net: use a per task frag allocator

We currently use a per socket order-0 page cache for tcp_sendmsg()
operations.

This page is used to build fragments for skbs.

Its done to increase probability of coalescing small write() into
single segments in skbs still in write queue (not yet sent)

But it wastes a lot of memory for applications handling many mostly
idle sockets, since each socket holds one page in sk->sk_sndmsg_page

Its also quite inefficient to build TSO 64KB packets, because we need
about 16 pages per skb on arches where PAGE_SIZE = 4096, so we hit
page allocator more than wanted.

This patch adds a per task frag allocator and uses bigger pages,
if available. An automatic fallback is done in case of memory pressure.

(up to 32768 bytes per frag, thats order-3 pages on x86)

This increases TCP stream performance by 20% on loopback device,
but also benefits on other network devices, since 8x less frags are
mapped on transmit and unmapped on tx completion. Alexander Duyck
mentioned a probable performance win on systems with IOMMU enabled.

Its possible some SG enabled hardware cant cope with bigger fragments,
but their ndo_start_xmit() should already handle this, splitting a
fragment in sub fragments, since some arches have PAGE_SIZE=65536

Successfully tested on various ethernet devices.
(ixgbe, igb, bnx2x, tg3, mellanox mlx4)

Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Ben Hutchings <bhutchings@solarflare.com>
Cc: Vijay Subramanian <subramanian.vijay@gmail.com>
Cc: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Vijay Subramanian <subramanian.vijay@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

include/linux/sched.h

@@ -1530,6 +1530,9 @@ struct task_struct {
 	 * cache last used pipe for splice
 	 */
 	struct pipe_inode_info *splice_pipe;
+
+	struct page_frag task_frag;
+
 #ifdef	CONFIG_TASK_DELAY_ACCT
 	struct task_delay_info *delays;
 #endif

include/net/inet_sock.h

@@ -101,10 +101,8 @@ struct inet_cork {
 	__be32			addr;
 	struct ip_options	*opt;
 	unsigned int		fragsize;
-	struct dst_entry	*dst;
 	int			length; /* Total length of all frames */
-	struct page		*page;
-	u32			off;
+	struct dst_entry	*dst;
 	u8			tx_flags;
 };
 

include/net/sock.h

@@ -247,8 +247,7 @@ struct cg_proto;
   *	@sk_stamp: time stamp of last packet received
   *	@sk_socket: Identd and reporting IO signals
   *	@sk_user_data: RPC layer private data
-  *	@sk_sndmsg_page: cached page for sendmsg
-  *	@sk_sndmsg_off: cached offset for sendmsg
+  *	@sk_frag: cached page frag
   *	@sk_peek_off: current peek_offset value
   *	@sk_send_head: front of stuff to transmit
   *	@sk_security: used by security modules
@@ -362,9 +361,8 @@ struct sock {
 	ktime_t			sk_stamp;
 	struct socket		*sk_socket;
 	void			*sk_user_data;
-	struct page		*sk_sndmsg_page;
+	struct page_frag	sk_frag;
 	struct sk_buff		*sk_send_head;
-	__u32			sk_sndmsg_off;
 	__s32			sk_peek_off;
 	int			sk_write_pending;
 #ifdef CONFIG_SECURITY
@@ -2034,18 +2032,23 @@ static inline void sk_stream_moderate_sndbuf(struct sock *sk)
 
 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp);
 
-static inline struct page *sk_stream_alloc_page(struct sock *sk)
+/**
+ * sk_page_frag - return an appropriate page_frag
+ * @sk: socket
+ *
+ * If socket allocation mode allows current thread to sleep, it means its
+ * safe to use the per task page_frag instead of the per socket one.
+ */
+static inline struct page_frag *sk_page_frag(struct sock *sk)
 {
-	struct page *page = NULL;
+	if (sk->sk_allocation & __GFP_WAIT)
+		return &current->task_frag;
 
-	page = alloc_pages(sk->sk_allocation, 0);
-	if (!page) {
-		sk_enter_memory_pressure(sk);
-		sk_stream_moderate_sndbuf(sk);
-	}
-	return page;
+	return &sk->sk_frag;
 }
 
+extern bool sk_page_frag_refill(struct sock *sk, struct page_frag *pfrag);
+
 /*
  *	Default write policy as shown to user space via poll/select/SIGIO
  */

kernel/exit.c

@@ -1046,6 +1046,9 @@ void do_exit(long code)
 	if (tsk->splice_pipe)
 		__free_pipe_info(tsk->splice_pipe);
 
+	if (tsk->task_frag.page)
+		put_page(tsk->task_frag.page);
+
 	validate_creds_for_do_exit(tsk);
 
 	preempt_disable();

kernel/fork.c

@@ -330,6 +330,7 @@ static struct task_struct *dup_task_struct(struct task_struct *orig)
 	tsk->btrace_seq = 0;
 #endif
 	tsk->splice_pipe = NULL;
+	tsk->task_frag.page = NULL;
 
 	account_kernel_stack(ti, 1);
 

net/core/skbuff.c

@@ -1655,38 +1655,19 @@ static struct page *linear_to_page(struct page *page, unsigned int *len,
 				   unsigned int *offset,
 				   struct sk_buff *skb, struct sock *sk)
 {
-	struct page *p = sk->sk_sndmsg_page;
-	unsigned int off;
+	struct page_frag *pfrag = sk_page_frag(sk);
 
