svcrdma: refactor marshalling logic
authorSteve Wise <swise@opengridcomputing.com>
Wed, 28 May 2014 20:12:01 +0000 (15:12 -0500)
committerJ. Bruce Fields <bfields@redhat.com>
Fri, 6 Jun 2014 23:22:50 +0000 (19:22 -0400)
This patch refactors the NFSRDMA server marshalling logic to
remove the intermediary map structures.  It also fixes an existing bug
where the NFSRDMA server was not minding the device fast register page
list length limitations.

Signed-off-by: Tom Tucker <tom@opengridcomputing.com>
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
include/linux/sunrpc/svc_rdma.h
net/sunrpc/xprtrdma/svc_rdma_recvfrom.c
net/sunrpc/xprtrdma/svc_rdma_sendto.c
net/sunrpc/xprtrdma/svc_rdma_transport.c

index 0b8e3e6bdacf0b6af2e32541d6ad83cf52af8f6e..5cf99a0163681ec2a1bee5c6a97171e82f361c53 100644 (file)
@@ -115,14 +115,13 @@ struct svc_rdma_fastreg_mr {
        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];
+               unsigned long lkey[RPCSVC_MAXPAGES];
        };
 };
-#define RDMACTXT_F_FAST_UNREG  1
 #define RDMACTXT_F_LAST_CTXT   2
 
 #define        SVCRDMA_DEVCAP_FAST_REG         1       /* fast mr registration */
index 8d904e4eef1551f1c3a3c9281b5aa7a7ae39fb66..52d9f2ce20b06c27fb3b31fa4af251c09fb05e1f 100644 (file)
@@ -1,4 +1,5 @@
 /*
+ * Copyright (c) 2014 Open Grid Computing, Inc. All rights reserved.
  * Copyright (c) 2005-2006 Network Appliance, Inc. All rights reserved.
  *
  * This software is available to you under a choice of one of two
@@ -69,7 +70,8 @@ static void rdma_build_arg_xdr(struct svc_rqst *rqstp,
 
        /* Set up the XDR head */
        rqstp->rq_arg.head[0].iov_base = page_address(page);
-       rqstp->rq_arg.head[0].iov_len = min(byte_count, ctxt->sge[0].length);
+       rqstp->rq_arg.head[0].iov_len =
+               min_t(size_t, byte_count, ctxt->sge[0].length);
        rqstp->rq_arg.len = byte_count;
        rqstp->rq_arg.buflen = byte_count;
 
@@ -85,7 +87,7 @@ static void rdma_build_arg_xdr(struct svc_rqst *rqstp,
                page = ctxt->pages[sge_no];
                put_page(rqstp->rq_pages[sge_no]);
                rqstp->rq_pages[sge_no] = page;
-               bc -= min(bc, ctxt->sge[sge_no].length);
+               bc -= min_t(u32, bc, ctxt->sge[sge_no].length);
                rqstp->rq_arg.buflen += ctxt->sge[sge_no].length;
                sge_no++;
        }
@@ -113,291 +115,265 @@ static void rdma_build_arg_xdr(struct svc_rqst *rqstp,
        rqstp->rq_arg.tail[0].iov_len = 0;
 }
 
-/* Encode a read-chunk-list as an array of IB SGE
- *
- * Assumptions:
- * - chunk[0]->position points to pages[0] at an offset of 0
- * - pages[] is not physically or virtually contiguous and consists of
- *   PAGE_SIZE elements.
- *
- * Output:
- * - sge array pointing into pages[] array.
- * - chunk_sge array specifying sge index and count for each
- *   chunk in the read list
- *
- */
-static int map_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)
+static int rdma_read_max_sge(struct svcxprt_rdma *xprt, int sge_count)
 {
-       int sge_no;
-       int sge_bytes;
-       int page_off;
-       int page_no;
-       int ch_bytes;
-       int ch_no;
-       struct rpcrdma_read_chunk *ch;
+       if (rdma_node_get_transport(xprt->sc_cm_id->device->node_type) ==
+            RDMA_TRANSPORT_IWARP)
+               return 1;
+       else
+               return min_t(int, sge_count, xprt->sc_max_sge);
+}
 
