IB/mthca: Fix access to MTT and MPT tables on non-cache-coherent CPUs

We allocate the MTT table with alloc_pages() and then do pci_map_sg(),
so we must call pci_dma_sync_sg() after the CPU writes to the MTT
table.  This works since the device will never write MTTs on mem-free
HCAs, once we get rid of the use of the WRITE_MTT firmware command.
This change is needed to make that work, and is an improvement for
now, since it gives FMRs a chance at working.

For MPTs, both the device and CPU might write there, so we must
allocate DMA coherent memory for these.
Signed-off-by: Michael S. Tsirkin <mst@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>

drivers/infiniband/hw/mthca/mthca_main.c

@@ -379,7 +379,7 @@ static int mthca_load_fw(struct mthca_dev *mdev)
 
 	mdev->fw.arbel.fw_icm =
 		mthca_alloc_icm(mdev, mdev->fw.arbel.fw_pages,
-				GFP_HIGHUSER | __GFP_NOWARN);
+				GFP_HIGHUSER | __GFP_NOWARN, 0);
 	if (!mdev->fw.arbel.fw_icm) {
 		mthca_err(mdev, "Couldn't allocate FW area, aborting.\n");
 		return -ENOMEM;
@@ -412,7 +412,7 @@ static int mthca_load_fw(struct mthca_dev *mdev)
 	mthca_UNMAP_FA(mdev, &status);
 
 err_free:
-	mthca_free_icm(mdev, mdev->fw.arbel.fw_icm);
+	mthca_free_icm(mdev, mdev->fw.arbel.fw_icm, 0);
 	return err;
 }
 
@@ -441,7 +441,7 @@ static int mthca_init_icm(struct mthca_dev *mdev,
 		  (unsigned long long) aux_pages << 2);
 
