mlx5: Add reserved_qpn allocation/deallocation support

Support allocate and deallocate a reserved QP number over DEVX APIs. Signed-off-by: Mark Zhang <[email protected]> Signed-off-by: Yishai Hadas <[email protected]>
linux-rdma · Jan 21, 2021 · 628fb16 · 628fb16
1 parent b8986e3
commit 628fb16
Show file tree

Hide file tree

Showing 10 changed files with 257 additions and 1 deletion.
diff --git a/debian/ibverbs-providers.symbols b/debian/ibverbs-providers.symbols
@@ -25,6 +25,7 @@ libmlx5.so.1 ibverbs-providers #MINVER#
  MLX5_1.15@MLX5_1.15 31
  MLX5_1.16@MLX5_1.16 32
  MLX5_1.17@MLX5_1.17 33
+ MLX5_1.18@MLX5_1.18 34
  mlx5dv_init_obj@MLX5_1.0 13
  mlx5dv_init_obj@MLX5_1.2 15
  mlx5dv_query_device@MLX5_1.0 13
@@ -121,6 +122,8 @@ libmlx5.so.1 ibverbs-providers #MINVER#
  mlx5dv_sched_node_create@MLX5_1.17 33
  mlx5dv_sched_node_destroy@MLX5_1.17 33
  mlx5dv_sched_node_modify@MLX5_1.17 33
+ mlx5dv_reserved_qpn_alloc@MLX5_1.18 34
+ mlx5dv_reserved_qpn_dealloc@MLX5_1.18 34
 libefa.so.1 ibverbs-providers #MINVER#
 * Build-Depends-Package: libibverbs-dev
  EFA_1.0@EFA_1.0 24

diff --git a/providers/mlx5/CMakeLists.txt b/providers/mlx5/CMakeLists.txt
@@ -11,7 +11,7 @@ if (MLX5_MW_DEBUG)
 endif()
 
 rdma_shared_provider(mlx5 libmlx5.map
- 1 1.17.${PACKAGE_VERSION}
+ 1 1.18.${PACKAGE_VERSION}
  buf.c
  cq.c
  dbrec.c

diff --git a/providers/mlx5/libmlx5.map b/providers/mlx5/libmlx5.map
@@ -168,3 +168,9 @@ MLX5_1.17 {
  mlx5dv_sched_node_destroy;
  mlx5dv_sched_node_modify;
 } MLX5_1.16;
+
+MLX5_1.18 {
+ global:
+ mlx5dv_reserved_qpn_alloc;
+ mlx5dv_reserved_qpn_dealloc;
+} MLX5_1.17;
diff --git a/providers/mlx5/man/CMakeLists.txt b/providers/mlx5/man/CMakeLists.txt
@@ -30,6 +30,7 @@ rdma_man_pages(
  mlx5dv_pp_alloc.3.md
  mlx5dv_query_device.3
  mlx5dv_query_qp_lag_port.3.md
+ mlx5dv_reserved_qpn_alloc.3.md
  mlx5dv_sched_node_create.3.md
  mlx5dv_ts_to_ns.3
  mlx5dv_wr_post.3.md
@@ -92,6 +93,7 @@ rdma_alias_man_pages(
  mlx5dv_dump.3 mlx5dv_dump_dr_rule.3
  mlx5dv_dump.3 mlx5dv_dump_dr_table.3
  mlx5dv_pp_alloc.3 mlx5dv_pp_free.3
+ mlx5dv_reserved_qpn_alloc.3 mlx5dv_reserved_qpn_dealloc.3
  mlx5dv_sched_node_create.3 mlx5dv_sched_leaf_create.3
  mlx5dv_sched_node_create.3 mlx5dv_sched_leaf_destroy.3
  mlx5dv_sched_node_create.3 mlx5dv_sched_leaf_modify.3

