zero-copy.9

实现linux平台下零拷贝技术的软件包。

 }
 
+static int __sock_zc_setup_seq(struct zsock *zsk, u32 seq)
+{
+	u32 off = 0;
+
+	printk("%s: setting seq=%u.\n", __func__, seq);
+	
+	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;
+		}
+	}
+
+	return 0;
+}
+
+int sock_zc_setup_seq(struct zsock *zsk, u32 seq)
+{
+	unsigned long flags;
+	int err;
+
+	write_lock_irqsave(&zsk->zc_lock, flags);
+	err = __sock_zc_setup_seq(zsk, seq);
+	write_unlock_irqrestore(&zsk->zc_lock, flags);
+
+	return err;
+}
+
+/*
+ * It's a tricky way to lose your data...
+ * To avoid this we must setup sequnce number
+ * using sock_zc_setup_seq() from internals of TCP state machine.
+ */
+static int zc_calc_index(struct zsock *zsk, u32 seq)
+{
+	int index, try = 0;
+	struct zc_page *zp;
+	
+	if (seq < zsk->zc_seq_first) {
+		__sock_zc_setup_seq(zsk, seq);
+		return 0;
+	}
+
+	index = (seq - zsk->zc_seq_first) / PAGE_SIZE;
+	
+	/*
+	 * Ugly? Yes, but modulo operation is too expensive on ppc32.
+	 */
+	while (index >= zsk->zc_page_num)
+		index -= zsk->zc_page_num;
+
+	while (index > 0 && index < zsk->zc_page_num && try < zsk->zc_page_num) {
+		zp = &zsk->zc_pages[index];
+		if (seq >= zp->seq && seq <= zp->seq + zp->size)
+			return index;
+		else if (seq > zp->seq + zp->size)
+			index++;
+		else if (seq < zp->seq)
+			index--;
+
+		try++;
+	}
+
+	if (index == zsk->zc_page_num || try == zsk->zc_page_num) {
+		__sock_zc_setup_seq(zsk, seq);
+		return 0;
+	}
+
+	return index;
+}
+
+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;
+	u32 seq, ack=0;
+
+	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);
+
+		seq = ntohl(th->seq);
+		ack = ntohl(th->ack_seq);
+
+		/*
+		 * Is it possible to be here from two different pathes?
+		 * This means that skb_alloc_zerocopy() is called from different IRQ handlers
+		 * on different CPUs simultaneously for the same zero-copy socket.
+		 *
+		 * If so, then sequence number setup must be done under write lock being held
+		 * using sock_zc_setup_seq() from internals of TCP state machine.
+		 */
+		if (!zsk->zc_seq_first)
+			__sock_zc_setup_seq(zsk, seq);
+	} else if (ip->protocol == IPPROTO_UDP) {
+		seq = zsk->zc_page_num * zsk->zc_page_index + zsk->zc_seq_first;
+	} else
+		goto out;
+
+	while (towrite) {
+		unsigned int sz;
+		
+		index = zc_calc_index(zsk, seq);
+
+		printk("%s: zsk=%p, towrite=%u, seq=%u, ack=%u, zc_seq_first=%u, index=%d.\n",
+				__func__, zsk, towrite, seq, ack, zsk->zc_seq_first, index);
+		
+		if (index >= zsk->zc_page_num || index < 0)
+			break;
+		
+		zp = &zsk->zc_pages[index];
+
+		printk("%s: zp=%p, zp->used=%u, zp->size=%u, zp->seq=%u, seq=%u.\n", 
+				__func__, zp, zp->used, zp->size, zp->seq, seq);
+
+		if (test_bit(ZC_PAGE_READY, &zp->flags))
+			break;
+
+		spin_lock(&zp->lock);
+		
+		err = -ENODEV;
+		if ((zp->size == zp->used) || (zp->size - zp->used < towrite && !zb->move_data))
+			goto unlock;
+
+		if (unlikely(skb_shinfo(skb)->nr_frags >= MAX_SKB_FRAGS)) {
+			err = -ENOMEM;
+			goto unlock;
+		}
+
