]> git.baikalelectronics.ru Git - kernel.git/commitdiff
IB/mlx5: Decouple MR allocation and population flows
authorIlya Lesokhin <ilyal@mellanox.com>
Thu, 17 Aug 2017 12:52:30 +0000 (15:52 +0300)
committerDoug Ledford <dledford@redhat.com>
Thu, 24 Aug 2017 21:47:34 +0000 (17:47 -0400)
mlx5 compatible devices have two ways of populating the MTT
table of an MKEY: using a FW command and using a UMR WQE.

A UMR is much faster, so it should be used whenever possible.
Unfortunately the code today uses UMR only if the MKEY was allocated
from the MR cache.

Fix the code to use UMR even for MKEYs that were allocated using
a FW command.

Signed-off-by: Ilya Lesokhin <ilyal@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>
drivers/infiniband/hw/mlx5/mr.c

index bc87016021e3586ae6f9a89483ea4563aaabf8b2..aa6f71570b77be7e3ca1b5f98e7ec1aa25a0e93c 100644 (file)
@@ -898,7 +898,8 @@ static int mlx5_ib_post_send_wait(struct mlx5_ib_dev *dev,
        return err;
 }
 
-static struct mlx5_ib_mr *reg_umr(struct ib_pd *pd, struct ib_umem *umem,
+static struct mlx5_ib_mr *alloc_mr_from_cache(
+                                 struct ib_pd *pd, struct ib_umem *umem,
                                  u64 virt_addr, u64 len, int npages,
                                  int page_shift, int order, int access_flags)
 {
@@ -930,16 +931,6 @@ static struct mlx5_ib_mr *reg_umr(struct ib_pd *pd, struct ib_umem *umem,
        mr->mmkey.size = len;
        mr->mmkey.pd = to_mpd(pd)->pdn;
 
-       err = mlx5_ib_update_xlt(mr, 0, npages, page_shift,
-                                MLX5_IB_UPD_XLT_ENABLE);
-
-       if (err) {
-               mlx5_mr_cache_free(dev, mr);
-               return ERR_PTR(err);
-       }
-
-       mr->live = 1;
-
        return mr;
 }
 
@@ -1105,7 +1096,8 @@ free_xlt:
 static struct mlx5_ib_mr *reg_create(struct ib_mr *ibmr, struct ib_pd *pd,
                                     u64 virt_addr, u64 length,
                                     struct ib_umem *umem, int npages,
-                                    int page_shift, int access_flags)
+                                    int page_shift, int access_flags,
+                                    bool populate)
 {
        struct mlx5_ib_dev *dev = to_mdev(pd->device);
        struct mlx5_ib_mr *mr;
@@ -1120,15 +1112,19 @@ static struct mlx5_ib_mr *reg_create(struct ib_mr *ibmr, struct ib_pd *pd,
        if (!mr)
                return ERR_PTR(-ENOMEM);
 
-       inlen = MLX5_ST_SZ_BYTES(create_mkey_in) +
-               sizeof(*pas) * ((npages + 1) / 2) * 2;
+       mr->ibmr.pd = pd;
+       mr->access_flags = access_flags;
+
+       inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
+       if (populate)
+               inlen += sizeof(*pas) * roundup(npages, 2);
        in = kvzalloc(inlen, GFP_KERNEL);
        if (!in) {
                err = -ENOMEM;
                goto err_1;
        }
        pas = (__be64 *)MLX5_ADDR_OF(create_mkey_in, in, klm_pas_mtt);
-       if (!(access_flags & IB_ACCESS_ON_DEMAND))
+       if (populate && !(access_flags & IB_ACCESS_ON_DEMAND))
                mlx5_ib_populate_pas(dev, umem, page_shift, pas,
                                     pg_cap ? MLX5_IB_MTT_PRESENT : 0);
 
@@ -1137,6 +1133,7 @@ static struct mlx5_ib_mr *reg_create(struct ib_mr *ibmr, struct ib_pd *pd,
        MLX5_SET(create_mkey_in, in, pg_access, !!(pg_cap));
 
        mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
+       MLX5_SET(mkc, mkc, free, !populate);
        MLX5_SET(mkc, mkc, access_mode, MLX5_MKC_ACCESS_MODE_MTT);
        MLX5_SET(mkc, mkc, a, !!(access_flags & IB_ACCESS_REMOTE_ATOMIC));
        MLX5_SET(mkc, mkc, rw, !!(access_flags & IB_ACCESS_REMOTE_WRITE));
@@ -1153,8 +1150,10 @@ static struct mlx5_ib_mr *reg_create(struct ib_mr *ibmr, struct ib_pd *pd,
                 get_octo_len(virt_addr, length, 1 << page_shift));
        MLX5_SET(mkc, mkc, log_page_size, page_shift);
        MLX5_SET(mkc, mkc, qpn, 0xffffff);
-       MLX5_SET(create_mkey_in, in, translations_octword_actual_size,
-                get_octo_len(virt_addr, length, 1 << page_shift));
+       if (populate) {
+               MLX5_SET(create_mkey_in, in, translations_octword_actual_size,
+                        get_octo_len(virt_addr, length, 1 << page_shift));
+       }
 