-       sge_no = 0;
-       page_no = 0;
-       page_off = 0;
-       ch = (struct rpcrdma_read_chunk *)&rmsgp->rm_body.rm_chunks[0];
-       ch_no = 0;
-       ch_bytes = ntohl(ch->rc_target.rs_length);
-       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 = ch_bytes;
-       head->arg.len = rqstp->rq_arg.len + ch_bytes;
-       head->arg.buflen = rqstp->rq_arg.buflen + ch_bytes;
-       head->count++;
-       chl_map->ch[0].start = 0;
-       while (byte_count) {
-               rpl_map->sge[sge_no].iov_base =
-                       page_address(rqstp->rq_arg.pages[page_no]) + page_off;
-               sge_bytes = min_t(int, PAGE_SIZE-page_off, ch_bytes);
-               rpl_map->sge[sge_no].iov_len = sge_bytes;
-               /*
-                * Don't bump head->count here because the same page
-                * may be used by multiple SGE.
-                */
-               head->arg.pages[page_no] = rqstp->rq_arg.pages[page_no];
-               rqstp->rq_respages = &rqstp->rq_arg.pages[page_no+1];
+typedef int (*rdma_reader_fn)(struct svcxprt_rdma *xprt,
+                             struct svc_rqst *rqstp,
+                             struct svc_rdma_op_ctxt *head,
+                             int *page_no,
+                             u32 *page_offset,
+                             u32 rs_handle,
+                             u32 rs_length,
+                             u64 rs_offset,
+                             int last);
+
+/* Issue an RDMA_READ using the local lkey to map the data sink */
+static int rdma_read_chunk_lcl(struct svcxprt_rdma *xprt,
+                              struct svc_rqst *rqstp,
+                              struct svc_rdma_op_ctxt *head,
+                              int *page_no,
+                              u32 *page_offset,
+                              u32 rs_handle,
+                              u32 rs_length,
+                              u64 rs_offset,
+                              int last)
+{
+       struct ib_send_wr read_wr;
+       int pages_needed = PAGE_ALIGN(*page_offset + rs_length) >> PAGE_SHIFT;
+       struct svc_rdma_op_ctxt *ctxt = svc_rdma_get_context(xprt);
+       int ret, read, pno;
+       u32 pg_off = *page_offset;
+       u32 pg_no = *page_no;
+
+       ctxt->direction = DMA_FROM_DEVICE;
+       ctxt->read_hdr = head;
+       pages_needed =
+               min_t(int, pages_needed, rdma_read_max_sge(xprt, pages_needed));
+       read = min_t(int, pages_needed << PAGE_SHIFT, rs_length);
+
+       for (pno = 0; pno < pages_needed; pno++) {
+               int len = min_t(int, rs_length, PAGE_SIZE - pg_off);
+
+               head->arg.pages[pg_no] = rqstp->rq_arg.pages[pg_no];
+               head->arg.page_len += len;
+               head->arg.len += len;
+               if (!pg_off)
+                       head->count++;
+               rqstp->rq_respages = &rqstp->rq_arg.pages[pg_no+1];
                rqstp->rq_next_page = rqstp->rq_respages + 1;
+               ctxt->sge[pno].addr =
+                       ib_dma_map_page(xprt->sc_cm_id->device,
+                                       head->arg.pages[pg_no], pg_off,
+                                       PAGE_SIZE - pg_off,
+                                       DMA_FROM_DEVICE);
+               ret = ib_dma_mapping_error(xprt->sc_cm_id->device,
+                                          ctxt->sge[pno].addr);
+               if (ret)
+                       goto err;
+               atomic_inc(&xprt->sc_dma_used);
 
-               byte_count -= sge_bytes;
-               ch_bytes -= sge_bytes;
-               sge_no++;
-               /*
-                * If all bytes for this chunk have been mapped to an
-                * SGE, move to the next SGE
-                */
-               if (ch_bytes == 0) {
-                       chl_map->ch[ch_no].count =
-                               sge_no - chl_map->ch[ch_no].start;
-                       ch_no++;
-                       ch++;
-                       chl_map->ch[ch_no].start = sge_no;
-                       ch_bytes = ntohl(ch->rc_target.rs_length);
-                       /* If bytes remaining account for next chunk */
-                       if (byte_count) {
-                               head->arg.page_len += ch_bytes;
-                               head->arg.len += ch_bytes;
-                               head->arg.buflen += ch_bytes;
-                       }
+               /* The lkey here is either a local dma lkey or a dma_mr lkey */
+               ctxt->sge[pno].lkey = xprt->sc_dma_lkey;
+               ctxt->sge[pno].length = len;
+               ctxt->count++;
+
+               /* adjust offset and wrap to next page if needed */
+               pg_off += len;
+               if (pg_off == PAGE_SIZE) {
+                       pg_off = 0;
+                       pg_no++;
                }
-               /*
-                * If this SGE consumed all of the page, move to the
-                * next page
-                */
-               if ((sge_bytes + page_off) == PAGE_SIZE) {
-                       page_no++;
-                       page_off = 0;
-                       /*
-                        * If there are still bytes left to map, bump
-                        * the page count
-                        */
-                       if (byte_count)
-                               head->count++;
-               } else
-                       page_off += sge_bytes;
+               rs_length -= len;
        }
-       BUG_ON(byte_count != 0);
-       return sge_no;
+
+       if (last && rs_length == 0)
+               set_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags);
+       else
+               clear_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags);
+
+       memset(&read_wr, 0, sizeof(read_wr));
+       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 = rs_handle;
+       read_wr.wr.rdma.remote_addr = rs_offset;
+       read_wr.sg_list = ctxt->sge;
+       read_wr.num_sge = pages_needed;
+
+       ret = svc_rdma_send(xprt, &read_wr);
+       if (ret) {
+               pr_err("svcrdma: Error %d posting RDMA_READ\n", ret);
+               set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
+               goto err;
+       }
+
+       /* return current location in page array */
+       *page_no = pg_no;
+       *page_offset = pg_off;
+       ret = read;
+       atomic_inc(&rdma_stat_read);
+       return ret;
+ err:
+       svc_rdma_unmap_dma(ctxt);
+       svc_rdma_put_context(ctxt, 0);
+       return ret;
 }
 
