svcrdma: Add class for RDMA backwards direction transport
authorChuck Lever <chuck.lever@oracle.com>
Thu, 7 Jan 2016 19:50:10 +0000 (14:50 -0500)
committerDoug Ledford <dledford@redhat.com>
Tue, 19 Jan 2016 20:30:48 +0000 (15:30 -0500)
To support the server-side of an NFSv4.1 backchannel on RDMA
connections, add a transport class that enables backward
direction messages on an existing forward channel connection.

Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Acked-by: Bruce Fields <bfields@fieldses.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>
include/linux/sunrpc/svc_rdma.h
net/sunrpc/xprt.c
net/sunrpc/xprtrdma/Makefile
net/sunrpc/xprtrdma/svc_rdma_backchannel.c [new file with mode: 0644]
net/sunrpc/xprtrdma/svc_rdma_recvfrom.c
net/sunrpc/xprtrdma/svc_rdma_transport.c
net/sunrpc/xprtrdma/transport.c
net/sunrpc/xprtrdma/xprt_rdma.h

index 9a2c418dc69005e4fbf8430522cd17b368382936..b13513a0caf463c99e7f9b65c9f802505ec30153 100644 (file)
@@ -195,6 +195,11 @@ struct svcxprt_rdma {
 
 #define RPCSVC_MAXPAYLOAD_RDMA RPCSVC_MAXPAYLOAD
 
+/* svc_rdma_backchannel.c */
+extern int svc_rdma_handle_bc_reply(struct rpc_xprt *xprt,
+                                   struct rpcrdma_msg *rmsgp,
+                                   struct xdr_buf *rcvbuf);
+
 /* svc_rdma_marshal.c */
 extern int svc_rdma_xdr_decode_req(struct rpcrdma_msg **, struct svc_rqst *);
 extern int svc_rdma_xdr_encode_error(struct svcxprt_rdma *,
index 2e98f4a243e57d4539b3f4048d54c4b53f3c12d5..37edea6fa92d9685570ad3c1b37cb5fda5c0b9b3 100644 (file)
@@ -1425,3 +1425,4 @@ void xprt_put(struct rpc_xprt *xprt)
        if (atomic_dec_and_test(&xprt->count))
                xprt_destroy(xprt);
 }
+EXPORT_SYMBOL_GPL(xprt_put);
index 33f99d3004f2718206ad6b245a214e20718ac540..dc9f3b513a05f3a2fc9027359f9e1e587f9d243b 100644 (file)
@@ -2,7 +2,7 @@ obj-$(CONFIG_SUNRPC_XPRT_RDMA) += rpcrdma.o
 
