1 files changed, 361 insertions, 271 deletions

diff --git a/drivers/infiniband/hw/mthca/mthca_mr.c b/drivers/infiniband/hw/mthca/mthca_mr.c
index 8960fc2306b..ed9a989e501 100644
--- a/drivers/infiniband/hw/mthca/mthca_mr.c
+++ b/drivers/infiniband/hw/mthca/mthca_mr.c
@@ -1,5 +1,6 @@
 /*
  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
  *
  * This software is available to you under a choice of one of two
  * licenses.  You may choose to be licensed under the terms of the GNU
@@ -28,34 +29,37 @@
  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
- *
- * $Id: mthca_mr.c 1349 2004-12-16 21:09:43Z roland $
  */
 
 #include <linux/slab.h>
-#include <linux/init.h>
 #include <linux/errno.h>
 
 #include "mthca_dev.h"
 #include "mthca_cmd.h"
 #include "mthca_memfree.h"
 
+struct mthca_mtt {
+	struct mthca_buddy *buddy;
+	int                 order;
+	u32                 first_seg;
+};
+
 /*
  * Must be packed because mtt_seg is 64 bits but only aligned to 32 bits.
  */
 struct mthca_mpt_entry {
-	u32 flags;
-	u32 page_size;
-	u32 key;
-	u32 pd;
-	u64 start;
-	u64 length;
-	u32 lkey;
-	u32 window_count;
-	u32 window_count_limit;
-	u64 mtt_seg;
-	u32 mtt_sz;		/* Arbel only */
-	u32 reserved[2];
+	__be32 flags;
+	__be32 page_size;
+	__be32 key;
+	__be32 pd;
+	__be64 start;
+	__be64 length;
+	__be32 lkey;
+	__be32 window_count;
+	__be32 window_count_limit;
+	__be64 mtt_seg;
+	__be32 mtt_sz;		/* Arbel only */
+	u32    reserved[2];
 } __attribute__((packed));
 
 #define MTHCA_MPT_FLAG_SW_OWNS       (0xfUL << 28)
@@ -69,6 +73,8 @@ struct mthca_mpt_entry {
 #define MTHCA_MPT_STATUS_SW 0xF0
 #define MTHCA_MPT_STATUS_HW 0x00
 
+#define SINAI_FMR_KEY_INC 0x1000000
+
 /*
  * Buddy allocator for MTT segments (currently not very efficient
  * since it doesn't keep a free list and just searches linearly
@@ -83,23 +89,26 @@ static u32 mthca_buddy_alloc(struct mthca_buddy *buddy, int order)
 
 	spin_lock(&buddy->lock);
 
-	for (o = order; o <= buddy->max_order; ++o) {
-		m = 1 << (buddy->max_order - o);
-		seg = find_first_bit(buddy->bits[o], m);
-		if (seg < m)
-			goto found;
-	}
+	for (o = order; o <= buddy->max_order; ++o)
+		if (buddy->num_free[o]) {
+			m = 1 << (buddy->max_order - o);
+			seg = find_first_bit(buddy->bits[o], m);
+			if (seg < m)
+				goto found;
+		}
 
 	spin_unlock(&buddy->lock);
 	return -1;
 
  found:
 	clear_bit(seg, buddy->bits[o]);
+	--buddy->num_free[o];
 
 	while (o > order) {
 		--o;
 		seg <<= 1;
 		set_bit(seg ^ 1, buddy->bits[o]);
+		++buddy->num_free[o];
 	}
 
 	spin_unlock(&buddy->lock);
@@ -117,29 +126,31 @@ static void mthca_buddy_free(struct mthca_buddy *buddy, u32 seg, int order)
 
 	while (test_bit(seg ^ 1, buddy->bits[order])) {
 		clear_bit(seg ^ 1, buddy->bits[order]);
+		--buddy->num_free[order];
 		seg >>= 1;
 		++order;
 	}
 
 	set_bit(seg, buddy->bits[order]);
+	++buddy->num_free[order];
 
 	spin_unlock(&buddy->lock);
 }
 
-static int __devinit mthca_buddy_init(struct mthca_buddy *buddy, int max_order)
+static int mthca_buddy_init(struct mthca_buddy *buddy, int max_order)
 {
 	int i, s;
 
 	buddy->max_order = max_order;
 	spin_lock_init(&buddy->lock);
 
-	buddy->bits = kmalloc((buddy->max_order + 1) * sizeof (long *),
+	buddy->bits = kzalloc((buddy->max_order + 1) * sizeof (long *),
 			      GFP_KERNEL);
-	if (!buddy->bits)
+	buddy->num_free = kcalloc((buddy->max_order + 1), sizeof *buddy->num_free,
+				  GFP_KERNEL);
+	if (!buddy->bits || !buddy->num_free)
 		goto err_out;
 
-	memset(buddy->bits, 0, (buddy->max_order + 1) * sizeof (long *));
-
 	for (i = 0; i <= buddy->max_order; ++i) {
 		s = BITS_TO_LONGS(1 << (buddy->max_order - i));
 		buddy->bits[i] = kmalloc(s * sizeof (long), GFP_KERNEL);
@@ -150,6 +161,7 @@ static int __devinit mthca_buddy_init(struct mthca_buddy *buddy, int max_order)
 	}
 
 	set_bit(0, buddy->bits[buddy->max_order]);
+	buddy->num_free[buddy->max_order] = 1;
 
 	return 0;
 
