mlx4_core: Support ICM tables in coherent memory

Enable having ICM tables in coherent memory, and use coherent memory
for the dMPT table.  This will allow writing MPT entries for MRs both
via the SW2HW_MPT command and also directly by the driver for FMR
remapping without needing to flush or worry about cacheline boundaries.

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>

4 files changed, 112 insertions, 49 deletions

diff --git a/drivers/net/mlx4/icm.c b/drivers/net/mlx4/icm.c
index b7a4aa8..250e248 100644
--- a/drivers/net/mlx4/icm.c
+++ b/drivers/net/mlx4/icm.c
@@ -34,6 +34,7 @@
 #include <linux/init.h>
 #include <linux/errno.h>
 #include <linux/mm.h>
+#include <linux/scatterlist.h>
 
 #include <linux/mlx4/cmd.h>
 
@@ -50,19 +51,41 @@ enum {
 	MLX4_TABLE_CHUNK_SIZE	= 1 << 18
 };
 
-void mlx4_free_icm(struct mlx4_dev *dev, struct mlx4_icm *icm)
+static void mlx4_free_icm_pages(struct mlx4_dev *dev, struct mlx4_icm_chunk *chunk)
 {
-	struct mlx4_icm_chunk *chunk, *tmp;
 	int i;
 
-	list_for_each_entry_safe(chunk, tmp, &icm->chunk_list, list) {
-		if (chunk->nsg > 0)
-			pci_unmap_sg(dev->pdev, chunk->mem, chunk->npages,
-				     PCI_DMA_BIDIRECTIONAL);
+	if (chunk->nsg > 0)
+		pci_unmap_sg(dev->pdev, chunk->mem, chunk->npages,
+			     PCI_DMA_BIDIRECTIONAL);
+
+	for (i = 0; i < chunk->npages; ++i)
+		__free_pages(chunk->mem[i].page,
+			     get_order(chunk->mem[i].length));
+}
 
-		for (i = 0; i < chunk->npages; ++i)
-			__free_pages(chunk->mem[i].page,
-				     get_order(chunk->mem[i].length));
+static void mlx4_free_icm_coherent(struct mlx4_dev *dev, struct mlx4_icm_chunk *chunk)
+{
+	int i;
+
+	for (i = 0; i < chunk->npages; ++i)
+		dma_free_coherent(&dev->pdev->dev, chunk->mem[i].length,
+				  lowmem_page_address(chunk->mem[i].page),
+				  sg_dma_address(&chunk->mem[i]));
+}
+
+void mlx4_free_icm(struct mlx4_dev *dev, struct mlx4_icm *icm, int coherent)
+{
+	struct mlx4_icm_chunk *chunk, *tmp;
+
+	if (!icm)
+		return;
+
+	list_for_each_entry_safe(chunk, tmp, &icm->chunk_list, list) {
+		if (coherent)
+			mlx4_free_icm_coherent(dev, chunk);
+		else
+			mlx4_free_icm_pages(dev, chunk);
 
 		kfree(chunk);
 	}
@@ -70,16 +93,45 @@ void mlx4_free_icm(struct mlx4_dev *dev, struct mlx4_icm *icm)
 	kfree(icm);
 }
 
+static int mlx4_alloc_icm_pages(struct scatterlist *mem, int order, gfp_t gfp_mask)
+{
+	mem->page = alloc_pages(gfp_mask, order);
+	if (!mem->page)
+		return -ENOMEM;
+
+	mem->length = PAGE_SIZE << order;
+	mem->offset = 0;
+	return 0;
+}
+
+static int mlx4_alloc_icm_coherent(struct device *dev, struct scatterlist *mem,
+				    int order, gfp_t gfp_mask)
+{
+	void *buf = dma_alloc_coherent(dev, PAGE_SIZE << order,
+				       &sg_dma_address(mem), gfp_mask);
+	if (!buf)
+		return -ENOMEM;
+
+	sg_set_buf(mem, buf, PAGE_SIZE << order);
+	BUG_ON(mem->offset);
+	sg_dma_len(mem) = PAGE_SIZE << order;
+	return 0;
+}
+
 struct mlx4_icm *mlx4_alloc_icm(struct mlx4_dev *dev, int npages,
-				gfp_t gfp_mask)
+				gfp_t gfp_mask, int coherent)
 {
 	struct mlx4_icm *icm;
 	struct mlx4_icm_chunk *chunk = NULL;
 	int cur_order;
+	int ret;
+
+	/* We use sg_set_buf for coherent allocs, which assumes low memory */
+	BUG_ON(coherent && (gfp_mask & __GFP_HIGHMEM));
 
 	icm = kmalloc(sizeof *icm, gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
 	if (!icm)
-		return icm;
+		return NULL;
 
 	icm->refcount = 0;
 	INIT_LIST_HEAD(&icm->chunk_list);
@@ -101,12 +153,20 @@ struct mlx4_icm *mlx4_alloc_icm(struct mlx4_dev *dev, int npages,
 		while (1 << cur_order > npages)
 			--cur_order;
 
