2 * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the BSD-type
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
14 * Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
17 * Redistributions in binary form must reproduce the above
18 * copyright notice, this list of conditions and the following
19 * disclaimer in the documentation and/or other materials provided
20 * with the distribution.
22 * Neither the name of the Network Appliance, Inc. nor the names of
23 * its contributors may be used to endorse or promote products
24 * derived from this software without specific prior written
27 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
28 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
29 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
30 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
31 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
32 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
33 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
34 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
35 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
36 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
37 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
43 * This file contains the top-level implementation of an RPC RDMA
46 * Naming convention: functions beginning with xprt_ are part of the
47 * transport switch. All others are RPC RDMA internal.
50 #include <linux/module.h>
51 #include <linux/slab.h>
52 #include <linux/seq_file.h>
53 #include <linux/sunrpc/addr.h>
55 #include "xprt_rdma.h"
57 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
58 # define RPCDBG_FACILITY RPCDBG_TRANS
65 static unsigned int xprt_rdma_slot_table_entries = RPCRDMA_DEF_SLOT_TABLE;
66 unsigned int xprt_rdma_max_inline_read = RPCRDMA_DEF_INLINE;
67 static unsigned int xprt_rdma_max_inline_write = RPCRDMA_DEF_INLINE;
68 static unsigned int xprt_rdma_inline_write_padding;
69 static unsigned int xprt_rdma_memreg_strategy = RPCRDMA_FRMR;
70 int xprt_rdma_pad_optimize = 0;
72 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
74 static unsigned int min_slot_table_size = RPCRDMA_MIN_SLOT_TABLE;
75 static unsigned int max_slot_table_size = RPCRDMA_MAX_SLOT_TABLE;
76 static unsigned int min_inline_size = RPCRDMA_MIN_INLINE;
77 static unsigned int max_inline_size = RPCRDMA_MAX_INLINE;
78 static unsigned int zero;
79 static unsigned int max_padding = PAGE_SIZE;
80 static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS;
81 static unsigned int max_memreg = RPCRDMA_LAST - 1;
83 static struct ctl_table_header *sunrpc_table_header;
85 static struct ctl_table xr_tunables_table[] = {
87 .procname = "rdma_slot_table_entries",
88 .data = &xprt_rdma_slot_table_entries,
89 .maxlen = sizeof(unsigned int),
91 .proc_handler = proc_dointvec_minmax,
92 .extra1 = &min_slot_table_size,
93 .extra2 = &max_slot_table_size
96 .procname = "rdma_max_inline_read",
97 .data = &xprt_rdma_max_inline_read,
98 .maxlen = sizeof(unsigned int),
100 .proc_handler = proc_dointvec_minmax,
101 .extra1 = &min_inline_size,
102 .extra2 = &max_inline_size,
105 .procname = "rdma_max_inline_write",
106 .data = &xprt_rdma_max_inline_write,
107 .maxlen = sizeof(unsigned int),
109 .proc_handler = proc_dointvec_minmax,
110 .extra1 = &min_inline_size,
111 .extra2 = &max_inline_size,
114 .procname = "rdma_inline_write_padding",
115 .data = &xprt_rdma_inline_write_padding,
116 .maxlen = sizeof(unsigned int),
118 .proc_handler = proc_dointvec_minmax,
120 .extra2 = &max_padding,
123 .procname = "rdma_memreg_strategy",
124 .data = &xprt_rdma_memreg_strategy,
125 .maxlen = sizeof(unsigned int),
127 .proc_handler = proc_dointvec_minmax,
128 .extra1 = &min_memreg,
129 .extra2 = &max_memreg,
132 .procname = "rdma_pad_optimize",
133 .data = &xprt_rdma_pad_optimize,
134 .maxlen = sizeof(unsigned int),
136 .proc_handler = proc_dointvec,
141 static struct ctl_table sunrpc_table[] = {
143 .procname = "sunrpc",
145 .child = xr_tunables_table
152 static struct rpc_xprt_ops xprt_rdma_procs; /*forward reference */
155 xprt_rdma_format_addresses4(struct rpc_xprt *xprt, struct sockaddr *sap)
157 struct sockaddr_in *sin = (struct sockaddr_in *)sap;
160 snprintf(buf, sizeof(buf), "%08x", ntohl(sin->sin_addr.s_addr));
161 xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
163 xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA;
167 xprt_rdma_format_addresses6(struct rpc_xprt *xprt, struct sockaddr *sap)
169 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap;
172 snprintf(buf, sizeof(buf), "%pi6", &sin6->sin6_addr);