-/* 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,
+/* Issue an RDMA_READ using an FRMR to map the data sink */
+static int rdma_read_chunk_frmr(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,
+                               u32 *page_offset,
+                               u32 rs_handle,
+                               u32 rs_length,
+                               u64 rs_offset,
+                               int last)
 {
-       int page_no;
-       int ch_no;
-       u32 offset;
-       struct rpcrdma_read_chunk *ch;
-       struct svc_rdma_fastreg_mr *frmr;
-       int ret = 0;
+       struct ib_send_wr read_wr;
+       struct ib_send_wr inv_wr;
+       struct ib_send_wr fastreg_wr;
+       u8 key;
+       int pages_needed = PAGE_ALIGN(*page_offset + rs_length) >> PAGE_SHIFT;
+       struct svc_rdma_op_ctxt *ctxt = svc_rdma_get_context(xprt);
+       struct svc_rdma_fastreg_mr *frmr = svc_rdma_get_frmr(xprt);
+       int ret, read, pno;
+       u32 pg_off = *page_offset;
+       u32 pg_no = *page_no;
 
-       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;
+       ctxt->direction = DMA_FROM_DEVICE;
+       ctxt->frmr = frmr;
+       pages_needed = min_t(int, pages_needed, xprt->sc_frmr_pg_list_len);
+       read = min_t(int, pages_needed << PAGE_SHIFT, rs_length);
 
-       /* Fast register the page list */
-       frmr->kva = page_address(rqstp->rq_arg.pages[0]);
+       frmr->kva = page_address(rqstp->rq_arg.pages[pg_no]);
        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] =
+       frmr->map_len = pages_needed << PAGE_SHIFT;
+       frmr->page_list_len = pages_needed;
+
+       for (pno = 0; pno < pages_needed; pno++) {
+               int len = min_t(int, rs_length, PAGE_SIZE - pg_off);
+
+               head->arg.pages[pg_no] = rqstp->rq_arg.pages[pg_no];
+               head->arg.page_len += len;
+               head->arg.len += len;
+               if (!pg_off)
+                       head->count++;
+               rqstp->rq_respages = &rqstp->rq_arg.pages[pg_no+1];
+               rqstp->rq_next_page = rqstp->rq_respages + 1;
+               frmr->page_list->page_list[pno] =
                        ib_dma_map_page(xprt->sc_cm_id->device,
-                                       rqstp->rq_arg.pages[page_no], 0,
+                                       head->arg.pages[pg_no], 0,
                                        PAGE_SIZE, DMA_FROM_DEVICE);
-               if (ib_dma_mapping_error(xprt->sc_cm_id->device,
-                                        frmr->page_list->page_list[page_no]))
-                       goto fatal_err;
+               ret = ib_dma_mapping_error(xprt->sc_cm_id->device,
+                                          frmr->page_list->page_list[pno]);
+               if (ret)
+                       goto 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];
-       rqstp->rq_next_page = rqstp->rq_respages + 1;
 
-       /* 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++) {
-               int len = ntohl(ch->rc_target.rs_length);
-               rpl_map->sge[ch_no].iov_base = frmr->kva + offset;
-               rpl_map->sge[ch_no].iov_len = len;
-               chl_map->ch[ch_no].count = 1;
-               chl_map->ch[ch_no].start = ch_no;
-               offset += len;
-               ch++;
+               /* adjust offset and wrap to next page if needed */
+               pg_off += len;
+               if (pg_off == PAGE_SIZE) {
+                       pg_off = 0;
+                       pg_no++;
+               }
+               rs_length -= len;
        }
 
