Skip to contentSkip to conversation

lacp rx/tx handlers fixes for bonding pmd

PATCH In review David Marchand 5 participants 16 messages Apr 10, 2019 – Aug 22, 2019

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

cover lacp rx/tx handlers fixes for bonding pmd

Another series with focus on the fast/normal rx/tx handlers for 802.3ad.

The first two patches make sure that the rx (resp. tx) fast and normal handlers are equivalent.

The third one will most likely have an impact on performance which I tried to mitigate with the last one. However, I have no benchmark to back those patches and I did not test those thoroughly, so they are more like RFC patches but sending anyway.

Signature · 16 lines
David Marchand

David Marchand (4):
  net/bonding: fix oob access in LACP mode when sending many packets
  net/bonding: fix LACP fast queue Rx handler
  net/bonding: fix unicast packets filtering when not in promisc
  net/bonding: prefer allmulti to promisc for LACP

 drivers/net/bonding/rte_eth_bond_8023ad.c         |  51 ++-
 drivers/net/bonding/rte_eth_bond_8023ad_private.h |   7 +
 drivers/net/bonding/rte_eth_bond_pmd.c            | 402 +++++++++-------------
 drivers/net/bonding/rte_eth_bond_private.h        |   3 -
 4 files changed, 224 insertions(+), 239 deletions(-)

-- 
1.8.3.1
#1 lore raw

PATCH 1/4 net/bonding: fix oob access in LACP mode when sending many packets

We'd better consolidate the fast queue and the normal tx burst functions under a common inline wrapper for maintenance.

But looking closer at the bufs_slave_port_idxs[] mapping array in those tx burst functions, its size is invalid since up to nb_bufs are handled here.
A previous patch [1] fixed this issue for balance tx burst function without mentionning it.

802.3ad and balance modes are functionally equivalent on transmit. The only difference is on the slave id distribution.
Add an additional inline wrapper to consolidate even more and fix this issue.

[1]: https://git.dpdk.org/dpdk/commit/?id=c5224f623431

Fixes 09150784a776 ("net/bonding: burst mode hash calculation")
---
 drivers/net/bonding/rte_eth_bond_pmd.c | 213 ++++++++-------------------------
 1 file changed, 51 insertions(+), 162 deletions(-)
drivers/net/bonding/rte_eth_bond_pmd.c +51 -162
index f30422a..c193d6d 100644
@@ -2,6 +2,7 @@
  * Copyright(c) 2010-2017 Intel Corporation
  */
 #include <stdlib.h>
+#include <stdbool.h>
 #include <netinet/in.h>
 
 #include <rte_mbuf.h>
@@ -295,97 +296,6 @@
 }
 
 static uint16_t
-bond_ethdev_tx_burst_8023ad_fast_queue(void *queue, struct rte_mbuf **bufs,
-		uint16_t nb_bufs)
-{
-	struct bond_tx_queue *bd_tx_q = (struct bond_tx_queue *)queue;
-	struct bond_dev_private *internals = bd_tx_q->dev_private;
-
-	uint16_t slave_port_ids[RTE_MAX_ETHPORTS];
-	uint16_t slave_count;
-
-	uint16_t dist_slave_port_ids[RTE_MAX_ETHPORTS];
-	uint16_t dist_slave_count;
-
-	/* 2-D array to sort mbufs for transmission on each slave into */
-	struct rte_mbuf *slave_bufs[RTE_MAX_ETHPORTS][nb_bufs];
-	/* Number of mbufs for transmission on each slave */
-	uint16_t slave_nb_bufs[RTE_MAX_ETHPORTS] = { 0 };
-	/* Mapping array generated by hash function to map mbufs to slaves */
-	uint16_t bufs_slave_port_idxs[RTE_MAX_ETHPORTS] = { 0 };
-
-	uint16_t slave_tx_count;
-	uint16_t total_tx_count = 0, total_tx_fail_count = 0;
-
-	uint16_t i;
-
-	if (unlikely(nb_bufs == 0))
-		return 0;
-
-	/* Copy slave list to protect against slave up/down changes during tx
-	 * bursting */
-	slave_count = internals->active_slave_count;
-	if (unlikely(slave_count < 1))
-		return 0;
-
-	memcpy(slave_port_ids, internals->active_slaves,
-			sizeof(slave_port_ids[0]) * slave_count);
-
-
-	dist_slave_count = 0;
-	for (i = 0; i < slave_count; i++) {
-		struct port *port = &bond_mode_8023ad_ports[slave_port_ids[i]];
-
-		if (ACTOR_STATE(port, DISTRIBUTING))
-			dist_slave_port_ids[dist_slave_count++] =
-					slave_port_ids[i];
-	}
-
-	if (unlikely(dist_slave_count < 1))
-		return 0;
-
-	/*
-	 * Populate slaves mbuf with the packets which are to be sent on it
-	 * selecting output slave using hash based on xmit policy
-	 */
-	internals->burst_xmit_hash(bufs, nb_bufs, dist_slave_count,
-			bufs_slave_port_idxs);
-
-	for (i = 0; i < nb_bufs; i++) {
-		/* Populate slave mbuf arrays with mbufs for that slave. */
-		uint16_t slave_idx = bufs_slave_port_idxs[i];
-
-		slave_bufs[slave_idx][slave_nb_bufs[slave_idx]++] = bufs[i];
-	}
-
-
-	/* Send packet burst on each slave device */
-	for (i = 0; i < dist_slave_count; i++) {
-		if (slave_nb_bufs[i] == 0)
-			continue;
-
-		slave_tx_count = rte_eth_tx_burst(dist_slave_port_ids[i],
-				bd_tx_q->queue_id, slave_bufs[i],
-				slave_nb_bufs[i]);
-
-		total_tx_count += slave_tx_count;
-
-		/* If tx burst fails move packets to end of bufs */
-		if (unlikely(slave_tx_count < slave_nb_bufs[i])) {
-			int slave_tx_fail_count = slave_nb_bufs[i] -
-					slave_tx_count;
-			total_tx_fail_count += slave_tx_fail_count;
-			memcpy(&bufs[nb_bufs - total_tx_fail_count],
-			       &slave_bufs[i][slave_tx_count],
-			       slave_tx_fail_count * sizeof(bufs[0]));
-		}
-	}
-
-	return total_tx_count;
-}
-
-
-static uint16_t
 bond_ethdev_rx_burst_8023ad(void *queue, struct rte_mbuf **bufs,
 		uint16_t nb_pkts)
 {
@@ -1200,16 +1110,13 @@ struct bwg_slave {
 	return num_tx_total;
 }
 
