RDMA/hns: Optimize the base address table config for MTR
authorXi Wang <wangxi11@huawei.com>
Fri, 21 May 2021 09:29:51 +0000 (17:29 +0800)
committerJason Gunthorpe <jgg@nvidia.com>
Fri, 28 May 2021 23:13:57 +0000 (20:13 -0300)
The base address table is allocated by dma allocator, and the size is
always aligned to PAGE_SIZE. If a fixed size is used to allocate the
table, the number of base address entries stored in the table will be
smaller than that can actually stored.

Link: https://lore.kernel.org/r/1621589395-2435-2-git-send-email-liweihang@huawei.com
Signed-off-by: Xi Wang <wangxi11@huawei.com>
Signed-off-by: Weihang Li <liweihang@huawei.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
drivers/infiniband/hw/hns/hns_roce_alloc.c
drivers/infiniband/hw/hns/hns_roce_cq.c
drivers/infiniband/hw/hns/hns_roce_device.h
drivers/infiniband/hw/hns/hns_roce_hw_v2.c
drivers/infiniband/hw/hns/hns_roce_mr.c
drivers/infiniband/hw/hns/hns_roce_qp.c
drivers/infiniband/hw/hns/hns_roce_srq.c

index 5d389ed..51374b6 100644 (file)
@@ -208,10 +208,10 @@ struct hns_roce_buf *hns_roce_buf_alloc(struct hns_roce_dev *hr_dev, u32 size,
 
        /* Calc the trunk size and num by required size and page_shift */
        if (flags & HNS_ROCE_BUF_DIRECT) {
-               buf->trunk_shift = ilog2(ALIGN(size, PAGE_SIZE));
+               buf->trunk_shift = order_base_2(ALIGN(size, PAGE_SIZE));
                ntrunk = 1;
        } else {
-               buf->trunk_shift = ilog2(ALIGN(page_size, PAGE_SIZE));
+               buf->trunk_shift = order_base_2(ALIGN(page_size, PAGE_SIZE));
                ntrunk = DIV_ROUND_UP(size, 1 << buf->trunk_shift);
        }
 
@@ -252,50 +252,41 @@ struct hns_roce_buf *hns_roce_buf_alloc(struct hns_roce_dev *hr_dev, u32 size,
 }
 
 int hns_roce_get_kmem_bufs(struct hns_roce_dev *hr_dev, dma_addr_t *bufs,
-                          int buf_cnt, int start, struct hns_roce_buf *buf)
+                          int buf_cnt, struct hns_roce_buf *buf,
+                          unsigned int page_shift)
 {
-       int i, end;
-       int total;
-
-       end = start + buf_cnt;
-       if (end > buf->npages) {
-               dev_err(hr_dev->dev,
-                       "failed to check kmem bufs, end %d + %d total %u!\n",
-                       start, buf_cnt, buf->npages);
+       unsigned int offset, max_size;
+       int total = 0;
+       int i;
+
+       if (page_shift > buf->trunk_shift) {
+               dev_err(hr_dev->dev, "failed to check kmem buf shift %u > %u\n",
+                       page_shift, buf->trunk_shift);
                return -EINVAL;
        }
 
-       total = 0;
-       for (i = start; i < end; i++)
-               bufs[total++] = hns_roce_buf_page(buf, i);
+       offset = 0;
+       max_size = buf->ntrunks << buf->trunk_shift;
+       for (i = 0; i < buf_cnt && offset < max_size; i++) {
+               bufs[total++] = hns_roce_buf_dma_addr(buf, offset);
+               offset += (1 << page_shift);
+       }
 
        return total;
 }
 
 int hns_roce_get_umem_bufs(struct hns_roce_dev *hr_dev, dma_addr_t *bufs,
-                          int buf_cnt, int start, struct ib_umem *umem,
+                          int buf_cnt, struct ib_umem *umem,
                           unsigned int page_shift)
 {
        struct ib_block_iter biter;
        int total = 0;
-       int idx = 0;
-       u64 addr;
-
-       if (page_shift < HNS_HW_PAGE_SHIFT) {
-               dev_err(hr_dev->dev, "failed to check umem page shift %u!\n",
-                       page_shift);
-               return -EINVAL;
-       }
 
