From: Avihai Horon <avihaih@xxxxxxxxxx>
Introduce the ibv_query_gid_table verb which queries the GID tables of
the given device and stores the queried data in a buffer provided by the
user.
If the kernel doesn't support ioctl or the needed uverbs method, the
API will try to query the GID tables via sysfs.
This API provides a faster way to query the GID tables of a device using
a single call over ioctl, instead of multiple calls to open, close and
read multiple sysfs files for a single GID table entry.
Signed-off-by: Avihai Horon <avihaih@xxxxxxxxxx>
Signed-off-by: Yishai Hadas <yishaih@xxxxxxxxxx>
---
debian/libibverbs1.symbols | 1 +
libibverbs/cmd_device.c | 117 +++++++++++++++++++++++++++++---
libibverbs/libibverbs.map.in | 1 +
libibverbs/man/CMakeLists.txt | 1 +
libibverbs/man/ibv_query_gid_table.3.md | 73 ++++++++++++++++++++
libibverbs/verbs.h | 16 +++++
6 files changed, 199 insertions(+), 10 deletions(-)
create mode 100644 libibverbs/man/ibv_query_gid_table.3.md
diff --git a/debian/libibverbs1.symbols b/debian/libibverbs1.symbols
index 536d543..99257de 100644
--- a/debian/libibverbs1.symbols
+++ b/debian/libibverbs1.symbols
@@ -11,6 +11,7 @@ libibverbs.so.1 libibverbs1 #MINVER#
IBVERBS_1.11@IBVERBS_1.11 32
(symver)IBVERBS_PRIVATE_25 25
_ibv_query_gid_ex@IBVERBS_1.11 32
+ _ibv_query_gid_table@IBVERBS_1.11 32
ibv_ack_async_event@IBVERBS_1.0 1.1.6
ibv_ack_async_event@IBVERBS_1.1 1.1.6
ibv_ack_cq_events@IBVERBS_1.0 1.1.6
diff --git a/libibverbs/cmd_device.c b/libibverbs/cmd_device.c
index 4f85010..8bcfbb4 100644
--- a/libibverbs/cmd_device.c
+++ b/libibverbs/cmd_device.c
@@ -321,7 +321,7 @@ static int query_sysfs_gid_type(struct ibv_context *context, uint8_t port_num,
static int query_sysfs_gid_entry(struct ibv_context *context, uint32_t port_num,
uint32_t gid_index,
struct ibv_gid_entry *entry,
- uint32_t attr_mask)
+ uint32_t attr_mask, int link_layer)
{
enum ibv_gid_type_sysfs gid_type;
struct ibv_port_attr port_attr = {};
@@ -342,14 +342,18 @@ static int query_sysfs_gid_entry(struct ibv_context *context, uint32_t port_num,
return EINVAL;
if (gid_type == IBV_GID_TYPE_SYSFS_IB_ROCE_V1) {
- ret = ibv_query_port(context, port_num, &port_attr);
- if (ret)
- goto out;
+ if (link_layer < 0) {
+ ret = ibv_query_port(context, port_num,
+ &port_attr);
+ if (ret)
+ goto out;
+
+ link_layer = port_attr.link_layer;
+ }
- if (port_attr.link_layer == IBV_LINK_LAYER_INFINIBAND) {
+ if (link_layer == IBV_LINK_LAYER_INFINIBAND) {
entry->gid_type = IBV_GID_TYPE_IB;
- } else if (port_attr.link_layer ==
- IBV_LINK_LAYER_ETHERNET) {
+ } else if (link_layer == IBV_LINK_LAYER_ETHERNET) {
entry->gid_type = IBV_GID_TYPE_ROCE_V1;
} else {
ret = EINVAL;
@@ -368,8 +372,64 @@ out:
return ret;
}
-/* Using async_event cmd_name because query_gid_ex is not in
- * verbs_context_ops while async_event is and doesn't use ioctl.