173 xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
175 xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA6;
179 xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap)
183 switch (sap->sa_family) {
185 xprt_rdma_format_addresses4(xprt, sap);
188 xprt_rdma_format_addresses6(xprt, sap);
191 pr_err("rpcrdma: Unrecognized address family\n");
195 (void)rpc_ntop(sap, buf, sizeof(buf));
196 xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL);
198 snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap));
199 xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
201 snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap));
202 xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
204 xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
208 xprt_rdma_free_addresses(struct rpc_xprt *xprt)
212 for (i = 0; i < RPC_DISPLAY_MAX; i++)
214 case RPC_DISPLAY_PROTO:
215 case RPC_DISPLAY_NETID:
218 kfree(xprt->address_strings[i]);
223 rpcrdma_conn_func(struct rpcrdma_ep *ep)
225 schedule_delayed_work(&ep->rep_connect_worker, 0);
229 rpcrdma_connect_worker(struct work_struct *work)
231 struct rpcrdma_ep *ep =
232 container_of(work, struct rpcrdma_ep, rep_connect_worker.work);
233 struct rpcrdma_xprt *r_xprt =
234 container_of(ep, struct rpcrdma_xprt, rx_ep);
235 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
237 spin_lock_bh(&xprt->transport_lock);
238 if (++xprt->connect_cookie == 0) /* maintain a reserved value */
239 ++xprt->connect_cookie;
240 if (ep->rep_connected > 0) {
241 if (!xprt_test_and_set_connected(xprt))
242 xprt_wake_pending_tasks(xprt, 0);
244 if (xprt_test_and_clear_connected(xprt))
245 xprt_wake_pending_tasks(xprt, -ENOTCONN);
247 spin_unlock_bh(&xprt->transport_lock);
251 xprt_rdma_connect_worker(struct work_struct *work)
253 struct rpcrdma_xprt *r_xprt = container_of(work, struct rpcrdma_xprt,
254 rx_connect_worker.work);
255 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
258 xprt_clear_connected(xprt);
260 dprintk("RPC: %s: %sconnect\n", __func__,
261 r_xprt->rx_ep.rep_connected != 0 ? "re" : "");
262 rc = rpcrdma_ep_connect(&r_xprt->rx_ep, &r_xprt->rx_ia);
264 xprt_wake_pending_tasks(xprt, rc);
266 dprintk("RPC: %s: exit\n", __func__);
267 xprt_clear_connecting(xprt);
271 xprt_rdma_inject_disconnect(struct rpc_xprt *xprt)
273 struct rpcrdma_xprt *r_xprt = container_of(xprt, struct rpcrdma_xprt,
276 pr_info("rpcrdma: injecting transport disconnect on xprt=%p\n", xprt);
277 rdma_disconnect(r_xprt->rx_ia.ri_id);
284 * Free all memory associated with the object, including its own.
285 * NOTE: none of the *destroy methods free memory for their top-level
286 * objects, even though they may have allocated it (they do free
287 * private memory). It's up to the caller to handle it. In this
288 * case (RDMA transport), all structure memory is inlined with the
289 * struct rpcrdma_xprt.
292 xprt_rdma_destroy(struct rpc_xprt *xprt)
294 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
296 dprintk("RPC: %s: called\n", __func__);
298 cancel_delayed_work_sync(&r_xprt->rx_connect_worker);
300 xprt_clear_connected(xprt);
302 rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia);
303 rpcrdma_buffer_destroy(&r_xprt->rx_buf);
304 rpcrdma_ia_close(&r_xprt->rx_ia);
306 xprt_rdma_free_addresses(xprt);
310 dprintk("RPC: %s: returning\n", __func__);
312 module_put(THIS_MODULE);
315 static const struct rpc_timeout xprt_rdma_default_timeout = {
316 .to_initval = 60 * HZ,
317 .to_maxval = 60 * HZ,
321 * xprt_setup_rdma - Set up transport to use RDMA
323 * @args: rpc transport arguments
325 static struct rpc_xprt *
326 xprt_setup_rdma(struct xprt_create *args)
328 struct rpcrdma_create_data_internal cdata;
329 struct rpc_xprt *xprt;
330 struct rpcrdma_xprt *new_xprt;
331 struct rpcrdma_ep *new_ep;
332 struct sockaddr *sap;
335 if (args->addrlen > sizeof(xprt->addr)) {
336 dprintk("RPC: %s: address too large\n", __func__);
337 return ERR_PTR(-EBADF);
340 xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt),
341 xprt_rdma_slot_table_entries,
342 xprt_rdma_slot_table_entries);
344 dprintk("RPC: %s: couldn't allocate rpcrdma_xprt\n",
346 return ERR_PTR(-ENOMEM);
349 /* 60 second timeout, no retries */
350 xprt->timeout = &xprt_rdma_default_timeout;
351 xprt->bind_timeout = RPCRDMA_BIND_TO;
352 xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
353 xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
355 xprt->resvport = 0; /* privileged port not needed */
356 xprt->tsh_size = 0; /* RPC-RDMA handles framing */
357 xprt->ops = &xprt_rdma_procs;
360 * Set up RDMA-specific connect data.
