xref: /openbmc/linux/net/sunrpc/xprtrdma/verbs.c (revision 2359ccdd)
1 /*
2  * Copyright (c) 2014-2017 Oracle.  All rights reserved.
3  * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
4  *
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
9  * license below:
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  *
15  *      Redistributions of source code must retain the above copyright
16  *      notice, this list of conditions and the following disclaimer.
17  *
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.
22  *
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
26  *      permission.
27  *
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.
39  */
40 
41 /*
42  * verbs.c
43  *
44  * Encapsulates the major functions managing:
45  *  o adapters
46  *  o endpoints
47  *  o connections
48  *  o buffer memory
49  */
50 
51 #include <linux/interrupt.h>
52 #include <linux/slab.h>
53 #include <linux/sunrpc/addr.h>
54 #include <linux/sunrpc/svc_rdma.h>
55 
56 #include <asm-generic/barrier.h>
57 #include <asm/bitops.h>
58 
59 #include <rdma/ib_cm.h>
60 
61 #include "xprt_rdma.h"
62 
63 /*
64  * Globals/Macros
65  */
66 
67 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
68 # define RPCDBG_FACILITY	RPCDBG_TRANS
69 #endif
70 
71 /*
72  * internal functions
73  */
74 static void rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt);
75 static void rpcrdma_mrs_destroy(struct rpcrdma_buffer *buf);
76 static void rpcrdma_dma_unmap_regbuf(struct rpcrdma_regbuf *rb);
77 
78 struct workqueue_struct *rpcrdma_receive_wq __read_mostly;
79 
80 int
81 rpcrdma_alloc_wq(void)
82 {
83 	struct workqueue_struct *recv_wq;
84 
85 	recv_wq = alloc_workqueue("xprtrdma_receive",
86 				  WQ_MEM_RECLAIM | WQ_HIGHPRI,
87 				  0);
88 	if (!recv_wq)
89 		return -ENOMEM;
90 
91 	rpcrdma_receive_wq = recv_wq;
92 	return 0;
93 }
94 
95 void
96 rpcrdma_destroy_wq(void)
97 {
98 	struct workqueue_struct *wq;
99 
100 	if (rpcrdma_receive_wq) {
101 		wq = rpcrdma_receive_wq;
102 		rpcrdma_receive_wq = NULL;
103 		destroy_workqueue(wq);
104 	}
105 }
106 
107 static void
108 rpcrdma_qp_async_error_upcall(struct ib_event *event, void *context)
109 {
110 	struct rpcrdma_ep *ep = context;
111 	struct rpcrdma_xprt *r_xprt = container_of(ep, struct rpcrdma_xprt,
112 						   rx_ep);
113 
114 	trace_xprtrdma_qp_error(r_xprt, event);
115 	pr_err("rpcrdma: %s on device %s ep %p\n",
116 	       ib_event_msg(event->event), event->device->name, context);
117 
118 	if (ep->rep_connected == 1) {
119 		ep->rep_connected = -EIO;
120 		rpcrdma_conn_func(ep);
121 		wake_up_all(&ep->rep_connect_wait);
122 	}
123 }
124 
125 /**
126  * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC
127  * @cq:	completion queue (ignored)
128  * @wc:	completed WR
129  *
130  */
131 static void
132 rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
133 {
134 	struct ib_cqe *cqe = wc->wr_cqe;
135 	struct rpcrdma_sendctx *sc =
136 		container_of(cqe, struct rpcrdma_sendctx, sc_cqe);
137 
138 	/* WARNING: Only wr_cqe and status are reliable at this point */
139 	trace_xprtrdma_wc_send(sc, wc);
140 	if (wc->status != IB_WC_SUCCESS && wc->status != IB_WC_WR_FLUSH_ERR)
141 		pr_err("rpcrdma: Send: %s (%u/0x%x)\n",
142 		       ib_wc_status_msg(wc->status),
143 		       wc->status, wc->vendor_err);
144 
145 	rpcrdma_sendctx_put_locked(sc);
146 }
147 
148 /**
149  * rpcrdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
150  * @cq:	completion queue (ignored)
151  * @wc:	completed WR
152  *
153  */
154 static void
155 rpcrdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
156 {
157 	struct ib_cqe *cqe = wc->wr_cqe;
158 	struct rpcrdma_rep *rep = container_of(cqe, struct rpcrdma_rep,
159 					       rr_cqe);
160 
161 	/* WARNING: Only wr_id and status are reliable at this point */
162 	trace_xprtrdma_wc_receive(rep, wc);
163 	if (wc->status != IB_WC_SUCCESS)
164 		goto out_fail;
165 
166 	/* status == SUCCESS means all fields in wc are trustworthy */
167 	rpcrdma_set_xdrlen(&rep->rr_hdrbuf, wc->byte_len);
168 	rep->rr_wc_flags = wc->wc_flags;
169 	rep->rr_inv_rkey = wc->ex.invalidate_rkey;
170 
171 	ib_dma_sync_single_for_cpu(rdmab_device(rep->rr_rdmabuf),
172 				   rdmab_addr(rep->rr_rdmabuf),
173 				   wc->byte_len, DMA_FROM_DEVICE);
174 
175 out_schedule:
176 	rpcrdma_reply_handler(rep);
177 	return;
178 
179 out_fail:
180 	if (wc->status != IB_WC_WR_FLUSH_ERR)
181 		pr_err("rpcrdma: Recv: %s (%u/0x%x)\n",
182 		       ib_wc_status_msg(wc->status),
183 		       wc->status, wc->vendor_err);
184 	rpcrdma_set_xdrlen(&rep->rr_hdrbuf, 0);
185 	goto out_schedule;
186 }
187 
188 static void
189 rpcrdma_update_connect_private(struct rpcrdma_xprt *r_xprt,
190 			       struct rdma_conn_param *param)
191 {
192 	struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
193 	const struct rpcrdma_connect_private *pmsg = param->private_data;
194 	unsigned int rsize, wsize;
195 
196 	/* Default settings for RPC-over-RDMA Version One */
197 	r_xprt->rx_ia.ri_implicit_roundup = xprt_rdma_pad_optimize;
198 	rsize = RPCRDMA_V1_DEF_INLINE_SIZE;
199 	wsize = RPCRDMA_V1_DEF_INLINE_SIZE;
200 
201 	if (pmsg &&
202 	    pmsg->cp_magic == rpcrdma_cmp_magic &&
203 	    pmsg->cp_version == RPCRDMA_CMP_VERSION) {
204 		r_xprt->rx_ia.ri_implicit_roundup = true;
205 		rsize = rpcrdma_decode_buffer_size(pmsg->cp_send_size);
206 		wsize = rpcrdma_decode_buffer_size(pmsg->cp_recv_size);
207 	}
208 
209 	if (rsize < cdata->inline_rsize)
210 		cdata->inline_rsize = rsize;
211 	if (wsize < cdata->inline_wsize)
212 		cdata->inline_wsize = wsize;
213 	dprintk("RPC:       %s: max send %u, max recv %u\n",
214 		__func__, cdata->inline_wsize, cdata->inline_rsize);
215 	rpcrdma_set_max_header_sizes(r_xprt);
216 }
217 
218 static int
219 rpcrdma_conn_upcall(struct rdma_cm_id *id, struct rdma_cm_event *event)
220 {
221 	struct rpcrdma_xprt *xprt = id->context;
222 	struct rpcrdma_ia *ia = &xprt->rx_ia;
223 	struct rpcrdma_ep *ep = &xprt->rx_ep;
224 	int connstate = 0;
225 
