bitkeeper revision 1.1159.79.10 (41505c58w6Hk7xUbFnsEdE5r462MrQ)

Rewrite our skb cache. Fix frontend driver to ask for sensibly-sized
skbuffs.

.rootkeys

@@ -126,6 +126,7 @@
 409ba2e7akOFqQUg6Qyg2s28xcXiMg linux-2.4.27-xen-sparse/mm/page_alloc.c
 3e5a4e683HKVU-sxtagrDasRB8eBVw linux-2.4.27-xen-sparse/mm/swapfile.c
 41180721bNns9Na7w1nJ0ZVt8bhUNA linux-2.4.27-xen-sparse/mm/vmalloc.c
+41505c57WAd5l1rlfCLNSCpx9J13vA linux-2.4.27-xen-sparse/net/core/skbuff.c
 413aa1d0oNP8HXLvfPuMe6cSroUfSA linux-2.6.8.1-patches/agpgart.patch
 413aa1d0ewvSv-ohnNnQQNGsbPTTNA linux-2.6.8.1-patches/drm.patch
 40f562372u3A7_kfbYYixPHJJxYUxA linux-2.6.8.1-xen-sparse/arch/xen/Kconfig
@@ -259,6 +260,7 @@
 40f56a0ddHCSs3501MY4hRf22tctOw linux-2.6.8.1-xen-sparse/mkbuildtree
 412f46c0LJuKAgSPGoC0Z1DEkLfuLA linux-2.6.8.1-xen-sparse/mm/memory.c
 410a94a4KT6I6X0LVc7djB39tRDp4g linux-2.6.8.1-xen-sparse/mm/page_alloc.c
+41505c572m-s9ATiO1LiD1GPznTTIg linux-2.6.8.1-xen-sparse/net/core/skbuff.c
 4149ec79wMpIHdvbntxqVGLRZZjPxw linux-2.6.8.1-xen-sparse/net/ipv4/raw.c
 413cb1e4zst25MDYjg63Y-NGC5_pLg netbsd-2.0-xen-sparse/Makefile
 413cb1e5c_Mkxf_X0zimEhTKI_l4DA netbsd-2.0-xen-sparse/mkbuildtree

linux-2.4.27-xen-sparse/include/linux/skbuff.h

@@ -231,6 +231,7 @@ struct sk_buff {
 
 extern void			__kfree_skb(struct sk_buff *skb);
 extern struct sk_buff *		alloc_skb(unsigned int size, int priority);
+extern struct sk_buff *		alloc_skb_from_cache(kmem_cache_t *cp, unsigned int size, int priority);
 extern void			kfree_skbmem(struct sk_buff *skb);
 extern struct sk_buff *		skb_clone(struct sk_buff *skb, int priority);
 extern struct sk_buff *		skb_copy(const struct sk_buff *skb, int priority);

linux-2.6.8.1-xen-sparse/arch/xen/kernel/skbuff.c

@@ -27,34 +27,9 @@ EXPORT_SYMBOL(__dev_alloc_skb);
 struct sk_buff *__dev_alloc_skb(unsigned int length, int gfp_mask)
 {
     struct sk_buff *skb;
-    u8             *new_data, *new_shinfo; 
-
-    /*
-     * Yuk! There is no way to get a skbuff head without allocating the
-     * data area using kmalloc(). So we do that and then replace the default
-     * data area with our own.
-     */
-    skb = alloc_skb(0, gfp_mask);
-    if ( unlikely(skb == NULL) )
-        return NULL;
-
-    new_data = kmem_cache_alloc(skbuff_cachep, gfp_mask);
-    if ( new_data == NULL )
-    {
-        dev_kfree_skb(skb);
-        return NULL;
-    }
-
-    new_shinfo = new_data + XEN_SKB_SIZE;
-    memcpy(new_shinfo, skb_shinfo(skb), sizeof(struct skb_shared_info));
-
-    kfree(skb->head);
-
-    skb->head = new_data;
-    skb->data = skb->tail = new_data + 16; /* __dev_alloc_skb does this */
-    skb->end  = new_shinfo;
-    skb->truesize = 1500;                  /* is this important? */
-
+    skb = alloc_skb_from_cache(skbuff_cachep, length + 16, gfp_mask);
+    if ( likely(skb != NULL) )
+        skb_reserve(skb, 16);
     return skb;
 }
 