363 sap = (struct sockaddr *)&cdata.addr;
364 memcpy(sap, args->dstaddr, args->addrlen);
366 /* Ensure xprt->addr holds valid server TCP (not RDMA)
367 * address, for any side protocols which peek at it */
368 xprt->prot = IPPROTO_TCP;
369 xprt->addrlen = args->addrlen;
370 memcpy(&xprt->addr, sap, xprt->addrlen);
372 if (rpc_get_port(sap))
373 xprt_set_bound(xprt);
375 cdata.max_requests = xprt->max_reqs;
377 cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */
378 cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */
380 cdata.inline_wsize = xprt_rdma_max_inline_write;
381 if (cdata.inline_wsize > cdata.wsize)
382 cdata.inline_wsize = cdata.wsize;
384 cdata.inline_rsize = xprt_rdma_max_inline_read;
385 if (cdata.inline_rsize > cdata.rsize)
386 cdata.inline_rsize = cdata.rsize;
388 cdata.padding = xprt_rdma_inline_write_padding;
391 * Create new transport instance, which includes initialized
397 new_xprt = rpcx_to_rdmax(xprt);
399 rc = rpcrdma_ia_open(new_xprt, sap, xprt_rdma_memreg_strategy);
404 * initialize and create ep
406 new_xprt->rx_data = cdata;
407 new_ep = &new_xprt->rx_ep;
408 new_ep->rep_remote_addr = cdata.addr;
410 rc = rpcrdma_ep_create(&new_xprt->rx_ep,
411 &new_xprt->rx_ia, &new_xprt->rx_data);
416 * Allocate pre-registered send and receive buffers for headers and
417 * any inline data. Also specify any padding which will be provided
418 * from a preregistered zero buffer.
420 rc = rpcrdma_buffer_create(new_xprt);
425 * Register a callback for connection events. This is necessary because
426 * connection loss notification is async. We also catch connection loss
427 * when reaping receives.
429 INIT_DELAYED_WORK(&new_xprt->rx_connect_worker,
430 xprt_rdma_connect_worker);
432 xprt_rdma_format_addresses(xprt, sap);
433 xprt->max_payload = new_xprt->rx_ia.ri_ops->ro_maxpages(new_xprt);
434 if (xprt->max_payload == 0)
436 xprt->max_payload <<= PAGE_SHIFT;
437 dprintk("RPC: %s: transport data payload maximum: %zu bytes\n",
438 __func__, xprt->max_payload);
440 if (!try_module_get(THIS_MODULE))
443 dprintk("RPC: %s: %s:%s\n", __func__,
444 xprt->address_strings[RPC_DISPLAY_ADDR],
445 xprt->address_strings[RPC_DISPLAY_PORT]);
449 xprt_rdma_free_addresses(xprt);
452 rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia);
454 rpcrdma_ia_close(&new_xprt->rx_ia);
461 * Close a connection, during shutdown or timeout/reconnect
464 xprt_rdma_close(struct rpc_xprt *xprt)
466 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
468 dprintk("RPC: %s: closing\n", __func__);
469 if (r_xprt->rx_ep.rep_connected > 0)
470 xprt->reestablish_timeout = 0;
471 xprt_disconnect_done(xprt);
472 rpcrdma_ep_disconnect(&r_xprt->rx_ep, &r_xprt->rx_ia);
476 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
478 struct sockaddr_in *sap;
480 sap = (struct sockaddr_in *)&xprt->addr;
481 sap->sin_port = htons(port);
482 sap = (struct sockaddr_in *)&rpcx_to_rdmad(xprt).addr;
483 sap->sin_port = htons(port);
484 dprintk("RPC: %s: %u\n", __func__, port);
488 * xprt_rdma_timer - invoked when an RPC times out
489 * @xprt: controlling RPC transport
490 * @task: RPC task that timed out
492 * Invoked when the transport is still connected, but an RPC
493 * retransmit timeout occurs.