-static uint16_t
-bond_ethdev_tx_burst_balance(void *queue, struct rte_mbuf **bufs,
-		uint16_t nb_bufs)
+static inline uint16_t
+tx_burst_balance(void *queue, struct rte_mbuf **bufs, uint16_t nb_bufs,
+		 uint16_t *slave_port_ids, uint16_t slave_count)
 {
 	struct bond_tx_queue *bd_tx_q = (struct bond_tx_queue *)queue;
 	struct bond_dev_private *internals = bd_tx_q->dev_private;
 
-	uint16_t slave_port_ids[RTE_MAX_ETHPORTS];
-	uint16_t slave_count;
-
 	/* Array to sort mbufs for transmission on each slave into */
 	struct rte_mbuf *slave_bufs[RTE_MAX_ETHPORTS][nb_bufs];
 	/* Number of mbufs for transmission on each slave */
@@ -1222,18 +1129,6 @@ struct bwg_slave {
 
 	uint16_t i;
 
-	if (unlikely(nb_bufs == 0))
-		return 0;
-
-	/* Copy slave list to protect against slave up/down changes during tx
-	 * bursting */
-	slave_count = internals->active_slave_count;
-	if (unlikely(slave_count < 1))
-		return 0;
-
-	memcpy(slave_port_ids, internals->active_slaves,
-			sizeof(slave_port_ids[0]) * slave_count);
-
 	/*
 	 * Populate slaves mbuf with the packets which are to be sent on it
 	 * selecting output slave using hash based on xmit policy
@@ -1274,7 +1169,7 @@ struct bwg_slave {
 }
 
 static uint16_t
-bond_ethdev_tx_burst_8023ad(void *queue, struct rte_mbuf **bufs,
+bond_ethdev_tx_burst_balance(void *queue, struct rte_mbuf **bufs,
 		uint16_t nb_bufs)
 {
 	struct bond_tx_queue *bd_tx_q = (struct bond_tx_queue *)queue;
@@ -1283,18 +1178,36 @@ struct bwg_slave {
 	uint16_t slave_port_ids[RTE_MAX_ETHPORTS];
 	uint16_t slave_count;
 
+	if (unlikely(nb_bufs == 0))
+		return 0;
+
+	/* Copy slave list to protect against slave up/down changes during tx
+	 * bursting
+	 */
+	slave_count = internals->active_slave_count;
+	if (unlikely(slave_count < 1))
+		return 0;
+
+	memcpy(slave_port_ids, internals->active_slaves,
+			sizeof(slave_port_ids[0]) * slave_count);
+	return tx_burst_balance(queue, bufs, nb_bufs, slave_port_ids,
+				slave_count);
+}
+
+static inline uint16_t
+tx_burst_8023ad(void *queue, struct rte_mbuf **bufs, uint16_t nb_bufs,
+		bool dedicated_txq)
+{
+	struct bond_tx_queue *bd_tx_q = (struct bond_tx_queue *)queue;
+	struct bond_dev_private *internals = bd_tx_q->dev_private;
+
+	uint16_t slave_port_ids[RTE_MAX_ETHPORTS];
+	uint16_t slave_count;
+
 	uint16_t dist_slave_port_ids[RTE_MAX_ETHPORTS];
 	uint16_t dist_slave_count;
 
-	/* 2-D array to sort mbufs for transmission on each slave into */
-	struct rte_mbuf *slave_bufs[RTE_MAX_ETHPORTS][nb_bufs];
-	/* Number of mbufs for transmission on each slave */
-	uint16_t slave_nb_bufs[RTE_MAX_ETHPORTS] = { 0 };
-	/* Mapping array generated by hash function to map mbufs to slaves */
-	uint16_t bufs_slave_port_idxs[RTE_MAX_ETHPORTS] = { 0 };
-
 	uint16_t slave_tx_count;
-	uint16_t total_tx_count = 0, total_tx_fail_count = 0;
 
 	uint16_t i;
 
