xprtrdma: Place registered MWs on a per-req list
authorChuck Lever <chuck.lever@oracle.com>
Wed, 29 Jun 2016 17:54:16 +0000 (13:54 -0400)
committerAnna Schumaker <Anna.Schumaker@Netapp.com>
Mon, 11 Jul 2016 19:50:43 +0000 (15:50 -0400)
Instead of placing registered MWs sparsely into the rl_segments
array, place these MWs on a per-req list.

ro_unmap_{sync,safe} can then simply pull those MWs off the list
instead of walking through the array.

This change significantly reduces the size of struct rpcrdma_req
by removing nsegs and rl_mw from every array element.

As an additional clean-up, chunk co-ordinates are returned in the
"*mw" output argument so they are no longer needed in every
array element.

Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Tested-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
net/sunrpc/xprtrdma/fmr_ops.c
net/sunrpc/xprtrdma/frwr_ops.c
net/sunrpc/xprtrdma/rpc_rdma.c
net/sunrpc/xprtrdma/transport.c
net/sunrpc/xprtrdma/verbs.c
net/sunrpc/xprtrdma/xprt_rdma.h

index 6521dceff63863c5750d066a7a9d182b6e24ae62..21cb3b150b371dc5e8aab5e02bc4af99a6a3d856 100644 (file)
@@ -101,6 +101,10 @@ fmr_op_release_mr(struct rpcrdma_mw *r)
        LIST_HEAD(unmap_list);
        int rc;
 
+       /* Ensure MW is not on any rl_registered list */
+       if (!list_empty(&r->mw_list))
+               list_del(&r->mw_list);
+
        kfree(r->fmr.fm_physaddrs);
        kfree(r->mw_sg);
 
@@ -176,17 +180,13 @@ fmr_op_maxpages(struct rpcrdma_xprt *r_xprt)
  */
 static int
 fmr_op_map(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mr_seg *seg,
-          int nsegs, bool writing)
+          int nsegs, bool writing, struct rpcrdma_mw **out)
 {
        struct rpcrdma_mr_seg *seg1 = seg;
        int len, pageoff, i, rc;
        struct rpcrdma_mw *mw;
        u64 *dma_pages;
 
-       mw = seg1->rl_mw;
-       seg1->rl_mw = NULL;
-       if (mw)
-               rpcrdma_defer_mr_recovery(mw);
        mw = rpcrdma_get_mw(r_xprt);
        if (!mw)
                return -ENOBUFS;
@@ -230,11 +230,11 @@ fmr_op_map(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mr_seg *seg,
        if (rc)
                goto out_maperr;
 
-       seg1->rl_mw = mw;
-       seg1->mr_rkey = mw->fmr.fm_mr->rkey;
-       seg1->mr_base = dma_pages[0] + pageoff;
-       seg1->mr_nsegs = mw->mw_nents;
-       seg1->mr_len = len;
+       mw->mw_handle = mw->fmr.fm_mr->rkey;
+       mw->mw_length = len;
+       mw->mw_offset = dma_pages[0] + pageoff;
+
+       *out = mw;
        return mw->mw_nents;
 
 out_dmamap_err:
@@ -255,13 +255,13 @@ out_maperr:
  *
  * Sleeps until it is safe for the host CPU to access the
  * previously mapped memory regions.
+ *
+ * Caller ensures that req->rl_registered is not empty.
  */
 static void
 fmr_op_unmap_sync(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
 {
-       struct rpcrdma_mr_seg *seg;
-       unsigned int i, nchunks;
-       struct rpcrdma_mw *mw;
+       struct rpcrdma_mw *mw, *tmp;
        LIST_HEAD(unmap_list);
        int rc;
 
