vhost: do not inline unlikely fragmented buffers code
[dpdk.git] / lib / librte_vhost / vhost.c
index 91026b3..981837b 100644 (file)
@@ -69,6 +69,170 @@ __vhost_iova_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
        return 0;
 }
 
+#define VHOST_LOG_PAGE 4096
+
+/*
+ * Atomically set a bit in memory.
+ */
+static __rte_always_inline void
+vhost_set_bit(unsigned int nr, volatile uint8_t *addr)
+{
+#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
+vhost_log_page(uint8_t *log_base, uint64_t page)
+{
+       vhost_set_bit(page % 8, &log_base[page / 8]);
+}
+
+void
+__vhost_log_write(struct virtio_net *dev, uint64_t addr, uint64_t len)
+{
+       uint64_t page;
+
+       if (unlikely(!dev->log_base || !len))
+               return;
+
+       if (unlikely(dev->log_size <= ((addr + len - 1) / VHOST_LOG_PAGE / 8)))
+               return;
+
+       /* To make sure guest memory updates are committed before logging */
+       rte_smp_wmb();
+
+       page = addr / VHOST_LOG_PAGE;
+       while (page * VHOST_LOG_PAGE < addr + len) {
+               vhost_log_page((uint8_t *)(uintptr_t)dev->log_base, page);
+               page += 1;
+       }
+}
+
+void
+__vhost_log_cache_sync(struct virtio_net *dev, struct vhost_virtqueue *vq)
+{
+       unsigned long *log_base;
+       int i;
+
+       if (unlikely(!dev->log_base))
+               return;
+
+       rte_smp_wmb();
+
+       log_base = (unsigned long *)(uintptr_t)dev->log_base;
+
+       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);
+       vq->log_cache_nb_elem++;
+}
+
+void
+__vhost_log_cache_write(struct virtio_net *dev, struct vhost_virtqueue *vq,
+                       uint64_t addr, uint64_t len)
+{
+       uint64_t page;
+
+       if (unlikely(!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;
+       }
+}
+
+void *
+vhost_alloc_copy_ind_table(struct virtio_net *dev, struct vhost_virtqueue *vq,
+               uint64_t desc_addr, uint64_t desc_len)
+{
+       void *idesc;
+       uint64_t src, dst;
+       uint64_t len, remain = desc_len;
+
+       idesc = rte_malloc(__func__, desc_len, 0);
+       if (unlikely(!idesc))
+               return NULL;
+
+       dst = (uint64_t)(uintptr_t)idesc;
+
+       while (remain) {
+               len = remain;
+               src = vhost_iova_to_vva(dev, vq, desc_addr, &len,
+                               VHOST_ACCESS_RO);
+               if (unlikely(!src || !len)) {
+                       rte_free(idesc);
+                       return NULL;
+               }
+
+               rte_memcpy((void *)(uintptr_t)dst, (void *)(uintptr_t)src, len);
+
+               remain -= len;
+               dst += len;
+               desc_addr += len;
+       }
+
+       return idesc;
+}
+
 void
 cleanup_vq(struct vhost_virtqueue *vq, int destroy)
 {
@@ -344,6 +508,7 @@ vhost_new_device(void)
        dev->flags = VIRTIO_DEV_BUILTIN_VIRTIO_NET;
        dev->slave_req_fd = -1;
        dev->vdpa_dev_id = -1;
+       dev->postcopy_ufd = -1;
        rte_spinlock_init(&dev->slave_req_lock);
 
        return i;
@@ -399,19 +564,6 @@ vhost_attach_vdpa_device(int vid, int did)
        dev->vdpa_dev_id = did;
 }
 
-void
-vhost_detach_vdpa_device(int vid)
-{
-       struct virtio_net *dev = get_device(vid);
-
-       if (dev == NULL)
-               return;
-
-       vhost_user_host_notifier_ctrl(vid, false);
-
-       dev->vdpa_dev_id = -1;
-}
-
 void
 vhost_set_ifname(int vid, const char *if_name, unsigned int if_len)
 {
@@ -459,7 +611,7 @@ rte_vhost_get_mtu(int vid, uint16_t *mtu)
 {
        struct virtio_net *dev = get_device(vid);
 
-       if (!dev)
+       if (dev == NULL || mtu == NULL)
                return -ENODEV;
 
        if (!(dev->flags & VIRTIO_DEV_READY))
@@ -527,7 +679,7 @@ rte_vhost_get_ifname(int vid, char *buf, size_t len)
 {
        struct virtio_net *dev = get_device(vid);
 
-       if (dev == NULL)
+       if (dev == NULL || buf == NULL)
                return -1;
 
        len = RTE_MIN(len, sizeof(dev->ifname));
@@ -544,7 +696,7 @@ rte_vhost_get_negotiated_features(int vid, uint64_t *features)
        struct virtio_net *dev;
 
        dev = get_device(vid);
-       if (!dev)
+       if (dev == NULL || features == NULL)
                return -1;
 