@@ -1307,6 +1220,9 @@ struct bwg_slave {
 	memcpy(slave_port_ids, internals->active_slaves,
 			sizeof(slave_port_ids[0]) * slave_count);
 
+	if (dedicated_txq)
+		goto skip_tx_ring;
+
 	/* Check for LACP control packets and send if available */
 	for (i = 0; i < slave_count; i++) {
 		struct port *port = &bond_mode_8023ad_ports[slave_port_ids[i]];
@@ -1328,6 +1244,7 @@ struct bwg_slave {
 		}
 	}
 
+skip_tx_ring:
 	if (unlikely(nb_bufs == 0))
 		return 0;
 
@@ -1340,53 +1257,25 @@ struct bwg_slave {
 					slave_port_ids[i];
 	}
 
-	if (likely(dist_slave_count > 0)) {
-
-		/*
-		 * Populate slaves mbuf with the packets which are to be sent
-		 * on it, selecting output slave using hash based on xmit policy
-		 */
-		internals->burst_xmit_hash(bufs, nb_bufs, dist_slave_count,
-				bufs_slave_port_idxs);
-
-		for (i = 0; i < nb_bufs; i++) {
-			/*
-			 * Populate slave mbuf arrays with mbufs for that
-			 * slave
-			 */
-			uint16_t slave_idx = bufs_slave_port_idxs[i];
-
-			slave_bufs[slave_idx][slave_nb_bufs[slave_idx]++] =
-					bufs[i];
-		}
-
-
-		/* Send packet burst on each slave device */
-		for (i = 0; i < dist_slave_count; i++) {
-			if (slave_nb_bufs[i] == 0)
-				continue;
-
-			slave_tx_count = rte_eth_tx_burst(
-					dist_slave_port_ids[i],
-					bd_tx_q->queue_id, slave_bufs[i],
-					slave_nb_bufs[i]);
-
-			total_tx_count += slave_tx_count;
+	if (unlikely(dist_slave_count < 1))
+		return 0;
 
-			/* If tx burst fails move packets to end of bufs */
-			if (unlikely(slave_tx_count < slave_nb_bufs[i])) {
-				int slave_tx_fail_count = slave_nb_bufs[i] -
-						slave_tx_count;
-				total_tx_fail_count += slave_tx_fail_count;
+	return tx_burst_balance(queue, bufs, nb_bufs, dist_slave_port_ids,
+				dist_slave_count);
+}
 
-				memcpy(&bufs[nb_bufs - total_tx_fail_count],
-				       &slave_bufs[i][slave_tx_count],
-				       slave_tx_fail_count * sizeof(bufs[0]));
-			}
-		}
-	}
+static uint16_t
+bond_ethdev_tx_burst_8023ad(void *queue, struct rte_mbuf **bufs,
+		uint16_t nb_bufs)
+{
+	return tx_burst_8023ad(queue, bufs, nb_bufs, false);
+}
 
-	return total_tx_count;
+static uint16_t
+bond_ethdev_tx_burst_8023ad_fast_queue(void *queue, struct rte_mbuf **bufs,
+		uint16_t nb_bufs)
+{
+	return tx_burst_8023ad(queue, bufs, nb_bufs, true);
 }
 
 static uint16_t
Signature · 1 line
1.8.3.1
#2 lore raw

PATCH 2/4 net/bonding: fix LACP fast queue Rx handler

In reply to David Marchand

fast queue Rx burst function is missing checks on promisc and the slave collecting state.
Define an inline wrapper to have a common base.

Fixes 112891cd27e5 ("net/bonding: add dedicated HW queues for LACP control")
---
 drivers/net/bonding/rte_eth_bond_pmd.c | 73 +++++++++++++---------------------
 1 file changed, 27 insertions(+), 46 deletions(-)
drivers/net/bonding/rte_eth_bond_pmd.c +27 -46
index c193d6d..989be5c 100644
@@ -256,48 +256,9 @@
 	return 0;
 }
 
