Skip to contentSkip to conversation

Support multiple path records

PATCH Applied Leon Romanovsky 3 participants 12 messages Sep 8, 2022 – Sep 23, 2022

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

cover Support multiple path records

From: Leon Romanovsky <[email protected]>

From Mark:

These patches allow IB core to receive multiple path records from user-space rdma netlink service.

Currently only one GMP PR is supported when doing a PR query. This means in a fabric with multiple subnets, when a packet goes out of the sender, it’s assigned with a LID to a router as its destination LID. The current solution selects a specific router per destination, means that there is no advanced routing from the sender towards the network.

This patchset supports to receive an inbound PR and an outbound PR along with the GMP PR. The LIDs in these 3 PRs can be used in this way:
1. GMP PR: used as the standard local/remote LIDs;
2. DLID of outbound PR: Used as the "dlid" field for outbound traffic;
3. DLID of inbound PR: Used as the "dlid" field for outbound traffic in responder side.

The inboundPR.dlid is passed to responder with the "primary_LID" filed in the ConnectRequest message.

With this, the user-space rdma netlink service can set special DLIDs in inbound/outbound PRs, to select specific routers for datapath between the 2 nodes.

The following cases were tested:
- New kernel with new netlink service that supports multiple path records;
- New kernel with old netlink service;
- Old kernel with new netlink service;
- Client side new kernel with new netlink service, server side with old kernel.

Thanks.

Mark Zhang (4):
RDMA/core: Rename rdma_route.num_paths field to num_pri_alt_paths RDMA/cma: Multiple path records support with netlink channel RDMA/cm: Use SLID in the work completion as the DLID in responder side RDMA/cm: Use DLID from inbound/outbound PathRecords as the datapath
DLID

 drivers/infiniband/core/cm.c              |  39 +++-
 drivers/infiniband/core/cma.c             |  88 ++++++--
 drivers/infiniband/core/sa_query.c        | 235 +++++++++++++++-------
 drivers/infiniband/core/ucma.c            |  10 +-
 drivers/infiniband/ulp/ipoib/ipoib_main.c |   2 +-
 drivers/infiniband/ulp/srp/ib_srp.c       |   2 +-
 include/rdma/ib_cm.h                      |   2 +
 include/rdma/ib_sa.h                      |   3 +-
 include/rdma/rdma_cm.h                    |  13 +-
 9 files changed, 284 insertions(+), 110 deletions(-)
Signature · 1 line
2.37.2
#1 lore raw

PATCH 1/4 RDMA/core: Rename rdma_route.num_paths field to num_pri_alt_paths

From: Mark Zhang <[email protected]>

This fields means the total number of primary and alternative paths, i.e.,:
0 - No primary nor alternate path is available;
1 - Only primary path is available;
2 - Both primary and alternate path are available.
Rename it to avoid confusion as with follow patches primary path will support multiple path records.

---
 drivers/infiniband/core/cma.c  | 18 +++++++++---------
 drivers/infiniband/core/ucma.c | 10 +++++-----
 include/rdma/rdma_cm.h         |  7 ++++++-
 3 files changed, 20 insertions(+), 15 deletions(-)
drivers/infiniband/core/cma.c +9 -9
index 46d06678dfbe..91e72a76d95e 100644
@@ -2241,14 +2241,14 @@ cma_ib_new_conn_id(const struct rdma_cm_id *listen_id,
 		goto err;
 
 	rt = &id->route;
-	rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
-	rt->path_rec = kmalloc_array(rt->num_paths, sizeof(*rt->path_rec),
-				     GFP_KERNEL);
+	rt->num_pri_alt_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
+	rt->path_rec = kmalloc_array(rt->num_pri_alt_paths,
+				     sizeof(*rt->path_rec), GFP_KERNEL);
 	if (!rt->path_rec)
 		goto err;
 
 	rt->path_rec[0] = *path;
-	if (rt->num_paths == 2)
+	if (rt->num_pri_alt_paths == 2)
 		rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
 
 	if (net_dev) {
@@ -2826,7 +2826,7 @@ static void cma_query_handler(int status, struct sa_path_rec *path_rec,
 	route = &work->id->id.route;
 
 	if (!status) {
-		route->num_paths = 1;
+		route->num_pri_alt_paths = 1;
 		*route->path_rec = *path_rec;
 	} else {
 		work->old_state = RDMA_CM_ROUTE_QUERY;
@@ -3081,7 +3081,7 @@ int rdma_set_ib_path(struct rdma_cm_id *id,
 		dev_put(ndev);
 	}
 
-	id->route.num_paths = 1;
+	id->route.num_pri_alt_paths = 1;
 	return 0;
 
 err_free:
@@ -3214,7 +3214,7 @@ static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
 		goto err1;
 	}
 
-	route->num_paths = 1;
+	route->num_pri_alt_paths = 1;
 
 	ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
 	if (!ndev) {
@@ -3274,7 +3274,7 @@ static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
 err2:
 	kfree(route->path_rec);
 	route->path_rec = NULL;
-	route->num_paths = 0;
+	route->num_pri_alt_paths = 0;
 err1:
 	kfree(work);
 	return ret;
@@ -4265,7 +4265,7 @@ static int cma_connect_ib(struct rdma_id_private *id_priv,
 	}
 
 	req.primary_path = &route->path_rec[0];
-	if (route->num_paths == 2)
+	if (route->num_pri_alt_paths == 2)
 		req.alternate_path = &route->path_rec[1];
 
 	req.ppath_sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
drivers/infiniband/core/ucma.c +5 -5
index 9d6ac9dff39a..bf42650f125b 100644
@@ -754,8 +754,8 @@ static void ucma_copy_ib_route(struct rdma_ucm_query_route_resp *resp,
 {
 	struct rdma_dev_addr *dev_addr;
 
-	resp->num_paths = route->num_paths;
-	switch (route->num_paths) {
+	resp->num_paths = route->num_pri_alt_paths;
+	switch (route->num_pri_alt_paths) {
 	case 0:
 		dev_addr = &route->addr.dev_addr;
 		rdma_addr_get_dgid(dev_addr,
@@ -781,8 +781,8 @@ static void ucma_copy_iboe_route(struct rdma_ucm_query_route_resp *resp,
 				 struct rdma_route *route)
 {
 