        /* convert system page cnt to hw page cnt */
        rdma_umem_for_each_dma_block(umem, &biter, 1 << page_shift) {
-               addr = rdma_block_iter_dma_address(&biter);
-               if (idx >= start) {
-                       bufs[total++] = addr;
-                       if (total >= buf_cnt)
-                               goto done;
-               }
-               idx++;
+               bufs[total++] = rdma_block_iter_dma_address(&biter);
+               if (total >= buf_cnt)
+                       goto done;
        }
 
 done:
index 488d86b..a5a74b6 100644 (file)
@@ -202,13 +202,13 @@ static int alloc_cq_buf(struct hns_roce_dev *hr_dev, struct hns_roce_cq *hr_cq,
        struct hns_roce_buf_attr buf_attr = {};
        int ret;
 
-       buf_attr.page_shift = hr_dev->caps.cqe_buf_pg_sz + HNS_HW_PAGE_SHIFT;
+       buf_attr.page_shift = hr_dev->caps.cqe_buf_pg_sz + PAGE_SHIFT;
        buf_attr.region[0].size = hr_cq->cq_depth * hr_cq->cqe_size;
        buf_attr.region[0].hopnum = hr_dev->caps.cqe_hop_num;
        buf_attr.region_count = 1;
 
        ret = hns_roce_mtr_create(hr_dev, &hr_cq->mtr, &buf_attr,
-                                 hr_dev->caps.cqe_ba_pg_sz + HNS_HW_PAGE_SHIFT,
+                                 hr_dev->caps.cqe_ba_pg_sz + PAGE_SHIFT,
                                  udata, addr);
        if (ret)
                ibdev_err(ibdev, "failed to alloc CQ mtr, ret = %d.\n", ret);
index 111cab5..02fffb7 100644 (file)
@@ -1060,14 +1060,18 @@ static inline void *hns_roce_buf_offset(struct hns_roce_buf *buf,
                        (offset & ((1 << buf->trunk_shift) - 1));
 }
 
-static inline dma_addr_t hns_roce_buf_page(struct hns_roce_buf *buf, u32 idx)
+static inline dma_addr_t hns_roce_buf_dma_addr(struct hns_roce_buf *buf,
+                                              unsigned int offset)
 {
-       unsigned int offset = idx << buf->page_shift;
-
        return buf->trunk_list[offset >> buf->trunk_shift].map +
                        (offset & ((1 << buf->trunk_shift) - 1));
 }
 
+static inline dma_addr_t hns_roce_buf_page(struct hns_roce_buf *buf, u32 idx)
+{
+       return hns_roce_buf_dma_addr(buf, idx << buf->page_shift);
+}
+
 #define hr_hw_page_align(x)            ALIGN(x, 1 << HNS_HW_PAGE_SHIFT)
 
 static inline u64 to_hr_hw_page_addr(u64 addr)
@@ -1204,9 +1208,10 @@ struct hns_roce_buf *hns_roce_buf_alloc(struct hns_roce_dev *hr_dev, u32 size,
                                        u32 page_shift, u32 flags);
 
 int hns_roce_get_kmem_bufs(struct hns_roce_dev *hr_dev, dma_addr_t *bufs,
-                          int buf_cnt, int start, struct hns_roce_buf *buf);
+                          int buf_cnt, struct hns_roce_buf *buf,
+                          unsigned int page_shift);
 int hns_roce_get_umem_bufs(struct hns_roce_dev *hr_dev, dma_addr_t *bufs,
-                          int buf_cnt, int start, struct ib_umem *umem,
+                          int buf_cnt, struct ib_umem *umem,
                           unsigned int page_shift);
 
 int hns_roce_create_srq(struct ib_srq *srq,
index d5e72e5..0c12d87 100644 (file)
@@ -2018,6 +2018,8 @@ static void set_hem_page_size(struct hns_roce_dev *hr_dev)
        caps->llm_buf_pg_sz = 0;
 
        /* MR */
+       caps->mpt_ba_pg_sz = 0;
+       caps->mpt_buf_pg_sz = 0;
        caps->pbl_ba_pg_sz = HNS_ROCE_BA_PG_SZ_SUPPORTED_16K;
        caps->pbl_buf_pg_sz = 0;
        calc_pg_sz(caps->num_mtpts, caps->mtpt_entry_sz, caps->mpt_hop_num,
@@ -2025,8 +2027,12 @@ static void set_hem_page_size(struct hns_roce_dev *hr_dev)
                   HEM_TYPE_MTPT);
 
