[PATCH RFC rdma-core 0/5] libhns: Add support for Dynamic Context Attachment
The HIP09 introduces the DCA(Dynamic Context Attachment) feature which supports many RC QPs to share the WQE buffer in a memory pool. If a QP enables DCA feature, the WQE's buffer will not be allocated when creating but when the users start to post WRs. This will reduce the memory consumption when there are too many QPs are inactive. Please note that we didn't find the right way to get user's configuration, so in #4 we still use environment variable to achieve this. We will be appreciated if anyone can provide some sugggestions. This series is associated with the kernel series "RDMA/hns: Add support for Dynamic Context Attachment", you can review it at: https://patchwork.kernel.org/project/linux-rdma/cover/1611394994-50363-1-git... Weihang Li (1): Update kernel headers Xi Wang (4): libhns: Introduce DCA for RC QP libhns: Add support for shrinking DCA memory pool libhns: Add support for attaching QP's WQE buffer libhns: Add support for configuring DCA kernel-headers/rdma/hns-abi.h | 64 +++++++ libibverbs/cmd_qp.c | 3 +- libibverbs/verbs.h | 1 + providers/hns/hns_roce_u.c | 89 +++++++++ providers/hns/hns_roce_u.h | 42 +++++ providers/hns/hns_roce_u_buf.c | 384 +++++++++++++++++++++++++++++++++++++++ providers/hns/hns_roce_u_hw_v2.c | 138 +++++++++++++- providers/hns/hns_roce_u_hw_v2.h | 7 + providers/hns/hns_roce_u_verbs.c | 44 ++++- 9 files changed, 754 insertions(+), 18 deletions(-) -- 2.8.1
To commit ?? ("RDMA/hns: Add method to query WQE buffer's address"). Signed-off-by: Weihang Li <liweihang@huawei.com> --- kernel-headers/rdma/hns-abi.h | 64 +++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 64 insertions(+) diff --git a/kernel-headers/rdma/hns-abi.h b/kernel-headers/rdma/hns-abi.h index 90b739d..b832f26 100644 --- a/kernel-headers/rdma/hns-abi.h +++ b/kernel-headers/rdma/hns-abi.h @@ -77,19 +77,83 @@ enum hns_roce_qp_cap_flags { HNS_ROCE_QP_CAP_RQ_RECORD_DB = 1 << 0, HNS_ROCE_QP_CAP_SQ_RECORD_DB = 1 << 1, HNS_ROCE_QP_CAP_OWNER_DB = 1 << 2, + HNS_ROCE_QP_CAP_DCA = 1 << 4, }; struct hns_roce_ib_create_qp_resp { __aligned_u64 cap_flags; }; +enum { + HNS_ROCE_CAP_FLAG_DCA_MODE = 1 << 15, +}; + struct hns_roce_ib_alloc_ucontext_resp { __u32 qp_tab_size; __u32 cqe_size; + __aligned_u64 cap_flags; }; struct hns_roce_ib_alloc_pd_resp { __u32 pdn; }; +#define UVERBS_ID_NS_MASK 0xF000 +#define UVERBS_ID_NS_SHIFT 12 + +enum hns_ib_objects { + HNS_IB_OBJECT_DCA_MEM = (1U << UVERBS_ID_NS_SHIFT), +}; + +enum hns_ib_dca_mem_methods { + HNS_IB_METHOD_DCA_MEM_REG = (1U << UVERBS_ID_NS_SHIFT), + HNS_IB_METHOD_DCA_MEM_DEREG, + HNS_IB_METHOD_DCA_MEM_SHRINK, + HNS_IB_METHOD_DCA_MEM_ATTACH, + HNS_IB_METHOD_DCA_MEM_DETACH, + HNS_IB_METHOD_DCA_MEM_QUERY, +}; + +enum hns_ib_dca_mem_reg_attrs { + HNS_IB_ATTR_DCA_MEM_REG_HANDLE = (1U << UVERBS_ID_NS_SHIFT), + HNS_IB_ATTR_DCA_MEM_REG_LEN, + HNS_IB_ATTR_DCA_MEM_REG_ADDR, + HNS_IB_ATTR_DCA_MEM_REG_KEY, +}; + +enum hns_ib_dca_mem_dereg_attrs { + HNS_IB_ATTR_DCA_MEM_DEREG_HANDLE = (1U << UVERBS_ID_NS_SHIFT), +}; + +enum hns_ib_dca_mem_shrink_attrs { + HNS_IB_ATTR_DCA_MEM_SHRINK_HANDLE = (1U << UVERBS_ID_NS_SHIFT), + HNS_IB_ATTR_DCA_MEM_SHRINK_RESERVED_SIZE, + HNS_IB_ATTR_DCA_MEM_SHRINK_OUT_FREE_KEY, + HNS_IB_ATTR_DCA_MEM_SHRINK_OUT_FREE_MEMS, +}; + +#define HNS_IB_ATTACH_FLAGS_NEW_BUFFER 1U + +enum hns_ib_dca_mem_attach_attrs { + HNS_IB_ATTR_DCA_MEM_ATTACH_HANDLE = (1U << UVERBS_ID_NS_SHIFT), + HNS_IB_ATTR_DCA_MEM_ATTACH_SQ_OFFSET, + HNS_IB_ATTR_DCA_MEM_ATTACH_SGE_OFFSET, + HNS_IB_ATTR_DCA_MEM_ATTACH_RQ_OFFSET, + HNS_IB_ATTR_DCA_MEM_ATTACH_OUT_ALLOC_FLAGS, + HNS_IB_ATTR_DCA_MEM_ATTACH_OUT_ALLOC_PAGES, +}; + +enum hns_ib_dca_mem_detach_attrs { + HNS_IB_ATTR_DCA_MEM_DETACH_HANDLE = (1U << UVERBS_ID_NS_SHIFT), + HNS_IB_ATTR_DCA_MEM_DETACH_SQ_INDEX, +}; + +enum hns_ib_dca_mem_query_attrs { + HNS_IB_ATTR_DCA_MEM_QUERY_HANDLE = (1U << UVERBS_ID_NS_SHIFT), + HNS_IB_ATTR_DCA_MEM_QUERY_PAGE_INDEX, + HNS_IB_ATTR_DCA_MEM_QUERY_OUT_KEY, + HNS_IB_ATTR_DCA_MEM_QUERY_OUT_OFFSET, + HNS_IB_ATTR_DCA_MEM_QUERY_OUT_PAGE_COUNT, +}; + #endif /* HNS_ABI_USER_H */ -- 2.8.1
From: Xi Wang <wangxi11@huawei.com> The HIP09 introduces the DCA(Dynamic context attachment) feature which supports many RC QPs to share the WQE buffer in a memory pool, this will reduce the memory consumption when there are too many QPs inactive. This patch wraps two functions for adding buffers to memory pool and removing buffers from memory pool by calling ib cmd implemented in hns kernel driver. If a QP enables DCA feature, the WQE's buffer will be attached to the memroy pool when the users start to post WRs and be detached when all CQEs has been polled. Signed-off-by: Xi Wang <wangxi11@huawei.com> Signed-off-by: Weihang Li <liweihang@huawei.com> --- providers/hns/hns_roce_u.c | 41 +++++++++++++ providers/hns/hns_roce_u.h | 18 ++++++ providers/hns/hns_roce_u_buf.c | 134 +++++++++++++++++++++++++++++++++++++++++ 3 files changed, 193 insertions(+) diff --git a/providers/hns/hns_roce_u.c b/providers/hns/hns_roce_u.c index e63ef32..28b1130 100644 --- a/providers/hns/hns_roce_u.c +++ b/providers/hns/hns_roce_u.c @@ -89,6 +89,40 @@ static const struct verbs_context_ops hns_common_ops = { .destroy_ah = hns_roce_u_destroy_ah, }; +static int init_dca_context(struct hns_roce_context *ctx, int page_size) +{ + struct hns_roce_dca_ctx *dca_ctx = &ctx->dca_ctx; + int ret; + + if (!(ctx->cap_flags & HNS_ROCE_CAP_FLAG_DCA_MODE)) + return 0; + + list_head_init(&dca_ctx->mem_list); + ret = pthread_spin_init(&dca_ctx->lock, PTHREAD_PROCESS_PRIVATE); + if (ret) + return ret; + + dca_ctx->unit_size = page_size * HNS_DCA_DEFAULT_UNIT_PAGES; + dca_ctx->max_size = HNS_DCA_MAX_MEM_SIZE; + dca_ctx->mem_cnt = 0; + + return 0; +} + +static void uninit_dca_context(struct hns_roce_context *ctx) +{ + struct hns_roce_dca_ctx *dca_ctx = &ctx->dca_ctx; + + if (!