zero-copy.8

实现linux平台下零拷贝技术的软件包。

+
+int zc_commit_data(struct zc_buf *zb)
+{
+	int err = -EINVAL;
+	
+	if (zb->zh)
+		err = zb->zh->commit_data(zb->zh, zb);
+	
+	return err;
+}
+
+void zc_cleanup(struct socket *sock, void *data, unsigned int size)
+{
+	struct sock_zc_setup_data *p = data;
+	int found = 0;
+	struct zc_handler *zh;
+
+	if (size <= sizeof(struct sock_zc_setup_data) || 
+			size != htonl(p->size) + sizeof(struct sock_zc_setup_data)) {
+		goto err_out_exit;
+	}
+
+	down(&zc_handler_lock);
+	list_for_each_entry(zh, &zc_handler_list, zc_entry) {
+		if (!zh->cleanup(zh, sock, p)) {
+			found = 1;
+			zc_handler_put(zh);
+			break;
+		}
+	}
+	up(&zc_handler_lock);
+
+err_out_exit:
+	return;
+
+}
+
+int zc_setup(struct socket *sock, void *data, unsigned int size)
+{
+	struct sock_zc_setup_data *p = data;
+	int found = 0;
+	struct zc_handler *zh;
+
+	if (size <= sizeof(struct sock_zc_setup_data) || 
+			size != htonl(p->size) + sizeof(struct sock_zc_setup_data)) {
+		goto err_out_exit;
+	}
+
+	down(&zc_handler_lock);
+	list_for_each_entry(zh, &zc_handler_list, zc_entry) {
+		if (!zh->setup(zh, sock, p)) {
+			found = 1;
+			break;
+		}
+	}
+	up(&zc_handler_lock);
+
+err_out_exit:
+	return (found)?0:-ENODEV;
+}
+
+int zc_add_handler(struct zc_handler *h)
+{
+	if (!h->alloc_data || !h->commit_data || !h->sock_bucket || !h->sock_bucket_number || 
+			!h->setup || !h->cleanup)
+		return -EINVAL;
+	
+	synchronize_rcu();
+
+	down(&zc_handler_lock);
+	list_add_rcu(&h->zc_entry, &zc_handler_list);
+	up(&zc_handler_lock);
+
+	return 0;
+}
+
+void zc_del_handler(struct zc_handler *h)
+{
+	synchronize_rcu();
+	
+	down(&zc_handler_lock);
+	list_del_rcu(&h->zc_entry);
+	up(&zc_handler_lock);
+}
+
+extern struct page * __grab_cache_page(struct address_space *mapping, unsigned long index,
+			struct page **cached_page, struct pagevec *lru_pvec);
+
+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);
+
+	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;
+}
+
+int prepare_page(struct zc_page *zp, struct zsock *zsk, 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, &zsk->zc_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;
+	zp->seq = zsk->zc_seq_first + pos_allocated;
+	clear_bit(ZC_PAGE_READY, &zp->flags);
+
+	printk("%s: zp=%p, seq=%u, page=%p, page_offset=%u, used=%u, size=%u has been prepared: err=%d.\n", 
+			__func__, zp, zp->seq, zp->page, zp->page_offset, zp->used, zp->size, err);
+
+	pos_allocated += bytes;
+
+	*ppos = pos_allocated;
+
+err_out_exit:
+	return err;
+}
+
+
+void sk_zc_fini(struct zsock *zsk)
+{
+	if (zsk) {
+		unsigned int zc_page_num;
+		struct zc_page *zc_pages;
+		unsigned long flags;
+		
+		spin_lock_irqsave(&zsk->zc_lock, flags);
+		zc_page_num 		= zsk->zc_page_num;
+		zc_pages 		= zsk->zc_pages;
+		
+		zsk->zc_pages 		= NULL;
+		zsk->zc_page_num 	= 0;
+		zsk->zc_page_index 	= 0;
+		zsk->zc_alloc_data	= NULL;
+		zsk->zc_commit_data	= NULL;
+		spin_unlock_irqrestore(&zsk->zc_lock, flags);
+
+		if (zc_page_num) {
+			struct address_space *mapping = zsk->zc_file->f_mapping;
+			int i;
+
+			printk("%s: zsk=%p, zc_page_num=%u, zc_pages=%p, refcnt=%d.\n", 
+					__func__, zsk, zc_page_num, zc_pages, atomic_read(&zsk->refcnt));
+			
+			/*
+			 * No new skbs can contribute data into VFS cache after this 
+			 * condition, so we only must care about those which are 
+			 * in socket queue already or will be inserted there after
+			 * allocation, but allocation itself will always fail
+			 * due to above locked changes.
