2 * Copyright (c) 2014-2017 Oracle. All rights reserved.
3 * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
5 * This software is available to you under a choice of one of two
6 * licenses. You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the BSD-type
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
15 * Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
18 * Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials provided
21 * with the distribution.
23 * Neither the name of the Network Appliance, Inc. nor the names of
24 * its contributors may be used to endorse or promote products
25 * derived from this software without specific prior written
28 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
29 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
30 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
31 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
32 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
33 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
34 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
35 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
36 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
37 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
38 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
44 * This file contains the top-level implementation of an RPC RDMA
47 * Naming convention: functions beginning with xprt_ are part of the
48 * transport switch. All others are RPC RDMA internal.
51 #include <linux/module.h>
52 #include <linux/slab.h>
53 #include <linux/seq_file.h>
54 #include <linux/sunrpc/addr.h>
55 #include <linux/smp.h>
57 #include "xprt_rdma.h"
59 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
60 # define RPCDBG_FACILITY RPCDBG_TRANS
67 static unsigned int xprt_rdma_slot_table_entries = RPCRDMA_DEF_SLOT_TABLE;
68 unsigned int xprt_rdma_max_inline_read = RPCRDMA_DEF_INLINE;
69 static unsigned int xprt_rdma_max_inline_write = RPCRDMA_DEF_INLINE;
70 unsigned int xprt_rdma_memreg_strategy = RPCRDMA_FRMR;
71 int xprt_rdma_pad_optimize;
73 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
75 static unsigned int min_slot_table_size = RPCRDMA_MIN_SLOT_TABLE;
76 static unsigned int max_slot_table_size = RPCRDMA_MAX_SLOT_TABLE;
77 static unsigned int min_inline_size = RPCRDMA_MIN_INLINE;
78 static unsigned int max_inline_size = RPCRDMA_MAX_INLINE;
79 static unsigned int zero;
80 static unsigned int max_padding = PAGE_SIZE;
81 static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS;
82 static unsigned int max_memreg = RPCRDMA_LAST - 1;
83 static unsigned int dummy;
85 static struct ctl_table_header *sunrpc_table_header;
87 static struct ctl_table xr_tunables_table[] = {
89 .procname = "rdma_slot_table_entries",
90 .data = &xprt_rdma_slot_table_entries,
91 .maxlen = sizeof(unsigned int),
93 .proc_handler = proc_dointvec_minmax,
94 .extra1 = &min_slot_table_size,
95 .extra2 = &max_slot_table_size
98 .procname = "rdma_max_inline_read",
99 .data = &xprt_rdma_max_inline_read,
100 .maxlen = sizeof(unsigned int),
102 .proc_handler = proc_dointvec_minmax,
103 .extra1 = &min_inline_size,
104 .extra2 = &max_inline_size,
107 .procname = "rdma_max_inline_write",
108 .data = &xprt_rdma_max_inline_write,
109 .maxlen = sizeof(unsigned int),
111 .proc_handler = proc_dointvec_minmax,
112 .extra1 = &min_inline_size,
113 .extra2 = &max_inline_size,
116 .procname = "rdma_inline_write_padding",
118 .maxlen = sizeof(unsigned int),
120 .proc_handler = proc_dointvec_minmax,
122 .extra2 = &max_padding,
125 .procname = "rdma_memreg_strategy",
126 .data = &xprt_rdma_memreg_strategy,
127 .maxlen = sizeof(unsigned int),
129 .proc_handler = proc_dointvec_minmax,
130 .extra1 = &min_memreg,
