[PATCH rdma-next v1 0/2] Add gratuitous ARP support to RDMA-CM

[PATCH rdma-next v1 0/2] Add gratuitous ARP support to RDMA-CM

[Leon Romanovsky]

From: Leon Romanovsky <leonro@nvidia.com>

Changelog:
v1: 
 * Removed special workqueue
 * Rewrote compare_netdev_and_ip()
v0: https://lore.kernel.org/all/cover.1649075034.git.leonro@nvidia.com

In this series, Patrisious adds gratuitous ARP support to RDMA-CM, in
order to speed up migration failover from one node to another.

Thanks

Patrisious Haddad (2):
  RDMA/core: Add an rb_tree that stores cm_ids sorted by ifindex and
    remote IP
  RDMA/core: Add a netevent notifier to cma

 drivers/infiniband/core/cma.c      | 234 +++++++++++++++++++++++++++--
 drivers/infiniband/core/cma_priv.h |   1 +
 include/rdma/rdma_cm.h             |   6 +
 3 files changed, 229 insertions(+), 12 deletions(-)

-- 
2.36.1

[PATCH rdma-next v1 1/2] RDMA/core: Add an rb_tree that stores cm_ids sorted by ifindex and remote IP

[Leon Romanovsky]

From: Patrisious Haddad <phaddad@nvidia.com>

Add to the cma, a tree that keeps track of all rdma_id_private channels
that were created while in RoCE mode.

The IDs are sorted first according to their netdevice ifindex then their
destination IP. And for IDs with matching IP they would be at the same node
in the tree, since the tree data is a list of all ids with matching destination IP.

The tree allows fast and efficient lookup of ids using an ifindex and
IP address which is useful for identifying relevant net_events promptly.

Signed-off-by: Patrisious Haddad <phaddad@nvidia.com>
Reviewed-by: Mark Zhang <markzhang@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
---
 drivers/infiniband/core/cma.c      | 151 ++++++++++++++++++++++++++---
 drivers/infiniband/core/cma_priv.h |   1 +
 2 files changed, 140 insertions(+), 12 deletions(-)

diff --git a/drivers/infiniband/core/cma.c b/drivers/infiniband/core/cma.c
index fabca5e51e3d..08bc3ea19716 100644
--- a/drivers/infiniband/core/cma.c
+++ b/drivers/infiniband/core/cma.c
@@ -11,6 +11,7 @@
 #include <linux/in6.h>
 #include <linux/mutex.h>
 #include <linux/random.h>
+#include <linux/rbtree.h>
 #include <linux/igmp.h>
 #include <linux/xarray.h>
 #include <linux/inetdevice.h>
@@ -168,6 +169,9 @@ static struct ib_sa_client sa_client;
 static LIST_HEAD(dev_list);
 static LIST_HEAD(listen_any_list);
 static DEFINE_MUTEX(lock);
+static struct rb_root id_table = RB_ROOT;
+/* Serialize operations of id_table tree */
+static DEFINE_SPINLOCK(id_table_lock);
 static struct workqueue_struct *cma_wq;
 static unsigned int cma_pernet_id;
 
@@ -202,6 +206,11 @@ struct xarray *cma_pernet_xa(struct net *net, enum rdma_ucm_port_space ps)
 	}
 }
 
+struct id_table_entry {
+	struct list_head id_list;
+	struct rb_node rb_node;
+};
+
 struct cma_device {
 	struct list_head	list;
 	struct ib_device	*device;
@@ -420,11 +429,21 @@ static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr)
 	return hdr->ip_version >> 4;
 }
 
-static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
+static void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 {
 	hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 }
 
+static struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
+{
+	return (struct sockaddr *)&id_priv->id.route.addr.src_addr;
+}
+
+static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
+{
+	return (struct sockaddr *)&id_priv->id.route.addr.dst_addr;
+}
+
 static int cma_igmp_send(struct net_device *ndev, union ib_gid *mgid, bool join)
 {
 	struct in_device *in_dev = NULL;
@@ -445,6 +464,119 @@ static int cma_igmp_send(struct net_device *ndev, union ib_gid *mgid, bool join)
 	return (in_dev) ? 0 : -ENODEV;
 }
 
