RDMA/mlx5: Lower setting the umem's PAS for SRQ

Some of the SRQ types are created using a WQ, and the WQ requires a
different parameter set to mlx5_umem_find_best_quantized_pgoff() as it has
a 5 bit page_offset.

Add the umem to the mlx5_srq_attr and defer computing the PAS data until
the code has figured out what kind of mailbox to use. Compute the PAS
directly from the umem for each of the four unique mailbox types.

This also avoids allocating memory to store the user PAS, instead it is
written directly to the mailbox as in most other cases.

Fixes: 01949d0109ee ("net/mlx5_core: Enable XRCs and SRQs when using ISSI > 0")
Link: https://lore.kernel.org/r/20201115114311.136250-8-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>

diff --git a/drivers/infiniband/hw/mlx5/srq.c b/drivers/infiniband/hw/mlx5/srq.c
index 7dfdc9e..fab6736 100644 (file)
--- a/drivers/infiniband/hw/mlx5/srq.c
+++ b/drivers/infiniband/hw/mlx5/srq.c
@@ -51,8 +51,6 @@ static int create_srq_user(struct ib_pd *pd, struct mlx5_ib_srq *srq,
                udata, struct mlx5_ib_ucontext, ibucontext);
        size_t ucmdlen;
        int err;
-       unsigned int page_offset_quantized;
-       unsigned int page_size;
        u32 uidx = MLX5_IB_DEFAULT_UIDX;
 
        ucmdlen = min(udata->inlen, sizeof(ucmd));
@@ -84,32 +82,14 @@ static int create_srq_user(struct ib_pd *pd, struct mlx5_ib_srq *srq,
                err = PTR_ERR(srq->umem);
                return err;
        }
-
-       page_size = mlx5_umem_find_best_quantized_pgoff(
-               srq->umem, srqc, log_page_size, MLX5_ADAPTER_PAGE_SHIFT,
-               page_offset, 64, &page_offset_quantized);
-       if (!page_size) {
-               mlx5_ib_warn(dev, "bad offset\n");
-               goto err_umem;
-       }
-
-       in->pas = kvcalloc(ib_umem_num_dma_blocks(srq->umem, page_size),
-                          sizeof(*in->pas), GFP_KERNEL);
-       if (!in->pas) {
-               err = -ENOMEM;
-               goto err_umem;
-       }
-
-       mlx5_ib_populate_pas(srq->umem, page_size, in->pas, 0);
+       in->umem = srq->umem;
 
        err = mlx5_ib_db_map_user(ucontext, udata, ucmd.db_addr, &srq->db);
        if (err) {
                mlx5_ib_dbg(dev, "map doorbell failed\n");
-               goto err_in;
+               goto err_umem;
        }
 
-       in->log_page_size = order_base_2(page_size) - MLX5_ADAPTER_PAGE_SHIFT;
-       in->page_offset = page_offset_quantized;
        in->uid = (in->type != IB_SRQT_XRC) ?  to_mpd(pd)->uid : 0;
        if (MLX5_CAP_GEN(dev->mdev, cqe_version) == MLX5_CQE_VERSION_V1 &&
            in->type != IB_SRQT_BASIC)
@@ -117,9 +97,6 @@ static int create_srq_user(struct ib_pd *pd, struct mlx5_ib_srq *srq,
 
        return 0;
 
-err_in:
-       kvfree(in->pas);
-
 err_umem:
        ib_umem_release(srq->umem);
 

diff --git a/drivers/infiniband/hw/mlx5/srq.h b/drivers/infiniband/hw/mlx5/srq.h
index 2c3627b..a7e3dc5 100644 (file)
--- a/drivers/infiniband/hw/mlx5/srq.h
+++ b/drivers/infiniband/hw/mlx5/srq.h
@@ -28,6 +28,7 @@ struct mlx5_srq_attr {
        u32 user_index;
        u64 db_record;
        __be64 *pas;
+       struct ib_umem *umem;
        u32 tm_log_list_size;
        u32 tm_next_tag;
        u32 tm_hw_phase_cnt;

diff --git a/drivers/infiniband/hw/mlx5/srq_cmd.c b/drivers/infiniband/hw/mlx5/srq_cmd.c
index db889ec..8b33853 100644 (file)
--- a/drivers/infiniband/hw/mlx5/srq_cmd.c
+++ b/drivers/infiniband/hw/mlx5/srq_cmd.c
@@ -92,6 +92,25 @@ struct mlx5_core_srq *mlx5_cmd_get_srq(struct mlx5_ib_dev *dev, u32 srqn)
        return srq;
 }
 
+static int __set_srq_page_size(struct mlx5_srq_attr *in,
+                              unsigned long page_size)
+{
+       if (!page_size)
+               return -EINVAL;
+       in->log_page_size = order_base_2(page_size) - MLX5_ADAPTER_PAGE_SHIFT;
+
+       if (WARN_ON(get_pas_size(in) !=
+                   ib_umem_num_dma_blocks(in->umem, page_size) * sizeof(u64)))
+               return -EINVAL;
+       return 0;
+}
+
+#define set_srq_page_size(in, typ, log_pgsz_fld)                               \
+       __set_srq_page_size(in, mlx5_umem_find_best_quantized_pgoff(           \
+                                       (in)->umem, typ, log_pgsz_fld,         \
+                                       MLX5_ADAPTER_PAGE_SHIFT, page_offset,  \
+                                       64, &(in)->page_offset))
+
 static int create_srq_cmd(struct mlx5_ib_dev *dev, struct mlx5_core_srq *srq,
                          struct mlx5_srq_attr *in)
 {
@@ -103,6 +122,12 @@ static int create_srq_cmd(struct mlx5_ib_dev *dev, struct mlx5_core_srq *srq,
        int inlen;
        int err;
 
