[PATCH net-next v4 af-packet 0/2] Enhance af-packet to provide (near zero)lossless packet capture functionality.

[PATCH net-next v4 af-packet 0/2] Enhance af-packet to provide (near zero)lossless packet capture functionality.

[Chetan Loke]

Changes in v4:
1) Used ALIGN macro                                     (Joe Perches)
2) Deleted duplicate field                              (Eric Dumazet)
3) Re-aligned tpacket fields for disk-capture

Changes in v3:
1) Stripped __packed__ attribute.                       (Dave Miller)
   Replaced with aligned_u64 and padding.
2) Added 'feature_request_word'.
3) Added rx_hash field to the v3-header.                (Chetan L)

Changes in v2:

1) Aligned bdqc members, pr_err to WARN, sob email      (Joe Perches)
2) Added tp_padding                                     (Eric Dumazet)
3) Nuked useless ;) white space                         (Stephen H)
4) Use __u types in headers                             (Ben Hutchings)
5) Added field for creating private area                (Chetan Loke)

This patch attempts to:
1)Improve network capture visibility by increasing packet density
2)Assist in analyzing multiple(aggregated) capture ports.

Benefits:
  B1) ~15-20% reduction in cpu-usage.
  B2) ~20% increase in packet capture rate.
  B3) ~2x  increase in packet density.
  B4) Port aggregation analysis.
  B5) Non static frame size to capture entire packet payload.

With the current af_packet->rx::mmap based approach, the element size
in the block needs to be statically configured. Nothing wrong with this
config/implementation. But the traffic profile cannot be known in advance.
And so it would be nice if that configuration wasn't static. Normally,
one would configure the element-size to be '2048' so that you can atleast
capture the entire 'MTU-size'.But if the traffic profile varies then we
would end up either i)wasting memory or ii) end up getting a sliced frame.
In other words the packet density will be much less in the first case.

Detailed description of the test-setup etc can be viewed at:
http://thread.gmane.org/gmane.linux.kernel/1158216


----------------------------------------

 include/linux/if_packet.h |  119 ++++++
 net/packet/af_packet.c    |  947 ++++++++++++++++++++++++++++++++++++++++++---
 2 files changed, 1020 insertions(+), 46 deletions(-)

-- 
1.7.5.2

[PATCH net-next v4 af-packet 1/2] Enhance af-packet to provide (near zero)lossless packet capture functionality.

[Chetan Loke]

Signed-off-by: Chetan Loke <loke.chetan@gmail.com>
---
 include/linux/if_packet.h |  119 +++++++++++++++++++++++++++++++++++++++++++++
 1 files changed, 119 insertions(+), 0 deletions(-)

diff --git a/include/linux/if_packet.h b/include/linux/if_packet.h
index c148606..5926d59 100644
--- a/include/linux/if_packet.h
+++ b/include/linux/if_packet.h
@@ -61,6 +61,17 @@ struct tpacket_stats {
 	unsigned int	tp_drops;
 };
 
+struct tpacket_stats_v3 {
+	unsigned int	tp_packets;
+	unsigned int	tp_drops;
+	unsigned int	tp_freeze_q_cnt;
+};
+
+union tpacket_stats_u {
+	struct tpacket_stats stats1;
+	struct tpacket_stats_v3 stats3;
+};
+
 struct tpacket_auxdata {
 	__u32		tp_status;
 	__u32		tp_len;
@@ -78,6 +89,7 @@ struct tpacket_auxdata {
 #define TP_STATUS_LOSING	0x4
 #define TP_STATUS_CSUMNOTREADY	0x8
 #define TP_STATUS_VLAN_VALID   0x10 /* auxdata has valid tp_vlan_tci */
+#define TP_STATUS_BLK_TMO	0x20
 
 /* Tx ring - header status */
 #define TP_STATUS_AVAILABLE	0x0
@@ -85,6 +97,9 @@ struct tpacket_auxdata {
 #define TP_STATUS_SENDING	0x2
 #define TP_STATUS_WRONG_FORMAT	0x4
 
+/* Rx ring - feature request bits */
+#define TP_FT_REQ_FILL_RXHASH	0x1
+
 struct tpacket_hdr {
 	unsigned long	tp_status;
 	unsigned int	tp_len;
@@ -111,11 +126,100 @@ struct tpacket2_hdr {
 	__u16		tp_padding;
 };
 
+struct hdr_variant1 {
+	__u32	tp_rxhash;
+	__u32	tp_vlan_tci;
+};
+
+struct tpacket3_hdr {
+	__u32		tp_next_offset;
+	__u32		tp_sec;
+	__u32		tp_nsec;
+	__u32		tp_snaplen;
+	__u32		tp_len;
+	__u32		tp_status;
+	__u16		tp_mac;
+	__u16		tp_net;
+	/* pkt_hdr variants */
+	union {
+		struct hdr_variant1 hv1;
+	};
+};
+
+struct bd_ts {
+	unsigned int ts_sec;
+	union {
+		unsigned int ts_usec;
+		unsigned int ts_nsec;
+	};
+};
+
+struct hdr_v1 {
+	__u32	block_status;
+	__u32	num_pkts;
+	__u32	offset_to_first_pkt;
+
+	/* Number of valid bytes (including padding)
+	 * blk_len <= tp_block_size
+	 */
+	__u32	blk_len;
+
+	/*
+	 * Quite a few uses of sequence number:
+	 * 1. Make sure cache flush etc worked.
+	 *    Well, one can argue - why not use the increasing ts below?
+	 *    But look at 2. below first.
+	 * 2. When you pass around blocks to other user space decoders,
+	 *    you can see which blk[s] is[are] outstanding etc.
+	 * 3. Validate kernel code.
+	 */
+	aligned_u64	seq_num;
+
+	/*
+	 * ts_last_pkt:
+	 *
+	 * Case 1.	Block has 'N'(N >=1) packets and TMO'd(timed out)
+	 *		ts_last_pkt == 'time-stamp of last packet' and NOT the
+	 *		time when the timer fired and the block was closed.
+	 *		By providing the ts of the last packet we can absolutely
+	 *		guarantee that time-stamp wise, the first packet in the
+	 *		next block will never precede the last packet of the
+	 *		previous block.
+	 * Case 2.	Block has zero packets and TMO'd
+	 *		ts_last_pkt = time when the timer fired and the block
+	 *		was closed.
+	 * Case 3.	Block has 'N' packets and NO TMO.
+	 *		ts_last_pkt = time-stamp of the last pkt in the block.
+	 *
+	 * ts_first_pkt:
+	 *		Is always the time-stamp when the block was opened.
+	 *		Case a)	ZERO packets
+	 *			No packets to deal with but atleast you know the
+	 *			time-interval of this block.
+	 *		Case b) Non-zero packets
+	 *			Use the ts of the first packet in the block.
+	 *
+	 */
+	struct bd_ts	ts_first_pkt, ts_last_pkt;
+};
+
+union bd_header_u {
+	struct hdr_v1 bh1;
+};
+
+struct block_desc {
+	__u32 version;
+	__u32 offset_to_priv;
+	union bd_header_u hdr;
+};
+
 #define TPACKET2_HDRLEN		(TPACKET_ALIGN(sizeof(struct tpacket2_hdr)) + sizeof(struct sockaddr_ll))
+#define TPACKET3_HDRLEN		(TPACKET_ALIGN(sizeof(struct tpacket3_hdr)) + sizeof(struct sockaddr_ll))
 
 enum tpacket_versions {
 	TPACKET_V1,
 	TPACKET_V2,
+	TPACKET_V3
 };
 
 /*
@@ -138,6 +242,21 @@ struct tpacket_req {
 	unsigned int	tp_frame_nr;	/* Total number of frames */
 };
 
+struct tpacket_req3 {
+	unsigned int	tp_block_size;	/* Minimal size of contiguous block */
+	unsigned int	tp_block_nr;	/* Number of blocks */
+	unsigned int	tp_frame_size;	/* Size of frame */
+	unsigned int	tp_frame_nr;	/* Total number of frames */
+	unsigned int	tp_retire_blk_tov; /* timeout in msecs */
+	unsigned int	tp_sizeof_priv; /* offset to private data area */
+	unsigned int	tp_feature_req_word;
+};
+
+union tpacket_req_u {
+	struct tpacket_req	req;
+	struct tpacket_req3	req3;
+};
+
 struct packet_mreq {
 	int		mr_ifindex;
 	unsigned short	mr_type;
-- 
1.7.5.2

[PATCH net-next v4 af-packet 2/2] Enhance af-packet to provide (near zero)lossless packet capture functionality.