226 	trace_xprtrdma_conn_upcall(xprt, event);
227 	switch (event->event) {
228 	case RDMA_CM_EVENT_ADDR_RESOLVED:
229 	case RDMA_CM_EVENT_ROUTE_RESOLVED:
230 		ia->ri_async_rc = 0;
231 		complete(&ia->ri_done);
232 		break;
233 	case RDMA_CM_EVENT_ADDR_ERROR:
234 		ia->ri_async_rc = -EHOSTUNREACH;
235 		complete(&ia->ri_done);
236 		break;
237 	case RDMA_CM_EVENT_ROUTE_ERROR:
238 		ia->ri_async_rc = -ENETUNREACH;
239 		complete(&ia->ri_done);
240 		break;
241 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
242 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
243 		pr_info("rpcrdma: removing device %s for %s:%s\n",
244 			ia->ri_device->name,
245 			rpcrdma_addrstr(xprt), rpcrdma_portstr(xprt));
246 #endif
247 		set_bit(RPCRDMA_IAF_REMOVING, &ia->ri_flags);
248 		ep->rep_connected = -ENODEV;
249 		xprt_force_disconnect(&xprt->rx_xprt);
250 		wait_for_completion(&ia->ri_remove_done);
251 
252 		ia->ri_id = NULL;
253 		ia->ri_device = NULL;
254 		/* Return 1 to ensure the core destroys the id. */
255 		return 1;
256 	case RDMA_CM_EVENT_ESTABLISHED:
257 		++xprt->rx_xprt.connect_cookie;
258 		connstate = 1;
259 		rpcrdma_update_connect_private(xprt, &event->param.conn);
260 		goto connected;
261 	case RDMA_CM_EVENT_CONNECT_ERROR:
262 		connstate = -ENOTCONN;
263 		goto connected;
264 	case RDMA_CM_EVENT_UNREACHABLE:
265 		connstate = -ENETDOWN;
266 		goto connected;
267 	case RDMA_CM_EVENT_REJECTED:
268 		dprintk("rpcrdma: connection to %s:%s rejected: %s\n",
269 			rpcrdma_addrstr(xprt), rpcrdma_portstr(xprt),
270 			rdma_reject_msg(id, event->status));
271 		connstate = -ECONNREFUSED;
272 		if (event->status == IB_CM_REJ_STALE_CONN)
273 			connstate = -EAGAIN;
274 		goto connected;
275 	case RDMA_CM_EVENT_DISCONNECTED:
276 		++xprt->rx_xprt.connect_cookie;
277 		connstate = -ECONNABORTED;
278 connected:
279 		xprt->rx_buf.rb_credits = 1;
280 		ep->rep_connected = connstate;
281 		rpcrdma_conn_func(ep);
282 		wake_up_all(&ep->rep_connect_wait);
283 		/*FALLTHROUGH*/
284 	default:
285 		dprintk("RPC:       %s: %s:%s on %s/%s (ep 0x%p): %s\n",
286 			__func__,
287 			rpcrdma_addrstr(xprt), rpcrdma_portstr(xprt),
288 			ia->ri_device->name, ia->ri_ops->ro_displayname,
289 			ep, rdma_event_msg(event->event));
290 		break;
291 	}
292 
293 	return 0;
294 }
295 
296 static struct rdma_cm_id *
297 rpcrdma_create_id(struct rpcrdma_xprt *xprt, struct rpcrdma_ia *ia)
298 {
299 	unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1;
300 	struct rdma_cm_id *id;
301 	int rc;
302 
303 	trace_xprtrdma_conn_start(xprt);
304 
305 	init_completion(&ia->ri_done);
306 	init_completion(&ia->ri_remove_done);
307 
308 	id = rdma_create_id(&init_net, rpcrdma_conn_upcall, xprt, RDMA_PS_TCP,
309 			    IB_QPT_RC);
310 	if (IS_ERR(id)) {
311 		rc = PTR_ERR(id);
312 		dprintk("RPC:       %s: rdma_create_id() failed %i\n",
313 			__func__, rc);
314 		return id;
315 	}
316 
317 	ia->ri_async_rc = -ETIMEDOUT;
318 	rc = rdma_resolve_addr(id, NULL,
319 			       (struct sockaddr *)&xprt->rx_xprt.addr,
320 			       RDMA_RESOLVE_TIMEOUT);
321 	if (rc) {
322 		dprintk("RPC:       %s: rdma_resolve_addr() failed %i\n",
323 			__func__, rc);
324 		goto out;
325 	}
326 	rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
327 	if (rc < 0) {
328 		trace_xprtrdma_conn_tout(xprt);
329 		goto out;
330 	}
331 
332 	rc = ia->ri_async_rc;
333 	if (rc)
334 		goto out;
335 
336 	ia->ri_async_rc = -ETIMEDOUT;
337 	rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
338 	if (rc) {
339 		dprintk("RPC:       %s: rdma_resolve_route() failed %i\n",
340 			__func__, rc);
341 		goto out;
342 	}
343 	rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
344 	if (rc < 0) {
345 		trace_xprtrdma_conn_tout(xprt);
346 		goto out;
347 	}
348 	rc = ia->ri_async_rc;
349 	if (rc)
350 		goto out;
351 
352 	return id;
353 
354 out:
355 	rdma_destroy_id(id);
356 	return ERR_PTR(rc);
357 }
358 
359 /*
360  * Exported functions.
361  */
362 
363 /**
364  * rpcrdma_ia_open - Open and initialize an Interface Adapter.
365  * @xprt: transport with IA to (re)initialize
366  *
367  * Returns 0 on success, negative errno if an appropriate
368  * Interface Adapter could not be found and opened.
369  */
370 int
371 rpcrdma_ia_open(struct rpcrdma_xprt *xprt)
372 {
373 	struct rpcrdma_ia *ia = &xprt->rx_ia;
374 	int rc;
375 
376 	ia->ri_id = rpcrdma_create_id(xprt, ia);
377 	if (IS_ERR(ia->ri_id)) {
378 		rc = PTR_ERR(ia->ri_id);
379 		goto out_err;
380 	}
381 	ia->ri_device = ia->ri_id->device;
382 
383 	ia->ri_pd = ib_alloc_pd(ia->ri_device, 0);
384 	if (IS_ERR(ia->ri_pd)) {
385 		rc = PTR_ERR(ia->ri_pd);
386 		pr_err("rpcrdma: ib_alloc_pd() returned %d\n", rc);
387 		goto out_err;
388 	}
389 
390 	switch (xprt_rdma_memreg_strategy) {
391 	case RPCRDMA_FRWR:
392 		if (frwr_is_supported(ia)) {
393 			ia->ri_ops = &rpcrdma_frwr_memreg_ops;
394 			break;
395 		}
396 		/*FALLTHROUGH*/
397 	case RPCRDMA_MTHCAFMR:
398 		if (fmr_is_supported(ia)) {
399 			ia->ri_ops = &rpcrdma_fmr_memreg_ops;
400 			break;
401 		}
402 		/*FALLTHROUGH*/
403 	default:
404 		pr_err("rpcrdma: Device %s does not support memreg mode %d\n",
405 		       ia->ri_device->name, xprt_rdma_memreg_strategy);
406 		rc = -EINVAL;
407 		goto out_err;
408 	}
409 
410 	return 0;
411 
412 out_err:
413 	rpcrdma_ia_close(ia);
414 	return rc;
415 }
416 
417 /**
418  * rpcrdma_ia_remove - Handle device driver unload
419  * @ia: interface adapter being removed
420  *
421  * Divest transport H/W resources associated with this adapter,
422  * but allow it to be restored later.
423  */
424 void
425 rpcrdma_ia_remove(struct rpcrdma_ia *ia)
426 {
427 	struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
428 						   rx_ia);
429 	struct rpcrdma_ep *ep = &r_xprt->rx_ep;
430 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
431 	struct rpcrdma_req *req;
432 	struct rpcrdma_rep *rep;
433 
434 	cancel_delayed_work_sync(&buf->rb_refresh_worker);
435 
436 	/* This is similar to rpcrdma_ep_destroy, but:
437 	 * - Don't cancel the connect worker.
