zero-copy.8

实现linux平台下零拷贝技术的软件包。

+}
+
+static int tcp_udp_v4_sendfile_cleanup(struct zc_handler *zh, struct socket *sock, struct sock_zc_setup_data *p)
+{
+	int err;
+
+	err = tcp_udp_v4_sendfile_check(zh, sock, p);
+	if (err)
+		return err;
+	
+	sk_zc_fini(sock->sk->zsk);
+	return 0;
+}
+
+int __init zc_add_tcp(void)
+{
+	int i;
+
+	for (i=0; i<tcp_udp_v4_zc_handler.sock_bucket_number; ++i) {
+		INIT_LIST_HEAD(&tcp_udp_v4_zc_handler.sock_bucket[i].list);
+		rwlock_init(&tcp_udp_v4_zc_handler.sock_bucket[i].lock);
+	}
+
+	atomic_set(&tcp_udp_v4_zc_handler.refcnt, 1);
+
+	return zc_add_handler(&tcp_udp_v4_zc_handler);
+}
+
+late_initcall(zc_add_tcp);
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:	Receiving zero-copy.
  *
  *
  *		This program is free software; you can redistribute it and/or
@@ -63,6 +64,7 @@
 #include <linux/smp_lock.h>
 #include <linux/socket.h>
 #include <linux/file.h>
+#include <linux/fs.h>
 #include <linux/net.h>
 #include <linux/interrupt.h>
 #include <linux/netdevice.h>
@@ -84,6 +86,11 @@
 #include <linux/compat.h>
 #include <linux/kmod.h>
 #include <linux/audit.h>
+#include <linux/pagemap.h>
+#include <linux/swap.h>
+#include <linux/writeback.h>
+#include <linux/ip.h>
+#include <linux/tcp.h>
 
 #ifdef CONFIG_NET_RADIO
 #include <linux/wireless.h>		/* Note : will define WIRELESS_EXT */
@@ -116,6 +123,7 @@ 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);
+static ssize_t sock_sendfile(struct file *file, loff_t *ppos, size_t count, read_actor_t actor, void *target);
 
 
 /*
@@ -136,7 +144,8 @@ 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 +735,372 @@ 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 zsock *zsk = zb->priv;
+	struct zc_page *zp;
+	int err = -ENODEV;
+	unsigned int towrite = zb->size;
+	struct sk_buff *skb = zb->skb;
+	struct ethhdr *eth;
+	struct iphdr *ip;
+	int index;
+
+	if (!zsk->zc_pages)
+		goto out;
+
+	eth = (struct ethhdr *)zb->header;
+	ip = (struct iphdr *)(eth+1);
+
+	if (ip->protocol == IPPROTO_TCP) {
+		struct tcphdr *th = (struct tcphdr *)(((u8 *)ip) + ip->ihl*4);
+		index = (ntohl(th->seq) - zsk->zc_seq_first) / zsk->zc_page_num;
+	} else if (ip->protocol == IPPROTO_UDP) {
+		index = zsk->zc_page_index;
+	} else
+		goto out;
+
+	if (index >= zsk->zc_page_num) {
+		zsk->zc_seq_first = zsk->zc_pages[0].seq;
+		index = 0;
+	}
+
+	while (towrite) {
+		zp = &zsk->zc_pages[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) {
+			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, size=%4u, used=%4u, towrite=%4u.\n", 
+				__func__, zsk->zc_page_index, skb_shinfo(skb)->nr_frags-1,
+				zp->page, 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) {
+				set_bit(ZC_PAGE_READY, &zp->flags);
+				
+				if (++zsk->zc_page_index == zsk->zc_page_num)
+					zsk->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;
+	}
+
+out:
+	return err;
+}
+