-	resp->num_paths = route->num_paths;
-	switch (route->num_paths) {
+	resp->num_paths = route->num_pri_alt_paths;
+	switch (route->num_pri_alt_paths) {
 	case 0:
 		rdma_ip2gid((struct sockaddr *)&route->addr.dst_addr,
 			    (union ib_gid *)&resp->ib_route[0].dgid);
@@ -921,7 +921,7 @@ static ssize_t ucma_query_path(struct ucma_context *ctx,
 	if (!resp)
 		return -ENOMEM;
 
-	resp->num_paths = ctx->cm_id->route.num_paths;
+	resp->num_paths = ctx->cm_id->route.num_pri_alt_paths;
 	for (i = 0, out_len -= sizeof(*resp);
 	     i < resp->num_paths && out_len > sizeof(struct ib_path_rec_data);
 	     i++, out_len -= sizeof(struct ib_path_rec_data)) {
include/rdma/rdma_cm.h +6 -1
index 5b18e2e36ee6..81916039ee24 100644
@@ -52,7 +52,12 @@ struct rdma_addr {
 struct rdma_route {
 	struct rdma_addr addr;
 	struct sa_path_rec *path_rec;
-	int num_paths;
+	/*
+	 * 0 - No primary nor alternate path is available
+	 * 1 - Only primary path is available
+	 * 2 - Both primary and alternate path are available
+	 */
+	int num_pri_alt_paths;
 };
 
 struct rdma_conn_param {
Signature · 1 line
2.37.2
#2 lore raw

PATCH 2/4 RDMA/cma: Multiple path records support with netlink channel

In reply to Leon Romanovsky

From: Mark Zhang <[email protected]>

Support receiving inbound and outbound IB path records (along with GMP PathRecord) from user-space service through the RDMA netlink channel. The LIDs in these 3 PRs can be used in this way:
1. GMP PR: used as the standard local/remote LIDs;
2. DLID of outbound PR: Used as the "dlid" field for outbound traffic;
3. DLID of inbound PR: Used as the "dlid" field for outbound traffic in responder side.

This is aimed to support adaptive routing. With current IB routing solution when a packet goes out it's assigned with a fixed DLID per target, meaning a fixed router will be used.
The LIDs in inbound/outbound path records can be used to identify group of routers that allow communication with another subnet's entity. With them packets from an inter-subnet connection may travel through any router in the set to reach the target.

As confirmed with Jason, when sending a netlink request, kernel uses LS_RESOLVE_PATH_USE_ALL so that the service knows kernel supports multiple PRs.

---
 drivers/infiniband/core/cma.c             |  70 ++++++-
 drivers/infiniband/core/sa_query.c        | 235 +++++++++++++++-------
 drivers/infiniband/ulp/ipoib/ipoib_main.c |   2 +-
 drivers/infiniband/ulp/srp/ib_srp.c       |   2 +-
 include/rdma/ib_sa.h                      |   3 +-
 include/rdma/rdma_cm.h                    |   6 +
 6 files changed, 231 insertions(+), 87 deletions(-)
drivers/infiniband/core/cma.c +59 -11
index 91e72a76d95e..a3efc462305d 100644
@@ -2026,6 +2026,8 @@ static void _destroy_id(struct rdma_id_private *id_priv,
 		cma_id_put(id_priv->id.context);
 
 	kfree(id_priv->id.route.path_rec);
+	kfree(id_priv->id.route.path_rec_inbound);
+	kfree(id_priv->id.route.path_rec_outbound);
 
 	put_net(id_priv->id.route.addr.dev_addr.net);
 	kfree(id_priv);
@@ -2817,26 +2819,72 @@ int rdma_set_min_rnr_timer(struct rdma_cm_id *id, u8 min_rnr_timer)
 }
 EXPORT_SYMBOL(rdma_set_min_rnr_timer);
 
+static void route_set_path_rec_inbound(struct cma_work *work,
+				       struct sa_path_rec *path_rec)
+{
+	struct rdma_route *route = &work->id->id.route;
+
+	if (!route->path_rec_inbound) {
+		route->path_rec_inbound =
+			kzalloc(sizeof(*route->path_rec_inbound), GFP_KERNEL);
+		if (!route->path_rec_inbound)
+			return;
+	}
+
+	*route->path_rec_inbound = *path_rec;
+}
+
+static void route_set_path_rec_outbound(struct cma_work *work,
+					struct sa_path_rec *path_rec)
+{
+	struct rdma_route *route = &work->id->id.route;
+
+	if (!route->path_rec_outbound) {
+		route->path_rec_outbound =
+			kzalloc(sizeof(*route->path_rec_outbound), GFP_KERNEL);
+		if (!route->path_rec_outbound)
+			return;
+	}
+
+	*route->path_rec_outbound = *path_rec;
+}
+
 static void cma_query_handler(int status, struct sa_path_rec *path_rec,
-			      void *context)
+			      int num_prs, void *context)
 {
 	struct cma_work *work = context;
 	struct rdma_route *route;
+	int i;
 
 	route = &work->id->id.route;
 
-	if (!status) {
-		route->num_pri_alt_paths = 1;
-		*route->path_rec = *path_rec;
-	} else {
-		work->old_state = RDMA_CM_ROUTE_QUERY;
-		work->new_state = RDMA_CM_ADDR_RESOLVED;
-		work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
-		work->event.status = status;
-		pr_debug_ratelimited("RDMA CM: ROUTE_ERROR: failed to query path. status %d\n",
-				     status);
+	if (status)
+		goto fail;
+
+	for (i = 0; i < num_prs; i++) {
+		if (!path_rec[i].flags || (path_rec[i].flags & IB_PATH_GMP))
+			*route->path_rec = path_rec[i];
+		else if (path_rec[i].flags & IB_PATH_INBOUND)
+			route_set_path_rec_inbound(work, &path_rec[i]);
+		else if (path_rec[i].flags & IB_PATH_OUTBOUND)
+			route_set_path_rec_outbound(work, &path_rec[i]);
+	}
+	if (!route->path_rec) {
+		status = -EINVAL;
+		goto fail;
 	}
 