 	mdev->fw.arbel.aux_icm = mthca_alloc_icm(mdev, aux_pages,
-						 GFP_HIGHUSER | __GFP_NOWARN);
+						 GFP_HIGHUSER | __GFP_NOWARN, 0);
 	if (!mdev->fw.arbel.aux_icm) {
 		mthca_err(mdev, "Couldn't allocate aux memory, aborting.\n");
 		return -ENOMEM;
@@ -471,7 +471,8 @@ static int mthca_init_icm(struct mthca_dev *mdev,
 	mdev->mr_table.mtt_table = mthca_alloc_icm_table(mdev, init_hca->mtt_base,
 							 MTHCA_MTT_SEG_SIZE,
 							 mdev->limits.num_mtt_segs,
-							 mdev->limits.reserved_mtts, 1);
+							 mdev->limits.reserved_mtts,
+							 1, 0);
 	if (!mdev->mr_table.mtt_table) {
 		mthca_err(mdev, "Failed to map MTT context memory, aborting.\n");
 		err = -ENOMEM;
@@ -481,7 +482,8 @@ static int mthca_init_icm(struct mthca_dev *mdev,
 	mdev->mr_table.mpt_table = mthca_alloc_icm_table(mdev, init_hca->mpt_base,
 							 dev_lim->mpt_entry_sz,
 							 mdev->limits.num_mpts,
-							 mdev->limits.reserved_mrws, 1);
+							 mdev->limits.reserved_mrws,
+							 1, 1);
 	if (!mdev->mr_table.mpt_table) {
 		mthca_err(mdev, "Failed to map MPT context memory, aborting.\n");
 		err = -ENOMEM;
@@ -491,7 +493,8 @@ static int mthca_init_icm(struct mthca_dev *mdev,
 	mdev->qp_table.qp_table = mthca_alloc_icm_table(mdev, init_hca->qpc_base,
 							dev_lim->qpc_entry_sz,
 							mdev->limits.num_qps,
-							mdev->limits.reserved_qps, 0);
+							mdev->limits.reserved_qps,
+							0, 0);
 	if (!mdev->qp_table.qp_table) {
 		mthca_err(mdev, "Failed to map QP context memory, aborting.\n");
 		err = -ENOMEM;
@@ -501,7 +504,8 @@ static int mthca_init_icm(struct mthca_dev *mdev,
 	mdev->qp_table.eqp_table = mthca_alloc_icm_table(mdev, init_hca->eqpc_base,
 							 dev_lim->eqpc_entry_sz,
 							 mdev->limits.num_qps,
-							 mdev->limits.reserved_qps, 0);
+							 mdev->limits.reserved_qps,
+							 0, 0);
 	if (!mdev->qp_table.eqp_table) {
 		mthca_err(mdev, "Failed to map EQP context memory, aborting.\n");
 		err = -ENOMEM;
@@ -511,7 +515,7 @@ static int mthca_init_icm(struct mthca_dev *mdev,
 	mdev->qp_table.rdb_table = mthca_alloc_icm_table(mdev, init_hca->rdb_base,
 							 MTHCA_RDB_ENTRY_SIZE,
 							 mdev->limits.num_qps <<
-							 mdev->qp_table.rdb_shift,
+							 mdev->qp_table.rdb_shift, 0,
 							 0, 0);
 	if (!mdev->qp_table.rdb_table) {
 		mthca_err(mdev, "Failed to map RDB context memory, aborting\n");
@@ -522,7 +526,8 @@ static int mthca_init_icm(struct mthca_dev *mdev,
        mdev->cq_table.table = mthca_alloc_icm_table(mdev, init_hca->cqc_base,
 						    dev_lim->cqc_entry_sz,
 						    mdev->limits.num_cqs,
-						    mdev->limits.reserved_cqs, 0);
+						    mdev->limits.reserved_cqs,
+						    0, 0);
 	if (!mdev->cq_table.table) {
 		mthca_err(mdev, "Failed to map CQ context memory, aborting.\n");
 		err = -ENOMEM;
@@ -534,7 +539,8 @@ static int mthca_init_icm(struct mthca_dev *mdev,
 			mthca_alloc_icm_table(mdev, init_hca->srqc_base,
 					      dev_lim->srq_entry_sz,
 					      mdev->limits.num_srqs,
-					      mdev->limits.reserved_srqs, 0);
+					      mdev->limits.reserved_srqs,
+					      0, 0);
 		if (!mdev->srq_table.table) {
 			mthca_err(mdev, "Failed to map SRQ context memory, "
 				  "aborting.\n");
@@ -554,7 +560,7 @@ static int mthca_init_icm(struct mthca_dev *mdev,
 						      mdev->limits.num_amgms,
 						      mdev->limits.num_mgms +
 						      mdev->limits.num_amgms,
-						      0);
+						      0, 0);
 	if (!mdev->mcg_table.table) {
 		mthca_err(mdev, "Failed to map MCG context memory, aborting.\n");
 		err = -ENOMEM;
@@ -592,7 +598,7 @@ static int mthca_init_icm(struct mthca_dev *mdev,
 	mthca_UNMAP_ICM_AUX(mdev, &status);
 
 err_free_aux:
-	mthca_free_icm(mdev, mdev->fw.arbel.aux_icm);
+	mthca_free_icm(mdev, mdev->fw.arbel.aux_icm, 0);
 
 	return err;
 }
@@ -613,7 +619,7 @@ static void mthca_free_icms(struct mthca_dev *mdev)
 	mthca_unmap_eq_icm(mdev);
 
 	mthca_UNMAP_ICM_AUX(mdev, &status);
-	mthca_free_icm(mdev, mdev->fw.arbel.aux_icm);
+	mthca_free_icm(mdev, mdev->fw.arbel.aux_icm, 0);
 }
 
 static int mthca_init_arbel(struct mthca_dev *mdev)
@@ -697,7 +703,7 @@ static int mthca_init_arbel(struct mthca_dev *mdev)
 
 err_stop_fw:
 	mthca_UNMAP_FA(mdev, &status);
-	mthca_free_icm(mdev, mdev->fw.arbel.fw_icm);
+	mthca_free_icm(mdev, mdev->fw.arbel.fw_icm, 0);
 