        err = mlx5_core_create_mkey(dev->mdev, &mr->mmkey, in, inlen);
        if (err) {
@@ -1163,9 +1162,7 @@ static struct mlx5_ib_mr *reg_create(struct ib_mr *ibmr, struct ib_pd *pd,
        }
        mr->mmkey.type = MLX5_MKEY_MR;
        mr->desc_size = sizeof(struct mlx5_mtt);
-       mr->umem = umem;
        mr->dev = dev;
-       mr->live = 1;
        kvfree(in);
 
        mlx5_ib_dbg(dev, "mkey = 0x%x\n", mr->mmkey.key);
@@ -1205,6 +1202,7 @@ struct ib_mr *mlx5_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
        int ncont;
        int order;
        int err;
+       bool use_umr = true;
 
        mlx5_ib_dbg(dev, "start 0x%llx, virt_addr 0x%llx, length 0x%llx, access_flags 0x%x\n",
                    start, virt_addr, length, access_flags);
@@ -1223,27 +1221,29 @@ struct ib_mr *mlx5_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
        err = mr_umem_get(pd, start, length, access_flags, &umem, &npages,
                           &page_shift, &ncont, &order);
 
-        if (err < 0)
+       if (err < 0)
                return ERR_PTR(err);
 
        if (order <= mr_cache_max_order(dev)) {
-               mr = reg_umr(pd, umem, virt_addr, length, ncont, page_shift,
-                            order, access_flags);
+               mr = alloc_mr_from_cache(pd, umem, virt_addr, length, ncont,
+                                        page_shift, order, access_flags);
                if (PTR_ERR(mr) == -EAGAIN) {
                        mlx5_ib_dbg(dev, "cache empty for order %d", order);
                        mr = NULL;
                }
-       } else if (access_flags & IB_ACCESS_ON_DEMAND &&
-                  !MLX5_CAP_GEN(dev->mdev, umr_extended_translation_offset)) {
-               err = -EINVAL;
-               pr_err("Got MR registration for ODP MR > 512MB, not supported for Connect-IB");
-               goto error;
+       } else if (!MLX5_CAP_GEN(dev->mdev, umr_extended_translation_offset)) {
+               if (access_flags & IB_ACCESS_ON_DEMAND) {
+                       err = -EINVAL;
+                       pr_err("Got MR registration for ODP MR > 512MB, not supported for Connect-IB");
+                       goto error;
+               }
+               use_umr = false;
        }
 
        if (!mr) {
                mutex_lock(&dev->slow_path_mutex);
                mr = reg_create(NULL, pd, virt_addr, length, umem, ncont,
-                               page_shift, access_flags);
+                               page_shift, access_flags, !use_umr);
                mutex_unlock(&dev->slow_path_mutex);
        }
 
@@ -1261,8 +1261,22 @@ struct ib_mr *mlx5_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
        update_odp_mr(mr);
 #endif
 
-       return &mr->ibmr;
+       if (use_umr) {
+               int update_xlt_flags = MLX5_IB_UPD_XLT_ENABLE;
+
+               if (access_flags & IB_ACCESS_ON_DEMAND)
+                       update_xlt_flags |= MLX5_IB_UPD_XLT_ZAP;
 
+               err = mlx5_ib_update_xlt(mr, 0, ncont, page_shift,
+                                        update_xlt_flags);
+               if (err) {
+                       mlx5_ib_dereg_mr(&mr->ibmr);
+                       return ERR_PTR(err);
+               }
+       }
+
+       mr->live = 1;
+       return &mr->ibmr;
 error:
        ib_umem_release(umem);
        return ERR_PTR(err);
@@ -1363,12 +1377,13 @@ int mlx5_ib_rereg_user_mr(struct ib_mr *ib_mr, int flags, u64 start,
                        return err;
 
                mr = reg_create(ib_mr, pd, addr, len, mr->umem, ncont,
-                               page_shift, access_flags);
+                               page_shift, access_flags, true);
 
                if (IS_ERR(mr))
                        return PTR_ERR(mr);
 
                mr->allocated_from_cache = 0;
+               mr->live = 1;
        } else {
                /*
                 * Send a UMR WQE