+	route->num_pri_alt_paths = 1;
+	queue_work(cma_wq, &work->work);
+	return;
+
+fail:
+	work->old_state = RDMA_CM_ROUTE_QUERY;
+	work->new_state = RDMA_CM_ADDR_RESOLVED;
+	work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
+	work->event.status = status;
+	pr_debug_ratelimited("RDMA CM: ROUTE_ERROR: failed to query path. status %d\n",
+			     status);
 	queue_work(cma_wq, &work->work);
 }
 
drivers/infiniband/core/sa_query.c +162 -73

337 lines not on this page · raw message

drivers/infiniband/ulp/ipoib/ipoib_main.c +1 -1
index a4904371e2db..ac25fc80fb33 100644
@@ -742,7 +742,7 @@ void ipoib_flush_paths(struct net_device *dev)
 
 static void path_rec_completion(int status,
 				struct sa_path_rec *pathrec,
-				void *path_ptr)
+				int num_prs, void *path_ptr)
 {
 	struct ipoib_path *path = path_ptr;
 	struct net_device *dev = path->dev;
drivers/infiniband/ulp/srp/ib_srp.c +1 -1
index 1e777b2043d6..e5ec5444b662 100644
@@ -699,7 +699,7 @@ static void srp_free_ch_ib(struct srp_target_port *target,
 
 static void srp_path_rec_completion(int status,
 				    struct sa_path_rec *pathrec,
-				    void *ch_ptr)
+				    int num_paths, void *ch_ptr)
 {
 	struct srp_rdma_ch *ch = ch_ptr;
 	struct srp_target_port *target = ch->target;
include/rdma/ib_sa.h +2 -1
index 3634d4cc7a56..e930bec33b31 100644
@@ -186,6 +186,7 @@ struct sa_path_rec {
 		struct sa_path_rec_opa opa;
 	};
 	enum sa_path_rec_type rec_type;
+	u32 flags;
 };
 
 static inline enum ib_gid_type
@@ -413,7 +414,7 @@ int ib_sa_path_rec_get(struct ib_sa_client *client, struct ib_device *device,
 		       ib_sa_comp_mask comp_mask, unsigned long timeout_ms,
 		       gfp_t gfp_mask,
 		       void (*callback)(int status, struct sa_path_rec *resp,
-					void *context),
+					int num_prs, void *context),
 		       void *context, struct ib_sa_query **query);
 
 struct ib_sa_multicast {
include/rdma/rdma_cm.h +6 -0
index 81916039ee24..cdc7cafab572 100644
@@ -49,9 +49,15 @@ struct rdma_addr {
 	struct rdma_dev_addr dev_addr;
 };
 
+#define RDMA_PRIMARY_PATH_MAX_REC_NUM 3
 struct rdma_route {
 	struct rdma_addr addr;
 	struct sa_path_rec *path_rec;
+
+	/* Optional path records of primary path */
+	struct sa_path_rec *path_rec_inbound;
+	struct sa_path_rec *path_rec_outbound;
+
 	/*
 	 * 0 - No primary nor alternate path is available
 	 * 1 - Only primary path is available
Signature · 1 line
2.37.2
maintainer
#3 lore raw

Re: [PATCH rdma-next 2/4] RDMA/cma: Multiple path records support with netlink channel

Quoted 4 lines · Leon Romanovsky
drivers/infiniband/core/cma.c +4 -0
+static void route_set_path_rec_inbound(struct cma_work *work,
+				       struct sa_path_rec *path_rec)
+{
+	struct rdma_route *route = &work->id->id.route;

drivers/infiniband/core/cma.c int rdma_set_min_rnr_timer(struct rdma_cm_id *id, u8 min_rnr_timer) · line 2835

Quoted 10 lines
drivers/infiniband/core/cma.c +10 -0
+
+	if (!route->path_rec_inbound) {
+		route->path_rec_inbound =
+			kzalloc(sizeof(*route->path_rec_inbound), GFP_KERNEL);
+		if (!route->path_rec_inbound)
+			return;
+	}
+
+	*route->path_rec_inbound = *path_rec;
+}

We are just ignoring these memory allocation failures??

drivers/infiniband/core/cma.c int rdma_set_min_rnr_timer(struct rdma_cm_id *id, u8 min_rnr_timer) · line 2853

Quoted 3 lines

static void cma_query_handler(int status, struct sa_path_rec *path_rec,

drivers/infiniband/core/cma.c +1 -1
-			      void *context)
+			      int num_prs, void *context)

This param should be "unsigned int num_prs"

Quoted 19 lines

{

drivers/infiniband/core/cma.c +3 -10
 	struct cma_work *work = context;
 	struct rdma_route *route;
+	int i;
 
 	route = &work->id->id.route;
 
-	if (!status) {
-		route->num_pri_alt_paths = 1;
-		*route->path_rec = *path_rec;
-	} else {
-		work->old_state = RDMA_CM_ROUTE_QUERY;
-		work->new_state = RDMA_CM_ADDR_RESOLVED;
-		work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
-		work->event.status = status;
-		pr_debug_ratelimited("RDMA CM: ROUTE_ERROR: failed to query path. status %d\n",
-				     status);
+	if (status)
+		goto fail;

drivers/infiniband/core/cma.c int rdma_set_min_rnr_timer(struct rdma_cm_id *id, u8 min_rnr_timer) · line 2872

Quoted 10 lines
drivers/infiniband/core/cma.c +10 -0
+
+	for (i = 0; i < num_prs; i++) {
+		if (!path_rec[i].flags || (path_rec[i].flags & IB_PATH_GMP))
+			*route->path_rec = path_rec[i];
+		else if (path_rec[i].flags & IB_PATH_INBOUND)
+			route_set_path_rec_inbound(work, &path_rec[i]);
+		else if (path_rec[i].flags & IB_PATH_OUTBOUND)
+			route_set_path_rec_outbound(work, &path_rec[i]);
+	}
+	if (!route->path_rec) {

Why is this needed? The for loop above will have already oops'd.

