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/init.h>
52 #include <linux/seq_file.h>
54 #include "xprt_rdma.h"
57 # define RPCDBG_FACILITY RPCDBG_TRANS
60 MODULE_LICENSE("Dual BSD/GPL");
62 MODULE_DESCRIPTION("RPC/RDMA Transport for Linux kernel NFS");
63 MODULE_AUTHOR("Network Appliance, Inc.");
69 static unsigned int xprt_rdma_slot_table_entries = RPCRDMA_DEF_SLOT_TABLE;
70 static unsigned int xprt_rdma_max_inline_read = RPCRDMA_DEF_INLINE;
71 static unsigned int xprt_rdma_max_inline_write = RPCRDMA_DEF_INLINE;
72 static unsigned int xprt_rdma_inline_write_padding;
73 static unsigned int xprt_rdma_memreg_strategy = RPCRDMA_FRMR;
74 int xprt_rdma_pad_optimize = 0;
78 static unsigned int min_slot_table_size = RPCRDMA_MIN_SLOT_TABLE;
79 static unsigned int max_slot_table_size = RPCRDMA_MAX_SLOT_TABLE;
80 static unsigned int zero;
81 static unsigned int max_padding = PAGE_SIZE;
82 static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS;
83 static unsigned int max_memreg = RPCRDMA_LAST - 1;
85 static struct ctl_table_header *sunrpc_table_header;
87 static ctl_table xr_tunables_table[] = {
89 .ctl_name = CTL_UNNUMBERED,
90 .procname = "rdma_slot_table_entries",
91 .data = &xprt_rdma_slot_table_entries,
92 .maxlen = sizeof(unsigned int),
94 .proc_handler = &proc_dointvec_minmax,
95 .strategy = &sysctl_intvec,
96 .extra1 = &min_slot_table_size,
97 .extra2 = &max_slot_table_size
100 .ctl_name = CTL_UNNUMBERED,
101 .procname = "rdma_max_inline_read",
102 .data = &xprt_rdma_max_inline_read,
103 .maxlen = sizeof(unsigned int),
105 .proc_handler = &proc_dointvec,
106 .strategy = &sysctl_intvec,
109 .ctl_name = CTL_UNNUMBERED,
110 .procname = "rdma_max_inline_write",
111 .data = &xprt_rdma_max_inline_write,
112 .maxlen = sizeof(unsigned int),
114 .proc_handler = &proc_dointvec,
115 .strategy = &sysctl_intvec,
118 .ctl_name = CTL_UNNUMBERED,
119 .procname = "rdma_inline_write_padding",
120 .data = &xprt_rdma_inline_write_padding,
121 .maxlen = sizeof(unsigned int),
123 .proc_handler = &proc_dointvec_minmax,
124 .strategy = &sysctl_intvec,
126 .extra2 = &max_padding,
129 .ctl_name = CTL_UNNUMBERED,
130 .procname = "rdma_memreg_strategy",
131 .data = &xprt_rdma_memreg_strategy,
132 .maxlen = sizeof(unsigned int),
134 .proc_handler = &proc_dointvec_minmax,
135 .strategy = &sysctl_intvec,
136 .extra1 = &min_memreg,
137 .extra2 = &max_memreg,
140 .ctl_name = CTL_UNNUMBERED,
141 .procname = "rdma_pad_optimize",
142 .data = &xprt_rdma_pad_optimize,
143 .maxlen = sizeof(unsigned int),
145 .proc_handler = &proc_dointvec,
152 static ctl_table sunrpc_table[] = {
154 .ctl_name = CTL_SUNRPC,
155 .procname = "sunrpc",
157 .child = xr_tunables_table
166 static struct rpc_xprt_ops xprt_rdma_procs; /* forward reference */
169 xprt_rdma_format_addresses(struct rpc_xprt *xprt)
171 struct sockaddr *sap = (struct sockaddr *)
172 &rpcx_to_rdmad(xprt).addr;
173 struct sockaddr_in *sin = (struct sockaddr_in *)sap;
176 (void)rpc_ntop(sap, buf, sizeof(buf));
177 xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL);
179 (void)snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap));
180 xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
182 xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
184 (void)snprintf(buf, sizeof(buf), "addr=%s port=%s proto=rdma",
185 xprt->address_strings[RPC_DISPLAY_ADDR],
186 xprt->address_strings[RPC_DISPLAY_PORT]);
187 xprt->address_strings[RPC_DISPLAY_ALL] = kstrdup(buf, GFP_KERNEL);