(ctx->cap_flags & HNS_ROCE_CAP_FLAG_DCA_MODE)) + return; + + pthread_spin_lock(&dca_ctx->lock); + hns_roce_cleanup_dca_mem(ctx); + pthread_spin_unlock(&dca_ctx->lock); + + pthread_spin_destroy(&dca_ctx->lock); +} + static struct verbs_context *hns_roce_alloc_context(struct ibv_device *ibdev, int cmd_fd, void *private_data) @@ -110,6 +144,8 @@ static struct verbs_context *hns_roce_alloc_context(struct ibv_device *ibdev, &resp.ibv_resp, sizeof(resp))) goto err_free; + context->cap_flags = resp.cap_flags; + context->num_qps = resp.qp_tab_size; context->qp_table_shift = ffs(context->num_qps) - 1 - HNS_ROCE_QP_TABLE_BITS; @@ -162,6 +198,9 @@ static struct verbs_context *hns_roce_alloc_context(struct ibv_device *ibdev, verbs_set_ops(&context->ibv_ctx, &hns_common_ops); verbs_set_ops(&context->ibv_ctx, &hr_dev->u_hw->hw_ops); + if (init_dca_context(context, hr_dev->page_size)) + goto db_free; + return &context->ibv_ctx; db_free: @@ -183,6 +222,8 @@ static void hns_roce_free_context(struct ibv_context *ibctx) if (hr_dev->hw_version == HNS_ROCE_HW_VER1) munmap(context->cq_tptr_base, HNS_ROCE_CQ_DB_BUF_SIZE); + uninit_dca_context(context); + verbs_uninit_context(&context->ibv_ctx); free(context); } diff --git a/providers/hns/hns_roce_u.h b/providers/hns/hns_roce_u.h index 67b4433..619b060 100644 --- a/providers/hns/hns_roce_u.h +++ b/providers/hns/hns_roce_u.h @@ -142,8 +142,21 @@ struct hns_roce_db_page { bitmap *bitmap; }; +#define HNS_DCA_MAX_MEM_SIZE ~0UL +#define HNS_DCA_DEFAULT_UNIT_PAGES 16 + +struct hns_roce_dca_ctx { + struct list_head mem_list; + pthread_spinlock_t lock; + uint64_t max_size; + uint64_t curr_size; + int mem_cnt; + unsigned int unit_size; +}; + struct hns_roce_context { struct verbs_context ibv_ctx; + uint32_t cap_flags; void *uar; pthread_spinlock_t uar_lock; @@ -167,6 +180,8 @@ struct hns_roce_context { unsigned int max_sge; int max_cqe; unsigned int cqe_size; + + struct hns_roce_dca_ctx dca_ctx; }; struct hns_roce_pd { @@ -395,6 +410,9 @@ void hns_roce_free_buf(struct hns_roce_buf *buf); void hns_roce_free_qp_buf(struct hns_roce_qp *qp, struct hns_roce_context *ctx); +void hns_roce_cleanup_dca_mem(struct hns_roce_context *ctx); +int hns_roce_add_dca_mem(struct hns_roce_context *ctx, uint32_t size); + void hns_roce_init_qp_indices(struct hns_roce_qp *qp); extern const struct hns_roce_u_hw hns_roce_u_hw_v1; diff --git a/providers/hns/hns_roce_u_buf.c b/providers/hns/hns_roce_u_buf.c index 471dd9c..424f916 100644 --- a/providers/hns/hns_roce_u_buf.c +++ b/providers/hns/hns_roce_u_buf.c @@ -60,3 +60,137 @@ void hns_roce_free_buf(struct hns_roce_buf *buf) munmap(buf->buf, buf->length); } + +struct hns_roce_dca_mem { + uint32_t handle; + struct list_node entry; + struct hns_roce_buf buf; + struct hns_roce_context *ctx; +}; + +static void free_dca_mem(struct hns_roce_context *ctx, + struct hns_roce_dca_mem *mem) +{ + hns_roce_free_buf(&mem->buf); + free(mem); +} + +static struct hns_roce_dca_mem *alloc_dca_mem(uint32_t size) +{ + struct hns_roce_dca_mem *mem = NULL; + int ret; + + mem = malloc(sizeof(struct hns_roce_dca_mem)); + if (!mem) { + errno = ENOMEM; + return NULL; + } + + ret = hns_roce_alloc_buf(&mem->buf, size, HNS_HW_PAGE_SIZE); + if (ret) { + errno = ENOMEM; + free(mem); + return NULL; + } + + return mem; +} + +static inline uint64_t dca_mem_to_key(struct hns_roce_dca_mem *dca_mem) +{ + return (uintptr_t)dca_mem; +} + +static inline void *dca_mem_addr(struct hns_roce_dca_mem *dca_mem, int offset) +{ + return dca_mem->buf.buf + offset; +} + +static int register_dca_mem(struct hns_roce_context *ctx, uint64_t key, + void *addr, uint32_t size, uint32_t *handle) +{ + struct ib_uverbs_attr *attr; + int ret; + + DECLARE_COMMAND_BUFFER(cmd, HNS_IB_OBJECT_DCA_MEM, + HNS_IB_METHOD_DCA_MEM_REG, 4); + fill_attr_in_uint32(cmd, HNS_IB_ATTR_DCA_MEM_REG_LEN, size); + fill_attr_in_uint64(cmd, HNS_IB_ATTR_DCA_MEM_REG_ADDR, (intptr_t)addr); + fill_attr_in_uint64(cmd, HNS_IB_ATTR_DCA_MEM_REG_KEY, key); + attr = fill_attr_out_obj(cmd, HNS_IB_ATTR_DCA_MEM_REG_HANDLE); + + ret = execute_ioctl(&ctx->ibv_ctx.context, cmd); + if (!ret) + *handle = read_attr_obj(HNS_IB_ATTR_DCA_MEM_REG_HANDLE, attr); + + return ret; +} + +static void deregister_dca_mem(struct hns_roce_context *ctx, uint32_t handle) +{ + DECLARE_COMMAND_BUFFER(cmd, HNS_IB_OBJECT_DCA_MEM, + HNS_IB_METHOD_DCA_MEM_DEREG, 1); + fill_attr_in_obj(cmd, HNS_IB_ATTR_DCA_MEM_DEREG_HANDLE, handle); + execute_ioctl(&ctx->ibv_ctx.context, cmd); +} + +void hns_roce_cleanup_dca_mem(struct hns_roce_context *ctx) +{ + struct hns_roce_dca_ctx *dca_ctx = &ctx->dca_ctx; + struct hns_roce_dca_mem *mem; + struct hns_roce_dca_mem *tmp; + + list_for_each_safe(&dca_ctx->mem_list, mem, tmp, entry) + deregister_dca_mem(ctx, mem->handle); +} + +static bool add_dca_mem_enabled(struct hns_roce_dca_ctx *ctx, + uint32_t alloc_size) +{ + bool enable; + + pthread_spin_lock(&ctx->lock); + + if (ctx->unit_size == 0) /* Pool size can't be increased */ + enable = false; + else if (ctx->max_size == HNS_DCA_MAX_MEM_SIZE) /* Pool size no limit */ + enable = true; + else /* Pool size doesn't exceed max size */ + enable = (ctx->curr_size + alloc_size) < ctx->max_size; + + pthread_spin_unlock(&ctx->lock); + + return enable; +} + +int hns_roce_add_dca_mem(struct hns_roce_context *ctx, uint32_t size) +{ + struct hns_roce_dca_ctx *dca_ctx = &ctx->dca_ctx; + struct hns_roce_dca_mem *mem; + int ret; + + if (!add_dca_mem_enabled(&ctx->dca_ctx, size)) + return -ENOMEM; + + /* Step 1: Alloc DCA mem address */ + mem = alloc_dca_mem(DIV_ROUND_UP(size, dca_ctx->unit_size)); + if (!mem) + return -ENOMEM; + + /* Step 2: Register DCA mem uobject to pin user address */ + ret = register_dca_mem(ctx, dca_mem_to_key(mem), dca_mem_addr(mem, 0), + mem->buf.length, &mem->handle); + if (ret) { + free_dca_mem(ctx, mem); + return ret; + } + + /* Step 3: Add DCA mem node to pool */ + pthread_spin_lock(&dca_ctx->lock); + list_add_tail(&dca_ctx->mem_list, &mem->entry); + dca_ctx->mem_cnt++; + dca_ctx->curr_size += mem->buf.length; + pthread_spin_unlock(&dca_ctx->lock); + + return 0; +} -- 2.8.1
