]> err.no Git - linux-2.6/commitdiff
tun: fallback if skb_alloc() fails on big packets
authorRusty Russell <rusty@rustcorp.com.au>
Fri, 15 Aug 2008 22:15:10 +0000 (15:15 -0700)
committerDavid S. Miller <davem@davemloft.net>
Sat, 16 Aug 2008 02:52:31 +0000 (19:52 -0700)
skb_alloc produces linear packets (using kmalloc()).  That can fail,
so should we fall back to making paged skbs.

My original version of this patch always allocate paged skbs for big
packets.  But that made performance drop from 8.4 seconds to 8.8
seconds on 1G lguest->Host TCP xmit.  So now we only do that as a
fallback.

Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Acked-by: Max Krasnyansky <maxk@qualcomm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/tun.c

index 95931a5a98838d5c21af7e7b3de414c0140af466..6daea0c91862ecddc2110712c8e62381e1e9639c 100644 (file)
@@ -358,6 +358,66 @@ static unsigned int tun_chr_poll(struct file *file, poll_table * wait)
        return mask;
 }
 
+/* prepad is the amount to reserve at front.  len is length after that.
+ * linear is a hint as to how much to copy (usually headers). */
+static struct sk_buff *tun_alloc_skb(size_t prepad, size_t len, size_t linear,
+                                    gfp_t gfp)
+{
+       struct sk_buff *skb;
+       unsigned int i;
+
+       skb = alloc_skb(prepad + len, gfp|__GFP_NOWARN);
+       if (skb) {
+               skb_reserve(skb, prepad);
+               skb_put(skb, len);
+               return skb;
+       }
+
+       /* Under a page?  Don't bother with paged skb. */
+       if (prepad + len < PAGE_SIZE)
+               return NULL;
+
+       /* Start with a normal skb, and add pages. */
+       skb = alloc_skb(prepad + linear, gfp);
+       if (!skb)
+               return NULL;
+
+       skb_reserve(skb, prepad);
+       skb_put(skb, linear);
+
+       len -= linear;
+
+       for (i = 0; i < MAX_SKB_FRAGS; i++) {
+               skb_frag_t *f = &skb_shinfo(skb)->frags[i];
+
+               f->page = alloc_page(gfp|__GFP_ZERO);
+               if (!f->page)
+                       break;
+
+               f->page_offset = 0;
+               f->size = PAGE_SIZE;
+
+               skb->data_len += PAGE_SIZE;
+               skb->len += PAGE_SIZE;
+               skb->truesize += PAGE_SIZE;
+               skb_shinfo(skb)->nr_frags++;
+
+               if (len < PAGE_SIZE) {
+                       len = 0;
+                       break;
+               }
+               len -= PAGE_SIZE;
+       }
+
+       /* Too large, or alloc fail? */
+       if (unlikely(len)) {
+               kfree_skb(skb);
+               skb = NULL;
+       }
+
+       return skb;
+}
+
 /* Get packet from user space buffer */
 static __inline__ ssize_t tun_get_user(struct tun_struct *tun, struct iovec *iv, size_t count)
 {
@@ -391,14 +451,12 @@ static __inline__ ssize_t tun_get_user(struct tun_struct *tun, struct iovec *iv,
                        return -EINVAL;
        }
 
-       if (!(skb = alloc_skb(len + align, GFP_KERNEL))) {
+       if (!(skb = tun_alloc_skb(align, len, gso.hdr_len, GFP_KERNEL))) {
                tun->dev->stats.rx_dropped++;
                return -ENOMEM;
        }
 
-       if (align)
-               skb_reserve(skb, align);
-       if (memcpy_fromiovec(skb_put(skb, len), iv, len)) {
+       if (skb_copy_datagram_from_iovec(skb, 0, iv, len)) {
                tun->dev->stats.rx_dropped++;
                kfree_skb(skb);
                return -EFAULT;