergeable_rx_buffer_size_read(struct file *file,
> +					     char __user *userbuf,
> +					     size_t count,
> +					     loff_t *ppos)
> +{
> +	struct receive_queue_stats *rq_stats = file->private_data;
> +	char buf[32];
> +	struct ewma avg;
> +	unsigned int start, len;
> +
> +	/* Don't allow partial reads. */
> +	if (*ppos)
> +		return 0;
> +	do {
> +		start = read_seqcount_begin(&rq_stats->dbg_seq);
> +		avg = rq_stats->avg_pkt_len;
> +	} while (read_seqcount_retry(&rq_stats->dbg_seq, start));
> +	len = scnprintf(buf, sizeof(buf), "%u\n",
get_mergeable_buf_len(&avg));
> +	return simple_read_from_buffer(userbuf, count, ppos, buf, len);
> +}
> +
> +void receive_queue_stats_free(struct kref *ref)
> +{
> +	struct receive_queue_stats *rq_stats;
> +
> +	rq_stats = container_of(ref, struct receive_queue_stats, refcount);
> +	kfree(rq_stats);
> +}
> +
> +static int receive_queue_stats_debugfs_open(struct inode *inode,
> +					    struct file *file)
> +{
> +	struct receive_queue_stats *rq_stats = inode->i_private;
> +	kref_get(&rq_stats->refcount);
> +	file->private_data = rq_stats;
> +	return 0;
> +}
> +
> +static int receive_queue_stats_debugfs_release(struct inode *inode,
> +					       struct file *file)
> +{
> +	struct receive_queue_stats *rq_stats = inode->i_private;
> +	kref_put(&rq_stats->refcount, receive_queue_stats_free);
> +	file->private_data = NULL;
> +	return 0;
> +}
> +
> +static const struct file_operations mergeable_rx_buffer_size_fops = {
> +	.owner = THIS_MODULE,
> +	.open = receive_queue_stats_debugfs_open,
> +	.read = mergeable_rx_buffer_size_read,
> +	.llseek = default_llseek,
> +	.release = receive_queue_stats_debugfs_release,
> +};
> +
> +static void receive_queue_debugfs_add(struct receive_queue *rq)
> +{
> +	struct virtnet_info *vi = rq->vq->vdev->priv;
> +	unsigned int rq_index = vq2rxq(rq->vq);
> +	struct receive_queue_stats *rq_stats = vi->rq_stats[rq_index];
> +	struct dentry *dentry;
> +	char name[32];
> +
> +	if (IS_ERR_OR_NULL(vi->dbg_dev_root))
> +		return;
> +	scnprintf(name, sizeof(name), "rx-%u", rq_index);
> +	dentry = debugfs_create_dir(name, vi->dbg_dev_root);
> +	if (IS_ERR_OR_NULL(dentry)) {
> +		pr_warn("%s: could not create %s rx queue debugfs dir\n",
> +			vi->dev->name, name);
> +		return;
> +	}
> +	rq_stats->dbg = dentry;
> +	if (vi->mergeable_rx_bufs)
> +		debugfs_create_file("mergeable_rx_buffer_size", S_IRUSR,
> +				rq_stats->dbg, rq_stats,
> +				&mergeable_rx_buffer_size_fops);
> +}
> +
> +static void receive_queue_debugfs_del(struct receive_queue *rq)
> +{
> +	struct virtnet_info *vi = rq->vq->vdev->priv;
> +	struct receive_queue_stats *rq_stats =
vi->rq_stats[vq2rxq(rq->vq)];
> +	debugfs_remove_recursive(rq_stats->dbg);
> +	rq_stats->dbg = NULL;
> +}
> +
>  /* TODO: Eliminate OOO packets during switching */
>  static int virtnet_set_channels(struct net_device *dev,
>  				struct ethtool_channels *channels)
>  {
>  	struct virtnet_info *vi = netdev_priv(dev);
> -	u16 queue_pairs = channels->combined_count;
> -	int err;
> +	u16 new_queue_pairs = channels->combined_count;
> +	u16 old_queue_pairs = vi->curr_queue_pairs;
> +	int err, i;
>  
>  	/* We don't support separate rx/tx channels.
>  	 * We don't allow setting 'other' channels.