131 .extra2 = &max_memreg,
134 .procname = "rdma_pad_optimize",
135 .data = &xprt_rdma_pad_optimize,
136 .maxlen = sizeof(unsigned int),
138 .proc_handler = proc_dointvec,
143 static struct ctl_table sunrpc_table[] = {
145 .procname = "sunrpc",
147 .child = xr_tunables_table
154 static const struct rpc_xprt_ops xprt_rdma_procs;
157 xprt_rdma_format_addresses4(struct rpc_xprt *xprt, struct sockaddr *sap)
159 struct sockaddr_in *sin = (struct sockaddr_in *)sap;
162 snprintf(buf, sizeof(buf), "%08x", ntohl(sin->sin_addr.s_addr));
163 xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
165 xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA;
169 xprt_rdma_format_addresses6(struct rpc_xprt *xprt, struct sockaddr *sap)
171 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap;
174 snprintf(buf, sizeof(buf), "%pi6", &sin6->sin6_addr);
175 xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
177 xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA6;
181 xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap)
185 switch (sap->sa_family) {
187 xprt_rdma_format_addresses4(xprt, sap);
190 xprt_rdma_format_addresses6(xprt, sap);
193 pr_err("rpcrdma: Unrecognized address family\n");
197 (void)rpc_ntop(sap, buf, sizeof(buf));
198 xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL);
200 snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap));
201 xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
203 snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap));
204 xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
206 xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
210 xprt_rdma_free_addresses(struct rpc_xprt *xprt)
214 for (i = 0; i < RPC_DISPLAY_MAX; i++)
216 case RPC_DISPLAY_PROTO:
217 case RPC_DISPLAY_NETID:
220 kfree(xprt->address_strings[i]);
225 rpcrdma_conn_func(struct rpcrdma_ep *ep)
227 schedule_delayed_work(&ep->rep_connect_worker, 0);
231 rpcrdma_connect_worker(struct work_struct *work)
233 struct rpcrdma_ep *ep =
234 container_of(work, struct rpcrdma_ep, rep_connect_worker.work);
235 struct rpcrdma_xprt *r_xprt =
236 container_of(ep, struct rpcrdma_xprt, rx_ep);
237 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
239 spin_lock_bh(&xprt->transport_lock);
240 if (++xprt->connect_cookie == 0) /* maintain a reserved value */
241 ++xprt->connect_cookie;
242 if (ep->rep_connected > 0) {
243 if (!xprt_test_and_set_connected(xprt))
244 xprt_wake_pending_tasks(xprt, 0);
246 if (xprt_test_and_clear_connected(xprt))
247 xprt_wake_pending_tasks(xprt, -ENOTCONN);
249 spin_unlock_bh(&xprt->transport_lock);
253 xprt_rdma_connect_worker(struct work_struct *work)
255 struct rpcrdma_xprt *r_xprt = container_of(work, struct rpcrdma_xprt,
256 rx_connect_worker.work);
257 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
260 xprt_clear_connected(xprt);
262 dprintk("RPC: %s: %sconnect\n", __func__,
263 r_xprt->rx_ep.rep_connected != 0 ? "re" : "");
264 rc = rpcrdma_ep_connect(&r_xprt->rx_ep, &r_xprt->rx_ia);
266 xprt_wake_pending_tasks(xprt, rc);
268 dprintk("RPC: %s: exit\n", __func__);
269 xprt_clear_connecting(xprt);
273 xprt_rdma_inject_disconnect(struct rpc_xprt *xprt)
275 struct rpcrdma_xprt *r_xprt = container_of(xprt, struct rpcrdma_xprt,
278 pr_info("rpcrdma: injecting transport disconnect on xprt=%p\n", xprt);
279 rdma_disconnect(r_xprt->rx_ia.ri_id);
286 * Free all memory associated with the object, including its own.
287 * NOTE: none of the *destroy methods free memory for their top-level
288 * objects, even though they may have allocated it (they do free
289 * private memory). It's up to the caller to handle it. In this
290 * case (RDMA transport), all structure memory is inlined with the
291 * struct rpcrdma_xprt.
