xref: /openbmc/linux/net/sunrpc/xprtrdma/verbs.c (revision 8730046c)
1 /*
2  * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
3  *
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
8  * license below:
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  *
14  *      Redistributions of source code must retain the above copyright
15  *      notice, this list of conditions and the following disclaimer.
16  *
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.
21  *
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
25  *      permission.
26  *
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.
38  */
39 
40 /*
41  * verbs.c
42  *
43  * Encapsulates the major functions managing:
44  *  o adapters
45  *  o endpoints
46  *  o connections
47  *  o buffer memory
48  */
49 
50 #include <linux/interrupt.h>
51 #include <linux/slab.h>
52 #include <linux/prefetch.h>
53 #include <linux/sunrpc/addr.h>
54 #include <linux/sunrpc/svc_rdma.h>
55 #include <asm/bitops.h>
56 #include <linux/module.h> /* try_module_get()/module_put() */
57 
58 #include "xprt_rdma.h"
59 
60 /*
61  * Globals/Macros
62  */
63 
64 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
65 # define RPCDBG_FACILITY	RPCDBG_TRANS
66 #endif
67 
68 /*
69  * internal functions
70  */
71 
72 static struct workqueue_struct *rpcrdma_receive_wq;
73 
74 int
75 rpcrdma_alloc_wq(void)
76 {
77 	struct workqueue_struct *recv_wq;
78 
79 	recv_wq = alloc_workqueue("xprtrdma_receive",
80 				  WQ_MEM_RECLAIM | WQ_UNBOUND | WQ_HIGHPRI,
81 				  0);
82 	if (!recv_wq)
83 		return -ENOMEM;
84 
85 	rpcrdma_receive_wq = recv_wq;
86 	return 0;
87 }
88 
89 void
90 rpcrdma_destroy_wq(void)
91 {
92 	struct workqueue_struct *wq;
93 
94 	if (rpcrdma_receive_wq) {
95 		wq = rpcrdma_receive_wq;
96 		rpcrdma_receive_wq = NULL;
97 		destroy_workqueue(wq);
98 	}
99 }
100 
101 static void
102 rpcrdma_qp_async_error_upcall(struct ib_event *event, void *context)
103 {
104 	struct rpcrdma_ep *ep = context;
105 
106 	pr_err("rpcrdma: %s on device %s ep %p\n",
107 	       ib_event_msg(event->event), event->device->name, context);
108 
109 	if (ep->rep_connected == 1) {
110 		ep->rep_connected = -EIO;
111 		rpcrdma_conn_func(ep);
112 		wake_up_all(&ep->rep_connect_wait);
113 	}
114 }
115 
116 /**
117  * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC
118  * @cq:	completion queue (ignored)
119  * @wc:	completed WR
120  *
121  */
122 static void
123 rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
124 {
125 	/* WARNING: Only wr_cqe and status are reliable at this point */
126 	if (wc->status != IB_WC_SUCCESS && wc->status != IB_WC_WR_FLUSH_ERR)
127 		pr_err("rpcrdma: Send: %s (%u/0x%x)\n",
128 		       ib_wc_status_msg(wc->status),
129 		       wc->status, wc->vendor_err);
130 }
131 
132 /* Perform basic sanity checking to avoid using garbage
133  * to update the credit grant value.
134  */
135 static void
136 rpcrdma_update_granted_credits(struct rpcrdma_rep *rep)
137 {
138 	struct rpcrdma_msg *rmsgp = rdmab_to_msg(rep->rr_rdmabuf);
139 	struct rpcrdma_buffer *buffer = &rep->rr_rxprt->rx_buf;
140 	u32 credits;
141 
142 	if (rep->rr_len < RPCRDMA_HDRLEN_ERR)
143 		return;
144 
145 	credits = be32_to_cpu(rmsgp->rm_credit);
146 	if (credits == 0)
147 		credits = 1;	/* don't deadlock */
148 	else if (credits > buffer->rb_max_requests)
149 		credits = buffer->rb_max_requests;
150 
151 	atomic_set(&buffer->rb_credits, credits);
152 }
153 
154 /**
155  * rpcrdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
156  * @cq:	completion queue (ignored)
157  * @wc:	completed WR
158  *
159  */
160 static void
161 rpcrdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
162 {
163 	struct ib_cqe *cqe = wc->wr_cqe;