-       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;
-       unsigned long off;
+       if (last && rs_length == 0)
+               set_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags);
+       else
+               clear_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags);
 
-       ctxt->count = count;
-       ctxt->direction = DMA_FROM_DEVICE;
-       for (i = 0; i < count; i++) {
-               ctxt->sge[i].length = 0; /* in case map fails */
-               if (!frmr) {
-                       BUG_ON(!virt_to_page(vec[i].iov_base));
-                       off = (unsigned long)vec[i].iov_base & ~PAGE_MASK;
-                       ctxt->sge[i].addr =
-                               ib_dma_map_page(xprt->sc_cm_id->device,
-                                               virt_to_page(vec[i].iov_base),
-                                               off,
-                                               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;
-               *sgl_offset = *sgl_offset + vec[i].iov_len;
+       /* Bump the key */
+       key = (u8)(frmr->mr->lkey & 0x000000FF);
+       ib_update_fast_reg_key(frmr->mr, ++key);
+
+       ctxt->sge[0].addr = (unsigned long)frmr->kva + *page_offset;
+       ctxt->sge[0].lkey = frmr->mr->lkey;
+       ctxt->sge[0].length = read;
+       ctxt->count = 1;
+       ctxt->read_hdr = head;
+
+       /* 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;
+       fastreg_wr.next = &read_wr;
+
+       /* Prepare RDMA_READ */
+       memset(&read_wr, 0, sizeof(read_wr));
+       read_wr.send_flags = IB_SEND_SIGNALED;
+       read_wr.wr.rdma.rkey = rs_handle;
+       read_wr.wr.rdma.remote_addr = rs_offset;
+       read_wr.sg_list = ctxt->sge;
+       read_wr.num_sge = 1;
+       if (xprt->sc_dev_caps & SVCRDMA_DEVCAP_READ_W_INV) {
+               read_wr.opcode = IB_WR_RDMA_READ_WITH_INV;
+               read_wr.wr_id = (unsigned long)ctxt;
+               read_wr.ex.invalidate_rkey = ctxt->frmr->mr->lkey;
+       } else {
+               read_wr.opcode = IB_WR_RDMA_READ;
+               read_wr.next = &inv_wr;
+               /* Prepare invalidate */
+               memset(&inv_wr, 0, sizeof(inv_wr));
+               inv_wr.wr_id = (unsigned long)ctxt;
+               inv_wr.opcode = IB_WR_LOCAL_INV;
+               inv_wr.send_flags = IB_SEND_SIGNALED;
+               inv_wr.ex.invalidate_rkey = frmr->mr->lkey;
+       }
+       ctxt->wr_op = read_wr.opcode;
+
+       /* Post the chain */
+       ret = svc_rdma_send(xprt, &fastreg_wr);
+       if (ret) {
+               pr_err("svcrdma: Error %d posting RDMA_READ\n", ret);
+               set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
+               goto err;
        }
-       return 0;
-}
 
-static int rdma_read_max_sge(struct svcxprt_rdma *xprt, int sge_count)
-{
-       if ((rdma_node_get_transport(xprt->sc_cm_id->device->node_type) ==
-            RDMA_TRANSPORT_IWARP) &&
-           sge_count > 1)
-               return 1;
-       else
-               return min_t(int, sge_count, xprt->sc_max_sge);
+       /* return current location in page array */
+       *page_no = pg_no;
+       *page_offset = pg_off;
+       ret = read;
+       atomic_inc(&rdma_stat_read);
+       return ret;
+ err:
+       svc_rdma_unmap_dma(ctxt);
+       svc_rdma_put_context(ctxt, 0);
+       svc_rdma_put_frmr(xprt, frmr);
+       return ret;
 }
 