-		chunk->mem[chunk->npages].page = alloc_pages(gfp_mask, cur_order);
-		if (chunk->mem[chunk->npages].page) {
-			chunk->mem[chunk->npages].length = PAGE_SIZE << cur_order;
-			chunk->mem[chunk->npages].offset = 0;
+		if (coherent)
+			ret = mlx4_alloc_icm_coherent(&dev->pdev->dev,
+						      &chunk->mem[chunk->npages],
+						      cur_order, gfp_mask);
+		else
+			ret = mlx4_alloc_icm_pages(&chunk->mem[chunk->npages],
+						   cur_order, gfp_mask);
+
+		if (!ret) {
+			++chunk->npages;
 
-			if (++chunk->npages == MLX4_ICM_CHUNK_LEN) {
+			if (coherent)
+				++chunk->nsg;
+			else if (chunk->npages == MLX4_ICM_CHUNK_LEN) {
 				chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
 							chunk->npages,
 							PCI_DMA_BIDIRECTIONAL);
@@ -125,7 +185,7 @@ struct mlx4_icm *mlx4_alloc_icm(struct mlx4_dev *dev, int npages,
 		}
 	}
 
-	if (chunk) {
+	if (!coherent && chunk) {
 		chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
 					chunk->npages,
 					PCI_DMA_BIDIRECTIONAL);
@@ -137,7 +197,7 @@ struct mlx4_icm *mlx4_alloc_icm(struct mlx4_dev *dev, int npages,
 	return icm;
 
 fail:
-	mlx4_free_icm(dev, icm);
+	mlx4_free_icm(dev, icm, coherent);
 	return NULL;
 }
 
@@ -202,7 +262,7 @@ int mlx4_table_get(struct mlx4_dev *dev, struct mlx4_icm_table *table, int obj)
 
 	table->icm[i] = mlx4_alloc_icm(dev, MLX4_TABLE_CHUNK_SIZE >> PAGE_SHIFT,
 				       (table->lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
-				       __GFP_NOWARN);
+				       __GFP_NOWARN, table->coherent);
 	if (!table->icm[i]) {
 		ret = -ENOMEM;
 		goto out;
@@ -210,7 +270,7 @@ int mlx4_table_get(struct mlx4_dev *dev, struct mlx4_icm_table *table, int obj)
 
 	if (mlx4_MAP_ICM(dev, table->icm[i], table->virt +
 			 (u64) i * MLX4_TABLE_CHUNK_SIZE)) {
-		mlx4_free_icm(dev, table->icm[i]);
+		mlx4_free_icm(dev, table->icm[i], table->coherent);
 		table->icm[i] = NULL;
 		ret = -ENOMEM;
 		goto out;
@@ -234,7 +294,7 @@ void mlx4_table_put(struct mlx4_dev *dev, struct mlx4_icm_table *table, int obj)
 	if (--table->icm[i]->refcount == 0) {
 		mlx4_UNMAP_ICM(dev, table->virt + i * MLX4_TABLE_CHUNK_SIZE,
 			       MLX4_TABLE_CHUNK_SIZE / MLX4_ICM_PAGE_SIZE);
-		mlx4_free_icm(dev, table->icm[i]);
+		mlx4_free_icm(dev, table->icm[i], table->coherent);
 		table->icm[i] = NULL;
 	}
 
@@ -309,7 +369,7 @@ void mlx4_table_put_range(struct mlx4_dev *dev, struct mlx4_icm_table *table,
 
 int mlx4_init_icm_table(struct mlx4_dev *dev, struct mlx4_icm_table *table,
 			u64 virt, int obj_size,	int nobj, int reserved,
-			int use_lowmem)
+			int use_lowmem, int use_coherent)
 {
 	int obj_per_chunk;
 	int num_icm;
@@ -327,6 +387,7 @@ int mlx4_init_icm_table(struct mlx4_dev *dev, struct mlx4_icm_table *table,
 	table->num_obj  = nobj;
 	table->obj_size = obj_size;
 	table->lowmem   = use_lowmem;
+	table->coherent = use_coherent;
 	mutex_init(&table->mutex);
 