 rpcrdma-y := transport.o rpc_rdma.o verbs.o \
        fmr_ops.o frwr_ops.o physical_ops.o \
-       svc_rdma.o svc_rdma_transport.o \
+       svc_rdma.o svc_rdma_backchannel.o svc_rdma_transport.o \
        svc_rdma_marshal.o svc_rdma_sendto.o svc_rdma_recvfrom.o \
        module.o
 rpcrdma-$(CONFIG_SUNRPC_BACKCHANNEL) += backchannel.o
diff --git a/net/sunrpc/xprtrdma/svc_rdma_backchannel.c b/net/sunrpc/xprtrdma/svc_rdma_backchannel.c
new file mode 100644 (file)
index 0000000..deff06a
--- /dev/null
@@ -0,0 +1,371 @@
+/*
+ * Copyright (c) 2015 Oracle.  All rights reserved.
+ *
+ * Support for backward direction RPCs on RPC/RDMA (server-side).
+ */
+
+#include <linux/sunrpc/svc_rdma.h>
+#include "xprt_rdma.h"
+
+#define RPCDBG_FACILITY        RPCDBG_SVCXPRT
+
+#undef SVCRDMA_BACKCHANNEL_DEBUG
+
+int svc_rdma_handle_bc_reply(struct rpc_xprt *xprt, struct rpcrdma_msg *rmsgp,
+                            struct xdr_buf *rcvbuf)
+{
+       struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
+       struct kvec *dst, *src = &rcvbuf->head[0];
+       struct rpc_rqst *req;
+       unsigned long cwnd;
+       u32 credits;
+       size_t len;
+       __be32 xid;
+       __be32 *p;
+       int ret;
+
+       p = (__be32 *)src->iov_base;
+       len = src->iov_len;
+       xid = rmsgp->rm_xid;
+
+#ifdef SVCRDMA_BACKCHANNEL_DEBUG
+       pr_info("%s: xid=%08x, length=%zu\n",
+               __func__, be32_to_cpu(xid), len);
+       pr_info("%s: RPC/RDMA: %*ph\n",
+               __func__, (int)RPCRDMA_HDRLEN_MIN, rmsgp);
+       pr_info("%s:      RPC: %*ph\n",
+               __func__, (int)len, p);
+#endif
+
+       ret = -EAGAIN;
+       if (src->iov_len < 24)
+               goto out_shortreply;
+
+       spin_lock_bh(&xprt->transport_lock);
+       req = xprt_lookup_rqst(xprt, xid);
+       if (!req)
+               goto out_notfound;
+
+       dst = &req->rq_private_buf.head[0];
+       memcpy(&req->rq_private_buf, &req->rq_rcv_buf, sizeof(struct xdr_buf));
+       if (dst->iov_len < len)
+               goto out_unlock;
+       memcpy(dst->iov_base, p, len);
+
+       credits = be32_to_cpu(rmsgp->rm_credit);
+       if (credits == 0)
+               credits = 1;    /* don't deadlock */
+       else if (credits > r_xprt->rx_buf.rb_bc_max_requests)
+               credits = r_xprt->rx_buf.rb_bc_max_requests;
+
+       cwnd = xprt->cwnd;
+       xprt->cwnd = credits << RPC_CWNDSHIFT;
+       if (xprt->cwnd > cwnd)
+               xprt_release_rqst_cong(req->rq_task);
+
+       ret = 0;
+       xprt_complete_rqst(req->rq_task, rcvbuf->len);
+       rcvbuf->len = 0;
+
+out_unlock:
+       spin_unlock_bh(&xprt->transport_lock);
+out:
+       return ret;
+
+out_shortreply:
+       dprintk("svcrdma: short bc reply: xprt=%p, len=%zu\n",
+               xprt, src->iov_len);
+       goto out;
+
+out_notfound:
+       dprintk("svcrdma: unrecognized bc reply: xprt=%p, xid=%08x\n",
+               xprt, be32_to_cpu(xid));
+
+       goto out_unlock;
+}
+
+/* Send a backwards direction RPC call.
+ *
+ * Caller holds the connection's mutex and has already marshaled
+ * the RPC/RDMA request.
+ *
+ * This is similar to svc_rdma_reply, but takes an rpc_rqst
+ * instead, does not support chunks, and avoids blocking memory
+ * allocation.
+ *
+ * XXX: There is still an opportunity to block in svc_rdma_send()
+ * if there are no SQ entries to post the Send. This may occur if
+ * the adapter has a small maximum SQ depth.
+ */
+static int svc_rdma_bc_sendto(struct svcxprt_rdma *rdma,
+                             struct rpc_rqst *rqst)
+{
+       struct xdr_buf *sndbuf = &rqst->rq_snd_buf;
+       struct svc_rdma_op_ctxt *ctxt;
+       struct svc_rdma_req_map *vec;
+       struct ib_send_wr send_wr;
+       int ret;
+
+       vec = svc_rdma_get_req_map(rdma);
+       ret = svc_rdma_map_xdr(rdma, sndbuf, vec);
+       if (ret)
+               goto out_err;
+
+       /* Post a recv buffer to handle the reply for this request. */
+       ret = svc_rdma_post_recv(rdma, GFP_NOIO);
+       if (ret) {
+               pr_err("svcrdma: Failed to post bc receive buffer, err=%d.\n",
+                      ret);
+               pr_err("svcrdma: closing transport %p.\n", rdma);
+               set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags);
+               ret = -ENOTCONN;
+               goto out_err;
+       }
+
+       ctxt = svc_rdma_get_context(rdma);
+       ctxt->pages[0] = virt_to_page(rqst->rq_buffer);
+       ctxt->count = 1;
+
+       ctxt->wr_op = IB_WR_SEND;
+       ctxt->direction = DMA_TO_DEVICE;
+       ctxt->sge[0].lkey = rdma->sc_dma_lkey;
+       ctxt->sge[0].length = sndbuf->len;
+       ctxt->sge[0].addr =
+           ib_dma_map_page(rdma->sc_cm_id->device, ctxt->pages[0], 0,
+                           sndbuf->len, DMA_TO_DEVICE);
+       if (ib_dma_mapping_error(rdma->sc_cm_id->device, ctxt->sge[0].addr)) {
+               ret = -EIO;
+               goto out_unmap;
+       }
+       atomic_inc(&rdma->sc_dma_used);
+
+       memset(&send_wr, 0, sizeof(send_wr));
+       send_wr.wr_id = (unsigned long)ctxt;
+       send_wr.sg_list = ctxt->sge;
+       send_wr.num_sge = 1;
+       send_wr.opcode = IB_WR_SEND;
+       send_wr.send_flags = IB_SEND_SIGNALED;
+
+       ret = svc_rdma_send(rdma, &send_wr);
+       if (ret) {
+               ret = -EIO;
+               goto out_unmap;
+       }
+
+out_err:
+       svc_rdma_put_req_map(rdma, vec);
+       dprintk("svcrdma: %s returns %d\n", __func__, ret);
+       return ret;
+
+out_unmap:
+       svc_rdma_unmap_dma(ctxt);
+       svc_rdma_put_context(ctxt, 1);
+       goto out_err;
+}
+
+/* Server-side transport endpoint wants a whole page for its send
+ * buffer. The client RPC code constructs the RPC header in this
+ * buffer before it invokes ->send_request.
+ *
+ * Returns NULL if there was a temporary allocation failure.
+ */
+static void *
+xprt_rdma_bc_allocate(struct rpc_task *task, size_t size)
+{
+       struct rpc_rqst *rqst = task->tk_rqstp;
+       struct svc_xprt *sxprt = rqst->rq_xprt->bc_xprt;
+       struct svcxprt_rdma *rdma;
+       struct page *page;
+
+       rdma = container_of(sxprt, struct svcxprt_rdma, sc_xprt);
+
+       /* Prevent an infinite loop: try to make this case work */
+       if (size > PAGE_SIZE)