> @@ -1288,14 +1426,21 @@ static int virtnet_set_channels(struct net_device
*dev,
>  	if (channels->rx_count || channels->tx_count ||
channels->other_count)
>  		return -EINVAL;
>  
> -	if (queue_pairs > vi->max_queue_pairs)
> +	if (new_queue_pairs > vi->max_queue_pairs)
>  		return -EINVAL;
>  
>  	get_online_cpus();
> -	err = virtnet_set_queues(vi, queue_pairs);
> +	err = virtnet_set_queues(vi, new_queue_pairs);
>  	if (!err) {
> -		netif_set_real_num_tx_queues(dev, queue_pairs);
> -		netif_set_real_num_rx_queues(dev, queue_pairs);
> +		if (new_queue_pairs < old_queue_pairs) {
> +			for (i = new_queue_pairs; i < old_queue_pairs; i++)
> +				receive_queue_debugfs_del(&vi->rq[i]);
> +		} else {
> +			for (i = old_queue_pairs; i < new_queue_pairs; i++)
> +				receive_queue_debugfs_add(&vi->rq[i]);
> +		}
> +		netif_set_real_num_tx_queues(dev, new_queue_pairs);
> +		netif_set_real_num_rx_queues(dev, new_queue_pairs);
>  
>  		virtnet_set_affinity(vi);
>  	}
> @@ -1336,7 +1481,44 @@ static int virtnet_change_mtu(struct net_device
*dev, int new_mtu)
>  	return 0;
>  }
>  
> +/* Must be called only after the net_device name has been expanded. */
> +static void virtnet_debugfs_init(struct virtnet_info *vi)
> +{
> +	int i;
> +
> +	if (IS_ERR_OR_NULL(virtnet_debugfs_root))
> +		return;
> +	vi->dbg_dev_root = debugfs_create_dir(vi->dev->name,
> +					      virtnet_debugfs_root);
> +	if (IS_ERR_OR_NULL(vi->dbg_dev_root)) {
> +		pr_warn("%s: could not create netdevice debugfs dir\n",
> +			vi->dev->name);
> +		return;
> +	}
> +	for (i = 0; i < vi->curr_queue_pairs; i++)
> +		receive_queue_debugfs_add(&vi->rq[i]);
> +}
> +
> +static void virtnet_debugfs_cleanup(struct virtnet_info *vi)
> +{
> +	int i;
> +
> +	for (i = 0; i < vi->max_queue_pairs; i++)
> +		receive_queue_debugfs_del(&vi->rq[i]);
> +	debugfs_remove_recursive(vi->dbg_dev_root);
> +	vi->dbg_dev_root = NULL;
> +}
> +
> +static int virtnet_init(struct net_device *dev)
> +{
> +	struct virtnet_info *vi = netdev_priv(dev);
> +
> +	virtnet_debugfs_init(vi);
> +	return 0;
> +}
> +
>  static const struct net_device_ops virtnet_netdev = {
> +	.ndo_init	     = virtnet_init,
>  	.ndo_open            = virtnet_open,
>  	.ndo_stop   	     = virtnet_close,
>  	.ndo_start_xmit      = start_xmit,
> @@ -1560,7 +1742,6 @@ static int virtnet_alloc_queues(struct virtnet_info
*vi)
>  			       napi_weight);
>  
>  		sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
> -		ewma_init(&vi->rq[i].mrg_avg_pkt_len, 1, RECEIVE_AVG_WEIGHT);
>  		sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
>  	}
>  
> @@ -1614,6 +1795,39 @@ err:
>  	return ret;
>  }
>  
> +static int virtnet_rename(struct notifier_block *this,
> +			  unsigned long event, void *ptr)
> +{
> +	struct net_device *dev = netdev_notifier_info_to_dev(ptr);
> +	struct virtnet_info *vi;
> +
> +	if (event != NETDEV_CHANGENAME || dev->netdev_ops !=
&virtnet_netdev)
> +		return NOTIFY_DONE;
> +	vi = netdev_priv(dev);
> +	if (IS_ERR_OR_NULL(vi->dbg_dev_root))
> +		return NOTIFY_DONE;
> +	if (IS_ERR_OR_NULL(debugfs_rename(virtnet_debugfs_root,
> +					  vi->dbg_dev_root,
> +					  virtnet_debugfs_root, dev->name))) {
> +		pr_warn("%s: failed debugfs rename, removing old debugfs
dir\n",
> +			dev->name);
> +		virtnet_debugfs_cleanup(vi);
> +	}
> +	return NOTIFY_DONE;
> +}
> +
> +static void virtnet_release_receive_queue_stats(struct virtnet_info *vi)
> +{
> +	int i;
> +
> +	for (i = 0; i < vi->max_queue_pairs; i++) {
> +		struct receive_queue_stats *rq_stats = vi->rq_stats[i];
> +		if (rq_stats)
> +			kref_put(&rq_stats->refcount, receive_queue_stats_free);
> +	}
> +	kfree(vi->rq_stats);
> +}
> +
>  static int virtnet_probe(struct virtio_device *vdev)
>  {
>  	int i, err;
> @@ -1723,10 +1937,24 @@ static int virtnet_probe(struct virtio_device
*vdev)
>  	vi->curr_queue_pairs = 1;
>  	vi->max_queue_pairs = max_queue_pairs;
>  
> +	vi->rq_stats = kzalloc(sizeof(vi->rq_stats[0]) *
> +			       vi->max_queue_pairs, GFP_KERNEL);
> +	if (!vi->rq_stats)
> +		goto free_dev_stats;
> +	for (i = 0; i < vi->max_queue_pairs; i++) {
> +		vi->rq_stats[i] = kzalloc(sizeof(*vi->rq_stats[0]), GFP_KERNEL);
> +		if (!vi->rq_stats[i])
> +			goto free_rq_stats;
> +		seqcount_init(&vi->rq_stats[i]->dbg_seq);
> +		kref_init(&vi->rq_stats[i]->refcount);
> +		ewma_init(&vi->rq_stats[i]->avg_pkt_len, 1,
> +			  RECEIVE_AVG_WEIGHT);
> +	}
> +
>  	/* Allocate/initialize the rx/tx queues, and invoke find_vqs */
>  	err = init_vqs(vi);
>  	if (err)
> -		goto free_stats;
> +		goto free_rq_stats;
>  
>  	netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
>  	netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
> @@ -1777,8 +2005,11 @@ free_recv_bufs:
>  free_vqs:
>  	cancel_delayed_work_sync(&vi->refill);
>  	free_receive_page_frags(vi);
> +	virtnet_debugfs_cleanup(vi);
>  	virtnet_del_vqs(vi);
> -free_stats:
> +free_rq_stats:
