From 4d8b1e6aa5d7ecfc1d2ee606b4bd838b4f1ac9d2 Mon Sep 17 00:00:00 2001

From: Maxime Coquelin <maxime.coquelin@redhat.com>

Date: Thu, 17 May 2018 13:44:47 +0200

Subject: [PATCH] vhost: improve dirty pages logging performance

[ upstream commit c16915b8710911a75f0fbdb1aa5243f4cdfaf26a ]

This patch caches all dirty pages logging until the used ring index

is updated.

The goal of this optimization is to fix a performance regression

introduced when the vhost library started to use atomic operations

to set bits in the shared dirty log map. While the fix was valid

as previous implementation wasn't safe against concurrent accesses,

contention was induced.

With this patch, during migration, we have:

1. Less atomic operations as only a single atomic OR operation

per 32 or 64 (depending on CPU) pages.

2. Less atomic operations as during a burst, the same page will

be marked dirty only once.

3. Less write memory barriers.

Fixes: 897f13a1f726 ("vhost: make page logging atomic")

Cc: stable@dpdk.org

Suggested-by: Michael S. Tsirkin <mst@redhat.com>

Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>

Reviewed-by: Tiwei Bie <tiwei.bie@intel.com>

---

 lib/librte_vhost/vhost.h      | 119 +++++++++++++++++++++++++++++++++++++++++-

 lib/librte_vhost/virtio_net.c |  29 ++++++----

 2 files changed, 137 insertions(+), 11 deletions(-)

diff --git a/lib/librte_vhost/vhost.h b/lib/librte_vhost/vhost.h

index 16d6b8913..42c6a3a75 100644

--- a/lib/librte_vhost/vhost.h

+++ b/lib/librte_vhost/vhost.h

@@ -59,6 +59,8 @@

 #define BUF_VECTOR_MAX 256

+#define VHOST_LOG_CACHE_NR 32

+

 /**

  * Structure contains buffer address, length and descriptor index

  * from vring to do scatter RX.

@@ -92,6 +94,14 @@ struct batch_copy_elem {

 	uint64_t log_addr;

 };

+/*

+ * Structure that contains the info for batched dirty logging.

+ */

+struct log_cache_entry {

+	uint32_t offset;

+	unsigned long val;

+};

+

 /**

  * Structure contains variables relevant to RX/TX virtqueues.

  */

@@ -133,6 +143,9 @@ struct vhost_virtqueue {

 	struct batch_copy_elem	*batch_copy_elems;

 	uint16_t		batch_copy_nb_elems;

+	struct log_cache_entry log_cache[VHOST_LOG_CACHE_NR];

+	uint16_t log_cache_nb_elem;

+

 	rte_rwlock_t	iotlb_lock;

 	rte_rwlock_t	iotlb_pending_lock;

 	struct rte_mempool *iotlb_pool;

@@ -266,7 +279,15 @@ struct virtio_net {

 static __rte_always_inline void

 vhost_set_bit(unsigned int nr, volatile uint8_t *addr)

 {

-	__sync_fetch_and_or_8(addr, (1U << nr));

+#if defined(RTE_TOOLCHAIN_GCC) && (GCC_VERSION < 70100)

+	/*

+	 * __sync_ built-ins are deprecated, but __atomic_ ones

+	 * are sub-optimized in older GCC versions.

+	 */

+	__sync_fetch_and_or_1(addr, (1U << nr));

+#else

+	__atomic_fetch_or(addr, (1U << nr), __ATOMIC_RELAXED);

+#endif

 }

 static __rte_always_inline void

@@ -297,6 +318,102 @@ vhost_log_write(struct virtio_net *dev, uint64_t addr, uint64_t len)

 	}

 }

+static __rte_always_inline void

+vhost_log_cache_sync(struct virtio_net *dev, struct vhost_virtqueue *vq)

+{

+	unsigned long *log_base;

+	int i;

+

+	if (likely(((dev->features & (1ULL << VHOST_F_LOG_ALL)) == 0) ||

+		   !dev->log_base))

+		return;

+

+	log_base = (unsigned long *)(uintptr_t)dev->log_base;

+

+	/*

+	 * It is expected a write memory barrier has been issued

+	 * before this function is called.