[Chetan Loke]

1) Blocks can now be configured with non-static frame format.
   Non-static frame format provides following benefits:
   1.1) Increases packet density by a factor of 2x.
   1.2) Ability to capture entire packet.
   1.3) Captures 99% 64-byte traffic as seen by the kernel.
2) Read/poll is now at a block-level rather than at packet level.
3) Added user-configurable timeout knob for timing out blocks on slow/bursty links.
4) Block level processing now allows monitoring multiple links as a single
   logical pipe.

Changes:
C1) tpacket_rcv()
    C1.1) packet_current_frame() is replaced by packet_current_rx_frame()
          The bulk of the processing is then moved in the following chain:
          packet_current_rx_frame()
            __packet_lookup_frame_in_block
              fill_curr_block()
              or
                retire_current_block
                dispatch_next_block
              or
              return NULL(queue is plugged/paused)


Signed-off-by: Chetan Loke <loke.chetan@gmail.com>
---
 net/packet/af_packet.c |  947 +++++++++++++++++++++++++++++++++++++++++++++---
 1 files changed, 901 insertions(+), 46 deletions(-)

diff --git a/net/packet/af_packet.c b/net/packet/af_packet.c
index c698cec..5fa19f0 100644
--- a/net/packet/af_packet.c
+++ b/net/packet/af_packet.c
@@ -40,6 +40,10 @@
  *					byte arrays at the end of sockaddr_ll
  *					and packet_mreq.
  *		Johann Baudy	:	Added TX RING.
+ *		Chetan Loke	:	Implemented TPACKET_V3 block abstraction
+ *					layer.
+ *					Copyright (C) 2011, <lokec@ccs.neu.edu>
+ *
  *
  *		This program is free software; you can redistribute it and/or
  *		modify it under the terms of the GNU General Public License
@@ -161,9 +165,61 @@ struct packet_mreq_max {
 	unsigned char	mr_address[MAX_ADDR_LEN];
 };
 
-static int packet_set_ring(struct sock *sk, struct tpacket_req *req,
+static int packet_set_ring(struct sock *sk, union tpacket_req_u *req_u,
 		int closing, int tx_ring);
 
+
+#define V3_ALIGNMENT	(8)
+
+#define BLK_HDR_LEN	(ALIGN(sizeof(struct block_desc), V3_ALIGNMENT))
+
+#define BLK_PLUS_PRIV(sz_of_priv) \
+	(BLK_HDR_LEN + ALIGN((sz_of_priv), V3_ALIGNMENT))
+
+struct kbdq_ft_ops {
+	int num_ops;
+	void (*ft_ops[2])(void *, void *);
+};
+
+/* kbdq - kernel block descriptor queue */
+struct kbdq_core {
+	struct pgv	*pkbdq;
+	unsigned int	hdrlen;
+	unsigned char	reset_pending_on_curr_blk;
+	unsigned char   delete_blk_timer;
+	unsigned short	kactive_blk_num;
+	unsigned short	blk_sizeof_priv;
+
+	/* last_kactive_blk_num:
+	 * trick to see if user-space has caught up
+	 * in order to avoid refreshing timer when every single pkt arrives.
+	 */
+	unsigned short	last_kactive_blk_num;
+
+	char		*pkblk_start;
+	char		*pkblk_end;
+	int		kblk_size;
+	unsigned int	knum_blocks;
+	uint64_t	knxt_seq_num;
+	char		*prev;
+	char		*nxt_offset;
+	struct sk_buff	*skb;
+	struct kbdq_ft_ops kfops;
+
+	atomic_t	blk_fill_in_prog;
+
+	/* Default is set to 8ms */
+#define DEFAULT_PRB_RETIRE_TOV	(8)
+
+	unsigned short  retire_blk_tov;
+	unsigned short  version;
+	unsigned long	tov_in_jiffies;
+
+	/* timer to retire an outstanding block */
+	struct timer_list retire_blk_timer;
+};
+
+#define PGV_FROM_VMALLOC 1
 struct pgv {
 	char *buffer;
 };
@@ -179,12 +235,40 @@ struct packet_ring_buffer {
 	unsigned int		pg_vec_pages;
 	unsigned int		pg_vec_len;
 
+	struct kbdq_core	prb_bdqc;
 	atomic_t		pending;
 };
 
+#define BLOCK_STATUS(x)	((x)->hdr.bh1.block_status)
+#define BLOCK_NUM_PKTS(x)	((x)->hdr.bh1.num_pkts)
+#define BLOCK_O2FP(x)		((x)->hdr.bh1.offset_to_first_pkt)
+#define BLOCK_LEN(x)		((x)->hdr.bh1.blk_len)
+#define BLOCK_SNUM(x)		((x)->hdr.bh1.seq_num)
+#define BLOCK_O2PRIV(x)	((x)->offset_to_priv)
+#define BLOCK_PRIV(x)		((void *)((char *)(x) + BLOCK_O2PRIV(x)))
+
 struct packet_sock;
 static int tpacket_snd(struct packet_sock *po, struct msghdr *msg);
 
+static void *packet_previous_frame(struct packet_sock *po,
+		struct packet_ring_buffer *rb,
+		int status);
+static void packet_increment_head(struct packet_ring_buffer *buff);
+static int prb_curr_blk_in_use(struct kbdq_core *,
+			struct block_desc *);
+static void *prb_dispatch_next_block(struct kbdq_core *,
+			struct packet_sock *);
+static void prb_retire_current_block(struct kbdq_core *,
+		struct packet_sock *, unsigned int status);
+static int prb_queue_frozen(struct kbdq_core *);
+static void prb_open_block(struct kbdq_core *, struct block_desc *);
+static void prb_retire_rx_blk_timer_expired(unsigned long);
+static void _prb_refresh_rx_retire_blk_timer(struct kbdq_core *);
+static void prb_init_blk_timer(struct packet_sock *, struct kbdq_core *,
+				void (*func) (unsigned long));
+static void prb_fill_rxhash(struct kbdq_core *, struct tpacket3_hdr *);
+static void prb_clear_rxhash(struct kbdq_core *, struct tpacket3_hdr *);
+static void prb_fill_vlan_info(struct kbdq_core *, struct tpacket3_hdr *);
 static void packet_flush_mclist(struct sock *sk);
 
 struct packet_fanout;
@@ -193,6 +277,7 @@ struct packet_sock {
 	struct sock		sk;
 	struct packet_fanout	*fanout;
 	struct tpacket_stats	stats;
+	union  tpacket_stats_u	stats_u;
 	struct packet_ring_buffer	rx_ring;
 	struct packet_ring_buffer	tx_ring;
 	int			copy_thresh;
@@ -242,6 +327,15 @@ struct packet_skb_cb {
 
 #define PACKET_SKB_CB(__skb)	((struct packet_skb_cb *)((__skb)->cb))
 