On Sun, Feb 07, 2021 at 11:12:51AM +0800, Weihang Li wrote:
+static int register_dca_mem(struct hns_roce_context *ctx, uint64_t key, + void *addr, uint32_t size, uint32_t *handle) +{ + struct ib_uverbs_attr *attr; + int ret; + + DECLARE_COMMAND_BUFFER(cmd, HNS_IB_OBJECT_DCA_MEM, + HNS_IB_METHOD_DCA_MEM_REG, 4); + fill_attr_in_uint32(cmd, HNS_IB_ATTR_DCA_MEM_REG_LEN, size); + fill_attr_in_uint64(cmd, HNS_IB_ATTR_DCA_MEM_REG_ADDR, (intptr_t)addr);
This should use ioctl_ptr_to_u64(), the place this was copied from should also be fixed
+ fill_attr_in_uint64(cmd, HNS_IB_ATTR_DCA_MEM_REG_KEY, key); + attr = fill_attr_out_obj(cmd, HNS_IB_ATTR_DCA_MEM_REG_HANDLE); + + ret = execute_ioctl(&ctx->ibv_ctx.context, cmd); + if (!ret) + *handle = read_attr_obj(HNS_IB_ATTR_DCA_MEM_REG_HANDLE, attr);
Success oriented flow everywhere please Jason
On 2021/2/10 3:44, Jason Gunthorpe wrote:
On Sun, Feb 07, 2021 at 11:12:51AM +0800, Weihang Li wrote:
+static int register_dca_mem(struct hns_roce_context *ctx, uint64_t key, + void *addr, uint32_t size, uint32_t *handle) +{ + struct ib_uverbs_attr *attr; + int ret; + + DECLARE_COMMAND_BUFFER(cmd, HNS_IB_OBJECT_DCA_MEM, + HNS_IB_METHOD_DCA_MEM_REG, 4); + fill_attr_in_uint32(cmd, HNS_IB_ATTR_DCA_MEM_REG_LEN, size); + fill_attr_in_uint64(cmd, HNS_IB_ATTR_DCA_MEM_REG_ADDR, (intptr_t)addr);
This should use ioctl_ptr_to_u64(), the place this was copied from should also be fixed
OK, I will fix it. The kernel part has been defined as u64, so it doesn't need to be modified.
+ fill_attr_in_uint64(cmd, HNS_IB_ATTR_DCA_MEM_REG_KEY, key); + attr = fill_attr_out_obj(cmd, HNS_IB_ATTR_DCA_MEM_REG_HANDLE); + + ret = execute_ioctl(&ctx->ibv_ctx.context, cmd); + if (!ret) + *handle = read_attr_obj(HNS_IB_ATTR_DCA_MEM_REG_HANDLE, attr);
Success oriented flow everywhere please
Jason
OK, thank you. Weihang
From: Xi Wang <wangxi11@huawei.com> The QP's WQE buffer may be detached after QP is modified or CQE is polled, and the state of DCA mem object may be changed as clean for no QP is using it. So shrink the clean DCA mem from the memory pool and destroy the DCA mem's buffer to reduce the memory consumption. Signed-off-by: Xi Wang <wangxi11@huawei.com> Signed-off-by: Weihang Li <liweihang@huawei.com> --- providers/hns/hns_roce_u.h | 2 + providers/hns/hns_roce_u_buf.c | 96 ++++++++++++++++++++++++++++++++++++++++ providers/hns/hns_roce_u_hw_v2.c | 7 +++ 3 files changed, 105 insertions(+) diff --git a/providers/hns/hns_roce_u.h b/providers/hns/hns_roce_u.h index 619b060..7dc4a1e 100644 --- a/providers/hns/hns_roce_u.h +++ b/providers/hns/hns_roce_u.h @@ -149,6 +149,7 @@ struct hns_roce_dca_ctx { struct list_head mem_list; pthread_spinlock_t lock; uint64_t max_size; + uint64_t min_size; uint64_t curr_size; int mem_cnt; unsigned int unit_size; @@ -410,6 +411,7 @@ void hns_roce_free_buf(struct hns_roce_buf *buf); void hns_roce_free_qp_buf(struct hns_roce_qp *qp, struct hns_roce_context *ctx); +void hns_roce_shrink_dca_mem(struct hns_roce_context *ctx); void hns_roce_cleanup_dca_mem(struct hns_roce_context *ctx); int hns_roce_add_dca_mem(struct hns_roce_context *ctx, uint32_t size); diff --git a/providers/hns/hns_roce_u_buf.c b/providers/hns/hns_roce_u_buf.c index 424f916..8f3d34a 100644 --- a/providers/hns/hns_roce_u_buf.c +++ b/providers/hns/hns_roce_u_buf.c @@ -101,6 +101,20 @@ static inline uint64_t dca_mem_to_key(struct hns_roce_dca_mem *dca_mem) return (uintptr_t)dca_mem; } +static struct hns_roce_dca_mem *key_to_dca_mem(struct hns_roce_dca_ctx *ctx, + uint64_t key) +{ + struct hns_roce_dca_mem *mem; + struct hns_roce_dca_mem *tmp; + + list_for_each_safe(&ctx->mem_list, mem, tmp, entry) { + if (dca_mem_to_key(mem) == key) + return mem; + } + + return NULL; +} + static inline void *dca_mem_addr(struct hns_roce_dca_mem *dca_mem, int offset) { return dca_mem->buf.buf + offset; @@ -144,6 +158,25 @@ void hns_roce_cleanup_dca_mem(struct hns_roce_context *ctx) deregister_dca_mem(ctx, mem->handle); } +struct hns_dca_mem_shrink_resp { + uint32_t free_mems; + uint64_t free_key; +}; + +static int shrink_dca_mem(struct hns_roce_context *ctx, uint32_t handle, + uint64_t size, struct hns_dca_mem_shrink_resp *resp) +{ + DECLARE_COMMAND_BUFFER(cmd, HNS_IB_OBJECT_DCA_MEM, + HNS_IB_METHOD_DCA_MEM_SHRINK, 4); + fill_attr_in_obj(cmd, HNS_IB_ATTR_DCA_MEM_SHRINK_HANDLE, handle); + fill_attr_in_uint64(cmd, HNS_IB_ATTR_DCA_MEM_SHRINK_RESERVED_SIZE, size); + fill_attr_out(cmd, HNS_IB_ATTR_DCA_MEM_SHRINK_OUT_FREE_KEY, + &resp->free_key, sizeof(resp->free_key)); + fill_attr_out(cmd, HNS_IB_ATTR_DCA_MEM_SHRINK_OUT_FREE_MEMS, + &resp->free_mems, sizeof(resp->free_mems)); + + return execute_ioctl(&ctx->ibv_ctx.context, cmd); +} static bool add_dca_mem_enabled(struct hns_roce_dca_ctx *ctx, uint32_t alloc_size) { @@ -194,3 +227,66 @@ int hns_roce_add_dca_mem(struct hns_roce_context *ctx, uint32_t size) return 0; } + +static bool shrink_dca_mem_enabled(struct hns_roce_dca_ctx *ctx) +{ + bool enable; + + pthread_spin_lock(&ctx->lock); + enable = ctx->mem_cnt > 0 && ctx->min_size < ctx->max_size; + pthread_spin_unlock(&ctx->lock); + + return enable; +} + +void hns_roce_shrink_dca_mem(struct hns_roce_context *ctx) +{ + struct hns_roce_dca_ctx *dca_ctx = &ctx->dca_ctx; + struct hns_dca_mem_shrink_resp resp = {}; + struct hns_roce_dca_mem *mem; + int dca_mem_cnt; + uint32_t handle; + int ret; + + pthread_spin_lock(&dca_ctx->lock); + dca_mem_cnt = ctx->dca_ctx.mem_cnt; + pthread_spin_unlock(&dca_ctx->lock); + while (dca_mem_cnt > 0 && shrink_dca_mem_enabled(dca_ctx)) { + resp.free_mems = 0; + /* Step 1: Use any DCA mem uobject to shrink pool */ + pthread_spin_lock(&dca_ctx->lock); + mem = list_tail(&dca_ctx->mem_list, + struct hns_roce_dca_mem, entry); + handle = mem ? mem->handle : 0; + pthread_spin_unlock(&dca_ctx->lock); + if (!mem) + break; + + ret = shrink_dca_mem(ctx, handle, dca_ctx->min_size, &resp); + if (ret || likely(resp.free_mems < 1)) + break; + + /* Step 2: Remove shrunk DCA mem node from pool */ + pthread_spin_lock(&dca_ctx->lock); + mem = key_to_dca_mem(dca_ctx, resp.free_key); + if (mem) { + list_del(&mem->entry); + dca_ctx->mem_cnt--; + dca_ctx->curr_size -= mem->buf.length; + } + + handle = mem ? mem->handle : 0; + pthread_spin_unlock(&dca_ctx->lock); + if (!mem) + break; + + /* Step 3: Destroy DCA mem uobject */ + deregister_dca_mem(ctx, handle); + free_dca_mem(ctx, mem); + /* No any free memory after deregister 1 DCA mem */ + if (resp.free_mems <= 1) + break; + + dca_mem_cnt--; + } +} diff --git a/providers/hns/hns_roce_u_hw_v2.c b/providers/hns/hns_roce_u_hw_v2.c index c8d273f..93f3546 100644 --- a/providers/hns/hns_roce_u_hw_v2.c +++ b/providers/hns/hns_roce_u_hw_v2.c @@ -651,6 +651,10 @@ static int hns_roce_u_v2_poll_cq(struct ibv_cq *ibvcq, int ne, pthread_spin_unlock(&cq->lock); + /* Try to shrink the DCA mem */ + if (ctx->dca_ctx.mem_cnt > 0) + hns_roce_shrink_dca_mem(ctx); + return err == V2_CQ_POLL_ERR ? err : npolled; } @@ -1478,6 +1482,9 @@ static int hns_roce_u_v2_destroy_qp(struct ibv_qp *ibqp) free(qp); + if (ctx->dca_ctx.mem_cnt > 0) + hns_roce_shrink_dca_mem(ctx); + return ret; } -- 2.8.1
From: Xi Wang <wangxi11@huawei.com> If a uQP works in DCA mode, the WQE's buffer will be split as many blocks and be stored into a list. The blocks are allocated from the DCA's memory pool before posting WRs and are dropped when the QP's CI is equal to PI after polling CQ. Signed-off-by: Xi Wang <wangxi11@huawei.com> Signed-off-by: Weihang Li <liweihang@huawei.com> --- providers/hns/hns_roce_u.h | 24 +++++- providers/hns/hns_roce_u_buf.c | 156 ++++++++++++++++++++++++++++++++++++++- providers/hns/hns_roce_u_hw_v2.c | 131 +++++++++++++++++++++++++++++--- providers/hns/hns_roce_u_hw_v2.h | 7 ++ providers/hns/hns_roce_u_verbs.c | 32 ++++++-- 5 files changed, 332 insertions(+), 18 deletions(-) diff --git a/providers/hns/hns_roce_u.h b/providers/hns/hns_roce_u.h index 7dc4a1e..ae72709 100644 --- a/providers/hns/hns_roce_u.h +++ b/providers/hns/hns_roce_u.h @@ -119,6 +119,12 @@ struct hns_roce_device { int hw_version; }; +struct hns_roce_buf_list { + void **bufs; + unsigned int max_cnt; + unsigned int shift; +}; + struct hns_roce_buf { void *buf; unsigned int length; @@ -268,9 +274,20 @@ struct hns_roce_rinl_buf { unsigned int wqe_cnt; }; +struct hns_roce_dca_attach_attr { + uint32_t sq_offset; + uint32_t sge_offset; + uint32_t rq_offset; +}; + +struct hns_roce_dca_detach_attr { + uint32_t sq_index; +}; + struct hns_roce_qp { struct verbs_qp verbs_qp; struct hns_roce_buf buf; + struct hns_roce_buf_list page_list; int max_inline_data; int buf_size; unsigned int sq_signal_bits; @@ -316,6 +333,7 @@ struct hns_roce_u_hw { * minimum page size. */ #define hr_hw_page_align(x) align(x, HNS_HW_PAGE_SIZE) +#define hr_hw_page_count(x) (hr_hw_page_align(x) / HNS_HW_PAGE_SIZE) static inline unsigned int to_hr_hem_entries_size(int count, int buf_shift) { @@ -411,9 +429,13 @@ void hns_roce_free_buf(struct hns_roce_buf *buf); void hns_roce_free_qp_buf(struct hns_roce_qp *qp, struct hns_roce_context *ctx); +int hns_roce_attach_dca_mem(struct hns_roce_context *ctx, uint32_t handle, + struct hns_roce_dca_attach_attr *attr, + uint32_t size, struct hns_roce_buf_list *buf_list); +void hns_roce_detach_dca_mem(struct hns_roce_context *ctx, uint32_t handle, + struct hns_roce_dca_detach_attr *attr); void hns_roce_shrink_dca_mem(struct hns_roce_context *ctx); void hns_roce_cleanup_dca_mem(struct hns_roce_context *ctx); -int hns_roce_add_dca_mem(struct hns_roce_context *ctx, uint32_t size); void hns_roce_init_qp_indices(struct hns_roce_qp *qp); diff --git a/providers/hns/hns_roce_u_buf.c b/providers/hns/hns_roce_u_buf.c index 8f3d34a..0bcb39b 100644 --- a/providers/hns/hns_roce_u_buf.c +++ b/providers/hns/hns_roce_u_buf.c @@ -177,6 +177,66 @@ static int shrink_dca_mem(struct hns_roce_context *ctx, uint32_t handle, return execute_ioctl(&ctx->ibv_ctx.context, cmd); } + +struct hns_dca_mem_query_resp { + uint64_t key; + uint32_t offset; + uint32_t page_count; +}; + +static int query_dca_mem(struct hns_roce_context *ctx, uint32_t handle, + uint32_t index, struct hns_dca_mem_query_resp *resp) +{ + DECLARE_COMMAND_BUFFER(cmd, HNS_IB_OBJECT_DCA_MEM, + HNS_IB_METHOD_DCA_MEM_QUERY, 5); + fill_attr_in_obj(cmd, HNS_IB_ATTR_DCA_MEM_QUERY_HANDLE, handle); + fill_attr_in_uint32(cmd, HNS_IB_ATTR_DCA_MEM_QUERY_PAGE_INDEX, index); + fill_attr_out(cmd, HNS_IB_ATTR_DCA_MEM_QUERY_OUT_KEY, + &resp->key, sizeof(resp->key)); + fill_attr_out(cmd, HNS_IB_ATTR_DCA_MEM_QUERY_OUT_OFFSET, + &resp->offset, sizeof(resp->offset)); + fill_attr_out(cmd, HNS_IB_ATTR_DCA_MEM_QUERY_OUT_PAGE_COUNT, + &resp->page_count, sizeof(resp->page_count)); + return execute_ioctl(&ctx->ibv_ctx.context, cmd); +} + +static int detach_dca_mem(struct hns_roce_context *ctx, uint32_t handle, + struct hns_roce_dca_detach_attr *attr) +{ + DECLARE_COMMAND_BUFFER(cmd, HNS_IB_OBJECT_DCA_MEM, + HNS_IB_METHOD_DCA_MEM_DETACH, 4); + fill_attr_in_obj(cmd, HNS_IB_ATTR_DCA_MEM_DETACH_HANDLE, handle); + fill_attr_in_uint32(cmd, HNS_IB_ATTR_DCA_MEM_DETACH_SQ_INDEX, + attr->sq_index); + return execute_ioctl(&ctx->ibv_ctx.context, cmd); +} + +struct hns_dca_mem_attach_resp { +#define HNS_DCA_ATTACH_OUT_FLAGS_NEW_BUFFER BIT(0) + uint32_t alloc_flags; + uint32_t alloc_pages; +}; + +static int attach_dca_mem(struct hns_roce_context *ctx, uint32_t handle, + struct hns_roce_dca_attach_attr *attr, + struct hns_dca_mem_attach_resp *resp) +{ + DECLARE_COMMAND_BUFFER(cmd, HNS_IB_OBJECT_DCA_MEM, + HNS_IB_METHOD_DCA_MEM_ATTACH, 