[PATCH libibverbs 0/2] RoCE V2 support for UD traffic

[PATCH libibverbs 0/2] RoCE V2 support for UD traffic

[Yishai Hadas]

Hi Doug,

This patch-set from Noa enables user application to work
properly with RoCE V2 when UD traffic is used.

The series was tested successfully with mlx5 driver (lib, kernel)
and can be accessed also from my openfabrics GIT at:
git://openfabrics.org/~yishaih/libibverbs.git branch: rocev2_v0

No change is required in the application side, all is done transparently
to the application.

Will send shortly some libmlx5 changes that are required to
work properly in its side.

Yishai

In General:
Currently, UD traffic is not supported over RoCE V2 in libibverbs,
since libibverbs can't differ between V1 and V2 GIDs and can't
select GID index properly.

This series contains two patches targeted to solve this:
- The first one introduces an internal helper function,
  ibv_query_gid_type, to be used by libibverbs and its vendors libs
  in order to select the correct GID index.
 
- The second patch changes init_ah_from_wc to use the first patch and
  set the GID index according to the RoCE version used.

Noa Osherovich (2):
  Add ibv_query_gid_type to support RoCE v2 UD traffic
  Add support for UD traffic on RoCE v2

 include/infiniband/driver.h |   7 ++
 src/libibverbs.map          |   2 +
 src/verbs.c                 | 243 ++++++++++++++++++++++++++++++++++++++++----
 3 files changed, 233 insertions(+), 19 deletions(-)

-- 
1.8.3.1

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH libibverbs 1/2] Add ibv_query_gid_type to support RoCE v2 UD traffic

[Yishai Hadas]

From: Noa Osherovich <noaos-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>

Currently, libibverbs does not support UD traffic for RoCE v2 since
it can't differ between v1 and v2 GIDs (both have the same GID, only
the version is different). This means that GID index can't be
selected correctly.

This patch introduces ibv_query_gid_type helper function to be used
by libibverbs and its vendors to return GID type based on its GID
index by using the relevant sysfs.

Signed-off-by: Noa Osherovich <noaos-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>
Reviewed-by: Yishai Hadas <yishaih-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>
---
 include/infiniband/driver.h |  7 +++++
 src/libibverbs.map          |  2 ++
 src/verbs.c                 | 65 +++++++++++++++++++++++++++++++++++++++++++++
 3 files changed, 74 insertions(+)

diff --git a/include/infiniband/driver.h b/include/infiniband/driver.h
index 65fa44f..d4b6f58 100644
--- a/include/infiniband/driver.h
+++ b/include/infiniband/driver.h
@@ -86,6 +86,11 @@ enum verbs_qp_mask {
 	VERBS_QP_RESERVED	= 1 << 1
 };
 
+enum ibv_gid_type {
+	IBV_GID_TYPE_IB_ROCE_V1,
+	IBV_GID_TYPE_ROCE_V2,
+};
+
 struct verbs_qp {
 	struct ibv_qp		qp;
 	uint32_t		comp_mask;
@@ -258,4 +263,6 @@ static inline int verbs_get_srq_num(struct ibv_srq *srq, uint32_t *srq_num)
 	return ENOSYS;
 }
 
+int ibv_query_gid_type(struct ibv_context *context, uint8_t port_num,
+		       unsigned int index, enum ibv_gid_type *type);
 #endif /* INFINIBAND_DRIVER_H */
diff --git a/src/libibverbs.map b/src/libibverbs.map
index 5134bd9..a7d3339 100644
--- a/src/libibverbs.map
+++ b/src/libibverbs.map
@@ -119,4 +119,6 @@ IBVERBS_1.1 {
 		ibv_cmd_open_qp;
 		ibv_cmd_rereg_mr;
 
+		ibv_query_gid_type;
+
 } IBVERBS_1.0;
diff --git a/src/verbs.c b/src/verbs.c
index 68888c3..08f03e3 100644
--- a/src/verbs.c
+++ b/src/verbs.c
@@ -41,6 +41,7 @@
 #include <stdlib.h>
 #include <errno.h>
 #include <string.h>
+#include <dirent.h>
 