Quoted 10 lines
drivers/infiniband/core/sa_query.c +10 -0
+/**
+ * ib_sa_pr_callback_multiple() - Parse path records then do callback.
+ *
+ * In a multiple-PR case the PRs are saved in "query->resp_pr_data"
+ * (instead of"mad->data") and with "ib_path_rec_data" structure format,
+ * so that rec->flags can be set to indicate the type of PR.
+ * This is valid only in IB fabric.
+ */
+static void ib_sa_pr_callback_multiple(struct ib_sa_path_query *query,
+				       int status, int num_prs,

drivers/infiniband/core/sa_query.c static int opa_pr_query_possible(struct ib_sa_client *client, · line 1494

Quoted 10 lines
drivers/infiniband/core/sa_query.c +10 -0
+				       struct ib_path_rec_data *rec_data)
+{
+	struct sa_path_rec *rec;
+	int i;
+
+	rec = kvcalloc(num_prs, sizeof(*rec), GFP_KERNEL);
+	if (!rec) {
+		query->callback(-ENOMEM, NULL, 0, query->context);
+		return;
+	}

This all seems really wild, why are we allocating memory so many times on this path? Have ib_nl_process_good_resolve_rsp unpack the mad instead of storing the raw format

It would also be good to make resp_pr_data always valid so all these special paths don't need to exist.

include/rdma/rdma_cm.h struct rdma_addr { · line 52

Quoted 9 lines
include/rdma/rdma_cm.h +1 -0
index 81916039ee24..cdc7cafab572 100644
@@ -49,9 +49,15 @@ struct rdma_addr {
 	struct rdma_dev_addr dev_addr;
 };
 
+#define RDMA_PRIMARY_PATH_MAX_REC_NUM 3

This is a strange place for this define, it should be in sa_query.c?

Jason

#4 lore raw

Re: [PATCH rdma-next 2/4] RDMA/cma: Multiple path records support with netlink channel

Quoted 19 lines · Jason Gunthorpe
Quoted 4 lines · Leon Romanovsky
drivers/infiniband/core/cma.c +4 -0
+static void route_set_path_rec_inbound(struct cma_work *work,
+				       struct sa_path_rec *path_rec)
+{
+	struct rdma_route *route = &work->id->id.route;

drivers/infiniband/core/cma.c int rdma_set_min_rnr_timer(struct rdma_cm_id *id, u8 min_rnr_timer) · line 2835

Quoted 10 lines
drivers/infiniband/core/cma.c +10 -0
+
+	if (!route->path_rec_inbound) {
+		route->path_rec_inbound =
+			kzalloc(sizeof(*route->path_rec_inbound), GFP_KERNEL);
+		if (!route->path_rec_inbound)
+			return;
+	}
+
+	*route->path_rec_inbound = *path_rec;
+}

We are just ignoring these memory allocation failures??

Inside "if" statement if kzalloc fails here then we don't set route->path_rec_inbound or outbound;

Quoted 5 lines

drivers/infiniband/core/cma.c int rdma_set_min_rnr_timer(struct rdma_cm_id *id, u8 min_rnr_timer) · line 2853

Quoted 3 lines

static void cma_query_handler(int status, struct sa_path_rec *path_rec,

drivers/infiniband/core/cma.c +1 -1
-			      void *context)
+			      int num_prs, void *context)

This param should be "unsigned int num_prs"

Ack

Quoted 24 lines
Quoted 24 lines
  {
  	struct cma_work *work = context;
  	struct rdma_route *route;
drivers/infiniband/core/cma.c +7 -10
+	int i;
 
  	route = &work->id->id.route;
 
-	if (!status) {
-		route->num_pri_alt_paths = 1;
-		*route->path_rec = *path_rec;
-	} else {
-		work->old_state = RDMA_CM_ROUTE_QUERY;
-		work->new_state = RDMA_CM_ADDR_RESOLVED;
-		work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
-		work->event.status = status;
-		pr_debug_ratelimited("RDMA CM: ROUTE_ERROR: failed to query path. status %d\n",
-				     status);
+	if (status)
+		goto fail;
+
+	for (i = 0; i < num_prs; i++) {
+		if (!path_rec[i].flags || (path_rec[i].flags & IB_PATH_GMP))
+			*route->path_rec = path_rec[i];
Quoted 8 lines

drivers/infiniband/core/cma.c int rdma_set_min_rnr_timer(struct rdma_cm_id *id, u8 min_rnr_timer) · line 2872

Quoted 6 lines
drivers/infiniband/core/cma.c +6 -0
+		else if (path_rec[i].flags & IB_PATH_INBOUND)
+			route_set_path_rec_inbound(work, &path_rec[i]);
+		else if (path_rec[i].flags & IB_PATH_OUTBOUND)
+			route_set_path_rec_outbound(work, &path_rec[i]);
+	}
+	if (!route->path_rec) {

Why is this needed? The for loop above will have already oops'd.

Right, this "if" is no needed. We don't need to check if route->path_rec is valid in this function because it is allocated in cma_resolve_ib_route()

Quoted 26 lines
Quoted 11 lines
drivers/infiniband/core/sa_query.c +10 -0
+/**
+ * ib_sa_pr_callback_multiple() - Parse path records then do callback.
+ *
+ * In a multiple-PR case the PRs are saved in "query->resp_pr_data"
+ * (instead of"mad->data") and with "ib_path_rec_data" structure format,
+ * so that rec->flags can be set to indicate the type of PR.
+ * This is valid only in IB fabric.
+ */
+static void ib_sa_pr_callback_multiple(struct ib_sa_path_query *query,
+				       int status, int num_prs,

drivers/infiniband/core/sa_query.c static int opa_pr_query_possible(struct ib_sa_client *client, · line 1494

Quoted 10 lines
drivers/infiniband/core/sa_query.c +10 -0
+				       struct ib_path_rec_data *rec_data)
+{
+	struct sa_path_rec *rec;
+	int i;
+
+	rec = kvcalloc(num_prs, sizeof(*rec), GFP_KERNEL);
+	if (!rec) {
+		query->callback(-ENOMEM, NULL, 0, query->context);
+		return;
+	}

This all seems really wild, why are we allocating memory so many times on this path? Have ib_nl_process_good_resolve_rsp unpack the mad instead of storing the raw format

Quoted 2 lines

It would also be good to make resp_pr_data always valid so all these special paths don't need to exist.