-/*
- * Use RDMA_READ to read data from the advertised client buffer into the
- * XDR stream starting at rq_arg.head[0].iov_base.
- * Each chunk in the array
- * contains the following fields:
- * discrim      - '1', This isn't used for data placement
- * position     - The xdr stream offset (the same for every chunk)
- * handle       - RMR for client memory region
- * length       - data transfer length
- * offset       - 64 bit tagged offset in remote memory region
- *
- * On our side, we need to read into a pagelist. The first page immediately
- * follows the RPC header.
- *
- * This function returns:
- * 0 - No error and no read-list found.
- *
- * 1 - Successful read-list processing. The data is not yet in
- * the pagelist and therefore the RPC request must be deferred. The
- * I/O completion will enqueue the transport again and
- * svc_rdma_recvfrom will complete the request.
- *
- * <0 - Error processing/posting read-list.
- *
- * NOTE: The ctxt must not be touched after the last WR has been posted
- * because the I/O completion processing may occur on another
- * processor and free / modify the context. Ne touche pas!
- */
-static int rdma_read_xdr(struct svcxprt_rdma *xprt,
-                        struct rpcrdma_msg *rmsgp,
-                        struct svc_rqst *rqstp,
-                        struct svc_rdma_op_ctxt *hdr_ctxt)
+static int rdma_read_chunks(struct svcxprt_rdma *xprt,
+                           struct rpcrdma_msg *rmsgp,
+                           struct svc_rqst *rqstp,
+                           struct svc_rdma_op_ctxt *head)
 {
-       struct ib_send_wr read_wr;
-       struct ib_send_wr inv_wr;
-       int err = 0;
-       int ch_no;
-       int ch_count;
-       int byte_count;
-       int sge_count;
-       u64 sgl_offset;
+       int page_no, ch_count, ret;
        struct rpcrdma_read_chunk *ch;
-       struct svc_rdma_op_ctxt *ctxt = NULL;
-       struct svc_rdma_req_map *rpl_map;
-       struct svc_rdma_req_map *chl_map;
+       u32 page_offset, byte_count;
+       u64 rs_offset;
+       rdma_reader_fn reader;
 
        /* If no read list is present, return 0 */
        ch = svc_rdma_get_read_chunk(rmsgp);
@@ -408,122 +384,55 @@ static int rdma_read_xdr(struct svcxprt_rdma *xprt,
        if (ch_count > RPCSVC_MAXPAGES)
                return -EINVAL;
 
-       /* Allocate temporary reply and chunk maps */
-       rpl_map = svc_rdma_get_req_map();
-       chl_map = svc_rdma_get_req_map();
+       /* The request is completed when the RDMA_READs complete. The
+        * head context keeps all the pages that comprise the
+        * request.
+        */
+       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;
+       head->arg.page_base = 0;
+       head->arg.page_len = 0;
+       head->arg.len = rqstp->rq_arg.len;
+       head->arg.buflen = rqstp->rq_arg.buflen;
 
-       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);
+       /* Use FRMR if supported */
+       if (xprt->sc_dev_caps & SVCRDMA_DEVCAP_FAST_REG)
+               reader = rdma_read_chunk_frmr;
        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;
+               reader = rdma_read_chunk_lcl;
 
+       page_no = 0; page_offset = 0;
        for (ch = (struct rpcrdma_read_chunk *)&rmsgp->rm_body.rm_chunks[0];
-            ch->rc_discrim != 0; ch++, ch_no++) {
-               u64 rs_offset;
-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);
+            ch->rc_discrim != 0; ch++) {
 
-               /* Prepare READ WR */
-               memset(&read_wr, 0, sizeof read_wr);
-               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 = ntohl(ch->rc_target.rs_handle);
                xdr_decode_hyper((__be32 *)&ch->rc_target.rs_offset,
                                 &rs_offset);
-               read_wr.wr.rdma.remote_addr = rs_offset + sgl_offset;
-               read_wr.sg_list = ctxt->sge;
-               read_wr.num_sge =
-                       rdma_read_max_sge(xprt, chl_map->ch[ch_no].count);
-               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)) {
-                       /*
-                        * Mark the last RDMA_READ with a bit to
-                        * indicate all RPC data has been fetched from
-                        * 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 */
-               err = svc_rdma_send(xprt, &read_wr);
-               if (err) {
-                       printk(KERN_ERR "svcrdma: Error %d posting RDMA_READ\n",
-                              err);
-                       set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
-                       svc_rdma_unmap_dma(ctxt);
-                       svc_rdma_put_context(ctxt, 0);
-                       goto out;
+               byte_count = ntohl(ch->rc_target.rs_length);
+
+               while (byte_count > 0) {
+                       ret = reader(xprt, rqstp, head,
+                                    &page_no, &page_offset,
+                                    ntohl(ch->rc_target.rs_handle),
+                                    byte_count, rs_offset,
+                                    ((ch+1)->rc_discrim == 0) /* last */
+                                    );
+                       if (ret < 0)
+                               goto err;
+                       byte_count -= ret;
+                       rs_offset += ret;
+                       head->arg.buflen += ret;
                }
-               atomic_inc(&rdma_stat_read);
-
-               if (read_wr.num_sge < chl_map->ch[ch_no].count) {
-                       chl_map->ch[ch_no].count -= read_wr.num_sge;
-                       chl_map->ch[ch_no].start += read_wr.num_sge;
-                       goto next_sge;
-               }
-               sgl_offset = 0;
-               err = 1;
        }
-
- out:
-       svc_rdma_put_req_map(rpl_map);
-       svc_rdma_put_req_map(chl_map);
-
+       ret = 1;
+ err:
        /* Detach arg pages. svc_recv will replenish them */
-       for (ch_no = 0; &rqstp->rq_pages[ch_no] < rqstp->rq_respages; ch_no++)
-               rqstp->rq_pages[ch_no] = NULL;
+       for (page_no = 0;
+            &rqstp->rq_pages[page_no] < rqstp->rq_respages; page_no++)
+               rqstp->rq_pages[page_no] = NULL;
 