189 (void)snprintf(buf, sizeof(buf), "%02x%02x%02x%02x",
190 NIPQUAD(sin->sin_addr.s_addr));
191 xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
193 (void)snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap));
194 xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
197 xprt->address_strings[RPC_DISPLAY_NETID] = "rdma";
201 xprt_rdma_free_addresses(struct rpc_xprt *xprt)
205 for (i = 0; i < RPC_DISPLAY_MAX; i++)
207 case RPC_DISPLAY_PROTO:
208 case RPC_DISPLAY_NETID:
211 kfree(xprt->address_strings[i]);
216 xprt_rdma_connect_worker(struct work_struct *work)
218 struct rpcrdma_xprt *r_xprt =
219 container_of(work, struct rpcrdma_xprt, rdma_connect.work);
220 struct rpc_xprt *xprt = &r_xprt->xprt;
223 if (!xprt->shutdown) {
224 xprt_clear_connected(xprt);
226 dprintk("RPC: %s: %sconnect\n", __func__,
227 r_xprt->rx_ep.rep_connected != 0 ? "re" : "");
228 rc = rpcrdma_ep_connect(&r_xprt->rx_ep, &r_xprt->rx_ia);
235 xprt_wake_pending_tasks(xprt, rc);
238 dprintk("RPC: %s: exit\n", __func__);
239 xprt_clear_connecting(xprt);
246 * Free all memory associated with the object, including its own.
247 * NOTE: none of the *destroy methods free memory for their top-level
248 * objects, even though they may have allocated it (they do free
249 * private memory). It's up to the caller to handle it. In this
250 * case (RDMA transport), all structure memory is inlined with the
251 * struct rpcrdma_xprt.
254 xprt_rdma_destroy(struct rpc_xprt *xprt)
256 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
259 dprintk("RPC: %s: called\n", __func__);
261 cancel_delayed_work(&r_xprt->rdma_connect);
262 flush_scheduled_work();
264 xprt_clear_connected(xprt);
266 rpcrdma_buffer_destroy(&r_xprt->rx_buf);
267 rc = rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia);
269 dprintk("RPC: %s: rpcrdma_ep_destroy returned %i\n",
271 rpcrdma_ia_close(&r_xprt->rx_ia);
273 xprt_rdma_free_addresses(xprt);
279 dprintk("RPC: %s: returning\n", __func__);
281 module_put(THIS_MODULE);
284 static const struct rpc_timeout xprt_rdma_default_timeout = {
285 .to_initval = 60 * HZ,
286 .to_maxval = 60 * HZ,
290 * xprt_setup_rdma - Set up transport to use RDMA
292 * @args: rpc transport arguments
294 static struct rpc_xprt *
295 xprt_setup_rdma(struct xprt_create *args)
297 struct rpcrdma_create_data_internal cdata;
298 struct rpc_xprt *xprt;
299 struct rpcrdma_xprt *new_xprt;
300 struct rpcrdma_ep *new_ep;
301 struct sockaddr_in *sin;
304 if (args->addrlen > sizeof(xprt->addr)) {
305 dprintk("RPC: %s: address too large\n", __func__);
306 return ERR_PTR(-EBADF);
309 xprt = kzalloc(sizeof(struct rpcrdma_xprt), GFP_KERNEL);
311 dprintk("RPC: %s: couldn't allocate rpcrdma_xprt\n",
313 return ERR_PTR(-ENOMEM);
316 xprt->max_reqs = xprt_rdma_slot_table_entries;
317 xprt->slot = kcalloc(xprt->max_reqs,
318 sizeof(struct rpc_rqst), GFP_KERNEL);
319 if (xprt->slot == NULL) {
320 dprintk("RPC: %s: couldn't allocate %d slots\n",
321 __func__, xprt->max_reqs);
323 return ERR_PTR(-ENOMEM);
326 /* 60 second timeout, no retries */
327 xprt->timeout = &xprt_rdma_default_timeout;
328 xprt->bind_timeout = (60U * HZ);
329 xprt->connect_timeout = (60U * HZ);
330 xprt->reestablish_timeout = (5U * HZ);
331 xprt->idle_timeout = (5U * 60 * HZ);
333 xprt->resvport = 0; /* privileged port not needed */
334 xprt->tsh_size = 0; /* RPC-RDMA handles framing */
335 xprt->max_payload = RPCRDMA_MAX_DATA_SEGS * PAGE_SIZE;
336 xprt->ops = &xprt_rdma_procs;
339 * Set up RDMA-specific connect data.