-	if (!p) {
-new_page:
-		p = sk->sk_sndmsg_page = alloc_pages(sk->sk_allocation, 0);
-		if (!p)
-			return NULL;
-
-		off = sk->sk_sndmsg_off = 0;
-		/* hold one ref to this page until it's full */
-	} else {
-		unsigned int mlen;
-
-		/* If we are the only user of the page, we can reset offset */
-		if (page_count(p) == 1)
-			sk->sk_sndmsg_off = 0;
-		off = sk->sk_sndmsg_off;
-		mlen = PAGE_SIZE - off;
-		if (mlen < 64 && mlen < *len) {
-			put_page(p);
-			goto new_page;
-		}
+	if (!sk_page_frag_refill(sk, pfrag))
+		return NULL;
 
-		*len = min_t(unsigned int, *len, mlen);
-	}
+	*len = min_t(unsigned int, *len, pfrag->size - pfrag->offset);
 
-	memcpy(page_address(p) + off, page_address(page) + *offset, *len);
-	sk->sk_sndmsg_off += *len;
-	*offset = off;
+	memcpy(page_address(pfrag->page) + pfrag->offset,
+	       page_address(page) + *offset, *len);
+	*offset = pfrag->offset;
+	pfrag->offset += *len;
 
-	return p;
+	return pfrag->page;
 }
 
 static bool spd_can_coalesce(const struct splice_pipe_desc *spd,

net/core/sock.c

@@ -1744,6 +1744,45 @@ struct sk_buff *sock_alloc_send_skb(struct sock *sk, unsigned long size,
 }
 EXPORT_SYMBOL(sock_alloc_send_skb);
 
+/* On 32bit arches, an skb frag is limited to 2^15 */
+#define SKB_FRAG_PAGE_ORDER	get_order(32768)
+
+bool sk_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
+{
+	int order;
+
+	if (pfrag->page) {
+		if (atomic_read(&pfrag->page->_count) == 1) {
+			pfrag->offset = 0;
+			return true;
+		}
+		if (pfrag->offset < pfrag->size)
+			return true;
+		put_page(pfrag->page);
+	}
+
+	/* We restrict high order allocations to users that can afford to wait */
+	order = (sk->sk_allocation & __GFP_WAIT) ? SKB_FRAG_PAGE_ORDER : 0;
+
+	do {
+		gfp_t gfp = sk->sk_allocation;
+
+		if (order)
+			gfp |= __GFP_COMP | __GFP_NOWARN;
+		pfrag->page = alloc_pages(gfp, order);
+		if (likely(pfrag->page)) {
+			pfrag->offset = 0;
+			pfrag->size = PAGE_SIZE << order;
+			return true;
+		}
+	} while (--order >= 0);
+
+	sk_enter_memory_pressure(sk);
+	sk_stream_moderate_sndbuf(sk);
+	return false;
+}
+EXPORT_SYMBOL(sk_page_frag_refill);
+
 static void __lock_sock(struct sock *sk)
 	__releases(&sk->sk_lock.slock)
 	__acquires(&sk->sk_lock.slock)
@@ -2173,8 +2212,8 @@ void sock_init_data(struct socket *sock, struct sock *sk)
 	sk->sk_error_report	=	sock_def_error_report;
 	sk->sk_destruct		=	sock_def_destruct;
 
-	sk->sk_sndmsg_page	=	NULL;
-	sk->sk_sndmsg_off	=	0;
+	sk->sk_frag.page	=	NULL;
+	sk->sk_frag.offset	=	0;
 	sk->sk_peek_off		=	-1;
 
 	sk->sk_peer_pid 	=	NULL;
@@ -2417,6 +2456,12 @@ void sk_common_release(struct sock *sk)
 	xfrm_sk_free_policy(sk);
 
 	sk_refcnt_debug_release(sk);
+
+	if (sk->sk_frag.page) {
+		put_page(sk->sk_frag.page);
+		sk->sk_frag.page = NULL;
+	}
+
 	sock_put(sk);
 }
 EXPORT_SYMBOL(sk_common_release);