+			 */
+
+			if (zsk->zc_cached_page) {
+				page_cache_release(zsk->zc_cached_page);
+				zsk->zc_cached_page = NULL;
+			}
+
+			for (i=0; i<zc_page_num; ++i)
+				commit_page(&zc_pages[i], zsk->zc_file, mapping);
+
+			zsk->zc_file->f_mode &= ~FMODE_ZEROCOPY;
+			fput(zsk->zc_file);
+			zsk->zc_file = NULL;
+			
+			kfree(zc_pages);
+			printk("%s: sk=%p has been released.\n", __func__, zsk);
+
+			zsk_put(zsk);
+		}
+	}
+}
+
+void sk_zc_init(struct zsock *zsk)
+{
+	spin_lock_init(&zsk->zc_lock);
+	init_waitqueue_head(&zsk->zc_data_ready);
+	zsk->zc_pages		= NULL;
+	zsk->zc_page_num	= 0;
+	zsk->zc_page_index	= 0;
+	zsk->zc_alloc_data	= NULL;
+	zsk->zc_commit_data	= NULL;
+	zsk->zc_file		= NULL;
+	zsk->zc_cached_page	= NULL;
+}
+
+struct zsock *zsk_alloc(struct zc_handler *handler, void *priv, unsigned int priv_size, int (* insert)(struct zsock *zsk), gfp_t gfp_mask)
+{
+	struct zsock *zsk;
+
+	zsk = kzalloc(sizeof(struct zsock) + priv_size, gfp_mask);
+	if (!zsk)
+		return NULL;
+
+	atomic_set(&zsk->refcnt, 1);
+	zsk->handler 	= handler;
+	zsk->priv_size 	= priv_size;
+	if (priv_size) {
+		zsk->priv = zsk+1;
+		memcpy(zsk->priv, priv, priv_size);
+	} else
+		zsk->priv = NULL;
+
+	zc_handler_get(handler);
+	
+	sk_zc_init(zsk);
+
+	if (insert) {
+		int err;
+
+		err = insert(zsk);
+		if (err) {
+			zc_handler_put(handler);
+			zsk_free(zsk);
+			return NULL;
+		}
+	}
+
+	return zsk;
+}
+
+void zsk_free(struct zsock *zsk)
+{
+	zc_handler_put(zsk->handler);
+	kfree(zsk);
+}
+
+static inline u32 tcp_udp_v4_hash(unsigned int bucket_number, const u32 src, const u16 sport, const u32 dst, const u16 dport)
+{
+	return inet_ehashfn(src, sport, dst, dport) & (bucket_number - 1);
+}
+
+int tcp_udp_v4_zc_sock_insert(struct zsock *zsk)
+{
+	u32 hash;
+	unsigned long flags;
+	struct tcp_udp_v4_priv *priv = zsk_priv(zsk);
+	struct zc_sock_bucket *b;
+
+	if (!priv)
+		return -ENODEV;
+	
+	hash = tcp_udp_v4_hash(zsk->handler->sock_bucket_number, priv->src, priv->sport, priv->dst, priv->dport);
+
+	b = &zsk->handler->sock_bucket[hash];
+
+	printk("%s: hash=%x, b=%p.\n", __func__, hash, b);
+
+	write_lock_irqsave(&b->lock, flags);
+	list_add_rcu(&zsk->zc_entry, &b->list);
+	write_unlock_irqrestore(&b->lock, flags);
+
+	printk("%s: done.\n", __func__);
+
+	return 0;
+}
+
+int tcp_udp_v4_zc_sock_remove(struct zsock *zsk)
+{
+	u32 hash;
+	unsigned long flags;
+	struct tcp_udp_v4_priv *priv = zsk_priv(zsk);
+	struct zc_sock_bucket *b;
+
+	if (!priv)
+		return -ENODEV;
+	
+	hash = tcp_udp_v4_hash(zsk->handler->sock_bucket_number, priv->src, priv->sport, priv->dst, priv->dport);
+
+	b = &zsk->handler->sock_bucket[hash];
+
+	write_lock_irqsave(&b->lock, flags);
+	list_del_rcu(&zsk->zc_entry);
+	write_unlock_irqrestore(&b->lock, flags);
+
+	return 0;
+}
+
+/*
+ * Must be called under RCU cover and with interrupts disabled. 