342 /* Put server RDMA address in local cdata */
343 memcpy(&cdata.addr, args->dstaddr, args->addrlen);
345 /* Ensure xprt->addr holds valid server TCP (not RDMA)
346 * address, for any side protocols which peek at it */
347 xprt->prot = IPPROTO_TCP;
348 xprt->addrlen = args->addrlen;
349 memcpy(&xprt->addr, &cdata.addr, xprt->addrlen);
351 sin = (struct sockaddr_in *)&cdata.addr;
352 if (ntohs(sin->sin_port) != 0)
353 xprt_set_bound(xprt);
355 dprintk("RPC: %s: %pI4:%u\n",
356 __func__, &sin->sin_addr.s_addr, ntohs(sin->sin_port));
358 /* Set max requests */
359 cdata.max_requests = xprt->max_reqs;
361 /* Set some length limits */
362 cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */
363 cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */
365 cdata.inline_wsize = xprt_rdma_max_inline_write;
366 if (cdata.inline_wsize > cdata.wsize)
367 cdata.inline_wsize = cdata.wsize;
369 cdata.inline_rsize = xprt_rdma_max_inline_read;
370 if (cdata.inline_rsize > cdata.rsize)
371 cdata.inline_rsize = cdata.rsize;
373 cdata.padding = xprt_rdma_inline_write_padding;
376 * Create new transport instance, which includes initialized
382 new_xprt = rpcx_to_rdmax(xprt);
384 rc = rpcrdma_ia_open(new_xprt, (struct sockaddr *) &cdata.addr,
385 xprt_rdma_memreg_strategy);
390 * initialize and create ep
392 new_xprt->rx_data = cdata;
393 new_ep = &new_xprt->rx_ep;
394 new_ep->rep_remote_addr = cdata.addr;
396 rc = rpcrdma_ep_create(&new_xprt->rx_ep,
397 &new_xprt->rx_ia, &new_xprt->rx_data);
402 * Allocate pre-registered send and receive buffers for headers and
403 * any inline data. Also specify any padding which will be provided
404 * from a preregistered zero buffer.
406 rc = rpcrdma_buffer_create(&new_xprt->rx_buf, new_ep, &new_xprt->rx_ia,
412 * Register a callback for connection events. This is necessary because
413 * connection loss notification is async. We also catch connection loss
414 * when reaping receives.
416 INIT_DELAYED_WORK(&new_xprt->rdma_connect, xprt_rdma_connect_worker);
417 new_ep->rep_func = rpcrdma_conn_func;
418 new_ep->rep_xprt = xprt;
420 xprt_rdma_format_addresses(xprt);
422 if (!try_module_get(THIS_MODULE))
428 xprt_rdma_free_addresses(xprt);
431 (void) rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia);
433 rpcrdma_ia_close(&new_xprt->rx_ia);
441 * Close a connection, during shutdown or timeout/reconnect
444 xprt_rdma_close(struct rpc_xprt *xprt)
446 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
448 dprintk("RPC: %s: closing\n", __func__);
449 if (r_xprt->rx_ep.rep_connected > 0)
450 xprt->reestablish_timeout = 0;
451 xprt_disconnect_done(xprt);
452 (void) rpcrdma_ep_disconnect(&r_xprt->rx_ep, &r_xprt->rx_ia);
456 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
458 struct sockaddr_in *sap;
460 sap = (struct sockaddr_in *)&xprt->addr;
461 sap->sin_port = htons(port);