diff --git a/providers/mlx5/man/mlx5dv_reserved_qpn_alloc.3.md b/providers/mlx5/man/mlx5dv_reserved_qpn_alloc.3.md
@@ -0,0 +1,54 @@
+---
+layout: page
+title: mlx5dv_reserved_qpn_alloc / dealloc
+section: 3
+tagline: Verbs
+date: 2020-12-29
+header: "mlx5 Programmer's Manual"
+footer: mlx5
+---
+
+# NAME
+
+mlx5dv_reserved_qpn_alloc - Allocate a reserved QP number from device
+
+mlx5dv_reserved_qpn_dealloc - Release the reserved QP number
+
+# SYNOPSIS
+
+```c
+#include <infiniband/mlx5dv.h>
+
+int mlx5dv_reserved_qpn_alloc(struct ibv_context *ctx, uint32_t *qpn);
+
+int mlx5dv_reserved_qpn_dealloc(struct ibv_context *ctx, uint32_t qpn);
+```
+
+# DESCRIPTION
+
+When work with RDMA_CM RDMA_TCP_PS + external QP support, a client node needs GUID level unique QP numbers to comply with the CM's timewait logic.
+
+If a real unique QP is not allocated, a device global QPN value is required and can be allocated via this interface.
+
+The mlx5 DCI QP is such an example, which could connect to the remote DCT's multiple times as long as the application provides unique QPN for each new RDMA_CM connection.
+
+These 2 APIs provide the allocation/deallocation of a unique QP number from/to device. This qpn can be used with
+DC QPN in RDMA_CM connection establishment, which will comply with the CM timewait kernel logic.
+
+# ARGUMENTS
+
+*ctx*
+: The device context to issue the action on.
+
+*qpn*
+: The allocated QP number (for alloc API), or the QP number to be deallocated (for dealloc API).
+
+# RETURN VALUE
+
+0 on success; EOPNOTSUPP if not supported, or other errno value on other failures.
+
+# AUTHOR
+
+Mark Zhang <[email protected]>
+
+Alex Rosenbaum <[email protected]>
diff --git a/providers/mlx5/mlx5.c b/providers/mlx5/mlx5.c
@@ -1557,6 +1557,160 @@ int mlx5dv_modify_qp_sched_elem(struct ibv_qp *qp,
  }
 }
 