linux-2.6.8.1-xen-sparse/drivers/xen/netback/netback.c

@@ -144,11 +144,12 @@ int netif_be_start_xmit(struct sk_buff *skb, struct net_device *dev)
      */
     if ( skb_shared(skb) || skb_cloned(skb) || !is_xen_skb(skb) )
     {
-        struct sk_buff *nskb = dev_alloc_skb(PAGE_SIZE);
         int hlen = skb->data - skb->head;
+        struct sk_buff *nskb = dev_alloc_skb(hlen + skb->len);
         if ( unlikely(nskb == NULL) )
             goto drop;
-        skb_reserve(nskb, hlen);
+        /* Account for any reservation already made by dev_alloc_skb(). */
+        skb_reserve(nskb, hlen - (nskb->data - nskb->head));
         __skb_put(nskb, skb->len);
         (void)skb_copy_bits(skb, -hlen, nskb->head, hlen + skb->len);
         nskb->dev = skb->dev;

linux-2.6.8.1-xen-sparse/drivers/xen/netfront/netfront.c

@@ -39,7 +39,17 @@
 #ifndef __GFP_NOWARN
 #define __GFP_NOWARN 0
 #endif
-#define alloc_skb_page() __dev_alloc_skb(PAGE_SIZE, GFP_ATOMIC|__GFP_NOWARN)
+#define alloc_xen_skb(_l) __dev_alloc_skb((_l), GFP_ATOMIC|__GFP_NOWARN)
+
+#define init_skb_shinfo(_skb)                         \
+    do {                                              \
+        atomic_set(&(skb_shinfo(_skb)->dataref), 1);  \
+        skb_shinfo(_skb)->nr_frags = 0;               \
+        skb_shinfo(_skb)->frag_list = NULL;           \
+    } while ( 0 )
+
+/* Allow headroom on each rx pkt for Ethernet header, alignment padding, ... */
+#define RX_HEADROOM 100
 
 /*
  * If the backend driver is pipelining transmit requests then we can be very
@@ -249,7 +259,7 @@ static void network_tx_buf_gc(struct net_device *dev)
             id  = np->tx->ring[MASK_NETIF_TX_IDX(i)].resp.id;
             skb = np->tx_skbs[id];
             ADD_ID_TO_FREELIST(np->tx_skbs, id);
-            dev_kfree_skb_any(skb);
+            dev_kfree_skb_irq(skb);
         }
         
         np->tx_resp_cons = prod;
@@ -292,7 +302,7 @@ static void network_alloc_rx_buffers(struct net_device *dev)
         return;
 
     do {
-        if ( unlikely((skb = alloc_skb_page()) == NULL) )
+        if ( unlikely((skb = alloc_xen_skb(dev->mtu + RX_HEADROOM)) == NULL) )
             break;
 
         skb->dev = dev;
@@ -368,7 +378,7 @@ static int network_start_xmit(struct sk_buff *skb, struct net_device *dev)
                   PAGE_SIZE) )
     {
         struct sk_buff *new_skb;
-        if ( unlikely((new_skb = alloc_skb_page()) == NULL) )
+        if ( unlikely((new_skb = alloc_xen_skb(skb->len)) == NULL) )
             goto drop;
         skb_put(new_skb, skb->len);
         memcpy(new_skb->data, skb->data, skb->len);
@@ -446,7 +456,7 @@ static irqreturn_t netif_int(int irq, void *dev_id, struct pt_regs *ptregs)
 static int netif_poll(struct net_device *dev, int *pbudget)
 {
     struct net_private *np = dev->priv;
-    struct sk_buff *skb;
+    struct sk_buff *skb, *nskb;
     netif_rx_response_t *rx;
     NETIF_RING_IDX i, rp;
     mmu_update_t *mmu = rx_mmu;
@@ -494,8 +504,10 @@ static int netif_poll(struct net_device *dev, int *pbudget)
         skb = np->rx_skbs[rx->id];
         ADD_ID_TO_FREELIST(np->rx_skbs, rx->id);
 