438 	 * - Don't call rpcrdma_ep_disconnect, which waits
439 	 *   for another conn upcall, which will deadlock.
440 	 * - rdma_disconnect is unneeded, the underlying
441 	 *   connection is already gone.
442 	 */
443 	if (ia->ri_id->qp) {
444 		ib_drain_qp(ia->ri_id->qp);
445 		rdma_destroy_qp(ia->ri_id);
446 		ia->ri_id->qp = NULL;
447 	}
448 	ib_free_cq(ep->rep_attr.recv_cq);
449 	ep->rep_attr.recv_cq = NULL;
450 	ib_free_cq(ep->rep_attr.send_cq);
451 	ep->rep_attr.send_cq = NULL;
452 
453 	/* The ULP is responsible for ensuring all DMA
454 	 * mappings and MRs are gone.
455 	 */
456 	list_for_each_entry(rep, &buf->rb_recv_bufs, rr_list)
457 		rpcrdma_dma_unmap_regbuf(rep->rr_rdmabuf);
458 	list_for_each_entry(req, &buf->rb_allreqs, rl_all) {
459 		rpcrdma_dma_unmap_regbuf(req->rl_rdmabuf);
460 		rpcrdma_dma_unmap_regbuf(req->rl_sendbuf);
461 		rpcrdma_dma_unmap_regbuf(req->rl_recvbuf);
462 	}
463 	rpcrdma_mrs_destroy(buf);
464 	ib_dealloc_pd(ia->ri_pd);
465 	ia->ri_pd = NULL;
466 
467 	/* Allow waiters to continue */
468 	complete(&ia->ri_remove_done);
469 
470 	trace_xprtrdma_remove(r_xprt);
471 }
472 
473 /**
474  * rpcrdma_ia_close - Clean up/close an IA.
475  * @ia: interface adapter to close
476  *
477  */
478 void
479 rpcrdma_ia_close(struct rpcrdma_ia *ia)
480 {
481 	if (ia->ri_id != NULL && !IS_ERR(ia->ri_id)) {
482 		if (ia->ri_id->qp)
483 			rdma_destroy_qp(ia->ri_id);
484 		rdma_destroy_id(ia->ri_id);
485 	}
486 	ia->ri_id = NULL;
487 	ia->ri_device = NULL;
488 
489 	/* If the pd is still busy, xprtrdma missed freeing a resource */
490 	if (ia->ri_pd && !IS_ERR(ia->ri_pd))
491 		ib_dealloc_pd(ia->ri_pd);
492 	ia->ri_pd = NULL;
493 }
494 
495 /*
496  * Create unconnected endpoint.
497  */
498 int
499 rpcrdma_ep_create(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia,
500 		  struct rpcrdma_create_data_internal *cdata)
501 {
502 	struct rpcrdma_connect_private *pmsg = &ep->rep_cm_private;
503 	unsigned int max_qp_wr, max_sge;
504 	struct ib_cq *sendcq, *recvcq;
505 	int rc;
506 
507 	max_sge = min_t(unsigned int, ia->ri_device->attrs.max_sge,
508 			RPCRDMA_MAX_SEND_SGES);
509 	if (max_sge < RPCRDMA_MIN_SEND_SGES) {
510 		pr_warn("rpcrdma: HCA provides only %d send SGEs\n", max_sge);
511 		return -ENOMEM;
512 	}
513 	ia->ri_max_send_sges = max_sge;
514 
515 	if (ia->ri_device->attrs.max_qp_wr <= RPCRDMA_BACKWARD_WRS) {
516 		dprintk("RPC:       %s: insufficient wqe's available\n",
517 			__func__);
518 		return -ENOMEM;
519 	}
520 	max_qp_wr = ia->ri_device->attrs.max_qp_wr - RPCRDMA_BACKWARD_WRS - 1;
521 
522 	/* check provider's send/recv wr limits */
523 	if (cdata->max_requests > max_qp_wr)
524 		cdata->max_requests = max_qp_wr;
525 
526 	ep->rep_attr.event_handler = rpcrdma_qp_async_error_upcall;
527 	ep->rep_attr.qp_context = ep;
528 	ep->rep_attr.srq = NULL;
529 	ep->rep_attr.cap.max_send_wr = cdata->max_requests;
530 	ep->rep_attr.cap.max_send_wr += RPCRDMA_BACKWARD_WRS;
531 	ep->rep_attr.cap.max_send_wr += 1;	/* drain cqe */
532 	rc = ia->ri_ops->ro_open(ia, ep, cdata);
533 	if (rc)
534 		return rc;
535 	ep->rep_attr.cap.max_recv_wr = cdata->max_requests;
536 	ep->rep_attr.cap.max_recv_wr += RPCRDMA_BACKWARD_WRS;
537 	ep->rep_attr.cap.max_recv_wr += 1;	/* drain cqe */
538 	ep->rep_attr.cap.max_send_sge = max_sge;
539 	ep->rep_attr.cap.max_recv_sge = 1;
540 	ep->rep_attr.cap.max_inline_data = 0;
541 	ep->rep_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
542 	ep->rep_attr.qp_type = IB_QPT_RC;
543 	ep->rep_attr.port_num = ~0;
544 
545 	dprintk("RPC:       %s: requested max: dtos: send %d recv %d; "
546 		"iovs: send %d recv %d\n",
547 		__func__,
548 		ep->rep_attr.cap.max_send_wr,
549 		ep->rep_attr.cap.max_recv_wr,
550 		ep->rep_attr.cap.max_send_sge,
551 		ep->rep_attr.cap.max_recv_sge);
552 
553 	/* set trigger for requesting send completion */
554 	ep->rep_send_batch = min_t(unsigned int, RPCRDMA_MAX_SEND_BATCH,
555 				   cdata->max_requests >> 2);
556 	ep->rep_send_count = ep->rep_send_batch;
557 	init_waitqueue_head(&ep->rep_connect_wait);
558 	INIT_DELAYED_WORK(&ep->rep_connect_worker, rpcrdma_connect_worker);
559 
560 	sendcq = ib_alloc_cq(ia->ri_device, NULL,
561 			     ep->rep_attr.cap.max_send_wr + 1,
562 			     1, IB_POLL_WORKQUEUE);
563 	if (IS_ERR(sendcq)) {
564 		rc = PTR_ERR(sendcq);
565 		dprintk("RPC:       %s: failed to create send CQ: %i\n",
566 			__func__, rc);
567 		goto out1;
568 	}
569 
570 	recvcq = ib_alloc_cq(ia->ri_device, NULL,
571 			     ep->rep_attr.cap.max_recv_wr + 1,
572 			     0, IB_POLL_WORKQUEUE);
573 	if (IS_ERR(recvcq)) {
574 		rc = PTR_ERR(recvcq);
575 		dprintk("RPC:       %s: failed to create recv CQ: %i\n",
576 			__func__, rc);
577 		goto out2;
578 	}
579 
580 	ep->rep_attr.send_cq = sendcq;
581 	ep->rep_attr.recv_cq = recvcq;
582 
583 	/* Initialize cma parameters */
584 	memset(&ep->rep_remote_cma, 0, sizeof(ep->rep_remote_cma));
585 
586 	/* Prepare RDMA-CM private message */
587 	pmsg->cp_magic = rpcrdma_cmp_magic;
588 	pmsg->cp_version = RPCRDMA_CMP_VERSION;
589 	pmsg->cp_flags |= ia->ri_ops->ro_send_w_inv_ok;
590 	pmsg->cp_send_size = rpcrdma_encode_buffer_size(cdata->inline_wsize);
591 	pmsg->cp_recv_size = rpcrdma_encode_buffer_size(cdata->inline_rsize);
592 	ep->rep_remote_cma.private_data = pmsg;
593 	ep->rep_remote_cma.private_data_len = sizeof(*pmsg);
594 
595 	/* Client offers RDMA Read but does not initiate */
596 	ep->rep_remote_cma.initiator_depth = 0;
597 	ep->rep_remote_cma.responder_resources =
598 		min_t(int, U8_MAX, ia->ri_device->attrs.max_qp_rd_atom);
599 
600 	/* Limit transport retries so client can detect server
601 	 * GID changes quickly. RPC layer handles re-establishing
602 	 * transport connection and retransmission.