@@ -157,13 +169,14 @@ err_out_free:
 	for (i = 0; i <= buddy->max_order; ++i)
 		kfree(buddy->bits[i]);
 
+err_out:
 	kfree(buddy->bits);
+	kfree(buddy->num_free);
 
-err_out:
 	return -ENOMEM;
 }
 
-static void __devexit mthca_buddy_cleanup(struct mthca_buddy *buddy)
+static void mthca_buddy_cleanup(struct mthca_buddy *buddy)
 {
 	int i;
 
@@ -171,10 +184,11 @@ static void __devexit mthca_buddy_cleanup(struct mthca_buddy *buddy)
 		kfree(buddy->bits[i]);
 
 	kfree(buddy->bits);
+	kfree(buddy->num_free);
 }
 
-static u32 mthca_alloc_mtt(struct mthca_dev *dev, int order,
-			   struct mthca_buddy *buddy)
+static u32 mthca_alloc_mtt_range(struct mthca_dev *dev, int order,
+				 struct mthca_buddy *buddy)
 {
 	u32 seg = mthca_buddy_alloc(buddy, order);
 
@@ -191,14 +205,181 @@ static u32 mthca_alloc_mtt(struct mthca_dev *dev, int order,
 	return seg;
 }
 
-static void mthca_free_mtt(struct mthca_dev *dev, u32 seg, int order,
-			   struct mthca_buddy* buddy)
+static struct mthca_mtt *__mthca_alloc_mtt(struct mthca_dev *dev, int size,
+					   struct mthca_buddy *buddy)
 {
-	mthca_buddy_free(buddy, seg, order);
+	struct mthca_mtt *mtt;
+	int i;
 
-	if (mthca_is_memfree(dev))
-		mthca_table_put_range(dev, dev->mr_table.mtt_table, seg,
-				      seg + (1 << order) - 1);
+	if (size <= 0)
+		return ERR_PTR(-EINVAL);
+
+	mtt = kmalloc(sizeof *mtt, GFP_KERNEL);
+	if (!mtt)
+		return ERR_PTR(-ENOMEM);
+
+	mtt->buddy = buddy;
+	mtt->order = 0;
+	for (i = dev->limits.mtt_seg_size / 8; i < size; i <<= 1)
+		++mtt->order;
+
+	mtt->first_seg = mthca_alloc_mtt_range(dev, mtt->order, buddy);
+	if (mtt->first_seg == -1) {
+		kfree(mtt);
+		return ERR_PTR(-ENOMEM);
+	}
+
+	return mtt;
+}
+
+struct mthca_mtt *mthca_alloc_mtt(struct mthca_dev *dev, int size)
+{
+	return __mthca_alloc_mtt(dev, size, &dev->mr_table.mtt_buddy);
+}
+
+void mthca_free_mtt(struct mthca_dev *dev, struct mthca_mtt *mtt)
+{
+	if (!mtt)
+		return;
+
+	mthca_buddy_free(mtt->buddy, mtt->first_seg, mtt->order);
+
+	mthca_table_put_range(dev, dev->mr_table.mtt_table,
+			      mtt->first_seg,
+			      mtt->first_seg + (1 << mtt->order) - 1);
+
+	kfree(mtt);
+}
+
+static int __mthca_write_mtt(struct mthca_dev *dev, struct mthca_mtt *mtt,
+			     int start_index, u64 *buffer_list, int list_len)
+{
+	struct mthca_mailbox *mailbox;
+	__be64 *mtt_entry;
+	int err = 0;
+	int i;
+
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		return PTR_ERR(mailbox);
+	mtt_entry = mailbox->buf;
+
+	while (list_len > 0) {
+		mtt_entry[0] = cpu_to_be64(dev->mr_table.mtt_base +
+					   mtt->first_seg * dev->limits.mtt_seg_size +
+					   start_index * 8);
+		mtt_entry[1] = 0;
+		for (i = 0; i < list_len && i < MTHCA_MAILBOX_SIZE / 8 - 2; ++i)
+			mtt_entry[i + 2] = cpu_to_be64(buffer_list[i] |
+						       MTHCA_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 = mthca_WRITE_MTT(dev, mailbox, (i + 1) & ~1);
+		if (err) {
+			mthca_warn(dev, "WRITE_MTT failed (%d)\n", err);
+			goto out;
+		}
+
+		list_len    -= i;
+		start_index += i;
+		buffer_list += i;
+	}
+
+out:
+	mthca_free_mailbox(dev, mailbox);
+	return err;
+}
+
+int mthca_write_mtt_size(struct mthca_dev *dev)
+{
+	if (dev->mr_table.fmr_mtt_buddy != &dev->mr_table.mtt_buddy ||
+	    !(dev->mthca_flags & MTHCA_FLAG_FMR))
+		/*
+		 * Be friendly to WRITE_MTT command
+		 * and leave two empty slots for the
+		 * index and reserved fields of the
+		 * mailbox.
+		 */
+		return PAGE_SIZE / sizeof (u64) - 2;
+
+	/* For Arbel, all MTTs must fit in the same page. */
+	return mthca_is_memfree(dev) ? (PAGE_SIZE / sizeof (u64)) : 0x7ffffff;
+}
+
+static void mthca_tavor_write_mtt_seg(struct mthca_dev *dev,
+				      struct mthca_mtt *mtt, int start_index,
+				      u64 *buffer_list, int list_len)
+{
+	u64 __iomem *mtts;
+	int i;
+
+	mtts = dev->mr_table.tavor_fmr.mtt_base + mtt->first_seg * dev->limits.mtt_seg_size +
+		start_index * sizeof (u64);
+	for (i = 0; i < list_len; ++i)
+		mthca_write64_raw(cpu_to_be64(buffer_list[i] | MTHCA_MTT_FLAG_PRESENT),
+				  mtts + i);
+}
+
+static void mthca_arbel_write_mtt_seg(struct mthca_dev *dev,
+				      struct mthca_mtt *mtt, int start_index,
+				      u64 *buffer_list, int list_len)
+{
+	__be64 *mtts;
+	dma_addr_t dma_handle;
+	int i;
+	int s = start_index * sizeof (u64);
+
+	/* For Arbel, all MTTs must fit in the same page. */
+	BUG_ON(s / PAGE_SIZE != (s + list_len * sizeof(u64) - 1) / PAGE_SIZE);
+	/* Require full segments */
+	BUG_ON(s % dev->limits.mtt_seg_size);
+
+	mtts = mthca_table_find(dev->mr_table.mtt_table, mtt->first_seg +
+				s / dev->limits.mtt_seg_size, &dma_handle);
+
+	BUG_ON(!mtts);
+
+	dma_sync_single_for_cpu(&dev->pdev->dev, dma_handle,
+				list_len * sizeof (u64), DMA_TO_DEVICE);
+
+	for (i = 0; i < list_len; ++i)
+		mtts[i] = cpu_to_be64(buffer_list[i] | MTHCA_MTT_FLAG_PRESENT);
+
+	dma_sync_single_for_device(&dev->pdev->dev, dma_handle,
+				   list_len * sizeof (u64), DMA_TO_DEVICE);
+}
+
+int mthca_write_mtt(struct mthca_dev *dev, struct mthca_mtt *mtt,
+		    int start_index, u64 *buffer_list, int list_len)
+{
+	int size = mthca_write_mtt_size(dev);
+	int chunk;
+
+	if (dev->mr_table.fmr_mtt_buddy != &dev->mr_table.mtt_buddy ||
+	    !(dev->mthca_flags & MTHCA_FLAG_FMR))
+		return __mthca_write_mtt(dev, mtt, start_index, buffer_list, list_len);
+
+	while (list_len > 0) {
+		chunk = min(size, list_len);
+		if (mthca_is_memfree(dev))
+			mthca_arbel_write_mtt_seg(dev, mtt, start_index,
+						  buffer_list, chunk);
+		else
+			mthca_tavor_write_mtt_seg(dev, mtt, start_index,
+						  buffer_list, chunk);
+
+		list_len    -= chunk;
+		start_index += chunk;
+		buffer_list += chunk;
+	}
+
+	return 0;
 }
 