+static struct reserved_qpn_blk *reserved_qpn_blk_alloc(struct mlx5_context *mctx)
+{
+ uint32_t out[DEVX_ST_SZ_DW(general_obj_out_cmd_hdr)] = {};
+ uint32_t in[DEVX_ST_SZ_DW(create_reserved_qpn_in)] = {};
+ struct reserved_qpn_blk *blk;
+ void *attr;
+
+ blk = calloc(1, sizeof(*blk));
+ if (!blk) {
+ errno = ENOMEM;
+ return NULL;
+ }
+
+ blk->bmp = bitmap_alloc0(1 << mctx->hca_cap_2_caps.log_reserved_qpns_per_obj);
+ if (!blk->bmp) {
+ errno = ENOMEM;
+ goto bmp_alloc_fail;
+ }
+
+ attr = DEVX_ADDR_OF(create_reserved_qpn_in, in, hdr);
+ DEVX_SET(general_obj_in_cmd_hdr,
+ attr, opcode, MLX5_CMD_OP_CREATE_GENERAL_OBJECT);
+ DEVX_SET(general_obj_in_cmd_hdr,
+ attr, obj_type, MLX5_OBJ_TYPE_RESERVED_QPN);
+ DEVX_SET(general_obj_in_cmd_hdr,
+ attr, log_obj_range, mctx->hca_cap_2_caps.log_reserved_qpns_per_obj);
+
+ blk->obj = mlx5dv_devx_obj_create(&mctx->ibv_ctx.context,
+ in, sizeof(in), out, sizeof(out));
+ if (!blk->obj)
+ goto obj_alloc_fail;
+
+ blk->first_qpn = blk->obj->object_id;
+ blk->next_avail_slot = 0;
+
+ return blk;
+
+obj_alloc_fail:
+ free(blk->bmp);
+
+bmp_alloc_fail:
+ free(blk);
+ return NULL;
+}
+
+static void reserved_qpn_blk_dealloc(struct reserved_qpn_blk *blk)
+{
+ if (mlx5dv_devx_obj_destroy(blk->obj))
+ assert(false);
+
+ free(blk->bmp);
+ free(blk);
+}
+
+static void reserved_qpn_blks_free(struct mlx5_context *mctx)
+{
+ struct reserved_qpn_blk *blk, *tmp;
+
+ pthread_mutex_lock(&mctx->reserved_qpns.mutex);
+
+ list_for_each_safe(&mctx->reserved_qpns.blk_list,
+ blk, tmp, entry) {
+ list_del(&blk->entry);
+ reserved_qpn_blk_dealloc(blk);
+ }
+
+ pthread_mutex_unlock(&mctx->reserved_qpns.mutex);
+}
+
+/**
+ * Allocate a reserved QPN either from the last FW object allocated,
+ * or by allocating a new one. When find a free QPN in an object, it
+ * always starts from last allocation position, to make sure the QPN
+ * always move forward to prevent stale QPN.
+ */
+int mlx5dv_reserved_qpn_alloc(struct ibv_context *ctx, uint32_t *qpn)
+{
+ struct mlx5_context *mctx = to_mctx(ctx);
+ struct reserved_qpn_blk *blk;
+ uint32_t qpns_per_obj;
+ int ret = 0;
+
+ if (!is_mlx5_dev(ctx->device) ||
+ !(mctx->general_obj_types_caps & (1ULL << MLX5_OBJ_TYPE_RESERVED_QPN)))
+ return EOPNOTSUPP;
+
+ qpns_per_obj = 1 << mctx->hca_cap_2_caps.log_reserved_qpns_per_obj;
+
+ pthread_mutex_lock(&mctx->reserved_qpns.mutex);
+
+ blk = list_tail(&mctx->reserved_qpns.blk_list,
+ struct reserved_qpn_blk, entry);
+ if (!blk ||
+ (blk->next_avail_slot >= qpns_per_obj)) {
+ blk = reserved_qpn_blk_alloc(mctx);
+ if (!blk) {
+ ret = errno;
+ goto end;
+ }
+ list_add_tail(&mctx->reserved_qpns.blk_list, &blk->entry);
+ }
+
+ *qpn = blk->first_qpn + blk->next_avail_slot;
+ bitmap_set_bit(blk->bmp, blk->next_avail_slot);
+ blk->next_avail_slot++;
+
+end:
+ pthread_mutex_unlock(&mctx->reserved_qpns.mutex);
+ return ret;
+}
+
+/**
+ * Deallocate a reserved QPN. The FW object is destroyed only when all QPNs
+ * in this object were used and freed.
+ */
+int mlx5dv_reserved_qpn_dealloc(struct ibv_context *ctx, uint32_t qpn)
+{
+ struct mlx5_context *mctx = to_mctx(ctx);
+ struct reserved_qpn_blk *blk, *tmp;
+ uint32_t qpns_per_obj;
+ bool found = false;
+ int ret = 0;
+
+ qpns_per_obj = 1 << mctx->hca_cap_2_caps.log_reserved_qpns_per_obj;
+
+ pthread_mutex_lock(&mctx->reserved_qpns.mutex);
+
+ list_for_each_safe(&mctx->reserved_qpns.blk_list,
+ blk, tmp, entry) {
+ if ((qpn >= blk->first_qpn) &&
+ (qpn < blk->first_qpn + qpns_per_obj)) {
+ found = true;
+ break;
+ }
+ }
+
+ if (!found || !bitmap_test_bit(blk->bmp, qpn - blk->first_qpn)) {
+ errno = EINVAL;
+ ret = errno;
+ goto end;
+ }
+
+ bitmap_clear_bit(blk->bmp, qpn - blk->first_qpn);
+ if ((blk->next_avail_slot >= qpns_per_obj) &&
+ (bitmap_empty(blk->bmp, qpns_per_obj))) {
+ list_del(&blk->entry);
+ reserved_qpn_blk_dealloc(blk);
+ }
+
+end:
+ pthread_mutex_unlock(&mctx->reserved_qpns.mutex);
+ return ret;
+}
+
 LATEST_SYMVER_FUNC(mlx5dv_init_obj, 1_2, "MLX5_1.2",
  int,
  struct mlx5dv_obj *obj, uint64_t obj_type)
@@ -1961,6 +2115,9 @@ static int mlx5_set_context(struct mlx5_context *context,
  mlx5_set_singleton_nc_uar(&v_ctx->context);
  context->cq_uar_reg = context->nc_uar ? context->nc_uar->uar : context->uar[0].reg;
 