 	for (i = 0; i * MLX4_TABLE_CHUNK_SIZE < reserved * obj_size; ++i) {
@@ -336,11 +397,11 @@ int mlx4_init_icm_table(struct mlx4_dev *dev, struct mlx4_icm_table *table,
 
 		table->icm[i] = mlx4_alloc_icm(dev, chunk_size >> PAGE_SHIFT,
 					       (use_lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
-					       __GFP_NOWARN);
+					       __GFP_NOWARN, use_coherent);
 		if (!table->icm[i])
 			goto err;
 		if (mlx4_MAP_ICM(dev, table->icm[i], virt + i * MLX4_TABLE_CHUNK_SIZE)) {
-			mlx4_free_icm(dev, table->icm[i]);
+			mlx4_free_icm(dev, table->icm[i], use_coherent);
 			table->icm[i] = NULL;
 			goto err;
 		}
@@ -359,7 +420,7 @@ err:
 		if (table->icm[i]) {
 			mlx4_UNMAP_ICM(dev, virt + i * MLX4_TABLE_CHUNK_SIZE,
 				       MLX4_TABLE_CHUNK_SIZE / MLX4_ICM_PAGE_SIZE);
-			mlx4_free_icm(dev, table->icm[i]);
+			mlx4_free_icm(dev, table->icm[i], use_coherent);
 		}
 
 	return -ENOMEM;
@@ -373,7 +434,7 @@ void mlx4_cleanup_icm_table(struct mlx4_dev *dev, struct mlx4_icm_table *table)
 		if (table->icm[i]) {
 			mlx4_UNMAP_ICM(dev, table->virt + i * MLX4_TABLE_CHUNK_SIZE,
 				       MLX4_TABLE_CHUNK_SIZE / MLX4_ICM_PAGE_SIZE);
-			mlx4_free_icm(dev, table->icm[i]);
+			mlx4_free_icm(dev, table->icm[i], table->coherent);
 		}
 
 	kfree(table->icm);
diff --git a/drivers/net/mlx4/icm.h b/drivers/net/mlx4/icm.h
index bea223d..a77db6d 100644
--- a/drivers/net/mlx4/icm.h
+++ b/drivers/net/mlx4/icm.h
@@ -67,8 +67,9 @@ struct mlx4_icm_iter {
 
 struct mlx4_dev;
 
-struct mlx4_icm *mlx4_alloc_icm(struct mlx4_dev *dev, int npages, gfp_t gfp_mask);
-void mlx4_free_icm(struct mlx4_dev *dev, struct mlx4_icm *icm);
+struct mlx4_icm *mlx4_alloc_icm(struct mlx4_dev *dev, int npages,
+				gfp_t gfp_mask, int coherent);
+void mlx4_free_icm(struct mlx4_dev *dev, struct mlx4_icm *icm, int coherent);
 
 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);
@@ -78,7 +79,7 @@ void mlx4_table_put_range(struct mlx4_dev *dev, struct mlx4_icm_table *table,
 			  int start, int end);
 int mlx4_init_icm_table(struct mlx4_dev *dev, struct mlx4_icm_table *table,
 			u64 virt, int obj_size,	int nobj, int reserved,
-			int use_lowmem);
+			int use_lowmem, int use_coherent);
 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);
diff --git a/drivers/net/mlx4/main.c b/drivers/net/mlx4/main.c
index 9e590e1..07c2847 100644
--- a/drivers/net/mlx4/main.c
+++ b/drivers/net/mlx4/main.c
@@ -168,7 +168,7 @@ static int __devinit mlx4_load_fw(struct mlx4_dev *dev)
 	int err;
 
 	priv->fw.fw_icm = mlx4_alloc_icm(dev, priv->fw.fw_pages,
-					 GFP_HIGHUSER | __GFP_NOWARN);
+					 GFP_HIGHUSER | __GFP_NOWARN, 0);
 	if (!priv->fw.fw_icm) {
 		mlx4_err(dev, "Couldn't allocate FW area, aborting.\n");
 		return -ENOMEM;
@@ -192,7 +192,7 @@ err_unmap_fa:
 	mlx4_UNMAP_FA(dev);
 
 err_free:
-	mlx4_free_icm(dev, priv->fw.fw_icm);
+	mlx4_free_icm(dev, priv->fw.fw_icm, 0);
 	return err;
 }
 