+	 */

+

+	for (i = 0; i < vq->log_cache_nb_elem; i++) {

+		struct log_cache_entry *elem = vq->log_cache + i;

+

+#if defined(RTE_TOOLCHAIN_GCC) && (GCC_VERSION < 70100)

+		/*

+		 * '__sync' builtins are deprecated, but '__atomic' ones

+		 * are sub-optimized in older GCC versions.

+		 */

+		__sync_fetch_and_or(log_base + elem->offset, elem->val);

+#else

+		__atomic_fetch_or(log_base + elem->offset, elem->val,

+				__ATOMIC_RELAXED);

+#endif

+	}

+

+	rte_smp_wmb();

+

+	vq->log_cache_nb_elem = 0;

+}

+

+static __rte_always_inline void

+vhost_log_cache_page(struct virtio_net *dev, struct vhost_virtqueue *vq,

+			uint64_t page)

+{

+	uint32_t bit_nr = page % (sizeof(unsigned long) << 3);

+	uint32_t offset = page / (sizeof(unsigned long) << 3);

+	int i;

+

+	for (i = 0; i < vq->log_cache_nb_elem; i++) {

+		struct log_cache_entry *elem = vq->log_cache + i;

+

+		if (elem->offset == offset) {

+			elem->val |= (1UL << bit_nr);

+			return;

+		}

+	}

+

+	if (unlikely(i >= VHOST_LOG_CACHE_NR)) {

+		/*

+		 * No more room for a new log cache entry,

+		 * so write the dirty log map directly.

+		 */

+		rte_smp_wmb();

+		vhost_log_page((uint8_t *)(uintptr_t)dev->log_base, page);

+

+		return;

+	}

+

+	vq->log_cache[i].offset = offset;

+	vq->log_cache[i].val = (1UL << bit_nr);

+}

+

+static __rte_always_inline void

+vhost_log_cache_write(struct virtio_net *dev, struct vhost_virtqueue *vq,

+			uint64_t addr, uint64_t len)

+{

+	uint64_t page;

+

+	if (likely(((dev->features & (1ULL << VHOST_F_LOG_ALL)) == 0) ||

+		   !dev->log_base || !len))

+		return;

+

+	if (unlikely(dev->log_size <= ((addr + len - 1) / VHOST_LOG_PAGE / 8)))

+		return;

+

+	page = addr / VHOST_LOG_PAGE;

+	while (page * VHOST_LOG_PAGE < addr + len) {

+		vhost_log_cache_page(dev, vq, page);

+		page += 1;

+	}

+}

+

+static __rte_always_inline void

+vhost_log_cache_used_vring(struct virtio_net *dev, struct vhost_virtqueue *vq,

+			uint64_t offset, uint64_t len)

+{

+	vhost_log_cache_write(dev, vq, vq->log_guest_addr + offset, len);

+}

+

 static __rte_always_inline void

 vhost_log_used_vring(struct virtio_net *dev, struct vhost_virtqueue *vq,

 		     uint64_t offset, uint64_t len)

diff --git a/lib/librte_vhost/virtio_net.c b/lib/librte_vhost/virtio_net.c

index a013c07b0..5f8763d3a 100644

--- a/lib/librte_vhost/virtio_net.c

+++ b/lib/librte_vhost/virtio_net.c

@@ -107,7 +107,7 @@ do_flush_shadow_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,

 	rte_memcpy(&vq->used->ring[to],

 			&vq->shadow_used_ring[from],

 			size * sizeof(struct vring_used_elem));

-	vhost_log_used_vring(dev, vq,

+	vhost_log_cache_used_vring(dev, vq,

 			offsetof(struct vring_used, ring[to]),

 			size * sizeof(struct vring_used_elem));

 }

@@ -135,6 +135,8 @@ flush_shadow_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq)

 	rte_smp_wmb();

+	vhost_log_cache_sync(dev, vq);

+

 	*(volatile uint16_t *)&vq->used->idx += vq->shadow_used_idx;

 	vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),

 		sizeof(vq->used->idx));

@@ -159,7 +161,7 @@ do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)

 	for (i = 0; i < count; i++) {

 		rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);