+#define GET_PBDQC_FROM_RB(x)	((struct kbdq_core *)(&(x)->prb_bdqc))
+#define GET_PBLOCK_DESC(x, bid)	\
+	((struct block_desc *)((x)->pkbdq[(bid)].buffer))
+#define GET_CURR_PBLOCK_DESC_FROM_CORE(x)	\
+	((struct block_desc *)((x)->pkbdq[(x)->kactive_blk_num].buffer))
+#define GET_NEXT_PRB_BLK_NUM(x) \
+	(((x)->kactive_blk_num < ((x)->knum_blocks-1)) ? \
+	((x)->kactive_blk_num+1) : 0)
+
 static inline struct packet_sock *pkt_sk(struct sock *sk)
 {
 	return (struct packet_sock *)sk;
@@ -325,8 +419,9 @@ static void __packet_set_status(struct packet_sock *po, void *frame, int status)
 		h.h2->tp_status = status;
 		flush_dcache_page(pgv_to_page(&h.h2->tp_status));
 		break;
+	case TPACKET_V3:
 	default:
-		pr_err("TPACKET version not supported\n");
+		WARN(1, "TPACKET version not supported.\n");
 		BUG();
 	}
 
@@ -351,8 +446,9 @@ static int __packet_get_status(struct packet_sock *po, void *frame)
 	case TPACKET_V2:
 		flush_dcache_page(pgv_to_page(&h.h2->tp_status));
 		return h.h2->tp_status;
+	case TPACKET_V3:
 	default:
-		pr_err("TPACKET version not supported\n");
+		WARN(1, "TPACKET version not supported.\n");
 		BUG();
 		return 0;
 	}
@@ -389,6 +485,670 @@ static inline void *packet_current_frame(struct packet_sock *po,
 	return packet_lookup_frame(po, rb, rb->head, status);
 }
 