+int zc_sock_commit_data(struct zc_buf *zb)
+{
+	struct zsock *zsk = zb->priv;
+	struct zc_page *zp;
+
+	if (!zsk->zc_pages)
+		return -1;
+
+	zp = &zsk->zc_pages[zsk->zc_page_index];
+
+	if (unlikely(zb->size != zp->size))
+		return 1;
+
+	if (zp->used == zp->size) {
+		set_bit(ZC_PAGE_READY, &zp->flags);
+		if (++zsk->zc_page_index == zsk->zc_page_num)
+			zsk->zc_page_index = 0;
+	}
+
+	return 0;
+}
+
+/*
+ * 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];
+
+static int receive_message(struct socket *sock, unsigned int ack_size)
+{
+	struct msghdr msg;
+	struct kvec iov;
+	int err;
+
+	sock->sk->sk_allocation |= GFP_NOIO;
+	iov.iov_base = message_buf;
+	iov.iov_len = min(ack_size, (unsigned int)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_NOSIGNAL | MSG_DONTWAIT;
+
+	err = kernel_recvmsg(sock, &msg, &iov, 1, iov.iov_len, msg.msg_flags);
+
+	printk("%s: kernel_recvmsg returned %d, ack_size=%u.\n", __func__, err, ack_size);
+
+	return err;
+}
+
+int tcp_udp_v4_sock_zc_init(struct socket *sock, struct tcp_udp_v4_priv *priv)
+{
+	struct file *file;
+	struct zsock *zsk;
+	struct zc_page *zc_pages, *zp;
+	int pnum_max, err, i;
+	unsigned long flags;
+	struct address_space *mapping;
+	struct inode *inode;
+	size_t count;
+
+	/*
+	 * Sane setup.
+	 */
+	count = INT_MAX;
+	pnum_max = 16;
+
+	if (!sock->sk)
+		return -EINVAL;
+
+	err = -EBADF;
+	file = fget(priv->fd);
+	if (!file)
+		goto err_out_exit;
+	if (!(file->f_mode & FMODE_WRITE))
+		goto err_out_fput;
+	err = -ETXTBSY;
+	if (file->f_mode & FMODE_ZEROCOPY)
+		goto err_out_fput;
+	err = -EINVAL;
+	if (!file->f_op)
+		goto err_out_fput;
+
+	err = rw_verify_area(WRITE, file, &file->f_pos, count);
+	if (err)
+		goto err_out_fput;
+
+	err = security_file_permission(file, MAY_WRITE);
+	if (err)
+		goto err_out_fput;
+
+	err = -ENOMEM;
+	zsk = zsk_alloc(&tcp_udp_v4_zc_handler, priv, sizeof(*priv), NULL, GFP_KERNEL);
+	if (!zsk)
+		goto err_out_fput;
+
+	mapping = file->f_mapping;
+	inode = mapping->host;
+	
+	zc_pages = kzalloc(sizeof(struct zc_page) * pnum_max, GFP_KERNEL);
+	if (!zc_pages) {
+		err = -ENOMEM;
+		goto err_out_zsk_put;
+	}
+
+	pagevec_init(&zsk->zc_lru_pvec, 0);
+
+	err = 0;
+	for (i=0; i<pnum_max; ++i) {
+		zp = &zc_pages[i];
+
+		err = prepare_page(zp, zsk, file, mapping, &zsk->zc_pos, count, &zsk->zc_lru_pvec);
+		if (unlikely(err))
+			goto err_out_commit_pages;
+	}
+
+	file->f_mode |= FMODE_ZEROCOPY;
+
+	spin_lock_irqsave(&zsk->zc_lock, flags);
+	zsk->zc_file		= file;
+	zsk->zc_pages 		= zc_pages;
+	zsk->zc_page_num 	= pnum_max;
+	zsk->zc_page_index 	= 0;
+	zsk->zc_alloc_data	= &zc_sock_alloc_data;
+	zsk->zc_commit_data	= &zc_sock_commit_data;
+	spin_unlock_irqrestore(&zsk->zc_lock, flags);
+
+	err = tcp_udp_v4_zc_sock_insert(zsk);
+	if (err) {
+		i = pnum_max;
+		goto err_out_commit_pages;
+	}
+	
+	sock->sk->zsk = zsk;
+