-        skb->data = skb->tail = skb->head + (rx->addr & ~PAGE_MASK);
-        skb_put(skb, rx->status);
+        /* NB. We handle skb overflow later. */
+        skb->data = skb->head + (rx->addr & ~PAGE_MASK);
+        skb->len  = rx->status;
+        skb->tail = skb->data + skb->len;
 
         np->stats.rx_packets++;
         np->stats.rx_bytes += rx->status;
@@ -529,10 +541,39 @@ static int netif_poll(struct net_device *dev, int *pbudget)
 
     while ( (skb = __skb_dequeue(&rxq)) != NULL )
     {
+        /*
+         * Enough room in skbuff for the data we were passed? Also, Linux 
+         * expects at least 16 bytes headroom in each receive buffer.
+         */
+        if ( unlikely(skb->tail > skb->end) ||
+             unlikely((skb->data - skb->head) < 16) )
+        {
+            nskb = NULL;
+
+            /* Only copy the packet if it fits in the current MTU. */
+            if ( skb->len <= (dev->mtu + ETH_HLEN) )
+            {
+                if ( (nskb = alloc_xen_skb(skb->len + 2)) != NULL )
+                {
+                    skb_reserve(nskb, 2);
+                    skb_put(nskb, skb->len);
+                    memcpy(nskb->data, skb->data, skb->len);
+                }
+            }
+
+            /* Reinitialise and then destroy the old skbuff. */
+            skb->len  = 0;
+            skb->tail = skb->data;
+            init_skb_shinfo(skb);
+            dev_kfree_skb(skb);
+
+            /* Switch old for new, if we copied the buffer. */
+            if ( (skb = nskb) == NULL )
+                continue;
+        }
+        
         /* Set the shared-info area, which is hidden behind the real data. */
-        atomic_set(&(skb_shinfo(skb)->dataref), 1);
-        skb_shinfo(skb)->nr_frags = 0;
-        skb_shinfo(skb)->frag_list = NULL;
+        init_skb_shinfo(skb);
 
         /* Ethernet-specific work. Delayed to here as it peeks the header. */
         skb->protocol = eth_type_trans(skb, dev);
@@ -596,8 +637,8 @@ static void network_connect(struct net_device *dev,
     netif_tx_request_t *tx;
 
     np = dev->priv;
-    spin_lock_irq(&np->rx_lock);
-    spin_lock(&np->tx_lock);
+    spin_lock_irq(&np->tx_lock);
+    spin_lock(&np->rx_lock);
 
     /* Recovery procedure: */
 
@@ -664,8 +705,8 @@ printk(KERN_ALERT"Netfront recovered tx=%d rxfree=%d\n",
     if ( np->user_state == UST_OPEN )
         netif_start_queue(dev);
 
-    spin_unlock(&np->tx_lock);
-    spin_unlock_irq(&np->rx_lock);
+    spin_unlock(&np->rx_lock);
+    spin_unlock_irq(&np->tx_lock);
 }
 
 static void netif_status_change(netif_fe_interface_status_changed_t *status)

linux-2.6.8.1-xen-sparse/include/linux/skbuff.h

@@ -293,6 +293,8 @@ struct sk_buff {
 
 extern void	       __kfree_skb(struct sk_buff *skb);
 extern struct sk_buff *alloc_skb(unsigned int size, int priority);
+extern struct sk_buff *alloc_skb_from_cache(kmem_cache_t *cp,
+					    unsigned int size, int priority);
 extern void	       kfree_skbmem(struct sk_buff *skb);
 extern struct sk_buff *skb_clone(struct sk_buff *skb, int priority);
 extern struct sk_buff *skb_copy(const struct sk_buff *skb, int priority);