6); + fill_attr_in_obj(cmd, HNS_IB_ATTR_DCA_MEM_ATTACH_HANDLE, handle); + fill_attr_in_uint32(cmd, HNS_IB_ATTR_DCA_MEM_ATTACH_SQ_OFFSET, + attr->sq_offset); + fill_attr_in_uint32(cmd, HNS_IB_ATTR_DCA_MEM_ATTACH_SGE_OFFSET, + attr->sge_offset); + fill_attr_in_uint32(cmd, HNS_IB_ATTR_DCA_MEM_ATTACH_RQ_OFFSET, + attr->rq_offset); + fill_attr_out(cmd, HNS_IB_ATTR_DCA_MEM_ATTACH_OUT_ALLOC_FLAGS, + &resp->alloc_flags, sizeof(resp->alloc_flags)); + fill_attr_out(cmd, HNS_IB_ATTR_DCA_MEM_ATTACH_OUT_ALLOC_PAGES, + &resp->alloc_pages, sizeof(resp->alloc_pages)); + return execute_ioctl(&ctx->ibv_ctx.context, cmd); +} + static bool add_dca_mem_enabled(struct hns_roce_dca_ctx *ctx, uint32_t alloc_size) { @@ -196,7 +256,7 @@ static bool add_dca_mem_enabled(struct hns_roce_dca_ctx *ctx, return enable; } -int hns_roce_add_dca_mem(struct hns_roce_context *ctx, uint32_t size) +static int add_dca_mem(struct hns_roce_context *ctx, uint32_t size) { struct hns_roce_dca_ctx *dca_ctx = &ctx->dca_ctx; struct hns_roce_dca_mem *mem; @@ -290,3 +350,97 @@ void hns_roce_shrink_dca_mem(struct hns_roce_context *ctx) dca_mem_cnt--; } } + +static void config_page_list(void *addr, struct hns_roce_buf_list *page_list, + uint32_t page_index, int page_count) +{ + void **bufs = &page_list->bufs[page_index]; + int page_size = 1 << page_list->shift; + int i; + + for (i = 0; i < page_count; i++) { + bufs[i] = addr; + addr += page_size; + } +} + +static int setup_dca_buf_list(struct hns_roce_context *ctx, uint32_t handle, + struct hns_roce_buf_list *buf_list, + uint32_t page_count) +{ + struct hns_roce_dca_ctx *dca_ctx = &ctx->dca_ctx; + struct hns_dca_mem_query_resp resp = {}; + struct hns_roce_dca_mem *mem; + uint32_t idx = 0; + int ret; + + while (idx < page_count && idx < buf_list->max_cnt) { + resp.page_count = 0; + ret = query_dca_mem(ctx, handle, idx, &resp); + if (ret) + return -ENOMEM; + + if (resp.page_count < 1) + break; + + pthread_spin_lock(&dca_ctx->lock); + mem = key_to_dca_mem(dca_ctx, resp.key); + if (mem && resp.offset < mem->buf.length) { + config_page_list(dca_mem_addr(mem, resp.offset), + buf_list, idx, resp.page_count); + } else { + pthread_spin_unlock(&dca_ctx->lock); + break; + } + + pthread_spin_unlock(&dca_ctx->lock); + + idx += resp.page_count; + } + + return (idx >= page_count) ? 0 : ENOMEM; +} + +#define DCA_EXPAND_MEM_TRY_TIMES 3 +int hns_roce_attach_dca_mem(struct hns_roce_context *ctx, uint32_t handle, + struct hns_roce_dca_attach_attr *attr, + uint32_t size, struct hns_roce_buf_list *buf_list) +{ + uint32_t buf_pages = size >> buf_list->shift; + struct hns_dca_mem_attach_resp resp = {}; + bool is_new_buf = true; + int try_times = 0; + int ret; + + do { + resp.alloc_pages = 0; + ret = attach_dca_mem(ctx, handle, attr, &resp); + if (ret) + break; + + if (resp.alloc_pages >= buf_pages) { + is_new_buf = !!(resp.alloc_flags & + HNS_DCA_ATTACH_OUT_FLAGS_NEW_BUFFER); + break; + } + + ret = add_dca_mem(ctx, size); + if (ret) + break; + } while (try_times++ < DCA_EXPAND_MEM_TRY_TIMES); + + if (ret || resp.alloc_pages < buf_pages) + return -ENOMEM; + + /* DCA config not changed */ + if (!is_new_buf && buf_list->bufs[0]) + return 0; + + return setup_dca_buf_list(ctx, handle, buf_list, buf_pages); +} + +void hns_roce_detach_dca_mem(struct hns_roce_context *ctx, uint32_t handle, + struct hns_roce_dca_detach_attr *attr) +{ + detach_dca_mem(ctx, handle, attr); +} diff --git a/providers/hns/hns_roce_u_hw_v2.c b/providers/hns/hns_roce_u_hw_v2.c index 93f3546..fb065d1 100644 --- a/providers/hns/hns_roce_u_hw_v2.c +++ b/providers/hns/hns_roce_u_hw_v2.c @@ -230,19 +230,28 @@ static struct hns_roce_v2_cqe *next_cqe_sw_v2(struct hns_roce_cq *cq) return get_sw_cqe_v2(cq, cq->cons_index); } +static inline void *get_wqe(struct hns_roce_qp *qp, unsigned int offset) +{ + if (qp->page_list.bufs) + return qp->page_list.bufs[offset >> qp->page_list.shift] + + (offset & ((1 << qp->page_list.shift) - 1)); + else + return qp->buf.buf + offset; +} + static void *get_recv_wqe_v2(struct hns_roce_qp *qp, unsigned int n) { - return qp->buf.buf + qp->rq.offset + (n << qp->rq.wqe_shift); + return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift)); } static void *get_send_wqe(struct hns_roce_qp *qp, unsigned int n) { - return qp->buf.buf + qp->sq.offset + (n << qp->sq.wqe_shift); + return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift)); } static void *get_send_sge_ex(struct hns_roce_qp *qp, unsigned int n) { - return qp->buf.buf + qp->ex_sge.offset + (n << qp->ex_sge.sge_shift); + return get_wqe(qp, qp->ex_sge.offset + (n << qp->ex_sge.sge_shift)); } static void *get_srq_wqe(struct hns_roce_srq *srq, int n) @@ -507,6 +516,62 @@ static int hns_roce_handle_recv_inl_wqe(struct hns_roce_v2_cqe *cqe, return V2_CQ_OK; } +static inline bool check_qp_dca_enable(struct hns_roce_qp *qp) +{ + return !!qp->page_list.bufs && (qp->flags & HNS_ROCE_QP_CAP_DCA); +} + +static int dca_attach_qp_buf(struct hns_roce_context *ctx, + struct hns_roce_qp *qp) +{ + struct hns_roce_dca_attach_attr attr = {}; + uint32_t idx; + + pthread_spin_lock(&qp->sq.lock); + pthread_spin_lock(&qp->rq.lock); + + if (qp->sq.wqe_cnt > 0) { + idx = qp->sq.head & (qp->sq.wqe_cnt - 1); + attr.sq_offset = idx << qp->sq.wqe_shift; + } + + if (qp->ex_sge.sge_cnt > 0) { + idx = qp->next_sge & (qp->ex_sge.sge_cnt - 1); + attr.sge_offset = idx << qp->ex_sge.sge_shift; + } + + if (qp->rq.wqe_cnt > 0) { + idx = qp->rq.head & (qp->rq.wqe_cnt - 1); + attr.rq_offset = idx << qp->rq.wqe_shift; + } + + pthread_spin_unlock(&qp->rq.lock); + pthread_spin_unlock(&qp->sq.lock); + + return hns_roce_attach_dca_mem(ctx, qp->verbs_qp.qp.handle, &attr, + qp->buf_size, &qp->page_list); +} + +static void dca_detach_qp_buf(struct hns_roce_context *ctx, + struct hns_roce_qp *qp) +{ + struct hns_roce_dca_detach_attr attr; + bool