603 	 */
604 	ep->rep_remote_cma.retry_count = 6;
605 
606 	/* RPC-over-RDMA handles its own flow control. In addition,
607 	 * make all RNR NAKs visible so we know that RPC-over-RDMA
608 	 * flow control is working correctly (no NAKs should be seen).
609 	 */
610 	ep->rep_remote_cma.flow_control = 0;
611 	ep->rep_remote_cma.rnr_retry_count = 0;
612 
613 	return 0;
614 
615 out2:
616 	ib_free_cq(sendcq);
617 out1:
618 	return rc;
619 }
620 
621 /*
622  * rpcrdma_ep_destroy
623  *
624  * Disconnect and destroy endpoint. After this, the only
625  * valid operations on the ep are to free it (if dynamically
626  * allocated) or re-create it.
627  */
628 void
629 rpcrdma_ep_destroy(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
630 {
631 	cancel_delayed_work_sync(&ep->rep_connect_worker);
632 
633 	if (ia->ri_id && ia->ri_id->qp) {
634 		rpcrdma_ep_disconnect(ep, ia);
635 		rdma_destroy_qp(ia->ri_id);
636 		ia->ri_id->qp = NULL;
637 	}
638 
639 	if (ep->rep_attr.recv_cq)
640 		ib_free_cq(ep->rep_attr.recv_cq);
641 	if (ep->rep_attr.send_cq)
642 		ib_free_cq(ep->rep_attr.send_cq);
643 }
644 
645 /* Re-establish a connection after a device removal event.
646  * Unlike a normal reconnection, a fresh PD and a new set
647  * of MRs and buffers is needed.
648  */
649 static int
650 rpcrdma_ep_recreate_xprt(struct rpcrdma_xprt *r_xprt,
651 			 struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
652 {
653 	int rc, err;
654 
655 	trace_xprtrdma_reinsert(r_xprt);
656 
657 	rc = -EHOSTUNREACH;
658 	if (rpcrdma_ia_open(r_xprt))
659 		goto out1;
660 
661 	rc = -ENOMEM;
662 	err = rpcrdma_ep_create(ep, ia, &r_xprt->rx_data);
663 	if (err) {
664 		pr_err("rpcrdma: rpcrdma_ep_create returned %d\n", err);
665 		goto out2;
666 	}
667 
668 	rc = -ENETUNREACH;
669 	err = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
670 	if (err) {
671 		pr_err("rpcrdma: rdma_create_qp returned %d\n", err);
672 		goto out3;
673 	}
674 
675 	rpcrdma_mrs_create(r_xprt);
676 	return 0;
677 
678 out3:
679 	rpcrdma_ep_destroy(ep, ia);
680 out2:
681 	rpcrdma_ia_close(ia);
682 out1:
683 	return rc;
684 }
685 
686 static int
687 rpcrdma_ep_reconnect(struct rpcrdma_xprt *r_xprt, struct rpcrdma_ep *ep,
688 		     struct rpcrdma_ia *ia)
689 {
690 	struct rdma_cm_id *id, *old;
691 	int err, rc;
692 
693 	trace_xprtrdma_reconnect(r_xprt);
694 
695 	rpcrdma_ep_disconnect(ep, ia);
696 
697 	rc = -EHOSTUNREACH;
698 	id = rpcrdma_create_id(r_xprt, ia);
699 	if (IS_ERR(id))
700 		goto out;
701 
702 	/* As long as the new ID points to the same device as the
703 	 * old ID, we can reuse the transport's existing PD and all
704 	 * previously allocated MRs. Also, the same device means
705 	 * the transport's previous DMA mappings are still valid.
706 	 *
707 	 * This is a sanity check only. There should be no way these
708 	 * point to two different devices here.
709 	 */
710 	old = id;
711 	rc = -ENETUNREACH;
712 	if (ia->ri_device != id->device) {
713 		pr_err("rpcrdma: can't reconnect on different device!\n");
714 		goto out_destroy;
715 	}
716 
717 	err = rdma_create_qp(id, ia->ri_pd, &ep->rep_attr);
718 	if (err) {
719 		dprintk("RPC:       %s: rdma_create_qp returned %d\n",
720 			__func__, err);
721 		goto out_destroy;
722 	}
723 
724 	/* Atomically replace the transport's ID and QP. */
725 	rc = 0;
726 	old = ia->ri_id;
727 	ia->ri_id = id;
728 	rdma_destroy_qp(old);
729 
730 out_destroy:
731 	rdma_destroy_id(old);
732 out:
733 	return rc;
734 }
735 
736 /*
737  * Connect unconnected endpoint.
738  */
739 int
740 rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
741 {
742 	struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
743 						   rx_ia);
744 	unsigned int extras;
745 	int rc;
746 
747 retry:
748 	switch (ep->rep_connected) {
749 	case 0:
750 		dprintk("RPC:       %s: connecting...\n", __func__);
751 		rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
752 		if (rc) {
753 			dprintk("RPC:       %s: rdma_create_qp failed %i\n",
754 				__func__, rc);
755 			rc = -ENETUNREACH;
756 			goto out_noupdate;
757 		}
758 		break;
759 	case -ENODEV:
760 		rc = rpcrdma_ep_recreate_xprt(r_xprt, ep, ia);
761 		if (rc)
762 			goto out_noupdate;
763 		break;
764 	default:
765 		rc = rpcrdma_ep_reconnect(r_xprt, ep, ia);
766 		if (rc)
767 			goto out;
768 	}
769 
770 	ep->rep_connected = 0;
771 
772 	rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma);
773 	if (rc) {
774 		dprintk("RPC:       %s: rdma_connect() failed with %i\n",
775 				__func__, rc);
776 		goto out;
777 	}
778 
779 	wait_event_interruptible(ep->rep_connect_wait, ep->rep_connected != 0);
780 	if (ep->rep_connected <= 0) {
781 		if (ep->rep_connected == -EAGAIN)
782 			goto retry;
783 		rc = ep->rep_connected;
784 		goto out;
785 	}
786 
787 	dprintk("RPC:       %s: connected\n", __func__);
788 	extras = r_xprt->rx_buf.rb_bc_srv_max_requests;
789 	if (extras)
790 		rpcrdma_ep_post_extra_recv(r_xprt, extras);
791 
792 out:
793 	if (rc)
794 		ep->rep_connected = rc;
795 
796 out_noupdate:
797 	return rc;
798 }
799 
800 /*
801  * rpcrdma_ep_disconnect
802  *
803  * This is separate from destroy to facilitate the ability
804  * to reconnect without recreating the endpoint.
805  *
806  * This call is not reentrant, and must not be made in parallel
807  * on the same endpoint.
808  */
809 void
810 rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
811 {
812 	int rc;
813 
814 	rc = rdma_disconnect(ia->ri_id);
815 	if (!rc)
816 		/* returns without wait if not connected */
817 		wait_event_interruptible(ep->rep_connect_wait,
818 							ep->rep_connected != 1);
819 	else
820 		ep->rep_connected = rc;
821 	trace_xprtrdma_disconnect(container_of(ep, struct rpcrdma_xprt,
822 					       rx_ep), rc);
823 
824 	ib_drain_qp(ia->ri_id->qp);
825 }
826 
827 /* Fixed-size circular FIFO queue. This implementation is wait-free and
828  * lock-free.