+static int compare_netdev_and_ip(int ifindex_a, struct sockaddr *sa,
+				 struct id_table_entry *entry_b)
+{
+	struct rdma_id_private *id_priv = list_first_entry(
+		&entry_b->id_list, struct rdma_id_private, id_list_entry);
+	int ifindex_b = id_priv->id.route.addr.dev_addr.bound_dev_if;
+	struct sockaddr *sb = cma_dst_addr(id_priv);
+
+	if (ifindex_a != ifindex_b)
+		return (ifindex_a > ifindex_b) ? 1 : -1;
+
+	if (sa->sa_family != sb->sa_family)
+		return sa->sa_family - sb->sa_family;
+
+	if (sa->sa_family == AF_INET)
+		return (int)__be32_to_cpu(
+			       ((struct sockaddr_in *)sa)->sin_addr.s_addr) -
+		       (int)__be32_to_cpu(
+			       ((struct sockaddr_in *)sb)->sin_addr.s_addr);
+
+	return memcmp((char *)&((struct sockaddr_in6 *)sa)->sin6_addr,
+		      (char *)&((struct sockaddr_in6 *)sb)->sin6_addr,
+		      sizeof(((struct sockaddr_in6 *)sa)->sin6_addr));
+}
+
+static int cma_add_id_to_tree(struct rdma_id_private *node_id_priv)
+{
+	struct rb_node **new, *parent = NULL;
+	struct id_table_entry *this, *node;
+	unsigned long flags;
+	int result;
+
+	node = kzalloc(sizeof(*node), GFP_KERNEL);
+	if (!node)
+		return -ENOMEM;
+
+	spin_lock_irqsave(&id_table_lock, flags);
+	new = &id_table.rb_node;
+	while (*new) {
+		this = container_of(*new, struct id_table_entry, rb_node);
+		result = compare_netdev_and_ip(
+			node_id_priv->id.route.addr.dev_addr.bound_dev_if,
+			cma_dst_addr(node_id_priv), this);
+
+		parent = *new;
+		if (result < 0)
+			new = &((*new)->rb_left);
+		else if (result > 0)
+			new = &((*new)->rb_right);
+		else {
+			list_add_tail(&node_id_priv->id_list_entry,
+				      &this->id_list);
+			kfree(node);
+			goto unlock;
+		}
+	}
+
+	INIT_LIST_HEAD(&node->id_list);
+	list_add_tail(&node_id_priv->id_list_entry, &node->id_list);
+
+	rb_link_node(&node->rb_node, parent, new);
+	rb_insert_color(&node->rb_node, &id_table);
+
+unlock:
+	spin_unlock_irqrestore(&id_table_lock, flags);
+	return 0;
+}
+
+static struct id_table_entry *
+node_from_ndev_ip(struct rb_root *root, int ifindex, struct sockaddr *sa)
+{
+	struct rb_node *node = root->rb_node;
+	struct id_table_entry *data;
+	int result;
+
+	while (node) {
+		data = container_of(node, struct id_table_entry, rb_node);
+		result = compare_netdev_and_ip(ifindex, sa, data);
+		if (result < 0)
+			node = node->rb_left;
+		else if (result > 0)
+			node = node->rb_right;
+		else
+			return data;
+	}
+
+	return NULL;
+}
+
+static void cma_remove_id_from_tree(struct rdma_id_private *id_priv)
+{
+	struct id_table_entry *data;
+	unsigned long flags;
+
+	spin_lock_irqsave(&id_table_lock, flags);
+	if (list_empty(&id_priv->id_list_entry))
+		goto out;
+
+	data = node_from_ndev_ip(&id_table,
+				 id_priv->id.route.addr.dev_addr.bound_dev_if,
+				 cma_dst_addr(id_priv));
+	if (!data)
+		goto out;
+
+	list_del_init(&id_priv->id_list_entry);
+	if (list_empty(&data->id_list)) {
+		rb_erase(&data->rb_node, &id_table);
+		kfree(data);
+	}
+out:
+	spin_unlock_irqrestore(&id_table_lock, flags);
+}
+
 static void _cma_attach_to_dev(struct rdma_id_private *id_priv,
 			       struct cma_device *cma_dev)
 {
@@ -481,16 +613,6 @@ static void cma_release_dev(struct rdma_id_private *id_priv)
 	mutex_unlock(&lock);
 }
 