is_empty; + + pthread_spin_lock(&qp->sq.lock); + pthread_spin_lock(&qp->rq.lock); + + is_empty = qp->sq.head == qp->sq.tail && qp->rq.head == qp->rq.tail; + if (is_empty && qp->sq.wqe_cnt > 0) + attr.sq_index = qp->sq.head & (qp->sq.wqe_cnt - 1); + + pthread_spin_unlock(&qp->rq.lock); + pthread_spin_unlock(&qp->sq.lock); + + if (is_empty) + hns_roce_detach_dca_mem(ctx, qp->verbs_qp.qp.handle, &attr); +} + static int hns_roce_v2_poll_one(struct hns_roce_cq *cq, struct hns_roce_qp **cur_qp, struct ibv_wc *wc) { @@ -636,6 +701,9 @@ static int hns_roce_u_v2_poll_cq(struct ibv_cq *ibvcq, int ne, for (npolled = 0; npolled < ne; ++npolled) { err = hns_roce_v2_poll_one(cq, &qp, wc + npolled); + if (qp && check_qp_dca_enable(qp)) + dca_detach_qp_buf(ctx, qp); + if (err != V2_CQ_OK) break; } @@ -1003,6 +1071,16 @@ static int set_rc_inl(struct hns_roce_qp *qp, const struct ibv_send_wr *wr, return 0; } +static inline void fill_rc_dca_fields(uint32_t qp_num, + struct hns_roce_rc_sq_wqe *wqe) +{ + roce_set_field(wqe->byte_4, RC_SQ_WQE_BYTE_4_SQPN_L_M, + RC_SQ_WQE_BYTE_4_SQPN_L_S, qp_num); + roce_set_field(wqe->byte_4, RC_SQ_WQE_BYTE_4_SQPN_H_M, + RC_SQ_WQE_BYTE_4_SQPN_H_S, + qp_num >> RC_SQ_WQE_BYTE_4_SQPN_L_W); +} + static void set_bind_mw_seg(struct hns_roce_rc_sq_wqe *wqe, const struct ibv_send_wr *wr) { @@ -1120,6 +1198,9 @@ static int set_rc_wqe(void *wqe, struct hns_roce_qp *qp, struct ibv_send_wr *wr, return ret; wqe_valid: + if (check_qp_dca_enable(qp)) + fill_rc_dca_fields(qp->verbs_qp.qp.qp_num, rc_sq_wqe); + /* * The pipeline can sequentially post all valid WQEs into WQ buffer, * including new WQEs waiting for the doorbell to update the PI again. @@ -1135,6 +1216,21 @@ wqe_valid: return 0; } +static int check_qp_send(struct hns_roce_qp *qp, struct hns_roce_context *ctx) +{ + struct ibv_qp *ibvqp = &qp->verbs_qp.qp; + + /* check that state is OK to post receive */ + if (ibvqp->state == IBV_QPS_RESET || ibvqp->state == IBV_QPS_INIT || + ibvqp->state == IBV_QPS_RTR) + return -EINVAL; + + if (check_qp_dca_enable(qp)) + return dca_attach_qp_buf(ctx, qp); + + return 0; +} + int hns_roce_u_v2_post_send(struct ibv_qp *ibvqp, struct ibv_send_wr *wr, struct ibv_send_wr **bad_wr) { @@ -1143,14 +1239,14 @@ int hns_roce_u_v2_post_send(struct ibv_qp *ibvqp, struct ibv_send_wr *wr, struct hns_roce_sge_info sge_info = {}; unsigned int wqe_idx, nreq; struct ibv_qp_attr attr; - int ret = 0; void *wqe; + int ret; /* check that state is OK to post send */ - if (ibvqp->state == IBV_QPS_RESET || ibvqp->state == IBV_QPS_INIT || - ibvqp->state == IBV_QPS_RTR) { + ret = check_qp_send(qp, ctx); + if (unlikely(ret)) { *bad_wr = wr; - return EINVAL; + return ret; } pthread_spin_lock(&qp->sq.lock); @@ -1216,6 +1312,20 @@ out: return ret; } +static int check_qp_recv(struct hns_roce_qp *qp, struct hns_roce_context *ctx) +{ + struct ibv_qp *ibvqp = &qp->verbs_qp.qp; + + /* check that state is OK to post receive */ + if (ibvqp->state == IBV_QPS_RESET) + return -EINVAL; + + if (check_qp_dca_enable(qp)) + return dca_attach_qp_buf(ctx, qp); + + return 0; +} + static void fill_rq_wqe(struct hns_roce_qp *qp, struct ibv_recv_wr *wr, unsigned int wqe_idx) { @@ -1253,12 +1363,13 @@ static int hns_roce_u_v2_post_recv(struct ibv_qp *ibvqp, struct ibv_recv_wr *wr, struct hns_roce_qp *qp = to_hr_qp(ibvqp); unsigned int wqe_idx, nreq; struct ibv_qp_attr attr; - int ret = 0; + int ret; /* check that state is OK to post receive */ - if (ibvqp->state == IBV_QPS_RESET) { + ret = check_qp_recv(qp, ctx); + if (ret) { *bad_wr = wr; - return EINVAL; + return ret; } pthread_spin_lock(&qp->rq.lock); diff --git a/providers/hns/hns_roce_u_hw_v2.h b/providers/hns/hns_roce_u_hw_v2.h index f5e6402..00716c6 100644 --- a/providers/hns/hns_roce_u_hw_v2.h +++ b/providers/hns/hns_roce_u_hw_v2.h @@ -250,6 +250,13 @@ struct hns_roce_rc_sq_wqe { #define RC_SQ_WQE_BYTE_4_RDMA_WRITE_S 22 +#define RC_SQ_WQE_BYTE_4_SQPN_L_W 2 +#define RC_SQ_WQE_BYTE_4_SQPN_L_S 5 +#define RC_SQ_WQE_BYTE_4_SQPN_L_M GENMASK(5, 2) + +#define RC_SQ_WQE_BYTE_4_SQPN_H_S 13 +#define RC_SQ_WQE_BYTE_4_SQPN_H_M GENMASK(18, 13) + #define RC_SQ_WQE_BYTE_16_XRC_SRQN_S 0 #define RC_SQ_WQE_BYTE_16_XRC_SRQN_M \ (((1UL << 24) - 1) << RC_SQ_WQE_BYTE_16_XRC_SRQN_S) diff --git a/providers/hns/hns_roce_u_verbs.c b/providers/hns/hns_roce_u_verbs.c index f3d6fc1..6ec8b12 100644 --- a/providers/hns/hns_roce_u_verbs.c +++ b/providers/hns/hns_roce_u_verbs.c @@ -688,6 +688,14 @@ static int calc_qp_buff_size(struct hns_roce_device *hr_dev, return 0; } +static inline bool check_qp_support_dca(bool pool_en, enum ibv_qp_type qp_type) +{ + if (pool_en && (qp_type == IBV_QPT_RC || qp_type == IBV_QPT_XRC_SEND)) + return true; + + return false; +} + static void qp_free_wqe(struct hns_roce_qp *qp) { qp_free_recv_inl_buf(qp); @@ -699,8 +707,8 @@ static void qp_free_wqe(struct hns_roce_qp *qp) hns_roce_free_buf(&qp->buf); } -static int qp_alloc_wqe(struct ibv_qp_cap *cap, struct hns_roce_qp *qp, - struct hns_roce_context *ctx) +static int qp_alloc_wqe(struct ibv_qp_init_attr_ex *attr, + struct hns_roce_qp *qp, struct hns_roce_context *ctx) { struct hns_roce_device *hr_dev = to_hr_dev(ctx->ibv_ctx.context.device); @@ -718,12 +726,24 @@ static int qp_alloc_wqe(struct ibv_qp_cap *cap, struct hns_roce_qp *qp, } if (qp->rq_rinl_buf.wqe_cnt) { - if (qp_alloc_recv_inl_buf(cap, qp)) + if (qp_alloc_recv_inl_buf(&attr->cap, qp)) goto err_alloc; } - if (hns_roce_alloc_buf(&qp->buf, qp->buf_size, HNS_HW_PAGE_SIZE)) - goto err_alloc; + if (check_qp_support_dca(ctx->dca_ctx.max_size != 0, attr->qp_type)) { + /* when DCA is enabled, use a buffer list to store page addr */ + qp->buf.buf = NULL; + qp->page_list.max_cnt = hr_hw_page_count(qp->buf_size); + qp->page_list.shift = HNS_HW_PAGE_SHIFT; + qp->page_list.bufs = calloc(qp->page_list.max_cnt, + sizeof(void *)); + if (!qp->page_list.bufs) + goto err_alloc; + } else { + if (hns_roce_alloc_buf(&qp->buf, qp->buf_size, + HNS_HW_PAGE_SIZE)) + goto err_alloc; + } return 0; @@ -928,7 +948,7 @@ static int hns_roce_alloc_qp_buf(struct ibv_qp_init_attr_ex *attr, pthread_spin_init(&qp->rq.lock, PTHREAD_PROCESS_PRIVATE)) return -ENOMEM; - ret = qp_alloc_wqe(&attr->cap, qp, ctx); + ret = qp_alloc_wqe(attr, qp, ctx); if (ret) return ret; -- 2.8.1