462 sap = (struct sockaddr_in *)&rpcx_to_rdmad(xprt).addr;
463 sap->sin_port = htons(port);
464 dprintk("RPC: %s: %u\n", __func__, port);
468 xprt_rdma_connect(struct rpc_task *task)
470 struct rpc_xprt *xprt = (struct rpc_xprt *)task->tk_xprt;
471 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
473 if (!xprt_test_and_set_connecting(xprt)) {
474 if (r_xprt->rx_ep.rep_connected != 0) {
476 schedule_delayed_work(&r_xprt->rdma_connect,
477 xprt->reestablish_timeout);
478 xprt->reestablish_timeout <<= 1;
479 if (xprt->reestablish_timeout > (30 * HZ))
480 xprt->reestablish_timeout = (30 * HZ);
481 else if (xprt->reestablish_timeout < (5 * HZ))
482 xprt->reestablish_timeout = (5 * HZ);
484 schedule_delayed_work(&r_xprt->rdma_connect, 0);
485 if (!RPC_IS_ASYNC(task))
486 flush_scheduled_work();
492 xprt_rdma_reserve_xprt(struct rpc_task *task)
494 struct rpc_xprt *xprt = task->tk_xprt;
495 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
496 int credits = atomic_read(&r_xprt->rx_buf.rb_credits);
498 /* == RPC_CWNDSCALE @ init, but *after* setup */
499 if (r_xprt->rx_buf.rb_cwndscale == 0UL) {
500 r_xprt->rx_buf.rb_cwndscale = xprt->cwnd;
501 dprintk("RPC: %s: cwndscale %lu\n", __func__,
502 r_xprt->rx_buf.rb_cwndscale);
503 BUG_ON(r_xprt->rx_buf.rb_cwndscale <= 0);
505 xprt->cwnd = credits * r_xprt->rx_buf.rb_cwndscale;
506 return xprt_reserve_xprt_cong(task);
510 * The RDMA allocate/free functions need the task structure as a place
511 * to hide the struct rpcrdma_req, which is necessary for the actual send/recv
512 * sequence. For this reason, the recv buffers are attached to send
513 * buffers for portions of the RPC. Note that the RPC layer allocates
514 * both send and receive buffers in the same call. We may register
515 * the receive buffer portion when using reply chunks.
518 xprt_rdma_allocate(struct rpc_task *task, size_t size)
520 struct rpc_xprt *xprt = task->tk_xprt;
521 struct rpcrdma_req *req, *nreq;
523 req = rpcrdma_buffer_get(&rpcx_to_rdmax(xprt)->rx_buf);
526 if (size > req->rl_size) {
527 dprintk("RPC: %s: size %zd too large for buffer[%zd]: "
528 "prog %d vers %d proc %d\n",
529 __func__, size, req->rl_size,
530 task->tk_client->cl_prog, task->tk_client->cl_vers,
531 task->tk_msg.rpc_proc->p_proc);
533 * Outgoing length shortage. Our inline write max must have
534 * been configured to perform direct i/o.
536 * This is therefore a large metadata operation, and the
537 * allocate call was made on the maximum possible message,
538 * e.g. containing long filename(s) or symlink data. In
539 * fact, while these metadata operations *might* carry
540 * large outgoing payloads, they rarely *do*. However, we
541 * have to commit to the request here, so reallocate and
542 * register it now. The data path will never require this
545 * If the allocation or registration fails, the RPC framework
546 * will (doggedly) retry.
