Merge branch 'from-tomtucker' into for-2.6.28

include/linux/sunrpc/svc_rdma.h

@@ -72,6 +72,7 @@ extern atomic_t rdma_stat_sq_prod;
  */
 struct svc_rdma_op_ctxt {
 	struct svc_rdma_op_ctxt *read_hdr;
+	struct svc_rdma_fastreg_mr *frmr;
 	int hdr_count;
 	struct xdr_buf arg;
 	struct list_head dto_q;
@@ -103,16 +104,30 @@ struct svc_rdma_chunk_sge {
 	int start;		/* sge no for this chunk */
 	int count;		/* sge count for this chunk */
 };
+struct svc_rdma_fastreg_mr {
+	struct ib_mr *mr;
+	void *kva;
+	struct ib_fast_reg_page_list *page_list;
+	int page_list_len;
+	unsigned long access_flags;
+	unsigned long map_len;
+	enum dma_data_direction direction;
+	struct list_head frmr_list;
+};
 struct svc_rdma_req_map {
+	struct svc_rdma_fastreg_mr *frmr;
 	unsigned long count;
 	union {
 		struct kvec sge[RPCSVC_MAXPAGES];
 		struct svc_rdma_chunk_sge ch[RPCSVC_MAXPAGES];
 	};
 };
-
+#define RDMACTXT_F_FAST_UNREG	1
 #define RDMACTXT_F_LAST_CTXT	2
 
+#define	SVCRDMA_DEVCAP_FAST_REG		1	/* fast mr registration */
+#define	SVCRDMA_DEVCAP_READ_W_INV	2	/* read w/ invalidate */
+
 struct svcxprt_rdma {
 	struct svc_xprt      sc_xprt;		/* SVC transport structure */
 	struct rdma_cm_id    *sc_cm_id;		/* RDMA connection id */
@@ -136,6 +151,11 @@ struct svcxprt_rdma {
 	struct ib_cq         *sc_rq_cq;
 	struct ib_cq         *sc_sq_cq;
 	struct ib_mr         *sc_phys_mr;	/* MR for server memory */
+	u32		     sc_dev_caps;	/* distilled device caps */
+	u32		     sc_dma_lkey;	/* local dma key */
+	unsigned int	     sc_frmr_pg_list_len;
+	struct list_head     sc_frmr_q;
+	spinlock_t	     sc_frmr_q_lock;
 
 	spinlock_t	     sc_lock;		/* transport lock */
 
@@ -192,8 +212,13 @@ extern int svc_rdma_post_recv(struct svcxprt_rdma *);
 extern int svc_rdma_create_listen(struct svc_serv *, int, struct sockaddr *);
 extern struct svc_rdma_op_ctxt *svc_rdma_get_context(struct svcxprt_rdma *);
 extern void svc_rdma_put_context(struct svc_rdma_op_ctxt *, int);
+extern void svc_rdma_unmap_dma(struct svc_rdma_op_ctxt *ctxt);
 extern struct svc_rdma_req_map *svc_rdma_get_req_map(void);
 extern void svc_rdma_put_req_map(struct svc_rdma_req_map *);
+extern int svc_rdma_fastreg(struct svcxprt_rdma *, struct svc_rdma_fastreg_mr *);
+extern struct svc_rdma_fastreg_mr *svc_rdma_get_frmr(struct svcxprt_rdma *);
+extern void svc_rdma_put_frmr(struct svcxprt_rdma *,
+			      struct svc_rdma_fastreg_mr *);
 extern void svc_sq_reap(struct svcxprt_rdma *);
 extern void svc_rq_reap(struct svcxprt_rdma *);
 extern struct svc_xprt_class svc_rdma_class;

net/sunrpc/xprtrdma/svc_rdma_recvfrom.c

@@ -116,7 +116,7 @@ static void rdma_build_arg_xdr(struct svc_rqst *rqstp,
  *
  * Assumptions:
  * - chunk[0]->position points to pages[0] at an offset of 0
- * - pages[] is not physically or virtually contigous and consists of
+ * - pages[] is not physically or virtually contiguous and consists of
  *   PAGE_SIZE elements.
  *
  * Output:
@@ -125,7 +125,7 @@ static void rdma_build_arg_xdr(struct svc_rqst *rqstp,
  *   chunk in the read list
  *
  */
-static int rdma_rcl_to_sge(struct svcxprt_rdma *xprt,
+static int map_read_chunks(struct svcxprt_rdma *xprt,
 			   struct svc_rqst *rqstp,
 			   struct svc_rdma_op_ctxt *head,
 			   struct rpcrdma_msg *rmsgp,
@@ -211,26 +211,128 @@ static int rdma_rcl_to_sge(struct svcxprt_rdma *xprt,
 	return sge_no;
 }
 