164 	struct rpcrdma_rep *rep = container_of(cqe, struct rpcrdma_rep,
165 					       rr_cqe);
166 
167 	/* WARNING: Only wr_id and status are reliable at this point */
168 	if (wc->status != IB_WC_SUCCESS)
169 		goto out_fail;
170 
171 	/* status == SUCCESS means all fields in wc are trustworthy */
172 	if (wc->opcode != IB_WC_RECV)
173 		return;
174 
175 	dprintk("RPC:       %s: rep %p opcode 'recv', length %u: success\n",
176 		__func__, rep, wc->byte_len);
177 
178 	rep->rr_len = wc->byte_len;
179 	rep->rr_wc_flags = wc->wc_flags;
180 	rep->rr_inv_rkey = wc->ex.invalidate_rkey;
181 
182 	ib_dma_sync_single_for_cpu(rep->rr_device,
183 				   rdmab_addr(rep->rr_rdmabuf),
184 				   rep->rr_len, DMA_FROM_DEVICE);
185 
186 	rpcrdma_update_granted_credits(rep);
187 
188 out_schedule:
189 	queue_work(rpcrdma_receive_wq, &rep->rr_work);
190 	return;
191 
192 out_fail:
193 	if (wc->status != IB_WC_WR_FLUSH_ERR)
194 		pr_err("rpcrdma: Recv: %s (%u/0x%x)\n",
195 		       ib_wc_status_msg(wc->status),
196 		       wc->status, wc->vendor_err);
197 	rep->rr_len = RPCRDMA_BAD_LEN;
198 	goto out_schedule;
199 }
200 
201 static void
202 rpcrdma_update_connect_private(struct rpcrdma_xprt *r_xprt,
203 			       struct rdma_conn_param *param)
204 {
205 	struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
206 	const struct rpcrdma_connect_private *pmsg = param->private_data;
207 	unsigned int rsize, wsize;
208 
209 	/* Default settings for RPC-over-RDMA Version One */
210 	r_xprt->rx_ia.ri_reminv_expected = false;
211 	rsize = RPCRDMA_V1_DEF_INLINE_SIZE;
212 	wsize = RPCRDMA_V1_DEF_INLINE_SIZE;
213 
214 	if (pmsg &&
215 	    pmsg->cp_magic == rpcrdma_cmp_magic &&
216 	    pmsg->cp_version == RPCRDMA_CMP_VERSION) {
217 		r_xprt->rx_ia.ri_reminv_expected = true;
218 		rsize = rpcrdma_decode_buffer_size(pmsg->cp_send_size);
219 		wsize = rpcrdma_decode_buffer_size(pmsg->cp_recv_size);
220 	}
221 
222 	if (rsize < cdata->inline_rsize)
223 		cdata->inline_rsize = rsize;
224 	if (wsize < cdata->inline_wsize)
225 		cdata->inline_wsize = wsize;
226 	dprintk("RPC:       %s: max send %u, max recv %u\n",
227 		__func__, cdata->inline_wsize, cdata->inline_rsize);
228 	rpcrdma_set_max_header_sizes(r_xprt);
229 }
230 
231 static int
232 rpcrdma_conn_upcall(struct rdma_cm_id *id, struct rdma_cm_event *event)
233 {
234 	struct rpcrdma_xprt *xprt = id->context;
235 	struct rpcrdma_ia *ia = &xprt->rx_ia;
236 	struct rpcrdma_ep *ep = &xprt->rx_ep;
237 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
238 	struct sockaddr *sap = (struct sockaddr *)&ep->rep_remote_addr;
239 #endif
240 	struct ib_qp_attr *attr = &ia->ri_qp_attr;
241 	struct ib_qp_init_attr *iattr = &ia->ri_qp_init_attr;
242 	int connstate = 0;
243 
244 	switch (event->event) {
245 	case RDMA_CM_EVENT_ADDR_RESOLVED:
246 	case RDMA_CM_EVENT_ROUTE_RESOLVED:
247 		ia->ri_async_rc = 0;
248 		complete(&ia->ri_done);
249 		break;
250 	case RDMA_CM_EVENT_ADDR_ERROR:
251 		ia->ri_async_rc = -EHOSTUNREACH;
252 		dprintk("RPC:       %s: CM address resolution error, ep 0x%p\n",
253 			__func__, ep);
254 		complete(&ia->ri_done);
255 		break;
256 	case RDMA_CM_EVENT_ROUTE_ERROR:
257 		ia->ri_async_rc = -ENETUNREACH;
258 		dprintk("RPC:       %s: CM route resolution error, ep 0x%p\n",
259 			__func__, ep);
260 		complete(&ia->ri_done);
261 		break;
262 	case RDMA_CM_EVENT_ESTABLISHED:
263 		connstate = 1;
264 		ib_query_qp(ia->ri_id->qp, attr,
265 			    IB_QP_MAX_QP_RD_ATOMIC | IB_QP_MAX_DEST_RD_ATOMIC,
266 			    iattr);
267 		dprintk("RPC:       %s: %d responder resources"
268 			" (%d initiator)\n",
269 			__func__, attr->max_dest_rd_atomic,
270 			attr->max_rd_atomic);
271 		rpcrdma_update_connect_private(xprt, &event->param.conn);
272 		goto connected;