548 if (rpcx_to_rdmax(xprt)->rx_ia.ri_memreg_strategy ==
549 RPCRDMA_BOUNCEBUFFERS) {
550 /* forced to "pure inline" */
551 dprintk("RPC: %s: too much data (%zd) for inline "
552 "(r/w max %d/%d)\n", __func__, size,
553 rpcx_to_rdmad(xprt).inline_rsize,
554 rpcx_to_rdmad(xprt).inline_wsize);
556 rpc_exit(task, -EIO); /* fail the operation */
557 rpcx_to_rdmax(xprt)->rx_stats.failed_marshal_count++;
560 if (task->tk_flags & RPC_TASK_SWAPPER)
561 nreq = kmalloc(sizeof *req + size, GFP_ATOMIC);
563 nreq = kmalloc(sizeof *req + size, GFP_NOFS);
567 if (rpcrdma_register_internal(&rpcx_to_rdmax(xprt)->rx_ia,
568 nreq->rl_base, size + sizeof(struct rpcrdma_req)
569 - offsetof(struct rpcrdma_req, rl_base),
570 &nreq->rl_handle, &nreq->rl_iov)) {
574 rpcx_to_rdmax(xprt)->rx_stats.hardway_register_count += size;
575 nreq->rl_size = size;
577 nreq->rl_nchunks = 0;
578 nreq->rl_buffer = (struct rpcrdma_buffer *)req;
579 nreq->rl_reply = req->rl_reply;
580 memcpy(nreq->rl_segments,
581 req->rl_segments, sizeof nreq->rl_segments);
582 /* flag the swap with an unused field */
583 nreq->rl_iov.length = 0;
584 req->rl_reply = NULL;
587 dprintk("RPC: %s: size %zd, request 0x%p\n", __func__, size, req);
589 req->rl_connect_cookie = 0; /* our reserved value */
590 return req->rl_xdr_buf;
593 rpcrdma_buffer_put(req);
594 rpcx_to_rdmax(xprt)->rx_stats.failed_marshal_count++;
599 * This function returns all RDMA resources to the pool.
602 xprt_rdma_free(void *buffer)
604 struct rpcrdma_req *req;
605 struct rpcrdma_xprt *r_xprt;
606 struct rpcrdma_rep *rep;
612 req = container_of(buffer, struct rpcrdma_req, rl_xdr_buf[0]);
613 if (req->rl_iov.length == 0) { /* see allocate above */
614 r_xprt = container_of(((struct rpcrdma_req *) req->rl_buffer)->rl_buffer,
615 struct rpcrdma_xprt, rx_buf);
617 r_xprt = container_of(req->rl_buffer, struct rpcrdma_xprt, rx_buf);
620 dprintk("RPC: %s: called on 0x%p%s\n",
621 __func__, rep, (rep && rep->rr_func) ? " (with waiter)" : "");
624 * Finish the deregistration. When using mw bind, this was
625 * begun in rpcrdma_reply_handler(). In all other modes, we
626 * do it here, in thread context. The process is considered
627 * complete when the rr_func vector becomes NULL - this
628 * was put in place during rpcrdma_reply_handler() - the wait
629 * call below will not block if the dereg is "done". If
630 * interrupted, our framework will clean up.
632 for (i = 0; req->rl_nchunks;) {
634 i += rpcrdma_deregister_external(
635 &req->rl_segments[i], r_xprt, NULL);
638 if (rep && wait_event_interruptible(rep->rr_unbind, !rep->rr_func)) {
639 rep->rr_func = NULL; /* abandon the callback */
640 req->rl_reply = NULL;
643 if (req->rl_iov.length == 0) { /* see allocate above */
644 struct rpcrdma_req *oreq = (struct rpcrdma_req *)req->rl_buffer;
645 oreq->rl_reply = req->rl_reply;
646 (void) rpcrdma_deregister_internal(&r_xprt->rx_ia,
653 /* Put back request+reply buffers */
654 rpcrdma_buffer_put(req);
658 * send_request invokes the meat of RPC RDMA. It must do the following:
659 * 1. Marshal the RPC request into an RPC RDMA request, which means
660 * putting a header in front of data, and creating IOVs for RDMA
661 * from those in the request.
662 * 2. In marshaling, detect opportunities for RDMA, and use them.
663 * 3. Post a recv message to set up asynch completion, then send
664 * the request (rpcrdma_ep_post).
665 * 4. No partial sends are possible in the RPC-RDMA protocol (as in UDP).