-static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
-{
-	return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
-}
-
-static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
-{
-	return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
-}
-
 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
 {
 	return id_priv->id.route.addr.src_addr.ss_family;
@@ -861,6 +983,7 @@ __rdma_create_id(struct net *net, rdma_cm_event_handler event_handler,
 	refcount_set(&id_priv->refcount, 1);
 	mutex_init(&id_priv->handler_mutex);
 	INIT_LIST_HEAD(&id_priv->device_item);
+	INIT_LIST_HEAD(&id_priv->id_list_entry);
 	INIT_LIST_HEAD(&id_priv->listen_list);
 	INIT_LIST_HEAD(&id_priv->mc_list);
 	get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
@@ -1883,6 +2006,7 @@ static void _destroy_id(struct rdma_id_private *id_priv,
 	cma_cancel_operation(id_priv, state);
 
 	rdma_restrack_del(&id_priv->res);
+	cma_remove_id_from_tree(id_priv);
 	if (id_priv->cma_dev) {
 		if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
 			if (id_priv->cm_id.ib)
@@ -3172,8 +3296,11 @@ int rdma_resolve_route(struct rdma_cm_id *id, unsigned long timeout_ms)
 	cma_id_get(id_priv);
 	if (rdma_cap_ib_sa(id->device, id->port_num))
 		ret = cma_resolve_ib_route(id_priv, timeout_ms);
-	else if (rdma_protocol_roce(id->device, id->port_num))
+	else if (rdma_protocol_roce(id->device, id->port_num)) {
 		ret = cma_resolve_iboe_route(id_priv);
+		if (!ret)
+			cma_add_id_to_tree(id_priv);
+	}
 	else if (rdma_protocol_iwarp(id->device, id->port_num))
 		ret = cma_resolve_iw_route(id_priv);
 	else
diff --git a/drivers/infiniband/core/cma_priv.h b/drivers/infiniband/core/cma_priv.h
index 757a0ef79872..b7354c94cf1b 100644
--- a/drivers/infiniband/core/cma_priv.h
+++ b/drivers/infiniband/core/cma_priv.h
@@ -64,6 +64,7 @@ struct rdma_id_private {
 		struct list_head listen_item;
 		struct list_head listen_list;
 	};
+	struct list_head        id_list_entry;
 	struct cma_device	*cma_dev;
 	struct list_head	mc_list;
 
-- 
2.36.1

[PATCH rdma-next v1 2/2] RDMA/core: Add a netevent notifier to cma

[Leon Romanovsky]

From: Patrisious Haddad <phaddad@nvidia.com>

Add a netevent callback for cma, mainly to catch NETEVENT_NEIGH_UPDATE.

Previously, when a system with failover MAC mechanism change its MAC address
during a CM connection attempt, the RDMA-CM would take a lot of time till
it disconnects and timesout due to the incorrect MAC address.

Now when we get a NETEVENT_NEIGH_UPDATE we check if it is due to a failover
MAC change and if so, we instantly destroy the CM and notify the user in order
to spare the unnecessary waiting for the timeout.

Signed-off-by: Patrisious Haddad <phaddad@nvidia.com>
Reviewed-by: Mark Zhang <markzhang@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
---
 drivers/infiniband/core/cma.c | 83 +++++++++++++++++++++++++++++++++++
 include/rdma/rdma_cm.h        |  6 +++
 2 files changed, 89 insertions(+)

diff --git a/drivers/infiniband/core/cma.c b/drivers/infiniband/core/cma.c
index 08bc3ea19716..644f5f1e1f46 100644
--- a/drivers/infiniband/core/cma.c
+++ b/drivers/infiniband/core/cma.c
@@ -21,6 +21,7 @@
 
 #include <net/net_namespace.h>
 #include <net/netns/generic.h>
+#include <net/netevent.h>
 #include <net/tcp.h>
 #include <net/ipv6.h>
 #include <net/ip_fib.h>
@@ -5049,10 +5050,89 @@ static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
 	return ret;
 }
 