+	printk("%s: sk=%p, zsk=%p, %d pages have been set up.\n", __func__, sock->sk, zsk, pnum_max);
+
+	return 0;
+
+err_out_commit_pages:
+	for (--i; i>=0; --i)
+		commit_page(&zc_pages[i], file, mapping);
+
+	kfree(zc_pages);
+err_out_zsk_put:
+	zsk_put(zsk);
+err_out_fput:
+	file->f_mode &= ~FMODE_ZEROCOPY;
+	fput(file);
+err_out_exit:
+	return err;
+}
+
+int sock_zc_setup_seq(struct zsock *zsk, u32 seq)
+{
+	unsigned long flags;
+	u32 off = 0;
+
+	spin_lock_irqsave(&zsk->zc_lock, flags);
+	zsk->zc_seq_first = seq;
+	
+	if (zsk->zc_pages) {
+		int i;
+		struct zc_page *zp;
+
+		for (i=0; i<zsk->zc_page_num; ++i) {
+			zp = &zsk->zc_pages[i];
+
+			zp->seq = zsk->zc_seq_first + off;
+			off += zp->size;
+		}
+	}
+	spin_unlock_irqrestore(&zsk->zc_lock, flags);
+	
+	printk("%s: zc_seq_first=%u, off=%u.\n", __func__, zsk->zc_seq_first, off);
+
+	return 0;
+}
+
+
+static 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;
+	unsigned long flags;
+	int i;
+	unsigned int ack_size, zc_page_index = 0;
+	struct zsock *zsk;
+
+	if (!count)
+		return 0;
+
+	sock = SOCKET_I(in_file->f_dentry->d_inode);
+
+	if (!sock || !sock->sk || !sock->sk->zsk) {
+		err = -ENODEV;
+		goto err_out_exit;
+	}
+	sk = sock->sk;
+	zsk = sk->zsk;
+
+	err = generic_write_checks(file, &zsk->zc_pos, &count, S_ISBLK(inode->i_mode));
+	if (err)
+		goto err_out_exit;
+
+	spin_lock_irqsave(&zsk->zc_lock, flags);
+	if (!zsk->zc_pages) {
+		spin_unlock_irqrestore(&zsk->zc_lock, flags);
+		err = -EINVAL;
+		goto err_out_exit;
+	}
+	spin_unlock_irqrestore(&zsk->zc_lock, flags);
+
+	while (count) {
+		struct zc_page *zp;
+
+		wait_event_interruptible_timeout(zsk->zc_data_ready, zsk->zc_page_index != zc_page_index, 5*HZ);
+		zc_page_index = zsk->zc_page_index;
+
+		printk("%s: wakeup: zsk->zc_page_index=%d.\n", __func__, zsk->zc_page_index);
+
+		ack_size = 0;
+		for (i=0; i<zsk->zc_page_num; ++i) {
+			zp = &zsk->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, written=%zx.\n", 
+						__func__, zp, i, zp->page, zp->flags, zp->page_offset, zp->size, zp->used, written);
+
+				err = commit_page(zp, file, mapping);
+				if (err)
+					goto err_out_release_all_pages;
+
+				count -= zp->used;
+				written += zp->used;
+				ack_size += zp->used;
+
+				err = prepare_page(zp, zsk, file, mapping, &zsk->zc_pos, count, &zsk->zc_lru_pvec);
+			}
+		}
+
+		printk("%s: Going to ack %u bytes.\n", __func__, ack_size);
+
+		while (ack_size > 0) {
+			err = receive_message(sock, ack_size);
+			if (err > 0) {
+				ack_size -= err;
+			} else
+				break;
+		}
+
+		if (signal_pending(current))
+			break;
+	}
+
+	pagevec_lru_add(&zsk->zc_lru_pvec);
+
+	*ppos = written;
+	err = written;
+
+err_out_release_all_pages:
+	printk("%s: Releasing zero-copy socket zsk=%p, sk=%p.\n", __func__, zsk, sk);
+
+	synchronize_rcu();
+	sk_zc_fini(zsk);
+	sk->zsk = NULL;
+
+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)
 {