 err_disable:
 	if (!(mdev->mthca_flags & MTHCA_FLAG_NO_LAM))
@@ -716,7 +722,7 @@ static void mthca_close_hca(struct mthca_dev *mdev)
 		mthca_free_icms(mdev);
 
 		mthca_UNMAP_FA(mdev, &status);
-		mthca_free_icm(mdev, mdev->fw.arbel.fw_icm);
+		mthca_free_icm(mdev, mdev->fw.arbel.fw_icm, 0);
 
 		if (!(mdev->mthca_flags & MTHCA_FLAG_NO_LAM))
 			mthca_DISABLE_LAM(mdev, &status);

drivers/infiniband/hw/mthca/mthca_memfree.c

@@ -35,6 +35,9 @@
  */
 
 #include <linux/mm.h>
+#include <linux/scatterlist.h>
+
+#include <asm/page.h>
 
 #include "mthca_memfree.h"
 #include "mthca_dev.h"
@@ -58,22 +61,42 @@ struct mthca_user_db_table {
 	}                page[0];
 };
 
-void mthca_free_icm(struct mthca_dev *dev, struct mthca_icm *icm)
+static void mthca_free_icm_pages(struct mthca_dev *dev, struct mthca_icm_chunk *chunk)
+{
+	int i;
+
+	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));
+}
+
+static void mthca_free_icm_coherent(struct mthca_dev *dev, struct mthca_icm_chunk *chunk)
 {
-	struct mthca_icm_chunk *chunk, *tmp;
 	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 mthca_free_icm(struct mthca_dev *dev, struct mthca_icm *icm, int coherent)
+{
+	struct mthca_icm_chunk *chunk, *tmp;
+
 	if (!icm)
 		return;
 
 	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);
-
-		for (i = 0; i < chunk->npages; ++i)
-			__free_pages(chunk->mem[i].page,
-				     get_order(chunk->mem[i].length));
+		if (coherent)
+			mthca_free_icm_coherent(dev, chunk);
+		else
+			mthca_free_icm_pages(dev, chunk);
 
 		kfree(chunk);
 	}
@@ -81,12 +104,41 @@ void mthca_free_icm(struct mthca_dev *dev, struct mthca_icm *icm)
 	kfree(icm);
 }
 
+static int mthca_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 mthca_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 mthca_icm *mthca_alloc_icm(struct mthca_dev *dev, int npages,
-				  gfp_t gfp_mask)
+				  gfp_t gfp_mask, int coherent)
 {
 	struct mthca_icm *icm;
 	struct mthca_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)
@@ -112,21 +164,28 @@ struct mthca_icm *mthca_alloc_icm(struct mthca_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 = mthca_alloc_icm_coherent(&dev->pdev->dev,
+						       &chunk->mem[chunk->npages],
+						       cur_order, gfp_mask);
+		else
+			ret = mthca_alloc_icm_pages(&chunk->mem[chunk->npages],
+						    cur_order, gfp_mask);
 
-			if (++chunk->npages == MTHCA_ICM_CHUNK_LEN) {
+		if (!ret) {
+			++chunk->npages;
+
+			if (!coherent && chunk->npages == MTHCA_ICM_CHUNK_LEN) {
 				chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
 							chunk->npages,
 							PCI_DMA_BIDIRECTIONAL);
 
 				if (chunk->nsg <= 0)
 					goto fail;
+			}
 
+			if (chunk->npages == MTHCA_ICM_CHUNK_LEN)
 				chunk = NULL;
-			}
 
 			npages -= 1 << cur_order;
 		} else {
@@ -136,7 +195,7 @@ struct mthca_icm *mthca_alloc_icm(struct mthca_dev *dev, int npages,
 		}
 	}
 
-	if (chunk) {
+	if (!coherent && chunk) {
 		chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
 					chunk->npages,
 					PCI_DMA_BIDIRECTIONAL);
@@ -148,7 +207,7 @@ struct mthca_icm *mthca_alloc_icm(struct mthca_dev *dev, int npages,
 	return icm;
 
 fail:
-	mthca_free_icm(dev, icm);
+	mthca_free_icm(dev, icm, coherent);
 	return NULL;
 }
 