273 	case RDMA_CM_EVENT_CONNECT_ERROR:
274 		connstate = -ENOTCONN;
275 		goto connected;
276 	case RDMA_CM_EVENT_UNREACHABLE:
277 		connstate = -ENETDOWN;
278 		goto connected;
279 	case RDMA_CM_EVENT_REJECTED:
280 		connstate = -ECONNREFUSED;
281 		goto connected;
282 	case RDMA_CM_EVENT_DISCONNECTED:
283 		connstate = -ECONNABORTED;
284 		goto connected;
285 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
286 		connstate = -ENODEV;
287 connected:
288 		dprintk("RPC:       %s: %sconnected\n",
289 					__func__, connstate > 0 ? "" : "dis");
290 		atomic_set(&xprt->rx_buf.rb_credits, 1);
291 		ep->rep_connected = connstate;
292 		rpcrdma_conn_func(ep);
293 		wake_up_all(&ep->rep_connect_wait);
294 		/*FALLTHROUGH*/
295 	default:
296 		dprintk("RPC:       %s: %pIS:%u (ep 0x%p): %s\n",
297 			__func__, sap, rpc_get_port(sap), ep,
298 			rdma_event_msg(event->event));
299 		break;
300 	}
301 
302 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
303 	if (connstate == 1) {
304 		int ird = attr->max_dest_rd_atomic;
305 		int tird = ep->rep_remote_cma.responder_resources;
306 
307 		pr_info("rpcrdma: connection to %pIS:%u on %s, memreg '%s', %d credits, %d responders%s\n",
308 			sap, rpc_get_port(sap),
309 			ia->ri_device->name,
310 			ia->ri_ops->ro_displayname,
311 			xprt->rx_buf.rb_max_requests,
312 			ird, ird < 4 && ird < tird / 2 ? " (low!)" : "");
313 	} else if (connstate < 0) {
314 		pr_info("rpcrdma: connection to %pIS:%u closed (%d)\n",
315 			sap, rpc_get_port(sap), connstate);
316 	}
317 #endif
318 
319 	return 0;
320 }
321 
322 static void rpcrdma_destroy_id(struct rdma_cm_id *id)
323 {
324 	if (id) {
325 		module_put(id->device->owner);
326 		rdma_destroy_id(id);
327 	}
328 }
329 
330 static struct rdma_cm_id *
331 rpcrdma_create_id(struct rpcrdma_xprt *xprt,
332 			struct rpcrdma_ia *ia, struct sockaddr *addr)
333 {
334 	unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1;
335 	struct rdma_cm_id *id;
336 	int rc;
337 
338 	init_completion(&ia->ri_done);
339 
340 	id = rdma_create_id(&init_net, rpcrdma_conn_upcall, xprt, RDMA_PS_TCP,
341 			    IB_QPT_RC);
342 	if (IS_ERR(id)) {
343 		rc = PTR_ERR(id);
344 		dprintk("RPC:       %s: rdma_create_id() failed %i\n",
345 			__func__, rc);
346 		return id;
347 	}
348 
349 	ia->ri_async_rc = -ETIMEDOUT;
350 	rc = rdma_resolve_addr(id, NULL, addr, RDMA_RESOLVE_TIMEOUT);
351 	if (rc) {
352 		dprintk("RPC:       %s: rdma_resolve_addr() failed %i\n",
353 			__func__, rc);
354 		goto out;
355 	}
356 	rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
357 	if (rc < 0) {
358 		dprintk("RPC:       %s: wait() exited: %i\n",
359 			__func__, rc);
360 		goto out;
361 	}
362 
363 	/* FIXME:
364 	 * Until xprtrdma supports DEVICE_REMOVAL, the provider must
365 	 * be pinned while there are active NFS/RDMA mounts to prevent
366 	 * hangs and crashes at umount time.
367 	 */
368 	if (!ia->ri_async_rc && !try_module_get(id->device->owner)) {
369 		dprintk("RPC:       %s: Failed to get device module\n",
370 			__func__);
371 		ia->ri_async_rc = -ENODEV;
372 	}
373 	rc = ia->ri_async_rc;
374 	if (rc)
375 		goto out;
376 
377 	ia->ri_async_rc = -ETIMEDOUT;
378 	rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
379 	if (rc) {
380 		dprintk("RPC:       %s: rdma_resolve_route() failed %i\n",
381 			__func__, rc);
382 		goto put;
383 	}
384 	rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
385 	if (rc < 0) {
386 		dprintk("RPC:       %s: wait() exited: %i\n",
387 			__func__, rc);
388 		goto put;
389 	}
390 	rc = ia->ri_async_rc;
391 	if (rc)
392 		goto put;
393 
394 	return id;
395 put:
396 	module_put(id->device->owner);
397 out:
398 	rdma_destroy_id(id);
399 	return ERR_PTR(rc);
400 }
401 
402 /*
403  * Exported functions.