-		vhost_log_write(dev, elem[i].log_addr, elem[i].len);

+		vhost_log_cache_write(dev, vq, elem[i].log_addr, elem[i].len);

 		PRINT_PACKET(dev, (uintptr_t)elem[i].dst, elem[i].len, 0);

 	}

 }

@@ -275,7 +277,7 @@ copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,

 		virtio_enqueue_offload(m,

 				(struct virtio_net_hdr *)(uintptr_t)desc_addr);

 		PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);

-		vhost_log_write(dev, desc_gaddr, dev->vhost_hlen);

+		vhost_log_cache_write(dev, vq, desc_gaddr, dev->vhost_hlen);

 	} else {

 		struct virtio_net_hdr vnet_hdr;

 		uint64_t remain = dev->vhost_hlen;

@@ -298,7 +300,7 @@ copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,

 					(void *)(uintptr_t)src, len);

 			PRINT_PACKET(dev, (uintptr_t)dst, len, 0);

-			vhost_log_write(dev, guest_addr, len);

+			vhost_log_cache_write(dev, vq, guest_addr, len);

 			remain -= len;

 			guest_addr += len;

 			dst += len;

@@ -379,7 +381,8 @@ copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,

 							desc_offset)),

 				rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),

 				cpy_len);

-			vhost_log_write(dev, desc_gaddr + desc_offset, cpy_len);

+			vhost_log_cache_write(dev, vq, desc_gaddr + desc_offset,

+					cpy_len);

 			PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),

 				     cpy_len, 0);

 		} else {

@@ -468,7 +471,7 @@ virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,

 		vq->used->ring[used_idx].id = desc_indexes[i];

 		vq->used->ring[used_idx].len = pkts[i]->pkt_len +

 					       dev->vhost_hlen;

-		vhost_log_used_vring(dev, vq,

+		vhost_log_cache_used_vring(dev, vq,

 			offsetof(struct vring_used, ring[used_idx]),

 			sizeof(vq->used->ring[used_idx]));

 	}

@@ -528,6 +531,8 @@ virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,

 	rte_smp_wmb();

+	vhost_log_cache_sync(dev, vq);

+

 	*(volatile uint16_t *)&vq->used->idx += count;

 	vq->last_used_idx += count;

 	vhost_log_used_vring(dev, vq,

@@ -797,7 +802,8 @@ copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,

 					PRINT_PACKET(dev, (uintptr_t)dst,

 							len, 0);

-					vhost_log_write(dev, guest_addr, len);

+					vhost_log_cache_write(dev, vq,

+							guest_addr, len);

 					remain -= len;

 					guest_addr += len;

@@ -806,7 +812,7 @@ copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,

 			} else {

 				PRINT_PACKET(dev, (uintptr_t)hdr_addr,

 						dev->vhost_hlen, 0);

-				vhost_log_write(dev, hdr_phys_addr,

+				vhost_log_cache_write(dev, vq, hdr_phys_addr,

 						dev->vhost_hlen);

 			}

@@ -820,7 +826,8 @@ copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,

 							desc_offset)),

 				rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),

 				cpy_len);

-			vhost_log_write(dev, desc_gaddr + desc_offset, cpy_len);

+			vhost_log_cache_write(dev, vq, desc_gaddr + desc_offset,

+					cpy_len);

 			PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),

 				cpy_len, 0);

 		} else {

@@ -1384,7 +1391,7 @@ update_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,

 {

 	vq->used->ring[used_idx].id  = desc_idx;

 	vq->used->ring[used_idx].len = 0;

-	vhost_log_used_vring(dev, vq,

+	vhost_log_cache_used_vring(dev, vq,

 			offsetof(struct vring_used, ring[used_idx]),

 			sizeof(vq->used->ring[used_idx]));

 }

@@ -1399,6 +1406,8 @@ update_used_idx(struct virtio_net *dev, struct vhost_virtqueue *vq,

 	rte_smp_wmb();

 	rte_smp_rmb();

+	vhost_log_cache_sync(dev, vq);

+

 	vq->used->idx += count;

 	vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),

 			sizeof(vq->used->idx));

-- 

2.14.3