+       if (in->umem) {
+               err = set_srq_page_size(in, srqc, log_page_size);
+               if (err)
+                       return err;
+       }
+
        pas_size  = get_pas_size(in);
        inlen     = MLX5_ST_SZ_BYTES(create_srq_in) + pas_size;
        create_in = kvzalloc(inlen, GFP_KERNEL);
@@ -114,7 +139,13 @@ static int create_srq_cmd(struct mlx5_ib_dev *dev, struct mlx5_core_srq *srq,
        pas = MLX5_ADDR_OF(create_srq_in, create_in, pas);
 
        set_srqc(srqc, in);
-       memcpy(pas, in->pas, pas_size);
+       if (in->umem)
+               mlx5_ib_populate_pas(
+                       in->umem,
+                       1UL << (in->log_page_size + MLX5_ADAPTER_PAGE_SHIFT),
+                       pas, 0);
+       else
+               memcpy(pas, in->pas, pas_size);
 
        MLX5_SET(create_srq_in, create_in, opcode,
                 MLX5_CMD_OP_CREATE_SRQ);
@@ -194,6 +225,12 @@ static int create_xrc_srq_cmd(struct mlx5_ib_dev *dev,
        int inlen;
        int err;
 
+       if (in->umem) {
+               err = set_srq_page_size(in, xrc_srqc, log_page_size);
+               if (err)
+                       return err;
+       }
+
        pas_size  = get_pas_size(in);
        inlen     = MLX5_ST_SZ_BYTES(create_xrc_srq_in) + pas_size;
        create_in = kvzalloc(inlen, GFP_KERNEL);
@@ -207,7 +244,13 @@ static int create_xrc_srq_cmd(struct mlx5_ib_dev *dev,
 
        set_srqc(xrc_srqc, in);
        MLX5_SET(xrc_srqc, xrc_srqc, user_index, in->user_index);
-       memcpy(pas, in->pas, pas_size);
+       if (in->umem)
+               mlx5_ib_populate_pas(
+                       in->umem,
+                       1UL << (in->log_page_size + MLX5_ADAPTER_PAGE_SHIFT),
+                       pas, 0);
+       else
+               memcpy(pas, in->pas, pas_size);
        MLX5_SET(create_xrc_srq_in, create_in, opcode,
                 MLX5_CMD_OP_CREATE_XRC_SRQ);
 
@@ -289,11 +332,18 @@ static int create_rmp_cmd(struct mlx5_ib_dev *dev, struct mlx5_core_srq *srq,
        void *create_in = NULL;
        void *rmpc;
        void *wq;
+       void *pas;
        int pas_size;
        int outlen;
        int inlen;
        int err;
 
+       if (in->umem) {
+               err = set_srq_page_size(in, wq, log_wq_pg_sz);
+               if (err)
+                       return err;
+       }
+
        pas_size = get_pas_size(in);
        inlen = MLX5_ST_SZ_BYTES(create_rmp_in) + pas_size;
        outlen = MLX5_ST_SZ_BYTES(create_rmp_out);
@@ -309,8 +359,16 @@ static int create_rmp_cmd(struct mlx5_ib_dev *dev, struct mlx5_core_srq *srq,
 
        MLX5_SET(rmpc, rmpc, state, MLX5_RMPC_STATE_RDY);
        MLX5_SET(create_rmp_in, create_in, uid, in->uid);
+       pas = MLX5_ADDR_OF(rmpc, rmpc, wq.pas);
+
        set_wq(wq, in);
-       memcpy(MLX5_ADDR_OF(rmpc, rmpc, wq.pas), in->pas, pas_size);
+       if (in->umem)
+               mlx5_ib_populate_pas(
+                       in->umem,
+                       1UL << (in->log_page_size + MLX5_ADAPTER_PAGE_SHIFT),
+                       pas, 0);
+       else
+               memcpy(pas, in->pas, pas_size);
 
        MLX5_SET(create_rmp_in, create_in, opcode, MLX5_CMD_OP_CREATE_RMP);
        err = mlx5_cmd_exec(dev->mdev, create_in, inlen, create_out, outlen);
@@ -421,10 +479,17 @@ static int create_xrq_cmd(struct mlx5_ib_dev *dev, struct mlx5_core_srq *srq,
        void *create_in;
        void *xrqc;
        void *wq;
+       void *pas;
        int pas_size;
        int inlen;
        int err;
 
+       if (in->umem) {
+               err = set_srq_page_size(in, wq, log_wq_pg_sz);
+               if (err)
+                       return err;
+       }
+
        pas_size = get_pas_size(in);
        inlen = MLX5_ST_SZ_BYTES(create_xrq_in) + pas_size;
        create_in = kvzalloc(inlen, GFP_KERNEL);
@@ -433,9 +498,16 @@ static int create_xrq_cmd(struct mlx5_ib_dev *dev, struct mlx5_core_srq *srq,
 
        xrqc = MLX5_ADDR_OF(create_xrq_in, create_in, xrq_context);
        wq = MLX5_ADDR_OF(xrqc, xrqc, wq);
+       pas = MLX5_ADDR_OF(xrqc, xrqc, wq.pas);
 
        set_wq(wq, in);
-       memcpy(MLX5_ADDR_OF(xrqc, xrqc, wq.pas), in->pas, pas_size);
+       if (in->umem)
+               mlx5_ib_populate_pas(
+                       in->umem,
+                       1UL << (in->log_page_size + MLX5_ADAPTER_PAGE_SHIFT),
+                       pas, 0);
+       else
+               memcpy(pas, in->pas, pas_size);
 
        if (in->type == IB_SRQT_TM) {
                MLX5_SET(xrqc, xrqc, topology, MLX5_XRQC_TOPOLOGY_TAG_MATCHING);