+               WARN_ONCE(1, "svcrdma: large bc buffer request (size %zu)\n",
+                         size);
+
+       page = alloc_page(RPCRDMA_DEF_GFP);
+       if (!page)
+               return NULL;
+
+       return page_address(page);
+}
+
+static void
+xprt_rdma_bc_free(void *buffer)
+{
+       /* No-op: ctxt and page have already been freed. */
+}
+
+static int
+rpcrdma_bc_send_request(struct svcxprt_rdma *rdma, struct rpc_rqst *rqst)
+{
+       struct rpc_xprt *xprt = rqst->rq_xprt;
+       struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
+       struct rpcrdma_msg *headerp = (struct rpcrdma_msg *)rqst->rq_buffer;
+       int rc;
+
+       /* Space in the send buffer for an RPC/RDMA header is reserved
+        * via xprt->tsh_size.
+        */
+       headerp->rm_xid = rqst->rq_xid;
+       headerp->rm_vers = rpcrdma_version;
+       headerp->rm_credit = cpu_to_be32(r_xprt->rx_buf.rb_bc_max_requests);
+       headerp->rm_type = rdma_msg;
+       headerp->rm_body.rm_chunks[0] = xdr_zero;
+       headerp->rm_body.rm_chunks[1] = xdr_zero;
+       headerp->rm_body.rm_chunks[2] = xdr_zero;
+
+#ifdef SVCRDMA_BACKCHANNEL_DEBUG
+       pr_info("%s: %*ph\n", __func__, 64, rqst->rq_buffer);
+#endif
+
+       rc = svc_rdma_bc_sendto(rdma, rqst);
+       if (rc)
+               goto drop_connection;
+       return rc;
+
+drop_connection:
+       dprintk("svcrdma: failed to send bc call\n");
+       xprt_disconnect_done(xprt);
+       return -ENOTCONN;
+}
+
+/* Send an RPC call on the passive end of a transport
+ * connection.
+ */
+static int
+xprt_rdma_bc_send_request(struct rpc_task *task)
+{
+       struct rpc_rqst *rqst = task->tk_rqstp;
+       struct svc_xprt *sxprt = rqst->rq_xprt->bc_xprt;
+       struct svcxprt_rdma *rdma;
+       int ret;
+
+       dprintk("svcrdma: sending bc call with xid: %08x\n",
+               be32_to_cpu(rqst->rq_xid));
+
+       if (!mutex_trylock(&sxprt->xpt_mutex)) {
+               rpc_sleep_on(&sxprt->xpt_bc_pending, task, NULL);
+               if (!mutex_trylock(&sxprt->xpt_mutex))
+                       return -EAGAIN;
+               rpc_wake_up_queued_task(&sxprt->xpt_bc_pending, task);
+       }
+
+       ret = -ENOTCONN;
+       rdma = container_of(sxprt, struct svcxprt_rdma, sc_xprt);
+       if (!test_bit(XPT_DEAD, &sxprt->xpt_flags))
+               ret = rpcrdma_bc_send_request(rdma, rqst);
+
+       mutex_unlock(&sxprt->xpt_mutex);
+
+       if (ret < 0)
+               return ret;
+       return 0;
+}
+
+static void
+xprt_rdma_bc_close(struct rpc_xprt *xprt)
+{
+       dprintk("svcrdma: %s: xprt %p\n", __func__, xprt);
+}
+
+static void
+xprt_rdma_bc_put(struct rpc_xprt *xprt)
+{
+       dprintk("svcrdma: %s: xprt %p\n", __func__, xprt);
+
+       xprt_free(xprt);
+       module_put(THIS_MODULE);
+}
+
+static struct rpc_xprt_ops xprt_rdma_bc_procs = {
+       .reserve_xprt           = xprt_reserve_xprt_cong,
+       .release_xprt           = xprt_release_xprt_cong,
+       .alloc_slot             = xprt_alloc_slot,
+       .release_request        = xprt_release_rqst_cong,
+       .buf_alloc              = xprt_rdma_bc_allocate,
+       .buf_free               = xprt_rdma_bc_free,
+       .send_request           = xprt_rdma_bc_send_request,
+       .set_retrans_timeout    = xprt_set_retrans_timeout_def,
+       .close                  = xprt_rdma_bc_close,
+       .destroy                = xprt_rdma_bc_put,
+       .print_stats            = xprt_rdma_print_stats
+};
+
+static const struct rpc_timeout xprt_rdma_bc_timeout = {
+       .to_initval = 60 * HZ,
+       .to_maxval = 60 * HZ,
+};
+
+/* It shouldn't matter if the number of backchannel session slots
+ * doesn't match the number of RPC/RDMA credits. That just means
+ * one or the other will have extra slots that aren't used.
+ */
+static struct rpc_xprt *
+xprt_setup_rdma_bc(struct xprt_create *args)
+{
+       struct rpc_xprt *xprt;
+       struct rpcrdma_xprt *new_xprt;
+
+       if (args->addrlen > sizeof(xprt->addr)) {
+               dprintk("RPC:       %s: address too large\n", __func__);
+               return ERR_PTR(-EBADF);
+       }
+
+       xprt = xprt_alloc(args->net, sizeof(*new_xprt),
+                         RPCRDMA_MAX_BC_REQUESTS,
+                         RPCRDMA_MAX_BC_REQUESTS);
+       if (!xprt) {
+               dprintk("RPC:       %s: couldn't allocate rpc_xprt\n",
+                       __func__);
+               return ERR_PTR(-ENOMEM);
+       }
+
+       xprt->timeout = &xprt_rdma_bc_timeout;
+       xprt_set_bound(xprt);
+       xprt_set_connected(xprt);
+       xprt->bind_timeout = RPCRDMA_BIND_TO;
+       xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
+       xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
+
+       xprt->prot = XPRT_TRANSPORT_BC_RDMA;
+       xprt->tsh_size = RPCRDMA_HDRLEN_MIN / sizeof(__be32);
+       xprt->ops = &xprt_rdma_bc_procs;
+
+       memcpy(&xprt->addr, args->dstaddr, args->addrlen);
+       xprt->addrlen = args->addrlen;
+       xprt_rdma_format_addresses(xprt, (struct sockaddr *)&xprt->addr);
+       xprt->resvport = 0;
+
+       xprt->max_payload = xprt_rdma_max_inline_read;
+
+       new_xprt = rpcx_to_rdmax(xprt);
+       new_xprt->rx_buf.rb_bc_max_requests = xprt->max_reqs;
+
+       xprt_get(xprt);
+       args->bc_xprt->xpt_bc_xprt = xprt;
+       xprt->bc_xprt = args->bc_xprt;
+
+       if (!try_module_get(THIS_MODULE))
+               goto out_fail;
+
+       /* Final put for backchannel xprt is in __svc_rdma_free */
+       xprt_get(xprt);
+       return xprt;
+
+out_fail:
+       xprt_rdma_free_addresses(xprt);
+       args->bc_xprt->xpt_bc_xprt = NULL;
+       xprt_put(xprt);
+       xprt_free(xprt);
+       return ERR_PTR(-EINVAL);
+}
+
+struct xprt_class xprt_rdma_bc = {
+       .list                   = LIST_HEAD_INIT(xprt_rdma_bc.list),
+       .name                   = "rdma backchannel",
+       .owner                  = THIS_MODULE,
+       .ident                  = XPRT_TRANSPORT_BC_RDMA,
+       .setup                  = xprt_setup_rdma_bc,
+};
index ff4f01e527ecc08a1480ecba8f00d41a90a76571..3dfe4642ec9224a0bee9734c03b5ff4e819ff5de 100644 (file)
@@ -567,6 +567,38 @@ static int rdma_read_complete(struct svc_rqst *rqstp,
        return ret;
 }
 
