Merge pull request #1426 from hginjgerx/dv

libhns: Introduce hns direct verbs and support userspace congestion control algorithm configuration with dv api

CMakeLists.txt

@@ -737,6 +737,7 @@ add_subdirectory(providers/efa)
 add_subdirectory(providers/efa/man)
 add_subdirectory(providers/erdma)
 add_subdirectory(providers/hns)
+add_subdirectory(providers/hns/man)
 add_subdirectory(providers/irdma)
 add_subdirectory(providers/mana)
 add_subdirectory(providers/mana/man)

debian/ibverbs-providers.install

@@ -1,6 +1,7 @@
 etc/libibverbs.d/
 usr/lib/*/libefa.so.*
 usr/lib/*/libibverbs/lib*-rdmav*.so
+usr/lib/*/libhns.so.*
 usr/lib/*/libmana.so.*
 usr/lib/*/libmlx4.so.*
 usr/lib/*/libmlx5.so.*

debian/ibverbs-providers.lintian-overrides

@@ -1,2 +1,2 @@
-# libefa, libmana, libmlx4 and libmlx5 are ibverbs provider that provides more functions.
-ibverbs-providers: package-name-doesnt-match-sonames libefa1 libmana1 libmlx4-1 libmlx5-1
+# libefa, libhns, libmana, libmlx4 and libmlx5 are ibverbs provider that provides more functions.
+ibverbs-providers: package-name-doesnt-match-sonames libefa1 libhns-1 libmana1 libmlx4-1 libmlx5-1

debian/ibverbs-providers.symbols

@@ -174,6 +174,12 @@ libefa.so.1 ibverbs-providers #MINVER#
  efadv_cq_from_ibv_cq_ex@EFA_1.2 43
  efadv_create_cq@EFA_1.2 43
  efadv_query_mr@EFA_1.3 50
+libhns.so.1 ibverbs-providers #MINVER#
+* Build-Depends-Package: libibverbs-dev
+ HNS_1.0@HNS_1.0 51
+ hnsdv_is_supported@HNS_1.0 51
+ hnsdv_create_qp@HNS_1.0 51
+ hnsdv_query_device@HNS_1.0 51
 libmana.so.1 ibverbs-providers #MINVER#
 * Build-Depends-Package: libibverbs-dev
  MANA_1.0@MANA_1.0 41

debian/libibverbs-dev.install

@@ -1,5 +1,6 @@
 usr/include/infiniband/arch.h
 usr/include/infiniband/efadv.h
+usr/include/infiniband/hnsdv.h
 usr/include/infiniband/ib_user_ioctl_verbs.h
 usr/include/infiniband/manadv.h
 usr/include/infiniband/mlx4dv.h
@@ -15,6 +16,8 @@ usr/include/infiniband/verbs_api.h
 usr/lib/*/lib*-rdmav*.a
 usr/lib/*/libefa.a
 usr/lib/*/libefa.so
+usr/lib/*/libhns.a
+usr/lib/*/libhns.so
 usr/lib/*/libibverbs*.so
 usr/lib/*/libibverbs.a
 usr/lib/*/libmana.a
@@ -24,18 +27,21 @@ usr/lib/*/libmlx4.so
 usr/lib/*/libmlx5.a
 usr/lib/*/libmlx5.so
 usr/lib/*/pkgconfig/libefa.pc
+usr/lib/*/pkgconfig/libhns.pc
 usr/lib/*/pkgconfig/libibverbs.pc
 usr/lib/*/pkgconfig/libmana.pc
 usr/lib/*/pkgconfig/libmlx4.pc
 usr/lib/*/pkgconfig/libmlx5.pc
 usr/share/man/man3/efadv_*.3
+usr/share/man/man3/hnsdv_*.3
 usr/share/man/man3/ibv_*
 usr/share/man/man3/mbps_to_ibv_rate.3
 usr/share/man/man3/manadv_*.3
 usr/share/man/man3/mlx4dv_*.3
 usr/share/man/man3/mlx5dv_*.3
 usr/share/man/man3/mult_to_ibv_rate.3
 usr/share/man/man7/efadv.7
+usr/share/man/man7/hnsdv.7
 usr/share/man/man7/manadv.7
 usr/share/man/man7/mlx4dv.7
 usr/share/man/man7/mlx5dv.7

kernel-headers/rdma/hns-abi.h

@@ -73,6 +73,17 @@ struct hns_roce_ib_create_srq_resp {
  __u32 cap_flags; /* Use enum hns_roce_srq_cap_flags */
 };
 