829  *
830  * Consumer is the code path that posts Sends. This path dequeues a
831  * sendctx for use by a Send operation. Multiple consumer threads
832  * are serialized by the RPC transport lock, which allows only one
833  * ->send_request call at a time.
834  *
835  * Producer is the code path that handles Send completions. This path
836  * enqueues a sendctx that has been completed. Multiple producer
837  * threads are serialized by the ib_poll_cq() function.
838  */
839 
840 /* rpcrdma_sendctxs_destroy() assumes caller has already quiesced
841  * queue activity, and ib_drain_qp has flushed all remaining Send
842  * requests.
843  */
844 static void rpcrdma_sendctxs_destroy(struct rpcrdma_buffer *buf)
845 {
846 	unsigned long i;
847 
848 	for (i = 0; i <= buf->rb_sc_last; i++)
849 		kfree(buf->rb_sc_ctxs[i]);
850 	kfree(buf->rb_sc_ctxs);
851 }
852 
853 static struct rpcrdma_sendctx *rpcrdma_sendctx_create(struct rpcrdma_ia *ia)
854 {
855 	struct rpcrdma_sendctx *sc;
856 
857 	sc = kzalloc(sizeof(*sc) +
858 		     ia->ri_max_send_sges * sizeof(struct ib_sge),
859 		     GFP_KERNEL);
860 	if (!sc)
861 		return NULL;
862 
863 	sc->sc_wr.wr_cqe = &sc->sc_cqe;
864 	sc->sc_wr.sg_list = sc->sc_sges;
865 	sc->sc_wr.opcode = IB_WR_SEND;
866 	sc->sc_cqe.done = rpcrdma_wc_send;
867 	return sc;
868 }
869 
870 static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt)
871 {
872 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
873 	struct rpcrdma_sendctx *sc;
874 	unsigned long i;
875 
876 	/* Maximum number of concurrent outstanding Send WRs. Capping
877 	 * the circular queue size stops Send Queue overflow by causing
878 	 * the ->send_request call to fail temporarily before too many
879 	 * Sends are posted.
880 	 */
881 	i = buf->rb_max_requests + RPCRDMA_MAX_BC_REQUESTS;
882 	dprintk("RPC:       %s: allocating %lu send_ctxs\n", __func__, i);
883 	buf->rb_sc_ctxs = kcalloc(i, sizeof(sc), GFP_KERNEL);
884 	if (!buf->rb_sc_ctxs)
885 		return -ENOMEM;
886 
887 	buf->rb_sc_last = i - 1;
888 	for (i = 0; i <= buf->rb_sc_last; i++) {
889 		sc = rpcrdma_sendctx_create(&r_xprt->rx_ia);
890 		if (!sc)
891 			goto out_destroy;
892 
893 		sc->sc_xprt = r_xprt;
894 		buf->rb_sc_ctxs[i] = sc;
895 	}
896 
897 	return 0;
898 
899 out_destroy:
900 	rpcrdma_sendctxs_destroy(buf);
901 	return -ENOMEM;
902 }
903 
904 /* The sendctx queue is not guaranteed to have a size that is a
905  * power of two, thus the helpers in circ_buf.h cannot be used.
906  * The other option is to use modulus (%), which can be expensive.
907  */
908 static unsigned long rpcrdma_sendctx_next(struct rpcrdma_buffer *buf,
909 					  unsigned long item)
910 {
911 	return likely(item < buf->rb_sc_last) ? item + 1 : 0;
912 }
913 
914 /**
915  * rpcrdma_sendctx_get_locked - Acquire a send context
916  * @buf: transport buffers from which to acquire an unused context
917  *
918  * Returns pointer to a free send completion context; or NULL if
919  * the queue is empty.
920  *
921  * Usage: Called to acquire an SGE array before preparing a Send WR.
922  *
923  * The caller serializes calls to this function (per rpcrdma_buffer),
924  * and provides an effective memory barrier that flushes the new value
925  * of rb_sc_head.
926  */
927 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_buffer *buf)
928 {
929 	struct rpcrdma_xprt *r_xprt;
930 	struct rpcrdma_sendctx *sc;
931 	unsigned long next_head;
932 
933 	next_head = rpcrdma_sendctx_next(buf, buf->rb_sc_head);
934 
935 	if (next_head == READ_ONCE(buf->rb_sc_tail))
936 		goto out_emptyq;
937 
938 	/* ORDER: item must be accessed _before_ head is updated */
939 	sc = buf->rb_sc_ctxs[next_head];
940 
941 	/* Releasing the lock in the caller acts as a memory
942 	 * barrier that flushes rb_sc_head.
943 	 */
944 	buf->rb_sc_head = next_head;
945 
946 	return sc;
947 
948 out_emptyq:
949 	/* The queue is "empty" if there have not been enough Send
950 	 * completions recently. This is a sign the Send Queue is
951 	 * backing up. Cause the caller to pause and try again.
952 	 */
953 	dprintk("RPC:       %s: empty sendctx queue\n", __func__);
954 	r_xprt = container_of(buf, struct rpcrdma_xprt, rx_buf);
955 	r_xprt->rx_stats.empty_sendctx_q++;
956 	return NULL;
957 }
958 
959 /**
960  * rpcrdma_sendctx_put_locked - Release a send context
961  * @sc: send context to release
962  *
963  * Usage: Called from Send completion to return a sendctxt
964  * to the queue.
965  *
966  * The caller serializes calls to this function (per rpcrdma_buffer).
967  */
968 void rpcrdma_sendctx_put_locked(struct rpcrdma_sendctx *sc)
969 {
970 	struct rpcrdma_buffer *buf = &sc->sc_xprt->rx_buf;
971 	unsigned long next_tail;
972 
973 	/* Unmap SGEs of previously completed by unsignaled
974 	 * Sends by walking up the queue until @sc is found.