294 xprt_rdma_destroy(struct rpc_xprt *xprt)
296 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
298 dprintk("RPC: %s: called\n", __func__);
300 cancel_delayed_work_sync(&r_xprt->rx_connect_worker);
302 xprt_clear_connected(xprt);
304 rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia);
305 rpcrdma_buffer_destroy(&r_xprt->rx_buf);
306 rpcrdma_ia_close(&r_xprt->rx_ia);
308 xprt_rdma_free_addresses(xprt);
312 dprintk("RPC: %s: returning\n", __func__);
314 module_put(THIS_MODULE);
317 static const struct rpc_timeout xprt_rdma_default_timeout = {
318 .to_initval = 60 * HZ,
319 .to_maxval = 60 * HZ,
323 * xprt_setup_rdma - Set up transport to use RDMA
325 * @args: rpc transport arguments
327 static struct rpc_xprt *
328 xprt_setup_rdma(struct xprt_create *args)
330 struct rpcrdma_create_data_internal cdata;
331 struct rpc_xprt *xprt;
332 struct rpcrdma_xprt *new_xprt;
333 struct rpcrdma_ep *new_ep;
334 struct sockaddr *sap;
337 if (args->addrlen > sizeof(xprt->addr)) {
338 dprintk("RPC: %s: address too large\n", __func__);
339 return ERR_PTR(-EBADF);
342 xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt),
343 xprt_rdma_slot_table_entries,
344 xprt_rdma_slot_table_entries);
346 dprintk("RPC: %s: couldn't allocate rpcrdma_xprt\n",
348 return ERR_PTR(-ENOMEM);
351 /* 60 second timeout, no retries */
352 xprt->timeout = &xprt_rdma_default_timeout;
353 xprt->bind_timeout = RPCRDMA_BIND_TO;
354 xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
355 xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
357 xprt->resvport = 0; /* privileged port not needed */
358 xprt->tsh_size = 0; /* RPC-RDMA handles framing */
359 xprt->ops = &xprt_rdma_procs;
362 * Set up RDMA-specific connect data.
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);
374 xprt_rdma_format_addresses(xprt, sap);
376 cdata.max_requests = xprt->max_reqs;
378 cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */
379 cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */
381 cdata.inline_wsize = xprt_rdma_max_inline_write;
382 if (cdata.inline_wsize > cdata.wsize)
383 cdata.inline_wsize = cdata.wsize;
385 cdata.inline_rsize = xprt_rdma_max_inline_read;
386 if (cdata.inline_rsize > cdata.rsize)
387 cdata.inline_rsize = cdata.rsize;
390 * Create new transport instance, which includes initialized
396 new_xprt = rpcx_to_rdmax(xprt);
398 rc = rpcrdma_ia_open(new_xprt);
403 * initialize and create ep
405 new_xprt->rx_data = cdata;
406 new_ep = &new_xprt->rx_ep;
408 rc = rpcrdma_ep_create(&new_xprt->rx_ep,
409 &new_xprt->rx_ia, &new_xprt->rx_data);
413 rc = rpcrdma_buffer_create(new_xprt);
417 INIT_DELAYED_WORK(&new_xprt->rx_connect_worker,
418 xprt_rdma_connect_worker);
420 xprt->max_payload = new_xprt->rx_ia.ri_ops->ro_maxpages(new_xprt);
421 if (xprt->max_payload == 0)
423 xprt->max_payload <<= PAGE_SHIFT;
424 dprintk("RPC: %s: transport data payload maximum: %zu bytes\n",
425 __func__, xprt->max_payload);
427 if (!try_module_get(THIS_MODULE))
430 dprintk("RPC: %s: %s:%s\n", __func__,
431 xprt->address_strings[RPC_DISPLAY_ADDR],
432 xprt->address_strings[RPC_DISPLAY_PORT]);
436 rpcrdma_buffer_destroy(&new_xprt->rx_buf);
439 rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia);
441 rpcrdma_ia_close(&new_xprt->rx_ia);
443 xprt_rdma_free_addresses(xprt);
449 * xprt_rdma_close - Close down RDMA connection
450 * @xprt: generic transport to be closed
452 * Called during transport shutdown reconnect, or device
453 * removal. Caller holds the transport's write lock.