+static void cma_netevent_work_handler(struct work_struct *_work)
+{
+	struct cma_netevent_work *network =
+		container_of(_work, struct cma_netevent_work, work);
+	struct rdma_cm_event event = {};
+
+	mutex_lock(&network->id_priv->handler_mutex);
+
+	if (READ_ONCE(network->id_priv->state) == RDMA_CM_DESTROYING ||
+	    READ_ONCE(network->id_priv->state) == RDMA_CM_DEVICE_REMOVAL)
+		goto out_unlock;
+
+	event.event = RDMA_CM_EVENT_UNREACHABLE;
+	event.status = -ETIMEDOUT;
+
+	if (cma_cm_event_handler(network->id_priv, &event)) {
+		__acquire(&network->id_priv->handler_mutex);
+		network->id_priv->cm_id.ib = NULL;
+		cma_id_put(network->id_priv);
+		destroy_id_handler_unlock(network->id_priv);
+		return;
+	}
+
+out_unlock:
+	mutex_unlock(&network->id_priv->handler_mutex);
+	cma_id_put(network->id_priv);
+}
+
+static int cma_netevent_callback(struct notifier_block *self,
+				 unsigned long event, void *ctx)
+{
+	struct id_table_entry *ips_node = NULL;
+	struct rdma_id_private *current_id;
+	struct neighbour *neigh = ctx;
+	unsigned long flags;
+
+	if (event != NETEVENT_NEIGH_UPDATE)
+		return NOTIFY_DONE;
+
+	spin_lock_irqsave(&id_table_lock, flags);
+	if (neigh->tbl->family == AF_INET6) {
+		struct sockaddr_in6 neigh_sock_6;
+
+		neigh_sock_6.sin6_family = AF_INET6;
+		neigh_sock_6.sin6_addr = *(struct in6_addr *)neigh->primary_key;
+		ips_node = node_from_ndev_ip(&id_table, neigh->dev->ifindex,
+					     (struct sockaddr *)&neigh_sock_6);
+	} else if (neigh->tbl->family == AF_INET) {
+		struct sockaddr_in neigh_sock_4;
+
+		neigh_sock_4.sin_family = AF_INET;
+		neigh_sock_4.sin_addr.s_addr = *(__be32 *)(neigh->primary_key);
+		ips_node = node_from_ndev_ip(&id_table, neigh->dev->ifindex,
+					     (struct sockaddr *)&neigh_sock_4);
+	} else
+		goto out;
+
+	if (!ips_node)
+		goto out;
+
+	list_for_each_entry(current_id, &ips_node->id_list, id_list_entry) {
+		if (!memcmp(current_id->id.route.addr.dev_addr.dst_dev_addr,
+			   neigh->ha, ETH_ALEN))
+			continue;
+		INIT_WORK(&current_id->id.network.work,
+			  cma_netevent_work_handler);
+		current_id->id.network.id_priv = current_id;
+		cma_id_get(current_id);
+		queue_work(cma_wq, &current_id->id.network.work);
+	}
+out:
+	spin_unlock_irqrestore(&id_table_lock, flags);
+	return NOTIFY_DONE;
+}
+
 static struct notifier_block cma_nb = {
 	.notifier_call = cma_netdev_callback
 };
 
+static struct notifier_block cma_netevent_cb = {
+	.notifier_call = cma_netevent_callback
+};
+
 static void cma_send_device_removal_put(struct rdma_id_private *id_priv)
 {
 	struct rdma_cm_event event = { .event = RDMA_CM_EVENT_DEVICE_REMOVAL };
@@ -5275,6 +5355,7 @@ static int __init cma_init(void)
 
 	ib_sa_register_client(&sa_client);
 	register_netdevice_notifier(&cma_nb);
+	register_netevent_notifier(&cma_netevent_cb);
 
 	ret = ib_register_client(&cma_client);
 	if (ret)
@@ -5289,6 +5370,7 @@ static int __init cma_init(void)
 err_ib:
 	ib_unregister_client(&cma_client);
 err:
+	unregister_netevent_notifier(&cma_netevent_cb);
 	unregister_netdevice_notifier(&cma_nb);
 	ib_sa_unregister_client(&sa_client);
 	unregister_pernet_subsys(&cma_pernet_operations);
@@ -5301,6 +5383,7 @@ static void __exit cma_cleanup(void)
 {
 	cma_configfs_exit();
 	ib_unregister_client(&cma_client);
+	unregister_netevent_notifier(&cma_netevent_cb);
 	unregister_netdevice_notifier(&cma_nb);
 	ib_sa_unregister_client(&sa_client);
 	unregister_pernet_subsys(&cma_pernet_operations);
diff --git a/include/rdma/rdma_cm.h b/include/rdma/rdma_cm.h
index d989f030fae0..d7de0958b76a 100644
--- a/include/rdma/rdma_cm.h
+++ b/include/rdma/rdma_cm.h
@@ -87,6 +87,11 @@ struct rdma_cm_event {
 	struct rdma_ucm_ece ece;
 };
 