+static void prb_del_retire_blk_timer(struct kbdq_core *pkc)
+{
+	del_timer_sync(&pkc->retire_blk_timer);
+}
+
+static void prb_shutdown_retire_blk_timer(struct packet_sock *po,
+		int tx_ring,
+		struct sk_buff_head *rb_queue)
+{
+	struct kbdq_core *pkc;
+
+	pkc = tx_ring ? &po->tx_ring.prb_bdqc : &po->rx_ring.prb_bdqc;
+
+	spin_lock(&rb_queue->lock);
+	pkc->delete_blk_timer = 1;
+	spin_unlock(&rb_queue->lock);
+
+	prb_del_retire_blk_timer(pkc);
+}
+
+static void prb_init_blk_timer(struct packet_sock *po,
+		struct kbdq_core *pkc,
+		void (*func) (unsigned long))
+{
+	init_timer(&pkc->retire_blk_timer);
+	pkc->retire_blk_timer.data = (long)po;
+	pkc->retire_blk_timer.function = func;
+	pkc->retire_blk_timer.expires = jiffies;
+}
+
+static void prb_setup_retire_blk_timer(struct packet_sock *po, int tx_ring)
+{
+	struct kbdq_core *pkc;
+
+	if (tx_ring)
+		BUG();
+
+	pkc = tx_ring ? &po->tx_ring.prb_bdqc : &po->rx_ring.prb_bdqc;
+	prb_init_blk_timer(po, pkc, prb_retire_rx_blk_timer_expired);
+}
+
+static int prb_calc_retire_blk_tmo(struct packet_sock *po,
+				int blk_size_in_bytes)
+{
+	struct net_device *dev;
+	unsigned int mbits = 0, msec = 0, div = 0, tmo = 0;
+
+	dev = dev_get_by_index(sock_net(&po->sk), po->ifindex);
+	if (unlikely(dev == NULL))
+		return DEFAULT_PRB_RETIRE_TOV;
+
+	if (dev->ethtool_ops && dev->ethtool_ops->get_settings) {
+		struct ethtool_cmd ecmd = { .cmd = ETHTOOL_GSET, };
+
+		if (!dev->ethtool_ops->get_settings(dev, &ecmd)) {
+			switch (ecmd.speed) {
+			case SPEED_10000:
+				msec = 1;
+				div = 10000/1000;
+				break;
+			case SPEED_1000:
+				msec = 1;
+				div = 1000/1000;
+				break;
+			/*
+			 * If the link speed is so slow you don't really
+			 * need to worry about perf anyways
+			 */
+			case SPEED_100:
+			case SPEED_10:
+			default:
+				return DEFAULT_PRB_RETIRE_TOV;
+			}
+		}
+	}
+
+	mbits = (blk_size_in_bytes * 8) / (1024 * 1024);
+
+	if (div)
+		mbits /= div;
+
+	tmo = mbits * msec;
+
+	if (div)
+		return tmo+1;
+	return tmo;
+}
+
+static void prb_init_ft_ops(struct kbdq_core *p1,
+			union tpacket_req_u *req_u)
+{
+	p1->kfops.ft_ops[p1->kfops.num_ops++] = prb_fill_vlan_info;
+
+	if (req_u->req3.tp_feature_req_word) {
+		if (req_u->req3.tp_feature_req_word & TP_FT_REQ_FILL_RXHASH)
+			p1->kfops.ft_ops[p1->kfops.num_ops++] = prb_fill_rxhash;
+		else
+			p1->kfops.ft_ops[p1->kfops.num_ops++] =
+			prb_clear_rxhash;
+	}
+}
+
+static void init_prb_bdqc(struct packet_sock *po,
+			struct packet_ring_buffer *rb,
+			struct pgv *pg_vec,
+			union tpacket_req_u *req_u, int tx_ring)
+{
+	struct kbdq_core *p1 = &rb->prb_bdqc;
+	struct block_desc *pbd;
+
+	memset(p1, 0x0, sizeof(*p1));
+
+	p1->knxt_seq_num = 1;
+	p1->pkbdq = pg_vec;
+	pbd = (struct block_desc *)pg_vec[0].buffer;
+	p1->pkblk_start	= (char *)pg_vec[0].buffer;
+	p1->kblk_size = req_u->req3.tp_block_size;
+	p1->knum_blocks	= req_u->req3.tp_block_nr;
+	p1->hdrlen = po->tp_hdrlen;
+	p1->version = po->tp_version;
+	p1->last_kactive_blk_num = 0;
+	po->stats_u.stats3.tp_freeze_q_cnt = 0;
+	if (req_u->req3.tp_retire_blk_tov)
+		p1->retire_blk_tov = req_u->req3.tp_retire_blk_tov;
+	else
+		p1->retire_blk_tov = prb_calc_retire_blk_tmo(po,
+						req_u->req3.tp_block_size);
+	p1->tov_in_jiffies = msecs_to_jiffies(p1->retire_blk_tov);
+	p1->blk_sizeof_priv = req_u->req3.tp_sizeof_priv;
+
+	prb_init_ft_ops(p1, req_u);
+	prb_setup_retire_blk_timer(po, tx_ring);
+	prb_open_block(p1, pbd);
+}
+
+/*  Do NOT update the last_blk_num first.
+ *  Assumes sk_buff_head lock is held.
+ */
+static void _prb_refresh_rx_retire_blk_timer(struct kbdq_core *pkc)
+{
+	mod_timer(&pkc->retire_blk_timer,
+			jiffies + pkc->tov_in_jiffies);
+	pkc->last_kactive_blk_num = pkc->kactive_blk_num;
+}
+
+/*
+ * Timer logic:
+ * 1) We refresh the timer only when we open a block.
+ *    By doing this we don't waste cycles refreshing the timer
+ *	  on packet-by-packet basis.
+ *
+ * With a 1MB block-size, on a 1Gbps line, it will take
+ * i) ~8 ms to fill a block + ii) memcpy etc.
+ * In this cut we are not accounting for the memcpy time.
+ *
+ * So, if the user sets the 'tmo' to 10ms then the timer
+ * will never fire while the block is still getting filled
+ * (which is what we want). However, the user could choose
+ * to close a block early and that's fine.
+ *
+ * But when the timer does fire, we check whether or not to refresh it.
+ * Since the tmo granularity is in msecs, it is not too expensive
+ * to refresh the timer, lets say every '8' msecs.
+ * Either the user can set the 'tmo' or we can derive it based on
+ * a) line-speed and b) block-size.
+ * prb_calc_retire_blk_tmo() calculates the tmo.
+ *
+ */
+static void prb_retire_rx_blk_timer_expired(unsigned long data)
+{
+	struct packet_sock *po = (struct packet_sock *)data;
+	struct kbdq_core *pkc = &po->rx_ring.prb_bdqc;
+	unsigned int frozen;
+	struct block_desc *pbd;
+
+	spin_lock(&po->sk.sk_receive_queue.lock);
+
+	frozen = prb_queue_frozen(pkc);
+	pbd = GET_CURR_PBLOCK_DESC_FROM_CORE(pkc);
+
+	if (unlikely(pkc->delete_blk_timer))
+		goto out;
+
+	/* We only need to plug the race when the block is partially filled.
+	 * tpacket_rcv:
+	 *		lock(); increment BLOCK_NUM_PKTS; unlock()
+	 *		copy_bits() is in progress ...
+	 *		timer fires on other cpu:
+	 *		we can't retire the current block because copy_bits
+	 *		is in progress.
+	 *
+	 */
+	if (BLOCK_NUM_PKTS(pbd)) {
+		while (atomic_read(&pkc->blk_fill_in_prog)) {
+			/* Waiting for skb_copy_bits to finish... */
+			cpu_relax();
+		}
+	}
+
+	if (pkc->last_kactive_blk_num == pkc->kactive_blk_num) {
+		if (!frozen) {
+			prb_retire_current_block(pkc, po, TP_STATUS_BLK_TMO);
+			if (!prb_dispatch_next_block(pkc, po))
+				goto refresh_timer;
+			else
+				goto out;
+		} else {
+			/* Case 1. Queue was frozen because user-space was
+			 *	   lagging behind.
+			 */
+			if (prb_curr_blk_in_use(pkc, pbd)) {
+				/*
+				 * Ok, user-space is still behind.
+				 * So just refresh the timer.
+				 */
+				goto refresh_timer;
+			} else {
+			       /* Case 2. queue was frozen,user-space caught up,
+				* now the link went idle && the timer fired.
+				* We don't have a block to close.So we open this
+				* block and restart the timer.
+				* opening a block thaws the queue,restarts timer
+				* Thawing/timer-refresh is a side effect.
+				*/
+				prb_open_block(pkc, pbd);
+				goto out;
+			}
+		}
+	}
+
+refresh_timer:
+	_prb_refresh_rx_retire_blk_timer(pkc);
+
+out:
+	spin_unlock(&po->sk.sk_receive_queue.lock);
+}
+
+static inline void prb_flush_block(struct kbdq_core *pkc1,
+		struct block_desc *pbd1, __u32 status)
+{
+	/* Flush everything minus the block header */
+
+#if ARCH_IMPLEMENTS_FLUSH_DCACHE_PAGE == 1
+	u8 *start, *end;
+
+	start = (u8 *)pbd1;
+
+	/* Skip the block header(we know header WILL fit in 4K) */
+	start += PAGE_SIZE;
+
+	end = (u8 *)PAGE_ALIGN((unsigned long)pkc1->pkblk_end);
+	for (; start < end; start += PAGE_SIZE)
+		flush_dcache_page(pgv_to_page(start));
+
+	smp_wmb();
+#endif
+
+	/* Now update the block status. */
+
+	BLOCK_STATUS(pbd1) = status;
+
+	/* Flush the block header */
+
+#if ARCH_IMPLEMENTS_FLUSH_DCACHE_PAGE == 1
+	start = (u8 *)pbd1;
+	flush_dcache_page(pgv_to_page(start));
+
+	smp_wmb();
+#endif
+}
+
+/*
+ * Side effect:
+ *
+ * 1) flush the block
+ * 2) Increment active_blk_num
+ *
+ * Note:We DONT refresh the timer on purpose.
+ *	Because almost always the next block will be opened.
+ */
+static void prb_close_block(struct kbdq_core *pkc1, struct block_desc *pbd1,
+		struct packet_sock *po, unsigned int stat)
+{
+	__u32 status = TP_STATUS_USER | stat;
+
+	struct tpacket3_hdr *last_pkt;
+	struct hdr_v1 *h1 = &pbd1->hdr.bh1;
+
+	if (po->stats.tp_drops)
+		status |= TP_STATUS_LOSING;
+
+	last_pkt = (struct tpacket3_hdr *)pkc1->prev;
+	last_pkt->tp_next_offset = 0;
+
+	/* Get the ts of the last pkt */
+	if (BLOCK_NUM_PKTS(pbd1)) {
+		h1->ts_last_pkt.ts_sec = last_pkt->tp_sec;
+		h1->ts_last_pkt.ts_nsec	= last_pkt->tp_nsec;
+	} else {
+		/* Ok, we tmo'd - so get the current time */
+		struct timespec ts;
+		getnstimeofday(&ts);
+		h1->ts_last_pkt.ts_sec = ts.tv_sec;
+		h1->ts_last_pkt.ts_nsec	= ts.tv_nsec;
+	}
+
+	smp_wmb();
+
+	/* Flush the block */
+	prb_flush_block(pkc1, pbd1, status);
+
+	pkc1->kactive_blk_num = GET_NEXT_PRB_BLK_NUM(pkc1);
+}
+
+static inline void prb_thaw_queue(struct kbdq_core *pkc)
+{
+	pkc->reset_pending_on_curr_blk = 0;
+}
+
+/*
+ * Side effect of opening a block:
+ *
+ * 1) prb_queue is thawed.
+ * 2) retire_blk_timer is refreshed.
+ *
+ */
+static void prb_open_block(struct kbdq_core *pkc1, struct block_desc *pbd1)
+{
+	struct timespec ts;
+	struct hdr_v1 *h1 = &pbd1->hdr.bh1;
+
+	smp_rmb();
+
+	if (likely(TP_STATUS_KERNEL == BLOCK_STATUS(pbd1))) {
+
+		/* We could have just memset this but we will lose the
+		 * flexibility of making the priv area sticky
+		 */
+		BLOCK_SNUM(pbd1) = pkc1->knxt_seq_num++;
+		BLOCK_NUM_PKTS(pbd1) = 0;
+		BLOCK_LEN(pbd1) = BLK_PLUS_PRIV(pkc1->blk_sizeof_priv);
+		getnstimeofday(&ts);
+		h1->ts_first_pkt.ts_sec = ts.tv_sec;
+		h1->ts_first_pkt.ts_nsec = ts.tv_nsec;
+		pkc1->pkblk_start = (char *)pbd1;
+		pkc1->nxt_offset = (char *)(pkc1->pkblk_start +
+		BLK_PLUS_PRIV(pkc1->blk_sizeof_priv));
+		BLOCK_O2FP(pbd1) = (__u32)BLK_PLUS_PRIV(pkc1->blk_sizeof_priv);
+		BLOCK_O2PRIV(pbd1) = BLK_HDR_LEN;
+		pbd1->version = pkc1->version;
+		pkc1->prev = pkc1->nxt_offset;
+		pkc1->pkblk_end = pkc1->pkblk_start + pkc1->kblk_size;
+		prb_thaw_queue(pkc1);
+		_prb_refresh_rx_retire_blk_timer(pkc1);
+
+		smp_wmb();
+
+		return;
+	}
+
+	WARN(1, "ERROR block:%p is NOT FREE status:%d kactive_blk_num:%d\n",
+		pbd1, BLOCK_STATUS(pbd1), pkc1->kactive_blk_num);
+	dump_stack();
+	BUG();
+}
+
+/*
+ * Queue freeze logic:
+ * 1) Assume tp_block_nr = 8 blocks.
+ * 2) At time 't0', user opens Rx ring.
+ * 3) Some time past 't0', kernel starts filling blocks starting from 0 .. 7
+ * 4) user-space is either sleeping or processing block '0'.
+ * 5) tpacket_rcv is currently filling block '7', since there is no space left,
+ *    it will close block-7,loop around and try to fill block '0'.
+ *    call-flow:
+ *    __packet_lookup_frame_in_block
+ *      prb_retire_current_block()
+ *      prb_dispatch_next_block()
+ *        |->(BLOCK_STATUS == USER) evaluates to true