> +	virtnet_release_receive_queue_stats(vi);
> +free_dev_stats:
>  	free_percpu(vi->stats);
>  free:
>  	free_netdev(dev);
> @@ -1812,10 +2043,12 @@ static void virtnet_remove(struct virtio_device
*vdev)
>  
>  	unregister_netdev(vi->dev);
>  
> +	virtnet_debugfs_cleanup(vi);
>  	remove_vq_common(vi);
>  
>  	flush_work(&vi->config_work);
>  
> +	virtnet_release_receive_queue_stats(vi);
>  	free_percpu(vi->stats);
>  	free_netdev(vi->dev);
>  }
> @@ -1884,6 +2117,19 @@ static int virtnet_restore(struct virtio_device
*vdev)
>  }
>  #endif
>  
> +static void virtnet_register_debugfs(void)
> +{
> +	virtnet_debugfs_root = debugfs_create_dir("virtio-net", NULL);
> +	if (IS_ERR_OR_NULL(virtnet_debugfs_root))
> +		pr_warn("Could not create virtio-net debugfs dir\n");
> +}
> +
> +static void virtnet_unregister_debugfs(void)
> +{
> +	debugfs_remove_recursive(virtnet_debugfs_root);
> +	virtnet_debugfs_root = NULL;
> +}
> +
>  static struct virtio_device_id id_table[] = {
>  	{ VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
>  	{ 0 },
> @@ -1917,7 +2163,39 @@ static struct virtio_driver virtio_net_driver = {
>  #endif
>  };
>  
> -module_virtio_driver(virtio_net_driver);
> +static struct notifier_block virtnet_rename_notifier = {
> +	.notifier_call = virtnet_rename,
> +};
> +
> +static int __init init(void)
> +{
> +	int err;
> +
> +	virtnet_register_debugfs();
> +	err = register_netdevice_notifier(&virtnet_rename_notifier);
> +	if (err)
> +		goto free_debugfs;
> +	err = register_virtio_driver(&virtio_net_driver);
> +	if (err)
> +		goto free_notifier;
> +	return 0;
> +
> +free_notifier:
> +	unregister_netdevice_notifier(&virtnet_rename_notifier);
> +free_debugfs:
> +	virtnet_unregister_debugfs();
> +	return err;
> +}
> +
> +static void __exit cleanup(void)
> +{
> +	unregister_virtio_driver(&virtio_net_driver);
> +	unregister_netdevice_notifier(&virtnet_rename_notifier);
> +	virtnet_unregister_debugfs();
> +}
> +
> +module_init(init);
> +module_exit(cleanup);
>  
>  MODULE_DEVICE_TABLE(virtio, id_table);
>  MODULE_DESCRIPTION("Virtio network driver");
> -- 
> 1.8.5.1

On Wed, 2014-01-08 at 20:08 +0200, Michael S. Tsirkin wrote:
> Eric said we also need a patch to add __GFP_NORETRY, right?
> Probably before this one in series.
Nope, this __GFP_NORETRY has nothing to do with this.

I am not yet convinced we want it.

This needs mm guys advice, as its a tradeoff for mm layer more than
networking...

On Mon, Jan 06, 2014 at 09:25:54PM -0800, Michael Dalton
wrote:
> Commit 2613af0ed18a ("virtio_net: migrate mergeable rx buffers to page
frag
> allocators") changed the mergeable receive buffer size from PAGE_SIZE
to
> MTU-size, introducing a single-stream regression for benchmarks with large
> average packet size. There is no single optimal buffer size for all
> workloads.  For workloads with packet size <= MTU bytes, MTU +
virtio-net
> header-sized buffers are preferred as larger buffers reduce the TCP window
> due to SKB truesize. However, single-stream workloads with large average
> packet sizes have higher throughput if larger (e.g., PAGE_SIZE) buffers
> are used.
> 
> This commit auto-tunes the mergeable receiver buffer packet size by
> choosing the packet buffer size based on an EWMA of the recent packet
> sizes for the receive queue. Packet buffer sizes range from MTU_SIZE +
> virtio-net header len to PAGE_SIZE. This improves throughput for
> large packet workloads, as any workload with average packet size >>
PAGE_SIZE will use PAGE_SIZE buffers.
> 
> These optimizations interact positively with recent commit
> ba275241030c ("virtio-net: coalesce rx frags when possible during
rx"),
> which coalesces adjacent RX SKB fragments in virtio_net. The coalescing
> optimizations benefit buffers of any size.
> 
> Benchmarks taken from an average of 5 netperf 30-second TCP_STREAM runs
> between two QEMU VMs on a single physical machine. Each VM has two VCPUs
> with all offloads & vhost enabled. All VMs and vhost threads run in a
> single 4 CPU cgroup cpuset, using cgroups to ensure that other processes
> in the system will not be scheduled on the benchmark CPUs. Trunk includes
> SKB rx frag coalescing.
> 
> net-next w/ virtio_net before 2613af0ed18a (PAGE_SIZE bufs): 14642.85Gb/s
> net-next (MTU-size bufs):  13170.01Gb/s
> net-next + auto-tune: 14555.94Gb/s
> 
> Jason Wang also reported a throughput increase on mlx4 from 22Gb/s
> using MTU-sized buffers to about 26Gb/s using auto-tuning.
> 
> Signed-off-by: Michael Dalton <mwdalton at google.com>
I like where this series is going.

There are a couple of minor comments by Jason worth addressing, I'm guessing
you are going to post v3 anyway?
> ---
> v2: Add per-receive queue metadata ring to track precise truesize for
>     mergeable receive buffers. Remove all truesize approximation. Never
>     try to fill a full RX ring (required for metadata ring in v2).