The ib_sa_pr_callback_single() uses stack variable "rec" but ib_sa_pr_callback_multiple() uses malloc because there are multiple PRs.

ib_sa_path_rec_callback is also used by ib_post_send_mad(), which always have single PR and saved in mad->data, so always set resp_pr_data in netlink case is not enough.

Quoted 11 lines

include/rdma/rdma_cm.h struct rdma_addr { · line 52

Quoted 9 lines

include/rdma/rdma_cm.h

Quoted 1 line
include/rdma/rdma_cm.h +0 -0

index 81916039ee24..cdc7cafab572 100644

include/rdma/rdma_cm.h +1 -0
@@ -49,9 +49,15 @@ struct rdma_addr {
  	struct rdma_dev_addr dev_addr;
  };
 
+#define RDMA_PRIMARY_PATH_MAX_REC_NUM 3

This is a strange place for this define, it should be in sa_query.c?

That's because path_rec, path_rec_inbound and path_rec_outbound are defined here, but yes it is only used in sa_query.c, so maybe better move it to there.

Thanks Jason.

maintainer
#5 lore raw

Re: [PATCH rdma-next 2/4] RDMA/cma: Multiple path records support with netlink channel

Quoted 20 lines · Mark Zhang
Quoted 19 lines · Jason Gunthorpe
Quoted 4 lines · Leon Romanovsky
drivers/infiniband/core/cma.c +4 -0
+static void route_set_path_rec_inbound(struct cma_work *work,
+				       struct sa_path_rec *path_rec)
+{
+	struct rdma_route *route = &work->id->id.route;

drivers/infiniband/core/cma.c int rdma_set_min_rnr_timer(struct rdma_cm_id *id, u8 min_rnr_timer) · line 2835

Quoted 10 lines
drivers/infiniband/core/cma.c +10 -0
+
+	if (!route->path_rec_inbound) {
+		route->path_rec_inbound =
+			kzalloc(sizeof(*route->path_rec_inbound), GFP_KERNEL);
+		if (!route->path_rec_inbound)
+			return;
+	}
+
+	*route->path_rec_inbound = *path_rec;
+}

We are just ignoring these memory allocation failures??

Quoted 2 lines

Inside "if" statement if kzalloc fails here then we don't set route->path_rec_inbound or outbound;

But why don't we propogate a ENOMEM failure code?

Quoted 20 lines
Quoted 17 lines

drivers/infiniband/core/sa_query.c static int opa_pr_query_possible(struct ib_sa_client *client, · line 1494

Quoted 12 lines
drivers/infiniband/core/sa_query.c +12 -0
+static void ib_sa_pr_callback_multiple(struct ib_sa_path_query *query,
+				       int status, int num_prs,
+				       struct ib_path_rec_data *rec_data)
+{
+	struct sa_path_rec *rec;
+	int i;
+
+	rec = kvcalloc(num_prs, sizeof(*rec), GFP_KERNEL);
+	if (!rec) {
+		query->callback(-ENOMEM, NULL, 0, query->context);
+		return;
+	}

This all seems really wild, why are we allocating memory so many times on this path? Have ib_nl_process_good_resolve_rsp unpack the mad instead of storing the raw format

Quoted 2 lines

It would also be good to make resp_pr_data always valid so all these special paths don't need to exist.

Quoted 6 lines

The ib_sa_pr_callback_single() uses stack variable "rec" but ib_sa_pr_callback_multiple() uses malloc because there are multiple PRs.

ib_sa_path_rec_callback is also used by ib_post_send_mad(), which always have single PR and saved in mad->data, so always set resp_pr_data in netlink case is not enough.

We should always be able to point resp_pr_data to some kind of storage, even if it is stack storage.

Jason

#6 lore raw

Re: [PATCH rdma-next 2/4] RDMA/cma: Multiple path records support with netlink channel

Quoted 21 lines · Jason Gunthorpe
Quoted 20 lines · Mark Zhang
Quoted 19 lines · Jason Gunthorpe
Quoted 4 lines · Leon Romanovsky
drivers/infiniband/core/cma.c +4 -0
+static void route_set_path_rec_inbound(struct cma_work *work,
+				       struct sa_path_rec *path_rec)
+{
+	struct rdma_route *route = &work->id->id.route;

drivers/infiniband/core/cma.c int rdma_set_min_rnr_timer(struct rdma_cm_id *id, u8 min_rnr_timer) · line 2835

Quoted 10 lines
drivers/infiniband/core/cma.c +10 -0
+
+	if (!route->path_rec_inbound) {
+		route->path_rec_inbound =
+			kzalloc(sizeof(*route->path_rec_inbound), GFP_KERNEL);
+		if (!route->path_rec_inbound)
+			return;
+	}
+
+	*route->path_rec_inbound = *path_rec;
+}

We are just ignoring these memory allocation failures??

Quoted 4 lines
Quoted 2 lines

Inside "if" statement if kzalloc fails here then we don't set route->path_rec_inbound or outbound;

But why don't we propogate a ENOMEM failure code?

Because inbound/outbound PRs are optional, so even they are provided they can still be ignored if cma is not able to set them (e.g. memory allocation failure in this case).

