]> asedeno.scripts.mit.edu Git - linux.git/commitdiff
i40iw: Utilize physically mapped memory regions
authorHenry Orosco <henry.orosco@intel.com>
Wed, 30 Nov 2016 20:56:14 +0000 (14:56 -0600)
committerDoug Ledford <dledford@redhat.com>
Mon, 5 Dec 2016 21:09:39 +0000 (16:09 -0500)
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 <shiraz.saleem@intel.com>
Signed-off-by: Henry Orosco <henry.orosco@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
drivers/infiniband/hw/i40iw/i40iw_verbs.c

index dcf08b8994ecd5f298424d8d65d1b691115dec0e..43bae5b14a71d03c824dd5789af75a9f6f600400 100644 (file)
@@ -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);