mlx4_core: Write MTTs from CPU instead with of WRITE_MTT FW command

Write MTT entries directly to ICM from the driver (eliminating use of
WRITE_MTT command).  This reduces the number of FW commands needed to
register an MR by at least a factor of 2 and speeds up memory
registration significantly.  This code will also be used to implement
FMRs.

Signed-off-by: Jack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: Michael S. Tsirkin <mst@dev.mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>

5 files changed, 67 insertions, 47 deletions

diff --git a/drivers/infiniband/hw/mlx4/mr.c b/drivers/infiniband/hw/mlx4/mr.c
index 85ae906f1d12..734ec2bd15cd 100644
--- a/drivers/infiniband/hw/mlx4/mr.c
+++ b/drivers/infiniband/hw/mlx4/mr.c
@@ -96,11 +96,10 @@ int mlx4_ib_umem_write_mtt(struct mlx4_ib_dev *dev, struct mlx4_mtt *mtt,
                                 pages[i++] = sg_dma_address(&chunk->page_list[j]) +
                                         umem->page_size * k;
                                 /*
-                                 * Be friendly to WRITE_MTT firmware
-                                 * command, and pass it chunks of
-                                 * appropriate size.
+                                 * Be friendly to mlx4_write_mtt() and
+                                 * pass it chunks of appropriate size.
                                  */
-                                if (i == PAGE_SIZE / sizeof (u64) - 2) {
+                                if (i == PAGE_SIZE / sizeof (u64)) {
                                         err = mlx4_write_mtt(dev->dev, mtt, n,
                                                              i, pages);
                                         if (err)
diff --git a/drivers/net/mlx4/icm.c b/drivers/net/mlx4/icm.c
index 250e24887578..4b3c109d5eae 100644
--- a/drivers/net/mlx4/icm.c
+++ b/drivers/net/mlx4/icm.c
@@ -301,9 +301,9 @@ void mlx4_table_put(struct mlx4_dev *dev, struct mlx4_icm_table *table, int obj)
         mutex_unlock(&table->mutex);
 }
 
-void *mlx4_table_find(struct mlx4_icm_table *table, int obj)
+void *mlx4_table_find(struct mlx4_icm_table *table, int obj, dma_addr_t *dma_handle)
 {
-        int idx, offset, i;
+        int idx, offset, dma_offset, i;
         struct mlx4_icm_chunk *chunk;
         struct mlx4_icm *icm;
         struct page *page = NULL;
@@ -313,15 +313,26 @@ void *mlx4_table_find(struct mlx4_icm_table *table, int obj)
 
         mutex_lock(&table->mutex);
 
-        idx = obj & (table->num_obj - 1);
-        icm = table->icm[idx / (MLX4_TABLE_CHUNK_SIZE / table->obj_size)];
-        offset = idx % (MLX4_TABLE_CHUNK_SIZE / table->obj_size);
+        idx = (obj & (table->num_obj - 1)) * table->obj_size;
+        icm = table->icm[idx / MLX4_TABLE_CHUNK_SIZE];
+        dma_offset = offset = idx % MLX4_TABLE_CHUNK_SIZE;
 
         if (!icm)
                 goto out;
 
         list_for_each_entry(chunk, &icm->chunk_list, list) {
                 for (i = 0; i < chunk->npages; ++i) {
+                        if (dma_handle && dma_offset >= 0) {
+                                if (sg_dma_len(&chunk->mem[i]) > dma_offset)
+                                        *dma_handle = sg_dma_address(&chunk->mem[i]) +
+                                                dma_offset;
+                                dma_offset -= sg_dma_len(&chunk->mem[i]);
+                        }
+                        /*
+                         * DMA mapping can merge pages but not split them,
+                         * so if we found the page, dma_handle has already
+                         * been assigned to.
+                         */
                         if (chunk->mem[i].length > offset) {
                                 page = chunk->mem[i].page;
                                 goto out;
diff --git a/drivers/net/mlx4/icm.h b/drivers/net/mlx4/icm.h
index a77db6de8597..6c44edf35847 100644
--- a/drivers/net/mlx4/icm.h
+++ b/drivers/net/mlx4/icm.h
@@ -83,7 +83,7 @@ int mlx4_init_icm_table(struct mlx4_dev *dev, struct mlx4_icm_table *table,
 void mlx4_cleanup_icm_table(struct mlx4_dev *dev, struct mlx4_icm_table *table);
 int mlx4_table_get(struct mlx4_dev *dev, struct mlx4_icm_table *table, int obj);
 void mlx4_table_put(struct mlx4_dev *dev, struct mlx4_icm_table *table, int obj);
-void *mlx4_table_find(struct mlx4_icm_table *table, int obj);
+void *mlx4_table_find(struct mlx4_icm_table *table, int obj, dma_addr_t *dma_handle);
 int mlx4_table_get_range(struct mlx4_dev *dev, struct mlx4_icm_table *table,
                          int start, int end);
 void mlx4_table_put_range(struct mlx4_dev *dev, struct mlx4_icm_table *table,
diff --git a/drivers/net/mlx4/main.c b/drivers/net/mlx4/main.c
index ed7e8d76cde0..478b3ba74edc 100644
--- a/drivers/net/mlx4/main.c
+++ b/drivers/net/mlx4/main.c
@@ -300,6 +300,17 @@ static int __devinit mlx4_init_icm(struct mlx4_dev *dev,
                 goto err_unmap_cmpt;
         }
 