> 
>  drivers/net/virtio_net.c | 145
++++++++++++++++++++++++++++++++++-------------
>  1 file changed, 107 insertions(+), 38 deletions(-)
> 
> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> index 526dfd8..f6e1ee0 100644
> --- a/drivers/net/virtio_net.c
> +++ b/drivers/net/virtio_net.c
> @@ -26,6 +26,7 @@
>  #include <linux/if_vlan.h>
>  #include <linux/slab.h>
>  #include <linux/cpu.h>
> +#include <linux/average.h>
>  
>  static int napi_weight = NAPI_POLL_WEIGHT;
>  module_param(napi_weight, int, 0444);
> @@ -36,11 +37,15 @@ module_param(gso, bool, 0444);
>  
>  /* FIXME: MTU in config. */
>  #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
> -#define MERGE_BUFFER_LEN (ALIGN(GOOD_PACKET_LEN + \
> -                                sizeof(struct virtio_net_hdr_mrg_rxbuf), \
> -                                L1_CACHE_BYTES))
>  #define GOOD_COPY_LEN	128
>  
> +/* Weight used for the RX packet size EWMA. The average packet size is
used to
> + * determine the packet buffer size when refilling RX rings. As the entire
RX
> + * ring may be refilled at once, the weight is chosen so that the EWMA
will be
> + * insensitive to short-term, transient changes in packet size.
> + */
> +#define RECEIVE_AVG_WEIGHT 64
> +
>  #define VIRTNET_DRIVER_VERSION "1.0.0"
>  
>  struct virtnet_stats {
> @@ -65,11 +70,30 @@ struct send_queue {
>  	char name[40];
>  };
>  
> +/* Per-packet buffer context for mergeable receive buffers. */
> +struct mergeable_receive_buf_ctx {
> +	/* Packet buffer base address. */
> +	void *buf;
> +
> +	/* Original size of the packet buffer for use in SKB truesize. Does not
> +	 * include any padding space used to avoid internal fragmentation.
> +	 */
> +	unsigned int truesize;
> +};
> +
>  /* Internal representation of a receive virtqueue */
>  struct receive_queue {
>  	/* Virtqueue associated with this receive_queue */
>  	struct virtqueue *vq;
>  
> +	/* Circular buffer of mergeable rxbuf contexts. */
> +	struct mergeable_receive_buf_ctx *mrg_buf_ctx;
> +
> +	/* Number of elements & head index of mrg_buf_ctx. Size must be
> +	 * equal to the associated virtqueue's vring size.
> +	 */
> +	unsigned int mrg_buf_ctx_size, mrg_buf_ctx_head;
> +
>  	struct napi_struct napi;
>  
>  	/* Number of input buffers, and max we've ever had. */
> @@ -78,6 +102,9 @@ struct receive_queue {
>  	/* Chain pages by the private ptr. */
>  	struct page *pages;
>  
> +	/* Average packet length for mergeable receive buffers. */
> +	struct ewma mrg_avg_pkt_len;
> +
>  	/* Page frag for packet buffer allocation. */
>  	struct page_frag alloc_frag;
>  
> @@ -327,32 +354,32 @@ err:
>  
>  static struct sk_buff *receive_mergeable(struct net_device *dev,
>  					 struct receive_queue *rq,
> -					 void *buf,
> +					 struct mergeable_receive_buf_ctx *ctx,
>  					 unsigned int len)
>  {
> -	struct skb_vnet_hdr *hdr = buf;
> +	struct skb_vnet_hdr *hdr = ctx->buf;
>  	int num_buf = hdr->mhdr.num_buffers;
> -	struct page *page = virt_to_head_page(buf);
> -	int offset = buf - page_address(page);
> -	unsigned int truesize = max_t(unsigned int, len, MERGE_BUFFER_LEN);
> +	struct page *page = virt_to_head_page(ctx->buf);
> +	int offset = ctx->buf - page_address(page);
> +	unsigned int truesize = max(len, ctx->truesize);
> +
>  	struct sk_buff *head_skb = page_to_skb(rq, page, offset, len, truesize);
>  	struct sk_buff *curr_skb = head_skb;
>  
>  	if (unlikely(!curr_skb))
>  		goto err_skb;
> -
>  	while (--num_buf) {
>  		int num_skb_frags;
>  
> -		buf = virtqueue_get_buf(rq->vq, &len);
> -		if (unlikely(!buf)) {
> +		ctx = virtqueue_get_buf(rq->vq, &len);
> +		if (unlikely(!ctx)) {
>  			pr_debug("%s: rx error: %d buffers out of %d missing\n",
>  				 dev->name, num_buf, hdr->mhdr.num_buffers);
>  			dev->stats.rx_length_errors++;
>  			goto err_buf;
>  		}
>  
> -		page = virt_to_head_page(buf);
> +		page = virt_to_head_page(ctx->buf);
>  		--rq->num;
>  
>  		num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
> @@ -369,13 +396,13 @@ static struct sk_buff *receive_mergeable(struct
net_device *dev,
>  			head_skb->truesize += nskb->truesize;
>  			num_skb_frags = 0;
>  		}
> -		truesize = max_t(unsigned int, len, MERGE_BUFFER_LEN);
> +		truesize = max(len, ctx->truesize);
>  		if (curr_skb != head_skb) {
>  			head_skb->data_len += len;
>  			head_skb->len += len;
>  			head_skb->truesize += truesize;
>  		}
> -		offset = buf - page_address(page);
> +		offset = ctx->buf - page_address(page);
>  		if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
>  			put_page(page);
>  			skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
> @@ -386,19 +413,20 @@ static struct sk_buff *receive_mergeable(struct
net_device *dev,
>  		}
>  	}
>  
> +	ewma_add(&rq->mrg_avg_pkt_len, head_skb->len);
>  	return head_skb;
>  
>  err_skb:
>  	put_page(page);
>  	while (--num_buf) {