+ pthread_mutex_init(&context->reserved_qpns.mutex, NULL);
+ list_head_init(&context->reserved_qpns.blk_list);
+
  return 0;
 
 err_free_bf:
@@ -2089,6 +2246,7 @@ static void mlx5_free_context(struct ibv_context *ibctx)
  munmap((void *)context->clock_info_page, page_size);
  close_debug_file(context);
  clean_dyn_uars(ibctx);
+ reserved_qpn_blks_free(context);
 
  verbs_uninit_context(&context->ibv_ctx);
  free(context);

diff --git a/providers/mlx5/mlx5.h b/providers/mlx5/mlx5.h
@@ -43,6 +43,7 @@
 #include <util/udma_barrier.h>
 #include <util/util.h>
 #include "mlx5-abi.h"
+#include <ccan/bitmap.h>
 #include <ccan/list.h>
 #include "bitmap.h"
 #include <ccan/minmax.h>
@@ -276,6 +277,19 @@ struct mlx5_hca_cap_2_caps {
  uint32_t log_reserved_qpns_per_obj;
 };
 
+struct reserved_qpn_blk {
+ bitmap *bmp;
+ uint32_t first_qpn;
+ struct list_node entry;
+ unsigned int next_avail_slot;
+ struct mlx5dv_devx_obj *obj;
+};
+
+struct mlx5_reserved_qpns {
+ struct list_head blk_list;
+ pthread_mutex_t mutex;
+};
+
 struct mlx5_context {
  struct verbs_context ibv_ctx;
  int max_num_qps;
@@ -347,6 +361,7 @@ struct mlx5_context {
  struct mlx5_entropy_caps entropy_caps;
  struct mlx5_qos_caps qos_caps;
  struct mlx5_hca_cap_2_caps hca_cap_2_caps;
+ uint64_t general_obj_types_caps;
  uint8_t qpc_extension_cap:1;
  pthread_mutex_t dyn_bfregs_mutex; /* protects the dynamic bfregs allocation */
  uint32_t num_dyn_bfregs;
@@ -367,6 +382,7 @@ struct mlx5_context {
  uint16_t qp_alloc_shared_uuars;
  struct mlx5_bf *nc_uar;
  void *cq_uar_reg;
+ struct mlx5_reserved_qpns reserved_qpns;
 };
 
 struct mlx5_bitmap {

diff --git a/providers/mlx5/mlx5_ifc.h b/providers/mlx5/mlx5_ifc.h
@@ -2446,6 +2446,7 @@ enum {
  MLX5_OBJ_TYPE_ASO_FLOW_METER = 0x0024,
  MLX5_OBJ_TYPE_ASO_FIRST_HIT = 0x0025,
  MLX5_OBJ_TYPE_SCHEDULING_ELEMENT = 0x0026,
+ MLX5_OBJ_TYPE_RESERVED_QPN = 0x002C,
 };
 
 struct mlx5_ifc_general_obj_in_cmd_hdr_bits {
@@ -3487,4 +3488,13 @@ struct mlx5_ifc_modify_sq_in_bits {
  struct mlx5_ifc_sqc_bits sq_context;
 };
 
+struct mlx5_ifc_reserved_qpn_bits {
+ u8 reserved_at_0[0x80];
+};
+
+struct mlx5_ifc_create_reserved_qpn_in_bits {
+ struct mlx5_ifc_general_obj_in_cmd_hdr_bits hdr;
+ struct mlx5_ifc_reserved_qpn_bits rqpns;
+};
+
 #endif /* MLX5_IFC_H */
diff --git a/providers/mlx5/mlx5dv.h b/providers/mlx5/mlx5dv.h
@@ -1657,6 +1657,9 @@ int mlx5dv_modify_qp_sched_elem(struct ibv_qp *qp,
  const struct mlx5dv_sched_leaf *requestor,
  const struct mlx5dv_sched_leaf *responder);
 
+int mlx5dv_reserved_qpn_alloc(struct ibv_context *ctx, uint32_t *qpn);
+int mlx5dv_reserved_qpn_dealloc(struct ibv_context *ctx, uint32_t qpn);
+
 #ifdef __cplusplus
 }
 #endif

diff --git a/providers/mlx5/verbs.c b/providers/mlx5/verbs.c
@@ -3488,6 +3488,10 @@ static void get_hca_general_caps(struct mlx5_context *mctx)
  DEVX_GET(query_hca_cap_out, out,
  capability.cmd_hca_cap.qpc_extension);
 
+ mctx->general_obj_types_caps =
+ DEVX_GET64(query_hca_cap_out, out,
+ capability.cmd_hca_cap.general_obj_types);
+
  if (DEVX_GET(query_hca_cap_out, out,
  capability.cmd_hca_cap.hca_cap_2))
  get_hca_general_caps_2(mctx);