+/* By convention, backchannel calls arrive via rdma_msg type
+ * messages, and never populate the chunk lists. This makes
+ * the RPC/RDMA header small and fixed in size, so it is
+ * straightforward to check the RPC header's direction field.
+ */
+static bool
+svc_rdma_is_backchannel_reply(struct svc_xprt *xprt, struct rpcrdma_msg *rmsgp)
+{
+       __be32 *p = (__be32 *)rmsgp;
+
+       if (!xprt->xpt_bc_xprt)
+               return false;
+
+       if (rmsgp->rm_type != rdma_msg)
+               return false;
+       if (rmsgp->rm_body.rm_chunks[0] != xdr_zero)
+               return false;
+       if (rmsgp->rm_body.rm_chunks[1] != xdr_zero)
+               return false;
+       if (rmsgp->rm_body.rm_chunks[2] != xdr_zero)
+               return false;
+
+       /* sanity */
+       if (p[7] != rmsgp->rm_xid)
+               return false;
+       /* call direction */
+       if (p[8] == cpu_to_be32(RPC_CALL))
+               return false;
+
+       return true;
+}
+
 /*
  * Set up the rqstp thread context to point to the RQ buffer. If
  * necessary, pull additional data from the client with an RDMA_READ
@@ -632,6 +664,15 @@ int svc_rdma_recvfrom(struct svc_rqst *rqstp)
                goto close_out;
        }
 