> -		buf = virtqueue_get_buf(rq->vq, &len);
> -		if (unlikely(!buf)) {
> +		ctx = virtqueue_get_buf(rq->vq, &len);
> +		if (unlikely(!ctx)) {
>  			pr_debug("%s: rx error: %d buffers missing\n",
>  				 dev->name, num_buf);
>  			dev->stats.rx_length_errors++;
>  			break;
>  		}
> -		page = virt_to_head_page(buf);
> +		page = virt_to_head_page(ctx->buf);
>  		put_page(page);
>  		--rq->num;
>  	}
> @@ -419,12 +447,14 @@ static void receive_buf(struct receive_queue *rq,
void *buf, unsigned int len)
>  	if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
>  		pr_debug("%s: short packet %i\n", dev->name, len);
>  		dev->stats.rx_length_errors++;
> -		if (vi->mergeable_rx_bufs)
> -			put_page(virt_to_head_page(buf));
> -		else if (vi->big_packets)
> +		if (vi->mergeable_rx_bufs) {
> +			struct mergeable_receive_buf_ctx *ctx = buf;
> +			put_page(virt_to_head_page(ctx->buf));
> +		} else if (vi->big_packets) {
>  			give_pages(rq, buf);
> -		else
> +		} else {
>  			dev_kfree_skb(buf);
> +		}
>  		return;
>  	}
>  
> @@ -572,29 +602,43 @@ static int add_recvbuf_big(struct receive_queue *rq,
gfp_t gfp)
>  
>  static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
>  {
> +	const unsigned int ring_size = rq->mrg_buf_ctx_size;
> +	const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
>  	struct page_frag *alloc_frag = &rq->alloc_frag;
> -	char *buf;
> +	struct mergeable_receive_buf_ctx *ctx;
>  	int err;
>  	unsigned int len, hole;
>  
> -	if (unlikely(!skb_page_frag_refill(MERGE_BUFFER_LEN, alloc_frag, gfp)))
> +	len = hdr_len + clamp_t(unsigned int,
ewma_read(&rq->mrg_avg_pkt_len),
> +				GOOD_PACKET_LEN, PAGE_SIZE - hdr_len);
> +	len = ALIGN(len, L1_CACHE_BYTES);
> +	if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
>  		return -ENOMEM;
> -	buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
> +
> +	ctx = &rq->mrg_buf_ctx[rq->mrg_buf_ctx_head];
> +	ctx->buf = (char *)page_address(alloc_frag->page) +
alloc_frag->offset;
> +	ctx->truesize = len;
>  	get_page(alloc_frag->page);
> -	len = MERGE_BUFFER_LEN;
>  	alloc_frag->offset += len;
>  	hole = alloc_frag->size - alloc_frag->offset;
> -	if (hole < MERGE_BUFFER_LEN) {
> +	if (hole < len) {
> +		/* To avoid internal fragmentation, if there is very likely not
> +		 * enough space for another buffer, add the remaining space to
> +		 * the current buffer. This extra space is not included in
> +		 * ctx->truesize.
> +		 */
>  		len += hole;
>  		alloc_frag->offset += hole;
>  	}
>  
> -	sg_init_one(rq->sg, buf, len);
> -	err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, buf, gfp);
> -	if (err < 0)
> -		put_page(virt_to_head_page(buf));
> -
> -	return err;
> +	sg_init_one(rq->sg, ctx->buf, len);
> +	err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, ctx, gfp);
> +	if (err < 0) {
> +		put_page(virt_to_head_page(ctx->buf));
> +		return err;
> +	}
> +	rq->mrg_buf_ctx_head = (rq->mrg_buf_ctx_head + 1) & (ring_size
- 1);
> +	return 0;
>  }
>  
>  /*
> @@ -610,6 +654,9 @@ static bool try_fill_recv(struct receive_queue *rq,
gfp_t gfp)
>  	int err;
>  	bool oom;
>  
> +	/* Do not attempt to add a buffer if the RX ring is full. */
> +	if (unlikely(!rq->vq->num_free))
> +		return true;
>  	gfp |= __GFP_COLD;
>  	do {
>  		if (vi->mergeable_rx_bufs)
> @@ -1354,8 +1401,10 @@ static void virtnet_free_queues(struct virtnet_info
*vi)
>  {
>  	int i;
>  
> -	for (i = 0; i < vi->max_queue_pairs; i++)
> +	for (i = 0; i < vi->max_queue_pairs; i++) {
>  		netif_napi_del(&vi->rq[i].napi);
> +		kfree(vi->rq[i].mrg_buf_ctx);
> +	}
>  
>  	kfree(vi->rq);
>  	kfree(vi->sq);
> @@ -1394,12 +1443,14 @@ static void free_unused_bufs(struct virtnet_info
*vi)
>  		struct virtqueue *vq = vi->rq[i].vq;
>  
>  		while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
> -			if (vi->mergeable_rx_bufs)
> -				put_page(virt_to_head_page(buf));
> -			else if (vi->big_packets)
> +			if (vi->mergeable_rx_bufs) {
> +				struct mergeable_receive_buf_ctx *ctx = buf;
> +				put_page(virt_to_head_page(ctx->buf));
> +			} else if (vi->big_packets) {
>  				give_pages(&vi->rq[i], buf);
> -			else
> +			} else {
>  				dev_kfree_skb(buf);
> +			}
>  			--vi->rq[i].num;
>  		}
>  		BUG_ON(vi->rq[i].num != 0);
> @@ -1509,6 +1560,7 @@ static int virtnet_alloc_queues(struct virtnet_info
*vi)
>  			       napi_weight);
>  
>  		sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
> +		ewma_init(&vi->rq[i].mrg_avg_pkt_len, 1, RECEIVE_AVG_WEIGHT);
>  		sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
>  	}
>  
> @@ -1522,7 +1574,8 @@ err_sq:
>  
>  static int init_vqs(struct virtnet_info *vi)
>  {
> -	int ret;
> +	struct virtio_device *vdev = vi->vdev;
> +	int i, ret;
>  
>  	/* Allocate send & receive queues */
>  	ret = virtnet_alloc_queues(vi);
> @@ -1533,12 +1586,28 @@ static int init_vqs(struct virtnet_info *vi)
>  	if (ret)
>  		goto err_free;
>  