404  */
405 
406 /*
407  * Open and initialize an Interface Adapter.
408  *  o initializes fields of struct rpcrdma_ia, including
409  *    interface and provider attributes and protection zone.
410  */
411 int
412 rpcrdma_ia_open(struct rpcrdma_xprt *xprt, struct sockaddr *addr, int memreg)
413 {
414 	struct rpcrdma_ia *ia = &xprt->rx_ia;
415 	int rc;
416 
417 	ia->ri_id = rpcrdma_create_id(xprt, ia, addr);
418 	if (IS_ERR(ia->ri_id)) {
419 		rc = PTR_ERR(ia->ri_id);
420 		goto out1;
421 	}
422 	ia->ri_device = ia->ri_id->device;
423 
424 	ia->ri_pd = ib_alloc_pd(ia->ri_device, 0);
425 	if (IS_ERR(ia->ri_pd)) {
426 		rc = PTR_ERR(ia->ri_pd);
427 		pr_err("rpcrdma: ib_alloc_pd() returned %d\n", rc);
428 		goto out2;
429 	}
430 
431 	switch (memreg) {
432 	case RPCRDMA_FRMR:
433 		if (frwr_is_supported(ia)) {
434 			ia->ri_ops = &rpcrdma_frwr_memreg_ops;
435 			break;
436 		}
437 		/*FALLTHROUGH*/
438 	case RPCRDMA_MTHCAFMR:
439 		if (fmr_is_supported(ia)) {
440 			ia->ri_ops = &rpcrdma_fmr_memreg_ops;
441 			break;
442 		}
443 		/*FALLTHROUGH*/
444 	default:
445 		pr_err("rpcrdma: Unsupported memory registration mode: %d\n",
446 		       memreg);
447 		rc = -EINVAL;
448 		goto out3;
449 	}
450 
451 	return 0;
452 
453 out3:
454 	ib_dealloc_pd(ia->ri_pd);
455 	ia->ri_pd = NULL;
456 out2:
457 	rpcrdma_destroy_id(ia->ri_id);
458 	ia->ri_id = NULL;
459 out1:
460 	return rc;
461 }
462 
463 /*
464  * Clean up/close an IA.
465  *   o if event handles and PD have been initialized, free them.
466  *   o close the IA
467  */
468 void
469 rpcrdma_ia_close(struct rpcrdma_ia *ia)
470 {
471 	dprintk("RPC:       %s: entering\n", __func__);
472 	if (ia->ri_id != NULL && !IS_ERR(ia->ri_id)) {
473 		if (ia->ri_id->qp)
474 			rdma_destroy_qp(ia->ri_id);
475 		rpcrdma_destroy_id(ia->ri_id);
476 		ia->ri_id = NULL;
477 	}
478 
479 	/* If the pd is still busy, xprtrdma missed freeing a resource */
480 	if (ia->ri_pd && !IS_ERR(ia->ri_pd))
481 		ib_dealloc_pd(ia->ri_pd);
482 }
483 
484 /*
485  * Create unconnected endpoint.
