nfs: new subdir Documentation/filesystems/nfs
[safe/jmp/linux-2.6] / drivers / virtio / virtio_ring.c
index 9859213..f536005 100644 (file)
  */
 #include <linux/virtio.h>
 #include <linux/virtio_ring.h>
+#include <linux/virtio_config.h>
 #include <linux/device.h>
 
 #ifdef DEBUG
 /* For development, we want to crash whenever the ring is screwed. */
-#define BAD_RING(vq, fmt...)                   \
-       do { dev_err(&vq->vq.vdev->dev, fmt); BUG(); } while(0)
-#define START_USE(vq) \
-       do { if ((vq)->in_use) panic("in_use = %i\n", (vq)->in_use); (vq)->in_use = __LINE__; mb(); } while(0)
-#define END_USE(vq) \
-       do { BUG_ON(!(vq)->in_use); (vq)->in_use = 0; mb(); } while(0)
+#define BAD_RING(_vq, fmt, args...)                            \
+       do {                                                    \
+               dev_err(&(_vq)->vq.vdev->dev,                   \
+                       "%s:"fmt, (_vq)->vq.name, ##args);      \
+               BUG();                                          \
+       } while (0)
+/* Caller is supposed to guarantee no reentry. */
+#define START_USE(_vq)                                         \
+       do {                                                    \
+               if ((_vq)->in_use)                              \
+                       panic("%s:in_use = %i\n",               \
+                             (_vq)->vq.name, (_vq)->in_use);   \
+               (_vq)->in_use = __LINE__;                       \
+               mb();                                           \
+       } while (0)
+#define END_USE(_vq) \
+       do { BUG_ON(!(_vq)->in_use); (_vq)->in_use = 0; mb(); } while(0)
 #else
-#define BAD_RING(vq, fmt...)                   \
-       do { dev_err(&vq->vq.vdev->dev, fmt); (vq)->broken = true; } while(0)
+#define BAD_RING(_vq, fmt, args...)                            \
+       do {                                                    \
+               dev_err(&_vq->vq.vdev->dev,                     \
+                       "%s:"fmt, (_vq)->vq.name, ##args);      \
+               (_vq)->broken = true;                           \
+       } while (0)
 #define START_USE(vq)
 #define END_USE(vq)
 #endif
@@ -45,6 +61,9 @@ struct vring_virtqueue
        /* Other side has made a mess, don't try any more. */
        bool broken;
 
+       /* Host supports indirect buffers */
+       bool indirect;
+
        /* Number of free buffers */
        unsigned int num_free;
        /* Head of free buffer list. */
@@ -69,6 +88,55 @@ struct vring_virtqueue
 
 #define to_vvq(_vq) container_of(_vq, struct vring_virtqueue, vq)
 
+/* Set up an indirect table of descriptors and add it to the queue. */
+static int vring_add_indirect(struct vring_virtqueue *vq,
+                             struct scatterlist sg[],
+                             unsigned int out,
+                             unsigned int in)
+{
+       struct vring_desc *desc;
+       unsigned head;
+       int i;
+
+       desc = kmalloc((out + in) * sizeof(struct vring_desc), GFP_ATOMIC);
+       if (!desc)
+               return vq->vring.num;
+
+       /* Transfer entries from the sg list into the indirect page */
+       for (i = 0; i < out; i++) {
+               desc[i].flags = VRING_DESC_F_NEXT;
+               desc[i].addr = sg_phys(sg);
+               desc[i].len = sg->length;
+               desc[i].next = i+1;
+               sg++;
+       }
+       for (; i < (out + in); i++) {
+               desc[i].flags = VRING_DESC_F_NEXT|VRING_DESC_F_WRITE;
+               desc[i].addr = sg_phys(sg);
+               desc[i].len = sg->length;
+               desc[i].next = i+1;
+               sg++;
+       }
+
+       /* Last one doesn't continue. */
+       desc[i-1].flags &= ~VRING_DESC_F_NEXT;
+       desc[i-1].next = 0;
+
+       /* We're about to use a buffer */
+       vq->num_free--;
+
+       /* Use a single buffer which doesn't continue */
+       head = vq->free_head;
+       vq->vring.desc[head].flags = VRING_DESC_F_INDIRECT;
+       vq->vring.desc[head].addr = virt_to_phys(desc);
+       vq->vring.desc[head].len = i * sizeof(struct vring_desc);
+
+       /* Update free pointer */
+       vq->free_head = vq->vring.desc[head].next;
+
+       return head;
+}
+
 static int vring_add_buf(struct virtqueue *_vq,
                         struct scatterlist sg[],
                         unsigned int out,
@@ -78,17 +146,29 @@ static int vring_add_buf(struct virtqueue *_vq,
        struct vring_virtqueue *vq = to_vvq(_vq);
        unsigned int i, avail, head, uninitialized_var(prev);
 