net/ipv4/ip_output.c

@@ -793,6 +793,7 @@ static int __ip_append_data(struct sock *sk,
 			    struct flowi4 *fl4,
 			    struct sk_buff_head *queue,
 			    struct inet_cork *cork,
+			    struct page_frag *pfrag,
 			    int getfrag(void *from, char *to, int offset,
 					int len, int odd, struct sk_buff *skb),
 			    void *from, int length, int transhdrlen,
@@ -987,47 +988,30 @@ static int __ip_append_data(struct sock *sk,
 			}
 		} else {
 			int i = skb_shinfo(skb)->nr_frags;
-			skb_frag_t *frag = &skb_shinfo(skb)->frags[i-1];
-			struct page *page = cork->page;
-			int off = cork->off;
-			unsigned int left;
-
-			if (page && (left = PAGE_SIZE - off) > 0) {
-				if (copy >= left)
-					copy = left;
-				if (page != skb_frag_page(frag)) {
-					if (i == MAX_SKB_FRAGS) {
-						err = -EMSGSIZE;
-						goto error;
-					}
-					skb_fill_page_desc(skb, i, page, off, 0);
-					skb_frag_ref(skb, i);
-					frag = &skb_shinfo(skb)->frags[i];
-				}
-			} else if (i < MAX_SKB_FRAGS) {
-				if (copy > PAGE_SIZE)
-					copy = PAGE_SIZE;
-				page = alloc_pages(sk->sk_allocation, 0);
-				if (page == NULL)  {
-					err = -ENOMEM;
-					goto error;
-				}
-				cork->page = page;
-				cork->off = 0;
 
-				skb_fill_page_desc(skb, i, page, 0, 0);
-				frag = &skb_shinfo(skb)->frags[i];
-			} else {
-				err = -EMSGSIZE;
-				goto error;
-			}
-			if (getfrag(from, skb_frag_address(frag)+skb_frag_size(frag),
-				    offset, copy, skb->len, skb) < 0) {
-				err = -EFAULT;
+			err = -ENOMEM;
+			if (!sk_page_frag_refill(sk, pfrag))
 				goto error;
+
+			if (!skb_can_coalesce(skb, i, pfrag->page,
+					      pfrag->offset)) {
+				err = -EMSGSIZE;
+				if (i == MAX_SKB_FRAGS)
+					goto error;
+
+				__skb_fill_page_desc(skb, i, pfrag->page,
+						     pfrag->offset, 0);
+				skb_shinfo(skb)->nr_frags = ++i;
+				get_page(pfrag->page);
 			}
-			cork->off += copy;
-			skb_frag_size_add(frag, copy);
+			copy = min_t(int, copy, pfrag->size - pfrag->offset);
+			if (getfrag(from,
+				    page_address(pfrag->page) + pfrag->offset,
+				    offset, copy, skb->len, skb) < 0)
+				goto error_efault;
+
+			pfrag->offset += copy;
+			skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
 			skb->len += copy;
 			skb->data_len += copy;
 			skb->truesize += copy;
@@ -1039,6 +1023,8 @@ static int __ip_append_data(struct sock *sk,
 
 	return 0;
 
+error_efault:
+	err = -EFAULT;
 error:
 	cork->length -= length;
 	IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
@@ -1079,8 +1065,6 @@ static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
 	cork->dst = &rt->dst;
 	cork->length = 0;
 	cork->tx_flags = ipc->tx_flags;
-	cork->page = NULL;
-	cork->off = 0;
 
 	return 0;
 }
@@ -1117,7 +1101,8 @@ int ip_append_data(struct sock *sk, struct flowi4 *fl4,
 		transhdrlen = 0;
 	}
 
-	return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base, getfrag,
+	return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
+				sk_page_frag(sk), getfrag,
 				from, length, transhdrlen, flags);
 }
 
@@ -1439,7 +1424,8 @@ struct sk_buff *ip_make_skb(struct sock *sk,
 	if (err)
 		return ERR_PTR(err);
 
-	err = __ip_append_data(sk, fl4, &queue, &cork, getfrag,
+	err = __ip_append_data(sk, fl4, &queue, &cork,
+			       &current->task_frag, getfrag,
 			       from, length, transhdrlen, flags);
 	if (err) {
 		__ip_flush_pending_frames(sk, &queue, &cork);

net/ipv4/raw.c

@@ -131,18 +131,23 @@ static struct sock *__raw_v4_lookup(struct net *net, struct sock *sk,
  *	0 - deliver
  *	1 - block
  */
-static __inline__ int icmp_filter(struct sock *sk, struct sk_buff *skb)
+static int icmp_filter(const struct sock *sk, const struct sk_buff *skb)
 {
-	int type;
-
-	if (!pskb_may_pull(skb, sizeof(struct icmphdr)))
+	struct icmphdr _hdr;
+	const struct icmphdr *hdr;
+
+	pr_err("icmp_filter skb_transport_offset %d data-head %ld len %d/%d\n",
+		skb_transport_offset(skb), skb->data - skb->head, skb->len, skb->data_len);
+	hdr = skb_header_pointer(skb, skb_transport_offset(skb),
+				 sizeof(_hdr), &_hdr);
+	pr_err("head %p data %p hdr %p type %d\n", skb->head, skb->data, hdr, hdr ? hdr->type : -1);
+	if (!hdr)
 		return 1;
 
-	type = icmp_hdr(skb)->type;
-	if (type < 32) {
+	if (hdr->type < 32) {
 		__u32 data = raw_sk(sk)->filter.data;
 
-		return ((1 << type) & data) != 0;
+		return ((1U << hdr->type) & data) != 0;
 	}
 
 	/* Do not block unknown ICMP types */