+		if (ip->protocol == IPPROTO_TCP && zp->seq + zp->used > seq) {
+			towrite = 0;
+			err = 0;
+			goto unlock;
+		}
+		
+		sz = min(zp->size - zp->used, towrite);
+		
+		/*
+		 * 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+seq-zp->seq, sz);
+
+		if (zb->move_data) {
+			err = zb->move_data(zb, sz);
+			if (err <= 0)
+				goto unlock;
+
+			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;
+
+		seq += err;
+
+		printk("%s: written=%u, zp=%p, index=%d, seq=%u.\n", __func__, err, zp, index, seq);
+
+		err = 0;
+
+unlock:
+		spin_unlock(&zp->lock);
+		if (err < 0)
+			break;
+	}
+
+	if (towrite != zb->size)
+		set_bit(ZSK_DATA_READY, &zsk->zc_flags);
+
+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;
+	}
+		
+	set_bit(ZSK_DATA_READY, &zsk->zc_flags);
+
+	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: ack_size=%u, ret=%d.\n", __func__, ack_size, err);
+
+	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];
+
+		spin_lock_init(&zp->lock);
+		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;
+
+	write_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;
+	zsk->sk			= sock->sk;
+	write_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;
+
+	write_lock_irqsave(&zsk->zc_lock, flags);
+	zsk->zc_file		= NULL;
+	zsk->zc_pages 		= 0;
+	zsk->zc_page_num 	= 0;
+	zsk->zc_page_index 	= 0;
+	zsk->zc_alloc_data	= NULL;
+	zsk->zc_commit_data	= NULL;
+	zsk->sk			= NULL;
+	write_unlock_irqrestore(&zsk->zc_lock, flags);
+
+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;
+}
+
+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;
+	int i;
+	unsigned int ack_size;
+	struct zsock *zsk;
+
+	if (!count)
+		return 0;
+
+	sock = SOCKET_I(in_file->f_dentry->d_inode);
+
+	if (!sock || !sock->sk || !sock->sk->zsk)
+		return -ENODEV;
+	sk = sock->sk;
+	zsk = sk->zsk;
+	sk->zsk = NULL;
+
+	err = generic_write_checks(file, &zsk->zc_pos, &count, S_ISBLK(inode->i_mode));
+	if (err)
+		goto err_out_exit;
+
+	if (!zsk->zc_pages) {
+		err = -EINVAL;
+		goto err_out_exit;
+	}
+
+	zsk_get(zsk);
+
+	while (count) {
+		struct zc_page *zp;
+
+		wait_event_interruptible_timeout(zsk->zc_data_ready, test_and_clear_bit(ZSK_DATA_READY, &zsk->zc_flags), 5*HZ);
+
+		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: zsk=%p: committing page %p [%d]: page=%p, flags=%08lx, page_offset=%08x, size=%08x, used=%08x, written=%zx.\n", 
+						__func__, zsk, 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);
+			}
+		}
+#if 0
+		if (!ack_size)
+			ack_size = zxk->zc_page_num * sizeof(message_buf);
+#endif
+		while ((err = receive_message(sock, zsk->zc_page_num * sizeof(message_buf))) > 0) {
+			ack_size -= err;
+		}
+
+		if (signal_pending(current))
+			break;
+	}
+
+	pagevec_lru_add(&zsk->zc_lru_pvec);
+
+	*ppos = written;
+	err = written;
+
+	printk("%s: written=%Lu.\n", __func__, written);
+
+err_out_release_all_pages:
+
+err_out_exit:
+	sk_zc_fini(zsk);
+	zc_cleanup(zsk);
+	return err;
+}
+
 static ssize_t sock_sendpage(struct file *file, struct page *page,
 			     int offset, size_t size, loff_t *ppos, int more)
 {