 static inline u32 tavor_hw_index_to_key(u32 ind)
@@ -237,96 +418,29 @@ static inline u32 key_to_hw_index(struct mthca_dev *dev, u32 key)
 		return tavor_key_to_hw_index(key);
 }
 
-int mthca_mr_alloc_notrans(struct mthca_dev *dev, u32 pd,
-			   u32 access, struct mthca_mr *mr)
+static inline u32 adjust_key(struct mthca_dev *dev, u32 key)
 {
-	void *mailbox = NULL;
-	struct mthca_mpt_entry *mpt_entry;
-	u32 key;
-	int err;
-	u8 status;
-
-	might_sleep();
-
-	mr->order = -1;
-	key = mthca_alloc(&dev->mr_table.mpt_alloc);
-	if (key == -1)
-		return -ENOMEM;
-	mr->ibmr.rkey = mr->ibmr.lkey = hw_index_to_key(dev, key);
-
-	if (mthca_is_memfree(dev)) {
-		err = mthca_table_get(dev, dev->mr_table.mpt_table, key);
-		if (err)
-			goto err_out_mpt_free;
-	}
-
-	mailbox = kmalloc(sizeof *mpt_entry + MTHCA_CMD_MAILBOX_EXTRA,
-			  GFP_KERNEL);
-	if (!mailbox) {
-		err = -ENOMEM;
-		goto err_out_table;
-	}
-	mpt_entry = MAILBOX_ALIGN(mailbox);
-
-	mpt_entry->flags = cpu_to_be32(MTHCA_MPT_FLAG_SW_OWNS     |
-				       MTHCA_MPT_FLAG_MIO         |
-				       MTHCA_MPT_FLAG_PHYSICAL    |
-				       MTHCA_MPT_FLAG_REGION      |
-				       access);
-	mpt_entry->page_size = 0;
-	mpt_entry->key       = cpu_to_be32(key);
-	mpt_entry->pd        = cpu_to_be32(pd);
-	mpt_entry->start     = 0;
-	mpt_entry->length    = ~0ULL;
-
-	memset(&mpt_entry->lkey, 0,
-	       sizeof *mpt_entry - offsetof(struct mthca_mpt_entry, lkey));
-
-	err = mthca_SW2HW_MPT(dev, mpt_entry,
-			      key & (dev->limits.num_mpts - 1),
-			      &status);
-	if (err) {
-		mthca_warn(dev, "SW2HW_MPT failed (%d)\n", err);
-		goto err_out_table;
-	} else if (status) {
-		mthca_warn(dev, "SW2HW_MPT returned status 0x%02x\n",
-			   status);
-		err = -EINVAL;
-		goto err_out_table;
-	}
-
-	kfree(mailbox);
-	return err;
-
-err_out_table:
-	if (mthca_is_memfree(dev))
-		mthca_table_put(dev, dev->mr_table.mpt_table, key);
-
-err_out_mpt_free:
-	mthca_free(&dev->mr_table.mpt_alloc, key);
-	kfree(mailbox);
-	return err;
+	if (dev->mthca_flags & MTHCA_FLAG_SINAI_OPT)
+		return ((key << 20) & 0x800000) | (key & 0x7fffff);
+	else
+		return key;
 }
 