@@ -272,14 +272,8 @@ fmr_op_unmap_sync(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
         * ib_unmap_fmr() is slow, so use a single call instead
         * of one call per mapped FMR.
         */
-       for (i = 0, nchunks = req->rl_nchunks; nchunks; nchunks--) {
-               seg = &req->rl_segments[i];
-               mw = seg->rl_mw;
-
+       list_for_each_entry(mw, &req->rl_registered, mw_list)
                list_add_tail(&mw->fmr.fm_mr->list, &unmap_list);
-
-               i += seg->mr_nsegs;
-       }
        rc = ib_unmap_fmr(&unmap_list);
        if (rc)
                goto out_reset;
@@ -287,34 +281,22 @@ fmr_op_unmap_sync(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
        /* ORDER: Now DMA unmap all of the req's MRs, and return
         * them to the free MW list.
         */
-       for (i = 0, nchunks = req->rl_nchunks; nchunks; nchunks--) {
-               seg = &req->rl_segments[i];
-               mw = seg->rl_mw;
-
+       list_for_each_entry_safe(mw, tmp, &req->rl_registered, mw_list) {
+               list_del_init(&mw->mw_list);
                list_del_init(&mw->fmr.fm_mr->list);
                ib_dma_unmap_sg(r_xprt->rx_ia.ri_device,
                                mw->mw_sg, mw->mw_nents, mw->mw_dir);
                rpcrdma_put_mw(r_xprt, mw);
-
-               i += seg->mr_nsegs;
-               seg->mr_nsegs = 0;
-               seg->rl_mw = NULL;
        }
 
-       req->rl_nchunks = 0;
        return;
 
 out_reset:
        pr_err("rpcrdma: ib_unmap_fmr failed (%i)\n", rc);
 
-       for (i = 0, nchunks = req->rl_nchunks; nchunks; nchunks--) {
-               seg = &req->rl_segments[i];
-               mw = seg->rl_mw;
-
+       list_for_each_entry_safe(mw, tmp, &req->rl_registered, mw_list) {
                list_del_init(&mw->fmr.fm_mr->list);
                fmr_op_recover_mr(mw);
-
-               i += seg->mr_nsegs;
        }
 }
 
@@ -325,22 +307,17 @@ static void
 fmr_op_unmap_safe(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
                  bool sync)
 {
-       struct rpcrdma_mr_seg *seg;
        struct rpcrdma_mw *mw;
-       unsigned int i;
 
-       for (i = 0; req->rl_nchunks; req->rl_nchunks--) {
-               seg = &req->rl_segments[i];
-               mw = seg->rl_mw;
+       while (!list_empty(&req->rl_registered)) {
+               mw = list_first_entry(&req->rl_registered,
+                                     struct rpcrdma_mw, mw_list);
+               list_del_init(&mw->mw_list);
 
                if (sync)
                        fmr_op_recover_mr(mw);
                else
                        rpcrdma_defer_mr_recovery(mw);
-
-               i += seg->mr_nsegs;
-               seg->mr_nsegs = 0;
-               seg->rl_mw = NULL;
        }
 }
 
index f4c06c8ba6220d49f415274012f776602a34a7ee..892b5e1d9b099b217fba65a26b47105a0c027c51 100644 (file)
@@ -128,6 +128,10 @@ frwr_op_release_mr(struct rpcrdma_mw *r)
 {
        int rc;
 
+       /* Ensure MW is not on any rl_registered list */
+       if (!list_empty(&r->mw_list))
+               list_del(&r->mw_list);
+
        rc = ib_dereg_mr(r->frmr.fr_mr);
        if (rc)
                pr_err("rpcrdma: final ib_dereg_mr for %p returned %i\n",
@@ -333,10 +337,9 @@ frwr_wc_localinv_wake(struct ib_cq *cq, struct ib_wc *wc)
  */
 static int
 frwr_op_map(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mr_seg *seg,
-           int nsegs, bool writing)
+           int nsegs, bool writing, struct rpcrdma_mw **out)
 {
        struct rpcrdma_ia *ia = &r_xprt->rx_ia;
-       struct rpcrdma_mr_seg *seg1 = seg;
        struct rpcrdma_mw *mw;
        struct rpcrdma_frmr *frmr;
        struct ib_mr *mr;
@@ -345,8 +348,7 @@ frwr_op_map(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mr_seg *seg,
        int rc, i, n, dma_nents;
        u8 key;
 