+struct cma_netevent_work {
+	struct work_struct work;
+	struct rdma_id_private *id_priv;
+};
+
 struct rdma_cm_id;
 
 /**
@@ -108,6 +113,7 @@ struct rdma_cm_id {
 	enum rdma_ucm_port_space ps;
 	enum ib_qp_type		 qp_type;
 	u32			 port_num;
+	struct cma_netevent_work network;
 };
 
 struct rdma_cm_id *
-- 
2.36.1

Re: [PATCH rdma-next v1 1/2] RDMA/core: Add an rb_tree that stores cm_ids sorted by ifindex and remote IP

[Jason Gunthorpe]

On Thu, May 19, 2022 at 07:41:22AM +0300, Leon Romanovsky wrote:

> +static int compare_netdev_and_ip(int ifindex_a, struct sockaddr *sa,
> +				 struct id_table_entry *entry_b)
> +{
> +	struct rdma_id_private *id_priv = list_first_entry(
> +		&entry_b->id_list, struct rdma_id_private, id_list_entry);
> +	int ifindex_b = id_priv->id.route.addr.dev_addr.bound_dev_if;
> +	struct sockaddr *sb = cma_dst_addr(id_priv);
> +
> +	if (ifindex_a != ifindex_b)
> +		return (ifindex_a > ifindex_b) ? 1 : -1;
> +
> +	if (sa->sa_family != sb->sa_family)
> +		return sa->sa_family - sb->sa_family;
> +
> +	if (sa->sa_family == AF_INET)
> +		return (int)__be32_to_cpu(
> +			       ((struct sockaddr_in *)sa)->sin_addr.s_addr) -
> +		       (int)__be32_to_cpu(
> +			       ((struct sockaddr_in *)sb)->sin_addr.s_addr);

This still overflows, just use memcmp

> +	return memcmp((char *)&((struct sockaddr_in6 *)sa)->sin6_addr,
> +		      (char *)&((struct sockaddr_in6 *)sb)->sin6_addr,
> +		      sizeof(((struct sockaddr_in6 *)sa)->sin6_addr));

This is ipv6_addr_cmp()

Jason

Re: [PATCH rdma-next v1 2/2] RDMA/core: Add a netevent notifier to cma

[Jason Gunthorpe]

On Thu, May 19, 2022 at 07:41:23AM +0300, Leon Romanovsky wrote:
> From: Patrisious Haddad <phaddad@nvidia.com>
> 
> Add a netevent callback for cma, mainly to catch NETEVENT_NEIGH_UPDATE.
> 
> Previously, when a system with failover MAC mechanism change its MAC address
> during a CM connection attempt, the RDMA-CM would take a lot of time till
> it disconnects and timesout due to the incorrect MAC address.
> 
> Now when we get a NETEVENT_NEIGH_UPDATE we check if it is due to a failover
> MAC change and if so, we instantly destroy the CM and notify the user in order
> to spare the unnecessary waiting for the timeout.
> 
> Signed-off-by: Patrisious Haddad <phaddad@nvidia.com>
> Reviewed-by: Mark Zhang <markzhang@nvidia.com>
> Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
>  drivers/infiniband/core/cma.c | 83 +++++++++++++++++++++++++++++++++++
>  include/rdma/rdma_cm.h        |  6 +++
>  2 files changed, 89 insertions(+)
> 
> diff --git a/drivers/infiniband/core/cma.c b/drivers/infiniband/core/cma.c
> index 08bc3ea19716..644f5f1e1f46 100644
> +++ b/drivers/infiniband/core/cma.c
> @@ -21,6 +21,7 @@
>  
>  #include <net/net_namespace.h>
>  #include <net/netns/generic.h>
> +#include <net/netevent.h>
>  #include <net/tcp.h>
>  #include <net/ipv6.h>
>  #include <net/ip_fib.h>
> @@ -5049,10 +5050,89 @@ static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
>  	return ret;
>  }
>  
> +static void cma_netevent_work_handler(struct work_struct *_work)
> +{
> +	struct cma_netevent_work *network =
> +		container_of(_work, struct cma_netevent_work, work);

This is just

	struct rdma_id_private *id_priv =
		container_of(_work, struct rdma_id_private, id.net_work);

> +struct cma_netevent_work {
> +	struct work_struct work;
> +	struct rdma_id_private *id_priv;
> +};

And this isn't needed.

Jason