-int mthca_mr_alloc_phys(struct mthca_dev *dev, u32 pd,
-			u64 *buffer_list, int buffer_size_shift,
-			int list_len, u64 iova, u64 total_size,
-			u32 access, struct mthca_mr *mr)
+int mthca_mr_alloc(struct mthca_dev *dev, u32 pd, int buffer_size_shift,
+		   u64 iova, u64 total_size, u32 access, struct mthca_mr *mr)
 {
-	void *mailbox;
-	u64 *mtt_entry;
+	struct mthca_mailbox *mailbox;
 	struct mthca_mpt_entry *mpt_entry;
 	u32 key;
-	int err = -ENOMEM;
-	u8 status;
 	int i;
+	int err;
 
-	might_sleep();
 	WARN_ON(buffer_size_shift >= 32);
 
 	key = mthca_alloc(&dev->mr_table.mpt_alloc);
 	if (key == -1)
 		return -ENOMEM;
+	key = adjust_key(dev, key);
 	mr->ibmr.rkey = mr->ibmr.lkey = hw_index_to_key(dev, key);
 
 	if (mthca_is_memfree(dev)) {
@@ -335,126 +449,103 @@ int mthca_mr_alloc_phys(struct mthca_dev *dev, u32 pd,
 			goto err_out_mpt_free;
 	}
 
-	for (i = MTHCA_MTT_SEG_SIZE / 8, mr->order = 0;
-	     i < list_len;
-	     i <<= 1, ++mr->order)
-		; /* nothing */
-
-	mr->first_seg = mthca_alloc_mtt(dev, mr->order,
-				       	&dev->mr_table.mtt_buddy);
-	if (mr->first_seg == -1)
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox)) {
+		err = PTR_ERR(mailbox);
 		goto err_out_table;
-
-	/*
-	 * If list_len is odd, we add one more dummy entry for
-	 * firmware efficiency.
-	 */
-	mailbox = kmalloc(max(sizeof *mpt_entry,
-			      (size_t) 8 * (list_len + (list_len & 1) + 2)) +
-			  MTHCA_CMD_MAILBOX_EXTRA,
-			  GFP_KERNEL);
-	if (!mailbox)
-		goto err_out_free_mtt;
-
-	mtt_entry = MAILBOX_ALIGN(mailbox);
-
-	mtt_entry[0] = cpu_to_be64(dev->mr_table.mtt_base +
-				   mr->first_seg * MTHCA_MTT_SEG_SIZE);
-	mtt_entry[1] = 0;
-	for (i = 0; i < list_len; ++i)
-		mtt_entry[i + 2] = cpu_to_be64(buffer_list[i] |
-					       MTHCA_MTT_FLAG_PRESENT);
-	if (list_len & 1) {
-		mtt_entry[i + 2] = 0;
-		++list_len;
 	}
-
-	if (0) {
-		mthca_dbg(dev, "Dumping MPT entry\n");
-		for (i = 0; i < list_len + 2; ++i)
-			printk(KERN_ERR "[%2d] %016llx\n",
-			       i, (unsigned long long) be64_to_cpu(mtt_entry[i]));
-	}
-
-	err = mthca_WRITE_MTT(dev, mtt_entry, list_len, &status);
-	if (err) {
-		mthca_warn(dev, "WRITE_MTT failed (%d)\n", err);
-		goto err_out_mailbox_free;
-	}
-	if (status) {
-		mthca_warn(dev, "WRITE_MTT returned status 0x%02x\n",
-			   status);
-		err = -EINVAL;
-		goto err_out_mailbox_free;
-	}
-
-	mpt_entry = MAILBOX_ALIGN(mailbox);
+	mpt_entry = mailbox->buf;
 
 	mpt_entry->flags = cpu_to_be32(MTHCA_MPT_FLAG_SW_OWNS     |
 				       MTHCA_MPT_FLAG_MIO         |
 				       MTHCA_MPT_FLAG_REGION      |
 				       access);
+	if (!mr->mtt)
+		mpt_entry->flags |= cpu_to_be32(MTHCA_MPT_FLAG_PHYSICAL);
 