+ *    5.1) Since block-0 is currently in-use, we just freeze the queue.
+ * 6) Now there are two cases:
+ *    6.1) Link goes idle right after the queue is frozen.
+ *         But remember, the last open_block() refreshed the timer.
+ *         When this timer expires,it will refresh itself so that we can
+ *         re-open block-0 in near future.
+ *    6.2) Link is busy and keeps on receiving packets. This is a simple
+ *         case and __packet_lookup_frame_in_block will check if block-0
+ *         is free and can now be re-used.
+ */
+static inline void prb_freeze_queue(struct kbdq_core *pkc,
+				  struct packet_sock *po)
+{
+	pkc->reset_pending_on_curr_blk = 1;
+	po->stats_u.stats3.tp_freeze_q_cnt++;
+}
+
+#define TOTAL_PKT_LEN_INCL_ALIGN(length) (ALIGN((length), V3_ALIGNMENT))
+
+/*
+ * If the next block is free then we will dispatch it
+ * and return a good offset.
+ * Else, we will freeze the queue.
+ * So, caller must check the return value.
+ */
+static void *prb_dispatch_next_block(struct kbdq_core *pkc,
+		struct packet_sock *po)
+{
+	struct block_desc *pbd;
+
+	smp_rmb();
+
+	/* 1. Get current block num */
+	pbd = GET_CURR_PBLOCK_DESC_FROM_CORE(pkc);
+
+	/* 2. If this block is currently in_use then freeze the queue */
+	if (TP_STATUS_USER & BLOCK_STATUS(pbd)) {
+		prb_freeze_queue(pkc, po);
+		return NULL;
+	}
+
+	/*
+	 * 3.
+	 * open this block and return the offset where the first packet
+	 * needs to get stored.
+	 */
+	prb_open_block(pkc, pbd);
+	return (void *)pkc->nxt_offset;
+}
+
+static void prb_retire_current_block(struct kbdq_core *pkc,
+		struct packet_sock *po, unsigned int status)
+{
+	struct block_desc *pbd = GET_CURR_PBLOCK_DESC_FROM_CORE(pkc);
+
+	/* retire/close the current block */
+	if (likely(TP_STATUS_KERNEL == BLOCK_STATUS(pbd))) {
+		/*
+		 * Plug the case where copy_bits() is in progress on
+		 * cpu-0 and tpacket_rcv() got invoked on cpu-1, didn't
+		 * have space to copy the pkt in the current block and
+		 * called prb_retire_current_block()
+		 *
+		 * We don't need to worry about the TMO case because
+		 * the timer-handler already handled this case.
+		 */
+		if (!(status & TP_STATUS_BLK_TMO)) {
+			while (atomic_read(&pkc->blk_fill_in_prog)) {
+				/* Waiting for skb_copy_bits to finish... */
+				cpu_relax();
+			}
+		}
+		prb_close_block(pkc, pbd, po, status);
+		return;
+	}
+
+	WARN(1, "ERROR-pbd[%d]:%p\n", pkc->kactive_blk_num, pbd);
+	dump_stack();
+	BUG();
+}
+
+static inline int prb_curr_blk_in_use(struct kbdq_core *pkc,
+				      struct block_desc *pbd)
+{
+	return TP_STATUS_USER & BLOCK_STATUS(pbd);
+}
+
+static inline int prb_queue_frozen(struct kbdq_core *pkc)
+{
+	return pkc->reset_pending_on_curr_blk;
+}
+
+static inline void prb_clear_blk_fill_status(struct packet_ring_buffer *rb)
+{
+	struct kbdq_core *pkc  = GET_PBDQC_FROM_RB(rb);
+	atomic_dec(&pkc->blk_fill_in_prog);
+}
+
+static void prb_fill_rxhash(struct kbdq_core *pkc,
+			struct tpacket3_hdr *ppd)
+{
+	ppd->hv1.tp_rxhash = skb_get_rxhash(pkc->skb);
+}
+
+static void prb_clear_rxhash(struct kbdq_core *pkc,
+			struct tpacket3_hdr *ppd)
+{
+	ppd->hv1.tp_rxhash = 0;
+}
+
+static void prb_fill_vlan_info(struct kbdq_core *pkc,
+			struct tpacket3_hdr *ppd)
+{
+	if (vlan_tx_tag_present(pkc->skb)) {
+		ppd->hv1.tp_vlan_tci = vlan_tx_tag_get(pkc->skb);
+		ppd->tp_status = TP_STATUS_VLAN_VALID;
+	} else {
+		ppd->hv1.tp_vlan_tci = ppd->tp_status = 0;
+	}
+}
+
+static void prb_run_all_ft_ops(struct kbdq_core *pkc,
+			struct tpacket3_hdr *ppd)
+{
+	int ops;
+	for (ops = 0; ops < pkc->kfops.num_ops; ops++)
+		pkc->kfops.ft_ops[ops](pkc, ppd);
+}
+
+static inline void prb_fill_curr_block(char *curr, struct kbdq_core *pkc,
+				struct block_desc *pbd,
+				unsigned int len)
+{
+	struct tpacket3_hdr *ppd;
+
+	ppd  = (struct tpacket3_hdr *)curr;
+	ppd->tp_next_offset = TOTAL_PKT_LEN_INCL_ALIGN(len);
+	pkc->prev = curr;
+	pkc->nxt_offset += TOTAL_PKT_LEN_INCL_ALIGN(len);
+	BLOCK_LEN(pbd) += TOTAL_PKT_LEN_INCL_ALIGN(len);
+	BLOCK_NUM_PKTS(pbd) += 1;
+	atomic_inc(&pkc->blk_fill_in_prog);
+	prb_run_all_ft_ops(pkc, ppd);
+}
+
+/* Assumes caller has the sk->rx_queue.lock */
+static void *__packet_lookup_frame_in_block(struct packet_sock *po,
+					    struct sk_buff *skb,
+						int status,
+					    unsigned int len
+					    )
+{
+	struct kbdq_core *pkc;
+	struct block_desc *pbd;
+	char *curr, *end;
+
+	pkc = GET_PBDQC_FROM_RB(((struct packet_ring_buffer *)&po->rx_ring));
+	pbd = GET_CURR_PBLOCK_DESC_FROM_CORE(pkc);
+
+	/* Queue is frozen when user space is lagging behind */
+	if (prb_queue_frozen(pkc)) {
+		/*
+		 * Check if that last block which caused the queue to freeze,
+		 * is still in_use by user-space.
+		 */
+		if (prb_curr_blk_in_use(pkc, pbd)) {
+			/* Can't record this packet */
+			return NULL;
+		} else {
+			/*
+			 * Ok, the block was released by user-space.
+			 * Now let's open that block.
+			 * opening a block also thaws the queue.
+			 * Thawing is a side effect.
+			 */
+			prb_open_block(pkc, pbd);
+		}
+	}
+
+	smp_mb();
+	curr = pkc->nxt_offset;
+	pkc->skb = skb;
+	end = (char *) ((char *)pbd + pkc->kblk_size);
+
+	/* first try the current block */
+	if (curr+TOTAL_PKT_LEN_INCL_ALIGN(len) < end) {
+		prb_fill_curr_block(curr, pkc, pbd, len);
+		return (void *)curr;
+	}
+
+	/* Ok, close the current block */
+	prb_retire_current_block(pkc, po, 0);
+
+	/* Now, try to dispatch the next block */
+	curr = (char *)prb_dispatch_next_block(pkc, po);
+	if (curr) {
+		pbd = GET_CURR_PBLOCK_DESC_FROM_CORE(pkc);
+		prb_fill_curr_block(curr, pkc, pbd, len);
+		return (void *)curr;
+	}
+
+	/*
+	 * No free blocks are available.user_space hasn't caught up yet.
+	 * Queue was just frozen and now this packet will get dropped.
+	 */
+	return NULL;
+}
+
+static inline void *packet_current_rx_frame(struct packet_sock *po,
+					    struct sk_buff *skb,
+					    int status, unsigned int len)
+{
+	char *curr = NULL;
+	switch (po->tp_version) {
+	case TPACKET_V1:
+	case TPACKET_V2:
+		curr = packet_lookup_frame(po, &po->rx_ring,
+					po->rx_ring.head, status);
+		return curr;
+	case TPACKET_V3:
+		return __packet_lookup_frame_in_block(po, skb, status, len);
+	default:
+		WARN(1, "TPACKET version not supported\n");
+		BUG();
+		return 0;
+	}
+}
+
+static inline void *prb_lookup_block(struct packet_sock *po,
+				     struct packet_ring_buffer *rb,
+				     unsigned int previous,
+				     int status)
+{
+	struct kbdq_core *pkc  = GET_PBDQC_FROM_RB(rb);
+	struct block_desc *pbd = GET_PBLOCK_DESC(pkc, previous);
+
+	if (status != BLOCK_STATUS(pbd))
+		return NULL;
+	return pbd;
+}
+
+static inline int prb_previous_blk_num(struct packet_ring_buffer *rb)
+{
+	unsigned int prev;
+	if (rb->prb_bdqc.kactive_blk_num)
+		prev = rb->prb_bdqc.kactive_blk_num-1;
+	else
+		prev = rb->prb_bdqc.knum_blocks-1;
+	return prev;
+}
+
+/* Assumes caller has held the rx_queue.lock */
+static inline void *__prb_previous_block(struct packet_sock *po,
+					 struct packet_ring_buffer *rb,
+					 int status)
+{
+	unsigned int previous = prb_previous_blk_num(rb);
+	return prb_lookup_block(po, rb, previous, status);
+}
+
+static inline void *packet_previous_rx_frame(struct packet_sock *po,
+					     struct packet_ring_buffer *rb,
+					     int status)
+{
+	if (po->tp_version <= TPACKET_V2)
+		return packet_previous_frame(po, rb, status);
+
+	return __prb_previous_block(po, rb, status);
+}
+
+static inline void packet_increment_rx_head(struct packet_sock *po,
+					    struct packet_ring_buffer *rb)
+{
+	switch (po->tp_version) {
+	case TPACKET_V1:
+	case TPACKET_V2:
+		return packet_increment_head(rb);
+	case TPACKET_V3:
+	default:
+		WARN(1, "TPACKET version not supported.\n");
+		BUG();
+		return;
+	}
+}
+
 static inline void *packet_previous_frame(struct packet_sock *po,
 		struct packet_ring_buffer *rb,
 		int status)
@@ -982,12 +1742,13 @@ static int tpacket_rcv(struct sk_buff *skb, struct net_device *dev,
 	union {
 		struct tpacket_hdr *h1;
 		struct tpacket2_hdr *h2;
+		struct tpacket3_hdr *h3;
 		void *raw;
 	} h;
 	u8 *skb_head = skb->data;
 	int skb_len = skb->len;
 	unsigned int snaplen, res;
-	unsigned long status = TP_STATUS_LOSING|TP_STATUS_USER;
+	unsigned long status = TP_STATUS_USER;
 	unsigned short macoff, netoff, hdrlen;
 	struct sk_buff *copy_skb = NULL;
 	struct timeval tv;
@@ -1033,37 +1794,46 @@ static int tpacket_rcv(struct sk_buff *skb, struct net_device *dev,
 			po->tp_reserve;
 		macoff = netoff - maclen;
 	}
-
-	if (macoff + snaplen > po->rx_ring.frame_size) {
-		if (po->copy_thresh &&
-		    atomic_read(&sk->sk_rmem_alloc) + skb->truesize <
-		    (unsigned)sk->sk_rcvbuf) {
-			if (skb_shared(skb)) {
-				copy_skb = skb_clone(skb, GFP_ATOMIC);
-			} else {
-				copy_skb = skb_get(skb);
-				skb_head = skb->data;
+	if (po->tp_version <= TPACKET_V2) {
+		if (macoff + snaplen > po->rx_ring.frame_size) {
+			if (po->copy_thresh &&
+				atomic_read(&sk->sk_rmem_alloc) + skb->truesize
+				< (unsigned)sk->sk_rcvbuf) {
+				if (skb_shared(skb)) {
+					copy_skb = skb_clone(skb, GFP_ATOMIC);
+				} else {
+					copy_skb = skb_get(skb);
+					skb_head = skb->data;
+				}
+				if (copy_skb)
+					skb_set_owner_r(copy_skb, sk);
 			}
-			if (copy_skb)
-				skb_set_owner_r(copy_skb, sk);
+			snaplen = po->rx_ring.frame_size - macoff;
+			if ((int)snaplen < 0)
+				snaplen = 0;
 		}
-		snaplen = po->rx_ring.frame_size - macoff;
-		if ((int)snaplen < 0)
-			snaplen = 0;
 	}
-
 	spin_lock(&sk->sk_receive_queue.lock);
-	h.raw = packet_current_frame(po, &po->rx_ring, TP_STATUS_KERNEL);
+	h.raw = packet_current_rx_frame(po, skb,
+					TP_STATUS_KERNEL, (macoff+snaplen));
 	if (!h.raw)
 		goto ring_is_full;
-	packet_increment_head(&po->rx_ring);
+	if (po->tp_version <= TPACKET_V2) {
+		packet_increment_rx_head(po, &po->rx_ring);
+	/*
+	 * LOSING will be reported till you read the stats,
+	 * because it's COR - Clear On Read.
+	 * Anyways, moving it for V1/V2 only as V3 doesn't need this
+	 * at packet level.
+	 */
+		if (po->stats.tp_drops)
+			status |= TP_STATUS_LOSING;
+	}
 	po->stats.tp_packets++;
 	if (copy_skb) {
 		status |= TP_STATUS_COPY;
 		__skb_queue_tail(&sk->sk_receive_queue, copy_skb);
 	}
-	if (!po->stats.tp_drops)
-		status &= ~TP_STATUS_LOSING;
 	spin_unlock(&sk->sk_receive_queue.lock);
 