-       mw = seg1->rl_mw;
-       seg1->rl_mw = NULL;
+       mw = NULL;
        do {
                if (mw)
                        rpcrdma_defer_mr_recovery(mw);
@@ -416,12 +418,11 @@ frwr_op_map(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mr_seg *seg,
        if (rc)
                goto out_senderr;
 
-       seg1->rl_mw = mw;
-       seg1->mr_rkey = mr->rkey;
-       seg1->mr_base = mr->iova;
-       seg1->mr_nsegs = mw->mw_nents;
-       seg1->mr_len = mr->length;
+       mw->mw_handle = mr->rkey;
+       mw->mw_length = mr->length;
+       mw->mw_offset = mr->iova;
 
+       *out = mw;
        return mw->mw_nents;
 
 out_dmamap_err:
@@ -443,9 +444,8 @@ out_senderr:
 }
 
 static struct ib_send_wr *
-__frwr_prepare_linv_wr(struct rpcrdma_mr_seg *seg)
+__frwr_prepare_linv_wr(struct rpcrdma_mw *mw)
 {
-       struct rpcrdma_mw *mw = seg->rl_mw;
        struct rpcrdma_frmr *f = &mw->frmr;
        struct ib_send_wr *invalidate_wr;
 
@@ -465,16 +465,16 @@ __frwr_prepare_linv_wr(struct rpcrdma_mr_seg *seg)
  *
  * Sleeps until it is safe for the host CPU to access the
  * previously mapped memory regions.
+ *
+ * Caller ensures that req->rl_registered is not empty.
  */
 static void
 frwr_op_unmap_sync(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
 {
        struct ib_send_wr *invalidate_wrs, *pos, *prev, *bad_wr;
        struct rpcrdma_ia *ia = &r_xprt->rx_ia;
-       struct rpcrdma_mr_seg *seg;
-       unsigned int i, nchunks;
+       struct rpcrdma_mw *mw, *tmp;
        struct rpcrdma_frmr *f;
-       struct rpcrdma_mw *mw;
        int rc;
 
        dprintk("RPC:       %s: req %p\n", __func__, req);
@@ -484,22 +484,18 @@ frwr_op_unmap_sync(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
         * Chain the LOCAL_INV Work Requests and post them with
         * a single ib_post_send() call.
         */
+       f = NULL;
        invalidate_wrs = pos = prev = NULL;
-       seg = NULL;
-       for (i = 0, nchunks = req->rl_nchunks; nchunks; nchunks--) {
-               seg = &req->rl_segments[i];
-
-               pos = __frwr_prepare_linv_wr(seg);
+       list_for_each_entry(mw, &req->rl_registered, mw_list) {
+               pos = __frwr_prepare_linv_wr(mw);
 
                if (!invalidate_wrs)
                        invalidate_wrs = pos;
                else
                        prev->next = pos;
                prev = pos;
-
-               i += seg->mr_nsegs;
+               f = &mw->frmr;
        }
-       f = &seg->rl_mw->frmr;
 
        /* Strong send queue ordering guarantees that when the
         * last WR in the chain completes, all WRs in the chain
@@ -524,20 +520,12 @@ frwr_op_unmap_sync(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
         * them to the free MW list.
         */
 unmap:
-       for (i = 0, nchunks = req->rl_nchunks; nchunks; nchunks--) {
-               seg = &req->rl_segments[i];
-               mw = seg->rl_mw;
-               seg->rl_mw = NULL;
-
+       list_for_each_entry_safe(mw, tmp, &req->rl_registered, mw_list) {
+               list_del_init(&mw->mw_list);
                ib_dma_unmap_sg(ia->ri_device,
                                mw->mw_sg, mw->mw_nents, mw->mw_dir);
                rpcrdma_put_mw(r_xprt, mw);
-
-               i += seg->mr_nsegs;
-               seg->mr_nsegs = 0;
        }
-
-       req->rl_nchunks = 0;
        return;
 