+       if (svc_rdma_is_backchannel_reply(xprt, rmsgp)) {
+               ret = svc_rdma_handle_bc_reply(xprt->xpt_bc_xprt, rmsgp,
+                                              &rqstp->rq_arg);
+               svc_rdma_put_context(ctxt, 0);
+               if (ret)
+                       goto repost;
+               return ret;
+       }
+
        /* Read read-list data. */
        ret = rdma_read_chunks(rdma_xprt, rmsgp, rqstp, ctxt);
        if (ret > 0) {
@@ -668,4 +709,15 @@ int svc_rdma_recvfrom(struct svc_rqst *rqstp)
        set_bit(XPT_CLOSE, &xprt->xpt_flags);
 defer:
        return 0;
+
+repost:
+       ret = svc_rdma_post_recv(rdma_xprt, GFP_KERNEL);
+       if (ret) {
+               pr_err("svcrdma: could not post a receive buffer, err=%d.\n",
+                      ret);
+               pr_err("svcrdma: closing transport %p.\n", rdma_xprt);
+               set_bit(XPT_CLOSE, &rdma_xprt->sc_xprt.xpt_flags);
+               ret = -ENOTCONN;
+       }
+       return ret;
 }
index af86dfeceb4a610af3e50720afd4a07f2406c42a..7fd23955f1d42aa5c1ef907f3d912ff47aecc1cf 100644 (file)
@@ -1287,12 +1287,14 @@ static void __svc_rdma_free(struct work_struct *work)
 {
        struct svcxprt_rdma *rdma =
                container_of(work, struct svcxprt_rdma, sc_work);
-       dprintk("svcrdma: svc_rdma_free(%p)\n", rdma);
+       struct svc_xprt *xprt = &rdma->sc_xprt;
+
+       dprintk("svcrdma: %s(%p)\n", __func__, rdma);
 