@@ -167,7 +226,7 @@ int mthca_table_get(struct mthca_dev *dev, struct mthca_icm_table *table, int ob
 
 	table->icm[i] = mthca_alloc_icm(dev, MTHCA_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;
@@ -175,7 +234,7 @@ int mthca_table_get(struct mthca_dev *dev, struct mthca_icm_table *table, int ob
 
 	if (mthca_MAP_ICM(dev, table->icm[i], table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
 			  &status) || status) {
-		mthca_free_icm(dev, table->icm[i]);
+		mthca_free_icm(dev, table->icm[i], table->coherent);
 		table->icm[i] = NULL;
 		ret = -ENOMEM;
 		goto out;
@@ -204,16 +263,16 @@ void mthca_table_put(struct mthca_dev *dev, struct mthca_icm_table *table, int o
 		mthca_UNMAP_ICM(dev, table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
 				MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE,
 				&status);
-		mthca_free_icm(dev, table->icm[i]);
+		mthca_free_icm(dev, table->icm[i], table->coherent);
 		table->icm[i] = NULL;
 	}
 
 	mutex_unlock(&table->mutex);
 }
 
-void *mthca_table_find(struct mthca_icm_table *table, int obj)
+void *mthca_table_find(struct mthca_icm_table *table, int obj, dma_addr_t *dma_handle)
 {
-	int idx, offset, i;
+	int idx, offset, dma_offset, i;
 	struct mthca_icm_chunk *chunk;
 	struct mthca_icm *icm;
 	struct page *page = NULL;
@@ -225,13 +284,22 @@ void *mthca_table_find(struct mthca_icm_table *table, int obj)
 
 	idx = (obj & (table->num_obj - 1)) * table->obj_size;
 	icm = table->icm[idx / MTHCA_TABLE_CHUNK_SIZE];
-	offset = idx % MTHCA_TABLE_CHUNK_SIZE;
+	dma_offset = offset = idx % MTHCA_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;
@@ -283,7 +351,7 @@ void mthca_table_put_range(struct mthca_dev *dev, struct mthca_icm_table *table,
 struct mthca_icm_table *mthca_alloc_icm_table(struct mthca_dev *dev,
 					      u64 virt, int obj_size,
 					      int nobj, int reserved,
-					      int use_lowmem)
+					      int use_lowmem, int use_coherent)
 {
 	struct mthca_icm_table *table;
 	int num_icm;
@@ -302,6 +370,7 @@ struct mthca_icm_table *mthca_alloc_icm_table(struct mthca_dev *dev,
 	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 < num_icm; ++i)
@@ -314,12 +383,12 @@ struct mthca_icm_table *mthca_alloc_icm_table(struct mthca_dev *dev,
 
 		table->icm[i] = mthca_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 (mthca_MAP_ICM(dev, table->icm[i], virt + i * MTHCA_TABLE_CHUNK_SIZE,
 				  &status) || status) {
-			mthca_free_icm(dev, table->icm[i]);
+			mthca_free_icm(dev, table->icm[i], table->coherent);
 			table->icm[i] = NULL;
 			goto err;
 		}
@@ -339,7 +408,7 @@ struct mthca_icm_table *mthca_alloc_icm_table(struct mthca_dev *dev,
 			mthca_UNMAP_ICM(dev, virt + i * MTHCA_TABLE_CHUNK_SIZE,
 					MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE,
 				        &status);
-			mthca_free_icm(dev, table->icm[i]);
+			mthca_free_icm(dev, table->icm[i], table->coherent);
 		}
 
 	kfree(table);
@@ -357,7 +426,7 @@ void mthca_free_icm_table(struct mthca_dev *dev, struct mthca_icm_table *table)
 			mthca_UNMAP_ICM(dev, table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
 					MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE,
 					&status);
-			mthca_free_icm(dev, table->icm[i]);
+			mthca_free_icm(dev, table->icm[i], table->coherent);
 		}
 
 	kfree(table);