495 * Since RDMA connections don't have a keep-alive, forcibly
496 * disconnect and retry to connect. This drives full
497 * detection of the network path, and retransmissions of
501 xprt_rdma_timer(struct rpc_xprt *xprt, struct rpc_task *task)
503 dprintk("RPC: %5u %s: xprt = %p\n", task->tk_pid, __func__, xprt);
505 xprt_force_disconnect(xprt);
509 xprt_rdma_connect(struct rpc_xprt *xprt, struct rpc_task *task)
511 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
513 if (r_xprt->rx_ep.rep_connected != 0) {
515 schedule_delayed_work(&r_xprt->rx_connect_worker,
516 xprt->reestablish_timeout);
517 xprt->reestablish_timeout <<= 1;
518 if (xprt->reestablish_timeout > RPCRDMA_MAX_REEST_TO)
519 xprt->reestablish_timeout = RPCRDMA_MAX_REEST_TO;
520 else if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
521 xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
523 schedule_delayed_work(&r_xprt->rx_connect_worker, 0);
524 if (!RPC_IS_ASYNC(task))
525 flush_delayed_work(&r_xprt->rx_connect_worker);
529 /* Allocate a fixed-size buffer in which to construct and send the
530 * RPC-over-RDMA header for this request.
533 rpcrdma_get_rdmabuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
536 size_t size = RPCRDMA_HDRBUF_SIZE;
537 struct rpcrdma_regbuf *rb;
542 rb = rpcrdma_alloc_regbuf(size, DMA_TO_DEVICE, flags);
546 r_xprt->rx_stats.hardway_register_count += size;
547 req->rl_rdmabuf = rb;
552 rpcrdma_get_sendbuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
553 size_t size, gfp_t flags)
555 struct rpcrdma_regbuf *rb;
557 if (req->rl_sendbuf && rdmab_length(req->rl_sendbuf) >= size)
560 rb = rpcrdma_alloc_regbuf(size, DMA_TO_DEVICE, flags);
564 rpcrdma_free_regbuf(req->rl_sendbuf);
565 r_xprt->rx_stats.hardway_register_count += size;
566 req->rl_sendbuf = rb;
570 /* The rq_rcv_buf is used only if a Reply chunk is necessary.
571 * The decision to use a Reply chunk is made later in
572 * rpcrdma_marshal_req. This buffer is registered at that time.
574 * Otherwise, the associated RPC Reply arrives in a separate
575 * Receive buffer, arbitrarily chosen by the HCA. The buffer
576 * allocated here for the RPC Reply is not utilized in that
577 * case. See rpcrdma_inline_fixup.
579 * A regbuf is used here to remember the buffer size.
582 rpcrdma_get_recvbuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
583 size_t size, gfp_t flags)
585 struct rpcrdma_regbuf *rb;
587 if (req->rl_recvbuf && rdmab_length(req->rl_recvbuf) >= size)
590 rb = rpcrdma_alloc_regbuf(size, DMA_NONE, flags);
594 rpcrdma_free_regbuf(req->rl_recvbuf);
595 r_xprt->rx_stats.hardway_register_count += size;
596 req->rl_recvbuf = rb;
601 * xprt_rdma_allocate - allocate transport resources for an RPC
605 * 0: Success; rq_buffer points to RPC buffer to use
606 * ENOMEM: Out of memory, call again later
607 * EIO: A permanent error occurred, do not retry
609 * The RDMA allocate/free functions need the task structure as a place
610 * to hide the struct rpcrdma_req, which is necessary for the actual
611 * send/recv sequence.
613 * xprt_rdma_allocate provides buffers that are already mapped for
614 * DMA, and a local DMA lkey is provided for each.