456 xprt_rdma_close(struct rpc_xprt *xprt)
458 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
459 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
460 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
462 dprintk("RPC: %s: closing xprt %p\n", __func__, xprt);
464 if (test_and_clear_bit(RPCRDMA_IAF_REMOVING, &ia->ri_flags)) {
465 xprt_clear_connected(xprt);
466 rpcrdma_ia_remove(ia);
469 if (ep->rep_connected == -ENODEV)
471 if (ep->rep_connected > 0)
472 xprt->reestablish_timeout = 0;
473 xprt_disconnect_done(xprt);
474 rpcrdma_ep_disconnect(ep, ia);
478 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
480 struct sockaddr_in *sap;
482 sap = (struct sockaddr_in *)&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;
548 xdr_buf_init(&req->rl_hdrbuf, rb->rg_base, rdmab_length(rb));
553 rpcrdma_get_sendbuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
554 size_t size, gfp_t flags)
556 struct rpcrdma_regbuf *rb;
558 if (req->rl_sendbuf && rdmab_length(req->rl_sendbuf) >= size)
561 rb = rpcrdma_alloc_regbuf(size, DMA_TO_DEVICE, flags);
565 rpcrdma_free_regbuf(req->rl_sendbuf);
566 r_xprt->rx_stats.hardway_register_count += size;
567 req->rl_sendbuf = rb;
571 /* The rq_rcv_buf is used only if a Reply chunk is necessary.
572 * The decision to use a Reply chunk is made later in
573 * rpcrdma_marshal_req. This buffer is registered at that time.
575 * Otherwise, the associated RPC Reply arrives in a separate
576 * Receive buffer, arbitrarily chosen by the HCA. The buffer
577 * allocated here for the RPC Reply is not utilized in that
578 * case. See rpcrdma_inline_fixup.
580 * A regbuf is used here to remember the buffer size.
583 rpcrdma_get_recvbuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
584 size_t size, gfp_t flags)
586 struct rpcrdma_regbuf *rb;
588 if (req->rl_recvbuf && rdmab_length(req->rl_recvbuf) >= size)
591 rb = rpcrdma_alloc_regbuf(size, DMA_NONE, flags);
595 rpcrdma_free_regbuf(req->rl_recvbuf);
596 r_xprt->rx_stats.hardway_register_count += size;
597 req->rl_recvbuf = rb;
602 * xprt_rdma_allocate - allocate transport resources for an RPC
606 * 0: Success; rq_buffer points to RPC buffer to use
607 * ENOMEM: Out of memory, call again later
608 * EIO: A permanent error occurred, do not retry
610 * The RDMA allocate/free functions need the task structure as a place
611 * to hide the struct rpcrdma_req, which is necessary for the actual
612 * send/recv sequence.
614 * xprt_rdma_allocate provides buffers that are already mapped for
615 * DMA, and a local DMA lkey is provided for each.
618 xprt_rdma_allocate(struct rpc_task *task)
620 struct rpc_rqst *rqst = task->tk_rqstp;
621 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
622 struct rpcrdma_req *req;
625 req = rpcrdma_buffer_get(&r_xprt->rx_buf);
629 flags = RPCRDMA_DEF_GFP;
630 if (RPC_IS_SWAPPER(task))
631 flags = __GFP_MEMALLOC | GFP_NOWAIT | __GFP_NOWARN;
633 if (!rpcrdma_get_rdmabuf(r_xprt, req, flags))
635 if (!rpcrdma_get_sendbuf(r_xprt, req, rqst->rq_callsize, flags))
637 if (!rpcrdma_get_recvbuf(r_xprt, req, rqst->rq_rcvsize, flags))
640 dprintk("RPC: %5u %s: send size = %zd, recv size = %zd, req = %p\n",
641 task->tk_pid, __func__, rqst->rq_callsize,
642 rqst->rq_rcvsize, req);
644 req->rl_cpu = smp_processor_id();
645 req->rl_connect_cookie = 0; /* our reserved value */
646 rpcrdma_set_xprtdata(rqst, req);
647 rqst->rq_buffer = req->rl_sendbuf->rg_base;
648 rqst->rq_rbuffer = req->rl_recvbuf->rg_base;
652 rpcrdma_buffer_put(req);
657 * xprt_rdma_free - release resources allocated by xprt_rdma_allocate
660 * Caller guarantees rqst->rq_buffer is non-NULL.