-       return err;
+       return ret;
 }
 
 static int rdma_read_complete(struct svc_rqst *rqstp,
@@ -595,13 +504,9 @@ int svc_rdma_recvfrom(struct svc_rqst *rqstp)
                                  struct svc_rdma_op_ctxt,
                                  dto_q);
                list_del_init(&ctxt->dto_q);
-       }
-       if (ctxt) {
                spin_unlock_bh(&rdma_xprt->sc_rq_dto_lock);
                return rdma_read_complete(rqstp, ctxt);
-       }
-
-       if (!list_empty(&rdma_xprt->sc_rq_dto_q)) {
+       } else if (!list_empty(&rdma_xprt->sc_rq_dto_q)) {
                ctxt = list_entry(rdma_xprt->sc_rq_dto_q.next,
                                  struct svc_rdma_op_ctxt,
                                  dto_q);
@@ -621,7 +526,6 @@ int svc_rdma_recvfrom(struct svc_rqst *rqstp)
                if (test_bit(XPT_CLOSE, &xprt->xpt_flags))
                        goto close_out;
 
-               BUG_ON(ret);
                goto out;
        }
        dprintk("svcrdma: processing ctxt=%p on xprt=%p, rqstp=%p, status=%d\n",
@@ -644,12 +548,11 @@ int svc_rdma_recvfrom(struct svc_rqst *rqstp)
        }
 
        /* Read read-list data. */
-       ret = rdma_read_xdr(rdma_xprt, rmsgp, rqstp, ctxt);
+       ret = rdma_read_chunks(rdma_xprt, rmsgp, rqstp, ctxt);
        if (ret > 0) {
                /* read-list posted, defer until data received from client. */
                goto defer;
-       }
-       if (ret < 0) {
+       } else if (ret < 0) {
                /* Post of read-list failed, free context. */
                svc_rdma_put_context(ctxt, 1);
                return 0;
index 7e024a51617e034811a950bb245d1f615e9809a9..49fd21a5c2157375937f9a60cc9c69fa6f2c765e 100644 (file)
@@ -1,4 +1,5 @@
 /*
+ * Copyright (c) 2014 Open Grid Computing, Inc. All rights reserved.
  * Copyright (c) 2005-2006 Network Appliance, Inc. All rights reserved.
  *
  * This software is available to you under a choice of one of two
 
 #define RPCDBG_FACILITY        RPCDBG_SVCXPRT
 
-/* Encode an XDR as an array of IB SGE
- *
- * Assumptions:
- * - head[0] is physically contiguous.
- * - tail[0] is physically contiguous.
- * - pages[] is not physically or virtually contiguous and consists of
- *   PAGE_SIZE elements.
- *
- * Output:
- * SGE[0]              reserved for RCPRDMA header
- * SGE[1]              data from xdr->head[]
- * SGE[2..sge_count-2] data from xdr->pages[]
- * SGE[sge_count-1]    data from xdr->tail.
- *
- * The max SGE we need is the length of the XDR / pagesize + one for
- * head + one for tail + one for RPCRDMA header. Since RPCSVC_MAXPAGES
- * reserves a page for both the request and the reply header, and this
- * array is only concerned with the reply we are assured that we have
- * on extra page for the RPCRMDA header.
- */
-static 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++;
-
-       /* Map the XDR head */
-       frmr->kva = frva;
-       frmr->direction = DMA_TO_DEVICE;
-       frmr->access_flags = 0;
-       frmr->map_len = PAGE_SIZE;
-       frmr->page_list_len = 1;
-       page_off = (unsigned long)xdr->head[0].iov_base & ~PAGE_MASK;
-       frmr->page_list->page_list[page_no] =
-               ib_dma_map_page(xprt->sc_cm_id->device,
-                               virt_to_page(xdr->head[0].iov_base),
-                               page_off,
-                               PAGE_SIZE - page_off,
-                               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);
-
-       /* Map the XDR page list */
-       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, page_off,
-                                       sge_bytes, 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_page(xprt->sc_cm_id->device, virt_to_page(va),
-                                   page_off,
-                                   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);
-       vec->frmr = NULL;
-       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)
@@ -208,9 +63,6 @@ static int map_xdr(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;
 