-static uint16_t
-bond_ethdev_rx_burst_8023ad_fast_queue(void *queue, struct rte_mbuf **bufs,
-		uint16_t nb_pkts)
-{
-	struct bond_rx_queue *bd_rx_q = (struct bond_rx_queue *)queue;
-	struct bond_dev_private *internals = bd_rx_q->dev_private;
-	uint16_t num_rx_total = 0;	/* Total number of received packets */
-	uint16_t slaves[RTE_MAX_ETHPORTS];
-	uint16_t slave_count;
-	uint16_t active_slave;
-	uint16_t i;
-
-	/* Copy slave list to protect against slave up/down changes during tx
-	 * bursting */
-	slave_count = internals->active_slave_count;
-	active_slave = internals->active_slave;
-	memcpy(slaves, internals->active_slaves,
-			sizeof(internals->active_slaves[0]) * slave_count);
-
-	for (i = 0; i < slave_count && nb_pkts; i++) {
-		uint16_t num_rx_slave;
-
-		/* Read packets from this slave */
-		num_rx_slave = rte_eth_rx_burst(slaves[active_slave],
-						bd_rx_q->queue_id,
-						bufs + num_rx_total, nb_pkts);
-		num_rx_total += num_rx_slave;
-		nb_pkts -= num_rx_slave;
-
-		if (++active_slave == slave_count)
-			active_slave = 0;
-	}
-
-	if (++internals->active_slave >= slave_count)
-		internals->active_slave = 0;
-
-	return num_rx_total;
-}
-
-static uint16_t
-bond_ethdev_rx_burst_8023ad(void *queue, struct rte_mbuf **bufs,
-		uint16_t nb_pkts)
+static inline uint16_t
+rx_burst_8023ad(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts,
+		bool dedicated_rxq)
 {
 	/* Cast to structure, containing bonded device's port id and queue id */
 	struct bond_rx_queue *bd_rx_q = (struct bond_rx_queue *)queue;
@@ -357,10 +318,16 @@
 			hdr = rte_pktmbuf_mtod(bufs[j], struct ether_hdr *);
 			subtype = ((struct slow_protocol_frame *)hdr)->slow_protocol.subtype;
 
-			/* Remove packet from array if it is slow packet or slave is not
-			 * in collecting state or bonding interface is not in promiscuous
-			 * mode and packet address does not match. */
-			if (unlikely(is_lacp_packets(hdr->ether_type, subtype, bufs[j]) ||
+			/* Remove packet from array if:
+			 * - it is slow packet but no dedicated rxq is present,
+			 * - slave is not in collecting state,
+			 * - bonding interface is not in promiscuous mode and
+			 *   packet is not multicast and address does not match,
+			 */
+			if (unlikely(
+				(!dedicated_rxq &&
+				 is_lacp_packets(hdr->ether_type, subtype,
+						 bufs[j])) ||
 				!collecting ||
 				(!promisc &&
 				 !is_multicast_ether_addr(&hdr->d_addr) &&
@@ -392,6 +359,20 @@
 	return num_rx_total;
 }
 
+static uint16_t
+bond_ethdev_rx_burst_8023ad(void *queue, struct rte_mbuf **bufs,
+		uint16_t nb_pkts)
+{
+	return rx_burst_8023ad(queue, bufs, nb_pkts, false);
+}
+
+static uint16_t
+bond_ethdev_rx_burst_8023ad_fast_queue(void *queue, struct rte_mbuf **bufs,
+		uint16_t nb_pkts)
+{
+	return rx_burst_8023ad(queue, bufs, nb_pkts, true);
+}
+
 #if defined(RTE_LIBRTE_BOND_DEBUG_ALB) || defined(RTE_LIBRTE_BOND_DEBUG_ALB_L1)
 uint32_t burstnumberRX;
 uint32_t burstnumberTX;
Signature · 1 line
1.8.3.1
#3 lore raw

Re: [dpdk-dev] [PATCH 2/4] net/bonding: fix LACP fast queue Rx handler

I should have some time this weekend to run these patches through our regression system.

Quoted 73 lines · David Marchand

fast queue Rx burst function is missing checks on promisc and the slave collecting state.
Define an inline wrapper to have a common base.

Fixes 112891cd27e5 ("net/bonding: add dedicated HW queues for LACP control")
---
  drivers/net/bonding/rte_eth_bond_pmd.c | 73 +++++++++++++---------------------
  1 file changed, 27 insertions(+), 46 deletions(-)
drivers/net/bonding/rte_eth_bond_pmd.c +3 -43
index c193d6d..989be5c 100644
@@ -256,48 +256,9 @@
  	return 0;
  }
 
-static uint16_t
-bond_ethdev_rx_burst_8023ad_fast_queue(void *queue, struct rte_mbuf **bufs,
-		uint16_t nb_pkts)
-{
-	struct bond_rx_queue *bd_rx_q = (struct bond_rx_queue *)queue;
-	struct bond_dev_private *internals = bd_rx_q->dev_private;
-	uint16_t num_rx_total = 0;	/* Total number of received packets */
-	uint16_t slaves[RTE_MAX_ETHPORTS];
-	uint16_t slave_count;
-	uint16_t active_slave;
-	uint16_t i;
-
-	/* Copy slave list to protect against slave up/down changes during tx
-	 * bursting */
-	slave_count = internals->active_slave_count;
-	active_slave = internals->active_slave;
-	memcpy(slaves, internals->active_slaves,
-			sizeof(internals->active_slaves[0]) * slave_count);
-
-	for (i = 0; i < slave_count && nb_pkts; i++) {
-		uint16_t num_rx_slave;
-
-		/* Read packets from this slave */
-		num_rx_slave = rte_eth_rx_burst(slaves[active_slave],
-						bd_rx_q->queue_id,
-						bufs + num_rx_total, nb_pkts);
-		num_rx_total += num_rx_slave;
-		nb_pkts -= num_rx_slave;
-
-		if (++active_slave == slave_count)
-			active_slave = 0;
-	}
-
-	if (++internals->active_slave >= slave_count)
-		internals->active_slave = 0;
-
-	return num_rx_total;
-}
-
-static uint16_t
-bond_ethdev_rx_burst_8023ad(void *queue, struct rte_mbuf **bufs,
-		uint16_t nb_pkts)
+static inline uint16_t
+rx_burst_8023ad(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts,
+		bool dedicated_rxq)
  {
  	/* Cast to structure, containing bonded device's port id and queue id */
  	struct bond_rx_queue *bd_rx_q = (struct bond_rx_queue *)queue;
@@ -357,10 +318,16 @@
  			hdr = rte_pktmbuf_mtod(bufs[j], struct ether_hdr *);
  			subtype = ((struct slow_protocol_frame *)hdr)->slow_protocol.subtype;
 