+        /*
+         * Reserved MTT entries must be aligned up to a cacheline
+         * boundary, since the FW will write to them, while the driver
+         * writes to all other MTT entries. (The variable
+         * dev->caps.mtt_entry_sz below is really the MTT segment
+         * size, not the raw entry size)
+         */
+        dev->caps.reserved_mtts =
+                ALIGN(dev->caps.reserved_mtts * dev->caps.mtt_entry_sz,
+                      dma_get_cache_alignment()) / dev->caps.mtt_entry_sz;
+
         err = mlx4_init_icm_table(dev, &priv->mr_table.mtt_table,
                                   init_hca->mtt_base,
                                   dev->caps.mtt_entry_sz,
diff --git a/drivers/net/mlx4/mr.c b/drivers/net/mlx4/mr.c
index 60a6ee27cfd4..3cc98c699aaf 100644
--- a/drivers/net/mlx4/mr.c
+++ b/drivers/net/mlx4/mr.c
@@ -349,58 +349,57 @@ err_table:
 }
 EXPORT_SYMBOL_GPL(mlx4_mr_enable);
 
-static int mlx4_WRITE_MTT(struct mlx4_dev *dev, struct mlx4_cmd_mailbox *mailbox,
-                          int num_mtt)
+static int mlx4_write_mtt_chunk(struct mlx4_dev *dev, struct mlx4_mtt *mtt,
+                                int start_index, int npages, u64 *page_list)
 {
-        return mlx4_cmd(dev, mailbox->dma, num_mtt, 0, MLX4_CMD_WRITE_MTT,
-                        MLX4_CMD_TIME_CLASS_B);
+        struct mlx4_priv *priv = mlx4_priv(dev);
+        __be64 *mtts;
+        dma_addr_t dma_handle;
+        int i;
+        int s = start_index * sizeof (u64);
+
+        /* All MTTs must fit in the same page */
+        if (start_index / (PAGE_SIZE / sizeof (u64)) !=
+            (start_index + npages - 1) / (PAGE_SIZE / sizeof (u64)))
+                return -EINVAL;
+
+        if (start_index & (MLX4_MTT_ENTRY_PER_SEG - 1))
+                return -EINVAL;
+
+        mtts = mlx4_table_find(&priv->mr_table.mtt_table, mtt->first_seg +
+                                s / dev->caps.mtt_entry_sz, &dma_handle);
+        if (!mtts)
+                return -ENOMEM;
+
+        for (i = 0; i < npages; ++i)
+                mtts[i] = cpu_to_be64(page_list[i] | MLX4_MTT_FLAG_PRESENT);
+
+        dma_sync_single(&dev->pdev->dev, dma_handle, npages * sizeof (u64), DMA_TO_DEVICE);
+
+        return 0;
 }
 
 int mlx4_write_mtt(struct mlx4_dev *dev, struct mlx4_mtt *mtt,
                    int start_index, int npages, u64 *page_list)
 {
-        struct mlx4_cmd_mailbox *mailbox;
-        __be64 *mtt_entry;
-        int i;
-        int err = 0;
+        int chunk;
+        int err;
 
         if (mtt->order < 0)
                 return -EINVAL;
 
-        mailbox = mlx4_alloc_cmd_mailbox(dev);
-        if (IS_ERR(mailbox))
-                return PTR_ERR(mailbox);
-
-        mtt_entry = mailbox->buf;
-
         while (npages > 0) {
-                mtt_entry[0] = cpu_to_be64(mlx4_mtt_addr(dev, mtt) + start_index * 8);
-                mtt_entry[1] = 0;
-
-                for (i = 0; i < npages && i < MLX4_MAILBOX_SIZE / 8 - 2; ++i)
-                        mtt_entry[i + 2] = cpu_to_be64(page_list[i] |
-                                                       MLX4_MTT_FLAG_PRESENT);
-
-                /*
-                 * If we have an odd number of entries to write, add
-                 * one more dummy entry for firmware efficiency.
-                 */
-                if (i & 1)
-                        mtt_entry[i + 2] = 0;
-
-                err = mlx4_WRITE_MTT(dev, mailbox, (i + 1) & ~1);
+                chunk = min_t(int, PAGE_SIZE / sizeof(u64), npages);
+                err = mlx4_write_mtt_chunk(dev, mtt, start_index, chunk, page_list);
                 if (err)
-                        goto out;
+                        return err;
 
-                npages      -= i;
-                start_index += i;
-                page_list   += i;
+                npages      -= chunk;
+                start_index += chunk;
+                page_list   += chunk;
         }
 
-out:
-        mlx4_free_cmd_mailbox(dev, mailbox);
-
-        return err;
+        return 0;
 }
 EXPORT_SYMBOL_GPL(mlx4_write_mtt);