486  */
487 int
488 rpcrdma_ep_create(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia,
489 				struct rpcrdma_create_data_internal *cdata)
490 {
491 	struct rpcrdma_connect_private *pmsg = &ep->rep_cm_private;
492 	struct ib_cq *sendcq, *recvcq;
493 	unsigned int max_qp_wr;
494 	int rc;
495 
496 	if (ia->ri_device->attrs.max_sge < RPCRDMA_MAX_SEND_SGES) {
497 		dprintk("RPC:       %s: insufficient sge's available\n",
498 			__func__);
499 		return -ENOMEM;
500 	}
501 
502 	if (ia->ri_device->attrs.max_qp_wr <= RPCRDMA_BACKWARD_WRS) {
503 		dprintk("RPC:       %s: insufficient wqe's available\n",
504 			__func__);
505 		return -ENOMEM;
506 	}
507 	max_qp_wr = ia->ri_device->attrs.max_qp_wr - RPCRDMA_BACKWARD_WRS - 1;
508 
509 	/* check provider's send/recv wr limits */
510 	if (cdata->max_requests > max_qp_wr)
511 		cdata->max_requests = max_qp_wr;
512 
513 	ep->rep_attr.event_handler = rpcrdma_qp_async_error_upcall;
514 	ep->rep_attr.qp_context = ep;
515 	ep->rep_attr.srq = NULL;
516 	ep->rep_attr.cap.max_send_wr = cdata->max_requests;
517 	ep->rep_attr.cap.max_send_wr += RPCRDMA_BACKWARD_WRS;
518 	ep->rep_attr.cap.max_send_wr += 1;	/* drain cqe */
519 	rc = ia->ri_ops->ro_open(ia, ep, cdata);
520 	if (rc)
521 		return rc;
522 	ep->rep_attr.cap.max_recv_wr = cdata->max_requests;
523 	ep->rep_attr.cap.max_recv_wr += RPCRDMA_BACKWARD_WRS;
524 	ep->rep_attr.cap.max_recv_wr += 1;	/* drain cqe */
525 	ep->rep_attr.cap.max_send_sge = RPCRDMA_MAX_SEND_SGES;
526 	ep->rep_attr.cap.max_recv_sge = 1;
527 	ep->rep_attr.cap.max_inline_data = 0;
528 	ep->rep_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
529 	ep->rep_attr.qp_type = IB_QPT_RC;
530 	ep->rep_attr.port_num = ~0;
531 
532 	dprintk("RPC:       %s: requested max: dtos: send %d recv %d; "
533 		"iovs: send %d recv %d\n",
534 		__func__,
535 		ep->rep_attr.cap.max_send_wr,
536 		ep->rep_attr.cap.max_recv_wr,
537 		ep->rep_attr.cap.max_send_sge,
538 		ep->rep_attr.cap.max_recv_sge);
539 
540 	/* set trigger for requesting send completion */
541 	ep->rep_cqinit = ep->rep_attr.cap.max_send_wr/2 - 1;
542 	if (ep->rep_cqinit <= 2)
543 		ep->rep_cqinit = 0;	/* always signal? */
544 	rpcrdma_init_cqcount(ep, 0);
545 	init_waitqueue_head(&ep->rep_connect_wait);
546 	INIT_DELAYED_WORK(&ep->rep_connect_worker, rpcrdma_connect_worker);
547 
548 	sendcq = ib_alloc_cq(ia->ri_device, NULL,
549 			     ep->rep_attr.cap.max_send_wr + 1,
550 			     0, IB_POLL_SOFTIRQ);
551 	if (IS_ERR(sendcq)) {
552 		rc = PTR_ERR(sendcq);
553 		dprintk("RPC:       %s: failed to create send CQ: %i\n",
554 			__func__, rc);
555 		goto out1;
556 	}
557 
558 	recvcq = ib_alloc_cq(ia->ri_device, NULL,
559 			     ep->rep_attr.cap.max_recv_wr + 1,
560 			     0, IB_POLL_SOFTIRQ);
561 	if (IS_ERR(recvcq)) {
562 		rc = PTR_ERR(recvcq);
563 		dprintk("RPC:       %s: failed to create recv CQ: %i\n",
564 			__func__, rc);
565 		goto out2;
566 	}
567 
568 	ep->rep_attr.send_cq = sendcq;
569 	ep->rep_attr.recv_cq = recvcq;
570 
571 	/* Initialize cma parameters */
572 	memset(&ep->rep_remote_cma, 0, sizeof(ep->rep_remote_cma));
573 
574 	/* Prepare RDMA-CM private message */
575 	pmsg->cp_magic = rpcrdma_cmp_magic;
576 	pmsg->cp_version = RPCRDMA_CMP_VERSION;
577 	pmsg->cp_flags |= ia->ri_ops->ro_send_w_inv_ok;
578 	pmsg->cp_send_size = rpcrdma_encode_buffer_size(cdata->inline_wsize);
579 	pmsg->cp_recv_size = rpcrdma_encode_buffer_size(cdata->inline_rsize);
580 	ep->rep_remote_cma.private_data = pmsg;
581 	ep->rep_remote_cma.private_data_len = sizeof(*pmsg);
582 
583 	/* Client offers RDMA Read but does not initiate */
584 	ep->rep_remote_cma.initiator_depth = 0;
585 	if (ia->ri_device->attrs.max_qp_rd_atom > 32)	/* arbitrary but <= 255 */
586 		ep->rep_remote_cma.responder_resources = 32;
587 	else
588 		ep->rep_remote_cma.responder_resources =
589 						ia->ri_device->attrs.max_qp_rd_atom;
590 
591 	/* Limit transport retries so client can detect server
592 	 * GID changes quickly. RPC layer handles re-establishing
593 	 * transport connection and retransmission.