+       START_USE(vq);
+
        BUG_ON(data == NULL);
+
+       /* If the host supports indirect descriptor tables, and we have multiple
+        * buffers, then go indirect. FIXME: tune this threshold */
+       if (vq->indirect && (out + in) > 1 && vq->num_free) {
+               head = vring_add_indirect(vq, sg, out, in);
+               if (head != vq->vring.num)
+                       goto add_head;
+       }
+
        BUG_ON(out + in > vq->vring.num);
        BUG_ON(out + in == 0);
 
-       START_USE(vq);
-
        if (vq->num_free < out + in) {
                pr_debug("Can't add buf len %i - avail = %i\n",
                         out + in, vq->num_free);
-               /* We notify *even if* VRING_USED_F_NO_NOTIFY is set here. */
-               vq->notify(&vq->vq);
+               /* FIXME: for historical reasons, we force a notify here if
+                * there are outgoing parts to the buffer.  Presumably the
+                * host should service the ring ASAP. */
+               if (out)
+                       vq->notify(&vq->vq);
                END_USE(vq);
                return -ENOSPC;
        }
@@ -99,16 +179,14 @@ static int vring_add_buf(struct virtqueue *_vq,
        head = vq->free_head;
        for (i = vq->free_head; out; i = vq->vring.desc[i].next, out--) {
                vq->vring.desc[i].flags = VRING_DESC_F_NEXT;
-               vq->vring.desc[i].addr = (page_to_pfn(sg_page(sg))<<PAGE_SHIFT)
-                       + sg->offset;
+               vq->vring.desc[i].addr = sg_phys(sg);
                vq->vring.desc[i].len = sg->length;
                prev = i;
                sg++;
        }
        for (; in; i = vq->vring.desc[i].next, in--) {
                vq->vring.desc[i].flags = VRING_DESC_F_NEXT|VRING_DESC_F_WRITE;
-               vq->vring.desc[i].addr = (page_to_pfn(sg_page(sg))<<PAGE_SHIFT)
-                       + sg->offset;
+               vq->vring.desc[i].addr = sg_phys(sg);
                vq->vring.desc[i].len = sg->length;
                prev = i;
                sg++;
@@ -119,6 +197,7 @@ static int vring_add_buf(struct virtqueue *_vq,
        /* Update free pointer */
        vq->free_head = i;
 
+add_head:
        /* Set token. */
        vq->data[head] = data;
 
@@ -129,7 +208,11 @@ static int vring_add_buf(struct virtqueue *_vq,
 
        pr_debug("Added buffer head %i to %p\n", head, vq);
        END_USE(vq);
-       return 0;
+
+       /* If we're indirect, we can fit many (assuming not OOM). */
+       if (vq->indirect)
+               return vq->num_free ? vq->vring.num : 0;
+       return vq->num_free;
 }
 
 static void vring_kick(struct virtqueue *_vq)
@@ -162,6 +245,11 @@ static void detach_buf(struct vring_virtqueue *vq, unsigned int head)
 
        /* Put back on free list: find end */
        i = head;
+
+       /* Free the indirect table */
+       if (vq->vring.desc[i].flags & VRING_DESC_F_INDIRECT)
+               kfree(phys_to_virt(vq->vring.desc[i].addr));
+
        while (vq->vring.desc[i].flags & VRING_DESC_F_NEXT) {
                i = vq->vring.desc[i].next;
                vq->num_free++;
@@ -186,6 +274,11 @@ static void *vring_get_buf(struct virtqueue *_vq, unsigned int *len)
 
        START_USE(vq);
 
+       if (unlikely(vq->broken)) {
+               END_USE(vq);
+               return NULL;
+       }
+
        if (!more_used(vq)) {
                pr_debug("No more buffers in queue\n");
                END_USE(vq);
@@ -216,10 +309,7 @@ static void vring_disable_cb(struct virtqueue *_vq)
 {
        struct vring_virtqueue *vq = to_vvq(_vq);
 