 	skb_copy_bits(skb, 0, h.raw + macoff, snaplen);
@@ -1114,6 +1884,29 @@ static int tpacket_rcv(struct sk_buff *skb, struct net_device *dev,
 		h.h2->tp_padding = 0;
 		hdrlen = sizeof(*h.h2);
 		break;
+	case TPACKET_V3:
+		/* tp_nxt_offset,vlan are already populated above.
+		 * So DONT clear those fields here
+		 */
+		h.h3->tp_status |= status;
+		h.h3->tp_len = skb->len;
+		h.h3->tp_snaplen = snaplen;
+		h.h3->tp_mac = macoff;
+		h.h3->tp_net = netoff;
+		if ((po->tp_tstamp & SOF_TIMESTAMPING_SYS_HARDWARE)
+				&& shhwtstamps->syststamp.tv64)
+			ts = ktime_to_timespec(shhwtstamps->syststamp);
+		else if ((po->tp_tstamp & SOF_TIMESTAMPING_RAW_HARDWARE)
+				&& shhwtstamps->hwtstamp.tv64)
+			ts = ktime_to_timespec(shhwtstamps->hwtstamp);
+		else if (skb->tstamp.tv64)
+			ts = ktime_to_timespec(skb->tstamp);
+		else
+			getnstimeofday(&ts);
+		h.h3->tp_sec  = ts.tv_sec;
+		h.h3->tp_nsec = ts.tv_nsec;
+		hdrlen = sizeof(*h.h3);
+		break;
 	default:
 		BUG();
 	}
@@ -1134,13 +1927,19 @@ static int tpacket_rcv(struct sk_buff *skb, struct net_device *dev,
 	{
 		u8 *start, *end;
 