 	mpt_entry->page_size = cpu_to_be32(buffer_size_shift - 12);
 	mpt_entry->key       = cpu_to_be32(key);
 	mpt_entry->pd        = cpu_to_be32(pd);
 	mpt_entry->start     = cpu_to_be64(iova);
 	mpt_entry->length    = cpu_to_be64(total_size);
+
 	memset(&mpt_entry->lkey, 0,
 	       sizeof *mpt_entry - offsetof(struct mthca_mpt_entry, lkey));
-	mpt_entry->mtt_seg   = cpu_to_be64(dev->mr_table.mtt_base +
-					   mr->first_seg * MTHCA_MTT_SEG_SIZE);
+
+	if (mr->mtt)
+		mpt_entry->mtt_seg =
+			cpu_to_be64(dev->mr_table.mtt_base +
+				    mr->mtt->first_seg * dev->limits.mtt_seg_size);
 
 	if (0) {
 		mthca_dbg(dev, "Dumping MPT entry %08x:\n", mr->ibmr.lkey);
 		for (i = 0; i < sizeof (struct mthca_mpt_entry) / 4; ++i) {
 			if (i % 4 == 0)
 				printk("[%02x] ", i * 4);
-			printk(" %08x", be32_to_cpu(((u32 *) mpt_entry)[i]));
+			printk(" %08x", be32_to_cpu(((__be32 *) mpt_entry)[i]));
 			if ((i + 1) % 4 == 0)
 				printk("\n");
 		}
 	}
 
-	err = mthca_SW2HW_MPT(dev, mpt_entry,
-			      key & (dev->limits.num_mpts - 1),
-			      &status);
-	if (err)
+	err = mthca_SW2HW_MPT(dev, mailbox,
+			      key & (dev->limits.num_mpts - 1));
+	if (err) {
 		mthca_warn(dev, "SW2HW_MPT failed (%d)\n", err);
-	else if (status) {
-		mthca_warn(dev, "SW2HW_MPT returned status 0x%02x\n",
-			   status);
-		err = -EINVAL;
+		goto err_out_mailbox;
 	}
 
-	kfree(mailbox);
+	mthca_free_mailbox(dev, mailbox);
 	return err;
 
-err_out_mailbox_free:
-	kfree(mailbox);
-
-err_out_free_mtt:
-	mthca_free_mtt(dev, mr->first_seg, mr->order, &dev->mr_table.mtt_buddy);
+err_out_mailbox:
+	mthca_free_mailbox(dev, mailbox);
 
 err_out_table:
-	if (mthca_is_memfree(dev))
-		mthca_table_put(dev, dev->mr_table.mpt_table, key);
+	mthca_table_put(dev, dev->mr_table.mpt_table, key);
 
 err_out_mpt_free:
 	mthca_free(&dev->mr_table.mpt_alloc, key);
 	return err;
 }
 
-/* Free mr or fmr */
-static void mthca_free_region(struct mthca_dev *dev, u32 lkey, int order,
-			      u32 first_seg, struct mthca_buddy *buddy)
+int mthca_mr_alloc_notrans(struct mthca_dev *dev, u32 pd,
+			   u32 access, struct mthca_mr *mr)
 {
-	if (order >= 0)
-		mthca_free_mtt(dev, first_seg, order, buddy);
+	mr->mtt = NULL;
+	return mthca_mr_alloc(dev, pd, 12, 0, ~0ULL, access, mr);
+}
 
-	if (mthca_is_memfree(dev))
-		mthca_table_put(dev, dev->mr_table.mpt_table,
-				arbel_key_to_hw_index(lkey));
+int mthca_mr_alloc_phys(struct mthca_dev *dev, u32 pd,
+			u64 *buffer_list, int buffer_size_shift,
+			int list_len, u64 iova, u64 total_size,
+			u32 access, struct mthca_mr *mr)
+{
+	int err;
+
+	mr->mtt = mthca_alloc_mtt(dev, list_len);
+	if (IS_ERR(mr->mtt))
+		return PTR_ERR(mr->mtt);
+
+	err = mthca_write_mtt(dev, mr->mtt, 0, buffer_list, list_len);
+	if (err) {
+		mthca_free_mtt(dev, mr->mtt);
+		return err;
+	}
+
+	err = mthca_mr_alloc(dev, pd, buffer_size_shift, iova,
+			     total_size, access, mr);
+	if (err)
+		mthca_free_mtt(dev, mr->mtt);
+
+	return err;
+}
+
+/* Free mr or fmr */
+static void mthca_free_region(struct mthca_dev *dev, u32 lkey)
+{
+	mthca_table_put(dev, dev->mr_table.mpt_table,
+			key_to_hw_index(dev, lkey));
 
 	mthca_free(&dev->mr_table.mpt_alloc, key_to_hw_index(dev, lkey));
 }
@@ -462,39 +553,29 @@ static void mthca_free_region(struct mthca_dev *dev, u32 lkey, int order,
 void mthca_free_mr(struct mthca_dev *dev, struct mthca_mr *mr)
 {
 	int err;
-	u8 status;
-
-	might_sleep();
 
 	err = mthca_HW2SW_MPT(dev, NULL,
 			      key_to_hw_index(dev, mr->ibmr.lkey) &
-			      (dev->limits.num_mpts - 1),
-			      &status);
+			      (dev->limits.num_mpts - 1));
 	if (err)
 		mthca_warn(dev, "HW2SW_MPT failed (%d)\n", err);
-	else if (status)
-		mthca_warn(dev, "HW2SW_MPT returned status 0x%02x\n",
-			   status);
 
