zero-copy.5

实现linux平台下零拷贝技术的软件包。

+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
+ */
+
+#include <linux/spinlock.h>
+#include <linux/list.h>
+#include <linux/skbuff.h>
+
+#include <net/inet_hashtables.h>
+
+static int tcp_sendfile_alloc_data(struct zc_buf *zb);
+static int tcp_sendfile_commit_data(struct zc_buf *zb);
+
+static struct zc_handler zc_tcp_sendfile_handler = {
+	.alloc_data		= &tcp_sendfile_alloc_data,
+	.commit_data		= &tcp_sendfile_commit_data,
+};
+
+static DEFINE_SPINLOCK(zc_lock);
+static LIST_HEAD(zc_list);
+
+int zc_alloc_data(struct zc_buf *zb)
+{
+	struct zc_handler *zh;
+	int err = -ENODEV;
+
+	if (unlikely(zb->size > PAGE_SIZE))
+		return err;
+	
+	rcu_read_lock();
+	list_for_each_entry_rcu(zh, &zc_list, zc_entry) {
+		err = zh->alloc_data(zb);
+		if (!err) {
+			zb->zh = zh;
+			break;
+		}
+	}
+	rcu_read_unlock();
+
+	return err;
+}
+
+int zc_commit_data(struct zc_buf *zb)
+{
+	int err = -EINVAL;
+	
+	if (zb->zh)
+		err = zb->zh->commit_data(zb);
+	
+	return err;
+}
+
+int zc_add_handler(struct zc_handler *h)
+{
+	if (!h->alloc_data || !h->commit_data)
+		return -EINVAL;
+
+	spin_lock(&zc_lock);
+	list_add_rcu(&h->zc_entry, &zc_list);
+	spin_unlock(&zc_lock);
+
+	return 0;
+}
+
+void zc_del_handler(struct zc_handler *h)
+{
+	spin_lock(&zc_lock);
+	list_del_rcu(&h->zc_entry);
+	spin_unlock(&zc_lock);
+
+	synchronize_rcu();
+}
+
+extern struct inet_hashinfo __cacheline_aligned tcp_hashinfo;
+
+static int tcp_sendfile_alloc_data(struct zc_buf *zb)
+{
+	struct ethhdr *eth;
+	struct iphdr *iph;
+	struct tcphdr *tcph;
+	struct sock *sk;
+	int dif, err = -EINVAL;
+	u32 saddr, daddr;
+	u16 sport, dport;
+	
+	if (zb->header_size < sizeof(struct ethhdr) + sizeof(struct iphdr) + sizeof(struct tcphdr)) {
+		printk("%s: wrong size %u, must be %zu.\n", 
+				__func__, zb->header_size, 
+				sizeof(struct ethhdr) + sizeof(struct iphdr) + sizeof(struct tcphdr));
+		goto err_out_exit;
+	}
+
+	eth = zb->header;
+	iph = (struct iphdr *)(eth + 1);
+
+	if (iph->protocol != IPPROTO_TCP)
+		goto err_out_exit;
+	
+	tcph = (struct tcphdr *)(iph + 1);
+
+	dif = 0;
+	
+	saddr = iph->saddr;
+	sport = tcph->source;
+	daddr = iph->daddr;
+	dport = tcph->dest;
+	
+	/*
+	 * I suspect it is not enough to disable BHs,
+	 * since it can be [and is] called from hard IRQ context.
+	 * Must do something with bound devices.
+	 */
+	sk = inet_lookup(&tcp_hashinfo, saddr, sport, daddr, dport, dif);
+
+	if (sk && sk->zc_alloc_data) {
+		zb->priv = sk;
+		err = sk->zc_alloc_data(zb);
+		zb->status = (err)?1:0;
+	}
+#if 0
+	printk("%s: sk=%p, %u.%u.%u.%u:%u -> %u.%u.%u.%u:%u, data=%p, status=%d, err=%d.\n", 
+			__func__, sk, 
+			NIPQUAD(saddr), ntohs(sport),
+			NIPQUAD(daddr), ntohs(dport),
+			zb->skb->data, zb->status, err);
+#endif	
+err_out_exit:
+	return err;
+}
+
+static int tcp_sendfile_commit_data(struct zc_buf *zb)
+{
+	struct sock *sk = zb->priv;
+	int err;
+
+	err = sk->zc_commit_data(zb);
+	
+	printk("%s: commiting data, sk=%p, size=%4u, err=%d.\n", __func__, sk, zb->size, err);
+
+	return err;
+}
+
+int __init zc_add_tcp(void)
+{
+	return zc_add_handler(&zc_tcp_sendfile_handler);
+}
+
+late_initcall(zc_add_tcp);
diff --git a/net/ipv4/tcp_ipv4.c b/net/ipv4/tcp_ipv4.c
--- a/net/ipv4/tcp_ipv4.c
+++ b/net/ipv4/tcp_ipv4.c
@@ -1242,6 +1242,10 @@ int tcp_v4_rcv(struct sk_buff *skb)
 		goto no_tcp_socket;
 