-		end = (u8 *)PAGE_ALIGN((unsigned long)h.raw + macoff + snaplen);
-		for (start = h.raw; start < end; start += PAGE_SIZE)
-			flush_dcache_page(pgv_to_page(start));
+		if (po->tp_version <= TPACKET_V2) {
+			end = (u8 *)PAGE_ALIGN((unsigned long)h.raw
+				+ macoff + snaplen);
+			for (start = h.raw; start < end; start += PAGE_SIZE)
+				flush_dcache_page(pgv_to_page(start));
+		}
 		smp_wmb();
 	}
 #endif
-	__packet_set_status(po, h.raw, status);
+	if (po->tp_version <= TPACKET_V2)
+		__packet_set_status(po, h.raw, status);
+	else
+		prb_clear_blk_fill_status(&po->rx_ring);
 
 	sk->sk_data_ready(sk, 0);
 
@@ -1631,7 +2430,7 @@ static int packet_release(struct socket *sock)
 	struct sock *sk = sock->sk;
 	struct packet_sock *po;
 	struct net *net;
-	struct tpacket_req req;
+	union tpacket_req_u req_u;
 
 	if (!sk)
 		return 0;
@@ -1654,13 +2453,13 @@ static int packet_release(struct socket *sock)
 
 	packet_flush_mclist(sk);
 
-	memset(&req, 0, sizeof(req));
+	memset(&req_u, 0, sizeof(req_u));
 
 	if (po->rx_ring.pg_vec)
-		packet_set_ring(sk, &req, 1, 0);
+		packet_set_ring(sk, &req_u, 1, 0);
 
 	if (po->tx_ring.pg_vec)
-		packet_set_ring(sk, &req, 1, 1);
+		packet_set_ring(sk, &req_u, 1, 1);
 
 	fanout_release(sk);
 
@@ -2280,15 +3079,27 @@ packet_setsockopt(struct socket *sock, int level, int optname, char __user *optv
 	case PACKET_RX_RING:
 	case PACKET_TX_RING:
 	{
-		struct tpacket_req req;
+		union tpacket_req_u req_u;
+		int len;
 
-		if (optlen < sizeof(req))
+		switch (po->tp_version) {
+		case TPACKET_V1:
+		case TPACKET_V2:
+			len = sizeof(req_u.req);
+			break;
+		case TPACKET_V3:
+		default:
+			len = sizeof(req_u.req3);
+			break;
+		}
+		if (optlen < len)
 			return -EINVAL;
 		if (pkt_sk(sk)->has_vnet_hdr)
 			return -EINVAL;
-		if (copy_from_user(&req, optval, sizeof(req)))
+		if (copy_from_user(&req_u.req, optval, len))
 			return -EFAULT;
-		return packet_set_ring(sk, &req, 0, optname == PACKET_TX_RING);
+		return packet_set_ring(sk, &req_u, 0,
+			optname == PACKET_TX_RING);
 	}
 	case PACKET_COPY_THRESH:
 	{
@@ -2315,6 +3126,7 @@ packet_setsockopt(struct socket *sock, int level, int optname, char __user *optv
 		switch (val) {
 		case TPACKET_V1:
 		case TPACKET_V2:
+		case TPACKET_V3:
 			po->tp_version = val;
 			return 0;
 		default:
@@ -2424,6 +3236,7 @@ static int packet_getsockopt(struct socket *sock, int level, int optname,
 	struct packet_sock *po = pkt_sk(sk);
 	void *data;
 	struct tpacket_stats st;
+	union tpacket_stats_u st_u;
 
 	if (level != SOL_PACKET)
 		return -ENOPROTOOPT;
@@ -2436,15 +3249,27 @@ static int packet_getsockopt(struct socket *sock, int level, int optname,
 
 	switch (optname) {
 	case PACKET_STATISTICS:
-		if (len > sizeof(struct tpacket_stats))
-			len = sizeof(struct tpacket_stats);
+		if (po->tp_version == TPACKET_V3) {
+			len = sizeof(struct tpacket_stats_v3);
+		} else {
+			if (len > sizeof(struct tpacket_stats))
+				len = sizeof(struct tpacket_stats);
+		}
 		spin_lock_bh(&sk->sk_receive_queue.lock);
-		st = po->stats;
+		if (po->tp_version == TPACKET_V3) {
+			memcpy(&st_u.stats3, &po->stats,
+			sizeof(struct tpacket_stats));
+			st_u.stats3.tp_freeze_q_cnt =
+			po->stats_u.stats3.tp_freeze_q_cnt;
+			st_u.stats3.tp_packets += po->stats.tp_drops;
+			data = &st_u.stats3;
+		} else {
+			st = po->stats;
+			st.tp_packets += st.tp_drops;
+			data = &st;
+		}
 		memset(&po->stats, 0, sizeof(st));
 		spin_unlock_bh(&sk->sk_receive_queue.lock);
-		st.tp_packets += st.tp_drops;
-
-		data = &st;
 		break;
 	case PACKET_AUXDATA:
 		if (len > sizeof(int))
@@ -2485,6 +3310,9 @@ static int packet_getsockopt(struct socket *sock, int level, int optname,
 		case TPACKET_V2:
 			val = sizeof(struct tpacket2_hdr);
 			break;
+		case TPACKET_V3:
+			val = sizeof(struct tpacket3_hdr);
+			break;
 		default:
 			return -EINVAL;
 		}
@@ -2641,7 +3469,8 @@ static unsigned int packet_poll(struct file *file, struct socket *sock,
 
 	spin_lock_bh(&sk->sk_receive_queue.lock);
 	if (po->rx_ring.pg_vec) {
-		if (!packet_previous_frame(po, &po->rx_ring, TP_STATUS_KERNEL))
+		if (!packet_previous_rx_frame(po, &po->rx_ring,
+			TP_STATUS_KERNEL))
 			mask |= POLLIN | POLLRDNORM;
 	}
 	spin_unlock_bh(&sk->sk_receive_queue.lock);
@@ -2760,7 +3589,7 @@ out_free_pgvec:
 	goto out;
 }
 
-static int packet_set_ring(struct sock *sk, struct tpacket_req *req,
+static int packet_set_ring(struct sock *sk, union tpacket_req_u *req_u,
 		int closing, int tx_ring)
 {
 	struct pgv *pg_vec = NULL;
@@ -2769,7 +3598,15 @@ static int packet_set_ring(struct sock *sk, struct tpacket_req *req,
 	struct packet_ring_buffer *rb;
 	struct sk_buff_head *rb_queue;
 	__be16 num;
-	int err;
+	int err = -EINVAL;
+	/* Added to avoid minimal code churn */
+	struct tpacket_req *req = &req_u->req;
+
+	/* Opening a Tx-ring is NOT supported in TPACKET_V3 */
+	if (!closing && tx_ring && (po->tp_version > TPACKET_V2)) {
+		WARN(1, "Tx-ring is not supported.\n");
+		goto out;
+	}
 
 	rb = tx_ring ? &po->tx_ring : &po->rx_ring;
 	rb_queue = tx_ring ? &sk->sk_write_queue : &sk->sk_receive_queue;
@@ -2795,6 +3632,9 @@ static int packet_set_ring(struct sock *sk, struct tpacket_req *req,
 		case TPACKET_V2:
 			po->tp_hdrlen = TPACKET2_HDRLEN;
 			break;
+		case TPACKET_V3:
+			po->tp_hdrlen = TPACKET3_HDRLEN;
+			break;
 		}
 
 		err = -EINVAL;
@@ -2820,6 +3660,17 @@ static int packet_set_ring(struct sock *sk, struct tpacket_req *req,
 		pg_vec = alloc_pg_vec(req, order);
 		if (unlikely(!pg_vec))
 			goto out;
+		switch (po->tp_version) {
+		case TPACKET_V3:
+		/* Transmit path is not supported. We checked
+		 * it above but just being paranoid
+		 */
+			if (!tx_ring)
+				init_prb_bdqc(po, rb, pg_vec, req_u, tx_ring);
+				break;
+		default:
+			break;
+		}
 	}
 	/* Done */
 	else {
@@ -2872,7 +3723,11 @@ static int packet_set_ring(struct sock *sk, struct tpacket_req *req,
 		register_prot_hook(sk);
 	}
 	spin_unlock(&po->bind_lock);
-
+	if (closing && (po->tp_version > TPACKET_V2)) {
+		/* Because we don't support block-based V3 on tx-ring */
+		if (!tx_ring)
+			prb_shutdown_retire_blk_timer(po, tx_ring, rb_queue);
+	}
 	release_sock(sk);
 
 	if (pg_vec)
-- 
1.7.5.2

Re: [PATCH net-next v4 af-packet 2/2] Enhance af-packet to provide (near zero)lossless packet capture functionality.