        *features = dev->features;
@@ -559,7 +711,7 @@ rte_vhost_get_mem_table(int vid, struct rte_vhost_memory **mem)
        size_t size;
 
        dev = get_device(vid);
-       if (!dev)
+       if (dev == NULL || mem == NULL)
                return -1;
 
        size = dev->mem->nregions * sizeof(struct rte_vhost_mem_region);
@@ -582,7 +734,7 @@ rte_vhost_get_vhost_vring(int vid, uint16_t vring_idx,
        struct vhost_virtqueue *vq;
 
        dev = get_device(vid);
-       if (!dev)
+       if (dev == NULL || vring == NULL)
                return -1;
 
        if (vring_idx >= VHOST_MAX_VRING)
@@ -647,12 +799,18 @@ rte_vhost_avail_entries(int vid, uint16_t queue_id)
 }
 
 static inline void
-vhost_enable_notify_split(struct vhost_virtqueue *vq, int enable)
+vhost_enable_notify_split(struct virtio_net *dev,
+               struct vhost_virtqueue *vq, int enable)
 {
-       if (enable)
-               vq->used->flags &= ~VRING_USED_F_NO_NOTIFY;
-       else
-               vq->used->flags |= VRING_USED_F_NO_NOTIFY;
+       if (!(dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))) {
+               if (enable)
+                       vq->used->flags &= ~VRING_USED_F_NO_NOTIFY;
+               else
+                       vq->used->flags |= VRING_USED_F_NO_NOTIFY;
+       } else {
+               if (enable)
+                       vhost_avail_event(vq) = vq->last_avail_idx;
+       }
 }
 
 static inline void
@@ -661,8 +819,10 @@ vhost_enable_notify_packed(struct virtio_net *dev,
 {
        uint16_t flags;
 
-       if (!enable)
+       if (!enable) {
                vq->device_event->flags = VRING_EVENT_F_DISABLE;
+               return;
+       }
 
        flags = VRING_EVENT_F_ENABLE;
        if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX)) {
@@ -690,7 +850,7 @@ rte_vhost_enable_guest_notification(int vid, uint16_t queue_id, int enable)
        if (vq_is_packed(dev))
                vhost_enable_notify_packed(dev, vq, enable);
        else
-               vhost_enable_notify_split(vq, enable);
+               vhost_enable_notify_split(dev, vq, enable);
 
        return 0;
 }
@@ -767,16 +927,9 @@ int rte_vhost_get_log_base(int vid, uint64_t *log_base,
 {
        struct virtio_net *dev = get_device(vid);
 
-       if (!dev)
+       if (dev == NULL || log_base == NULL || log_size == NULL)
                return -1;
 
-       if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
-               RTE_LOG(ERR, VHOST_DATA,
-                       "(%d) %s: built-in vhost net backend is disabled.\n",
-                       dev->vid, __func__);
-               return -1;
-       }
-
        *log_base = dev->log_base;
        *log_size = dev->log_size;
 
@@ -788,15 +941,8 @@ int rte_vhost_get_vring_base(int vid, uint16_t queue_id,
 {
        struct virtio_net *dev = get_device(vid);
 
-       if (!dev)
-               return -1;
-
-       if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
-               RTE_LOG(ERR, VHOST_DATA,
-                       "(%d) %s: built-in vhost net backend is disabled.\n",
-                       dev->vid, __func__);
+       if (dev == NULL || last_avail_idx == NULL || last_used_idx == NULL)
                return -1;
-       }
 
        *last_avail_idx = dev->virtqueue[queue_id]->last_avail_idx;
        *last_used_idx = dev->virtqueue[queue_id]->last_used_idx;
@@ -812,15 +958,21 @@ int rte_vhost_set_vring_base(int vid, uint16_t queue_id,
        if (!dev)
                return -1;
 
-       if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
-               RTE_LOG(ERR, VHOST_DATA,
-                       "(%d) %s: built-in vhost net backend is disabled.\n",
-                       dev->vid, __func__);
-               return -1;
-       }
-
        dev->virtqueue[queue_id]->last_avail_idx = last_avail_idx;
        dev->virtqueue[queue_id]->last_used_idx = last_used_idx;
 
        return 0;
 }
+
+int rte_vhost_extern_callback_register(int vid,
+               struct rte_vhost_user_extern_ops const * const ops, void *ctx)
+{
+       struct virtio_net *dev = get_device(vid);
+
+       if (dev == NULL || ops == NULL)
+               return -1;
+
+       dev->extern_ops = *ops;
+       dev->extern_data = ctx;
+       return 0;
+}