@@ -282,8 +134,6 @@ static dma_addr_t dma_map_xdr(struct svcxprt_rdma *xprt,
 }
 
 /* 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,
@@ -327,23 +177,16 @@ static int send_write(struct svcxprt_rdma *xprt, struct svc_rqst *rqstp,
                sge_bytes = min_t(size_t,
                          bc, vec->sge[xdr_sge_no].iov_len-sge_off);
                sge[sge_no].length = sge_bytes;
-               if (!vec->frmr) {
-                       sge[sge_no].addr =
-                               dma_map_xdr(xprt, &rqstp->rq_res, xdr_off,
-                                           sge_bytes, DMA_TO_DEVICE);
-                       xdr_off += sge_bytes;
-                       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;
-               }
+               sge[sge_no].addr =
+                       dma_map_xdr(xprt, &rqstp->rq_res, xdr_off,
+                                   sge_bytes, DMA_TO_DEVICE);
+               xdr_off += sge_bytes;
+               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;
                ctxt->count++;
-               ctxt->frmr = vec->frmr;
                sge_off = 0;
                sge_no++;
                xdr_sge_no++;
@@ -369,7 +212,6 @@ static int send_write(struct svcxprt_rdma *xprt, struct svc_rqst *rqstp,
        return 0;
  err:
        svc_rdma_unmap_dma(ctxt);
-       svc_rdma_put_frmr(xprt, vec->frmr);
        svc_rdma_put_context(ctxt, 0);
        /* Fatal error, close transport */
        return -EIO;
@@ -397,10 +239,7 @@ static int send_write_chunks(struct svcxprt_rdma *xprt,
        res_ary = (struct rpcrdma_write_array *)
                &rdma_resp->rm_body.rm_chunks[1];
 
-       if (vec->frmr)
-               max_write = vec->frmr->map_len;
-       else
-               max_write = xprt->sc_max_sge * PAGE_SIZE;
+       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;
@@ -472,10 +311,7 @@ static int send_reply_chunks(struct svcxprt_rdma *xprt,
        res_ary = (struct rpcrdma_write_array *)
                &rdma_resp->rm_body.rm_chunks[2];
 
-       if (vec->frmr)
-               max_write = vec->frmr->map_len;
-       else
-               max_write = xprt->sc_max_sge * PAGE_SIZE;
+       max_write = xprt->sc_max_sge * PAGE_SIZE;
 
        /* xdr offset starts at RPC message */
        nchunks = ntohl(arg_ary->wc_nchunks);
@@ -545,7 +381,6 @@ 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;
@@ -559,7 +394,6 @@ static int send_reply(struct svcxprt_rdma *rdma,
                       "svcrdma: could not post a receive buffer, err=%d."
                       "Closing transport %p.\n", ret, rdma);
                set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags);
-               svc_rdma_put_frmr(rdma, vec->frmr);
                svc_rdma_put_context(ctxt, 0);
                return -ENOTCONN;
        }
@@ -567,11 +401,6 @@ 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 */
        ctxt->sge[0].lkey = rdma->sc_dma_lkey;
@@ -590,21 +419,15 @@ static int send_reply(struct svcxprt_rdma *rdma,
                int xdr_off = 0;
                sge_bytes = min_t(size_t, vec->sge[sge_no].iov_len, byte_count);
                byte_count -= sge_bytes;
-               if (!vec->frmr) {
-                       ctxt->sge[sge_no].addr =
-                               dma_map_xdr(rdma, &rqstp->rq_res, xdr_off,
-                                           sge_bytes, DMA_TO_DEVICE);
-                       xdr_off += sge_bytes;
-                       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].addr =
+                       dma_map_xdr(rdma, &rqstp->rq_res, xdr_off,
+                                   sge_bytes, DMA_TO_DEVICE);
+               xdr_off += sge_bytes;
+               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;
                ctxt->sge[sge_no].length = sge_bytes;
        }
        BUG_ON(byte_count != 0);
@@ -627,6 +450,7 @@ static int send_reply(struct svcxprt_rdma *rdma,
                        ctxt->sge[page_no+1].length = 0;
        }
        rqstp->rq_next_page = rqstp->rq_respages + 1;
+
        BUG_ON(sge_no > rdma->sc_max_sge);
        memset(&send_wr, 0, sizeof send_wr);
        ctxt->wr_op = IB_WR_SEND;
@@ -635,15 +459,6 @@ 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)
@@ -653,7 +468,6 @@ static int send_reply(struct svcxprt_rdma *rdma,
 
  err:
        svc_rdma_unmap_dma(ctxt);
-       svc_rdma_put_frmr(rdma, vec->frmr);
        svc_rdma_put_context(ctxt, 1);
        return -EIO;
 }
index 02db8d9cc994163b876f84b4cc5cdf0e39b5ebc0..e7323fbbd348c5e30a8a99763df73c37fa6b5a26 100644 (file)
@@ -1,4 +1,5 @@
 /*
+ * Copyright (c) 2014 Open Grid Computing, Inc. All rights reserved.
  * Copyright (c) 2005-2007 Network Appliance, Inc. All rights reserved.
  *
  * This software is available to you under a choice of one of two
@@ -162,7 +163,6 @@ 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;
 }
 