> +	if (vi->mergeable_rx_bufs) {
> +		for (i = 0; i < vi->max_queue_pairs; i++) {
> +			struct receive_queue *rq = &vi->rq[i];
> +			rq->mrg_buf_ctx_size = virtqueue_get_vring_size(rq->vq);
> +			rq->mrg_buf_ctx = kmalloc(sizeof(*rq->mrg_buf_ctx) *
> +						  rq->mrg_buf_ctx_size,
> +						  GFP_KERNEL);
> +			if (!rq->mrg_buf_ctx) {
> +				ret = -ENOMEM;
> +				goto err_del_vqs;
> +			}
> +		}
> +	}
> +
>  	get_online_cpus();
>  	virtnet_set_affinity(vi);
>  	put_online_cpus();
>  
>  	return 0;
>  
> +err_del_vqs:
> +	vdev->config->del_vqs(vdev);
>  err_free:
>  	virtnet_free_queues(vi);
>  err:
> -- 
> 1.8.5.1

On Mon, Jan 06, 2014 at 09:25:54PM -0800, Michael Dalton
wrote:
> Commit 2613af0ed18a ("virtio_net: migrate mergeable rx buffers to page
frag
> allocators") changed the mergeable receive buffer size from PAGE_SIZE
to
> MTU-size, introducing a single-stream regression for benchmarks with large
> average packet size. There is no single optimal buffer size for all
> workloads.  For workloads with packet size <= MTU bytes, MTU +
virtio-net
> header-sized buffers are preferred as larger buffers reduce the TCP window
> due to SKB truesize. However, single-stream workloads with large average
> packet sizes have higher throughput if larger (e.g., PAGE_SIZE) buffers
> are used.
> 
> This commit auto-tunes the mergeable receiver buffer packet size by
> choosing the packet buffer size based on an EWMA of the recent packet
> sizes for the receive queue. Packet buffer sizes range from MTU_SIZE +
> virtio-net header len to PAGE_SIZE. This improves throughput for
> large packet workloads, as any workload with average packet size >>
PAGE_SIZE will use PAGE_SIZE buffers.
> 
> These optimizations interact positively with recent commit
> ba275241030c ("virtio-net: coalesce rx frags when possible during
rx"),
> which coalesces adjacent RX SKB fragments in virtio_net. The coalescing
> optimizations benefit buffers of any size.
> 
> Benchmarks taken from an average of 5 netperf 30-second TCP_STREAM runs
> between two QEMU VMs on a single physical machine. Each VM has two VCPUs
> with all offloads & vhost enabled. All VMs and vhost threads run in a
> single 4 CPU cgroup cpuset, using cgroups to ensure that other processes
> in the system will not be scheduled on the benchmark CPUs. Trunk includes
> SKB rx frag coalescing.
> 
> net-next w/ virtio_net before 2613af0ed18a (PAGE_SIZE bufs): 14642.85Gb/s
> net-next (MTU-size bufs):  13170.01Gb/s
> net-next + auto-tune: 14555.94Gb/s
> 
> Jason Wang also reported a throughput increase on mlx4 from 22Gb/s
> using MTU-sized buffers to about 26Gb/s using auto-tuning.
> 
> Signed-off-by: Michael Dalton <mwdalton at google.com>

Sorry that I didn't notice early, but there seems to be a bug here.
See below.
Also, I think we can simplify code, see suggestion below.
> ---
> v2: Add per-receive queue metadata ring to track precise truesize for
>     mergeable receive buffers. Remove all truesize approximation. Never
>     try to fill a full RX ring (required for metadata ring in v2).
> 
>  drivers/net/virtio_net.c | 145
++++++++++++++++++++++++++++++++++-------------
>  1 file changed, 107 insertions(+), 38 deletions(-)
> 
> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> index 526dfd8..f6e1ee0 100644
> --- a/drivers/net/virtio_net.c
> +++ b/drivers/net/virtio_net.c
> @@ -26,6 +26,7 @@
>  #include <linux/if_vlan.h>
>  #include <linux/slab.h>
>  #include <linux/cpu.h>
> +#include <linux/average.h>
>  
>  static int napi_weight = NAPI_POLL_WEIGHT;
>  module_param(napi_weight, int, 0444);
> @@ -36,11 +37,15 @@ module_param(gso, bool, 0444);
>  
>  /* FIXME: MTU in config. */
>  #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
> -#define MERGE_BUFFER_LEN (ALIGN(GOOD_PACKET_LEN + \
> -                                sizeof(struct virtio_net_hdr_mrg_rxbuf), \
> -                                L1_CACHE_BYTES))
>  #define GOOD_COPY_LEN	128
>  
> +/* Weight used for the RX packet size EWMA. The average packet size is
used to
> + * determine the packet buffer size when refilling RX rings. As the entire
RX
> + * ring may be refilled at once, the weight is chosen so that the EWMA
will be
> + * insensitive to short-term, transient changes in packet size.
> + */
> +#define RECEIVE_AVG_WEIGHT 64
> +
>  #define VIRTNET_DRIVER_VERSION "1.0.0"
>  
>  struct virtnet_stats {
> @@ -65,11 +70,30 @@ struct send_queue {
>  	char name[40];
>  };
>  
> +/* Per-packet buffer context for mergeable receive buffers. */
> +struct mergeable_receive_buf_ctx {
> +	/* Packet buffer base address. */
> +	void *buf;
> +
> +	/* Original size of the packet buffer for use in SKB truesize. Does not
> +	 * include any padding space used to avoid internal fragmentation.
> +	 */
> +	unsigned int truesize;
Don't need full int really, it's up to 4K/cache line size,
1 byte would be enough, maximum 2 ...