-			/* Remove packet from array if it is slow packet or slave is not

drivers/net/bonding/rte_eth_bond_pmd.c line 327

Quoted 10 lines
drivers/net/bonding/rte_eth_bond_pmd.c +7 -3
-			 * in collecting state or bonding interface is not in promiscuous
-			 * mode and packet address does not match. */
-			if (unlikely(is_lacp_packets(hdr->ether_type, subtype, bufs[j]) ||
+			/* Remove packet from array if:
+			 * - it is slow packet but no dedicated rxq is present,
+			 * - slave is not in collecting state,
+			 * - bonding interface is not in promiscuous mode and
+			 *   packet is not multicast and address does not match,
+			 */
+			if (unlikely(

The coding style checker doesn't like this:

CHECK:OPEN_ENDED_LINE: Lines should not end with a '('

Quoted 28 lines
drivers/net/bonding/rte_eth_bond_pmd.c +17 -0
+				(!dedicated_rxq &&
+				 is_lacp_packets(hdr->ether_type, subtype,
+						 bufs[j])) ||
  				!collecting ||
  				(!promisc &&
  				 !is_multicast_ether_addr(&hdr->d_addr) &&
@@ -392,6 +359,20 @@
  	return num_rx_total;
  }
 
+static uint16_t
+bond_ethdev_rx_burst_8023ad(void *queue, struct rte_mbuf **bufs,
+		uint16_t nb_pkts)
+{
+	return rx_burst_8023ad(queue, bufs, nb_pkts, false);
+}
+
+static uint16_t
+bond_ethdev_rx_burst_8023ad_fast_queue(void *queue, struct rte_mbuf **bufs,
+		uint16_t nb_pkts)
+{
+	return rx_burst_8023ad(queue, bufs, nb_pkts, true);
+}
+
  #if defined(RTE_LIBRTE_BOND_DEBUG_ALB) || defined(RTE_LIBRTE_BOND_DEBUG_ALB_L1)
  uint32_t burstnumberRX;
  uint32_t burstnumberTX;
#4 lore raw

Re: [dpdk-dev] [PATCH 2/4] net/bonding: fix LACP fast queue Rx handler

Hello Chas,

Quoted 3 lines · Chas Williams

I should have some time this weekend to run these patches through our regression system.

Did you manage to run this series through your tests system ?

Quoted 20 lines · David Marchand

Review line 318

Quoted 2 lines
(quoted diff) +0 -0
@@ -357,10 +318,16 @@
                      hdr = rte_pktmbuf_mtod(bufs[j], struct ether_hdr

*);

Review line 319

Quoted 1 line
(quoted diff) +0 -0
                      subtype = ((struct slow_protocol_frame

*)hdr)->slow_protocol.subtype;

Review line 359

Quoted 2 lines
(quoted diff) +0 -1
-                     /* Remove packet from array if it is slow packet

or slave is not

Review line 360

Quoted 1 line
(quoted diff) +0 -1
-                      * in collecting state or bonding interface is not

in promiscuous

Review line 362

Quoted 2 lines
(quoted diff) +0 -2
-                      * mode and packet address does not match. */
-                     if (unlikely(is_lacp_packets(hdr->ether_type,

subtype, bufs[j]) ||

Review line 321

Quoted 2 lines
(quoted diff) +2 -0
+                     /* Remove packet from array if:
+                      * - it is slow packet but no dedicated rxq is

present,

Review line 323

Quoted 2 lines
(quoted diff) +2 -0
+                      * - slave is not in collecting state,
+                      * - bonding interface is not in promiscuous mode

and

Quoted 9 lines

Review line 324

Quoted 1 line
(quoted diff) +1 -0
+                      *   packet is not multicast and address does not

match,

drivers/net/bonding/rte_eth_bond_pmd.c line 327

Quoted 2 lines
drivers/net/bonding/rte_eth_bond_pmd.c +2 -0
+                      */
+                     if (unlikely(

The coding style checker doesn't like this:

CHECK:OPEN_ENDED_LINE: Lines should not end with a '('

Yes, I had seen this warning, just found it easier to read this way.