        /* We should only be called from kref_put */
-       if (atomic_read(&rdma->sc_xprt.xpt_ref.refcount) != 0)
+       if (atomic_read(&xprt->xpt_ref.refcount) != 0)
                pr_err("svcrdma: sc_xprt still in use? (%d)\n",
-                      atomic_read(&rdma->sc_xprt.xpt_ref.refcount));
+                      atomic_read(&xprt->xpt_ref.refcount));
 
        /*
         * Destroy queued, but not processed read completions. Note
@@ -1327,6 +1329,12 @@ static void __svc_rdma_free(struct work_struct *work)
                pr_err("svcrdma: dma still in use? (%d)\n",
                       atomic_read(&rdma->sc_dma_used));
 
+       /* Final put of backchannel client transport */
+       if (xprt->xpt_bc_xprt) {
+               xprt_put(xprt->xpt_bc_xprt);
+               xprt->xpt_bc_xprt = NULL;
+       }
+
        rdma_dealloc_frmr_q(rdma);
        svc_rdma_destroy_ctxts(rdma);
        svc_rdma_destroy_maps(rdma);
index 8c545f7d75257ef72135f8dec24a91ad86fbbc9f..5c7d235672fa666b8d21a34b96199b66bb27906b 100644 (file)
@@ -63,7 +63,7 @@
  */
 
 static unsigned int xprt_rdma_slot_table_entries = RPCRDMA_DEF_SLOT_TABLE;
-static unsigned int xprt_rdma_max_inline_read = RPCRDMA_DEF_INLINE;
+unsigned int xprt_rdma_max_inline_read = RPCRDMA_DEF_INLINE;
 static unsigned int xprt_rdma_max_inline_write = RPCRDMA_DEF_INLINE;
 static unsigned int xprt_rdma_inline_write_padding;
 static unsigned int xprt_rdma_memreg_strategy = RPCRDMA_FRMR;
@@ -143,12 +143,7 @@ static struct ctl_table sunrpc_table[] = {
 
 #endif
 
-#define RPCRDMA_BIND_TO                (60U * HZ)
-#define RPCRDMA_INIT_REEST_TO  (5U * HZ)
-#define RPCRDMA_MAX_REEST_TO   (30U * HZ)
-#define RPCRDMA_IDLE_DISC_TO   (5U * 60 * HZ)
-
-static struct rpc_xprt_ops xprt_rdma_procs;    /* forward reference */
+static struct rpc_xprt_ops xprt_rdma_procs;    /*forward reference */
 
 static void
 xprt_rdma_format_addresses4(struct rpc_xprt *xprt, struct sockaddr *sap)
@@ -174,7 +169,7 @@ xprt_rdma_format_addresses6(struct rpc_xprt *xprt, struct sockaddr *sap)
        xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA6;
 }
 
-static void
+void
 xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap)
 {
        char buf[128];
@@ -203,7 +198,7 @@ xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap)
        xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
 }
 