        /* QP */
-       caps->qpc_timer_ba_pg_sz  = 0;
+       caps->qpc_ba_pg_sz = 0;
+       caps->qpc_buf_pg_sz = 0;
+       caps->qpc_timer_ba_pg_sz = 0;
        caps->qpc_timer_buf_pg_sz = 0;
+       caps->sccc_ba_pg_sz = 0;
+       caps->sccc_buf_pg_sz = 0;
        caps->mtt_ba_pg_sz = 0;
        caps->mtt_buf_pg_sz = 0;
        calc_pg_sz(caps->num_qps, caps->qpc_sz, caps->qpc_hop_num,
@@ -2039,6 +2045,12 @@ static void set_hem_page_size(struct hns_roce_dev *hr_dev)
                           &caps->sccc_ba_pg_sz, HEM_TYPE_SCCC);
 
        /* CQ */
+       caps->cqc_ba_pg_sz = 0;
+       caps->cqc_buf_pg_sz = 0;
+       caps->cqc_timer_ba_pg_sz = 0;
+       caps->cqc_timer_buf_pg_sz = 0;
+       caps->cqe_ba_pg_sz = HNS_ROCE_BA_PG_SZ_SUPPORTED_256K;
+       caps->cqe_buf_pg_sz = 0;
        calc_pg_sz(caps->num_cqs, caps->cqc_entry_sz, caps->cqc_hop_num,
                   caps->cqc_bt_num, &caps->cqc_buf_pg_sz, &caps->cqc_ba_pg_sz,
                   HEM_TYPE_CQC);
@@ -2053,6 +2065,12 @@ static void set_hem_page_size(struct hns_roce_dev *hr_dev)
 
        /* SRQ */
        if (caps->flags & HNS_ROCE_CAP_FLAG_SRQ) {
+               caps->srqc_ba_pg_sz = 0;
+               caps->srqc_buf_pg_sz = 0;
+               caps->srqwqe_ba_pg_sz = 0;
+               caps->srqwqe_buf_pg_sz = 0;
+               caps->idx_ba_pg_sz = 0;
+               caps->idx_buf_pg_sz = 0;
                calc_pg_sz(caps->num_srqs, caps->srqc_entry_sz,
                           caps->srqc_hop_num, caps->srqc_bt_num,
                           &caps->srqc_buf_pg_sz, &caps->srqc_ba_pg_sz,
@@ -6161,14 +6179,14 @@ static int alloc_eq_buf(struct hns_roce_dev *hr_dev, struct hns_roce_eq *eq)
        else
                eq->hop_num = hr_dev->caps.eqe_hop_num;
 
-       buf_attr.page_shift = hr_dev->caps.eqe_buf_pg_sz + HNS_HW_PAGE_SHIFT;
+       buf_attr.page_shift = hr_dev->caps.eqe_buf_pg_sz + PAGE_SHIFT;
        buf_attr.region[0].size = eq->entries * eq->eqe_size;
        buf_attr.region[0].hopnum = eq->hop_num;
        buf_attr.region_count = 1;
 
        err = hns_roce_mtr_create(hr_dev, &eq->mtr, &buf_attr,
-                                 hr_dev->caps.eqe_ba_pg_sz +
-                                 HNS_HW_PAGE_SHIFT, NULL, 0);
+                                 hr_dev->caps.eqe_ba_pg_sz + PAGE_SHIFT, NULL,
+                                 0);
        if (err)
                dev_err(hr_dev->dev, "Failed to alloc EQE mtr, err %d\n", err);
 
index 79b3c30..8e6b1ae 100644 (file)
@@ -122,7 +122,7 @@ static int alloc_mr_pbl(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr,
        buf_attr.mtt_only = is_fast;
 
        err = hns_roce_mtr_create(hr_dev, &mr->pbl_mtr, &buf_attr,
-                                 hr_dev->caps.pbl_ba_pg_sz + HNS_HW_PAGE_SHIFT,
+                                 hr_dev->caps.pbl_ba_pg_sz + PAGE_SHIFT,
                                  udata, start);
        if (err)
                ibdev_err(ibdev, "failed to alloc pbl mtr, ret = %d.\n", err);
@@ -737,11 +737,11 @@ static int mtr_map_bufs(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr,
                return -ENOMEM;
 
        if (mtr->umem)
-               npage = hns_roce_get_umem_bufs(hr_dev, pages, page_count, 0,
+               npage = hns_roce_get_umem_bufs(hr_dev, pages, page_count,
                                               mtr->umem, page_shift);
        else
-               npage = hns_roce_get_kmem_bufs(hr_dev, pages, page_count, 0,
-                                              mtr->kmem);
+               npage = hns_roce_get_kmem_bufs(hr_dev, pages, page_count,
+                                              mtr->kmem, page_shift);
 