@@ -338,22 +338,21 @@ static void process_context(struct svcxprt_rdma *xprt,
 
        switch (ctxt->wr_op) {
        case IB_WR_SEND:
-               if (test_bit(RDMACTXT_F_FAST_UNREG, &ctxt->flags))
-                       svc_rdma_put_frmr(xprt, ctxt->frmr);
+               BUG_ON(ctxt->frmr);
                svc_rdma_put_context(ctxt, 1);
                break;
 
        case IB_WR_RDMA_WRITE:
+               BUG_ON(ctxt->frmr);
                svc_rdma_put_context(ctxt, 0);
                break;
 
        case IB_WR_RDMA_READ:
        case IB_WR_RDMA_READ_WITH_INV:
+               svc_rdma_put_frmr(xprt, ctxt->frmr);
                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,
@@ -365,6 +364,7 @@ static void process_context(struct svcxprt_rdma *xprt,
                break;
 
        default:
+               BUG_ON(1);
                printk(KERN_ERR "svcrdma: unexpected completion type, "
                       "opcode=%d\n",
                       ctxt->wr_op);
@@ -380,29 +380,42 @@ static void process_context(struct svcxprt_rdma *xprt,
 static void sq_cq_reap(struct svcxprt_rdma *xprt)
 {
        struct svc_rdma_op_ctxt *ctxt = NULL;
-       struct ib_wc wc;
+       struct ib_wc wc_a[6];
+       struct ib_wc *wc;
        struct ib_cq *cq = xprt->sc_sq_cq;
        int ret;
 
+       memset(wc_a, 0, sizeof(wc_a));
+
        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) {
-               if (wc.status != IB_WC_SUCCESS)
-                       /* Close the transport */
-                       set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
+       while ((ret = ib_poll_cq(cq, ARRAY_SIZE(wc_a), wc_a)) > 0) {
+               int i;
 
-               /* Decrement used SQ WR count */
-               atomic_dec(&xprt->sc_sq_count);
-               wake_up(&xprt->sc_send_wait);
+               for (i = 0; i < ret; i++) {
+                       wc = &wc_a[i];
+                       if (wc->status != IB_WC_SUCCESS) {
+                               dprintk("svcrdma: sq wc err status %d\n",
+                                       wc->status);
 
-               ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id;
-               if (ctxt)
-                       process_context(xprt, ctxt);
+                               /* Close the transport */
+                               set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
+                       }
 
-               svc_xprt_put(&xprt->sc_xprt);
+                       /* Decrement used SQ WR count */
+                       atomic_dec(&xprt->sc_sq_count);
+                       wake_up(&xprt->sc_send_wait);
+
+                       ctxt = (struct svc_rdma_op_ctxt *)
+                               (unsigned long)wc->wr_id;
+                       if (ctxt)
+                               process_context(xprt, ctxt);
+
+                       svc_xprt_put(&xprt->sc_xprt);
+               }
        }
 
        if (ctxt)
@@ -995,7 +1008,11 @@ static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
                        need_dma_mr = 0;
                break;
        case RDMA_TRANSPORT_IB:
-               if (!(devattr.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)) {
+               if (!(newxprt->sc_dev_caps & SVCRDMA_DEVCAP_FAST_REG)) {
+                       need_dma_mr = 1;
+                       dma_mr_acc = IB_ACCESS_LOCAL_WRITE;
+               } else if (!(devattr.device_cap_flags &
+                            IB_DEVICE_LOCAL_DMA_LKEY)) {
                        need_dma_mr = 1;
                        dma_mr_acc = IB_ACCESS_LOCAL_WRITE;
                } else
@@ -1192,14 +1209,7 @@ static int svc_rdma_has_wspace(struct svc_xprt *xprt)
                container_of(xprt, struct svcxprt_rdma, sc_xprt);
 
        /*
-        * If there are fewer SQ WR available than required to send a
-        * simple response, return false.
-        */
-       if ((rdma->sc_sq_depth - atomic_read(&rdma->sc_sq_count) < 3))
-               return 0;
-
-       /*
-        * ...or there are already waiters on the SQ,
+        * If there are already waiters on the SQ,
         * return false.
         */
        if (waitqueue_active(&rdma->sc_send_wait))