bitkeeper revision 1.22.1.9 (3e428de1nQwf1QUfwsp1nTiJP4ByhQ)

Zero-copy receive path now works over discontiguous memory with no guest-side pool.

4 files changed, 86 insertions, 32 deletions

diff --git a/xenolinux-2.4.16-sparse/arch/xeno/drivers/network/network.c b/xenolinux-2.4.16-sparse/arch/xeno/drivers/network/network.c
index 9c40ad0066..64f33c10ff 100644
--- a/xenolinux-2.4.16-sparse/arch/xeno/drivers/network/network.c
+++ b/xenolinux-2.4.16-sparse/arch/xeno/drivers/network/network.c
@@ -21,6 +21,7 @@
 #include <linux/skbuff.h>
 #include <linux/init.h>
 
+#include <asm/io.h>
 #include <net/sock.h>
 
 #define NET_TX_IRQ _EVENT_NET_TX
@@ -48,6 +49,8 @@ static void cleanup_module(void);
 
 static struct list_head dev_list;
 
+static unsigned int net_countx;
+
 /*
  * RX RING:   RX_IDX <= rx_cons <= rx_prod
  * TX RING:   TX_IDX <= tx_cons <= tx_prod
@@ -72,7 +75,7 @@ struct net_private
 static int network_open(struct net_device *dev)
 {
     struct net_private *np = dev->priv;
-    int error;
+    int error = 0;
     char *rxlabel, *txlabel;
 
     // This is inevitably not the right way to allocate a couple of static strings.
@@ -179,13 +182,59 @@ static void network_tx_buf_gc(struct net_device *dev)
     spin_unlock_irqrestore(&np->tx_lock, flags);
 }
 
+inline unsigned long get_ppte(unsigned long addr)
+{
+        unsigned long ppte = 0xdeadbeef;
+        pgd_t *pgd; pmd_t *pmd; pte_t *ptep;
+        pgd = pgd_offset_k(addr);
+
+        if (pgd_none(*pgd) || pgd_bad(*pgd)) BUG();
+        
+        pmd = pmd_offset(pgd, addr);
+        if (pmd_none(*pmd)) BUG(); 
+        if (pmd_bad(*pmd)) BUG();
+        
+        ptep = pte_offset(pmd, addr);
+        ppte = (unsigned long)phys_to_machine(virt_to_phys(ptep));
+
+        return ppte;
+}
+/*
+static void validate_free_list(void)
+{
+    unsigned long addr, ppfn, mpfn, mpfn2, flags;
+    struct list_head *i;
+    struct net_page_info *np;
+
+    printk(KERN_ALERT "Walking free pages:\n");
+   
+    spin_lock_irqsave(&net_page_list_lock, flags);
+    
+    list_for_each(i, &net_page_list) 
+    {
+        np = list_entry(i, struct net_page_info, list);
+        addr = np->virt_addr;
+        ppfn = virt_to_phys(addr) >> PAGE_SHIFT;
+        mpfn = get_ppte(addr);
+        mpfn2 = phys_to_machine_mapping[ppfn];
 
+        mpfn = (*(unsigned long *)phys_to_virt(machine_to_phys(mpfn))) >> PAGE_SHIFT;
+        if (mpfn != mpfn2) printk(KERN_ALERT "mpfn %lu != %lu\n", mpfn, mpfn2);
+
+        if (machine_to_phys_mapping[mpfn] != ppfn) printk(KERN_ALERT "ppfn %lu != %lu\n", machine_to_phys_mapping[mpfn], ppfn);
+    }
+
+    spin_unlock_irqrestore(&net_page_list_lock, flags);
+    
+}
+*/
 static void network_alloc_rx_buffers(struct net_device *dev)
 {
     unsigned int i;
     struct net_private *np = dev->priv;
     struct sk_buff *skb;
     unsigned int end = RX_RING_ADD(np->rx_idx, RX_MAX_ENTRIES);
+    
 