975 	 */
976 	next_tail = buf->rb_sc_tail;
977 	do {
978 		next_tail = rpcrdma_sendctx_next(buf, next_tail);
979 
980 		/* ORDER: item must be accessed _before_ tail is updated */
981 		rpcrdma_unmap_sendctx(buf->rb_sc_ctxs[next_tail]);
982 
983 	} while (buf->rb_sc_ctxs[next_tail] != sc);
984 
985 	/* Paired with READ_ONCE */
986 	smp_store_release(&buf->rb_sc_tail, next_tail);
987 }
988 
989 static void
990 rpcrdma_mr_recovery_worker(struct work_struct *work)
991 {
992 	struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
993 						  rb_recovery_worker.work);
994 	struct rpcrdma_mr *mr;
995 
996 	spin_lock(&buf->rb_recovery_lock);
997 	while (!list_empty(&buf->rb_stale_mrs)) {
998 		mr = rpcrdma_mr_pop(&buf->rb_stale_mrs);
999 		spin_unlock(&buf->rb_recovery_lock);
1000 
1001 		trace_xprtrdma_recover_mr(mr);
1002 		mr->mr_xprt->rx_ia.ri_ops->ro_recover_mr(mr);
1003 
1004 		spin_lock(&buf->rb_recovery_lock);
1005 	}
1006 	spin_unlock(&buf->rb_recovery_lock);
1007 }
1008 
1009 void
1010 rpcrdma_mr_defer_recovery(struct rpcrdma_mr *mr)
1011 {
1012 	struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
1013 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1014 
1015 	spin_lock(&buf->rb_recovery_lock);
1016 	rpcrdma_mr_push(mr, &buf->rb_stale_mrs);
1017 	spin_unlock(&buf->rb_recovery_lock);
1018 
1019 	schedule_delayed_work(&buf->rb_recovery_worker, 0);
1020 }
1021 
1022 static void
1023 rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt)
1024 {
1025 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1026 	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
1027 	unsigned int count;
1028 	LIST_HEAD(free);
1029 	LIST_HEAD(all);
1030 
1031 	for (count = 0; count < 3; count++) {
1032 		struct rpcrdma_mr *mr;
1033 		int rc;
1034 
1035 		mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1036 		if (!mr)
1037 			break;
1038 
1039 		rc = ia->ri_ops->ro_init_mr(ia, mr);
1040 		if (rc) {
1041 			kfree(mr);
1042 			break;
1043 		}
1044 
1045 		mr->mr_xprt = r_xprt;
1046 
1047 		list_add(&mr->mr_list, &free);
1048 		list_add(&mr->mr_all, &all);
1049 	}
1050 
1051 	spin_lock(&buf->rb_mrlock);
1052 	list_splice(&free, &buf->rb_mrs);
1053 	list_splice(&all, &buf->rb_all);
1054 	r_xprt->rx_stats.mrs_allocated += count;
1055 	spin_unlock(&buf->rb_mrlock);
1056 	trace_xprtrdma_createmrs(r_xprt, count);
1057 
1058 	xprt_write_space(&r_xprt->rx_xprt);
1059 }
1060 
1061 static void
1062 rpcrdma_mr_refresh_worker(struct work_struct *work)
1063 {
1064 	struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
1065 						  rb_refresh_worker.work);
1066 	struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
1067 						   rx_buf);
1068 
1069 	rpcrdma_mrs_create(r_xprt);
1070 }
1071 
1072 struct rpcrdma_req *
1073 rpcrdma_create_req(struct rpcrdma_xprt *r_xprt)
1074 {
1075 	struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
1076 	struct rpcrdma_regbuf *rb;
1077 	struct rpcrdma_req *req;
1078 
1079 	req = kzalloc(sizeof(*req), GFP_KERNEL);
1080 	if (req == NULL)
1081 		return ERR_PTR(-ENOMEM);
1082 
1083 	rb = rpcrdma_alloc_regbuf(RPCRDMA_HDRBUF_SIZE,
1084 				  DMA_TO_DEVICE, GFP_KERNEL);
1085 	if (IS_ERR(rb)) {
1086 		kfree(req);
1087 		return ERR_PTR(-ENOMEM);
1088 	}
1089 	req->rl_rdmabuf = rb;
1090 	xdr_buf_init(&req->rl_hdrbuf, rb->rg_base, rdmab_length(rb));
1091 	req->rl_buffer = buffer;
1092 	INIT_LIST_HEAD(&req->rl_registered);
1093 
1094 	spin_lock(&buffer->rb_reqslock);
1095 	list_add(&req->rl_all, &buffer->rb_allreqs);
1096 	spin_unlock(&buffer->rb_reqslock);
1097 	return req;
1098 }
1099 
1100 /**
1101  * rpcrdma_create_rep - Allocate an rpcrdma_rep object
1102  * @r_xprt: controlling transport
1103  *
1104  * Returns 0 on success or a negative errno on failure.
1105  */
1106 int
1107 rpcrdma_create_rep(struct rpcrdma_xprt *r_xprt)
1108 {
1109 	struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
1110 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1111 	struct rpcrdma_rep *rep;
1112 	int rc;
1113 
1114 	rc = -ENOMEM;
1115 	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1116 	if (rep == NULL)
1117 		goto out;
1118 
1119 	rep->rr_rdmabuf = rpcrdma_alloc_regbuf(cdata->inline_rsize,
1120 					       DMA_FROM_DEVICE, GFP_KERNEL);
1121 	if (IS_ERR(rep->rr_rdmabuf)) {
1122 		rc = PTR_ERR(rep->rr_rdmabuf);
1123 		goto out_free;
1124 	}
1125 	xdr_buf_init(&rep->rr_hdrbuf, rep->rr_rdmabuf->rg_base,
1126 		     rdmab_length(rep->rr_rdmabuf));
1127 
1128 	rep->rr_cqe.done = rpcrdma_wc_receive;
1129 	rep->rr_rxprt = r_xprt;
1130 	INIT_WORK(&rep->rr_work, rpcrdma_deferred_completion);
1131 	rep->rr_recv_wr.next = NULL;
1132 	rep->rr_recv_wr.wr_cqe = &rep->rr_cqe;
1133 	rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
1134 	rep->rr_recv_wr.num_sge = 1;
1135 
1136 	spin_lock(&buf->rb_lock);
1137 	list_add(&rep->rr_list, &buf->rb_recv_bufs);
1138 	spin_unlock(&buf->rb_lock);
1139 	return 0;
1140 
1141 out_free:
1142 	kfree(rep);
1143 out:
1144 	dprintk("RPC:       %s: reply buffer %d alloc failed\n",
1145 		__func__, rc);
1146 	return rc;
1147 }
1148 
1149 int
1150 rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
1151 {
1152 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1153 	int i, rc;
1154 
1155 	buf->rb_max_requests = r_xprt->rx_data.max_requests;
1156 	buf->rb_bc_srv_max_requests = 0;
1157 	spin_lock_init(&buf->rb_mrlock);
1158 	spin_lock_init(&buf->rb_lock);
1159 	spin_lock_init(&buf->rb_recovery_lock);
1160 	INIT_LIST_HEAD(&buf->rb_mrs);
1161 	INIT_LIST_HEAD(&buf->rb_all);
1162 	INIT_LIST_HEAD(&buf->rb_stale_mrs);
1163 	INIT_DELAYED_WORK(&buf->rb_refresh_worker,
1164 			  rpcrdma_mr_refresh_worker);
1165 	INIT_DELAYED_WORK(&buf->rb_recovery_worker,
1166 			  rpcrdma_mr_recovery_worker);
1167 
1168 	rpcrdma_mrs_create(r_xprt);
1169 
1170 	INIT_LIST_HEAD(&buf->rb_send_bufs);
1171 	INIT_LIST_HEAD(&buf->rb_allreqs);
1172 	spin_lock_init(&buf->rb_reqslock);
1173 	for (i = 0; i < buf->rb_max_requests; i++) {
1174 		struct rpcrdma_req *req;
1175 
1176 		req = rpcrdma_create_req(r_xprt);
1177 		if (IS_ERR(req)) {
1178 			dprintk("RPC:       %s: request buffer %d alloc"
1179 				" failed\n", __func__, i);
1180 			rc = PTR_ERR(req);
1181 			goto out;
1182 		}
1183 		list_add(&req->rl_list, &buf->rb_send_bufs);
1184 	}
1185 
1186 	INIT_LIST_HEAD(&buf->rb_recv_bufs);
1187 	for (i = 0; i <= buf->rb_max_requests; i++) {
1188 		rc = rpcrdma_create_rep(r_xprt);
1189 		if (rc)
1190 			goto out;
1191 	}
1192 
1193 	rc = rpcrdma_sendctxs_create(r_xprt);
1194 	if (rc)
1195 		goto out;
1196 
1197 	return 0;
1198 out:
1199 	rpcrdma_buffer_destroy(buf);
1200 	return rc;
1201 }
1202 
1203 static struct rpcrdma_req *
1204 rpcrdma_buffer_get_req_locked(struct rpcrdma_buffer *buf)
1205 {
1206 	struct rpcrdma_req *req;
1207 
1208 	req = list_first_entry(&buf->rb_send_bufs,
1209 			       struct rpcrdma_req, rl_list);
1210 	list_del_init(&req->rl_list);
1211 	return req;
1212 }
1213 
1214 static struct rpcrdma_rep *
1215 rpcrdma_buffer_get_rep_locked(struct rpcrdma_buffer *buf)
1216 {
1217 	struct rpcrdma_rep *rep;
1218 
1219 	rep = list_first_entry(&buf->rb_recv_bufs,
1220 			       struct rpcrdma_rep, rr_list);
1221 	list_del(&rep->rr_list);
1222 	return rep;
1223 }
1224 
1225 static void
1226 rpcrdma_destroy_rep(struct rpcrdma_rep *rep)
1227 {
1228 	rpcrdma_free_regbuf(rep->rr_rdmabuf);
1229 	kfree(rep);
1230 }
1231 
1232 void
1233 rpcrdma_destroy_req(struct rpcrdma_req *req)
1234 {
1235 	rpcrdma_free_regbuf(req->rl_recvbuf);
1236 	rpcrdma_free_regbuf(req->rl_sendbuf);
1237 	rpcrdma_free_regbuf(req->rl_rdmabuf);
1238 	kfree(req);
1239 }
1240 
1241 static void
1242 rpcrdma_mrs_destroy(struct rpcrdma_buffer *buf)
1243 {
1244 	struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
1245 						   rx_buf);
1246 	struct rpcrdma_ia *ia = rdmab_to_ia(buf);
1247 	struct rpcrdma_mr *mr;
1248 	unsigned int count;
1249 
1250 	count = 0;
1251 	spin_lock(&buf->rb_mrlock);
1252 	while (!list_empty(&buf->rb_all)) {
1253 		mr = list_entry(buf->rb_all.next, struct rpcrdma_mr, mr_all);
1254 		list_del(&mr->mr_all);
1255 
1256 		spin_unlock(&buf->rb_mrlock);
1257 		ia->ri_ops->ro_release_mr(mr);
1258 		count++;
1259 		spin_lock(&buf->rb_mrlock);
1260 	}
1261 	spin_unlock(&buf->rb_mrlock);
1262 	r_xprt->rx_stats.mrs_allocated = 0;
1263 
1264 	dprintk("RPC:       %s: released %u MRs\n", __func__, count);
1265 }
1266 
1267 void
1268 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1269 {
1270 	cancel_delayed_work_sync(&buf->rb_recovery_worker);
1271 	cancel_delayed_work_sync(&buf->rb_refresh_worker);
1272 
1273 	rpcrdma_sendctxs_destroy(buf);
1274 
1275 	while (!list_empty(&buf->rb_recv_bufs)) {
1276 		struct rpcrdma_rep *rep;
1277 
1278 		rep = rpcrdma_buffer_get_rep_locked(buf);
1279 		rpcrdma_destroy_rep(rep);
1280 	}
1281 	buf->rb_send_count = 0;
1282 
1283 	spin_lock(&buf->rb_reqslock);
1284 	while (!list_empty(&buf->rb_allreqs)) {
1285 		struct rpcrdma_req *req;
1286 
1287 		req = list_first_entry(&buf->rb_allreqs,
1288 				       struct rpcrdma_req, rl_all);
1289 		list_del(&req->rl_all);
1290 
1291 		spin_unlock(&buf->rb_reqslock);
1292 		rpcrdma_destroy_req(req);
1293 		spin_lock(&buf->rb_reqslock);
1294 	}
1295 	spin_unlock(&buf->rb_reqslock);
1296 	buf->rb_recv_count = 0;
1297 
1298 	rpcrdma_mrs_destroy(buf);
1299 }
1300 
1301 /**
1302  * rpcrdma_mr_get - Allocate an rpcrdma_mr object
1303  * @r_xprt: controlling transport
1304  *
1305  * Returns an initialized rpcrdma_mr or NULL if no free
1306  * rpcrdma_mr objects are available.
1307  */
1308 struct rpcrdma_mr *
1309 rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt)
1310 {
1311 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1312 	struct rpcrdma_mr *mr = NULL;
1313 
1314 	spin_lock(&buf->rb_mrlock);
1315 	if (!list_empty(&buf->rb_mrs))
1316 		mr = rpcrdma_mr_pop(&buf->rb_mrs);
1317 	spin_unlock(&buf->rb_mrlock);
1318 
1319 	if (!mr)
1320 		goto out_nomrs;
1321 	return mr;
1322 
1323 out_nomrs:
1324 	trace_xprtrdma_nomrs(r_xprt);
1325 	if (r_xprt->rx_ep.rep_connected != -ENODEV)
1326 		schedule_delayed_work(&buf->rb_refresh_worker, 0);
1327 
1328 	/* Allow the reply handler and refresh worker to run */
1329 	cond_resched();
1330 
1331 	return NULL;
1332 }
1333 
1334 static void
1335 __rpcrdma_mr_put(struct rpcrdma_buffer *buf, struct rpcrdma_mr *mr)
1336 {
1337 	spin_lock(&buf->rb_mrlock);
1338 	rpcrdma_mr_push(mr, &buf->rb_mrs);
1339 	spin_unlock(&buf->rb_mrlock);
1340 }
1341 
1342 /**
1343  * rpcrdma_mr_put - Release an rpcrdma_mr object
1344  * @mr: object to release
1345  *
1346  */
1347 void
1348 rpcrdma_mr_put(struct rpcrdma_mr *mr)
1349 {
1350 	__rpcrdma_mr_put(&mr->mr_xprt->rx_buf, mr);
1351 }
1352 
1353 /**
1354  * rpcrdma_mr_unmap_and_put - DMA unmap an MR and release it
1355  * @mr: object to release
1356  *
1357  */
1358 void
1359 rpcrdma_mr_unmap_and_put(struct rpcrdma_mr *mr)
1360 {
1361 	struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
1362 
1363 	trace_xprtrdma_dma_unmap(mr);
1364 	ib_dma_unmap_sg(r_xprt->rx_ia.ri_device,
1365 			mr->mr_sg, mr->mr_nents, mr->mr_dir);
1366 	__rpcrdma_mr_put(&r_xprt->rx_buf, mr);
1367 }
1368 
1369 static struct rpcrdma_rep *
1370 rpcrdma_buffer_get_rep(struct rpcrdma_buffer *buffers)
1371 {
1372 	/* If an RPC previously completed without a reply (say, a
1373 	 * credential problem or a soft timeout occurs) then hold off
1374 	 * on supplying more Receive buffers until the number of new
1375 	 * pending RPCs catches up to the number of posted Receives.
1376 	 */
1377 	if (unlikely(buffers->rb_send_count < buffers->rb_recv_count))
1378 		return NULL;
1379 
1380 	if (unlikely(list_empty(&buffers->rb_recv_bufs)))
1381 		return NULL;
1382 	buffers->rb_recv_count++;
1383 	return rpcrdma_buffer_get_rep_locked(buffers);
1384 }
1385 
1386 /*
1387  * Get a set of request/reply buffers.
1388  *
1389  * Reply buffer (if available) is attached to send buffer upon return.
1390  */
1391 struct rpcrdma_req *
1392 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1393 {
1394 	struct rpcrdma_req *req;
1395 
1396 	spin_lock(&buffers->rb_lock);
1397 	if (list_empty(&buffers->rb_send_bufs))
1398 		goto out_reqbuf;
1399 	buffers->rb_send_count++;
1400 	req = rpcrdma_buffer_get_req_locked(buffers);
1401 	req->rl_reply = rpcrdma_buffer_get_rep(buffers);
1402 	spin_unlock(&buffers->rb_lock);
1403 
1404 	return req;
1405 
1406 out_reqbuf:
1407 	spin_unlock(&buffers->rb_lock);
1408 	return NULL;
1409 }
1410 
1411 /*
1412  * Put request/reply buffers back into pool.
1413  * Pre-decrement counter/array index.
1414  */
1415 void
1416 rpcrdma_buffer_put(struct rpcrdma_req *req)
1417 {
1418 	struct rpcrdma_buffer *buffers = req->rl_buffer;
1419 	struct rpcrdma_rep *rep = req->rl_reply;
1420 
1421 	req->rl_reply = NULL;
1422 
1423 	spin_lock(&buffers->rb_lock);
1424 	buffers->rb_send_count--;
1425 	list_add_tail(&req->rl_list, &buffers->rb_send_bufs);
1426 	if (rep) {
1427 		buffers->rb_recv_count--;
1428 		list_add_tail(&rep->rr_list, &buffers->rb_recv_bufs);
1429 	}
1430 	spin_unlock(&buffers->rb_lock);
1431 }
1432 
1433 /*
1434  * Recover reply buffers from pool.