Quoted 23 lines
Quoted 23 lines
Quoted 17 lines

drivers/infiniband/core/sa_query.c static int opa_pr_query_possible(struct ib_sa_client *client, · line 1494

Quoted 12 lines
drivers/infiniband/core/sa_query.c +12 -0
+static void ib_sa_pr_callback_multiple(struct ib_sa_path_query *query,
+				       int status, int num_prs,
+				       struct ib_path_rec_data *rec_data)
+{
+	struct sa_path_rec *rec;
+	int i;
+
+	rec = kvcalloc(num_prs, sizeof(*rec), GFP_KERNEL);
+	if (!rec) {
+		query->callback(-ENOMEM, NULL, 0, query->context);
+		return;
+	}

This all seems really wild, why are we allocating memory so many times on this path? Have ib_nl_process_good_resolve_rsp unpack the mad instead of storing the raw format

Quoted 2 lines

It would also be good to make resp_pr_data always valid so all these special paths don't need to exist.

The ib_sa_pr_callback_single() uses stack variable "rec" but ib_sa_pr_callback_multiple() uses malloc because there are multiple PRs.

Quoted 6 lines
Quoted 3 lines

ib_sa_path_rec_callback is also used by ib_post_send_mad(), which always have single PR and saved in mad->data, so always set resp_pr_data in netlink case is not enough.

We should always be able to point resp_pr_data to some kind of storage, even if it is stack storage.

The idea is:
- Single PR: PR in mad->data; Used by both netlink and ib_post_send_mad();
- Multiple PRs: PRs in resp_pr_data, with "ib_path_rec_data" structure format; Currently used by netlink only.

So if we want to always use resp_pr_data then in single-PR case we need to copy mad->data to resp_pr_data in both netlink and ib_post_send_mad(), and turn it into "ib_path_rec_data" structure format. This adds complexity for single-PR, which should be most of situations?

Use malloc instead of stack for resp_pr_data and multiple-PR unpack is because sizeof(sa_path_rec)=72B, now we supports 3 and there might be more in future..

#7 lore raw

Re: [PATCH rdma-next 2/4] RDMA/cma: Multiple path records support with netlink channel

Quoted 66 lines · Mark Zhang

External email: Use caution opening links or attachments

Quoted 22 lines · Jason Gunthorpe
Quoted 21 lines · Mark Zhang
Quoted 20 lines · Jason Gunthorpe
Quoted 12 lines · Leon Romanovsky
drivers/infiniband/core/cma.c +8 -0
+static void route_set_path_rec_inbound(struct cma_work *work,
+                                 struct sa_path_rec *path_rec)
+{
+  struct rdma_route *route = &work->id->id.route;
+
+  if (!route->path_rec_inbound) {
+          route->path_rec_inbound =
+                  kzalloc(sizeof(*route->path_rec_inbound), 

GFP_KERNEL);
+          if (!route->path_rec_inbound)
+                  return;
+  }

drivers/infiniband/core/cma.c int rdma_set_min_rnr_timer(struct rdma_cm_id *id, u8 min_rnr_timer) · line 2835

Quoted 3 lines
drivers/infiniband/core/cma.c +3 -0
+
+  *route->path_rec_inbound = *path_rec;
+}

We are just ignoring these memory allocation failures??

Quoted 4 lines
Quoted 2 lines

Inside "if" statement if kzalloc fails here then we don't set route->path_rec_inbound or outbound;

But why don't we propogate a ENOMEM failure code?

Because inbound/outbound PRs are optional, so even they are provided they can still be ignored if cma is not able to set them (e.g. memory allocation failure in this case).

Quoted 24 lines
Quoted 24 lines
Quoted 18 lines

drivers/infiniband/core/sa_query.c static int opa_pr_query_possible(struct ib_sa_client *client, · line 1494

Quoted 13 lines

+static void ib_sa_pr_callback_multiple(struct ib_sa_path_query *query,
+                                 int status, int num_prs, +                                 struct ib_path_rec_data *rec_data)

drivers/infiniband/core/sa_query.c +9 -0
+{
+  struct sa_path_rec *rec;
+  int i;
+
+  rec = kvcalloc(num_prs, sizeof(*rec), GFP_KERNEL);
+  if (!rec) {
+          query->callback(-ENOMEM, NULL, 0, query->context);
+          return;
+  }

This all seems really wild, why are we allocating memory so many times on this path? Have ib_nl_process_good_resolve_rsp unpack the mad instead of storing the raw format

Quoted 2 lines

It would also be good to make resp_pr_data always valid so all these special paths don't need to exist.

The ib_sa_pr_callback_single() uses stack variable "rec" but ib_sa_pr_callback_multiple() uses malloc because there are multiple PRs.

Quoted 6 lines
Quoted 3 lines

ib_sa_path_rec_callback is also used by ib_post_send_mad(), which always have single PR and saved in mad->data, so always set resp_pr_data in netlink case is not enough.

We should always be able to point resp_pr_data to some kind of storage, even if it is stack storage.

Quoted 10 lines

The idea is:
- Single PR: PR in mad->data; Used by both netlink and ib_post_send_mad();
- Multiple PRs: PRs in resp_pr_data, with "ib_path_rec_data" structure format; Currently used by netlink only.
So if we want to always use resp_pr_data then in single-PR case we need to copy mad->data to resp_pr_data in both netlink and ib_post_send_mad(), and turn it into "ib_path_rec_data" structure format. This adds complexity for single-PR, which should be most of situations?

Sorry what I mean is resp_pr_data is used by netlink while mad->data is used by post_send_mad(), so if we want to always use resp_pr_data then in post_send_mad() we need to do malloc, copy and transfer. Besides if malloc failures then we still need to use mad->data, unless we always use stack no matter how many PRs there are.