617 xprt_rdma_allocate(struct rpc_task *task)
619 struct rpc_rqst *rqst = task->tk_rqstp;
620 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
621 struct rpcrdma_req *req;
624 req = rpcrdma_buffer_get(&r_xprt->rx_buf);
628 flags = RPCRDMA_DEF_GFP;
629 if (RPC_IS_SWAPPER(task))
630 flags = __GFP_MEMALLOC | GFP_NOWAIT | __GFP_NOWARN;
632 if (!rpcrdma_get_rdmabuf(r_xprt, req, flags))
634 if (!rpcrdma_get_sendbuf(r_xprt, req, rqst->rq_callsize, flags))
636 if (!rpcrdma_get_recvbuf(r_xprt, req, rqst->rq_rcvsize, flags))
639 dprintk("RPC: %5u %s: send size = %zd, recv size = %zd, req = %p\n",
640 task->tk_pid, __func__, rqst->rq_callsize,
641 rqst->rq_rcvsize, req);
643 req->rl_connect_cookie = 0; /* our reserved value */
644 rpcrdma_set_xprtdata(rqst, req);
645 rqst->rq_buffer = req->rl_sendbuf->rg_base;
646 rqst->rq_rbuffer = req->rl_recvbuf->rg_base;
650 rpcrdma_buffer_put(req);
655 * xprt_rdma_free - release resources allocated by xprt_rdma_allocate
658 * Caller guarantees rqst->rq_buffer is non-NULL.
661 xprt_rdma_free(struct rpc_task *task)
663 struct rpc_rqst *rqst = task->tk_rqstp;
664 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
665 struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
666 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
668 if (req->rl_backchannel)
671 dprintk("RPC: %s: called on 0x%p\n", __func__, req->rl_reply);
673 if (unlikely(!list_empty(&req->rl_registered)))
674 ia->ri_ops->ro_unmap_safe(r_xprt, req, !RPC_IS_ASYNC(task));
675 rpcrdma_unmap_sges(ia, req);
676 rpcrdma_buffer_put(req);
680 * xprt_rdma_send_request - marshal and send an RPC request
681 * @task: RPC task with an RPC message in rq_snd_buf
684 * 0: The request has been sent
685 * ENOTCONN: Caller needs to invoke connect logic then call again
686 * ENOBUFS: Call again later to send the request
687 * EIO: A permanent error occurred. The request was not sent,
688 * and don't try it again
690 * send_request invokes the meat of RPC RDMA. It must do the following:
692 * 1. Marshal the RPC request into an RPC RDMA request, which means
693 * putting a header in front of data, and creating IOVs for RDMA
694 * from those in the request.
695 * 2. In marshaling, detect opportunities for RDMA, and use them.
696 * 3. Post a recv message to set up asynch completion, then send
697 * the request (rpcrdma_ep_post).
698 * 4. No partial sends are possible in the RPC-RDMA protocol (as in UDP).
701 xprt_rdma_send_request(struct rpc_task *task)
703 struct rpc_rqst *rqst = task->tk_rqstp;
704 struct rpc_xprt *xprt = rqst->rq_xprt;
705 struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
706 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
709 /* On retransmit, remove any previously registered chunks */
710 if (unlikely(!list_empty(&req->rl_registered)))
711 r_xprt->rx_ia.ri_ops->ro_unmap_safe(r_xprt, req, false);
713 rc = rpcrdma_marshal_req(rqst);
717 if (req->rl_reply == NULL) /* e.g. reconnection */
718 rpcrdma_recv_buffer_get(req);
720 /* Must suppress retransmit to maintain credits */
721 if (req->rl_connect_cookie == xprt->connect_cookie)
722 goto drop_connection;
723 req->rl_connect_cookie = xprt->connect_cookie;
725 if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
726 goto drop_connection;
728 rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len;
729 rqst->rq_bytes_sent = 0;
736 xprt_disconnect_done(xprt);
737 return -ENOTCONN; /* implies disconnect */
740 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
742 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
745 if (xprt_connected(xprt))
746 idle_time = (long)(jiffies - xprt->last_used) / HZ;