drivers/infiniband/hw/mthca/mthca_memfree.h

@@ -69,6 +69,7 @@ struct mthca_icm_table {
 	int               num_obj;
 	int               obj_size;
 	int               lowmem;
+	int               coherent;
 	struct mutex      mutex;
 	struct mthca_icm *icm[0];
 };
@@ -82,17 +83,17 @@ struct mthca_icm_iter {
 struct mthca_dev;
 
 struct mthca_icm *mthca_alloc_icm(struct mthca_dev *dev, int npages,
-				  gfp_t gfp_mask);
-void mthca_free_icm(struct mthca_dev *dev, struct mthca_icm *icm);
+				  gfp_t gfp_mask, int coherent);
+void mthca_free_icm(struct mthca_dev *dev, struct mthca_icm *icm, int coherent);
 
 struct mthca_icm_table *mthca_alloc_icm_table(struct mthca_dev *dev,
 					      u64 virt, int obj_size,
 					      int nobj, int reserved,
-					      int use_lowmem);
+					      int use_lowmem, int use_coherent);
 void mthca_free_icm_table(struct mthca_dev *dev, struct mthca_icm_table *table);
 int mthca_table_get(struct mthca_dev *dev, struct mthca_icm_table *table, int obj);
 void mthca_table_put(struct mthca_dev *dev, struct mthca_icm_table *table, int obj);
-void *mthca_table_find(struct mthca_icm_table *table, int obj);
+void *mthca_table_find(struct mthca_icm_table *table, int obj, dma_addr_t *dma_handle);
 int mthca_table_get_range(struct mthca_dev *dev, struct mthca_icm_table *table,
 			  int start, int end);
 void mthca_table_put_range(struct mthca_dev *dev, struct mthca_icm_table *table,

drivers/infiniband/hw/mthca/mthca_mr.c

@@ -524,7 +524,7 @@ int mthca_fmr_alloc(struct mthca_dev *dev, u32 pd,
 		if (err)
 			goto err_out_mpt_free;
 
-		mr->mem.arbel.mpt = mthca_table_find(dev->mr_table.mpt_table, key);
+		mr->mem.arbel.mpt = mthca_table_find(dev->mr_table.mpt_table, key, NULL);
 		BUG_ON(!mr->mem.arbel.mpt);
 	} else
 		mr->mem.tavor.mpt = dev->mr_table.tavor_fmr.mpt_base +
@@ -538,7 +538,8 @@ int mthca_fmr_alloc(struct mthca_dev *dev, u32 pd,
 
 	if (mthca_is_memfree(dev)) {
 		mr->mem.arbel.mtts = mthca_table_find(dev->mr_table.mtt_table,
-						      mr->mtt->first_seg);
+						      mr->mtt->first_seg,
+						      &mr->mem.arbel.dma_handle);
 		BUG_ON(!mr->mem.arbel.mtts);
 	} else
 		mr->mem.tavor.mtts = dev->mr_table.tavor_fmr.mtt_base + mtt_seg;
@@ -712,6 +713,9 @@ int mthca_arbel_map_phys_fmr(struct ib_fmr *ibfmr, u64 *page_list,
 		fmr->mem.arbel.mtts[i] = cpu_to_be64(page_list[i] |
 						     MTHCA_MTT_FLAG_PRESENT);
 
+	dma_sync_single(&dev->pdev->dev, fmr->mem.arbel.dma_handle,
+			list_len * sizeof(u64), DMA_TO_DEVICE);
+
 	fmr->mem.arbel.mpt->key    = cpu_to_be32(key);
 	fmr->mem.arbel.mpt->lkey   = cpu_to_be32(key);
 	fmr->mem.arbel.mpt->length = cpu_to_be64(list_len * (1ull << fmr->attr.page_shift));

drivers/infiniband/hw/mthca/mthca_provider.h

@@ -89,6 +89,7 @@ struct mthca_fmr {
 		struct {
 			struct mthca_mpt_entry *mpt;
 			__be64 *mtts;
+			dma_addr_t dma_handle;
 		} arbel;
 	} mem;
 };