-static void rdma_set_ctxt_sge(struct svcxprt_rdma *xprt,
-			      struct svc_rdma_op_ctxt *ctxt,
-			      struct kvec *vec,
-			      u64 *sgl_offset,
-			      int count)
+/* Map a read-chunk-list to an XDR and fast register the page-list.
+ *
+ * Assumptions:
+ * - chunk[0]	position points to pages[0] at an offset of 0
+ * - pages[]	will be made physically contiguous by creating a one-off memory
+ *		region using the fastreg verb.
+ * - byte_count is # of bytes in read-chunk-list
+ * - ch_count	is # of chunks in read-chunk-list
+ *
+ * Output:
+ * - sge array pointing into pages[] array.
+ * - chunk_sge array specifying sge index and count for each
+ *   chunk in the read list
+ */
+static int fast_reg_read_chunks(struct svcxprt_rdma *xprt,
+				struct svc_rqst *rqstp,
+				struct svc_rdma_op_ctxt *head,
+				struct rpcrdma_msg *rmsgp,
+				struct svc_rdma_req_map *rpl_map,
+				struct svc_rdma_req_map *chl_map,
+				int ch_count,
+				int byte_count)
+{
+	int page_no;
+	int ch_no;
+	u32 offset;
+	struct rpcrdma_read_chunk *ch;
+	struct svc_rdma_fastreg_mr *frmr;
+	int ret = 0;
+
+	frmr = svc_rdma_get_frmr(xprt);
+	if (IS_ERR(frmr))
+		return -ENOMEM;
+
+	head->frmr = frmr;
+	head->arg.head[0] = rqstp->rq_arg.head[0];
+	head->arg.tail[0] = rqstp->rq_arg.tail[0];
+	head->arg.pages = &head->pages[head->count];
+	head->hdr_count = head->count; /* save count of hdr pages */
+	head->arg.page_base = 0;
+	head->arg.page_len = byte_count;
+	head->arg.len = rqstp->rq_arg.len + byte_count;
+	head->arg.buflen = rqstp->rq_arg.buflen + byte_count;
+
+	/* Fast register the page list */
+	frmr->kva = page_address(rqstp->rq_arg.pages[0]);
+	frmr->direction = DMA_FROM_DEVICE;
+	frmr->access_flags = (IB_ACCESS_LOCAL_WRITE|IB_ACCESS_REMOTE_WRITE);
+	frmr->map_len = byte_count;
+	frmr->page_list_len = PAGE_ALIGN(byte_count) >> PAGE_SHIFT;
+	for (page_no = 0; page_no < frmr->page_list_len; page_no++) {
+		frmr->page_list->page_list[page_no] =
+			ib_dma_map_single(xprt->sc_cm_id->device,
+					  page_address(rqstp->rq_arg.pages[page_no]),
+					  PAGE_SIZE, DMA_TO_DEVICE);
+		if (ib_dma_mapping_error(xprt->sc_cm_id->device,
+					 frmr->page_list->page_list[page_no]))
+			goto fatal_err;
+		atomic_inc(&xprt->sc_dma_used);
+		head->arg.pages[page_no] = rqstp->rq_arg.pages[page_no];
+	}
+	head->count += page_no;
+
+	/* rq_respages points one past arg pages */
+	rqstp->rq_respages = &rqstp->rq_arg.pages[page_no];
+
+	/* Create the reply and chunk maps */
+	offset = 0;
+	ch = (struct rpcrdma_read_chunk *)&rmsgp->rm_body.rm_chunks[0];
+	for (ch_no = 0; ch_no < ch_count; ch_no++) {
+		rpl_map->sge[ch_no].iov_base = frmr->kva + offset;
+		rpl_map->sge[ch_no].iov_len = ch->rc_target.rs_length;
+		chl_map->ch[ch_no].count = 1;
+		chl_map->ch[ch_no].start = ch_no;
+		offset += ch->rc_target.rs_length;
+		ch++;
+	}
+
+	ret = svc_rdma_fastreg(xprt, frmr);
+	if (ret)
+		goto fatal_err;
+
+	return ch_no;
+
+ fatal_err:
+	printk("svcrdma: error fast registering xdr for xprt %p", xprt);
+	svc_rdma_put_frmr(xprt, frmr);
+	return -EIO;
+}
+
+static int rdma_set_ctxt_sge(struct svcxprt_rdma *xprt,
+			     struct svc_rdma_op_ctxt *ctxt,
+			     struct svc_rdma_fastreg_mr *frmr,
+			     struct kvec *vec,
+			     u64 *sgl_offset,
+			     int count)
 {
 	int i;
 
 	ctxt->count = count;
 	ctxt->direction = DMA_FROM_DEVICE;
 	for (i = 0; i < count; i++) {
-		atomic_inc(&xprt->sc_dma_used);
-		ctxt->sge[i].addr =
-			ib_dma_map_single(xprt->sc_cm_id->device,
-					  vec[i].iov_base, vec[i].iov_len,
-					  DMA_FROM_DEVICE);
+		ctxt->sge[i].length = 0; /* in case map fails */
+		if (!frmr) {
+			ctxt->sge[i].addr =
+				ib_dma_map_single(xprt->sc_cm_id->device,
+						  vec[i].iov_base,
+						  vec[i].iov_len,
+						  DMA_FROM_DEVICE);
+			if (ib_dma_mapping_error(xprt->sc_cm_id->device,
+						 ctxt->sge[i].addr))
+				return -EINVAL;
+			ctxt->sge[i].lkey = xprt->sc_dma_lkey;
+			atomic_inc(&xprt->sc_dma_used);
+		} else {
+			ctxt->sge[i].addr = (unsigned long)vec[i].iov_base;
+			ctxt->sge[i].lkey = frmr->mr->lkey;
+		}
 		ctxt->sge[i].length = vec[i].iov_len;
-		ctxt->sge[i].lkey = xprt->sc_phys_mr->lkey;
 		*sgl_offset = *sgl_offset + vec[i].iov_len;
 	}
+	return 0;
 }
 
 static int rdma_read_max_sge(struct svcxprt_rdma *xprt, int sge_count)
@@ -278,6 +380,7 @@ static int rdma_read_xdr(struct svcxprt_rdma *xprt,
 			 struct svc_rdma_op_ctxt *hdr_ctxt)
 {
 	struct ib_send_wr read_wr;
+	struct ib_send_wr inv_wr;
 	int err = 0;
 	int ch_no;
 	int ch_count;
@@ -301,9 +404,20 @@ static int rdma_read_xdr(struct svcxprt_rdma *xprt,
 	svc_rdma_rcl_chunk_counts(ch, &ch_count, &byte_count);
 	if (ch_count > RPCSVC_MAXPAGES)
 		return -EINVAL;
-	sge_count = rdma_rcl_to_sge(xprt, rqstp, hdr_ctxt, rmsgp,
-				    rpl_map, chl_map,
-				    ch_count, byte_count);
+
+	if (!xprt->sc_frmr_pg_list_len)
+		sge_count = map_read_chunks(xprt, rqstp, hdr_ctxt, rmsgp,
+					    rpl_map, chl_map, ch_count,
+					    byte_count);
+	else
+		sge_count = fast_reg_read_chunks(xprt, rqstp, hdr_ctxt, rmsgp,
+						 rpl_map, chl_map, ch_count,
+						 byte_count);
+	if (sge_count < 0) {
+		err = -EIO;
+		goto out;
+	}
+
 	sgl_offset = 0;
 	ch_no = 0;
 
@@ -312,13 +426,16 @@ static int rdma_read_xdr(struct svcxprt_rdma *xprt,
 next_sge:
 		ctxt = svc_rdma_get_context(xprt);
 		ctxt->direction = DMA_FROM_DEVICE;
+		ctxt->frmr = hdr_ctxt->frmr;
+		ctxt->read_hdr = NULL;
 		clear_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags);
+		clear_bit(RDMACTXT_F_FAST_UNREG, &ctxt->flags);
 