-	mthca_free_region(dev, mr->ibmr.lkey, mr->order, mr->first_seg,
-			  &dev->mr_table.mtt_buddy);
+	mthca_free_region(dev, mr->ibmr.lkey);
+	mthca_free_mtt(dev, mr->mtt);
 }
 
 int mthca_fmr_alloc(struct mthca_dev *dev, u32 pd,
 		    u32 access, struct mthca_fmr *mr)
 {
 	struct mthca_mpt_entry *mpt_entry;
-	void *mailbox;
+	struct mthca_mailbox *mailbox;
 	u64 mtt_seg;
 	u32 key, idx;
-	u8 status;
 	int list_len = mr->attr.max_pages;
 	int err = -ENOMEM;
 	int i;
 
-	might_sleep();
-
-	if (mr->attr.page_size < 12 || mr->attr.page_size >= 32)
+	if (mr->attr.page_shift < 12 || mr->attr.page_shift >= 32)
 		return -EINVAL;
 
 	/* For Arbel, all MTTs must fit in the same page. */
@@ -507,6 +588,7 @@ int mthca_fmr_alloc(struct mthca_dev *dev, u32 pd,
 	key = mthca_alloc(&dev->mr_table.mpt_alloc);
 	if (key == -1)
 		return -ENOMEM;
+	key = adjust_key(dev, key);
 
 	idx = key & (dev->limits.num_mpts - 1);
 	mr->ibmr.rkey = mr->ibmr.lkey = hw_index_to_key(dev, key);
@@ -516,44 +598,42 @@ 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 +
-		       	sizeof *(mr->mem.tavor.mpt) * idx;
+			sizeof *(mr->mem.tavor.mpt) * idx;
 
-	for (i = MTHCA_MTT_SEG_SIZE / 8, mr->order = 0;
-	     i < list_len;
-	     i <<= 1, ++mr->order)
-		; /* nothing */
-
-	mr->first_seg = mthca_alloc_mtt(dev, mr->order,
-				       	dev->mr_table.fmr_mtt_buddy);
-	if (mr->first_seg == -1)
+	mr->mtt = __mthca_alloc_mtt(dev, list_len, dev->mr_table.fmr_mtt_buddy);
+	if (IS_ERR(mr->mtt)) {
+		err = PTR_ERR(mr->mtt);
 		goto err_out_table;
+	}
 
-	mtt_seg = mr->first_seg * MTHCA_MTT_SEG_SIZE;
+	mtt_seg = mr->mtt->first_seg * dev->limits.mtt_seg_size;
 
 	if (mthca_is_memfree(dev)) {
 		mr->mem.arbel.mtts = mthca_table_find(dev->mr_table.mtt_table,
-						      mr->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;
 
-	mailbox = kmalloc(sizeof *mpt_entry + MTHCA_CMD_MAILBOX_EXTRA,
-			  GFP_KERNEL);
-	if (!mailbox)
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox)) {
+		err = PTR_ERR(mailbox);
 		goto err_out_free_mtt;
+	}
 
-	mpt_entry = MAILBOX_ALIGN(mailbox);
+	mpt_entry = mailbox->buf;
 
 	mpt_entry->flags = cpu_to_be32(MTHCA_MPT_FLAG_SW_OWNS     |
 				       MTHCA_MPT_FLAG_MIO         |
 				       MTHCA_MPT_FLAG_REGION      |
 				       access);
 
-	mpt_entry->page_size = cpu_to_be32(mr->attr.page_size - 12);
+	mpt_entry->page_size = cpu_to_be32(mr->attr.page_shift - 12);
 	mpt_entry->key       = cpu_to_be32(key);
 	mpt_entry->pd        = cpu_to_be32(pd);
 	memset(&mpt_entry->start, 0,
@@ -565,42 +645,33 @@ int mthca_fmr_alloc(struct mthca_dev *dev, u32 pd,
 		for (i = 0; i < sizeof (struct mthca_mpt_entry) / 4; ++i) {
 			if (i % 4 == 0)
 				printk("[%02x] ", i * 4);
-			printk(" %08x", be32_to_cpu(((u32 *) mpt_entry)[i]));
+			printk(" %08x", be32_to_cpu(((__be32 *) mpt_entry)[i]));
 			if ((i + 1) % 4 == 0)
 				printk("\n");
 		}
 	}
 
-	err = mthca_SW2HW_MPT(dev, mpt_entry,
-			      key & (dev->limits.num_mpts - 1),
-			      &status);
+	err = mthca_SW2HW_MPT(dev, mailbox,
+			      key & (dev->limits.num_mpts - 1));
 	if (err) {
 		mthca_warn(dev, "SW2HW_MPT failed (%d)\n", err);
 		goto err_out_mailbox_free;
 	}
-	if (status) {
-		mthca_warn(dev, "SW2HW_MPT returned status 0x%02x\n",
-			   status);
-		err = -EINVAL;
-		goto err_out_mailbox_free;
-	}
 