Quoted 8 lines
Quoted 6 lines
drivers/net/bonding/rte_eth_bond_pmd.c +3 -0
+                             (!dedicated_rxq &&
+                              is_lacp_packets(hdr->ether_type, subtype,
+                                              bufs[j])) ||
                              !collecting ||
                              (!promisc &&
                               !is_multicast_ether_addr(&hdr->d_addr) &&
Signature · 1 line
David Marchand
#5 lore raw

Re: [dpdk-dev] [PATCH 2/4] net/bonding: fix LACP fast queue Rx handler

Quoted 9 lines · David Marchand

Hello Chas,

On Fri, Apr 12, 2019 at 4:02 PM Chas Williams <[email protected] <mailto:[email protected]>> wrote:

I should have some time this weekend to run these patches through our regression system.

Did you manage to run this series through your tests system ?

There were some other issues over the weekend. Hopefully this one.

Quoted 49 lines · David Marchand

> @@ -357,10 +318,16 @@
>                       hdr = rte_pktmbuf_mtod(bufs[j], struct
ether_hdr *);
>                       subtype = ((struct slow_protocol_frame
*)hdr)->slow_protocol.subtype;
>
> -                     /* Remove packet from array if it is slow
packet or slave is not
> -                      * in collecting state or bonding interface
is not in promiscuous
> -                      * mode and packet address does not match. */ > -                     if
(unlikely(is_lacp_packets(hdr->ether_type, subtype, bufs[j]) ||
> +                     /* Remove packet from array if:
> +                      * - it is slow packet but no dedicated rxq
is present,
> +                      * - slave is not in collecting state, > +                      * - bonding interface is not in
promiscuous mode and
> +                      *   packet is not multicast and address
does not match,
> +                      */
> +                     if (unlikely(

The coding style checker doesn't like this:

CHECK:OPEN_ENDED_LINE: Lines should not end with a '('

Yes, I had seen this warning, just found it easier to read this way.

     > +                             (!dedicated_rxq &&
     > +                              is_lacp_packets(hdr->ether_type,
    subtype,
     > +                                              bufs[j])) ||
     >                               !collecting ||
     >                               (!promisc &&
     >                               
    !is_multicast_ether_addr(&hdr->d_addr) &&
Signature · 1 line
David Marchand
#6 lore raw

Re: [dpdk-dev] [PATCH 2/4] net/bonding: fix LACP fast queue Rx handler

Hello Chas,

Quoted 6 lines · Chas Williams
Quoted 5 lines · David Marchand

Hello Chas,

On Fri, Apr 12, 2019 at 4:02 PM Chas Williams <[email protected] <mailto:[email protected]>> wrote:

Quoted 8 lines
Quoted 5 lines

I should have some time this weekend to run these patches through our regression system.

Did you manage to run this series through your tests system ?

There were some other issues over the weekend. Hopefully this one.

Any update ?
Thanks.

Signature · 1 line
David Marchand
#7 lore raw

Re: [dpdk-dev] [PATCH 2/4] net/bonding: fix LACP fast queue Rx handler

Quoted 14 lines · David Marchand

Hello Chas,

Quoted 10 lines · Chas Williams
Quoted 9 lines · David Marchand

Hello Chas,

On Fri, Apr 12, 2019 at 4:02 PM Chas Williams <[email protected] <mailto:[email protected]>> wrote:

I should have some time this weekend to run these patches through our regression system.

Quoted 9 lines
Quoted 4 lines
Quoted 1 line

Did you manage to run this series through your tests system ?

There were some other issues over the weekend. Hopefully this one.

Any update ?
Thanks.

Reminder of this patchset, if there is no objection in next a few days I will merge them.

Thanks,
ferruh

#8 lore raw

Re: [dpdk-dev] [PATCH 2/4] net/bonding: fix LACP fast queue Rx handler

Quoted 25 lines · Ferruh Yigit
Quoted 14 lines · David Marchand

Hello Chas,

Quoted 10 lines · Chas Williams
Quoted 9 lines · David Marchand

Hello Chas,

On Fri, Apr 12, 2019 at 4:02 PM Chas Williams <[email protected] <mailto:[email protected]>> wrote:

I should have some time this weekend to run these patches through our regression system.

Quoted 9 lines
Quoted 4 lines
Quoted 1 line

Did you manage to run this series through your tests system ?

There were some other issues over the weekend. Hopefully this one.

Any update ?
Thanks.

Quoted 6 lines

Reminder of this patchset, if there is no objection in next a few days I will merge them.

Thanks,
ferruh

OK, I was able to get a clean run for these patches through our regression system. I didn't see any failures with these patches applied. Consider the following:

David Marchand (4):
net/bonding: fix oob access in LACP mode when sending many packets net/bonding: fix LACP fast queue Rx handler net/bonding: fix unicast packets filtering when not in promisc net/bonding: prefer allmulti to promisc for LACP

Sorry this took so long!

#9 lore raw

Re: [dpdk-dev] [PATCH 2/4] net/bonding: fix LACP fast queue Rx handler

Quoted 47 lines · Chas Williams
Quoted 25 lines · Ferruh Yigit
Quoted 14 lines · David Marchand

Hello Chas,

Quoted 10 lines · Chas Williams
Quoted 9 lines · David Marchand

Hello Chas,

On Fri, Apr 12, 2019 at 4:02 PM Chas Williams <[email protected] <mailto:[email protected]>> wrote:

I should have some time this weekend to run these patches through our regression system.