1435  * This happens when recovering from disconnect.
1436  */
1437 void
1438 rpcrdma_recv_buffer_get(struct rpcrdma_req *req)
1439 {
1440 	struct rpcrdma_buffer *buffers = req->rl_buffer;
1441 
1442 	spin_lock(&buffers->rb_lock);
1443 	req->rl_reply = rpcrdma_buffer_get_rep(buffers);
1444 	spin_unlock(&buffers->rb_lock);
1445 }
1446 
1447 /*
1448  * Put reply buffers back into pool when not attached to
1449  * request. This happens in error conditions.
1450  */
1451 void
1452 rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1453 {
1454 	struct rpcrdma_buffer *buffers = &rep->rr_rxprt->rx_buf;
1455 
1456 	spin_lock(&buffers->rb_lock);
1457 	buffers->rb_recv_count--;
1458 	list_add_tail(&rep->rr_list, &buffers->rb_recv_bufs);
1459 	spin_unlock(&buffers->rb_lock);
1460 }
1461 
1462 /**
1463  * rpcrdma_alloc_regbuf - allocate and DMA-map memory for SEND/RECV buffers
1464  * @size: size of buffer to be allocated, in bytes
1465  * @direction: direction of data movement
1466  * @flags: GFP flags
1467  *
1468  * Returns an ERR_PTR, or a pointer to a regbuf, a buffer that
1469  * can be persistently DMA-mapped for I/O.
1470  *
1471  * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1472  * receiving the payload of RDMA RECV operations. During Long Calls
1473  * or Replies they may be registered externally via ro_map.
1474  */
1475 struct rpcrdma_regbuf *
1476 rpcrdma_alloc_regbuf(size_t size, enum dma_data_direction direction,
1477 		     gfp_t flags)
1478 {
1479 	struct rpcrdma_regbuf *rb;
1480 
1481 	rb = kmalloc(sizeof(*rb) + size, flags);
1482 	if (rb == NULL)
1483 		return ERR_PTR(-ENOMEM);
1484 
1485 	rb->rg_device = NULL;
1486 	rb->rg_direction = direction;
1487 	rb->rg_iov.length = size;
1488 
1489 	return rb;
1490 }
1491 
1492 /**
1493  * __rpcrdma_map_regbuf - DMA-map a regbuf
1494  * @ia: controlling rpcrdma_ia
1495  * @rb: regbuf to be mapped
1496  */
1497 bool
1498 __rpcrdma_dma_map_regbuf(struct rpcrdma_ia *ia, struct rpcrdma_regbuf *rb)
1499 {
1500 	struct ib_device *device = ia->ri_device;
1501 
1502 	if (rb->rg_direction == DMA_NONE)
1503 		return false;
1504 
1505 	rb->rg_iov.addr = ib_dma_map_single(device,
1506 					    (void *)rb->rg_base,
1507 					    rdmab_length(rb),
1508 					    rb->rg_direction);
1509 	if (ib_dma_mapping_error(device, rdmab_addr(rb)))
1510 		return false;
1511 
1512 	rb->rg_device = device;
1513 	rb->rg_iov.lkey = ia->ri_pd->local_dma_lkey;
1514 	return true;
1515 }
1516 
1517 static void
1518 rpcrdma_dma_unmap_regbuf(struct rpcrdma_regbuf *rb)
1519 {
1520 	if (!rb)
1521 		return;
1522 
1523 	if (!rpcrdma_regbuf_is_mapped(rb))
1524 		return;
1525 
1526 	ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb),
1527 			    rdmab_length(rb), rb->rg_direction);
1528 	rb->rg_device = NULL;
1529 }
1530 
1531 /**
1532  * rpcrdma_free_regbuf - deregister and free registered buffer
1533  * @rb: regbuf to be deregistered and freed
1534  */
1535 void
1536 rpcrdma_free_regbuf(struct rpcrdma_regbuf *rb)
1537 {
1538 	rpcrdma_dma_unmap_regbuf(rb);
1539 	kfree(rb);
1540 }
1541 
1542 /*
1543  * Prepost any receive buffer, then post send.
1544  *
1545  * Receive buffer is donated to hardware, reclaimed upon recv completion.
1546  */
1547 int
1548 rpcrdma_ep_post(struct rpcrdma_ia *ia,
1549 		struct rpcrdma_ep *ep,
1550 		struct rpcrdma_req *req)
1551 {
1552 	struct ib_send_wr *send_wr = &req->rl_sendctx->sc_wr;
1553 	int rc;
1554 
1555 	if (req->rl_reply) {
1556 		rc = rpcrdma_ep_post_recv(ia, req->rl_reply);
1557 		if (rc)
1558 			return rc;
1559 		req->rl_reply = NULL;
1560 	}
1561 
1562 	if (!ep->rep_send_count ||
1563 	    test_bit(RPCRDMA_REQ_F_TX_RESOURCES, &req->rl_flags)) {
1564 		send_wr->send_flags |= IB_SEND_SIGNALED;
1565 		ep->rep_send_count = ep->rep_send_batch;
1566 	} else {
1567 		send_wr->send_flags &= ~IB_SEND_SIGNALED;
1568 		--ep->rep_send_count;
1569 	}
1570 
1571 	rc = ia->ri_ops->ro_send(ia, req);
1572 	trace_xprtrdma_post_send(req, rc);
1573 	if (rc)
1574 		return -ENOTCONN;
1575 	return 0;
1576 }
1577 
1578 int
1579 rpcrdma_ep_post_recv(struct rpcrdma_ia *ia,
1580 		     struct rpcrdma_rep *rep)
1581 {
1582 	struct ib_recv_wr *recv_wr_fail;
1583 	int rc;
1584 
1585 	if (!rpcrdma_dma_map_regbuf(ia, rep->rr_rdmabuf))
1586 		goto out_map;
1587 	rc = ib_post_recv(ia->ri_id->qp, &rep->rr_recv_wr, &recv_wr_fail);
1588 	trace_xprtrdma_post_recv(rep, rc);
1589 	if (rc)
1590 		return -ENOTCONN;
1591 	return 0;
1592 
1593 out_map:
1594 	pr_err("rpcrdma: failed to DMA map the Receive buffer\n");
1595 	return -EIO;
1596 }
1597 
1598 /**
1599  * rpcrdma_ep_post_extra_recv - Post buffers for incoming backchannel requests
1600  * @r_xprt: transport associated with these backchannel resources
1601  * @count: minimum number of incoming requests expected
1602  *
1603  * Returns zero if all requested buffers were posted, or a negative errno.
1604  */
1605 int
1606 rpcrdma_ep_post_extra_recv(struct rpcrdma_xprt *r_xprt, unsigned int count)
1607 {
1608 	struct rpcrdma_buffer *buffers = &r_xprt->rx_buf;
1609 	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
1610 	struct rpcrdma_rep *rep;
1611 	int rc;
1612 
1613 	while (count--) {
1614 		spin_lock(&buffers->rb_lock);
1615 		if (list_empty(&buffers->rb_recv_bufs))
1616 			goto out_reqbuf;
1617 		rep = rpcrdma_buffer_get_rep_locked(buffers);
1618 		spin_unlock(&buffers->rb_lock);
1619 
1620 		rc = rpcrdma_ep_post_recv(ia, rep);
1621 		if (rc)
1622 			goto out_rc;
1623 	}
1624 
1625 	return 0;
1626 
1627 out_reqbuf:
1628 	spin_unlock(&buffers->rb_lock);
1629 	trace_xprtrdma_noreps(r_xprt);
1630 	return -ENOMEM;
1631 
1632 out_rc:
1633 	rpcrdma_recv_buffer_put(rep);
1634 	return rc;
1635 }
1636