-	kfree(mailbox);
+	mthca_free_mailbox(dev, mailbox);
 	return 0;
 
 err_out_mailbox_free:
-	kfree(mailbox);
+	mthca_free_mailbox(dev, mailbox);
 
 err_out_free_mtt:
-	mthca_free_mtt(dev, mr->first_seg, mr->order,
-		       dev->mr_table.fmr_mtt_buddy);
+	mthca_free_mtt(dev, mr->mtt);
 
 err_out_table:
-	if (mthca_is_memfree(dev))
-		mthca_table_put(dev, dev->mr_table.mpt_table, key);
+	mthca_table_put(dev, dev->mr_table.mpt_table, key);
 
 err_out_mpt_free:
-	mthca_free(&dev->mr_table.mpt_alloc, mr->ibmr.lkey);
+	mthca_free(&dev->mr_table.mpt_alloc, key);
 	return err;
 }
 
@@ -609,8 +680,9 @@ int mthca_free_fmr(struct mthca_dev *dev, struct mthca_fmr *fmr)
 	if (fmr->maps)
 		return -EBUSY;
 
-	mthca_free_region(dev, fmr->ibmr.lkey, fmr->order, fmr->first_seg,
-			  dev->mr_table.fmr_mtt_buddy);
+	mthca_free_region(dev, fmr->ibmr.lkey);
+	mthca_free_mtt(dev, fmr->mtt);
+
 	return 0;
 }
 
@@ -622,7 +694,7 @@ static inline int mthca_check_fmr(struct mthca_fmr *fmr, u64 *page_list,
 	if (list_len > fmr->attr.max_pages)
 		return -EINVAL;
 
-	page_mask = (1 << fmr->attr.page_size) - 1;
+	page_mask = (1 << fmr->attr.page_shift) - 1;
 
 	/* We are getting page lists, so va must be page aligned. */
 	if (iova & page_mask)
@@ -670,10 +742,10 @@ int mthca_tavor_map_phys_fmr(struct ib_fmr *ibfmr, u64 *page_list,
 	}
 
 	mpt_entry.lkey   = cpu_to_be32(key);
-	mpt_entry.length = cpu_to_be64(list_len * (1ull << fmr->attr.page_size));
+	mpt_entry.length = cpu_to_be64(list_len * (1ull << fmr->attr.page_shift));
 	mpt_entry.start  = cpu_to_be64(iova);
 
-	writel(mpt_entry.lkey, &fmr->mem.tavor.mpt->key);
+	__raw_writel((__force u32) mpt_entry.lkey, &fmr->mem.tavor.mpt->key);
 	memcpy_toio(&fmr->mem.tavor.mpt->start, &mpt_entry.start,
 		    offsetof(struct mthca_mpt_entry, window_count) -
 		    offsetof(struct mthca_mpt_entry, start));
@@ -698,20 +770,29 @@ int mthca_arbel_map_phys_fmr(struct ib_fmr *ibfmr, u64 *page_list,
 	++fmr->maps;
 
 	key = arbel_key_to_hw_index(fmr->ibmr.lkey);
-	key += dev->limits.num_mpts;
+	if (dev->mthca_flags & MTHCA_FLAG_SINAI_OPT)
+		key += SINAI_FMR_KEY_INC;
+	else
+		key += dev->limits.num_mpts;
 	fmr->ibmr.lkey = fmr->ibmr.rkey = arbel_hw_index_to_key(key);
 
 	*(u8 *) fmr->mem.arbel.mpt = MTHCA_MPT_STATUS_SW;
 
 	wmb();
 
+	dma_sync_single_for_cpu(&dev->pdev->dev, fmr->mem.arbel.dma_handle,
+				list_len * sizeof(u64), DMA_TO_DEVICE);
+
 	for (i = 0; i < list_len; ++i)
 		fmr->mem.arbel.mtts[i] = cpu_to_be64(page_list[i] |
 						     MTHCA_MTT_FLAG_PRESENT);
 
+	dma_sync_single_for_device(&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_size));
+	fmr->mem.arbel.mpt->length = cpu_to_be64(list_len * (1ull << fmr->attr.page_shift));
 	fmr->mem.arbel.mpt->start  = cpu_to_be64(iova);
 
 	wmb();
@@ -725,15 +806,9 @@ int mthca_arbel_map_phys_fmr(struct ib_fmr *ibfmr, u64 *page_list,
 
 void mthca_tavor_fmr_unmap(struct mthca_dev *dev, struct mthca_fmr *fmr)
 {
-	u32 key;
-
 	if (!fmr->maps)
 		return;
 
-	key = tavor_key_to_hw_index(fmr->ibmr.lkey);
-	key &= dev->limits.num_mpts - 1;
-	fmr->ibmr.lkey = fmr->ibmr.rkey = tavor_hw_index_to_key(key);
-
 	fmr->maps = 0;
 
 	writeb(MTHCA_MPT_STATUS_SW, fmr->mem.tavor.mpt);
@@ -741,23 +816,18 @@ void mthca_tavor_fmr_unmap(struct mthca_dev *dev, struct mthca_fmr *fmr)
 
 void mthca_arbel_fmr_unmap(struct mthca_dev *dev, struct mthca_fmr *fmr)
 {
-	u32 key;
-
 	if (!fmr->maps)
 		return;
 