 		/* Prepare READ WR */
 		memset(&read_wr, 0, sizeof read_wr);
-		ctxt->wr_op = IB_WR_RDMA_READ;
 		read_wr.wr_id = (unsigned long)ctxt;
 		read_wr.opcode = IB_WR_RDMA_READ;
+		ctxt->wr_op = read_wr.opcode;
 		read_wr.send_flags = IB_SEND_SIGNALED;
 		read_wr.wr.rdma.rkey = ch->rc_target.rs_handle;
 		read_wr.wr.rdma.remote_addr =
@@ -327,10 +444,15 @@ next_sge:
 		read_wr.sg_list = ctxt->sge;
 		read_wr.num_sge =
 			rdma_read_max_sge(xprt, chl_map->ch[ch_no].count);
-		rdma_set_ctxt_sge(xprt, ctxt,
-				  &rpl_map->sge[chl_map->ch[ch_no].start],
-				  &sgl_offset,
-				  read_wr.num_sge);
+		err = rdma_set_ctxt_sge(xprt, ctxt, hdr_ctxt->frmr,
+					&rpl_map->sge[chl_map->ch[ch_no].start],
+					&sgl_offset,
+					read_wr.num_sge);
+		if (err) {
+			svc_rdma_unmap_dma(ctxt);
+			svc_rdma_put_context(ctxt, 0);
+			goto out;
+		}
 		if (((ch+1)->rc_discrim == 0) &&
 		    (read_wr.num_sge == chl_map->ch[ch_no].count)) {
 			/*
@@ -339,6 +461,29 @@ next_sge:
 			 * the client and the RPC needs to be enqueued.
 			 */
 			set_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags);
+			if (hdr_ctxt->frmr) {
+				set_bit(RDMACTXT_F_FAST_UNREG, &ctxt->flags);
+				/*
+				 * Invalidate the local MR used to map the data
+				 * sink.
+				 */
+				if (xprt->sc_dev_caps &
+				    SVCRDMA_DEVCAP_READ_W_INV) {
+					read_wr.opcode =
+						IB_WR_RDMA_READ_WITH_INV;
+					ctxt->wr_op = read_wr.opcode;
+					read_wr.ex.invalidate_rkey =
+						ctxt->frmr->mr->lkey;
+				} else {
+					/* Prepare INVALIDATE WR */
+					memset(&inv_wr, 0, sizeof inv_wr);
+					inv_wr.opcode = IB_WR_LOCAL_INV;
+					inv_wr.send_flags = IB_SEND_SIGNALED;
+					inv_wr.ex.invalidate_rkey =
+						hdr_ctxt->frmr->mr->lkey;
+					read_wr.next = &inv_wr;
+				}
+			}
 			ctxt->read_hdr = hdr_ctxt;
 		}
 		/* Post the read */

net/sunrpc/xprtrdma/svc_rdma_sendto.c

@@ -69,9 +69,127 @@
  * array is only concerned with the reply we are assured that we have
  * on extra page for the RPCRMDA header.
  */
-static void xdr_to_sge(struct svcxprt_rdma *xprt,
-		       struct xdr_buf *xdr,
-		       struct svc_rdma_req_map *vec)
+int fast_reg_xdr(struct svcxprt_rdma *xprt,
+		 struct xdr_buf *xdr,
+		 struct svc_rdma_req_map *vec)
+{
+	int sge_no;
+	u32 sge_bytes;
+	u32 page_bytes;
+	u32 page_off;
+	int page_no = 0;
+	u8 *frva;
+	struct svc_rdma_fastreg_mr *frmr;
+
+	frmr = svc_rdma_get_frmr(xprt);
+	if (IS_ERR(frmr))
+		return -ENOMEM;
+	vec->frmr = frmr;
+
+	/* Skip the RPCRDMA header */
+	sge_no = 1;
+
+	/* Map the head. */
+	frva = (void *)((unsigned long)(xdr->head[0].iov_base) & PAGE_MASK);
+	vec->sge[sge_no].iov_base = xdr->head[0].iov_base;
+	vec->sge[sge_no].iov_len = xdr->head[0].iov_len;
+	vec->count = 2;
+	sge_no++;
+
+	/* Build the FRMR */
+	frmr->kva = frva;
+	frmr->direction = DMA_TO_DEVICE;
+	frmr->access_flags = 0;
+	frmr->map_len = PAGE_SIZE;
+	frmr->page_list_len = 1;
+	frmr->page_list->page_list[page_no] =
+		ib_dma_map_single(xprt->sc_cm_id->device,
+				  (void *)xdr->head[0].iov_base,
+				  PAGE_SIZE, DMA_TO_DEVICE);
+	if (ib_dma_mapping_error(xprt->sc_cm_id->device,
+				 frmr->page_list->page_list[page_no]))
+		goto fatal_err;
+	atomic_inc(&xprt->sc_dma_used);
+
+	page_off = xdr->page_base;
+	page_bytes = xdr->page_len + page_off;
+	if (!page_bytes)
+		goto encode_tail;
+
+	/* Map the pages */
+	vec->sge[sge_no].iov_base = frva + frmr->map_len + page_off;
+	vec->sge[sge_no].iov_len = page_bytes;
+	sge_no++;
+	while (page_bytes) {
+		struct page *page;
+
+		page = xdr->pages[page_no++];
+		sge_bytes = min_t(u32, page_bytes, (PAGE_SIZE - page_off));
+		page_bytes -= sge_bytes;
+
+		frmr->page_list->page_list[page_no] =
+			ib_dma_map_page(xprt->sc_cm_id->device, page, 0,
+					  PAGE_SIZE, DMA_TO_DEVICE);
+		if (ib_dma_mapping_error(xprt->sc_cm_id->device,
+					 frmr->page_list->page_list[page_no]))
+			goto fatal_err;
+
+		atomic_inc(&xprt->sc_dma_used);
+		page_off = 0; /* reset for next time through loop */
+		frmr->map_len += PAGE_SIZE;
+		frmr->page_list_len++;
+	}
+	vec->count++;
+
+ encode_tail:
+	/* Map tail */
+	if (0 == xdr->tail[0].iov_len)
+		goto done;
+
+	vec->count++;
+	vec->sge[sge_no].iov_len = xdr->tail[0].iov_len;
+
+	if (((unsigned long)xdr->tail[0].iov_base & PAGE_MASK) ==
+	    ((unsigned long)xdr->head[0].iov_base & PAGE_MASK)) {
+		/*
+		 * If head and tail use the same page, we don't need
+		 * to map it again.
+		 */
+		vec->sge[sge_no].iov_base = xdr->tail[0].iov_base;
+	} else {
+		void *va;
+
+		/* Map another page for the tail */
+		page_off = (unsigned long)xdr->tail[0].iov_base & ~PAGE_MASK;
+		va = (void *)((unsigned long)xdr->tail[0].iov_base & PAGE_MASK);
+		vec->sge[sge_no].iov_base = frva + frmr->map_len + page_off;
+
+		frmr->page_list->page_list[page_no] =
+			ib_dma_map_single(xprt->sc_cm_id->device, va, PAGE_SIZE,
+					  DMA_TO_DEVICE);
+		if (ib_dma_mapping_error(xprt->sc_cm_id->device,
+					 frmr->page_list->page_list[page_no]))
+			goto fatal_err;
+		atomic_inc(&xprt->sc_dma_used);
+		frmr->map_len += PAGE_SIZE;
+		frmr->page_list_len++;
+	}
+
+ done:
+	if (svc_rdma_fastreg(xprt, frmr))
+		goto fatal_err;
+
+	return 0;
+
+ fatal_err:
+	printk("svcrdma: Error fast registering memory for xprt %p\n", xprt);
+	svc_rdma_put_frmr(xprt, frmr);
+	return -EIO;
+}
+
+static int map_xdr(struct svcxprt_rdma *xprt,
+		   struct xdr_buf *xdr,
+		   struct svc_rdma_req_map *vec)
 {
 	int sge_max = (xdr->len+PAGE_SIZE-1) / PAGE_SIZE + 3;
 	int sge_no;
@@ -83,6 +201,9 @@ static void xdr_to_sge(struct svcxprt_rdma *xprt,
 	BUG_ON(xdr->len !=
 	       (xdr->head[0].iov_len + xdr->page_len + xdr->tail[0].iov_len));
 