 #include "ibverbs.h"
 #ifndef NRESOLVE_NEIGH
@@ -585,6 +586,70 @@ struct ibv_ah *__ibv_create_ah(struct ibv_pd *pd, struct ibv_ah_attr *attr)
 }
 default_symver(__ibv_create_ah, ibv_create_ah);
 
+/* GID types as appear in sysfs, no change is expected as of ABI
+ * compatibility.
+ */
+#define V1_TYPE "IB/RoCE v1"
+#define V2_TYPE "RoCE v2"
+int ibv_query_gid_type(struct ibv_context *context, uint8_t port_num,
+		       unsigned int index, enum ibv_gid_type *type)
+{
+	char name[32];
+	char buff[11];
+
+	snprintf(name, sizeof(name), "ports/%d/gid_attrs/types/%d", port_num,
+		 index);
+
+	/* Reset errno so that we can rely on its value upon any error flow in
+	 * ibv_read_sysfs_file.
+	 */
+	errno = 0;
+	if (ibv_read_sysfs_file(context->device->ibdev_path, name, buff,
+				sizeof(buff)) <= 0) {
+		char *dir_path;
+		DIR *dir;
+
+		if (errno == EINVAL) {
+			/* In IB, this file doesn't exist and the kernel sets
+			 * errno to -EINVAL.
+			 */
+			*type = IBV_GID_TYPE_IB_ROCE_V1;
+			return 0;
+		}
+		if (asprintf(&dir_path, "%s/%s/%d/%s/",
+			     context->device->ibdev_path, "ports", port_num,
+			     "gid_attrs") < 0)
+			return -1;
+		dir = opendir(dir_path);
+		free(dir_path);
+		if (!dir) {
+			if (errno == ENOENT)
+				/* Assuming that if gid_attrs doesn't exist,
+				 * we have an old kernel and all GIDs are
+				 * IB/RoCE v1
+				 */
+				*type = IBV_GID_TYPE_IB_ROCE_V1;
+			else
+				return -1;
+		} else {
+			closedir(dir);
+			errno = EFAULT;
+			return -1;
+		}
+	} else {
+		if (!strcmp(buff, V1_TYPE)) {
+			*type = IBV_GID_TYPE_IB_ROCE_V1;
+		} else if (!strcmp(buff, V2_TYPE)) {
+			*type = IBV_GID_TYPE_ROCE_V2;
+		} else {
+			errno = ENOTSUP;
+			return -1;
+		}
+	}
+
+	return 0;
+}
+
 static int ibv_find_gid_index(struct ibv_context *context, uint8_t port_num,
 			      union ibv_gid *gid)
 {
-- 
1.8.3.1

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [libibverbs, 1/2] Add ibv_query_gid_type to support RoCE v2 UD traffic

[Jason Gunthorpe]

On Tue, Aug 30, 2016 at 07:51:44PM +0300, Yishai Hadas wrote:

> diff --git a/src/libibverbs.map b/src/libibverbs.map
> index 5134bd9..a7d3339 100644
> +++ b/src/libibverbs.map
> @@ -119,4 +119,6 @@ IBVERBS_1.1 {
>  		ibv_cmd_open_qp;
>  		ibv_cmd_rereg_mr;
>  
> +		ibv_query_gid_type;
> +
>  } IBVERBS_1.0;

NAK on incorrect use of the map file.

Also for your other patches.

Use IBVERBS_1.3 as the symbol label.

I wrote this for rdma-plumbing to describe the appropriate approach.

[..]

# Shared Library Versions

The shared libraries use the typical semantic versioning scheme, eg
*libibumad* has a version like `3.1.11`.

The version number is broken up into three fields:
- '3' is called the SONAME and is embedded into the ELF:
   ```sh
   $ readelf -ds build/lib/libibumad.so.3.1.11
    0x000000000000000e (SONAME)             Library soname: [libibumad.so.3]
   ```

   We do not expect this value to ever change for our libraries. It indicates
   the overall ABI, changing it means the library will not dynamically to old
   programs link anymore.