Quoted 9 lines
Quoted 4 lines
Quoted 1 line

Did you manage to run this series through your tests system ?

There were some other issues over the weekend. Hopefully this one.

Any update ?
Thanks.

Quoted 6 lines

Reminder of this patchset, if there is no objection in next a few days I will merge them.

Thanks,
ferruh

OK, I was able to get a clean run for these patches through our regression system. I didn't see any failures with these patches applied. Consider the following:

David Marchand (4):
net/bonding: fix oob access in LACP mode when sending many packets net/bonding: fix LACP fast queue Rx handler net/bonding: fix unicast packets filtering when not in promisc net/bonding: prefer allmulti to promisc for LACP

Quoted 3 lines

Sorry this took so long!

Ah, cool.
Thanks a lot Chas.

Signature · 1 line
David Marchand
#10 lore raw

PATCH 3/4 net/bonding: fix unicast packets filtering when not in promisc

In reply to David Marchand

By default, the 802.3ad code enables promisc mode on all slaves. To avoid all packets going to the application (unless the application asked for promiscuous mode), all frames are supposed to be filtered in the rx burst handler.

However the incriminated commit broke this because non pure ethernet frames (basically any unicast Ether()/IP() packet) are not filtered anymore.

Fixes 71b7b37ec959 ("net/bonding: use ptype flags for LACP Rx filtering")
---
 drivers/net/bonding/rte_eth_bond_pmd.c | 7 -------
 1 file changed, 7 deletions(-)
drivers/net/bonding/rte_eth_bond_pmd.c +0 -7
index 989be5c..9ef0717 100644
@@ -305,13 +305,6 @@
 
 		/* Handle slow protocol packets. */
 		while (j < num_rx_total) {
-
-			/* If packet is not pure L2 and is known, skip it */
-			if ((bufs[j]->packet_type & ~RTE_PTYPE_L2_ETHER) != 0) {
-				j++;
-				continue;
-			}
-
 			if (j + 3 < num_rx_total)
 				rte_prefetch0(rte_pktmbuf_mtod(bufs[j + 3], void *));
 
Signature · 1 line
1.8.3.1
#11 lore raw

PATCH 4/4 net/bonding: prefer allmulti to promisc for LACP

In reply to David Marchand

Rather than the global promiscuous mode on all slaves, let's try to use allmulti.
To do this, we apply the same mechanism than for promiscuous and drop in rx_burst.

When adding a slave, we first try to use allmulti, else we try promiscuous.
Because of this, we must be able to handle allmulti on the bonding interface, so the associated dev_ops is added with checks on which mode has been applied on each slave.

Rather than add a new flag in the bond private structure, we can remove promiscuous_en since ethdev already maintains this information. When starting the bond device, there is no promisc/allmulti re-apply as, again, ethdev does this itself.

On reception, we must inspect if the packets are multicast or unicast and take a decision to drop based on which mode has been enabled on the bonding interface.

Note: in the future, if we define an API so that we can add/remove mc addresses to a port (rather than the current global set_mc_addr_list API), we can refine this to only register the LACP multicast mac address.

---
 drivers/net/bonding/rte_eth_bond_8023ad.c         |  51 +++++++++-
 drivers/net/bonding/rte_eth_bond_8023ad_private.h |   7 ++
 drivers/net/bonding/rte_eth_bond_pmd.c            | 113 +++++++++++++++++-----
 drivers/net/bonding/rte_eth_bond_private.h        |   3 -
 4 files changed, 148 insertions(+), 26 deletions(-)
drivers/net/bonding/rte_eth_bond_8023ad.c +50 -1
index 5004898..adf6b7e 100644
@@ -907,6 +907,49 @@
 			bond_mode_8023ad_periodic_cb, arg);
 }
 
+static int
+bond_mode_8023ad_register_lacp_mac(uint16_t slave_id)
+{
+	rte_eth_allmulticast_enable(slave_id);
+	if (rte_eth_allmulticast_get(slave_id)) {
+		RTE_BOND_LOG(DEBUG, "forced allmulti for port %u",
+			     slave_id);
+		bond_mode_8023ad_ports[slave_id].forced_rx_flags =
+				BOND_8023AD_FORCED_ALLMULTI;
+		return 0;
+	}
+
+	rte_eth_promiscuous_enable(slave_id);
+	if (rte_eth_promiscuous_get(slave_id)) {
+		RTE_BOND_LOG(DEBUG, "forced promiscuous for port %u",
+			     slave_id);
+		bond_mode_8023ad_ports[slave_id].forced_rx_flags =
+				BOND_8023AD_FORCED_PROMISC;
+		return 0;
+	}
+
+	return -1;
+}
+
+static void
+bond_mode_8023ad_unregister_lacp_mac(uint16_t slave_id)
+{
+	switch (bond_mode_8023ad_ports[slave_id].forced_rx_flags) {
+	case BOND_8023AD_FORCED_ALLMULTI:
+		RTE_BOND_LOG(DEBUG, "unset allmulti for port %u", slave_id);
+		rte_eth_allmulticast_disable(slave_id);
+		break;
+
+	case BOND_8023AD_FORCED_PROMISC:
+		RTE_BOND_LOG(DEBUG, "unset promisc for port %u", slave_id);
+		rte_eth_promiscuous_disable(slave_id);
+		break;
+
+	default:
+		break;
+	}
+}
+
 void
 bond_mode_8023ad_activate_slave(struct rte_eth_dev *bond_dev,
 				uint16_t slave_id)
@@ -948,7 +991,11 @@
 