+	if (xprt->sc_frmr_pg_list_len)
+		return fast_reg_xdr(xprt, xdr, vec);
+
 	/* Skip the first sge, this is for the RPCRDMA header */
 	sge_no = 1;
 
@@ -116,9 +237,12 @@ static void xdr_to_sge(struct svcxprt_rdma *xprt,
 
 	BUG_ON(sge_no > sge_max);
 	vec->count = sge_no;
+	return 0;
 }
 
 /* Assumptions:
+ * - We are using FRMR
+ *     - or -
  * - The specified write_len can be represented in sc_max_sge * PAGE_SIZE
  */
 static int send_write(struct svcxprt_rdma *xprt, struct svc_rqst *rqstp,
@@ -158,30 +282,35 @@ static int send_write(struct svcxprt_rdma *xprt, struct svc_rqst *rqstp,
 	sge_no = 0;
 
 	/* Copy the remaining SGE */
-	while (bc != 0 && xdr_sge_no < vec->count) {
-		sge[sge_no].lkey = xprt->sc_phys_mr->lkey;
-		sge_bytes = min((size_t)bc,
-				(size_t)(vec->sge[xdr_sge_no].iov_len-sge_off));
+	while (bc != 0) {
+		sge_bytes = min_t(size_t,
+			  bc, vec->sge[xdr_sge_no].iov_len-sge_off);
 		sge[sge_no].length = sge_bytes;
-		atomic_inc(&xprt->sc_dma_used);
-		sge[sge_no].addr =
-			ib_dma_map_single(xprt->sc_cm_id->device,
-					  (void *)
-					  vec->sge[xdr_sge_no].iov_base + sge_off,
-					  sge_bytes, DMA_TO_DEVICE);
-		if (dma_mapping_error(xprt->sc_cm_id->device->dma_device,
-					sge[sge_no].addr))
-			goto err;
+		if (!vec->frmr) {
+			sge[sge_no].addr =
+				ib_dma_map_single(xprt->sc_cm_id->device,
+						  (void *)
+						  vec->sge[xdr_sge_no].iov_base + sge_off,
+						  sge_bytes, DMA_TO_DEVICE);
+			if (ib_dma_mapping_error(xprt->sc_cm_id->device,
+						 sge[sge_no].addr))
+				goto err;
+			atomic_inc(&xprt->sc_dma_used);
+			sge[sge_no].lkey = xprt->sc_dma_lkey;
+		} else {
+			sge[sge_no].addr = (unsigned long)
+				vec->sge[xdr_sge_no].iov_base + sge_off;
+			sge[sge_no].lkey = vec->frmr->mr->lkey;
+		}
+		ctxt->count++;
+		ctxt->frmr = vec->frmr;
 		sge_off = 0;
 		sge_no++;
-		ctxt->count++;
 		xdr_sge_no++;
+		BUG_ON(xdr_sge_no > vec->count);
 		bc -= sge_bytes;
 	}
 
-	BUG_ON(bc != 0);
-	BUG_ON(xdr_sge_no > vec->count);
-
 	/* Prepare WRITE WR */
 	memset(&write_wr, 0, sizeof write_wr);
 	ctxt->wr_op = IB_WR_RDMA_WRITE;
@@ -226,7 +355,10 @@ static int send_write_chunks(struct svcxprt_rdma *xprt,
 	res_ary = (struct rpcrdma_write_array *)
 		&rdma_resp->rm_body.rm_chunks[1];
 
-	max_write = xprt->sc_max_sge * PAGE_SIZE;
+	if (vec->frmr)
+		max_write = vec->frmr->map_len;
+	else
+		max_write = xprt->sc_max_sge * PAGE_SIZE;
 
 	/* Write chunks start at the pagelist */
 	for (xdr_off = rqstp->rq_res.head[0].iov_len, chunk_no = 0;
@@ -297,7 +429,10 @@ static int send_reply_chunks(struct svcxprt_rdma *xprt,
 	res_ary = (struct rpcrdma_write_array *)
 		&rdma_resp->rm_body.rm_chunks[2];
 
-	max_write = xprt->sc_max_sge * PAGE_SIZE;
+	if (vec->frmr)
+		max_write = vec->frmr->map_len;
+	else
+		max_write = xprt->sc_max_sge * PAGE_SIZE;
 
 	/* xdr offset starts at RPC message */
 	for (xdr_off = 0, chunk_no = 0;
@@ -307,7 +442,6 @@ static int send_reply_chunks(struct svcxprt_rdma *xprt,
 		ch = &arg_ary->wc_array[chunk_no].wc_target;
 		write_len = min(xfer_len, ch->rs_length);
 
-
 		/* Prepare the reply chunk given the length actually
 		 * written */
 		rs_offset = get_unaligned(&(ch->rs_offset));
@@ -366,6 +500,7 @@ static int send_reply(struct svcxprt_rdma *rdma,
 		      int byte_count)
 {
 	struct ib_send_wr send_wr;
+	struct ib_send_wr inv_wr;
 	int sge_no;
 	int sge_bytes;
 	int page_no;
@@ -385,27 +520,45 @@ static int send_reply(struct svcxprt_rdma *rdma,
 	/* Prepare the context */
 	ctxt->pages[0] = page;
 	ctxt->count = 1;
+	ctxt->frmr = vec->frmr;
+	if (vec->frmr)
+		set_bit(RDMACTXT_F_FAST_UNREG, &ctxt->flags);
+	else
+		clear_bit(RDMACTXT_F_FAST_UNREG, &ctxt->flags);
 
 	/* Prepare the SGE for the RPCRDMA Header */
-	atomic_inc(&rdma->sc_dma_used);
 	ctxt->sge[0].addr =
 		ib_dma_map_page(rdma->sc_cm_id->device,
 				page, 0, PAGE_SIZE, DMA_TO_DEVICE);
+	if (ib_dma_mapping_error(rdma->sc_cm_id->device, ctxt->sge[0].addr))
+		goto err;
+	atomic_inc(&rdma->sc_dma_used);
+
 	ctxt->direction = DMA_TO_DEVICE;
+
 	ctxt->sge[0].length = svc_rdma_xdr_get_reply_hdr_len(rdma_resp);
-	ctxt->sge[0].lkey = rdma->sc_phys_mr->lkey;
+	ctxt->sge[0].lkey = rdma->sc_dma_lkey;
 