-       START_USE(vq);
-       BUG_ON(vq->vring.avail->flags & VRING_AVAIL_F_NO_INTERRUPT);
        vq->vring.avail->flags |= VRING_AVAIL_F_NO_INTERRUPT;
-       END_USE(vq);
 }
 
 static bool vring_enable_cb(struct virtqueue *_vq)
@@ -227,14 +317,12 @@ static bool vring_enable_cb(struct virtqueue *_vq)
        struct vring_virtqueue *vq = to_vvq(_vq);
 
        START_USE(vq);
-       BUG_ON(!(vq->vring.avail->flags & VRING_AVAIL_F_NO_INTERRUPT));
 
        /* We optimistically turn back on interrupts, then check if there was
         * more to do. */
        vq->vring.avail->flags &= ~VRING_AVAIL_F_NO_INTERRUPT;
        mb();
        if (unlikely(more_used(vq))) {
-               vq->vring.avail->flags |= VRING_AVAIL_F_NO_INTERRUPT;
                END_USE(vq);
                return false;
        }
@@ -255,19 +343,13 @@ irqreturn_t vring_interrupt(int irq, void *_vq)
        if (unlikely(vq->broken))
                return IRQ_HANDLED;
 
-       /* Other side may have missed us turning off the interrupt,
-        * but we should preserve disable semantic for virtio users. */
-       if (unlikely(vq->vring.avail->flags & VRING_AVAIL_F_NO_INTERRUPT)) {
-               pr_debug("virtqueue interrupt after disable for %p\n", vq);
-               return IRQ_HANDLED;
-       }
-
        pr_debug("virtqueue callback for %p (%p)\n", vq, vq->vq.callback);
        if (vq->vq.callback)
                vq->vq.callback(&vq->vq);
 
        return IRQ_HANDLED;
 }
+EXPORT_SYMBOL_GPL(vring_interrupt);
 
 static struct virtqueue_ops vring_vq_ops = {
        .add_buf = vring_add_buf,
@@ -278,10 +360,12 @@ static struct virtqueue_ops vring_vq_ops = {
 };
 
 struct virtqueue *vring_new_virtqueue(unsigned int num,
+                                     unsigned int vring_align,
                                      struct virtio_device *vdev,
                                      void *pages,
                                      void (*notify)(struct virtqueue *),
-                                     void (*callback)(struct virtqueue *))
+                                     void (*callback)(struct virtqueue *),
+                                     const char *name)
 {
        struct vring_virtqueue *vq;
        unsigned int i;
@@ -296,18 +380,22 @@ struct virtqueue *vring_new_virtqueue(unsigned int num,
        if (!vq)
                return NULL;
 
-       vring_init(&vq->vring, num, pages, PAGE_SIZE);
+       vring_init(&vq->vring, num, pages, vring_align);
        vq->vq.callback = callback;
        vq->vq.vdev = vdev;
        vq->vq.vq_ops = &vring_vq_ops;
+       vq->vq.name = name;
        vq->notify = notify;
        vq->broken = false;
        vq->last_used_idx = 0;
        vq->num_added = 0;
+       list_add_tail(&vq->vq.list, &vdev->vqs);
 #ifdef DEBUG
        vq->in_use = false;
 #endif
 
+       vq->indirect = virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC);
+
        /* No callback?  Tell other side not to bother us. */
        if (!callback)
                vq->vring.avail->flags |= VRING_AVAIL_F_NO_INTERRUPT;
@@ -320,9 +408,30 @@ struct virtqueue *vring_new_virtqueue(unsigned int num,
 
        return &vq->vq;
 }
+EXPORT_SYMBOL_GPL(vring_new_virtqueue);
 
 void vring_del_virtqueue(struct virtqueue *vq)
 {
+       list_del(&vq->list);
        kfree(to_vvq(vq));
 }
+EXPORT_SYMBOL_GPL(vring_del_virtqueue);
+
+/* Manipulates transport-specific feature bits. */
+void vring_transport_features(struct virtio_device *vdev)
+{
+       unsigned int i;
+
+       for (i = VIRTIO_TRANSPORT_F_START; i < VIRTIO_TRANSPORT_F_END; i++) {
+               switch (i) {
+               case VIRTIO_RING_F_INDIRECT_DESC:
+                       break;
+               default:
+                       /* We don't understand this bit. */
+                       clear_bit(i, vdev->features);
+               }
+       }
+}
+EXPORT_SYMBOL_GPL(vring_transport_features);
 
+MODULE_LICENSE("GPL");