 process:
+
+	if (skb->zerocopy)
+		printk("%s: good zero-copy TCP packet found.\n", __func__);
+
 	if (sk->sk_state == TCP_TIME_WAIT)
 		goto do_time_wait;
 
diff --git a/net/socket.c b/net/socket.c
--- a/net/socket.c
+++ b/net/socket.c
@@ -44,6 +44,7 @@
  *		Tigran Aivazian	:	sys_send(args) calls sys_sendto(args, NULL, 0)
  *		Tigran Aivazian	:	Made listen(2) backlog sanity checks 
  *					protocol-independent
+ *		Evgeniy Polyakov:	Added sock_sendfile().
  *
  *
  *		This program is free software; you can redistribute it and/or
@@ -84,6 +85,10 @@
 #include <linux/compat.h>
 #include <linux/kmod.h>
 #include <linux/audit.h>
+#include <linux/pagevec.h>
+#include <linux/pagemap.h>
+#include <linux/swap.h>
+#include <linux/writeback.h>
 
 #ifdef CONFIG_NET_RADIO
 #include <linux/wireless.h>		/* Note : will define WIRELESS_EXT */
@@ -116,6 +121,7 @@ static ssize_t sock_writev(struct file *
 			  unsigned long count, loff_t *ppos);
 static ssize_t sock_sendpage(struct file *file, struct page *page,
 			     int offset, size_t size, loff_t *ppos, int more);
+ssize_t sock_sendfile(struct file *file, loff_t *ppos, size_t count, read_actor_t actor, void *target);
 
 
 /*
@@ -136,7 +142,8 @@ static struct file_operations socket_fil
 	.fasync =	sock_fasync,
 	.readv =	sock_readv,
 	.writev =	sock_writev,
-	.sendpage =	sock_sendpage
+	.sendpage =	sock_sendpage,
+	.sendfile =	sock_sendfile,
 };
 
 /*
@@ -726,6 +733,343 @@ static ssize_t sock_aio_write(struct kio
 	return __sock_sendmsg(iocb, sock, &x->async_msg, size);
 }
 