 	/* Determine how many of our SGE are to be transmitted */
 	for (sge_no = 1; byte_count && sge_no < vec->count; sge_no++) {
 		sge_bytes = min_t(size_t, vec->sge[sge_no].iov_len, byte_count);
 		byte_count -= sge_bytes;
-		atomic_inc(&rdma->sc_dma_used);
-		ctxt->sge[sge_no].addr =
-			ib_dma_map_single(rdma->sc_cm_id->device,
-					  vec->sge[sge_no].iov_base,
-					  sge_bytes, DMA_TO_DEVICE);
+		if (!vec->frmr) {
+			ctxt->sge[sge_no].addr =
+				ib_dma_map_single(rdma->sc_cm_id->device,
+						  vec->sge[sge_no].iov_base,
+						  sge_bytes, DMA_TO_DEVICE);
+			if (ib_dma_mapping_error(rdma->sc_cm_id->device,
+						 ctxt->sge[sge_no].addr))
+				goto err;
+			atomic_inc(&rdma->sc_dma_used);
+			ctxt->sge[sge_no].lkey = rdma->sc_dma_lkey;
+		} else {
+			ctxt->sge[sge_no].addr = (unsigned long)
+				vec->sge[sge_no].iov_base;
+			ctxt->sge[sge_no].lkey = vec->frmr->mr->lkey;
+		}
 		ctxt->sge[sge_no].length = sge_bytes;
-		ctxt->sge[sge_no].lkey = rdma->sc_phys_mr->lkey;
 	}
 	BUG_ON(byte_count != 0);
 
@@ -417,11 +570,16 @@ static int send_reply(struct svcxprt_rdma *rdma,
 		ctxt->pages[page_no+1] = rqstp->rq_respages[page_no];
 		ctxt->count++;
 		rqstp->rq_respages[page_no] = NULL;
-		/* If there are more pages than SGE, terminate SGE list */
+		/*
+		 * If there are more pages than SGE, terminate SGE
+		 * list so that svc_rdma_unmap_dma doesn't attempt to
+		 * unmap garbage.
+		 */
 		if (page_no+1 >= sge_no)
 			ctxt->sge[page_no+1].length = 0;
 	}
 	BUG_ON(sge_no > rdma->sc_max_sge);
+	BUG_ON(sge_no > ctxt->count);
 	memset(&send_wr, 0, sizeof send_wr);
 	ctxt->wr_op = IB_WR_SEND;
 	send_wr.wr_id = (unsigned long)ctxt;
@@ -429,12 +587,26 @@ static int send_reply(struct svcxprt_rdma *rdma,
 	send_wr.num_sge = sge_no;
 	send_wr.opcode = IB_WR_SEND;
 	send_wr.send_flags =  IB_SEND_SIGNALED;
+	if (vec->frmr) {
+		/* Prepare INVALIDATE WR */
+		memset(&inv_wr, 0, sizeof inv_wr);
+		inv_wr.opcode = IB_WR_LOCAL_INV;
+		inv_wr.send_flags = IB_SEND_SIGNALED;
+		inv_wr.ex.invalidate_rkey =
+			vec->frmr->mr->lkey;
+		send_wr.next = &inv_wr;
+	}
 
 	ret = svc_rdma_send(rdma, &send_wr);
 	if (ret)
-		svc_rdma_put_context(ctxt, 1);
+		goto err;
 
-	return ret;
+	return 0;
+
+ err:
+	svc_rdma_put_frmr(rdma, vec->frmr);
+	svc_rdma_put_context(ctxt, 1);
+	return -EIO;
 }
 
 void svc_rdma_prep_reply_hdr(struct svc_rqst *rqstp)
@@ -477,8 +649,9 @@ int svc_rdma_sendto(struct svc_rqst *rqstp)
 	ctxt = svc_rdma_get_context(rdma);
 	ctxt->direction = DMA_TO_DEVICE;
 	vec = svc_rdma_get_req_map();
-	xdr_to_sge(rdma, &rqstp->rq_res, vec);
-
+	ret = map_xdr(rdma, &rqstp->rq_res, vec);
+	if (ret)
+		goto err0;
 	inline_bytes = rqstp->rq_res.len;
 
 	/* Create the RDMA response header */
@@ -498,7 +671,7 @@ int svc_rdma_sendto(struct svc_rqst *rqstp)
 	if (ret < 0) {
 		printk(KERN_ERR "svcrdma: failed to send write chunks, rc=%d\n",
 		       ret);
-		goto error;
+		goto err1;
 	}
 	inline_bytes -= ret;
 
@@ -508,7 +681,7 @@ int svc_rdma_sendto(struct svc_rqst *rqstp)
 	if (ret < 0) {
 		printk(KERN_ERR "svcrdma: failed to send reply chunks, rc=%d\n",
 		       ret);
-		goto error;
+		goto err1;
 	}
 	inline_bytes -= ret;
 
@@ -517,9 +690,11 @@ int svc_rdma_sendto(struct svc_rqst *rqstp)
 	svc_rdma_put_req_map(vec);
 	dprintk("svcrdma: send_reply returns %d\n", ret);
 	return ret;
- error:
+
+ err1:
+	put_page(res_page);
+ err0:
 	svc_rdma_put_req_map(vec);
 	svc_rdma_put_context(ctxt, 0);
-	put_page(res_page);
 	return ret;
 }

net/sunrpc/xprtrdma/svc_rdma_transport.c

@@ -100,20 +100,29 @@ struct svc_rdma_op_ctxt *svc_rdma_get_context(struct svcxprt_rdma *xprt)
 	ctxt->xprt = xprt;
 	INIT_LIST_HEAD(&ctxt->dto_q);
 	ctxt->count = 0;
+	ctxt->frmr = NULL;
 	atomic_inc(&xprt->sc_ctxt_used);
 	return ctxt;
 }
 
-static void svc_rdma_unmap_dma(struct svc_rdma_op_ctxt *ctxt)
+void svc_rdma_unmap_dma(struct svc_rdma_op_ctxt *ctxt)
 {
 	struct svcxprt_rdma *xprt = ctxt->xprt;
 	int i;
 	for (i = 0; i < ctxt->count && ctxt->sge[i].length; i++) {
-		atomic_dec(&xprt->sc_dma_used);
-		ib_dma_unmap_single(xprt->sc_cm_id->device,
-				    ctxt->sge[i].addr,
-				    ctxt->sge[i].length,
-				    ctxt->direction);
+		/*
+		 * Unmap the DMA addr in the SGE if the lkey matches
+		 * the sc_dma_lkey, otherwise, ignore it since it is
+		 * an FRMR lkey and will be unmapped later when the
+		 * last WR that uses it completes.
+		 */
+		if (ctxt->sge[i].lkey == xprt->sc_dma_lkey) {
+			atomic_dec(&xprt->sc_dma_used);
+			ib_dma_unmap_single(xprt->sc_cm_id->device,
+					    ctxt->sge[i].addr,
+					    ctxt->sge[i].length,
+					    ctxt->direction);
+		}
 	}
 }
 
@@ -150,6 +159,7 @@ struct svc_rdma_req_map *svc_rdma_get_req_map(void)
 		schedule_timeout_uninterruptible(msecs_to_jiffies(500));
 	}
 	map->count = 0;
+	map->frmr = NULL;
 	return map;
 }
 
@@ -315,6 +325,50 @@ static void rq_cq_reap(struct svcxprt_rdma *xprt)
 		svc_xprt_enqueue(&xprt->sc_xprt);
 }
 