So if all we want is extra 1-2 bytes per buffer, we don't really
need this extra level of indirection I think.
We can just allocate them before the header together with an skb.


> +};
> +
>  /* Internal representation of a receive virtqueue */
>  struct receive_queue {
>  	/* Virtqueue associated with this receive_queue */
>  	struct virtqueue *vq;
>  
> +	/* Circular buffer of mergeable rxbuf contexts. */
> +	struct mergeable_receive_buf_ctx *mrg_buf_ctx;
> +
> +	/* Number of elements & head index of mrg_buf_ctx. Size must be
> +	 * equal to the associated virtqueue's vring size.
> +	 */
> +	unsigned int mrg_buf_ctx_size, mrg_buf_ctx_head;
> +
>  	struct napi_struct napi;
>  
>  	/* Number of input buffers, and max we've ever had. */
> @@ -78,6 +102,9 @@ struct receive_queue {
>  	/* Chain pages by the private ptr. */
>  	struct page *pages;
>  
> +	/* Average packet length for mergeable receive buffers. */
> +	struct ewma mrg_avg_pkt_len;
> +
>  	/* Page frag for packet buffer allocation. */
>  	struct page_frag alloc_frag;
>  
> @@ -327,32 +354,32 @@ err:
>  
>  static struct sk_buff *receive_mergeable(struct net_device *dev,
>  					 struct receive_queue *rq,
> -					 void *buf,
> +					 struct mergeable_receive_buf_ctx *ctx,
>  					 unsigned int len)
>  {
> -	struct skb_vnet_hdr *hdr = buf;
> +	struct skb_vnet_hdr *hdr = ctx->buf;
>  	int num_buf = hdr->mhdr.num_buffers;
> -	struct page *page = virt_to_head_page(buf);
> -	int offset = buf - page_address(page);
> -	unsigned int truesize = max_t(unsigned int, len, MERGE_BUFFER_LEN);
> +	struct page *page = virt_to_head_page(ctx->buf);
> +	int offset = ctx->buf - page_address(page);
> +	unsigned int truesize = max(len, ctx->truesize);
> +
>  	struct sk_buff *head_skb = page_to_skb(rq, page, offset, len, truesize);
>  	struct sk_buff *curr_skb = head_skb;
>  
>  	if (unlikely(!curr_skb))
>  		goto err_skb;
> -
>  	while (--num_buf) {
>  		int num_skb_frags;
>  
> -		buf = virtqueue_get_buf(rq->vq, &len);
> -		if (unlikely(!buf)) {
> +		ctx = virtqueue_get_buf(rq->vq, &len);
> +		if (unlikely(!ctx)) {
>  			pr_debug("%s: rx error: %d buffers out of %d missing\n",
>  				 dev->name, num_buf, hdr->mhdr.num_buffers);
>  			dev->stats.rx_length_errors++;
>  			goto err_buf;
>  		}
>  
> -		page = virt_to_head_page(buf);
> +		page = virt_to_head_page(ctx->buf);
>  		--rq->num;
>  
>  		num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
> @@ -369,13 +396,13 @@ static struct sk_buff *receive_mergeable(struct
net_device *dev,
>  			head_skb->truesize += nskb->truesize;
>  			num_skb_frags = 0;
>  		}
> -		truesize = max_t(unsigned int, len, MERGE_BUFFER_LEN);
> +		truesize = max(len, ctx->truesize);
>  		if (curr_skb != head_skb) {
>  			head_skb->data_len += len;
>  			head_skb->len += len;
>  			head_skb->truesize += truesize;
>  		}
> -		offset = buf - page_address(page);
> +		offset = ctx->buf - page_address(page);
>  		if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
>  			put_page(page);
>  			skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
> @@ -386,19 +413,20 @@ static struct sk_buff *receive_mergeable(struct
net_device *dev,
>  		}
>  	}
>  
> +	ewma_add(&rq->mrg_avg_pkt_len, head_skb->len);
>  	return head_skb;
>  
>  err_skb:
>  	put_page(page);
>  	while (--num_buf) {
> -		buf = virtqueue_get_buf(rq->vq, &len);
> -		if (unlikely(!buf)) {
> +		ctx = virtqueue_get_buf(rq->vq, &len);
> +		if (unlikely(!ctx)) {
>  			pr_debug("%s: rx error: %d buffers missing\n",
>  				 dev->name, num_buf);
>  			dev->stats.rx_length_errors++;
>  			break;
>  		}
> -		page = virt_to_head_page(buf);
> +		page = virt_to_head_page(ctx->buf);
>  		put_page(page);
>  		--rq->num;
>  	}
> @@ -419,12 +447,14 @@ static void receive_buf(struct receive_queue *rq,
void *buf, unsigned int len)
>  	if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
>  		pr_debug("%s: short packet %i\n", dev->name, len);
>  		dev->stats.rx_length_errors++;
> -		if (vi->mergeable_rx_bufs)
> -			put_page(virt_to_head_page(buf));
> -		else if (vi->big_packets)
> +		if (vi->mergeable_rx_bufs) {
> +			struct mergeable_receive_buf_ctx *ctx = buf;
> +			put_page(virt_to_head_page(ctx->buf));
> +		} else if (vi->big_packets) {
>  			give_pages(rq, buf);
> -		else
> +		} else {
>  			dev_kfree_skb(buf);
> +		}
>  		return;
>  	}
>  
> @@ -572,29 +602,43 @@ static int add_recvbuf_big(struct receive_queue *rq,
gfp_t gfp)
>  
>  static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
>  {
> +	const unsigned int ring_size = rq->mrg_buf_ctx_size;
> +	const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
>  	struct page_frag *alloc_frag = &rq->alloc_frag;
> -	char *buf;
> +	struct mergeable_receive_buf_ctx *ctx;
>  	int err;
>  	unsigned int len, hole;
>  
> -	if (unlikely(!skb_page_frag_refill(MERGE_BUFFER_LEN, alloc_frag, gfp)))
> +	len = hdr_len + clamp_t(unsigned int,
ewma_read(&rq->mrg_avg_pkt_len),
> +				GOOD_PACKET_LEN, PAGE_SIZE - hdr_len);
> +	len = ALIGN(len, L1_CACHE_BYTES);
> +	if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
>  		return -ENOMEM;
> -	buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
> +
> +	ctx = &rq->mrg_buf_ctx[rq->mrg_buf_ctx_head];
> +	ctx->buf = (char *)page_address(alloc_frag->page) +
alloc_frag->offset;
> +	ctx->truesize = len;
>  	get_page(alloc_frag->page);
> -	len = MERGE_BUFFER_LEN;
>  	alloc_frag->offset += len;
>  	hole = alloc_frag->size - alloc_frag->offset;
> -	if (hole < MERGE_BUFFER_LEN) {
> +	if (hole < len) {
> +		/* To avoid internal fragmentation, if there is very likely not
> +		 * enough space for another buffer, add the remaining space to
> +		 * the current buffer. This extra space is not included in
> +		 * ctx->truesize.