+enum hns_roce_congest_type_flags {
+ HNS_ROCE_CREATE_QP_FLAGS_DCQCN,
+ HNS_ROCE_CREATE_QP_FLAGS_LDCP,
+ HNS_ROCE_CREATE_QP_FLAGS_HC3,
+ HNS_ROCE_CREATE_QP_FLAGS_DIP,
+};
+
+enum hns_roce_create_qp_comp_mask {
+ HNS_ROCE_CREATE_QP_MASK_CONGEST_TYPE = 1 << 0,
+};
+
 struct hns_roce_ib_create_qp {
  __aligned_u64 buf_addr;
  __aligned_u64 db_addr;
@@ -81,6 +92,9 @@ struct hns_roce_ib_create_qp {
  __u8 sq_no_prefetch;
  __u8 reserved[5];
  __aligned_u64 sdb_addr;
+ __aligned_u64 comp_mask; /* Use enum hns_roce_create_qp_comp_mask */
+ __aligned_u64 create_flags;
+ __aligned_u64 cong_type_flags;
 };
 
 enum hns_roce_qp_cap_flags {
@@ -114,6 +128,8 @@ struct hns_roce_ib_alloc_ucontext_resp {
  __u32 reserved;
  __u32 config;
  __u32 max_inline_data;
+ __u8 congest_type;
+ __u8 reserved0[7];
 };
 
 struct hns_roce_ib_alloc_ucontext {

providers/hns/CMakeLists.txt

@@ -1,7 +1,14 @@
-rdma_provider(hns
+rdma_shared_provider(hns libhns.map
+ 1 1.0.${PACKAGE_VERSION}
  hns_roce_u.c
  hns_roce_u_buf.c
  hns_roce_u_db.c
  hns_roce_u_hw_v2.c
  hns_roce_u_verbs.c
 )
+
+publish_headers(infiniband
+ hnsdv.h
+)
+
+rdma_pkg_config("hns" "libibverbs" "${CMAKE_THREAD_LIBS_INIT}")

providers/hns/hns_roce_u.c

@@ -97,50 +97,33 @@ static uint32_t calc_table_shift(uint32_t entry_count, uint32_t size_shift)
  return count_shift > size_shift ? count_shift - size_shift : 0;
 }
 
-static struct verbs_context *hns_roce_alloc_context(struct ibv_device *ibdev,
-   int cmd_fd,
-   void *private_data)
+static int set_context_attr(struct hns_roce_device *hr_dev,
+ struct hns_roce_context *context,
+ struct hns_roce_alloc_ucontext_resp *resp)
 {
- struct hns_roce_device *hr_dev = to_hr_dev(ibdev);
- struct hns_roce_alloc_ucontext_resp resp = {};
- struct hns_roce_alloc_ucontext cmd = {};
  struct ibv_device_attr dev_attrs;
- struct hns_roce_context *context;
  int i;
 
- context = verbs_init_and_alloc_context(ibdev, cmd_fd, context, ibv_ctx,
- RDMA_DRIVER_HNS);
- if (!context)
- return NULL;
-
- cmd.config |= HNS_ROCE_EXSGE_FLAGS | HNS_ROCE_RQ_INLINE_FLAGS |
- HNS_ROCE_CQE_INLINE_FLAGS;
- if (ibv_cmd_get_context(&context->ibv_ctx, &cmd.ibv_cmd, sizeof(cmd),
- &resp.ibv_resp, sizeof(resp)))
- goto err_free;
-
- if (!resp.cqe_size)
+ if (!resp->cqe_size)
  context->cqe_size = HNS_ROCE_CQE_SIZE;
- else if (resp.cqe_size <= HNS_ROCE_V3_CQE_SIZE)
- context->cqe_size = resp.cqe_size;
+ else if (resp->cqe_size <= HNS_ROCE_V3_CQE_SIZE)
+ context->cqe_size = resp->cqe_size;
  else
  context->cqe_size = HNS_ROCE_V3_CQE_SIZE;
 
- context->config = resp.config;
- if (resp.config & HNS_ROCE_RSP_EXSGE_FLAGS)
- context->max_inline_data = resp.max_inline_data;
+ context->config = resp->config;
+ if (resp->config & HNS_ROCE_RSP_EXSGE_FLAGS)
+ context->max_inline_data = resp->max_inline_data;
 
- context->qp_table_shift = calc_table_shift(resp.qp_tab_size,
+ context->qp_table_shift = calc_table_shift(resp->qp_tab_size,
  HNS_ROCE_QP_TABLE_BITS);
  context->qp_table_mask = (1 << context->qp_table_shift) - 1;
- pthread_mutex_init(&context->qp_table_mutex, NULL);
  for (i = 0; i < HNS_ROCE_QP_TABLE_SIZE; ++i)
  context->qp_table[i].refcnt = 0;
 