+/*
+ * Processs a completion context
+ */
+static void process_context(struct svcxprt_rdma *xprt,
+			    struct svc_rdma_op_ctxt *ctxt)
+{
+	svc_rdma_unmap_dma(ctxt);
+
+	switch (ctxt->wr_op) {
+	case IB_WR_SEND:
+		if (test_bit(RDMACTXT_F_FAST_UNREG, &ctxt->flags))
+			svc_rdma_put_frmr(xprt, ctxt->frmr);
+		svc_rdma_put_context(ctxt, 1);
+		break;
+
+	case IB_WR_RDMA_WRITE:
+		svc_rdma_put_context(ctxt, 0);
+		break;
+
+	case IB_WR_RDMA_READ:
+	case IB_WR_RDMA_READ_WITH_INV:
+		if (test_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags)) {
+			struct svc_rdma_op_ctxt *read_hdr = ctxt->read_hdr;
+			BUG_ON(!read_hdr);
+			if (test_bit(RDMACTXT_F_FAST_UNREG, &ctxt->flags))
+				svc_rdma_put_frmr(xprt, ctxt->frmr);
+			spin_lock_bh(&xprt->sc_rq_dto_lock);
+			set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
+			list_add_tail(&read_hdr->dto_q,
+				      &xprt->sc_read_complete_q);
+			spin_unlock_bh(&xprt->sc_rq_dto_lock);
+			svc_xprt_enqueue(&xprt->sc_xprt);
+		}
+		svc_rdma_put_context(ctxt, 0);
+		break;
+
+	default:
+		printk(KERN_ERR "svcrdma: unexpected completion type, "
+		       "opcode=%d\n",
+		       ctxt->wr_op);
+		break;
+	}
+}
+
 /*
  * Send Queue Completion Handler - potentially called on interrupt context.
  *
@@ -327,17 +381,12 @@ static void sq_cq_reap(struct svcxprt_rdma *xprt)
 	struct ib_cq *cq = xprt->sc_sq_cq;
 	int ret;
 
-
 	if (!test_and_clear_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags))
 		return;
 
 	ib_req_notify_cq(xprt->sc_sq_cq, IB_CQ_NEXT_COMP);
 	atomic_inc(&rdma_stat_sq_poll);
 	while ((ret = ib_poll_cq(cq, 1, &wc)) > 0) {
-		ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id;
-		xprt = ctxt->xprt;
-
-		svc_rdma_unmap_dma(ctxt);
 		if (wc.status != IB_WC_SUCCESS)
 			/* Close the transport */
 			set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
@@ -346,35 +395,10 @@ static void sq_cq_reap(struct svcxprt_rdma *xprt)
 		atomic_dec(&xprt->sc_sq_count);
 		wake_up(&xprt->sc_send_wait);
 
-		switch (ctxt->wr_op) {
-		case IB_WR_SEND:
-			svc_rdma_put_context(ctxt, 1);
-			break;
-
-		case IB_WR_RDMA_WRITE:
-			svc_rdma_put_context(ctxt, 0);
-			break;
-
-		case IB_WR_RDMA_READ:
-			if (test_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags)) {
-				struct svc_rdma_op_ctxt *read_hdr = ctxt->read_hdr;
-				BUG_ON(!read_hdr);
-				spin_lock_bh(&xprt->sc_rq_dto_lock);
-				set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
-				list_add_tail(&read_hdr->dto_q,
-					      &xprt->sc_read_complete_q);
-				spin_unlock_bh(&xprt->sc_rq_dto_lock);
-				svc_xprt_enqueue(&xprt->sc_xprt);
-			}
-			svc_rdma_put_context(ctxt, 0);
-			break;
+		ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id;
+		if (ctxt)
+			process_context(xprt, ctxt);
 
-		default:
-			printk(KERN_ERR "svcrdma: unexpected completion type, "
-			       "opcode=%d, status=%d\n",
-			       wc.opcode, wc.status);
-			break;
-		}
 		svc_xprt_put(&xprt->sc_xprt);
 	}
 
@@ -425,10 +449,12 @@ static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *serv,
 	INIT_LIST_HEAD(&cma_xprt->sc_dto_q);
 	INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q);
 	INIT_LIST_HEAD(&cma_xprt->sc_read_complete_q);
+	INIT_LIST_HEAD(&cma_xprt->sc_frmr_q);
 	init_waitqueue_head(&cma_xprt->sc_send_wait);
 
 	spin_lock_init(&cma_xprt->sc_lock);
 	spin_lock_init(&cma_xprt->sc_rq_dto_lock);
+	spin_lock_init(&cma_xprt->sc_frmr_q_lock);
 
 	cma_xprt->sc_ord = svcrdma_ord;
 
@@ -462,7 +488,7 @@ int svc_rdma_post_recv(struct svcxprt_rdma *xprt)
 	struct ib_recv_wr recv_wr, *bad_recv_wr;
 	struct svc_rdma_op_ctxt *ctxt;
 	struct page *page;
-	unsigned long pa;
+	dma_addr_t pa;
 	int sge_no;
 	int buflen;
 	int ret;
@@ -474,13 +500,15 @@ int svc_rdma_post_recv(struct svcxprt_rdma *xprt)
 		BUG_ON(sge_no >= xprt->sc_max_sge);
 		page = svc_rdma_get_page();
 		ctxt->pages[sge_no] = page;
-		atomic_inc(&xprt->sc_dma_used);
 		pa = ib_dma_map_page(xprt->sc_cm_id->device,
 				     page, 0, PAGE_SIZE,
 				     DMA_FROM_DEVICE);
+		if (ib_dma_mapping_error(xprt->sc_cm_id->device, pa))
+			goto err_put_ctxt;
+		atomic_inc(&xprt->sc_dma_used);
 		ctxt->sge[sge_no].addr = pa;
 		ctxt->sge[sge_no].length = PAGE_SIZE;
-		ctxt->sge[sge_no].lkey = xprt->sc_phys_mr->lkey;
+		ctxt->sge[sge_no].lkey = xprt->sc_dma_lkey;
 		buflen += PAGE_SIZE;
 	}
 	ctxt->count = sge_no;
@@ -496,6 +524,10 @@ int svc_rdma_post_recv(struct svcxprt_rdma *xprt)
 		svc_rdma_put_context(ctxt, 1);
 	}
 	return ret;
+
+ err_put_ctxt:
+	svc_rdma_put_context(ctxt, 1);
+	return -ENOMEM;
 }
 
 /*
@@ -566,7 +598,7 @@ static int rdma_listen_handler(struct rdma_cm_id *cma_id,
 		dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, "
 			"event=%d\n", cma_id, cma_id->context, event->event);
 		handle_connect_req(cma_id,
-				   event->param.conn.responder_resources);
+				   event->param.conn.initiator_depth);
 		break;
 
 	case RDMA_CM_EVENT_ESTABLISHED:
@@ -686,6 +718,97 @@ static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
 	return ERR_PTR(ret);
 }
 