-	key = arbel_key_to_hw_index(fmr->ibmr.lkey);
-	key &= dev->limits.num_mpts - 1;
-	fmr->ibmr.lkey = fmr->ibmr.rkey = arbel_hw_index_to_key(key);
-
 	fmr->maps = 0;
 
 	*(u8 *) fmr->mem.arbel.mpt = MTHCA_MPT_STATUS_SW;
 }
 
-int __devinit mthca_init_mr_table(struct mthca_dev *dev)
+int mthca_init_mr_table(struct mthca_dev *dev)
 {
-	int err, i;
+	phys_addr_t addr;
+	int mpts, mtts, err, i;
 
 	err = mthca_alloc_init(&dev->mr_table.mpt_alloc,
 			       dev->limits.num_mpts,
@@ -771,6 +841,9 @@ int __devinit mthca_init_mr_table(struct mthca_dev *dev)
 	else
 		dev->mthca_flags |= MTHCA_FLAG_FMR;
 
+	if (dev->mthca_flags & MTHCA_FLAG_SINAI_OPT)
+		mthca_dbg(dev, "Memory key throughput optimization activated.\n");
+
 	err = mthca_buddy_init(&dev->mr_table.mtt_buddy,
 			       fls(dev->limits.num_mtt_segs - 1));
 
@@ -788,10 +861,21 @@ int __devinit mthca_init_mr_table(struct mthca_dev *dev)
 			err = -EINVAL;
 			goto err_fmr_mpt;
 		}
+		mpts = mtts = 1 << i;
+	} else {
+		mtts = dev->limits.num_mtt_segs;
+		mpts = dev->limits.num_mpts;
+	}
+
+	if (!mthca_is_memfree(dev) &&
+	    (dev->mthca_flags & MTHCA_FLAG_FMR)) {
+
+		addr = pci_resource_start(dev->pdev, 4) +
+			((pci_resource_len(dev->pdev, 4) - 1) &
+			 dev->mr_table.mpt_base);
 
 		dev->mr_table.tavor_fmr.mpt_base =
-		       	ioremap(dev->mr_table.mpt_base,
-				(1 << i) * sizeof (struct mthca_mpt_entry));
+			ioremap(addr, mpts * sizeof(struct mthca_mpt_entry));
 
 		if (!dev->mr_table.tavor_fmr.mpt_base) {
 			mthca_warn(dev, "MPT ioremap for FMR failed.\n");
@@ -799,26 +883,31 @@ int __devinit mthca_init_mr_table(struct mthca_dev *dev)
 			goto err_fmr_mpt;
 		}
 
+		addr = pci_resource_start(dev->pdev, 4) +
+			((pci_resource_len(dev->pdev, 4) - 1) &
+			 dev->mr_table.mtt_base);
+
 		dev->mr_table.tavor_fmr.mtt_base =
-			ioremap(dev->mr_table.mtt_base,
-				(1 << i) * MTHCA_MTT_SEG_SIZE);
+			ioremap(addr, mtts * dev->limits.mtt_seg_size);
 		if (!dev->mr_table.tavor_fmr.mtt_base) {
 			mthca_warn(dev, "MTT ioremap for FMR failed.\n");
 			err = -ENOMEM;
 			goto err_fmr_mtt;
 		}
+	}
 
-		err = mthca_buddy_init(&dev->mr_table.tavor_fmr.mtt_buddy, i);
+	if (dev->limits.fmr_reserved_mtts) {
+		err = mthca_buddy_init(&dev->mr_table.tavor_fmr.mtt_buddy, fls(mtts - 1));
 		if (err)
 			goto err_fmr_mtt_buddy;
 
 		/* Prevent regular MRs from using FMR keys */
-		err = mthca_buddy_alloc(&dev->mr_table.mtt_buddy, i);
+		err = mthca_buddy_alloc(&dev->mr_table.mtt_buddy, fls(mtts - 1));
 		if (err)
 			goto err_reserve_fmr;
 
 		dev->mr_table.fmr_mtt_buddy =
-		       	&dev->mr_table.tavor_fmr.mtt_buddy;
+			&dev->mr_table.tavor_fmr.mtt_buddy;
 	} else
 		dev->mr_table.fmr_mtt_buddy = &dev->mr_table.mtt_buddy;
 
@@ -826,7 +915,8 @@ int __devinit mthca_init_mr_table(struct mthca_dev *dev)
 	if (dev->limits.reserved_mtts) {
 		i = fls(dev->limits.reserved_mtts - 1);
 
-		if (mthca_alloc_mtt(dev, i, dev->mr_table.fmr_mtt_buddy) == -1) {
+		if (mthca_alloc_mtt_range(dev, i,
+					  dev->mr_table.fmr_mtt_buddy) == -1) {
 			mthca_warn(dev, "MTT table of order %d is too small.\n",
 				  dev->mr_table.fmr_mtt_buddy->max_order);
 			err = -ENOMEM;
@@ -858,7 +948,7 @@ err_mtt_buddy:
 	return err;
 }
 
-void __devexit mthca_cleanup_mr_table(struct mthca_dev *dev)
+void mthca_cleanup_mr_table(struct mthca_dev *dev)
 {
 	/* XXX check if any MRs are still allocated? */
 	if (dev->limits.fmr_reserved_mtts)