[David Miller]

From: Chetan Loke <loke.chetan@gmail.com>
Date: Thu, 11 Aug 2011 22:31:00 -0400

> +struct kbdq_ft_ops {
> +	int num_ops;
> +	void (*ft_ops[2])(void *, void *);
> +};
...
> +	struct kbdq_ft_ops kfops;
 ...
> +static void prb_init_ft_ops(struct kbdq_core *p1,
> +			union tpacket_req_u *req_u)
> +{
> +	p1->kfops.ft_ops[p1->kfops.num_ops++] = prb_fill_vlan_info;
> +
> +	if (req_u->req3.tp_feature_req_word) {
> +		if (req_u->req3.tp_feature_req_word & TP_FT_REQ_FILL_RXHASH)
> +			p1->kfops.ft_ops[p1->kfops.num_ops++] = prb_fill_rxhash;
> +		else
> +			p1->kfops.ft_ops[p1->kfops.num_ops++] =
> +			prb_clear_rxhash;
> +	}
> +}

It is a lot cheaper to just test the flags in-line than do indirect calls.
Indirect calls are very expensive on many cpus.

In fact, since the first op (prb_fill_vlan_info) is unconditional, we eat
the indirect call cost for absolutely no reason at all.

This kfops stuff was not present in your previous changes.  And I'm
going to tell you that if you keep adding things each revision instead
of just fixing the specific items you've received feedback about, a
set of changes this invasive and of this size will never get merged.

Please resist the urge to further tinker with the code, and just
address the feedback we give you.

Re: [PATCH net-next v4 af-packet 2/2] Enhance af-packet to provide (near zero)lossless packet capture functionality.

[chetan loke]

On Fri, Aug 19, 2011 at 6:36 AM, David Miller <davem@davemloft.net> wrote:
> From: Chetan Loke <loke.chetan@gmail.com>
> Date: Thu, 11 Aug 2011 22:31:00 -0400
>
>> +struct kbdq_ft_ops {
>> +     int num_ops;
>> +     void (*ft_ops[2])(void *, void *);
>> +};
> ...
>> +     struct kbdq_ft_ops kfops;
>  ...
>> +static void prb_init_ft_ops(struct kbdq_core *p1,
>> +                     union tpacket_req_u *req_u)
>> +{
>> +     p1->kfops.ft_ops[p1->kfops.num_ops++] = prb_fill_vlan_info;
>> +
>> +     if (req_u->req3.tp_feature_req_word) {
>> +             if (req_u->req3.tp_feature_req_word & TP_FT_REQ_FILL_RXHASH)
>> +                     p1->kfops.ft_ops[p1->kfops.num_ops++] = prb_fill_rxhash;
>> +             else
>> +                     p1->kfops.ft_ops[p1->kfops.num_ops++] =
>> +                     prb_clear_rxhash;
>> +     }
>> +}
>
> It is a lot cheaper to just test the flags in-line than do indirect calls.
> Indirect calls are very expensive on many cpus.
>
> In fact, since the first op (prb_fill_vlan_info) is unconditional, we eat
> the indirect call cost for absolutely no reason at all.

Oki. Will test the flags in-line in prb_run_all_ft_ops().


>
> This kfops stuff was not present in your previous changes.  And I'm
> going to tell you that if you keep adding things each revision instead
> of just fixing the specific items you've received feedback about, a
> set of changes this invasive and of this size will never get merged.
>

First, thanks(to you and others) for reviewing the huge patchset. I
did think about breaking it into multiple patches but since it
involves changes in a single file it was a little difficult. And I
didn't want to create a separate file for this.

> Please resist the urge to further tinker with the code, and just
> address the feedback we give you.
>
Sorry won't happen. These changes should have been present in v3 when
I added the feature-request word because that was the intent of that
word. It was a screw up on my part.


Chetan Loke

Re: [PATCH net-next v4 af-packet 0/2] Enhance af-packet to provide (near zero)lossless packet capture functionality.

[David Miller]

From: Chetan Loke <loke.chetan@gmail.com>
Date: Thu, 11 Aug 2011 22:30:58 -0400

> Changes in v4:

Thanks for updating this patch, I will review it soon.

Re: [PATCH net-next v4 af-packet 0/2] Enhance af-packet to provide (near zero)lossless packet capture functionality.

[David Miller]

From: Chetan Loke <loke.chetan@gmail.com>
Date: Thu, 11 Aug 2011 22:30:58 -0400

> This patch attempts to:
> 1)Improve network capture visibility by increasing packet density
> 2)Assist in analyzing multiple(aggregated) capture ports.

Just some general comments before I say something about the actual
code.

Claiming loss-less, or almost loss-less, packet capture is something
which is highly dependent upon the hardware and traffic load.

You cannot claim near loss-less packet capture for an i386 with a
10gbit ethernet part going line rate, for example.

So I want you to describe this as what it is, an improvement.  More
specifically it's just a more flexible mmap ring buffer scheme that
reduces the wastage incurred by the fixed buffer size scheme we
have currently.

Next, you need to use a different subject line for the two patches.
Right now you say the same thing in both, and someone scanning the
changesets headers will have no idea what is different about one
change vs. the other.

Re: [PATCH net-next v4 af-packet 0/2] Enhance af-packet to provide (near zero)lossless packet capture functionality.

[chetan loke]

On Fri, Aug 19, 2011 at 6:33 AM, David Miller <davem@davemloft.net> wrote:
> From: Chetan Loke <loke.chetan@gmail.com>
> Date: Thu, 11 Aug 2011 22:30:58 -0400
>
>> This patch attempts to:
>> 1)Improve network capture visibility by increasing packet density
>> 2)Assist in analyzing multiple(aggregated) capture ports.
>
> Just some general comments before I say something about the actual
> code.
>
> Claiming loss-less, or almost loss-less, packet capture is something
> which is highly dependent upon the hardware and traffic load.

Totally agree. I have added the text in [Patch 2/2], bullet 1.3 -
capture 99% traffic as seen by the kernel. So lossless is relative to
'the traffic processed by the kernel' and not to the (inline)hardware.
But yes, the patch subject itself could be misleading. Not trying to
argue but just wanted to make sure we are all on the same page.


> So I want you to describe this as what it is, an improvement.  More
> specifically it's just a more flexible mmap ring buffer scheme that
> reduces the wastage incurred by the fixed buffer size scheme we
> have currently.
>
Oki. I will change the cover-subject to 'Enhance af-packet to provide
a flexible mmap ring buffer scheme eliminating fixed frame-size
requirement'.

> Next, you need to use a different subject line for the two patches.
> Right now you say the same thing in both, and someone scanning the
> changesets headers will have no idea what is different about one
> change vs. the other.
>
Done.


Chetan Loke