 	/* use this port as agregator */
 	port->aggregator_port_id = slave_id;
-	rte_eth_promiscuous_enable(slave_id);
+
+	if (bond_mode_8023ad_register_lacp_mac(slave_id) < 0) {
+		RTE_BOND_LOG(WARNING, "slave %u is most likely broken and won't receive LACP packets",
+			     slave_id);
+	}
 
 	timer_cancel(&port->warning_timer);
 
@@ -1022,6 +1069,8 @@
 	old_partner_state = port->partner_state;
 	record_default(port);
 
+	bond_mode_8023ad_unregister_lacp_mac(slave_id);
+
 	/* If partner timeout state changes then disable timer */
 	if (!((old_partner_state ^ port->partner_state) &
 			STATE_LACP_SHORT_TIMEOUT))
drivers/net/bonding/rte_eth_bond_8023ad_private.h +7 -0
index f91902e..edda841 100644
@@ -109,6 +109,13 @@ struct port {
 	uint16_t sm_flags;
 	enum rte_bond_8023ad_selection selected;
 
+	/** Indicates if either allmulti or promisc has been enforced on the
+	 * slave so that we can receive lacp packets
+	 */
+#define BOND_8023AD_FORCED_ALLMULTI (1 << 0)
+#define BOND_8023AD_FORCED_PROMISC (1 << 1)
+	uint8_t forced_rx_flags;
+
 	uint64_t current_while_timer;
 	uint64_t periodic_timer;
 	uint64_t wait_while_timer;
drivers/net/bonding/rte_eth_bond_pmd.c +91 -22

173 lines not on this page · raw message

drivers/net/bonding/rte_eth_bond_private.h +0 -3
index 8afef39..4154e17 100644
@@ -122,9 +122,6 @@ struct bond_dev_private {
 
 	uint8_t user_defined_mac;
 	/**< Flag for whether MAC address is user defined or not */
-	uint8_t promiscuous_en;
-	/**< Enabled/disable promiscuous mode on bonding device */
-
 
 	uint8_t link_status_polling_enabled;
 	uint32_t link_status_polling_interval_ms;
Signature · 1 line
1.8.3.1
#12 lore raw
In reply to David Marchand
Quoted 11 lines · David Marchand

Another series with focus on the fast/normal rx/tx handlers for 802.3ad.

The first two patches make sure that the rx (resp. tx) fast and normal handlers are equivalent.

The third one will most likely have an impact on performance which I tried to mitigate with the last one. However, I have no benchmark to back those patches and I did not test those thoroughly, so they are more like RFC patches but sending anyway.

Hi Chas,

Reminder of this patchset, it is waiting for a long time, if there is no objection I am for merging it.

Thanks,
ferruh

#13 lore raw
Quoted 13 lines · Ferruh Yigit
Quoted 11 lines · David Marchand

Another series with focus on the fast/normal rx/tx handlers for 802.3ad.

The first two patches make sure that the rx (resp. tx) fast and normal handlers are equivalent.

The third one will most likely have an impact on performance which I tried to mitigate with the last one. However, I have no benchmark to back those patches and I did not test those thoroughly, so they are more like RFC patches but sending anyway.

Quoted 4 lines

Hi Chas,

Reminder of this patchset, it is waiting for a long time, if there is no objection I am for merging it.

These patches caused some additional regressions in our local tests. I haven't had time to investigate it though.

Quoted 4 lines

Thanks,
ferruh

#14 lore raw
In reply to Ferruh Yigit
Quoted 12 lines · Ferruh Yigit
Quoted 11 lines · David Marchand

Another series with focus on the fast/normal rx/tx handlers for 802.3ad.

The first two patches make sure that the rx (resp. tx) fast and normal handlers are equivalent.

The third one will most likely have an impact on performance which I tried to mitigate with the last one. However, I have no benchmark to back those patches and I did not test those thoroughly, so they are more like RFC patches but sending anyway.

Quoted 4 lines

Hi Chas,

Reminder of this patchset, it is waiting for a long time, if there is no objection I am for merging it.

It caused some additional failure in our regression testing but I haven't had time to look at it.

Quoted 4 lines

Thanks,
ferruh

Acked-by Chas Williams on the v1 series
#15 lore raw
In reply to David Marchand
Quoted 11 lines · David Marchand

Another series with focus on the fast/normal rx/tx handlers for 802.3ad.

The first two patches make sure that the rx (resp. tx) fast and normal handlers are equivalent.

The third one will most likely have an impact on performance which I tried to mitigate with the last one. However, I have no benchmark to back those patches and I did not test those thoroughly, so they are more like RFC patches but sending anyway.

for series,

Series applied to dpdk-next-net/master, thanks.