From b6a529da69ce880ee4d0f3730ef46ead7f7cd0d3 Mon Sep 17 00:00:00 2001 From: Henry Orosco Date: Wed, 30 Nov 2016 14:56:14 -0600 Subject: [PATCH] i40iw: Utilize physically mapped memory regions Add support to use physically mapped WQ's and MR's if determined that the OS registered user-memory for the region is physically contiguous. This feature will eliminate the need for unnecessarily setting up and using PBL's when not required. Signed-off-by: Shiraz Saleem Signed-off-by: Henry Orosco Signed-off-by: Doug Ledford --- drivers/infiniband/hw/i40iw/i40iw_verbs.c | 102 ++++++++++++++++++++-- 1 file changed, 93 insertions(+), 9 deletions(-) diff --git a/drivers/infiniband/hw/i40iw/i40iw_verbs.c b/drivers/infiniband/hw/i40iw/i40iw_verbs.c index dcf08b8994ec..43bae5b14a71 100644 --- a/drivers/infiniband/hw/i40iw/i40iw_verbs.c +++ b/drivers/infiniband/hw/i40iw/i40iw_verbs.c @@ -1355,11 +1355,63 @@ static void i40iw_copy_user_pgaddrs(struct i40iw_mr *iwmr, } } +/** + * i40iw_check_mem_contiguous - check if pbls stored in arr are contiguous + * @arr: lvl1 pbl array + * @npages: page count + * pg_size: page size + * + */ +static bool i40iw_check_mem_contiguous(u64 *arr, u32 npages, u32 pg_size) +{ + u32 pg_idx; + + for (pg_idx = 0; pg_idx < npages; pg_idx++) { + if ((*arr + (pg_size * pg_idx)) != arr[pg_idx]) + return false; + } + return true; +} + +/** + * i40iw_check_mr_contiguous - check if MR is physically contiguous + * @palloc: pbl allocation struct + * pg_size: page size + */ +static bool i40iw_check_mr_contiguous(struct i40iw_pble_alloc *palloc, u32 pg_size) +{ + struct i40iw_pble_level2 *lvl2 = &palloc->level2; + struct i40iw_pble_info *leaf = lvl2->leaf; + u64 *arr = NULL; + u64 *start_addr = NULL; + int i; + bool ret; + + if (palloc->level == I40IW_LEVEL_1) { + arr = (u64 *)palloc->level1.addr; + ret = i40iw_check_mem_contiguous(arr, palloc->total_cnt, pg_size); + return ret; + } + + start_addr = (u64 *)leaf->addr; + + for (i = 0; i < lvl2->leaf_cnt; i++, leaf++) { + arr = (u64 *)leaf->addr; + if ((*start_addr + (i * pg_size * PBLE_PER_PAGE)) != *arr) + return false; + ret = i40iw_check_mem_contiguous(arr, leaf->cnt, pg_size); + if (!ret) + return false; + } + + return true; +} + /** * i40iw_setup_pbles - copy user pg address to pble's * @iwdev: iwarp device * @iwmr: mr pointer for this memory registration - * @use_pbles: flag if to use pble's or memory (level 0) + * @use_pbles: flag if to use pble's */ static int i40iw_setup_pbles(struct i40iw_device *iwdev, struct i40iw_mr *iwmr, @@ -1372,9 +1424,6 @@ static int i40iw_setup_pbles(struct i40iw_device *iwdev, enum i40iw_status_code status; enum i40iw_pble_level level = I40IW_LEVEL_1; - if (!use_pbles && (iwmr->page_cnt > MAX_SAVE_PAGE_ADDRS)) - return -ENOMEM; - if (use_pbles) { mutex_lock(&iwdev->pbl_mutex); status = i40iw_get_pble(&iwdev->sc_dev, iwdev->pble_rsrc, palloc, iwmr->page_cnt); @@ -1391,6 +1440,10 @@ static int i40iw_setup_pbles(struct i40iw_device *iwdev, } i40iw_copy_user_pgaddrs(iwmr, pbl, level); + + if (use_pbles) + iwmr->pgaddrmem[0] = *pbl; + return 