Quoted 4 lines

Use malloc instead of stack for resp_pr_data and multiple-PR unpack is because sizeof(sa_path_rec)=72B, now we supports 3 and there might be more in future..

maintainer
#8 lore raw

Re: [PATCH rdma-next 2/4] RDMA/cma: Multiple path records support with netlink channel

In reply to Mark Zhang
Quoted 22 lines · Mark Zhang
Quoted 21 lines · Jason Gunthorpe
Quoted 20 lines · Mark Zhang
Quoted 19 lines · Jason Gunthorpe
Quoted 4 lines · Leon Romanovsky
drivers/infiniband/core/cma.c +4 -0
+static void route_set_path_rec_inbound(struct cma_work *work,
+				       struct sa_path_rec *path_rec)
+{
+	struct rdma_route *route = &work->id->id.route;

drivers/infiniband/core/cma.c int rdma_set_min_rnr_timer(struct rdma_cm_id *id, u8 min_rnr_timer) · line 2835

Quoted 10 lines
drivers/infiniband/core/cma.c +10 -0
+
+	if (!route->path_rec_inbound) {
+		route->path_rec_inbound =
+			kzalloc(sizeof(*route->path_rec_inbound), GFP_KERNEL);
+		if (!route->path_rec_inbound)
+			return;
+	}
+
+	*route->path_rec_inbound = *path_rec;
+}

We are just ignoring these memory allocation failures??

Quoted 8 lines
Quoted 4 lines
Quoted 2 lines

Inside "if" statement if kzalloc fails here then we don't set route->path_rec_inbound or outbound;

But why don't we propogate a ENOMEM failure code?

Because inbound/outbound PRs are optional, so even they are provided they can still be ignored if cma is not able to set them (e.g. memory allocation failure in this case).

This isn't how we do things, the netlink operation had a failure, the failure should propogate. If we've run out of memory the CM connection is going to blow up anyhow very quickly.

Quoted 9 lines
Quoted 2 lines

We should always be able to point resp_pr_data to some kind of storage, even if it is stack storage.

The idea is:
- Single PR: PR in mad->data; Used by both netlink and ib_post_send_mad();
- Multiple PRs: PRs in resp_pr_data, with "ib_path_rec_data" structure format; Currently used by netlink only.

Quoted 4 lines

So if we want to always use resp_pr_data then in single-PR case we need to copy mad->data to resp_pr_data in both netlink and ib_post_send_mad(), and turn it into "ib_path_rec_data" structure format. This adds complexity for single-PR, which should be most of situations?

You don't copy it, you unpack it. We already have to unpack it, just to a (large!) stack var - unpack it to resp_pr_data instead.

Jason

#9 lore raw

PATCH 3/4 RDMA/cm: Use SLID in the work completion as the DLID in responder side

In reply to Leon Romanovsky

From: Mark Zhang <[email protected]>

The responder should always use WC's SLID as the dlid, to follow the IB SPEC section "13.5.4.2 COMMON RESPONSE ACTIONS":
A responder always takes the following actions in constructing a response packet:
- The SLID of the received packet is used as the DLID in the response packet.

Fixes ac3a949fb2ff ("IB/CM: Set appropriate slid and dlid when handling CM request")
---
 drivers/infiniband/core/cm.c | 14 +++++++-------
 1 file changed, 7 insertions(+), 7 deletions(-)
drivers/infiniband/core/cm.c +7 -7
index d7410ee2ade7..ade82752f9f7 100644
@@ -1614,14 +1614,13 @@ static void cm_path_set_rec_type(struct ib_device *ib_device, u32 port_num,
 
 static void cm_format_path_lid_from_req(struct cm_req_msg *req_msg,
 					struct sa_path_rec *primary_path,
-					struct sa_path_rec *alt_path)
+					struct sa_path_rec *alt_path,
+					struct ib_wc *wc)
 {
 	u32 lid;
 
 	if (primary_path->rec_type != SA_PATH_REC_TYPE_OPA) {
-		sa_path_set_dlid(primary_path,
-				 IBA_GET(CM_REQ_PRIMARY_LOCAL_PORT_LID,
-					 req_msg));
+		sa_path_set_dlid(primary_path, wc->slid);
 		sa_path_set_slid(primary_path,
 				 IBA_GET(CM_REQ_PRIMARY_REMOTE_PORT_LID,
 					 req_msg));
@@ -1658,7 +1657,8 @@ static void cm_format_path_lid_from_req(struct cm_req_msg *req_msg,
 
 static void cm_format_paths_from_req(struct cm_req_msg *req_msg,
 				     struct sa_path_rec *primary_path,
-				     struct sa_path_rec *alt_path)
+				     struct sa_path_rec *alt_path,
+				     struct ib_wc *wc)
 {
 	primary_path->dgid =
 		*IBA_GET_MEM_PTR(CM_REQ_PRIMARY_LOCAL_PORT_GID, req_msg);
@@ -1716,7 +1716,7 @@ static void cm_format_paths_from_req(struct cm_req_msg *req_msg,
 		if (sa_path_is_roce(alt_path))
 			alt_path->roce.route_resolved = false;
 	}
-	cm_format_path_lid_from_req(req_msg, primary_path, alt_path);
+	cm_format_path_lid_from_req(req_msg, primary_path, alt_path, wc);
 }
 
 static u16 cm_get_bth_pkey(struct cm_work *work)
@@ -2129,7 +2129,7 @@ static int cm_req_handler(struct cm_work *work)
 	if (cm_req_has_alt_path(req_msg))
 		work->path[1].rec_type = work->path[0].rec_type;
 	cm_format_paths_from_req(req_msg, &work->path[0],
-				 &work->path[1]);
+				 &work->path[1], work->mad_recv_wc->wc);
 	if (cm_id_priv->av.ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE)
 		sa_path_set_dmac(&work->path[0],
 				 cm_id_priv->av.ah_attr.roce.dmac);
Signature · 1 line
2.37.2
#10 lore raw

PATCH 4/4 RDMA/cm: Use DLID from inbound/outbound PathRecords as the datapath DLID

In reply to Leon Romanovsky

From: Mark Zhang <[email protected]>

In inter-subnet cases, when inbound/outbound PRs are available, outbound_PR.dlid is used as the requestor's datapath DLID and inbound_PR.dlid is used as the responder's DLID. The inbound_PR.dlid is passed to responder side with the "ConnectReq.Primary_Local_Port_LID" field. With this solution the PERMISSIVE_LID is no longer used in Primary Local LID field.