+static struct svc_rdma_fastreg_mr *rdma_alloc_frmr(struct svcxprt_rdma *xprt)
+{
+	struct ib_mr *mr;
+	struct ib_fast_reg_page_list *pl;
+	struct svc_rdma_fastreg_mr *frmr;
+
+	frmr = kmalloc(sizeof(*frmr), GFP_KERNEL);
+	if (!frmr)
+		goto err;
+
+	mr = ib_alloc_fast_reg_mr(xprt->sc_pd, RPCSVC_MAXPAGES);
+	if (!mr)
+		goto err_free_frmr;
+
+	pl = ib_alloc_fast_reg_page_list(xprt->sc_cm_id->device,
+					 RPCSVC_MAXPAGES);
+	if (!pl)
+		goto err_free_mr;
+
+	frmr->mr = mr;
+	frmr->page_list = pl;
+	INIT_LIST_HEAD(&frmr->frmr_list);
+	return frmr;
+
+ err_free_mr:
+	ib_dereg_mr(mr);
+ err_free_frmr:
+	kfree(frmr);
+ err:
+	return ERR_PTR(-ENOMEM);
+}
+
+static void rdma_dealloc_frmr_q(struct svcxprt_rdma *xprt)
+{
+	struct svc_rdma_fastreg_mr *frmr;
+
+	while (!list_empty(&xprt->sc_frmr_q)) {
+		frmr = list_entry(xprt->sc_frmr_q.next,
+				  struct svc_rdma_fastreg_mr, frmr_list);
+		list_del_init(&frmr->frmr_list);
+		ib_dereg_mr(frmr->mr);
+		ib_free_fast_reg_page_list(frmr->page_list);
+		kfree(frmr);
+	}
+}
+
+struct svc_rdma_fastreg_mr *svc_rdma_get_frmr(struct svcxprt_rdma *rdma)
+{
+	struct svc_rdma_fastreg_mr *frmr = NULL;
+
+	spin_lock_bh(&rdma->sc_frmr_q_lock);
+	if (!list_empty(&rdma->sc_frmr_q)) {
+		frmr = list_entry(rdma->sc_frmr_q.next,
+				  struct svc_rdma_fastreg_mr, frmr_list);
+		list_del_init(&frmr->frmr_list);
+		frmr->map_len = 0;
+		frmr->page_list_len = 0;
+	}
+	spin_unlock_bh(&rdma->sc_frmr_q_lock);
+	if (frmr)
+		return frmr;
+
+	return rdma_alloc_frmr(rdma);
+}
+
+static void frmr_unmap_dma(struct svcxprt_rdma *xprt,
+			   struct svc_rdma_fastreg_mr *frmr)
+{
+	int page_no;
+	for (page_no = 0; page_no < frmr->page_list_len; page_no++) {
+		dma_addr_t addr = frmr->page_list->page_list[page_no];
+		if (ib_dma_mapping_error(frmr->mr->device, addr))
+			continue;
+		atomic_dec(&xprt->sc_dma_used);
+		ib_dma_unmap_single(frmr->mr->device, addr, PAGE_SIZE,
+				    frmr->direction);
+	}
+}
+
+void svc_rdma_put_frmr(struct svcxprt_rdma *rdma,
+		       struct svc_rdma_fastreg_mr *frmr)
+{
+	if (frmr) {
+		frmr_unmap_dma(rdma, frmr);
+		spin_lock_bh(&rdma->sc_frmr_q_lock);
+		BUG_ON(!list_empty(&frmr->frmr_list));
+		list_add(&frmr->frmr_list, &rdma->sc_frmr_q);
+		spin_unlock_bh(&rdma->sc_frmr_q_lock);
+	}
+}
+
 /*
  * This is the xpo_recvfrom function for listening endpoints. Its
  * purpose is to accept incoming connections. The CMA callback handler
@@ -704,6 +827,8 @@ static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
 	struct rdma_conn_param conn_param;
 	struct ib_qp_init_attr qp_attr;
 	struct ib_device_attr devattr;
+	int dma_mr_acc;
+	int need_dma_mr;
 	int ret;
 	int i;
 
@@ -819,15 +944,77 @@ static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
 	}
 	newxprt->sc_qp = newxprt->sc_cm_id->qp;
 
-	/* Register all of physical memory */
-	newxprt->sc_phys_mr = ib_get_dma_mr(newxprt->sc_pd,
-					    IB_ACCESS_LOCAL_WRITE |
-					    IB_ACCESS_REMOTE_WRITE);
-	if (IS_ERR(newxprt->sc_phys_mr)) {
-		dprintk("svcrdma: Failed to create DMA MR ret=%d\n", ret);
+	/*
+	 * Use the most secure set of MR resources based on the
+	 * transport type and available memory management features in
+	 * the device. Here's the table implemented below:
+	 *
+	 *		Fast	Global	DMA	Remote WR
+	 *		Reg	LKEY	MR	Access
+	 *		Sup'd	Sup'd	Needed	Needed
+	 *
+	 * IWARP	N	N	Y	Y
+	 *		N	Y	Y	Y
+	 *		Y	N	Y	N
+	 *		Y	Y	N	-
+	 *
+	 * IB		N	N	Y	N
+	 *		N	Y	N	-
+	 *		Y	N	Y	N
+	 *		Y	Y	N	-
+	 *
+	 * NB:	iWARP requires remote write access for the data sink
+	 *	of an RDMA_READ. IB does not.
+	 */
+	if (devattr.device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
+		newxprt->sc_frmr_pg_list_len =
+			devattr.max_fast_reg_page_list_len;
+		newxprt->sc_dev_caps |= SVCRDMA_DEVCAP_FAST_REG;
+	}
+
+	/*
+	 * Determine if a DMA MR is required and if so, what privs are required
+	 */
+	switch (rdma_node_get_transport(newxprt->sc_cm_id->device->node_type)) {
+	case RDMA_TRANSPORT_IWARP:
+		newxprt->sc_dev_caps |= SVCRDMA_DEVCAP_READ_W_INV;
+		if (!(newxprt->sc_dev_caps & SVCRDMA_DEVCAP_FAST_REG)) {
+			need_dma_mr = 1;
+			dma_mr_acc =
+				(IB_ACCESS_LOCAL_WRITE |
+				 IB_ACCESS_REMOTE_WRITE);
+		} else if (!(devattr.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)) {
+			need_dma_mr = 1;
+			dma_mr_acc = IB_ACCESS_LOCAL_WRITE;
+		} else
+			need_dma_mr = 0;
+		break;
+	case RDMA_TRANSPORT_IB:
+		if (!(devattr.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)) {
+			need_dma_mr = 1;
+			dma_mr_acc = IB_ACCESS_LOCAL_WRITE;
+		} else
+			need_dma_mr = 0;
+		break;
+	default:
 		goto errout;
 	}
 