663 xprt_rdma_free(struct rpc_task *task)
665 struct rpc_rqst *rqst = task->tk_rqstp;
666 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
667 struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
669 if (test_bit(RPCRDMA_REQ_F_BACKCHANNEL, &req->rl_flags))
672 dprintk("RPC: %s: called on 0x%p\n", __func__, req->rl_reply);
674 if (test_bit(RPCRDMA_REQ_F_PENDING, &req->rl_flags))
675 rpcrdma_release_rqst(r_xprt, 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
683 * Caller holds the transport's write lock.
686 * 0: The request has been sent
687 * ENOTCONN: Caller needs to invoke connect logic then call again
688 * ENOBUFS: Call again later to send the request
689 * EIO: A permanent error occurred. The request was not sent,
690 * and don't try it again
692 * send_request invokes the meat of RPC RDMA. It must do the following:
694 * 1. Marshal the RPC request into an RPC RDMA request, which means
695 * putting a header in front of data, and creating IOVs for RDMA
696 * from those in the request.
697 * 2. In marshaling, detect opportunities for RDMA, and use them.
698 * 3. Post a recv message to set up asynch completion, then send
699 * the request (rpcrdma_ep_post).
700 * 4. No partial sends are possible in the RPC-RDMA protocol (as in UDP).
703 xprt_rdma_send_request(struct rpc_task *task)
705 struct rpc_rqst *rqst = task->tk_rqstp;
706 struct rpc_xprt *xprt = rqst->rq_xprt;
707 struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
708 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
711 if (!xprt_connected(xprt))
712 goto drop_connection;
714 /* On retransmit, remove any previously registered chunks */
715 if (unlikely(!list_empty(&req->rl_registered)))
716 r_xprt->rx_ia.ri_ops->ro_unmap_sync(r_xprt,
717 &req->rl_registered);
719 rc = rpcrdma_marshal_req(r_xprt, rqst);
723 if (req->rl_reply == NULL) /* e.g. reconnection */
724 rpcrdma_recv_buffer_get(req);
726 /* Must suppress retransmit to maintain credits */
727 if (req->rl_connect_cookie == xprt->connect_cookie)
728 goto drop_connection;
729 req->rl_connect_cookie = xprt->connect_cookie;
731 __set_bit(RPCRDMA_REQ_F_PENDING, &req->rl_flags);
732 if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
733 goto drop_connection;
735 rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len;
736 rqst->rq_bytes_sent = 0;
743 xprt_disconnect_done(xprt);
744 return -ENOTCONN; /* implies disconnect */
747 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
749 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
752 if (xprt_connected(xprt))
753 idle_time = (long)(jiffies - xprt->last_used) / HZ;
755 seq_puts(seq, "\txprt:\trdma ");
756 seq_printf(seq, "%u %lu %lu %lu %ld %lu %lu %lu %llu %llu ",
757 0, /* need a local port? */
758 xprt->stat.bind_count,
759 xprt->stat.connect_count,
760 xprt->stat.connect_time,
767 seq_printf(seq, "%lu %lu %lu %llu %llu %llu %llu %lu %lu %lu %lu ",
768 r_xprt->rx_stats.read_chunk_count,
769 r_xprt->rx_stats.write_chunk_count,
770 r_xprt->rx_stats.reply_chunk_count,
771 r_xprt->rx_stats.total_rdma_request,
772 r_xprt->rx_stats.total_rdma_reply,
773 r_xprt->rx_stats.pullup_copy_count,