+ */
+static struct zsock *tcp_udp_v4_zc_sock_lookup(const struct zc_sock_bucket *bucket, const unsigned int bucket_number, 
+		const u32 src, const u16 sport, const u32 dst, const u16 dport)
+{
+	u32 hash = tcp_udp_v4_hash(bucket_number, src, sport, dst, dport);
+	struct zsock *zsk;
+	struct tcp_udp_v4_priv *priv;
+
+	printk("%s: hash=%08x: %u.%u.%u.%u:%u -> %u.%u.%u.%u:%u.\n", 
+			__func__, hash, NIPQUAD(src), htons(sport), NIPQUAD(dst), htons(dport));
+	
+	list_for_each_entry_rcu(zsk, &bucket[hash].list, zc_entry) {
+		priv = zsk_priv(zsk);
+
+		/* May not happen for this kind of zc sockets, actually... */
+		if (unlikely(!priv))
+			continue;
+
+		printk("%s: zsk=%p: %u.%u.%u.%u:%u -> %u.%u.%u.%u:%u.\n", 
+			__func__, zsk, NIPQUAD(priv->src), htons(priv->sport), NIPQUAD(priv->dst), htons(priv->dport));
+
+		if (priv->sport == sport && priv->dport == dport && priv->src == src && priv->dst == dst) {
+			zsk_get(zsk);
+			return zsk;
+		}
+	}
+
+	return NULL;
+}
+
+static int tcp_udp_v4_sendfile_alloc_data(struct zc_handler *zh, struct zc_buf *zb)
+{
+	struct ethhdr *eth;
+	struct iphdr *iph;
+	struct zsock *zsk;
+	int err = -EINVAL;
+	u16 sport, dport;
+
+	if (zb->header_size < sizeof(struct ethhdr) + sizeof(struct iphdr))
+		goto err_out_exit;
+
+	eth = zb->header;
+
+	if (eth->h_proto != htons(ETH_P_IP))
+		goto err_out_exit;
+
+	iph = (struct iphdr *)(eth + 1);
+	
+	//if (iph->protocol != IPPROTO_TCP || iph->protocol != IPPROTO_UDP)
+	if (iph->protocol != IPPROTO_TCP)
+		goto err_out_exit;
+
+	if (iph->protocol == IPPROTO_TCP) {
+		struct tcphdr *tcph = (struct tcphdr *)(((u8 *)iph) + iph->ihl*4);
+		sport = tcph->source;
+		dport = tcph->dest;
+	} else {
+		struct udphdr *udph = (struct udphdr *)(((u8 *)iph) + iph->ihl*4);
+		sport = udph->source;
+		dport = udph->dest;
+	}
+
+	local_irq_disable();
+	rcu_read_lock();
+	zsk = tcp_udp_v4_zc_sock_lookup(zh->sock_bucket, zh->sock_bucket_number, iph->daddr, dport, iph->saddr, sport);
+	if (zsk) {
+		printk("%s: zsk=%p, zc_alloc_data=%p, refcnt=%d.\n", __func__, zsk, zsk->zc_alloc_data, atomic_read(&zsk->refcnt));
+		spin_lock(&zsk->zc_lock);
+		if (zsk->zc_alloc_data && zsk->zc_pages) {
+			zb->priv = zsk;
+			err = zsk->zc_alloc_data(zb);
+			zb->status = (err)?1:0;
+			wake_up(&zsk->zc_data_ready);
+		}
+		spin_unlock(&zsk->zc_lock);
+		zsk_put(zsk);
+	}
+	rcu_read_unlock();
+	local_irq_enable();
+
+err_out_exit:
+	return err;
+}
+
+static int tcp_udp_v4_sendfile_commit_data(struct zc_handler *zh, struct zc_buf *zb)
+{
+	struct zsock *zsk = zb->priv;
+	int err;
+	unsigned long flags;
+
+	spin_lock_irqsave(&zsk->zc_lock, flags);
+	err = zsk->zc_commit_data(zb);
+	spin_unlock_irqrestore(&zsk->zc_lock, flags);
+
+	wake_up(&zsk->zc_data_ready);
+
+	printk("%s: commiting data, zsk=%p, size=%4u, err=%d.\n", __func__, zsk, zb->size, err);
+
+	return err;
+}
+
+static int tcp_udp_v4_sendfile_check(struct zc_handler *zh, struct socket *sock, struct sock_zc_setup_data *p)
+{
+	struct tcp_udp_v4_priv *priv;
+	u32 type = ntohl(p->type);
+	u32 size = ntohl(p->size);
+
+	if (type != IPPROTO_TCP && type != IPPROTO_UDP)
+		return -EINVAL;
+
+	if (size != sizeof(struct tcp_udp_v4_priv))
+		return -EINVAL;
+
+	priv = (struct tcp_udp_v4_priv *)p->data;
+
+	printk("%s: %u.%u.%u.%u:%u -> %u.%u.%u.%u:%u.\n", 
+			__func__, NIPQUAD(priv->src), htons(priv->sport),
+			NIPQUAD(priv->dst), htons(priv->dport));
+
+	return 0;
+}
+
+static int tcp_udp_v4_sendfile_setup(struct zc_handler *zh, struct socket *sock, struct sock_zc_setup_data *p)
+{
+	struct tcp_udp_v4_priv *priv = (struct tcp_udp_v4_priv *)p->data;
+	int err;
+
+	err = tcp_udp_v4_sendfile_check(zh, sock, p);
+	if (err)
+		return err;
+
+	printk("%s: %u.%u.%u.%u:%u -> %u.%u.%u.%u:%u.\n", 
+			__func__, NIPQUAD(priv->src), htons(priv->sport),
+			NIPQUAD(priv->dst), htons(priv->dport));
+
+	return tcp_udp_v4_sock_zc_init(sock, priv);