 reset_mrs:
@@ -547,17 +535,12 @@ reset_mrs:
        /* Find and reset the MRs in the LOCAL_INV WRs that did not
         * get posted. This is synchronous, and slow.
         */
-       for (i = 0, nchunks = req->rl_nchunks; nchunks; nchunks--) {
-               seg = &req->rl_segments[i];
-               mw = seg->rl_mw;
+       list_for_each_entry(mw, &req->rl_registered, mw_list) {
                f = &mw->frmr;
-
                if (mw->frmr.fr_mr->rkey == bad_wr->ex.invalidate_rkey) {
                        __frwr_reset_mr(ia, mw);
                        bad_wr = bad_wr->next;
                }
-
-               i += seg->mr_nsegs;
        }
        goto unmap;
 }
@@ -569,22 +552,17 @@ static void
 frwr_op_unmap_safe(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
                   bool sync)
 {
-       struct rpcrdma_mr_seg *seg;
        struct rpcrdma_mw *mw;
-       unsigned int i;
 
-       for (i = 0; req->rl_nchunks; req->rl_nchunks--) {
-               seg = &req->rl_segments[i];
-               mw = seg->rl_mw;
+       while (!list_empty(&req->rl_registered)) {
+               mw = list_first_entry(&req->rl_registered,
+                                     struct rpcrdma_mw, mw_list);
+               list_del_init(&mw->mw_list);
 
                if (sync)
                        frwr_op_recover_mr(mw);
                else
                        rpcrdma_defer_mr_recovery(mw);
-
-               i += seg->mr_nsegs;
-               seg->mr_nsegs = 0;
-               seg->rl_mw = NULL;
        }
 }
 
index 8fde0ab3b6958d1663d71be0416748d78b28827d..6d34c1f7908aae73baea8cde88584b20d5764036 100644 (file)
@@ -286,11 +286,11 @@ rpcrdma_convert_iovs(struct xdr_buf *xdrbuf, unsigned int pos,
 }
 
 static inline __be32 *
-xdr_encode_rdma_segment(__be32 *iptr, struct rpcrdma_mr_seg *seg)
+xdr_encode_rdma_segment(__be32 *iptr, struct rpcrdma_mw *mw)
 {
-       *iptr++ = cpu_to_be32(seg->mr_rkey);
-       *iptr++ = cpu_to_be32(seg->mr_len);
-       return xdr_encode_hyper(iptr, seg->mr_base);
+       *iptr++ = cpu_to_be32(mw->mw_handle);
+       *iptr++ = cpu_to_be32(mw->mw_length);
+       return xdr_encode_hyper(iptr, mw->mw_offset);
 }
 
 /* XDR-encode the Read list. Supports encoding a list of read
@@ -311,6 +311,7 @@ rpcrdma_encode_read_list(struct rpcrdma_xprt *r_xprt,
                         __be32 *iptr, enum rpcrdma_chunktype rtype)
 {
        struct rpcrdma_mr_seg *seg = req->rl_nextseg;
+       struct rpcrdma_mw *mw;
        unsigned int pos;
        int n, nsegs;
 
@@ -328,9 +329,11 @@ rpcrdma_encode_read_list(struct rpcrdma_xprt *r_xprt,
                return ERR_PTR(nsegs);
 
        do {
-               n = r_xprt->rx_ia.ri_ops->ro_map(r_xprt, seg, nsegs, false);
+               n = r_xprt->rx_ia.ri_ops->ro_map(r_xprt, seg, nsegs,
+                                                false, &mw);
                if (n < 0)
                        return ERR_PTR(n);
+               list_add(&mw->mw_list, &req->rl_registered);
 
                *iptr++ = xdr_one;      /* item present */
 
@@ -338,13 +341,12 @@ rpcrdma_encode_read_list(struct rpcrdma_xprt *r_xprt,
                 * have the same "position".
                 */
                *iptr++ = cpu_to_be32(pos);
-               iptr = xdr_encode_rdma_segment(iptr, seg);
+               iptr = xdr_encode_rdma_segment(iptr, mw);
 
