vhost: don't bother copying iovecs in handle_rx(), kill memcpy_toiovecend()
authorAl Viro <viro@zeniv.linux.org.uk>
Wed, 10 Dec 2014 20:51:28 +0000 (15:51 -0500)
committerAl Viro <viro@zeniv.linux.org.uk>
Wed, 4 Feb 2015 06:34:16 +0000 (01:34 -0500)
Cc: Michael S. Tsirkin <mst@redhat.com>
Cc: kvm@vger.kernel.org
Cc: virtualization@lists.linux-foundation.org
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
drivers/vhost/net.c
include/linux/uio.h
lib/iovec.c

index d86cc9bb9ea4e16060fbaa414934f3f97bf85f9d..e022cc40303d69cd2664cab909157672cd58f744 100644 (file)
@@ -84,10 +84,6 @@ struct vhost_net_ubuf_ref {
 
 struct vhost_net_virtqueue {
        struct vhost_virtqueue vq;
-       /* hdr is used to store the virtio header.
-        * Since each iovec has >= 1 byte length, we never need more than
-        * header length entries to store the header. */
-       struct iovec hdr[sizeof(struct virtio_net_hdr_mrg_rxbuf)];
        size_t vhost_hlen;
        size_t sock_hlen;
        /* vhost zerocopy support fields below: */
@@ -235,44 +231,6 @@ static bool vhost_sock_zcopy(struct socket *sock)
                sock_flag(sock->sk, SOCK_ZEROCOPY);
 }
 
-/* Pop first len bytes from iovec. Return number of segments used. */
-static int move_iovec_hdr(struct iovec *from, struct iovec *to,
-                         size_t len, int iov_count)
-{
-       int seg = 0;
-       size_t size;
-
-       while (len && seg < iov_count) {
-               size = min(from->iov_len, len);
-               to->iov_base = from->iov_base;
-               to->iov_len = size;
-               from->iov_len -= size;
-               from->iov_base += size;
-               len -= size;
-               ++from;
-               ++to;
-               ++seg;
-       }
-       return seg;
-}
-/* Copy iovec entries for len bytes from iovec. */
-static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
-                          size_t len, int iovcount)
-{
-       int seg = 0;
-       size_t size;
-
-       while (len && seg < iovcount) {
-               size = min(from->iov_len, len);
-               to->iov_base = from->iov_base;
-               to->iov_len = size;
-               len -= size;
-               ++from;
-               ++to;
-               ++seg;
-       }
-}
-
 /* In case of DMA done not in order in lower device driver for some reason.
  * upend_idx is used to track end of used idx, done_idx is used to track head
  * of used idx. Once lower device DMA done contiguously, we will signal KVM
@@ -570,9 +528,9 @@ static void handle_rx(struct vhost_net *net)
                .msg_controllen = 0,
                .msg_flags = MSG_DONTWAIT,
        };
-       struct virtio_net_hdr_mrg_rxbuf hdr = {
-               .hdr.flags = 0,
-               .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
+       struct virtio_net_hdr hdr = {
+               .flags = 0,
+               .gso_type = VIRTIO_NET_HDR_GSO_NONE
        };
        size_t total_len = 0;
        int err, mergeable;
@@ -580,6 +538,7 @@ static void handle_rx(struct vhost_net *net)
        size_t vhost_hlen, sock_hlen;
        size_t vhost_len, sock_len;
        struct socket *sock;
+       struct iov_iter fixup;
 
        mutex_lock(&vq->mutex);
        sock = vq->private_data;
@@ -624,14 +583,19 @@ static void handle_rx(struct vhost_net *net)
                        break;
                }
                /* We don't need to be notified again. */
-               if (unlikely((vhost_hlen)))
-                       /* Skip header. TODO: support TSO. */
-                       move_iovec_hdr(vq->iov, nvq->hdr, vhost_hlen, in);
-               else
-                       /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
-                        * needed because recvmsg can modify msg_iov. */
-                       copy_iovec_hdr(vq->iov, nvq->hdr, sock_hlen, in);
-               iov_iter_init(&msg.msg_iter, READ, vq->iov, in, sock_len);
+               iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len);
+               fixup = msg.msg_iter;
+               if (unlikely((vhost_hlen))) {
+                       /* We will supply the header ourselves
+                        * TODO: support TSO.
+                        */
+                       iov_iter_advance(&msg.msg_iter, vhost_hlen);
+               } else {
+                       /* It'll come from socket; we'll need to patch
+                        * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
+                        */
+                       iov_iter_advance(&fixup, sizeof(hdr));
+               }
                err = sock->ops->recvmsg(NULL, sock, &msg,
                                         sock_len, MSG_DONTWAIT | MSG_TRUNC);
                /* Userspace might have consumed the packet meanwhile:
@@ -643,18 +607,18 @@ static void handle_rx(struct vhost_net *net)
                        vhost_discard_vq_desc(vq, headcount);
                        continue;
                }
+               /* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
                if (unlikely(vhost_hlen) &&
-                   memcpy_toiovecend(nvq->hdr, (unsigned char *)&hdr, 0,
-                                     vhost_hlen)) {
+                   copy_to_iter(&hdr, sizeof(hdr), &fixup) != sizeof(hdr)) {
                        vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
                               vq->iov->iov_base);
                        break;
                }
-               /* TODO: Should check and handle checksum. */
+               /* Supply (or replace) ->num_buffers if VIRTIO_NET_F_MRG_RXBUF
+                * TODO: Should check and handle checksum.
+                */
                if (likely(mergeable) &&
-                   memcpy_toiovecend(nvq->hdr, (unsigned char *)&headcount,
-                                     offsetof(typeof(hdr), num_buffers),
-                                     sizeof hdr.num_buffers)) {
+                   copy_to_iter(&headcount, 2, &fixup) != 2) {
                        vq_err(vq, "Failed num_buffers write");
                        vhost_discard_vq_desc(vq, headcount);
                        break;
index af3439f4ebf2501485f2dbbc0e71ecf2e1e00fac..02bd8a92038a8b6112abfbfb88e80d6ff3c3db22 100644 (file)
@@ -137,7 +137,4 @@ size_t csum_and_copy_from_iter(void *addr, size_t bytes, __wsum *csum, struct io
 
 int memcpy_fromiovecend(unsigned char *kdata, const struct iovec *iov,
                        int offset, int len);
-int memcpy_toiovecend(const struct iovec *v, unsigned char *kdata,
-                     int offset, int len);
-
 #endif
index 4a90875c64ae711fad4e112951fbfbf0524ad50c..d8f17a9b1ccf90e3033fd561d256da0ee0c818a1 100644 (file)
@@ -2,32 +2,6 @@
 #include <linux/export.h>
 #include <linux/uio.h>
 
-/*
- *     Copy kernel to iovec. Returns -EFAULT on error.
- */
-
-int memcpy_toiovecend(const struct iovec *iov, unsigned char *kdata,
-                     int offset, int len)
-{
-       int copy;
-       for (; len > 0; ++iov) {
-               /* Skip over the finished iovecs */
-               if (unlikely(offset >= iov->iov_len)) {
-                       offset -= iov->iov_len;
-                       continue;
-               }
-               copy = min_t(unsigned int, iov->iov_len - offset, len);
-               if (copy_to_user(iov->iov_base + offset, kdata, copy))
-                       return -EFAULT;
-               offset = 0;
-               kdata += copy;
-               len -= copy;
-       }
-
-       return 0;
-}
-EXPORT_SYMBOL(memcpy_toiovecend);
-
 /*
  *     Copy iovec to kernel. Returns -EFAULT on error.
  */