@@ -207,7 +207,7 @@ static int __devinit mlx4_init_cmpt_table(struct mlx4_dev *dev, u64 cmpt_base,
 				  ((u64) (MLX4_CMPT_TYPE_QP *
 					  cmpt_entry_sz) << MLX4_CMPT_SHIFT),
 				  cmpt_entry_sz, dev->caps.num_qps,
-				  dev->caps.reserved_qps, 0);
+				  dev->caps.reserved_qps, 0, 0);
 	if (err)
 		goto err;
 
@@ -216,7 +216,7 @@ static int __devinit mlx4_init_cmpt_table(struct mlx4_dev *dev, u64 cmpt_base,
 				  ((u64) (MLX4_CMPT_TYPE_SRQ *
 					  cmpt_entry_sz) << MLX4_CMPT_SHIFT),
 				  cmpt_entry_sz, dev->caps.num_srqs,
-				  dev->caps.reserved_srqs, 0);
+				  dev->caps.reserved_srqs, 0, 0);
 	if (err)
 		goto err_qp;
 
@@ -225,7 +225,7 @@ static int __devinit mlx4_init_cmpt_table(struct mlx4_dev *dev, u64 cmpt_base,
 				  ((u64) (MLX4_CMPT_TYPE_CQ *
 					  cmpt_entry_sz) << MLX4_CMPT_SHIFT),
 				  cmpt_entry_sz, dev->caps.num_cqs,
-				  dev->caps.reserved_cqs, 0);
+				  dev->caps.reserved_cqs, 0, 0);
 	if (err)
 		goto err_srq;
 
@@ -236,7 +236,7 @@ static int __devinit mlx4_init_cmpt_table(struct mlx4_dev *dev, u64 cmpt_base,
 				  cmpt_entry_sz,
 				  roundup_pow_of_two(MLX4_NUM_EQ +
 						     dev->caps.reserved_eqs),
-				  MLX4_NUM_EQ + dev->caps.reserved_eqs, 0);
+				  MLX4_NUM_EQ + dev->caps.reserved_eqs, 0, 0);
 	if (err)
 		goto err_cq;
 
@@ -275,7 +275,7 @@ static int __devinit mlx4_init_icm(struct mlx4_dev *dev,
 		 (unsigned long long) aux_pages << 2);
 