-               dprintk("RPC: %5u %s: read segment pos %u "
-                       "%d@0x%016llx:0x%08x (%s)\n",
+               dprintk("RPC: %5u %s: pos %u %u@0x%016llx:0x%08x (%s)\n",
                        rqst->rq_task->tk_pid, __func__, pos,
-                       seg->mr_len, (unsigned long long)seg->mr_base,
-                       seg->mr_rkey, n < nsegs ? "more" : "last");
+                       mw->mw_length, (unsigned long long)mw->mw_offset,
+                       mw->mw_handle, n < nsegs ? "more" : "last");
 
                r_xprt->rx_stats.read_chunk_count++;
                req->rl_nchunks++;
@@ -376,6 +378,7 @@ rpcrdma_encode_write_list(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
                          enum rpcrdma_chunktype wtype)
 {
        struct rpcrdma_mr_seg *seg = req->rl_nextseg;
+       struct rpcrdma_mw *mw;
        int n, nsegs, nchunks;
        __be32 *segcount;
 
@@ -396,17 +399,18 @@ rpcrdma_encode_write_list(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
 
        nchunks = 0;
        do {
-               n = r_xprt->rx_ia.ri_ops->ro_map(r_xprt, seg, nsegs, true);
+               n = r_xprt->rx_ia.ri_ops->ro_map(r_xprt, seg, nsegs,
+                                                true, &mw);
                if (n < 0)
                        return ERR_PTR(n);
+               list_add(&mw->mw_list, &req->rl_registered);
 
-               iptr = xdr_encode_rdma_segment(iptr, seg);
+               iptr = xdr_encode_rdma_segment(iptr, mw);
 
-               dprintk("RPC: %5u %s: write segment "
-                       "%d@0x016%llx:0x%08x (%s)\n",
+               dprintk("RPC: %5u %s: %u@0x016%llx:0x%08x (%s)\n",
                        rqst->rq_task->tk_pid, __func__,
-                       seg->mr_len, (unsigned long long)seg->mr_base,
-                       seg->mr_rkey, n < nsegs ? "more" : "last");
+                       mw->mw_length, (unsigned long long)mw->mw_offset,
+                       mw->mw_handle, n < nsegs ? "more" : "last");
 
                r_xprt->rx_stats.write_chunk_count++;
                r_xprt->rx_stats.total_rdma_request += seg->mr_len;
@@ -443,6 +447,7 @@ rpcrdma_encode_reply_chunk(struct rpcrdma_xprt *r_xprt,
                           __be32 *iptr, enum rpcrdma_chunktype wtype)
 {
        struct rpcrdma_mr_seg *seg = req->rl_nextseg;
+       struct rpcrdma_mw *mw;
        int n, nsegs, nchunks;
        __be32 *segcount;
 
@@ -461,17 +466,18 @@ rpcrdma_encode_reply_chunk(struct rpcrdma_xprt *r_xprt,
 
        nchunks = 0;
        do {
-               n = r_xprt->rx_ia.ri_ops->ro_map(r_xprt, seg, nsegs, true);
+               n = r_xprt->rx_ia.ri_ops->ro_map(r_xprt, seg, nsegs,
+                                                true, &mw);
                if (n < 0)
                        return ERR_PTR(n);
+               list_add(&mw->mw_list, &req->rl_registered);
 
-               iptr = xdr_encode_rdma_segment(iptr, seg);
+               iptr = xdr_encode_rdma_segment(iptr, mw);
 
-               dprintk("RPC: %5u %s: reply segment "
-                       "%d@0x%016llx:0x%08x (%s)\n",
+               dprintk("RPC: %5u %s: %u@0x%016llx:0x%08x (%s)\n",
                        rqst->rq_task->tk_pid, __func__,
-                       seg->mr_len, (unsigned long long)seg->mr_base,
-                       seg->mr_rkey, n < nsegs ? "more" : "last");
+                       mw->mw_length, (unsigned long long)mw->mw_offset,
+                       mw->mw_handle, n < nsegs ? "more" : "last");
 