From: Xi Wang <wangxi11@huawei.com> Add a group of environment variables for configuring DCA memory pool which is as follows: HNS_DCA_MAX_SIZE - the max size when expand the memory pool. HNS_DCA_MIN_SIZE - the reverved size when shrink the memory pool. HNS_DCA_UNIT_SIZE - the increase unit size when expand the memory pool. Also append 'IBV_QP_CREATE_DYNAMIC_CONTEXT_ATTACH' for creating a DCA QP. Signed-off-by: Xi Wang <wangxi11@huawei.com> Signed-off-by: Weihang Li <liweihang@huawei.com> --- libibverbs/cmd_qp.c | 3 ++- libibverbs/verbs.h | 1 + providers/hns/hns_roce_u.c | 52 ++++++++++++++++++++++++++++++++++++++-- providers/hns/hns_roce_u_verbs.c | 18 ++++++++++---- 4 files changed, 67 insertions(+), 7 deletions(-) diff --git a/libibverbs/cmd_qp.c b/libibverbs/cmd_qp.c index 056f397..f9899ad 100644 --- a/libibverbs/cmd_qp.c +++ b/libibverbs/cmd_qp.c @@ -38,7 +38,8 @@ enum { IBV_QP_CREATE_SCATTER_FCS | IBV_QP_CREATE_CVLAN_STRIPPING | IBV_QP_CREATE_SOURCE_QPN | - IBV_QP_CREATE_PCI_WRITE_END_PADDING + IBV_QP_CREATE_PCI_WRITE_END_PADDING | + IBV_QP_CREATE_DYNAMIC_CONTEXT_ATTACH }; diff --git a/libibverbs/verbs.h b/libibverbs/verbs.h index 656b0f9..db37dce 100644 --- a/libibverbs/verbs.h +++ b/libibverbs/verbs.h @@ -918,6 +918,7 @@ enum ibv_qp_create_flags { IBV_QP_CREATE_CVLAN_STRIPPING = 1 << 9, IBV_QP_CREATE_SOURCE_QPN = 1 << 10, IBV_QP_CREATE_PCI_WRITE_END_PADDING = 1 << 11, + IBV_QP_CREATE_DYNAMIC_CONTEXT_ATTACH = 1 << 13, }; enum ibv_qp_create_send_ops_flags { diff --git a/providers/hns/hns_roce_u.c b/providers/hns/hns_roce_u.c index 28b1130..3ba99ed 100644 --- a/providers/hns/hns_roce_u.c +++ b/providers/hns/hns_roce_u.c @@ -92,6 +92,10 @@ static const struct verbs_context_ops hns_common_ops = { static int init_dca_context(struct hns_roce_context *ctx, int page_size) { struct hns_roce_dca_ctx *dca_ctx = &ctx->dca_ctx; + int unit_size = 0; + long max_size = 0; + long min_size; + char *env; int ret; if (!(ctx->cap_flags & HNS_ROCE_CAP_FLAG_DCA_MODE)) @@ -102,8 +106,52 @@ static int init_dca_context(struct hns_roce_context *ctx, int page_size) if (ret) return ret; - dca_ctx->unit_size = page_size * HNS_DCA_DEFAULT_UNIT_PAGES; - dca_ctx->max_size = HNS_DCA_MAX_MEM_SIZE; + env = getenv("HNS_DCA_UNIT_SIZE"); + if (env) { + unit_size = atoi(env); + /* Disable DCA only for this process */ + if (unit_size == 0) + return 0; + } + + if (unit_size < 1) + unit_size = page_size * HNS_DCA_DEFAULT_UNIT_PAGES; + + unit_size = align(unit_size, page_size); + + /* + * not set OR 0: Unlimited memory pool increase. + * others: Maximum memory pool size to be increased. + */ + env = getenv("HNS_DCA_MAX_SIZE"); + if (env) + max_size = atol(env); + + if (max_size == 0) + max_size = HNS_DCA_MAX_MEM_SIZE; + else + max_size = DIV_ROUND_UP(max_size, unit_size); + + /* + * not set: The memory pool cannot be reduced. + * others: The size of free memory in the pool cannot exceed this value. + * 0: Always reduce the free memory in the pool. + */ + env = getenv("HNS_DCA_MIN_SIZE"); + if (env) { + min_size = atol(env); + if (min_size > 0) + min_size = DIV_ROUND_UP(min_size, unit_size); + else + min_size = 0; + } else { + min_size = HNS_DCA_MAX_MEM_SIZE; + } + + dca_ctx->unit_size = unit_size; + dca_ctx->max_size = max_size; + dca_ctx->min_size = min_size; + dca_ctx->mem_cnt = 0; return 0; diff --git a/providers/hns/hns_roce_u_verbs.c b/providers/hns/hns_roce_u_verbs.c index 6ec8b12..0697328 100644 --- a/providers/hns/hns_roce_u_verbs.c +++ b/providers/hns/hns_roce_u_verbs.c @@ -542,7 +542,7 @@ int hns_roce_u_destroy_srq(struct ibv_srq *srq) } enum { - CREATE_QP_SUP_CREATE_FLAGS = 0, + CREATE_QP_SUP_CREATE_FLAGS = IBV_QP_CREATE_DYNAMIC_CONTEXT_ATTACH, }; enum { @@ -688,10 +688,11 @@ static int calc_qp_buff_size(struct hns_roce_device *hr_dev, return 0; } -static inline bool check_qp_support_dca(bool pool_en, enum ibv_qp_type qp_type) +static inline bool check_qp_support_dca(bool pool_en, enum ibv_qp_type qp_type, + uint32_t create_flags) { if (pool_en && (qp_type == IBV_QPT_RC || qp_type == IBV_QPT_XRC_SEND)) - return true; + return !!(create_flags & IBV_QP_CREATE_DYNAMIC_CONTEXT_ATTACH); return false; } @@ -730,7 +731,8 @@ static int qp_alloc_wqe(struct ibv_qp_init_attr_ex *attr, goto err_alloc; } - if (check_qp_support_dca(ctx->dca_ctx.max_size != 0, attr->qp_type)) { + if (check_qp_support_dca(ctx->dca_ctx.max_size != 0, + attr->qp_type, attr->create_flags)) { /* when DCA is enabled, use a buffer list to store page addr */ qp->buf.buf = NULL; qp->page_list.max_cnt = hr_hw_page_count(qp->buf_size); @@ -901,6 +903,14 @@ static int qp_exec_create_cmd(struct ibv_qp_init_attr_ex *attr, struct hns_roce_create_qp_ex cmd_ex = {}; int ret; + /* + * When handling the command in kernel space, the user QP enable + * the DCA mode by checking whether the cmd.buf_addr is NULL but + * not the attr->create_flags has the DCA enable bit, so clear + * this bit before command is ready to run. + */ + attr->create_flags &= ~IBV_QP_CREATE_DYNAMIC_CONTEXT_ATTACH; + cmd_ex.sdb_addr = (uintptr_t)qp->sdb; cmd_ex.db_addr = (uintptr_t)qp->rdb; cmd_ex.buf_addr = (uintptr_t)qp->buf.buf; -- 2.8.1
On Sun, Feb 07, 2021 at 11:12:49AM +0800, Weihang Li wrote:
The HIP09 introduces the DCA(Dynamic Context Attachment) feature which supports many RC QPs to share the WQE buffer in a memory pool. If a QP enables DCA feature, the WQE's buffer will not be allocated when creating but when the users start to post WRs. This will reduce the memory consumption when there are too many QPs are inactive.