594 	 */
595 	ep->rep_remote_cma.retry_count = 6;
596 
597 	/* RPC-over-RDMA handles its own flow control. In addition,
598 	 * make all RNR NAKs visible so we know that RPC-over-RDMA
599 	 * flow control is working correctly (no NAKs should be seen).
600 	 */
601 	ep->rep_remote_cma.flow_control = 0;
602 	ep->rep_remote_cma.rnr_retry_count = 0;
603 
604 	return 0;
605 
606 out2:
607 	ib_free_cq(sendcq);
608 out1:
609 	return rc;
610 }
611 
612 /*
613  * rpcrdma_ep_destroy
614  *
615  * Disconnect and destroy endpoint. After this, the only
616  * valid operations on the ep are to free it (if dynamically
617  * allocated) or re-create it.
618  */
619 void
620 rpcrdma_ep_destroy(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
621 {
622 	dprintk("RPC:       %s: entering, connected is %d\n",
623 		__func__, ep->rep_connected);
624 
625 	cancel_delayed_work_sync(&ep->rep_connect_worker);
626 
627 	if (ia->ri_id->qp) {
628 		rpcrdma_ep_disconnect(ep, ia);
629 		rdma_destroy_qp(ia->ri_id);
630 		ia->ri_id->qp = NULL;
631 	}
632 
633 	ib_free_cq(ep->rep_attr.recv_cq);
634 	ib_free_cq(ep->rep_attr.send_cq);
635 }
636 
637 /*
638  * Connect unconnected endpoint.
639  */
640 int
641 rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
642 {
643 	struct rdma_cm_id *id, *old;
644 	int rc = 0;
645 	int retry_count = 0;
646 
647 	if (ep->rep_connected != 0) {
648 		struct rpcrdma_xprt *xprt;
649 retry:
650 		dprintk("RPC:       %s: reconnecting...\n", __func__);
651 
652 		rpcrdma_ep_disconnect(ep, ia);
653 
654 		xprt = container_of(ia, struct rpcrdma_xprt, rx_ia);
655 		id = rpcrdma_create_id(xprt, ia,
656 				(struct sockaddr *)&xprt->rx_data.addr);
657 		if (IS_ERR(id)) {
658 			rc = -EHOSTUNREACH;
659 			goto out;
660 		}
661 		/* TEMP TEMP TEMP - fail if new device:
662 		 * Deregister/remarshal *all* requests!
663 		 * Close and recreate adapter, pd, etc!
664 		 * Re-determine all attributes still sane!
665 		 * More stuff I haven't thought of!
666 		 * Rrrgh!
667 		 */
668 		if (ia->ri_device != id->device) {
669 			printk("RPC:       %s: can't reconnect on "
670 				"different device!\n", __func__);
671 			rpcrdma_destroy_id(id);
672 			rc = -ENETUNREACH;
673 			goto out;
674 		}
675 		/* END TEMP */
676 		rc = rdma_create_qp(id, ia->ri_pd, &ep->rep_attr);
677 		if (rc) {
678 			dprintk("RPC:       %s: rdma_create_qp failed %i\n",
679 				__func__, rc);
680 			rpcrdma_destroy_id(id);
681 			rc = -ENETUNREACH;
682 			goto out;
683 		}
684 
685 		old = ia->ri_id;
686 		ia->ri_id = id;
687 
688 		rdma_destroy_qp(old);
689 		rpcrdma_destroy_id(old);
690 	} else {
691 		dprintk("RPC:       %s: connecting...\n", __func__);
692 		rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
693 		if (rc) {
694 			dprintk("RPC:       %s: rdma_create_qp failed %i\n",
695 				__func__, rc);
696 			/* do not update ep->rep_connected */
697 			return -ENETUNREACH;
698 		}
699 	}
700 
701 	ep->rep_connected = 0;
702 
703 	rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma);
704 	if (rc) {
705 		dprintk("RPC:       %s: rdma_connect() failed with %i\n",
706 				__func__, rc);
707 		goto out;
708 	}
709 
710 	wait_event_interruptible(ep->rep_connect_wait, ep->rep_connected != 0);
711 
712 	/*
713 	 * Check state. A non-peer reject indicates no listener
714 	 * (ECONNREFUSED), which may be a transient state. All
715 	 * others indicate a transport condition which has already
716 	 * undergone a best-effort.