- '1' is called the ABI level and is used within the ELF as the last component
   symbol version tag.  This version must be changed every time a new symbol
   is introduced. It allows the user to see what version of the ABI the
   library provides.

- '11' is the overall release number and is copied from `PACKAGE_VERSION` This
  version increases with every package release, even if the library code did
  not change. It allows the user to see what upstream source was used to build
  the library.

This version is encoded into the filename `build/lib/libibumad.so.3.1.11` and
a symlink from `libibumad.so.3` to `build/lib/libibumad.so.3.1.11` is created.

## Shared Library Symbol Versions

Symbol versions are a linker technique that lets the library author provide
two symbols with different ABIs that have the same API name. The linker
differentiates the two cases internally. This allows the library author to
change the ABI that the API uses. This project typically does not make use of
this feature.

As a secondary feature, the symbol version is also used by package managers
like RPM to manage the ABI level. To make this work properly the ABI level
must be correctly encoded into the symbol version.

## Adding a new symbol

First, increase the ABI level of the library. It is safe to re-use the ABI
level for multiple new functions within a single release, but once a release
is tagged the ABI level becomes *immutable*. The maintainer can provide
guidence on what ABI level to use for each series.

```diff
 rdma_library(ibumad libibumad.map
   # See Documentation/versioning.md
-  3 3.1.${PACKAGE_VERSION}
+  3 3.2.${PACKAGE_VERSION}
```

Next, add your new symbol to the symbol version file:

```diff
+ IBUMAD_3.2 {
+ 	global:
+ 		umad_new_symbol;
+ } IBUMAD_1.0;
```

NOTE: Once a release is made the stanzas in the map file are *immutable* and
cannot be changed. Do not add your new symbol to old stanzas.

The new symbol should appear in the ELF:

```sh
$ readelf -s build/lib/libibumad.so.3.1.11
 35: 00000000000031e0   450 FUNC    GLOBAL DEFAULT   12 umad_new_symbol@@IBUMAD_3.2
```

Finally update the `debian/libibumad3.symbols` file.
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH libibverbs 2/2] Add support for UD traffic on RoCE v2

[Yishai Hadas]

From: Noa Osherovich <noaos-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>

When creating an Address Handle from a Work Completion, source GID index must
be set. To enable RoCE v2, libibverbs should select the right GID index for it.
RoCE v2 is using UDP/IP layers so a WC's GRH could be either an IB GRH, IPv4
header or an IPv6 header.

Libibverbs should be able to differ between them in a driver-agnostic way to
avoid API changes.

This patch is using the fact that for RoCE v2, the GRH is either an IPv6 header
or 20 garbled bytes followed by an IPv4 header, as defined in RoCE v2 annex.

The annex also specifies that for packets with IPv4 header, the version number
is 4, for packets with IPv6 header it's 6 or the packet is silently dropped.
This fact is also taken into account when parsing the GRH.

Signed-off-by: Noa Osherovich <noaos-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>
Reviewed-by: Yishai Hadas <yishaih-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>
---
 src/verbs.c | 178 +++++++++++++++++++++++++++++++++++++++++++++++++++++-------
 1 file changed, 159 insertions(+), 19 deletions(-)

diff --git a/src/verbs.c b/src/verbs.c
index 08f03e3..fc8b9f6 100644
--- a/src/verbs.c
+++ b/src/verbs.c
@@ -41,6 +41,7 @@
 #include <stdlib.h>
 #include <errno.h>
 #include <string.h>
+#include <linux/ip.h>
 #include <dirent.h>
 