     for ( i = np->net_ring->rx_prod; i != end; i = RX_RING_INC(i) )
     {
@@ -194,8 +243,9 @@ static void network_alloc_rx_buffers(struct net_device *dev)
         skb->dev = dev;
         //skb_reserve(skb, 2); /* word align the IP header */
         np->rx_skb_ring[i] = skb;
-        np->net_ring->rx_ring[i].addr = (unsigned long)skb->net_page->ppte; //data;
+        np->net_ring->rx_ring[i].addr = get_ppte(skb->head); 
         np->net_ring->rx_ring[i].size = RX_BUF_SIZE - 16; /* arbitrary */
+//printk(KERN_ALERT "[%p]\n", phys_to_machine(virt_to_phys(skb->page_ptr)));
     }
 
     np->net_ring->rx_prod = i;
@@ -273,6 +323,8 @@ static void network_rx_int(int irq, void *dev_id, struct pt_regs *ptregs)
     struct net_private *np = dev->priv;
     struct sk_buff *skb;
     
+    /*if (net_countx++ % 100 == 0) validate_free_list();*/
+    
  again:
     for ( i = np->rx_idx; i != np->net_ring->rx_cons; i = RX_RING_INC(i) )
     {
@@ -283,7 +335,14 @@ static void network_rx_int(int irq, void *dev_id, struct pt_regs *ptregs)
                 continue;
         }
         skb = np->rx_skb_ring[i];
-        
+
+//printk(KERN_ALERT "[%u]: ptmm[%lx] old:(%lx) new:(%lx)\n", i , virt_to_phys(skb->head) >> PAGE_SHIFT, phys_to_machine_mapping[virt_to_phys(skb->head) >> PAGE_SHIFT], (*(unsigned long *)phys_to_virt(machine_to_phys(np->net_ring->rx_ring[i].addr))) >> PAGE_SHIFT);
+
+        phys_to_machine_mapping[virt_to_phys(skb->head) >> PAGE_SHIFT] =
+            (*(unsigned long *)phys_to_virt(
+                    machine_to_phys(np->net_ring->rx_ring[i].addr))
+             ) >> PAGE_SHIFT;
+
         skb_put(skb, np->net_ring->rx_ring[i].size);
         skb->protocol = eth_type_trans(skb, dev);
         np->stats.rx_packets++;
diff --git a/xenolinux-2.4.16-sparse/include/asm-xeno/io.h b/xenolinux-2.4.16-sparse/include/asm-xeno/io.h
index 0f097342ba..1afc7e4d68 100644
--- a/xenolinux-2.4.16-sparse/include/asm-xeno/io.h
+++ b/xenolinux-2.4.16-sparse/include/asm-xeno/io.h
@@ -78,17 +78,6 @@ static inline void * phys_to_virt(unsigned long address)
  * These are equally trivial.
  */
 
-static inline unsigned long virt_to_mach(volatile void * address)
-{
-       return __pa(address) + (unsigned long) start_info.phys_base;
-}
-
-static inline void *mach_to_virt(unsigned long address)
-{
-        return __va(address) - (unsigned long) start_info.phys_base;
-}
-
-
 /*
  * Change "struct page" to physical address.
  */
diff --git a/xenolinux-2.4.16-sparse/include/linux/skbuff.h b/xenolinux-2.4.16-sparse/include/linux/skbuff.h
index 2d0a94dce2..54963090f6 100644
--- a/xenolinux-2.4.16-sparse/include/linux/skbuff.h
+++ b/xenolinux-2.4.16-sparse/include/linux/skbuff.h
@@ -43,7 +43,8 @@
 #define SKB_ZERO_COPY       1
 
 #define NUM_NET_PAGES       9 // about 1Meg of buffers. (2^9)
-struct net_page_info {
+
+/*struct net_page_info {
         struct list_head list;
         unsigned long   virt_addr;
         unsigned long   ppte;
@@ -54,7 +55,7 @@ extern struct net_page_info *net_page_table;
 extern struct list_head net_page_list;
 extern spinlock_t net_page_list_lock;
 extern unsigned int net_pages;
-
+*/
 /* End zero copy additions */
 
 #define HAVE_ALLOC_SKB		/* For the drivers to know */
diff --git a/xenolinux-2.4.16-sparse/net/core/skbuff.c b/xenolinux-2.4.16-sparse/net/core/skbuff.c
index ec76f00de0..45332f1ffd 100644
--- a/xenolinux-2.4.16-sparse/net/core/skbuff.c
+++ b/xenolinux-2.4.16-sparse/net/core/skbuff.c
@@ -64,12 +64,13 @@
 #include <asm/system.h>
 