 	priv->fw.aux_icm = mlx4_alloc_icm(dev, aux_pages,
-					  GFP_HIGHUSER | __GFP_NOWARN);
+					  GFP_HIGHUSER | __GFP_NOWARN, 0);
 	if (!priv->fw.aux_icm) {
 		mlx4_err(dev, "Couldn't allocate aux memory, aborting.\n");
 		return -ENOMEM;
@@ -303,7 +303,7 @@ static int __devinit mlx4_init_icm(struct mlx4_dev *dev,
 				  init_hca->mtt_base,
 				  dev->caps.mtt_entry_sz,
 				  dev->caps.num_mtt_segs,
-				  dev->caps.reserved_mtts, 1);
+				  dev->caps.reserved_mtts, 1, 0);
 	if (err) {
 		mlx4_err(dev, "Failed to map MTT context memory, aborting.\n");
 		goto err_unmap_eq;
@@ -313,7 +313,7 @@ static int __devinit mlx4_init_icm(struct mlx4_dev *dev,
 				  init_hca->dmpt_base,
 				  dev_cap->dmpt_entry_sz,
 				  dev->caps.num_mpts,
-				  dev->caps.reserved_mrws, 1);
+				  dev->caps.reserved_mrws, 1, 1);
 	if (err) {
 		mlx4_err(dev, "Failed to map dMPT context memory, aborting.\n");
 		goto err_unmap_mtt;
@@ -323,7 +323,7 @@ static int __devinit mlx4_init_icm(struct mlx4_dev *dev,
 				  init_hca->qpc_base,
 				  dev_cap->qpc_entry_sz,
 				  dev->caps.num_qps,
-				  dev->caps.reserved_qps, 0);
+				  dev->caps.reserved_qps, 0, 0);
 	if (err) {
 		mlx4_err(dev, "Failed to map QP context memory, aborting.\n");
 		goto err_unmap_dmpt;
@@ -333,7 +333,7 @@ static int __devinit mlx4_init_icm(struct mlx4_dev *dev,
 				  init_hca->auxc_base,
 				  dev_cap->aux_entry_sz,
 				  dev->caps.num_qps,
-				  dev->caps.reserved_qps, 0);
+				  dev->caps.reserved_qps, 0, 0);
 	if (err) {
 		mlx4_err(dev, "Failed to map AUXC context memory, aborting.\n");
 		goto err_unmap_qp;
@@ -343,7 +343,7 @@ static int __devinit mlx4_init_icm(struct mlx4_dev *dev,
 				  init_hca->altc_base,
 				  dev_cap->altc_entry_sz,
 				  dev->caps.num_qps,
-				  dev->caps.reserved_qps, 0);
+				  dev->caps.reserved_qps, 0, 0);
 	if (err) {
 		mlx4_err(dev, "Failed to map ALTC context memory, aborting.\n");
 		goto err_unmap_auxc;
@@ -353,7 +353,7 @@ static int __devinit mlx4_init_icm(struct mlx4_dev *dev,
 				  init_hca->rdmarc_base,
 				  dev_cap->rdmarc_entry_sz << priv->qp_table.rdmarc_shift,
 				  dev->caps.num_qps,
-				  dev->caps.reserved_qps, 0);
+				  dev->caps.reserved_qps, 0, 0);
 	if (err) {
 		mlx4_err(dev, "Failed to map RDMARC context memory, aborting\n");
 		goto err_unmap_altc;
@@ -363,7 +363,7 @@ static int __devinit mlx4_init_icm(struct mlx4_dev *dev,
 				  init_hca->cqc_base,
 				  dev_cap->cqc_entry_sz,
 				  dev->caps.num_cqs,
-				  dev->caps.reserved_cqs, 0);
+				  dev->caps.reserved_cqs, 0, 0);
 	if (err) {
 		mlx4_err(dev, "Failed to map CQ context memory, aborting.\n");
 		goto err_unmap_rdmarc;
@@ -373,7 +373,7 @@ static int __devinit mlx4_init_icm(struct mlx4_dev *dev,
 				  init_hca->srqc_base,
 				  dev_cap->srq_entry_sz,
 				  dev->caps.num_srqs,
-				  dev->caps.reserved_srqs, 0);
+				  dev->caps.reserved_srqs, 0, 0);
 	if (err) {
 		mlx4_err(dev, "Failed to map SRQ context memory, aborting.\n");
 		goto err_unmap_cq;
@@ -388,7 +388,7 @@ static int __devinit mlx4_init_icm(struct mlx4_dev *dev,
 				  init_hca->mc_base, MLX4_MGM_ENTRY_SIZE,
 				  dev->caps.num_mgms + dev->caps.num_amgms,
 				  dev->caps.num_mgms + dev->caps.num_amgms,
-				  0);
+				  0, 0);
 	if (err) {
 		mlx4_err(dev, "Failed to map MCG context memory, aborting.\n");
 		goto err_unmap_srq;
@@ -433,7 +433,7 @@ err_unmap_aux:
 	mlx4_UNMAP_ICM_AUX(dev);
 
 err_free_aux:
-	mlx4_free_icm(dev, priv->fw.aux_icm);
+	mlx4_free_icm(dev, priv->fw.aux_icm, 0);
 
 	return err;
 }
@@ -458,7 +458,7 @@ static void mlx4_free_icms(struct mlx4_dev *dev)
 	mlx4_unmap_eq_icm(dev);
 
 	mlx4_UNMAP_ICM_AUX(dev);
-	mlx4_free_icm(dev, priv->fw.aux_icm);
+	mlx4_free_icm(dev, priv->fw.aux_icm, 0);
 }
 
 static void mlx4_close_hca(struct mlx4_dev *dev)
@@ -466,7 +466,7 @@ static void mlx4_close_hca(struct mlx4_dev *dev)
 	mlx4_CLOSE_HCA(dev, 0);
 	mlx4_free_icms(dev);
 	mlx4_UNMAP_FA(dev);
-	mlx4_free_icm(dev, mlx4_priv(dev)->fw.fw_icm);
+	mlx4_free_icm(dev, mlx4_priv(dev)->fw.fw_icm, 0);
 }
 
 static int __devinit mlx4_init_hca(struct mlx4_dev *dev)
@@ -537,7 +537,7 @@ err_free_icm:
 
 err_stop_fw:
 	mlx4_UNMAP_FA(dev);
-	mlx4_free_icm(dev, priv->fw.fw_icm);
+	mlx4_free_icm(dev, priv->fw.fw_icm, 0);
 
 	return err;
 }
diff --git a/drivers/net/mlx4/mlx4.h b/drivers/net/mlx4/mlx4.h
index b9f8397..2bad045 100644
--- a/drivers/net/mlx4/mlx4.h
+++ b/drivers/net/mlx4/mlx4.h
@@ -129,6 +129,7 @@ struct mlx4_icm_table {
 	int			num_obj;
 	int			obj_size;
 	int			lowmem;
+	int			coherent;
 	struct mutex		mutex;
 	struct mlx4_icm	      **icm;
 };