Please note that we didn't find the right way to get user's configuration, so in #4 we still use environment variable to achieve this. We will be appreciated if anyone can provide some sugggestions.
That is definately not going to work.. It should be some dv thing, dv create qp or a dv customization of the parent domain spring to mind Jason
On 2021/2/10 3:39, Jason Gunthorpe wrote:
On Sun, Feb 07, 2021 at 11:12:49AM +0800, Weihang Li wrote:
The HIP09 introduces the DCA(Dynamic Context Attachment) feature which supports many RC QPs to share the WQE buffer in a memory pool. If a QP enables DCA feature, the WQE's buffer will not be allocated when creating but when the users start to post WRs. This will reduce the memory consumption when there are too many QPs are inactive.
Please note that we didn't find the right way to get user's configuration, so in #4 we still use environment variable to achieve this. We will be appreciated if anyone can provide some sugggestions.
That is definately not going to work.. It should be some dv thing, dv create qp or a dv customization of the parent domain spring to mind
Jason
Thank you, we will use the parameter of verbs_open_device() named private_data to allow user pass the DCA configurations into the driver according to the way of mlx5dv_open_device(). Weihang
On Sun, Feb 07, 2021 at 11:12:49AM +0800, Weihang Li wrote:
The HIP09 introduces the DCA(Dynamic Context Attachment) feature which supports many RC QPs to share the WQE buffer in a memory pool. If a QP enables DCA feature, the WQE's buffer will not be allocated when creating but when the users start to post WRs. This will reduce the memory consumption when there are too many QPs are inactive.
One a WQE buffer is allocated it still acts as a normal WQE ring buffer? So this DCA logic is to remap the send queue buffer based on demand for SQEs? How does it interact with the normal max send queue entries reported? Would like to see proper man pages explaining how this all works for rdma-core. Jason
On 2021/2/10 3:54, Jason Gunthorpe wrote:
On Sun, Feb 07, 2021 at 11:12:49AM +0800, Weihang Li wrote:
The HIP09 introduces the DCA(Dynamic Context Attachment) feature which supports many RC QPs to share the WQE buffer in a memory pool. If a QP enables DCA feature, the WQE's buffer will not be allocated when creating but when the users start to post WRs. This will reduce the memory consumption when there are too many QPs are inactive. One a WQE buffer is allocated it still acts as a normal WQE ring buffer? So this DCA logic is to remap the send queue buffer based on demand for SQEs? How does it interact with the normal max send queue entries reported?
Not exactly. If DCA is enabled, we first allocate a memory pool with a default size when opening device. Each time we trying to post WR(s) to a QP, the driver will check if current QP has WQE buffer. If not, the driver will check whether there is enough free memory in the DCA memory pool. If there is, the QP will get WQE buffer from the pool, including SQ buffer in size of max_send_wr, RQ buffer in size of max_recv_wr and extended sge buffer. If there is no enough space for the WQE buffer in the DCA pool, the driver will expand the size of pool and then assign the buffer to the QP. And if the WQE buffer of a QP is not used, this buffer would be recycled and the DCA memory pool would be shrinked.
Would like to see proper man pages explaining how this all works for rdma-core.
Jason
OK, will add man pages for DCA in next version. Thanks Weihang
On Sat, Feb 20, 2021 at 08:40:02AM +0000, liweihang wrote:
On 2021/2/10 3:54, Jason Gunthorpe wrote:
On Sun, Feb 07, 2021 at 11:12:49AM +0800, Weihang Li wrote:
The HIP09 introduces the DCA(Dynamic Context Attachment) feature which supports many RC QPs to share the WQE buffer in a memory pool. If a QP enables DCA feature, the WQE's buffer will not be allocated when creating but when the users start to post WRs. This will reduce the memory consumption when there are too many QPs are inactive. One a WQE buffer is allocated it still acts as a normal WQE ring buffer? So this DCA logic is to remap the send queue buffer based on demand for SQEs? How does it interact with the normal max send queue entries reported?
Not exactly. If DCA is enabled, we first allocate a memory pool with a default size when opening device. Each time we trying to post WR(s) to a QP, the driver will check if current QP has WQE buffer.
If not, the driver will check whether there is enough free memory in the DCA memory pool. If there is, the QP will get WQE buffer from the pool,
Does that mean the QP can have a non-contiguous list of buffers? Ie it isn't just a linear ring of memory? Jason
On 2021/2/23 0:02, Jason Gunthorpe wrote:
On Sat, Feb 20, 2021 at 08:40:02AM +0000, liweihang wrote:
On 2021/2/10 3:54, Jason Gunthorpe wrote:
On Sun, Feb 07, 2021 at 11:12:49AM +0800, Weihang Li wrote:
The HIP09 introduces the DCA(Dynamic Context Attachment) feature which supports many RC QPs to share the WQE buffer in a memory pool. If a QP enables DCA feature, the WQE's buffer will not be allocated when creating but when the users start to post WRs. This will reduce the memory consumption when there are too many QPs are inactive. One a WQE buffer is allocated it still acts as a normal WQE ring buffer? So this DCA logic is to remap the send queue buffer based on demand for SQEs? How does it interact with the normal max send queue entries reported?
Not exactly. If DCA is enabled, we first allocate a memory pool with a default size when opening device. Each time we trying to post WR(s) to a QP, the driver will check if current QP has WQE buffer.
If not, the driver will check whether there is enough free memory in the DCA memory pool. If there is, the QP will get WQE buffer from the pool,
Does that mean the QP can have a non-contiguous list of buffers? Ie it isn't just a linear ring of memory?
Jason
Yes, when using multilevel addressing, the virtual address of QP's buffers may be non-contiguous to avoid memory fragmentation. Weihang
On 2021/2/10 3:54, Jason Gunthorpe wrote:
On Sun, Feb 07, 2021 at 11:12:49AM +0800, Weihang Li wrote:
The HIP09 introduces the DCA(Dynamic Context Attachment) feature which supports many RC QPs to share the WQE buffer in a memory pool. If a QP enables DCA feature, the WQE's buffer will not be allocated when creating but when the users start to post WRs. This will reduce the memory consumption when there are too many QPs are inactive.
One a WQE buffer is allocated it still acts as a normal WQE ring buffer? So this DCA logic is to remap the send queue buffer based on demand for SQEs? How does it interact with the normal max send queue entries reported?
Would like to see proper man pages explaining how this all works for rdma-core.
Jason
Hi Jason, I'm confused about how to introduce DCA in man pages. Current man pages in rdma-core can be classifed into public and vendor-defined ones. For example, ibv_create_qp.3 in libibverbs/man and mlx5dv.7 in providers/mlx5/man, but most of them is a description for a single interface. If we want to explain how to use DCA and how does it work, should we put a hns_dca.x file in providers/hns/man? Or add a file about hns_dca_open_device() and introduce DCA in it? And another question, I know the files with a number suffix like ibv_create_qp.3 is for man pages in linux. What about the markdown files with .md suffix like ibv_fork_init.3.md? If we want to add a new one about DCA, which type should we choose? Thanks Weihang
On Wed, Feb 24, 2021 at 09:48:34AM +0000, liweihang wrote:
I'm confused about how to introduce DCA in man pages. Current man pages in rdma-core can be classifed into public and vendor-defined ones. For example, ibv_create_qp.3 in libibverbs/man and mlx5dv.7 in providers/mlx5/man, but most of them is a description for a single interface. If we want to explain how to use DCA and how does it work, should we put a hns_dca.x file in providers/hns/man?
You could do that, yes
And another question, I know the files with a number suffix like ibv_create_qp.3 is for man pages in linux. What about the markdown files with .md suffix like ibv_fork_init.3.md? If we want to add a new one about DCA, which type should we choose?
Always use the md files for new man pages please Jason
participants (4)
-
Jason Gunthorpe -
Jason Gunthorpe -
liweihang -
Weihang Li