 /* zc globals: */
+/*
 char *net_page_chunk;
 struct net_page_info *net_page_table;
 struct list_head net_page_list;
 spinlock_t net_page_list_lock = SPIN_LOCK_UNLOCKED;
 unsigned int net_pages;
-
+*/
 
 
 int sysctl_hot_list_len = 128;
@@ -229,7 +230,7 @@ nohead:
 }
 
 /* begin zc code additions: */
-
+/*
 void init_net_pages(unsigned long order_pages)
 {
         int i;
@@ -248,16 +249,16 @@ void init_net_pages(unsigned long order_pages)
                 np->virt_addr = (unsigned long)net_page_chunk + (i * PAGE_SIZE);
 
                 // now fill the pte pointer:
-                np->ppte = 0xdeadbeef;
-                pgd = pgd_offset_k(np->virt_addr);
-                if (pgd_none(*pgd) || pgd_bad(*pgd)) BUG();
+                //np->ppte = 0xdeadbeef;
+                //pgd = pgd_offset_k(np->virt_addr);
+                //if (pgd_none(*pgd) || pgd_bad(*pgd)) BUG();
 
-                if (pmd_none(*pmd)) BUG(); 
-                if (pmd_bad(*pmd)) BUG();
-
-                ptep = pte_offset(pmd, np->virt_addr);
-                np->ppte = (unsigned long)virt_to_mach(ptep);
+                //if (pmd_none(*pmd)) BUG(); 
+                //if (pmd_bad(*pmd)) BUG();
 
+                //ptep = pte_offset(pmd, np->virt_addr);
+                //np->ppte = phys_to_machine(virt_to_phys(ptep));
+                
                 list_add_tail(&np->list, &net_page_list);
         }
         net_pages = nr_pages;
@@ -267,6 +268,7 @@ void init_net_pages(unsigned long order_pages)
 
 struct net_page_info *get_net_page(void)
 {
+
     struct list_head *list_ptr;
     struct net_page_info *np;
     unsigned long flags;
@@ -301,7 +303,7 @@ void free_net_page(struct net_page_info *np)
     spin_unlock_irqrestore(&net_page_list_lock, flags);
 
 }
-
+*/
 struct sk_buff *alloc_zc_skb(unsigned int size,int gfp_mask)
 {
 	struct sk_buff *skb;
@@ -332,12 +334,13 @@ struct sk_buff *alloc_zc_skb(unsigned int size,int gfp_mask)
                 printk("alloc_zc_skb called with unruly size.\n");
                 size = PAGE_SIZE;
         }
-	skb->net_page = get_net_page();
+	/*skb->net_page = get_net_page();
         if (skb->net_page == NULL)
         {
                 goto nodata;
         }
-        data = (u8 *)skb->net_page->virt_addr;
+        data = (u8 *)skb->net_page->virt_addr;*/
+        data = (char *)__get_free_page(gfp_mask);
 	if (data == NULL)
 		goto nodata;
 	/* XXX: does not include slab overhead */ 
@@ -443,7 +446,9 @@ static void skb_release_data(struct sk_buff *skb)
                 {
 		    kfree(skb->head);
                 } else {// SKB_ZERO_COPY
-                    free_net_page(skb->net_page);
+                    //free_net_page(skb->net_page);
+//printk(KERN_ALERT "<%p>\n", phys_to_machine(virt_to_phys(skb->head)));
+                    free_page((void *)skb->head);
                 }
 	}
 
@@ -559,7 +564,7 @@ struct sk_buff *skb_clone(struct sk_buff *skb, int gfp_mask)
 	C(tc_index);
 #endif
         C(skb_type);
-        C(net_page);
+        //C(net_page);
 	atomic_inc(&(skb_shinfo(skb)->dataref));
 	skb->cloned = 1;
 #ifdef CONFIG_NETFILTER
@@ -1361,7 +1366,7 @@ void __init skb_init(void)
 	if (!skbuff_head_cache)
 		panic("cannot create skbuff cache");
 
-        init_net_pages(NUM_NET_PAGES);
+        //init_net_pages(NUM_NET_PAGES);
 
 	for (i=0; i<NR_CPUS; i++)
 		skb_queue_head_init(&skb_head_pool[i].list);