- context->srq_table_shift = calc_table_shift(resp.srq_tab_size,
+ context->srq_table_shift = calc_table_shift(resp->srq_tab_size,
  HNS_ROCE_SRQ_TABLE_BITS);
  context->srq_table_mask = (1 << context->srq_table_shift) - 1;
- pthread_mutex_init(&context->srq_table_mutex, NULL);
  for (i = 0; i < HNS_ROCE_SRQ_TABLE_SIZE; ++i)
  context->srq_table[i].refcnt = 0;
 
@@ -149,20 +132,49 @@ static struct verbs_context *hns_roce_alloc_context(struct ibv_device *ibdev,
  struct ibv_device_attr_ex,
  orig_attr),
  sizeof(dev_attrs)))
- goto err_free;
+ return EIO;
 
  hr_dev->hw_version = dev_attrs.hw_ver;
+ hr_dev->congest_cap = resp->congest_type;
  context->max_qp_wr = dev_attrs.max_qp_wr;
  context->max_sge = dev_attrs.max_sge;
  context->max_cqe = dev_attrs.max_cqe;
  context->max_srq_wr = dev_attrs.max_srq_wr;
  context->max_srq_sge = dev_attrs.max_srq_sge;
 
+ return 0;
+}
+
+static struct verbs_context *hns_roce_alloc_context(struct ibv_device *ibdev,
+ int cmd_fd,
+ void *private_data)
+{
+ struct hns_roce_device *hr_dev = to_hr_dev(ibdev);
+ struct hns_roce_alloc_ucontext_resp resp = {};
+ struct hns_roce_alloc_ucontext cmd = {};
+ struct hns_roce_context *context;
+
+ context = verbs_init_and_alloc_context(ibdev, cmd_fd, context, ibv_ctx,
+ RDMA_DRIVER_HNS);
+ if (!context)
+ return NULL;
+
+ cmd.config |= HNS_ROCE_EXSGE_FLAGS | HNS_ROCE_RQ_INLINE_FLAGS |
+ HNS_ROCE_CQE_INLINE_FLAGS;
+ if (ibv_cmd_get_context(&context->ibv_ctx, &cmd.ibv_cmd, sizeof(cmd),
+ &resp.ibv_resp, sizeof(resp)))
+ goto err_free;
+
+ if (set_context_attr(hr_dev, context, &resp))
+ goto err_free;
+
  context->uar = mmap(NULL, hr_dev->page_size, PROT_READ | PROT_WRITE,
  MAP_SHARED, cmd_fd, 0);
  if (context->uar == MAP_FAILED)
  goto err_free;
 
+ pthread_mutex_init(&context->qp_table_mutex, NULL);
+ pthread_mutex_init(&context->srq_table_mutex, NULL);
  pthread_spin_init(&context->uar_lock, PTHREAD_PROCESS_PRIVATE);
 
  verbs_set_ops(&context->ibv_ctx, &hns_common_ops);
@@ -216,4 +228,17 @@ static const struct verbs_device_ops hns_roce_dev_ops = {
  .uninit_device = hns_uninit_device,
  .alloc_context = hns_roce_alloc_context,
 };
+
+bool is_hns_dev(struct ibv_device *device)
+{
+ struct verbs_device *verbs_device = verbs_get_device(device);
+
+ return verbs_device->ops == &hns_roce_dev_ops;
+}
+
+bool hnsdv_is_supported(struct ibv_device *device)
+{
+ return is_hns_dev(device);
+}
+
 PROVIDER_DRIVER(hns, hns_roce_dev_ops);

providers/hns/hns_roce_u.h

@@ -158,6 +158,7 @@ struct hns_roce_device {
  int page_size;
  const struct hns_roce_u_hw *u_hw;
  int hw_version;
+ uint8_t congest_cap;
 };
 
 struct hns_roce_buf {
@@ -494,6 +495,8 @@ void hns_roce_free_qp_buf(struct hns_roce_qp *qp, struct hns_roce_context *ctx);
 
 void hns_roce_init_qp_indices(struct hns_roce_qp *qp);
 
+bool is_hns_dev(struct ibv_device *device);
+
 extern const struct hns_roce_u_hw hns_roce_u_hw_v2;
 
 #endif /* _HNS_ROCE_U_H */

providers/hns/hns_roce_u_abi.h

@@ -36,6 +36,7 @@
 #include <infiniband/kern-abi.h>
 #include <rdma/hns-abi.h>
 #include <kernel-abi/hns-abi.h>
+#include "hnsdv.h"
 
 DECLARE_DRV_CMD(hns_roce_alloc_pd, IB_USER_VERBS_CMD_ALLOC_PD,
  empty, hns_roce_ib_alloc_pd_resp);