> +		 */
>  		len += hole;
>  		alloc_frag->offset += hole;
>  	}
>  
> -	sg_init_one(rq->sg, buf, len);
> -	err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, buf, gfp);
> -	if (err < 0)
> -		put_page(virt_to_head_page(buf));
> -
> -	return err;
> +	sg_init_one(rq->sg, ctx->buf, len);
> +	err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, ctx, gfp);
> +	if (err < 0) {
> +		put_page(virt_to_head_page(ctx->buf));
> +		return err;
> +	}
> +	rq->mrg_buf_ctx_head = (rq->mrg_buf_ctx_head + 1) & (ring_size
- 1);
Wait a second. this assumes that buffers are consumes in order?
This happens to be the case but is not guaranteed by the spec
at all.
> +	return 0;
>  }
>  
>  /*
> @@ -610,6 +654,9 @@ static bool try_fill_recv(struct receive_queue *rq,
gfp_t gfp)
>  	int err;
>  	bool oom;
>  
> +	/* Do not attempt to add a buffer if the RX ring is full. */
> +	if (unlikely(!rq->vq->num_free))
> +		return true;
>  	gfp |= __GFP_COLD;
>  	do {
>  		if (vi->mergeable_rx_bufs)
> @@ -1354,8 +1401,10 @@ static void virtnet_free_queues(struct virtnet_info
*vi)
>  {
>  	int i;
>  
> -	for (i = 0; i < vi->max_queue_pairs; i++)
> +	for (i = 0; i < vi->max_queue_pairs; i++) {
>  		netif_napi_del(&vi->rq[i].napi);
> +		kfree(vi->rq[i].mrg_buf_ctx);
> +	}
>  
>  	kfree(vi->rq);
>  	kfree(vi->sq);
> @@ -1394,12 +1443,14 @@ static void free_unused_bufs(struct virtnet_info
*vi)
>  		struct virtqueue *vq = vi->rq[i].vq;
>  
>  		while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
> -			if (vi->mergeable_rx_bufs)
> -				put_page(virt_to_head_page(buf));
> -			else if (vi->big_packets)
> +			if (vi->mergeable_rx_bufs) {
> +				struct mergeable_receive_buf_ctx *ctx = buf;
> +				put_page(virt_to_head_page(ctx->buf));
> +			} else if (vi->big_packets) {
>  				give_pages(&vi->rq[i], buf);
> -			else
> +			} else {
>  				dev_kfree_skb(buf);
> +			}
>  			--vi->rq[i].num;
>  		}
>  		BUG_ON(vi->rq[i].num != 0);
> @@ -1509,6 +1560,7 @@ static int virtnet_alloc_queues(struct virtnet_info
*vi)
>  			       napi_weight);
>  
>  		sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
> +		ewma_init(&vi->rq[i].mrg_avg_pkt_len, 1, RECEIVE_AVG_WEIGHT);
>  		sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
>  	}
>  
> @@ -1522,7 +1574,8 @@ err_sq:
>  
>  static int init_vqs(struct virtnet_info *vi)
>  {
> -	int ret;
> +	struct virtio_device *vdev = vi->vdev;
> +	int i, ret;
>  
>  	/* Allocate send & receive queues */
>  	ret = virtnet_alloc_queues(vi);
> @@ -1533,12 +1586,28 @@ static int init_vqs(struct virtnet_info *vi)
>  	if (ret)
>  		goto err_free;
>  
> +	if (vi->mergeable_rx_bufs) {
> +		for (i = 0; i < vi->max_queue_pairs; i++) {
> +			struct receive_queue *rq = &vi->rq[i];
> +			rq->mrg_buf_ctx_size = virtqueue_get_vring_size(rq->vq);
> +			rq->mrg_buf_ctx = kmalloc(sizeof(*rq->mrg_buf_ctx) *
> +						  rq->mrg_buf_ctx_size,
> +						  GFP_KERNEL);
> +			if (!rq->mrg_buf_ctx) {
> +				ret = -ENOMEM;
> +				goto err_del_vqs;
> +			}
> +		}
> +	}
> +
>  	get_online_cpus();
>  	virtnet_set_affinity(vi);
>  	put_online_cpus();
>  
>  	return 0;
>  
> +err_del_vqs:
> +	vdev->config->del_vqs(vdev);
>  err_free:
>  	virtnet_free_queues(vi);
>  err:
> -- 
> 1.8.5.1