        if (npage != page_count) {
                ibdev_err(ibdev, "failed to get mtr page %d != %d.\n", npage,
@@ -753,8 +753,8 @@ static int mtr_map_bufs(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr,
        if (mtr->hem_cfg.is_direct && npage > 1) {
                ret = mtr_check_direct_pages(pages, npage, page_shift);
                if (ret) {
-                       ibdev_err(ibdev, "failed to check %s mtr, idx = %d.\n",
-                                 mtr->umem ? "user" : "kernel", ret);
+                       ibdev_err(ibdev, "failed to check %s page: %d / %d.\n",
+                                 mtr->umem ? "umtr" : "kmtr", ret, npage);
                        ret = -ENOBUFS;
                        goto err_alloc_list;
                }
@@ -799,7 +799,7 @@ int hns_roce_mtr_map(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr,
                if (r->offset + r->count > page_cnt) {
                        ret = -EINVAL;
                        ibdev_err(ibdev,
-                                 "failed to check mtr%u end %u + %u, max %u.\n",
+                                 "failed to check mtr%u count %u + %u > %u.\n",
                                  i, r->offset, r->count, page_cnt);
                        return ret;
                }
index c6e120e..9203cf1 100644 (file)
@@ -761,7 +761,7 @@ static int alloc_qp_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
                goto err_inline;
        }
        ret = hns_roce_mtr_create(hr_dev, &hr_qp->mtr, &buf_attr,
-                                 HNS_HW_PAGE_SHIFT + hr_dev->caps.mtt_ba_pg_sz,
+                                 PAGE_SHIFT + hr_dev->caps.mtt_ba_pg_sz,
                                  udata, addr);
        if (ret) {
                ibdev_err(ibdev, "failed to create WQE mtr, ret = %d.\n", ret);
index 546d182..c842210 100644 (file)
@@ -167,14 +167,14 @@ static int alloc_srq_idx(struct hns_roce_dev *hr_dev, struct hns_roce_srq *srq,
 
        srq->idx_que.entry_shift = ilog2(HNS_ROCE_IDX_QUE_ENTRY_SZ);
 
-       buf_attr.page_shift = hr_dev->caps.idx_buf_pg_sz + HNS_HW_PAGE_SHIFT;
+       buf_attr.page_shift = hr_dev->caps.idx_buf_pg_sz + PAGE_SHIFT;
        buf_attr.region[0].size = to_hr_hem_entries_size(srq->wqe_cnt,
                                        srq->idx_que.entry_shift);
        buf_attr.region[0].hopnum = hr_dev->caps.idx_hop_num;
        buf_attr.region_count = 1;
 
        ret = hns_roce_mtr_create(hr_dev, &idx_que->mtr, &buf_attr,
-                                 hr_dev->caps.idx_ba_pg_sz + HNS_HW_PAGE_SHIFT,
+                                 hr_dev->caps.idx_ba_pg_sz + PAGE_SHIFT,
                                  udata, addr);
        if (ret) {
                ibdev_err(ibdev,
@@ -222,15 +222,15 @@ static int alloc_srq_wqe_buf(struct hns_roce_dev *hr_dev,
                                                      HNS_ROCE_SGE_SIZE *
                                                      srq->max_gs)));
 
-       buf_attr.page_shift = hr_dev->caps.srqwqe_buf_pg_sz + HNS_HW_PAGE_SHIFT;
+       buf_attr.page_shift = hr_dev->caps.srqwqe_buf_pg_sz + PAGE_SHIFT;
        buf_attr.region[0].size = to_hr_hem_entries_size(srq->wqe_cnt,
                                                         srq->wqe_shift);
        buf_attr.region[0].hopnum = hr_dev->caps.srqwqe_hop_num;
        buf_attr.region_count = 1;
 
        ret = hns_roce_mtr_create(hr_dev, &srq->buf_mtr, &buf_attr,
-                                 hr_dev->caps.srqwqe_ba_pg_sz +
-                                 HNS_HW_PAGE_SHIFT, udata, addr);
+                                 hr_dev->caps.srqwqe_ba_pg_sz + PAGE_SHIFT,
+                                 udata, addr);
        if (ret)
                ibdev_err(ibdev,
                          "failed to alloc SRQ buf mtr, ret = %d.\n", ret);