+	/* Create the DMA MR if needed, otherwise, use the DMA LKEY */
+	if (need_dma_mr) {
+		/* Register all of physical memory */
+		newxprt->sc_phys_mr =
+			ib_get_dma_mr(newxprt->sc_pd, dma_mr_acc);
+		if (IS_ERR(newxprt->sc_phys_mr)) {
+			dprintk("svcrdma: Failed to create DMA MR ret=%d\n",
+				ret);
+			goto errout;
+		}
+		newxprt->sc_dma_lkey = newxprt->sc_phys_mr->lkey;
+	} else
+		newxprt->sc_dma_lkey =
+			newxprt->sc_cm_id->device->local_dma_lkey;
+
 	/* Post receive buffers */
 	for (i = 0; i < newxprt->sc_max_requests; i++) {
 		ret = svc_rdma_post_recv(newxprt);
@@ -961,6 +1148,9 @@ static void __svc_rdma_free(struct work_struct *work)
 	WARN_ON(atomic_read(&rdma->sc_ctxt_used) != 0);
 	WARN_ON(atomic_read(&rdma->sc_dma_used) != 0);
 
+	/* De-allocate fastreg mr */
+	rdma_dealloc_frmr_q(rdma);
+
 	/* Destroy the QP if present (not a listener) */
 	if (rdma->sc_qp && !IS_ERR(rdma->sc_qp))
 		ib_destroy_qp(rdma->sc_qp);
@@ -1014,21 +1204,59 @@ static int svc_rdma_has_wspace(struct svc_xprt *xprt)
 	return 1;
 }
 
+/*
+ * Attempt to register the kvec representing the RPC memory with the
+ * device.
+ *
+ * Returns:
+ *  NULL : The device does not support fastreg or there were no more
+ *         fastreg mr.
+ *  frmr : The kvec register request was successfully posted.
+ *    <0 : An error was encountered attempting to register the kvec.
+ */
+int svc_rdma_fastreg(struct svcxprt_rdma *xprt,
+		     struct svc_rdma_fastreg_mr *frmr)
+{
+	struct ib_send_wr fastreg_wr;
+	u8 key;
+
+	/* Bump the key */
+	key = (u8)(frmr->mr->lkey & 0x000000FF);
+	ib_update_fast_reg_key(frmr->mr, ++key);
+
+	/* Prepare FASTREG WR */
+	memset(&fastreg_wr, 0, sizeof fastreg_wr);
+	fastreg_wr.opcode = IB_WR_FAST_REG_MR;
+	fastreg_wr.send_flags = IB_SEND_SIGNALED;
+	fastreg_wr.wr.fast_reg.iova_start = (unsigned long)frmr->kva;
+	fastreg_wr.wr.fast_reg.page_list = frmr->page_list;
+	fastreg_wr.wr.fast_reg.page_list_len = frmr->page_list_len;
+	fastreg_wr.wr.fast_reg.page_shift = PAGE_SHIFT;
+	fastreg_wr.wr.fast_reg.length = frmr->map_len;
+	fastreg_wr.wr.fast_reg.access_flags = frmr->access_flags;
+	fastreg_wr.wr.fast_reg.rkey = frmr->mr->lkey;
+	return svc_rdma_send(xprt, &fastreg_wr);
+}
+
 int svc_rdma_send(struct svcxprt_rdma *xprt, struct ib_send_wr *wr)
 {
-	struct ib_send_wr *bad_wr;
+	struct ib_send_wr *bad_wr, *n_wr;
+	int wr_count;
+	int i;
 	int ret;
 
 	if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
 		return -ENOTCONN;
 
 	BUG_ON(wr->send_flags != IB_SEND_SIGNALED);
-	BUG_ON(((struct svc_rdma_op_ctxt *)(unsigned long)wr->wr_id)->wr_op !=
-		wr->opcode);
+	wr_count = 1;
+	for (n_wr = wr->next; n_wr; n_wr = n_wr->next)
+		wr_count++;
+
 	/* If the SQ is full, wait until an SQ entry is available */
 	while (1) {
 		spin_lock_bh(&xprt->sc_lock);
-		if (xprt->sc_sq_depth == atomic_read(&xprt->sc_sq_count)) {
+		if (xprt->sc_sq_depth < atomic_read(&xprt->sc_sq_count) + wr_count) {
 			spin_unlock_bh(&xprt->sc_lock);
 			atomic_inc(&rdma_stat_sq_starve);
 
@@ -1043,19 +1271,26 @@ int svc_rdma_send(struct svcxprt_rdma *xprt, struct ib_send_wr *wr)
 				return 0;
 			continue;
 		}
-		/* Bumped used SQ WR count and post */
-		svc_xprt_get(&xprt->sc_xprt);
+		/* Take a transport ref for each WR posted */
+		for (i = 0; i < wr_count; i++)
+			svc_xprt_get(&xprt->sc_xprt);
+
+		/* Bump used SQ WR count and post */
+		atomic_add(wr_count, &xprt->sc_sq_count);
 		ret = ib_post_send(xprt->sc_qp, wr, &bad_wr);
-		if (!ret)
-			atomic_inc(&xprt->sc_sq_count);
-		else {
-			svc_xprt_put(&xprt->sc_xprt);
+		if (ret) {
+			set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
+			atomic_sub(wr_count, &xprt->sc_sq_count);
+			for (i = 0; i < wr_count; i ++)
+				svc_xprt_put(&xprt->sc_xprt);
 			dprintk("svcrdma: failed to post SQ WR rc=%d, "
 			       "sc_sq_count=%d, sc_sq_depth=%d\n",
 			       ret, atomic_read(&xprt->sc_sq_count),
 			       xprt->sc_sq_depth);
 		}
 		spin_unlock_bh(&xprt->sc_lock);
+		if (ret)
+			wake_up(&xprt->sc_send_wait);
 		break;
 	}
 	return ret;
@@ -1079,10 +1314,14 @@ void svc_rdma_send_error(struct svcxprt_rdma *xprt, struct rpcrdma_msg *rmsgp,
 	length = svc_rdma_xdr_encode_error(xprt, rmsgp, err, va);
 
 	/* Prepare SGE for local address */
-	atomic_inc(&xprt->sc_dma_used);
 	sge.addr = ib_dma_map_page(xprt->sc_cm_id->device,
 				   p, 0, PAGE_SIZE, DMA_FROM_DEVICE);
-	sge.lkey = xprt->sc_phys_mr->lkey;
+	if (ib_dma_mapping_error(xprt->sc_cm_id->device, sge.addr)) {
+		put_page(p);
+		return;
+	}
+	atomic_inc(&xprt->sc_dma_used);
+	sge.lkey = xprt->sc_dma_lkey;
 	sge.length = length;
 
 	ctxt = svc_rdma_get_context(xprt);
@@ -1103,6 +1342,9 @@ void svc_rdma_send_error(struct svcxprt_rdma *xprt, struct rpcrdma_msg *rmsgp,
 	if (ret) {
 		dprintk("svcrdma: Error %d posting send for protocol error\n",
 			ret);
+		ib_dma_unmap_page(xprt->sc_cm_id->device,
+				  sge.addr, PAGE_SIZE,
+				  DMA_FROM_DEVICE);
 		svc_rdma_put_context(ctxt, 1);
 	}
 }