 #include "ibverbs.h"
@@ -651,48 +652,187 @@ int ibv_query_gid_type(struct ibv_context *context, uint8_t port_num,
 }
 
 static int ibv_find_gid_index(struct ibv_context *context, uint8_t port_num,
-			      union ibv_gid *gid)
+			      union ibv_gid *gid, enum ibv_gid_type gid_type)
 {
+	enum ibv_gid_type sgid_type = 0;
 	union ibv_gid sgid;
 	int i = 0, ret;
 
 	do {
-		ret = ibv_query_gid(context, port_num, i++, &sgid);
-	} while (!ret && memcmp(&sgid, gid, sizeof *gid));
+		ret = ibv_query_gid(context, port_num, i, &sgid);
+		if (!ret) {
+			ret = ibv_query_gid_type(context, port_num, i,
+						 &sgid_type);
+		}
+		i++;
+	} while (!ret && (memcmp(&sgid, gid, sizeof(*gid)) ||
+		 (gid_type != sgid_type)));
 
 	return ret ? ret : i - 1;
 }
 
-int ibv_init_ah_from_wc(struct ibv_context *context, uint8_t port_num,
-			struct ibv_wc *wc, struct ibv_grh *grh,
-			struct ibv_ah_attr *ah_attr)
+static inline void map_ipv4_addr_to_ipv6(__be32 ipv4, struct in6_addr *ipv6)
+{
+	ipv6->s6_addr32[0] = 0;
+	ipv6->s6_addr32[1] = 0;
+	ipv6->s6_addr32[2] = htonl(0x0000FFFF);
+	ipv6->s6_addr32[3] = ipv4;
+}
+
+static inline uint16_t ipv4_calc_hdr_csum(uint16_t *data, unsigned int num_hwords)
+{
+	unsigned int i = 0;
+	uint32_t sum = 0;
+
+	for (i = 0; i < num_hwords; i++)
+		sum += *(data++);
+
+	sum = (sum & 0xffff) + (sum >> 16);
+
+	return ~sum;
+}
+
+static inline int get_grh_header_version(struct ibv_grh *grh)
+{
+	int ip6h_version = (ntohl(grh->version_tclass_flow) >> 28) & 0xf;
+	struct iphdr *ip4h = (struct iphdr *)((void *)grh + 20);
+	struct iphdr ip4h_checked;
+
+	if (ip6h_version != 6) {
+		if (ip4h->version == 4)
+			return 4;
+		errno = EPROTONOSUPPORT;
+		return -1;
+	}
+	/* version may be 6 or 4 */
+	if (ip4h->ihl != 5) /* IPv4 header length must be 5 for RoCE v2. */
+		return 6;
+	/*
+	* Verify checksum.
+	* We can't write on scattered buffers so we have to copy to temp
+	* buffer.
+	*/
+	memcpy(&ip4h_checked, ip4h, sizeof(ip4h_checked));
+	/* Need to set the checksum field (check) to 0 before re-calculating
+	 * the checksum.
+	 */
+	ip4h_checked.check = 0;
+	ip4h_checked.check = ipv4_calc_hdr_csum((uint16_t *)&ip4h_checked, 10);
+	/* if IPv4 header checksum is OK, believe it */
+	if (ip4h->check == ip4h_checked.check)
+		return 4;
+	return 6;
+}
+
+static inline void set_ah_attr_generic_fields(struct ibv_ah_attr *ah_attr,
+					      struct ibv_wc *wc,
+					      struct ibv_grh *grh,
+					      uint8_t port_num)
 {
 	uint32_t flow_class;
-	int ret;
 
-	memset(ah_attr, 0, sizeof *ah_attr);
+	flow_class = ntohl(grh->version_tclass_flow);
+	ah_attr->grh.flow_label = flow_class & 0xFFFFF;
 	ah_attr->dlid = wc->slid;
 	ah_attr->sl = wc->sl;
 	ah_attr->src_path_bits = wc->dlid_path_bits;
 	ah_attr->port_num = port_num;
+}
 
-	if (wc->wc_flags & IBV_WC_GRH) {
-		ah_attr->is_global = 1;
-		ah_attr->grh.dgid = grh->sgid;
+static inline int set_ah_attr_by_ipv4(struct ibv_context *context,
+				      struct ibv_ah_attr *ah_attr,
+				      struct iphdr *ip4h, uint8_t port_num)
+{
+	union ibv_gid sgid;
+	int ret;
+
+	/* No point searching multicast GIDs in GID table */
+	if (IN_CLASSD(ntohl(ip4h->daddr))) {
+		errno = EINVAL;
+		return -1;
+	}
 