774 r_xprt->rx_stats.fixup_copy_count,
775 r_xprt->rx_stats.hardway_register_count,
776 r_xprt->rx_stats.failed_marshal_count,
777 r_xprt->rx_stats.bad_reply_count,
778 r_xprt->rx_stats.nomsg_call_count);
779 seq_printf(seq, "%lu %lu %lu %lu %lu %lu\n",
780 r_xprt->rx_stats.mrs_recovered,
781 r_xprt->rx_stats.mrs_orphaned,
782 r_xprt->rx_stats.mrs_allocated,
783 r_xprt->rx_stats.local_inv_needed,
784 r_xprt->rx_stats.empty_sendctx_q,
785 r_xprt->rx_stats.reply_waits_for_send);
789 xprt_rdma_enable_swap(struct rpc_xprt *xprt)
795 xprt_rdma_disable_swap(struct rpc_xprt *xprt)
800 * Plumbing for rpc transport switch and kernel module
803 static const struct rpc_xprt_ops xprt_rdma_procs = {
804 .reserve_xprt = xprt_reserve_xprt_cong,
805 .release_xprt = xprt_release_xprt_cong, /* sunrpc/xprt.c */
806 .alloc_slot = xprt_alloc_slot,
807 .release_request = xprt_release_rqst_cong, /* ditto */
808 .set_retrans_timeout = xprt_set_retrans_timeout_def, /* ditto */
809 .timer = xprt_rdma_timer,
810 .rpcbind = rpcb_getport_async, /* sunrpc/rpcb_clnt.c */
811 .set_port = xprt_rdma_set_port,
812 .connect = xprt_rdma_connect,
813 .buf_alloc = xprt_rdma_allocate,
814 .buf_free = xprt_rdma_free,
815 .send_request = xprt_rdma_send_request,
816 .close = xprt_rdma_close,
817 .destroy = xprt_rdma_destroy,
818 .print_stats = xprt_rdma_print_stats,
819 .enable_swap = xprt_rdma_enable_swap,
820 .disable_swap = xprt_rdma_disable_swap,
821 .inject_disconnect = xprt_rdma_inject_disconnect,
822 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
823 .bc_setup = xprt_rdma_bc_setup,
824 .bc_up = xprt_rdma_bc_up,
825 .bc_maxpayload = xprt_rdma_bc_maxpayload,
826 .bc_free_rqst = xprt_rdma_bc_free_rqst,
827 .bc_destroy = xprt_rdma_bc_destroy,
831 static struct xprt_class xprt_rdma = {
832 .list = LIST_HEAD_INIT(xprt_rdma.list),
834 .owner = THIS_MODULE,
835 .ident = XPRT_TRANSPORT_RDMA,
836 .setup = xprt_setup_rdma,
839 void xprt_rdma_cleanup(void)
843 dprintk("RPCRDMA Module Removed, deregister RPC RDMA transport\n");
844 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
845 if (sunrpc_table_header) {
846 unregister_sysctl_table(sunrpc_table_header);
847 sunrpc_table_header = NULL;
850 rc = xprt_unregister_transport(&xprt_rdma);
852 dprintk("RPC: %s: xprt_unregister returned %i\n",
855 rpcrdma_destroy_wq();
857 rc = xprt_unregister_transport(&xprt_rdma_bc);
859 dprintk("RPC: %s: xprt_unregister(bc) returned %i\n",
863 int xprt_rdma_init(void)
867 rc = rpcrdma_alloc_wq();
871 rc = xprt_register_transport(&xprt_rdma);
873 rpcrdma_destroy_wq();
877 rc = xprt_register_transport(&xprt_rdma_bc);
879 xprt_unregister_transport(&xprt_rdma);
880 rpcrdma_destroy_wq();
884 dprintk("RPCRDMA Module Init, register RPC RDMA transport\n");
886 dprintk("Defaults:\n");
887 dprintk("\tSlots %d\n"
888 "\tMaxInlineRead %d\n\tMaxInlineWrite %d\n",
889 xprt_rdma_slot_table_entries,
890 xprt_rdma_max_inline_read, xprt_rdma_max_inline_write);
891 dprintk("\tPadding 0\n\tMemreg %d\n", xprt_rdma_memreg_strategy);
893 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
894 if (!sunrpc_table_header)
895 sunrpc_table_header = register_sysctl_table(sunrpc_table);