-static void
+void
 xprt_rdma_free_addresses(struct rpc_xprt *xprt)
 {
        unsigned int i;
@@ -499,7 +494,7 @@ xprt_rdma_allocate(struct rpc_task *task, size_t size)
        if (req == NULL)
                return NULL;
 
-       flags = GFP_NOIO | __GFP_NOWARN;
+       flags = RPCRDMA_DEF_GFP;
        if (RPC_IS_SWAPPER(task))
                flags = __GFP_MEMALLOC | GFP_NOWAIT | __GFP_NOWARN;
 
@@ -639,7 +634,7 @@ drop_connection:
        return -ENOTCONN;       /* implies disconnect */
 }
 
-static void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
+void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
 {
        struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
        long idle_time = 0;
@@ -740,6 +735,11 @@ void xprt_rdma_cleanup(void)
 
        rpcrdma_destroy_wq();
        frwr_destroy_recovery_wq();
+
+       rc = xprt_unregister_transport(&xprt_rdma_bc);
+       if (rc)
+               dprintk("RPC:       %s: xprt_unregister(bc) returned %i\n",
+                       __func__, rc);
 }
 
 int xprt_rdma_init(void)
@@ -763,6 +763,14 @@ int xprt_rdma_init(void)
                return rc;
        }
 
+       rc = xprt_register_transport(&xprt_rdma_bc);
+       if (rc) {
+               xprt_unregister_transport(&xprt_rdma);
+               rpcrdma_destroy_wq();
+               frwr_destroy_recovery_wq();
+               return rc;
+       }
+
        dprintk("RPCRDMA Module Init, register RPC RDMA transport\n");
 
        dprintk("Defaults:\n");
index 72276c7907e452899fa2ba5dc37f3dc606e6bad6..5a38236d051625ce14b5456fc4464298b3ff8e5d 100644 (file)
 #define RDMA_RESOLVE_TIMEOUT   (5000)  /* 5 seconds */
 #define RDMA_CONNECT_RETRY_MAX (2)     /* retries if no listener backlog */
 
+#define RPCRDMA_BIND_TO                (60U * HZ)
+#define RPCRDMA_INIT_REEST_TO  (5U * HZ)
+#define RPCRDMA_MAX_REEST_TO   (30U * HZ)
+#define RPCRDMA_IDLE_DISC_TO   (5U * 60 * HZ)
+
 /*
  * Interface Adapter -- one per transport instance
  */
@@ -147,6 +152,8 @@ rdmab_to_msg(struct rpcrdma_regbuf *rb)
        return (struct rpcrdma_msg *)rb->rg_base;
 }
 
+#define RPCRDMA_DEF_GFP                (GFP_NOIO | __GFP_NOWARN)
+
 /*
  * struct rpcrdma_rep -- this structure encapsulates state required to recv
  * and complete a reply, asychronously. It needs several pieces of
@@ -308,6 +315,8 @@ struct rpcrdma_buffer {
        u32                     rb_bc_srv_max_requests;
        spinlock_t              rb_reqslock;    /* protect rb_allreqs */
        struct list_head        rb_allreqs;
+
+       u32                     rb_bc_max_requests;
 };
 #define rdmab_to_ia(b) (&container_of((b), struct rpcrdma_xprt, rx_buf)->rx_ia)
 
@@ -513,6 +522,10 @@ int rpcrdma_marshal_req(struct rpc_rqst *);
 
 /* RPC/RDMA module init - xprtrdma/transport.c
  */
+extern unsigned int xprt_rdma_max_inline_read;
+void xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap);
+void xprt_rdma_free_addresses(struct rpc_xprt *xprt);
+void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq);
 int xprt_rdma_init(void);
 void xprt_rdma_cleanup(void);
 
@@ -528,4 +541,6 @@ void xprt_rdma_bc_free_rqst(struct rpc_rqst *);
 void xprt_rdma_bc_destroy(struct rpc_xprt *, unsigned int);
 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
 
+extern struct xprt_class xprt_rdma_bc;
+
 #endif                         /* _LINUX_SUNRPC_XPRT_RDMA_H */