0; } @@ -1412,14 +1465,18 @@ static int i40iw_handle_q_mem(struct i40iw_device *iwdev, struct i40iw_cq_mr *cqmr = &iwpbl->cq_mr; struct i40iw_hmc_pble *hmc_p; u64 *arr = iwmr->pgaddrmem; + u32 pg_size; int err; int total; + bool ret = true; total = req->sq_pages + req->rq_pages + req->cq_pages; + pg_size = iwmr->region->page_size; err = i40iw_setup_pbles(iwdev, iwmr, use_pbles); if (err) return err; + if (use_pbles && (palloc->level != I40IW_LEVEL_1)) { i40iw_free_pble(iwdev->pble_rsrc, palloc); iwpbl->pbl_allocated = false; @@ -1428,26 +1485,44 @@ static int i40iw_handle_q_mem(struct i40iw_device *iwdev, if (use_pbles) arr = (u64 *)palloc->level1.addr; - if (req->reg_type == IW_MEMREG_TYPE_QP) { + + if (iwmr->type == IW_MEMREG_TYPE_QP) { hmc_p = &qpmr->sq_pbl; qpmr->shadow = (dma_addr_t)arr[total]; + if (use_pbles) { + ret = i40iw_check_mem_contiguous(arr, req->sq_pages, pg_size); + if (ret) + ret = i40iw_check_mem_contiguous(&arr[req->sq_pages], req->rq_pages, pg_size); + } + + if (!ret) { hmc_p->idx = palloc->level1.idx; hmc_p = &qpmr->rq_pbl; hmc_p->idx = palloc->level1.idx + req->sq_pages; } else { hmc_p->addr = arr[0]; hmc_p = &qpmr->rq_pbl; - hmc_p->addr = arr[1]; + hmc_p->addr = arr[req->sq_pages]; } } else { /* CQ */ hmc_p = &cqmr->cq_pbl; cqmr->shadow = (dma_addr_t)arr[total]; + if (use_pbles) + ret = i40iw_check_mem_contiguous(arr, req->cq_pages, pg_size); + + if (!ret) hmc_p->idx = palloc->level1.idx; else hmc_p->addr = arr[0]; } + + if (use_pbles && ret) { + i40iw_free_pble(iwdev->pble_rsrc, palloc); + iwpbl->pbl_allocated = false; + } + return err; } @@ -1646,7 +1721,7 @@ static int i40iw_hwreg_mr(struct i40iw_device *iwdev, stag_info->pd_id = iwpd->sc_pd.pd_id; stag_info->addr_type = I40IW_ADDR_TYPE_VA_BASED; - if (iwmr->page_cnt > 1) { + if (iwpbl->pbl_allocated) { if (palloc->level == I40IW_LEVEL_1) { stag_info->first_pm_pbl_index = palloc->level1.idx; stag_info->chunk_size = 1; @@ -1702,6 +1777,7 @@ static struct ib_mr *i40iw_reg_user_mr(struct ib_pd *pd, bool use_pbles = false; unsigned long flags; int err = -ENOSYS; + int ret; if (length > I40IW_MAX_MR_SIZE) return ERR_PTR(-EINVAL); @@ -1758,13 +1834,21 @@ static struct ib_mr *i40iw_reg_user_mr(struct ib_pd *pd, spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags); break; case IW_MEMREG_TYPE_MEM: + use_pbles = (iwmr->page_cnt != 1); access = I40IW_ACCESS_FLAGS_LOCALREAD; - use_pbles = (iwmr->page_cnt != 1); err = i40iw_setup_pbles(iwdev, iwmr, use_pbles); if (err) goto error; + if (use_pbles) { + ret = i40iw_check_mr_contiguous(palloc, region->page_size); + if (ret) { + i40iw_free_pble(iwdev->pble_rsrc, palloc); + iwpbl->pbl_allocated = false; + } + } + access |= i40iw_get_user_access(acc); stag = i40iw_create_stag(iwdev); if (!stag) { @@ -1792,7 +1876,7 @@ static struct ib_mr *i40iw_reg_user_mr(struct ib_pd *pd, return &iwmr->ibmr; error: - if (palloc->level != I40IW_LEVEL_0) + if (palloc->level != I40IW_LEVEL_0 && iwpbl->pbl_allocated) i40iw_free_pble(iwdev->pble_rsrc, palloc); ib_umem_release(region); kfree(iwmr); -- 2.39.5