+static int query_gid_table_fb(struct ibv_context *context,
+ struct ibv_gid_entry *entries, size_t max_entries,
+ uint64_t *num_entries, size_t entry_size)
+{
+ struct ibv_device_attr dev_attr = {};
+ struct ibv_port_attr port_attr = {};
+ struct ibv_gid_entry entry = {};
+ int attr_mask;
+ void *tmp;
+ int i, j;
+ int ret;
+
+ ret = ibv_query_device(context, &dev_attr);
+ if (ret)
+ goto out;
+
+ tmp = entries;
+ *num_entries = 0;
+ attr_mask = VERBS_QUERY_GID_ATTR_GID | VERBS_QUERY_GID_ATTR_TYPE |
+ VERBS_QUERY_GID_ATTR_NDEV_IFINDEX;
+ for (i = 0; i < dev_attr.phys_port_cnt; i++) {
+ ret = ibv_query_port(context, i + 1, &port_attr);
+ if (ret)
+ goto out;
+
+ for (j = 0; j < port_attr.gid_tbl_len; j++) {
+ /* In case we already reached max_entries, query to some
+ * temp entry, in case all other entries are zeros the
+ * API should succceed.
+ */
+ if (*num_entries == max_entries)
+ tmp = &entry;
+ ret = query_sysfs_gid_entry(context, i + 1, j,
+ tmp,
+ attr_mask,
+ port_attr.link_layer);
+ if (ret)
+ goto out;
+ if (is_zero_gid(&((struct ibv_gid_entry *)tmp)->gid))
+ continue;
+ if (*num_entries == max_entries) {
+ ret = EINVAL;
+ goto out;
+ }
+
+ (*num_entries)++;
+ tmp += entry_size;
+ }
+ }
+
+out:
+ return ret;
+}
+
+/* Using async_event cmd_name because query_gid_ex and query_gid_table are not
+ * in verbs_context_ops while async_event is and doesn't use ioctl.
+ * If one of them is not supported, so is the other. Hence, we can use a single
+ * cmd_name for both of them.
*/
#define query_gid_kernel_cap async_event
int __ibv_query_gid_ex(struct ibv_context *context, uint32_t port_num,
@@ -395,7 +455,7 @@ int __ibv_query_gid_ex(struct ibv_context *context, uint32_t port_num,
return EOPNOTSUPP;
ret = query_sysfs_gid_entry(context, port_num, gid_index,
- entry, fallback_attr_mask);
+ entry, fallback_attr_mask, -1);
if (ret)
return ret;
@@ -419,3 +479,40 @@ int _ibv_query_gid_ex(struct ibv_context *context, uint32_t port_num,
VERBS_QUERY_GID_ATTR_TYPE |
VERBS_QUERY_GID_ATTR_NDEV_IFINDEX);
}
+
+ssize_t _ibv_query_gid_table(struct ibv_context *context,
+ struct ibv_gid_entry *entries,
+ size_t max_entries, uint32_t flags,
+ size_t entry_size)
+{
+ DECLARE_COMMAND_BUFFER(cmdb, UVERBS_OBJECT_DEVICE,
+ UVERBS_METHOD_QUERY_GID_TABLE, 4);
+ uint64_t num_entries;
+ int ret;
+
+ fill_attr_const_in(cmdb, UVERBS_ATTR_QUERY_GID_TABLE_ENTRY_SIZE,
+ entry_size);
+ fill_attr_in_uint32(cmdb, UVERBS_ATTR_QUERY_GID_TABLE_FLAGS, flags);
+ fill_attr_out(cmdb, UVERBS_ATTR_QUERY_GID_TABLE_RESP_ENTRIES, entries,
+ _array_len(entry_size, max_entries));
+ fill_attr_out_ptr(cmdb, UVERBS_ATTR_QUERY_GID_TABLE_RESP_NUM_ENTRIES,
+ &num_entries);
+
+ switch (execute_ioctl_fallback(context, query_gid_kernel_cap, cmdb,
+ &ret)) {
+ case TRY_WRITE:
+ if (flags)
+ return -EOPNOTSUPP;
+
+ ret = query_gid_table_fb(context, entries, max_entries,
+ &num_entries, entry_size);
+ break;
+ default:
+ break;
+ }
+
+ if (ret)
+ return -ret;
+
+ return num_entries;
+}
diff --git a/libibverbs/libibverbs.map.in b/libibverbs/libibverbs.map.in
index dae4963..7429016 100644
--- a/libibverbs/libibverbs.map.in
+++ b/libibverbs/libibverbs.map.in
@@ -145,6 +145,7 @@ IBVERBS_1.10 {
IBVERBS_1.11 {
global:
_ibv_query_gid_ex;
+ _ibv_query_gid_table;
} IBVERBS_1.10;
/* If any symbols in this stanza change ABI then the entire staza gets a new symbol
diff --git a/libibverbs/man/CMakeLists.txt b/libibverbs/man/CMakeLists.txt