748 seq_puts(seq, "\txprt:\trdma ");
749 seq_printf(seq, "%u %lu %lu %lu %ld %lu %lu %lu %llu %llu ",
750 0, /* need a local port? */
751 xprt->stat.bind_count,
752 xprt->stat.connect_count,
753 xprt->stat.connect_time,
760 seq_printf(seq, "%lu %lu %lu %llu %llu %llu %llu %lu %lu %lu %lu ",
761 r_xprt->rx_stats.read_chunk_count,
762 r_xprt->rx_stats.write_chunk_count,
763 r_xprt->rx_stats.reply_chunk_count,
764 r_xprt->rx_stats.total_rdma_request,
765 r_xprt->rx_stats.total_rdma_reply,
766 r_xprt->rx_stats.pullup_copy_count,
767 r_xprt->rx_stats.fixup_copy_count,
768 r_xprt->rx_stats.hardway_register_count,
769 r_xprt->rx_stats.failed_marshal_count,
770 r_xprt->rx_stats.bad_reply_count,
771 r_xprt->rx_stats.nomsg_call_count);
772 seq_printf(seq, "%lu %lu %lu %lu\n",
773 r_xprt->rx_stats.mrs_recovered,
774 r_xprt->rx_stats.mrs_orphaned,
775 r_xprt->rx_stats.mrs_allocated,
776 r_xprt->rx_stats.local_inv_needed);
780 xprt_rdma_enable_swap(struct rpc_xprt *xprt)
786 xprt_rdma_disable_swap(struct rpc_xprt *xprt)
791 * Plumbing for rpc transport switch and kernel module
794 static struct rpc_xprt_ops xprt_rdma_procs = {
795 .reserve_xprt = xprt_reserve_xprt_cong,
796 .release_xprt = xprt_release_xprt_cong, /* sunrpc/xprt.c */
797 .alloc_slot = xprt_alloc_slot,
798 .release_request = xprt_release_rqst_cong, /* ditto */
799 .set_retrans_timeout = xprt_set_retrans_timeout_def, /* ditto */
800 .timer = xprt_rdma_timer,
801 .rpcbind = rpcb_getport_async, /* sunrpc/rpcb_clnt.c */
802 .set_port = xprt_rdma_set_port,
803 .connect = xprt_rdma_connect,
804 .buf_alloc = xprt_rdma_allocate,
805 .buf_free = xprt_rdma_free,
806 .send_request = xprt_rdma_send_request,
807 .close = xprt_rdma_close,
808 .destroy = xprt_rdma_destroy,
809 .print_stats = xprt_rdma_print_stats,
810 .enable_swap = xprt_rdma_enable_swap,
811 .disable_swap = xprt_rdma_disable_swap,
812 .inject_disconnect = xprt_rdma_inject_disconnect,
813 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
814 .bc_setup = xprt_rdma_bc_setup,
815 .bc_up = xprt_rdma_bc_up,
816 .bc_maxpayload = xprt_rdma_bc_maxpayload,
817 .bc_free_rqst = xprt_rdma_bc_free_rqst,
818 .bc_destroy = xprt_rdma_bc_destroy,
822 static struct xprt_class xprt_rdma = {
823 .list = LIST_HEAD_INIT(xprt_rdma.list),
825 .owner = THIS_MODULE,
826 .ident = XPRT_TRANSPORT_RDMA,
827 .setup = xprt_setup_rdma,
830 void xprt_rdma_cleanup(void)
834 dprintk("RPCRDMA Module Removed, deregister RPC RDMA transport\n");
835 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
836 if (sunrpc_table_header) {
837 unregister_sysctl_table(sunrpc_table_header);
838 sunrpc_table_header = NULL;
841 rc = xprt_unregister_transport(&xprt_rdma);
843 dprintk("RPC: %s: xprt_unregister returned %i\n",
846 rpcrdma_destroy_wq();
848 rc = xprt_unregister_transport(&xprt_rdma_bc);
850 dprintk("RPC: %s: xprt_unregister(bc) returned %i\n",
854 int xprt_rdma_init(void)
858 rc = rpcrdma_alloc_wq();
862 rc = xprt_register_transport(&xprt_rdma);
864 rpcrdma_destroy_wq();
868 rc = xprt_register_transport(&xprt_rdma_bc);
870 xprt_unregister_transport(&xprt_rdma);
871 rpcrdma_destroy_wq();
875 dprintk("RPCRDMA Module Init, register RPC RDMA transport\n");
877 dprintk("Defaults:\n");
878 dprintk("\tSlots %d\n"
879 "\tMaxInlineRead %d\n\tMaxInlineWrite %d\n",
880 xprt_rdma_slot_table_entries,
881 xprt_rdma_max_inline_read, xprt_rdma_max_inline_write);
882 dprintk("\tPadding %d\n\tMemreg %d\n",
883 xprt_rdma_inline_write_padding, xprt_rdma_memreg_strategy);
885 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
886 if (!sunrpc_table_header)
887 sunrpc_table_header = register_sysctl_table(sunrpc_table);