---
 drivers/infiniband/core/cm.c  | 25 +++++++++++++++++++++++--
 drivers/infiniband/core/cma.c |  2 ++
 include/rdma/ib_cm.h          |  2 ++
 3 files changed, 27 insertions(+), 2 deletions(-)
drivers/infiniband/core/cm.c +23 -2
index ade82752f9f7..1f9938a2c475 100644
@@ -175,6 +175,7 @@ struct cm_device {
 struct cm_av {
 	struct cm_port *port;
 	struct rdma_ah_attr ah_attr;
+	u16 dlid_datapath;
 	u16 pkey_index;
 	u8 timeout;
 };
@@ -1304,6 +1305,7 @@ static void cm_format_req(struct cm_req_msg *req_msg,
 	struct sa_path_rec *pri_path = param->primary_path;
 	struct sa_path_rec *alt_path = param->alternate_path;
 	bool pri_ext = false;
+	__be16 lid;
 
 	if (pri_path->rec_type == SA_PATH_REC_TYPE_OPA)
 		pri_ext = opa_is_extended_lid(pri_path->opa.dlid,
@@ -1363,9 +1365,16 @@ static void cm_format_req(struct cm_req_msg *req_msg,
 					      htons(ntohl(sa_path_get_dlid(
 						      pri_path)))));
 	} else {
+
+		if (param->primary_path_inbound) {
+			lid = param->primary_path_inbound->ib.dlid;
+			IBA_SET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg,
+				be16_to_cpu(lid));
+		} else
+			IBA_SET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg,
+				be16_to_cpu(IB_LID_PERMISSIVE));
+
 		/* Work-around until there's a way to obtain remote LID info */
-		IBA_SET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg,
-			be16_to_cpu(IB_LID_PERMISSIVE));
 		IBA_SET(CM_REQ_PRIMARY_REMOTE_PORT_LID, req_msg,
 			be16_to_cpu(IB_LID_PERMISSIVE));
 	}
@@ -1520,6 +1529,10 @@ int ib_send_cm_req(struct ib_cm_id *cm_id,
 	spin_lock_irqsave(&cm_id_priv->lock, flags);
 
 	cm_move_av_from_path(&cm_id_priv->av, &av);
+	if (param->primary_path_outbound)
+		cm_id_priv->av.dlid_datapath =
+			be16_to_cpu(param->primary_path_outbound->ib.dlid);
+
 	if (param->alternate_path)
 		cm_move_av_from_path(&cm_id_priv->alt_av, &alt_av);
 
@@ -2154,6 +2167,10 @@ static int cm_req_handler(struct cm_work *work)
 				       NULL, 0);
 		goto rejected;
 	}
+	if (cm_id_priv->av.ah_attr.type == RDMA_AH_ATTR_TYPE_IB)
+		cm_id_priv->av.dlid_datapath =
+			IBA_GET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg);
+
 	if (cm_req_has_alt_path(req_msg)) {
 		ret = cm_init_av_by_path(&work->path[1], NULL,
 					 &cm_id_priv->alt_av);
@@ -4113,6 +4130,10 @@ static int cm_init_qp_rtr_attr(struct cm_id_private *cm_id_priv,
 		*qp_attr_mask = IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU |
 				IB_QP_DEST_QPN | IB_QP_RQ_PSN;
 		qp_attr->ah_attr = cm_id_priv->av.ah_attr;
+		if ((qp_attr->ah_attr.type == RDMA_AH_ATTR_TYPE_IB) &&
+		    cm_id_priv->av.dlid_datapath &&
+		    (cm_id_priv->av.dlid_datapath != 0xffff))
+			qp_attr->ah_attr.ib.dlid = cm_id_priv->av.dlid_datapath;
 		qp_attr->path_mtu = cm_id_priv->path_mtu;
 		qp_attr->dest_qp_num = be32_to_cpu(cm_id_priv->remote_qpn);
 		qp_attr->rq_psn = be32_to_cpu(cm_id_priv->rq_psn);
drivers/infiniband/core/cma.c +2 -0
index a3efc462305d..7eacb23165fc 100644
@@ -4313,6 +4313,8 @@ static int cma_connect_ib(struct rdma_id_private *id_priv,
 	}
 
 	req.primary_path = &route->path_rec[0];
+	req.primary_path_inbound = route->path_rec_inbound;
+	req.primary_path_outbound = route->path_rec_outbound;
 	if (route->num_pri_alt_paths == 2)
 		req.alternate_path = &route->path_rec[1];
 
include/rdma/ib_cm.h +2 -0
index 8dae5847020a..a2ac62b4a6cf 100644
@@ -348,6 +348,8 @@ struct ib_cm_id *ib_cm_insert_listen(struct ib_device *device,
 
 struct ib_cm_req_param {
 	struct sa_path_rec	*primary_path;
+	struct sa_path_rec	*primary_path_inbound;
+	struct sa_path_rec	*primary_path_outbound;
 	struct sa_path_rec	*alternate_path;
 	const struct ib_gid_attr *ppath_sgid_attr;
 	__be64			service_id;
Signature · 1 line
2.37.2
author Applied
#11 lore raw
In reply to Leon Romanovsky
Quoted 8 lines · Leon Romanovsky

From: Leon Romanovsky <[email protected]>

From Mark:

These patches allow IB core to receive multiple path records from user-space rdma netlink service.

[...]

Applied, thanks!

[1/4] RDMA/core: Rename rdma_route.num_paths field to num_pri_alt_paths

      https://git.kernel.org/rdma/rdma/c/bf9a9928510a03
[2/4] RDMA/cma: Multiple path records support with netlink channel
      https://git.kernel.org/rdma/rdma/c/5a374949339427
[3/4] RDMA/cm: Use SLID in the work completion as the DLID in responder side
      https://git.kernel.org/rdma/rdma/c/b7d95040c13f61
[4/4] RDMA/cm: Use DLID from inbound/outbound PathRecords as the datapath DLID
      https://git.kernel.org/rdma/rdma/c/eb8336dbe373ed

Best regards,

Signature · 1 line
Leon Romanovsky <[email protected]>