717 	 */
718 	if (ep->rep_connected == -ECONNREFUSED &&
719 	    ++retry_count <= RDMA_CONNECT_RETRY_MAX) {
720 		dprintk("RPC:       %s: non-peer_reject, retry\n", __func__);
721 		goto retry;
722 	}
723 	if (ep->rep_connected <= 0) {
724 		/* Sometimes, the only way to reliably connect to remote
725 		 * CMs is to use same nonzero values for ORD and IRD. */
726 		if (retry_count++ <= RDMA_CONNECT_RETRY_MAX + 1 &&
727 		    (ep->rep_remote_cma.responder_resources == 0 ||
728 		     ep->rep_remote_cma.initiator_depth !=
729 				ep->rep_remote_cma.responder_resources)) {
730 			if (ep->rep_remote_cma.responder_resources == 0)
731 				ep->rep_remote_cma.responder_resources = 1;
732 			ep->rep_remote_cma.initiator_depth =
733 				ep->rep_remote_cma.responder_resources;
734 			goto retry;
735 		}
736 		rc = ep->rep_connected;
737 	} else {
738 		struct rpcrdma_xprt *r_xprt;
739 		unsigned int extras;
740 
741 		dprintk("RPC:       %s: connected\n", __func__);
742 
743 		r_xprt = container_of(ia, struct rpcrdma_xprt, rx_ia);
744 		extras = r_xprt->rx_buf.rb_bc_srv_max_requests;
745 
746 		if (extras) {
747 			rc = rpcrdma_ep_post_extra_recv(r_xprt, extras);
748 			if (rc) {
749 				pr_warn("%s: rpcrdma_ep_post_extra_recv: %i\n",
750 					__func__, rc);
751 				rc = 0;
752 			}
753 		}
754 	}
755 
756 out:
757 	if (rc)
758 		ep->rep_connected = rc;
759 	return rc;
760 }
761 
762 /*
763  * rpcrdma_ep_disconnect
764  *
765  * This is separate from destroy to facilitate the ability
766  * to reconnect without recreating the endpoint.
767  *
768  * This call is not reentrant, and must not be made in parallel
769  * on the same endpoint.
770  */
771 void
772 rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
773 {
774 	int rc;
775 
776 	rc = rdma_disconnect(ia->ri_id);
777 	if (!rc) {
778 		/* returns without wait if not connected */
779 		wait_event_interruptible(ep->rep_connect_wait,
780 							ep->rep_connected != 1);
781 		dprintk("RPC:       %s: after wait, %sconnected\n", __func__,
782 			(ep->rep_connected == 1) ? "still " : "dis");
783 	} else {
784 		dprintk("RPC:       %s: rdma_disconnect %i\n", __func__, rc);
785 		ep->rep_connected = rc;
786 	}
787 
788 	ib_drain_qp(ia->ri_id->qp);
789 }
790 
791 static void
792 rpcrdma_mr_recovery_worker(struct work_struct *work)
793 {
794 	struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
795 						  rb_recovery_worker.work);
796 	struct rpcrdma_mw *mw;
797 
798 	spin_lock(&buf->rb_recovery_lock);
799 	while (!list_empty(&buf->rb_stale_mrs)) {
800 		mw = list_first_entry(&buf->rb_stale_mrs,
801 				      struct rpcrdma_mw, mw_list);
802 		list_del_init(&mw->mw_list);
803 		spin_unlock(&buf->rb_recovery_lock);
804 
805 		dprintk("RPC:       %s: recovering MR %p\n", __func__, mw);
806 		mw->mw_xprt->rx_ia.ri_ops->ro_recover_mr(mw);
807 
808 		spin_lock(&buf->rb_recovery_lock);
809 	}
810 	spin_unlock(&buf->rb_recovery_lock);
811 }
812 
813 void
814 rpcrdma_defer_mr_recovery(struct rpcrdma_mw *mw)
815 {
816 	struct rpcrdma_xprt *r_xprt = mw->mw_xprt;
817 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
818 
819 	spin_lock(&buf->rb_recovery_lock);
820 	list_add(&mw->mw_list, &buf->rb_stale_mrs);
821 	spin_unlock(&buf->rb_recovery_lock);
822 
823 	schedule_delayed_work(&buf->rb_recovery_worker, 0);
824 }
825 
826 static void
827 rpcrdma_create_mrs(struct rpcrdma_xprt *r_xprt)
828 {
829 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
830 	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
831 	unsigned int count;
832 	LIST_HEAD(free);
833 	LIST_HEAD(all);
834 
835 	for (count = 0; count < 32; count++) {
836 		struct rpcrdma_mw *mw;
837 		int rc;
838 
839 		mw = kzalloc(sizeof(*mw), GFP_KERNEL);
840 		if (!mw)