index 2dea4ff..1fb5ac1 100644
--- a/libibverbs/man/CMakeLists.txt
+++ b/libibverbs/man/CMakeLists.txt
@@ -58,6 +58,7 @@ rdma_man_pages(
ibv_query_ece.3.md
ibv_query_gid.3.md
ibv_query_gid_ex.3.md
+ ibv_query_gid_table.3.md
ibv_query_pkey.3.md
ibv_query_port.3
ibv_query_qp.3
diff --git a/libibverbs/man/ibv_query_gid_table.3.md b/libibverbs/man/ibv_query_gid_table.3.md
new file mode 100644
index 0000000..e10f51c
--- /dev/null
+++ b/libibverbs/man/ibv_query_gid_table.3.md
@@ -0,0 +1,73 @@
+---
+date: 2020-04-24
+footer: libibverbs
+header: "Libibverbs Programmer's Manual"
+layout: page
+license: 'Licensed under the OpenIB.org BSD license (FreeBSD Variant) - See COPYING.md'
+section: 3
+title: IBV_QUERY_GID_TABLE
+---
+
+# NAME
+
+ibv_query_gid_table - query an InfiniBand device's GID table
+
+# SYNOPSIS
+
+```c
+#include <infiniband/verbs.h>
+
+ssize_t ibv_query_gid_table(struct ibv_context *context,
+ struct ibv_gid_entry *entries,
+ size_t max_entries,
+ uint32_t flags);
+```
+
+# DESCRIPTION
+
+**ibv_query_gid_table()** returns the valid GID table entries of the RDMA
+device context *context* at the pointer *entries*.
+
+A caller must allocate *entries* array for the GID table entries it
+desires to query. This API returns only valid GID table entries.
+
+A caller must pass non zero number of entries at *max_entries* that corresponds
+to the size of *entries* array.
+
+*entries* array must be allocated such that it can contain all the valid
+GID table entries of the device. If there are more valid GID entries than
+the provided value of *max_entries* and *entries* array, the call will fail.
+For example, if a RDMA device *context* has a total of 10 valid
+GID entries, *entries* should be allocated for at least 10 entries, and
+*max_entries* should be set appropriately.
+
+# ARGUMENTS
+
+*context*
+: The context of the device to query.
+
+*entries*
+: Array of ibv_gid_entry structs where the GID entries are returned.
+ Please see **ibv_query_gid_ex**(3) man page for *ibv_gid_entry*.
+
+*max_entries*
+: Maximum number of entries that can be returned.
+
+*flags*
+: Extra fields to query post *entries->ndev_ifindex*, for now must be 0.
+
+# RETURN VALUE
+
+**ibv_query_gid_table()** returns the number of entries that were read on success or negative errno value on error.
+Number of entries returned is <= max_entries.
+
+# SEE ALSO
+
+**ibv_open_device**(3),
+**ibv_query_device**(3),
+**ibv_query_port**(3),
+**ibv_query_gid_ex**(3)
+
+# AUTHOR
+
+Parav Pandit <parav@xxxxxxxxxx>
diff --git a/libibverbs/verbs.h b/libibverbs/verbs.h
index e5bf900..caf626c 100644
--- a/libibverbs/verbs.h
+++ b/libibverbs/verbs.h
@@ -43,6 +43,7 @@
#include <string.h>
#include <linux/types.h>
#include <stdint.h>
+#include <sys/types.h>
#include <infiniband/verbs_api.h>
#ifdef __cplusplus
@@ -2359,6 +2360,21 @@ static inline int ibv_query_gid_ex(struct ibv_context *context,
sizeof(*entry));
}
+ssize_t _ibv_query_gid_table(struct ibv_context *context,
+ struct ibv_gid_entry *entries, size_t max_entries,
+ uint32_t flags, size_t entry_size);
+
+/*
+ * ibv_query_gid_table - Get all valid GID table entries
+ */
+static inline ssize_t ibv_query_gid_table(struct ibv_context *context,
+ struct ibv_gid_entry *entries,
+ size_t max_entries, uint32_t flags)
+{
+ return _ibv_query_gid_table(context, entries, max_entries, flags,
+ sizeof(*entries));
+}
+
/**
* ibv_query_pkey - Get a P_Key table entry
*/
--
1.8.3.1