                r_xprt->rx_stats.reply_chunk_count++;
                r_xprt->rx_stats.total_rdma_request += seg->mr_len;
@@ -690,10 +696,7 @@ rpcrdma_marshal_req(struct rpc_rqst *rqst)
 out_overflow:
        pr_err("rpcrdma: send overflow: hdrlen %zd rpclen %zu %s/%s\n",
                hdrlen, rpclen, transfertypes[rtype], transfertypes[wtype]);
-       /* Terminate this RPC. Chunks registered above will be
-        * released by xprt_release -> xprt_rmda_free .
-        */
-       return -EIO;
+       iptr = ERR_PTR(-EIO);
 
 out_unmap:
        r_xprt->rx_ia.ri_ops->ro_unmap_safe(r_xprt, req, false);
@@ -705,15 +708,13 @@ out_unmap:
  * RDMA'd by server. See map at rpcrdma_create_chunks()! :-)
  */
 static int
-rpcrdma_count_chunks(struct rpcrdma_rep *rep, unsigned int max, int wrchunk, __be32 **iptrp)
+rpcrdma_count_chunks(struct rpcrdma_rep *rep, int wrchunk, __be32 **iptrp)
 {
        unsigned int i, total_len;
        struct rpcrdma_write_chunk *cur_wchunk;
        char *base = (char *)rdmab_to_msg(rep->rr_rdmabuf);
 
        i = be32_to_cpu(**iptrp);
-       if (i > max)
-               return -1;
        cur_wchunk = (struct rpcrdma_write_chunk *) (*iptrp + 1);
        total_len = 0;
        while (i--) {
@@ -960,14 +961,13 @@ rpcrdma_reply_handler(struct rpcrdma_rep *rep)
                    (headerp->rm_body.rm_chunks[1] == xdr_zero &&
                     headerp->rm_body.rm_chunks[2] != xdr_zero) ||
                    (headerp->rm_body.rm_chunks[1] != xdr_zero &&
-                    req->rl_nchunks == 0))
+                    list_empty(&req->rl_registered)))
                        goto badheader;
                if (headerp->rm_body.rm_chunks[1] != xdr_zero) {
                        /* count any expected write chunks in read reply */
                        /* start at write chunk array count */
                        iptr = &headerp->rm_body.rm_chunks[2];
-                       rdmalen = rpcrdma_count_chunks(rep,
-                                               req->rl_nchunks, 1, &iptr);
+                       rdmalen = rpcrdma_count_chunks(rep, 1, &iptr);
                        /* check for validity, and no reply chunk after */
                        if (rdmalen < 0 || *iptr++ != xdr_zero)
                                goto badheader;
@@ -997,11 +997,11 @@ rpcrdma_reply_handler(struct rpcrdma_rep *rep)
                if (headerp->rm_body.rm_chunks[0] != xdr_zero ||
                    headerp->rm_body.rm_chunks[1] != xdr_zero ||
                    headerp->rm_body.rm_chunks[2] != xdr_one ||
-                   req->rl_nchunks == 0)
+                   list_empty(&req->rl_registered))
                        goto badheader;
                iptr = (__be32 *)((unsigned char *)headerp +
                                                        RPCRDMA_HDRLEN_MIN);
-               rdmalen = rpcrdma_count_chunks(rep, req->rl_nchunks, 0, &iptr);
+               rdmalen = rpcrdma_count_chunks(rep, 0, &iptr);
                if (rdmalen < 0)
                        goto badheader;
                r_xprt->rx_stats.total_rdma_reply += rdmalen;
@@ -1014,14 +1014,9 @@ rpcrdma_reply_handler(struct rpcrdma_rep *rep)
 