+int zc_sock_alloc_data(struct zc_buf *zb)
+{
+	struct sock *sk = zb->priv;
+	unsigned long flags;
+	struct zc_page *zp;
+	int err = -ENODEV, need_wakeup = 0;
+	unsigned int towrite = zb->size;
+	struct sk_buff *skb = zb->skb;
+
+	if (!sk || !sk->zc_page_num || !skb)
+		goto out;
+
+	spin_lock_irqsave(&sk->zc_lock, flags);
+
+	BUG_ON(sk->zc_page_index + 1 > sk->zc_page_num);
+	
+	while (towrite) {
+		zp = &sk->zc_pages[sk->zc_page_index];
+		if (zp->size == zp->used || test_bit(ZC_PAGE_READY, &zp->flags))
+			break;
+		if (zp->size - zp->used < towrite && !zb->move_data)
+			break;
+
+		if (skb_shinfo(skb)->nr_frags >= MAX_SKB_FRAGS) {
+			printk("%s: [%1d] skb has grown too much: zp=%p, page=%p, size=%4u, used=%4u, towrite=%4u.\n",
+					__func__, sk->zc_page_index, 
+					zp, zp->page, zp->size, zp->used, towrite);
+			err = -ENOMEM;
+			break;
+		}
+		/*
+		 * Setup fragment with offset to point to the area where
+		 * we actually can write without overwriting old data.
+		 * Setup fragment size to be equal not to the real data size,
+		 * but size of the area where we actually can write data into.
+		 */
+		skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags, zp->page, zp->page_offset+zp->used, zp->size-zp->used);
+
+		printk("%s: [%1d.%1d] data=%p [%p], size=%4u, used=%4u, towrite=%4u.\n", 
+				__func__, sk->zc_page_index, skb_shinfo(skb)->nr_frags-1,
+				zp->page, page_address(zp->page) + zp->page_offset, 
+				zp->size, zp->used, towrite);
+		
+		if (zb->move_data) {
+			unsigned int sz = min(zp->size - zp->used, towrite);
+
+			err = zb->move_data(zb, sz);
+			if (err <= 0)
+				break;
+
+			if (zp->used + err == zp->size) {
+				need_wakeup = 1;
+				set_bit(ZC_PAGE_READY, &zp->flags);
+				if (++sk->zc_page_index == sk->zc_page_num)
+					sk->zc_page_index = 0;
+			}
+		} else 
+			err = zb->size;
+
+		skb->len	+= err;
+		skb->data_len	+= err;
+		skb->truesize	+= err;
+		
+		towrite 	-= err;
+		zp->used 	+= err;
+
+		err = 0;
+	}
+
+	spin_unlock_irqrestore(&sk->zc_lock, flags);
+out:
+	if (need_wakeup)
+		wake_up(&sk->zc_data_ready);
+	return err;
+}
+
+int zc_sock_commit_data(struct zc_buf *zb)
+{
+	struct sock *sk = zb->priv;
+	unsigned long flags;
+	struct zc_page *zp;
+
+	spin_lock_irqsave(&sk->zc_lock, flags);
+		
+	BUG_ON(sk->zc_page_index + 1 > sk->zc_page_num);
+	
+	zp = &sk->zc_pages[sk->zc_page_index];
+
+	if (unlikely(zb->size != zp->size)) {
+		spin_unlock_irqrestore(&sk->zc_lock, flags);
+		return 1;
+	}
+
+	if (zp->used == zp->size) {
+		set_bit(ZC_PAGE_READY, &zp->flags);
+		if (++sk->zc_page_index == sk->zc_page_num)
+			sk->zc_page_index = 0;
+	}
+	spin_unlock_irqrestore(&sk->zc_lock, flags);
+		
+	wake_up(&sk->zc_data_ready);
+
+	return 0;
+}
+
+extern struct page * __grab_cache_page(struct address_space *mapping, unsigned long index,
+			struct page **cached_page, struct pagevec *lru_pvec);
+
+static int commit_page(struct zc_page *zp, struct file *file, struct address_space *mapping)
+{
+	int err;
+	struct address_space_operations *a_ops = mapping->a_ops;
+
+	flush_dcache_page(zp->page);
+	err = a_ops->commit_write(file, zp->page, zp->page_offset, zp->page_offset+zp->used);
+	unlock_page(zp->page);
+	mark_page_accessed(zp->page);
+	page_cache_release(zp->page);
+	if (zp->cached_page) {
+		page_cache_release(zp->cached_page);
+		zp->cached_page = NULL;
+	}
+	
+	printk("%s: zp=%p, page=%p, page_offset=%u, used=%u, size=%u has been committed: err=%d.\n", 
+			__func__, zp, zp->page, zp->page_offset, zp->used, zp->size, err);
+	
+	if (err < 0)
+		goto err_out_exit;
+
+	balance_dirty_pages_ratelimited(mapping);
+
+err_out_exit:
+	return err;
+}
+
+static int prepare_page(struct zc_page *zp, struct file *file, struct address_space *mapping, 
+		loff_t *ppos, loff_t count, struct pagevec *lru_pvec)
+{
+	unsigned long index;
+	unsigned long page_offset;
+	unsigned long bytes;
+	struct address_space_operations *a_ops = mapping->a_ops;
+	loff_t pos_allocated = *ppos;
+	int err = 0;
+	
+	page_offset = (pos_allocated & (PAGE_CACHE_SIZE -1));
+	index = pos_allocated >> PAGE_CACHE_SHIFT;
+	bytes = PAGE_CACHE_SIZE - page_offset;
+	if (bytes > count)
+		bytes = count;
+	
+	zp->page = __grab_cache_page(mapping, index, &zp->cached_page, lru_pvec);
+	if (!zp->page) {
+		err = -ENOMEM;
+		goto err_out_exit;
+	}
+
+	err = a_ops->prepare_write(file, zp->page, page_offset, page_offset+bytes);
+	if (unlikely(err)) {