-		ret = ibv_find_gid_index(context, port_num, &grh->dgid);
-		if (ret < 0)
-			return ret;
+	map_ipv4_addr_to_ipv6(ip4h->daddr, (struct in6_addr *)&sgid);
+	ret = ibv_find_gid_index(context, port_num, &sgid,
+				 IBV_GID_TYPE_ROCE_V2);
+	if (ret < 0)
+		return ret;
 
-		ah_attr->grh.sgid_index = (uint8_t) ret;
-		flow_class = ntohl(grh->version_tclass_flow);
-		ah_attr->grh.flow_label = flow_class & 0xFFFFF;
-		ah_attr->grh.hop_limit = grh->hop_limit;
-		ah_attr->grh.traffic_class = (flow_class >> 20) & 0xFF;
+	map_ipv4_addr_to_ipv6(ip4h->saddr,
+			      (struct in6_addr *)&ah_attr->grh.dgid);
+	ah_attr->grh.sgid_index = (uint8_t) ret;
+	ah_attr->grh.hop_limit = ip4h->ttl;
+	ah_attr->grh.traffic_class = ip4h->tos;
+
+	return 0;
+}
+
+#define IB_NEXT_HDR    0x1b
+static inline int set_ah_attr_by_ipv6(struct ibv_context *context,
+				  struct ibv_ah_attr *ah_attr,
+				  struct ibv_grh *grh, uint8_t port_num)
+{
+	uint32_t flow_class;
+	uint32_t sgid_type;
+	int ret;
+
+	/* No point searching multicast GIDs in GID table */
+	if (grh->dgid.raw[0] == 0xFF) {
+		errno = EINVAL;
+		return -1;
+	}
+
+	ah_attr->grh.dgid = grh->sgid;
+	if (grh->next_hdr == IPPROTO_UDP) {
+		sgid_type = IBV_GID_TYPE_ROCE_V2;
+	} else if (grh->next_hdr == IB_NEXT_HDR) {
+		sgid_type = IBV_GID_TYPE_IB_ROCE_V1;
+	} else {
+		errno = EPROTONOSUPPORT;
+		return -1;
 	}
+
+	ret = ibv_find_gid_index(context, port_num, &grh->dgid,
+				 sgid_type);
+	if (ret < 0)
+		return ret;
+
+	ah_attr->grh.sgid_index = (uint8_t) ret;
+	flow_class = ntohl(grh->version_tclass_flow);
+	ah_attr->grh.hop_limit = grh->hop_limit;
+	ah_attr->grh.traffic_class = (flow_class >> 20) & 0xFF;
+
 	return 0;
 }
 
+int ibv_init_ah_from_wc(struct ibv_context *context, uint8_t port_num,
+			struct ibv_wc *wc, struct ibv_grh *grh,
+			struct ibv_ah_attr *ah_attr)
+{
+	int version;
+	int ret = 0;
+
+	memset(ah_attr, 0, sizeof *ah_attr);
+	set_ah_attr_generic_fields(ah_attr, wc, grh, port_num);
+
+	if (wc->wc_flags & IBV_WC_GRH) {
+		ah_attr->is_global = 1;
+		version = get_grh_header_version(grh);
+
+		if (version == 4)
+			ret = set_ah_attr_by_ipv4(context, ah_attr,
+						  (struct iphdr *)((void *)grh + 20),
+						  port_num);
+		else if (version == 6)
+			ret = set_ah_attr_by_ipv6(context, ah_attr, grh,
+						  port_num);
+		else
+			ret = -1;
+	}
+
+	return ret;
+}
+
 struct ibv_ah *ibv_create_ah_from_wc(struct ibv_pd *pd, struct ibv_wc *wc,
 				     struct ibv_grh *grh, uint8_t port_num)
 {
-- 
1.8.3.1

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html