 badheader:
        default:
-               dprintk("%s: invalid rpcrdma reply header (type %d):"
-                               " chunks[012] == %d %d %d"
-                               " expected chunks <= %d\n",
-                               __func__, be32_to_cpu(headerp->rm_type),
-                               headerp->rm_body.rm_chunks[0],
-                               headerp->rm_body.rm_chunks[1],
-                               headerp->rm_body.rm_chunks[2],
-                               req->rl_nchunks);
+               dprintk("RPC: %5u %s: invalid rpcrdma reply (type %u)\n",
+                       rqst->rq_task->tk_pid, __func__,
+                       be32_to_cpu(headerp->rm_type));
                status = -EIO;
                r_xprt->rx_stats.bad_reply_count++;
                break;
@@ -1035,7 +1030,7 @@ out:
         * control: waking the next RPC waits until this RPC has
         * relinquished all its Send Queue entries.
         */
-       if (req->rl_nchunks)
+       if (!list_empty(&req->rl_registered))
                r_xprt->rx_ia.ri_ops->ro_unmap_sync(r_xprt, req);
 
        spin_lock_bh(&xprt->transport_lock);
index b1dd42a934845d7e34831387aaf79708a2d2fb36..81f0e879f019e43d35cc9c85ead0dfd17ebc8d30 100644 (file)
@@ -619,6 +619,9 @@ xprt_rdma_send_request(struct rpc_task *task)
        struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
        int rc = 0;
 
+       /* On retransmit, remove any previously registered chunks */
+       r_xprt->rx_ia.ri_ops->ro_unmap_safe(r_xprt, req, false);
+
        rc = rpcrdma_marshal_req(rqst);
        if (rc < 0)
                goto failed_marshal;
index e8677eafb32923650b61b17fefdc34b966b3f5f0..a74d79d473c8366c907b2e2bb7bcaba63e9572d9 100644 (file)
@@ -847,6 +847,7 @@ rpcrdma_create_req(struct rpcrdma_xprt *r_xprt)
        spin_unlock(&buffer->rb_reqslock);
        req->rl_cqe.done = rpcrdma_wc_send;
        req->rl_buffer = &r_xprt->rx_buf;
+       INIT_LIST_HEAD(&req->rl_registered);
        return req;
 }
 
index 649d01dda3272053b0767548942bc15995e5e922..f5d05110de9ffd5fbbc71cf92875eaa762116e02 100644 (file)
@@ -245,6 +245,9 @@ struct rpcrdma_mw {
                struct rpcrdma_frmr     frmr;
        };
        struct rpcrdma_xprt     *mw_xprt;
+       u32                     mw_handle;
+       u32                     mw_length;
+       u64                     mw_offset;
        struct list_head        mw_all;
 };
 
@@ -272,11 +275,7 @@ struct rpcrdma_mw {
  */
 
 struct rpcrdma_mr_seg {                /* chunk descriptors */
-       struct rpcrdma_mw *rl_mw;       /* registered MR */
-       u64             mr_base;        /* registration result */
-       u32             mr_rkey;        /* registration result */
        u32             mr_len;         /* length of chunk or segment */
-       int             mr_nsegs;       /* number of segments in chunk or 0 */
        struct page     *mr_page;       /* owning page, if any */
        char            *mr_offset;     /* kva if no page, else offset */
 };
@@ -294,6 +293,7 @@ struct rpcrdma_req {
        struct ib_sge           rl_send_iov[RPCRDMA_MAX_IOVS];
        struct rpcrdma_regbuf   *rl_rdmabuf;
        struct rpcrdma_regbuf   *rl_sendbuf;
+       struct list_head        rl_registered;  /* registered segments */
        struct rpcrdma_mr_seg   rl_segments[RPCRDMA_MAX_SEGS];
        struct rpcrdma_mr_seg   *rl_nextseg;
 
@@ -397,7 +397,8 @@ struct rpcrdma_stats {
 struct rpcrdma_xprt;
 struct rpcrdma_memreg_ops {
        int             (*ro_map)(struct rpcrdma_xprt *,
-                                 struct rpcrdma_mr_seg *, int, bool);
+                                 struct rpcrdma_mr_seg *, int, bool,
+                                 struct rpcrdma_mw **);
        void            (*ro_unmap_sync)(struct rpcrdma_xprt *,
                                         struct rpcrdma_req *);
        void            (*ro_unmap_safe)(struct rpcrdma_xprt *,