+		unlock_page(zp->page);
+		page_cache_release(zp->page);
+		goto err_out_exit;
+	}
+
+	zp->page_offset = page_offset;
+	zp->size = bytes;
+	zp->used = 0;
+	clear_bit(ZC_PAGE_READY, &zp->flags);
+	
+	printk("%s: zp=%p, page=%p, page_offset=%u, used=%u, size=%u has been prepared: err=%d.\n", 
+			__func__, zp, zp->page, zp->page_offset, zp->used, zp->size, err);
+
+	pos_allocated += bytes;
+
+	*ppos = pos_allocated;
+
+err_out_exit:
+	return err;
+}
+
+/*
+ * This should process all socket's related stuff,
+ * for example emit TCP ACKs...
+ * Since zero-copy skb can only have valid header,
+ * this should process that header at skb->data.
+ * skb_copy_datagram_iovec() is changed to not even touch
+ * zero-copied skb.
+ */
+
+static u8 message_buf[PAGE_SIZE];
+
+int receive_message(struct socket *sock)
+{
+	struct msghdr msg;
+	struct kvec iov;
+	int err;
+	
+	sock->sk->sk_allocation |= GFP_NOIO;
+	iov.iov_base = message_buf;
+	iov.iov_len = sizeof(message_buf);
+	msg.msg_name = NULL;
+	msg.msg_namelen = 0;
+	msg.msg_control = NULL;
+	msg.msg_controllen = 0;
+	msg.msg_namelen = 0;
+	msg.msg_flags = MSG_DONTWAIT;
+	
+	err = kernel_recvmsg(sock, &msg, &iov, 1, iov.iov_len, 0);
+
+	printk("%s: kernel_recvmsg returned %d.\n", __func__, err);
+	
+	return (err<0)?err:0;
+}
+
+ssize_t sock_sendfile(struct file *in_file, loff_t *ppos, size_t count, read_actor_t actor, void *target)
+{
+	struct socket *sock;
+	struct sock *sk;
+	int err = 0;
+	size_t written = 0;
+	struct file *file = target;
+	struct address_space *mapping = file->f_mapping;
+	struct inode *inode = mapping->host;
+	loff_t pos, pos_allocated;
+	struct pagevec lru_pvec;
+	unsigned long flags;
+	int pnum_max = 16, i;
+	unsigned int zc_page_index;
+	struct zc_page *zc_pages, *zp;
+
+	if (!count)
+		return 0;
+
+	pos = pos_allocated = *ppos;
+	err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
+	if (err)
+		goto err_out_exit;
+
+	sock = SOCKET_I(in_file->f_dentry->d_inode);
+
+	if (!sock || !sock->sk) {
+		err = -ENODEV;
+		goto err_out_exit;
+	}
+	sk = sock->sk;
+
+	pnum_max = ((count >> PAGE_CACHE_SHIFT) > pnum_max)?pnum_max:(count >> PAGE_CACHE_SHIFT);
+	zc_pages = kzalloc(sizeof(struct zc_page) * pnum_max, GFP_KERNEL);
+	if (!zc_pages) {
+		err = -ENOMEM;
+		goto err_out_exit;
+	}
+
+	pagevec_init(&lru_pvec, 0);
+	
+	err = 0;
+	for (i=0; i<pnum_max; ++i) {
+		zp = &zc_pages[i];
+		
+		err = prepare_page(zp, file, mapping, &pos_allocated, count, &lru_pvec);
+		if (unlikely(err))
+			goto err_out_release_pages;
+	}
+
+	zc_page_index = 0;
+	
+	spin_lock_irqsave(&sk->zc_lock, flags);
+	sk->zc_pages 		= zc_pages;
+	sk->zc_page_num 	= pnum_max;
+	sk->zc_page_index 	= zc_page_index;
+	sk->zc_alloc_data	= &zc_sock_alloc_data;
+	sk->zc_commit_data	= &zc_sock_commit_data;
+	spin_unlock_irqrestore(&sk->zc_lock, flags);
+
+	printk("%s: sk=%p, %d pages have been set up.\n", __func__, sk, pnum_max);
+
+	while (count) {
+		struct zc_page *zp;
+		
+		interruptible_sleep_on(&sk->zc_data_ready);
+		
+		printk("%s: wakeup: zc_page_index=%d, sk->zc_page_index=%d.\n", __func__, zc_page_index, sk->zc_page_index);
+
+		spin_lock_irqsave(&sk->zc_lock, flags);
+		if (zc_page_index == sk->zc_page_index) {
+			spin_unlock_irqrestore(&sk->zc_lock, flags);
+			goto last_check;
+		}
+		spin_unlock_irqrestore(&sk->zc_lock, flags);
+
+		for (i=0; i<pnum_max; ++i) {
+			zp = &zc_pages[i];
+			
+			if (test_bit(ZC_PAGE_READY, &zp->flags)) {
+				printk("%s: checking page %p [%d]: page=%p, flags=%08lx, page_offset=%08x, size=%08x, used=%08x\n", 
+						__func__, zp, i, zp->page, zp->flags, zp->page_offset, zp->size, zp->used);
+				err = commit_page(zp, file, mapping);
+				if (err)
+					goto err_out_release_all_pages;
+				
+				count -= zp->used;
+				written += zp->used;
+				pos += zp->used;
+				
+				zc_page_index++;
+
+				err = prepare_page(zp, file, mapping, &pos_allocated, count, &lru_pvec);
+			}
+		}
+
+		do {
+			err = receive_message(sock);
+		} while (!err);
+
+last_check:
+		if (signal_pending(current))
+			break;
+	}
+	
+	pagevec_lru_add(&lru_pvec);
+
+	*ppos = pos;
+	err = written;
+
+err_out_release_all_pages:
+	i = pnum_max;
+err_out_release_pages:
+	for (--i; i>=0; --i)
+		commit_page(&zc_pages[i], file, mapping);
+
+	kfree(zc_pages);
+
+err_out_exit:
+	
+	return err;
+}
+
 static ssize_t sock_sendpage(struct file *file, struct page *page,
 			     int offset, size_t size, loff_t *ppos, int more)
 {