669 xprt_rdma_send_request(struct rpc_task *task)
671 struct rpc_rqst *rqst = task->tk_rqstp;
672 struct rpc_xprt *xprt = task->tk_xprt;
673 struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
674 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
676 /* marshal the send itself */
677 if (req->rl_niovs == 0 && rpcrdma_marshal_req(rqst) != 0) {
678 r_xprt->rx_stats.failed_marshal_count++;
679 dprintk("RPC: %s: rpcrdma_marshal_req failed\n",
684 if (req->rl_reply == NULL) /* e.g. reconnection */
685 rpcrdma_recv_buffer_get(req);
688 req->rl_reply->rr_func = rpcrdma_reply_handler;
689 /* this need only be done once, but... */
690 req->rl_reply->rr_xprt = xprt;
693 /* Must suppress retransmit to maintain credits */
694 if (req->rl_connect_cookie == xprt->connect_cookie)
695 goto drop_connection;
696 req->rl_connect_cookie = xprt->connect_cookie;
698 if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
699 goto drop_connection;
701 task->tk_bytes_sent += rqst->rq_snd_buf.len;
702 rqst->rq_bytes_sent = 0;
706 xprt_disconnect_done(xprt);
707 return -ENOTCONN; /* implies disconnect */
710 static void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
712 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
715 if (xprt_connected(xprt))
716 idle_time = (long)(jiffies - xprt->last_used) / HZ;
719 "\txprt:\trdma %u %lu %lu %lu %ld %lu %lu %lu %Lu %Lu "
720 "%lu %lu %lu %Lu %Lu %Lu %Lu %lu %lu %lu\n",
722 0, /* need a local port? */
723 xprt->stat.bind_count,
724 xprt->stat.connect_count,
725 xprt->stat.connect_time,
733 r_xprt->rx_stats.read_chunk_count,
734 r_xprt->rx_stats.write_chunk_count,
735 r_xprt->rx_stats.reply_chunk_count,
736 r_xprt->rx_stats.total_rdma_request,
737 r_xprt->rx_stats.total_rdma_reply,
738 r_xprt->rx_stats.pullup_copy_count,
739 r_xprt->rx_stats.fixup_copy_count,
740 r_xprt->rx_stats.hardway_register_count,
741 r_xprt->rx_stats.failed_marshal_count,
742 r_xprt->rx_stats.bad_reply_count);
746 * Plumbing for rpc transport switch and kernel module
749 static struct rpc_xprt_ops xprt_rdma_procs = {
750 .reserve_xprt = xprt_rdma_reserve_xprt,
751 .release_xprt = xprt_release_xprt_cong, /* sunrpc/xprt.c */
752 .release_request = xprt_release_rqst_cong, /* ditto */
753 .set_retrans_timeout = xprt_set_retrans_timeout_def, /* ditto */
754 .rpcbind = rpcb_getport_async, /* sunrpc/rpcb_clnt.c */
755 .set_port = xprt_rdma_set_port,
756 .connect = xprt_rdma_connect,
757 .buf_alloc = xprt_rdma_allocate,
758 .buf_free = xprt_rdma_free,
759 .send_request = xprt_rdma_send_request,
760 .close = xprt_rdma_close,
761 .destroy = xprt_rdma_destroy,
762 .print_stats = xprt_rdma_print_stats
765 static struct xprt_class xprt_rdma = {
766 .list = LIST_HEAD_INIT(xprt_rdma.list),
768 .owner = THIS_MODULE,
769 .ident = XPRT_TRANSPORT_RDMA,
770 .setup = xprt_setup_rdma,
773 static void __exit xprt_rdma_cleanup(void)
777 dprintk(KERN_INFO "RPCRDMA Module Removed, deregister RPC RDMA transport\n");
779 if (sunrpc_table_header) {
780 unregister_sysctl_table(sunrpc_table_header);
781 sunrpc_table_header = NULL;
784 rc = xprt_unregister_transport(&xprt_rdma);
786 dprintk("RPC: %s: xprt_unregister returned %i\n",
790 static int __init xprt_rdma_init(void)
794 rc = xprt_register_transport(&xprt_rdma);
799 dprintk(KERN_INFO "RPCRDMA Module Init, register RPC RDMA transport\n");
801 dprintk(KERN_INFO "Defaults:\n");
802 dprintk(KERN_INFO "\tSlots %d\n"
803 "\tMaxInlineRead %d\n\tMaxInlineWrite %d\n",
804 xprt_rdma_slot_table_entries,
805 xprt_rdma_max_inline_read, xprt_rdma_max_inline_write);
806 dprintk(KERN_INFO "\tPadding %d\n\tMemreg %d\n",
807 xprt_rdma_inline_write_padding, xprt_rdma_memreg_strategy);
810 if (!sunrpc_table_header)
811 sunrpc_table_header = register_sysctl_table(sunrpc_table);
816 module_init(xprt_rdma_init);
817 module_exit(xprt_rdma_cleanup);