841 			break;
842 
843 		rc = ia->ri_ops->ro_init_mr(ia, mw);
844 		if (rc) {
845 			kfree(mw);
846 			break;
847 		}
848 
849 		mw->mw_xprt = r_xprt;
850 
851 		list_add(&mw->mw_list, &free);
852 		list_add(&mw->mw_all, &all);
853 	}
854 
855 	spin_lock(&buf->rb_mwlock);
856 	list_splice(&free, &buf->rb_mws);
857 	list_splice(&all, &buf->rb_all);
858 	r_xprt->rx_stats.mrs_allocated += count;
859 	spin_unlock(&buf->rb_mwlock);
860 
861 	dprintk("RPC:       %s: created %u MRs\n", __func__, count);
862 }
863 
864 static void
865 rpcrdma_mr_refresh_worker(struct work_struct *work)
866 {
867 	struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
868 						  rb_refresh_worker.work);
869 	struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
870 						   rx_buf);
871 
872 	rpcrdma_create_mrs(r_xprt);
873 }
874 
875 struct rpcrdma_req *
876 rpcrdma_create_req(struct rpcrdma_xprt *r_xprt)
877 {
878 	struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
879 	struct rpcrdma_req *req;
880 
881 	req = kzalloc(sizeof(*req), GFP_KERNEL);
882 	if (req == NULL)
883 		return ERR_PTR(-ENOMEM);
884 
885 	INIT_LIST_HEAD(&req->rl_free);
886 	spin_lock(&buffer->rb_reqslock);
887 	list_add(&req->rl_all, &buffer->rb_allreqs);
888 	spin_unlock(&buffer->rb_reqslock);
889 	req->rl_cqe.done = rpcrdma_wc_send;
890 	req->rl_buffer = &r_xprt->rx_buf;
891 	INIT_LIST_HEAD(&req->rl_registered);
892 	req->rl_send_wr.next = NULL;
893 	req->rl_send_wr.wr_cqe = &req->rl_cqe;
894 	req->rl_send_wr.sg_list = req->rl_send_sge;
895 	req->rl_send_wr.opcode = IB_WR_SEND;
896 	return req;
897 }
898 
899 struct rpcrdma_rep *
900 rpcrdma_create_rep(struct rpcrdma_xprt *r_xprt)
901 {
902 	struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
903 	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
904 	struct rpcrdma_rep *rep;
905 	int rc;
906 
907 	rc = -ENOMEM;
908 	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
909 	if (rep == NULL)
910 		goto out;
911 
912 	rep->rr_rdmabuf = rpcrdma_alloc_regbuf(cdata->inline_rsize,
913 					       DMA_FROM_DEVICE, GFP_KERNEL);
914 	if (IS_ERR(rep->rr_rdmabuf)) {
915 		rc = PTR_ERR(rep->rr_rdmabuf);
916 		goto out_free;
917 	}
918 
919 	rep->rr_device = ia->ri_device;
920 	rep->rr_cqe.done = rpcrdma_wc_receive;
921 	rep->rr_rxprt = r_xprt;
922 	INIT_WORK(&rep->rr_work, rpcrdma_reply_handler);
923 	rep->rr_recv_wr.next = NULL;
924 	rep->rr_recv_wr.wr_cqe = &rep->rr_cqe;
925 	rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
926 	rep->rr_recv_wr.num_sge = 1;
927 	return rep;
928 
929 out_free:
930 	kfree(rep);
931 out:
932 	return ERR_PTR(rc);
933 }
934 
935 int
936 rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
937 {
938 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
939 	int i, rc;
940 
941 	buf->rb_max_requests = r_xprt->rx_data.max_requests;
942 	buf->rb_bc_srv_max_requests = 0;
943 	atomic_set(&buf->rb_credits, 1);
944 	spin_lock_init(&buf->rb_mwlock);
945 	spin_lock_init(&buf->rb_lock);
946 	spin_lock_init(&buf->rb_recovery_lock);
947 	INIT_LIST_HEAD(&buf->rb_mws);
948 	INIT_LIST_HEAD(&buf->rb_all);
949 	INIT_LIST_HEAD(&buf->rb_stale_mrs);
950 	INIT_DELAYED_WORK(&buf->rb_refresh_worker,
951 			  rpcrdma_mr_refresh_worker);
952 	INIT_DELAYED_WORK(&buf->rb_recovery_worker,
953 			  rpcrdma_mr_recovery_worker);
954 
955 	rpcrdma_create_mrs(r_xprt);
956 
957 	INIT_LIST_HEAD(&buf->rb_send_bufs);
958 	INIT_LIST_HEAD(&buf->rb_allreqs);
959 	spin_lock_init(&buf->rb_reqslock);
960 	for (i = 0; i < buf->rb_max_requests; i++) {
961 		struct rpcrdma_req *req;
962 
963 		req = rpcrdma_create_req(r_xprt);
964 		if (IS_ERR(req)) {
965 			dprintk("RPC:       %s: request buffer %d alloc"
966 				" failed\n", __func__, i);
967 			rc = PTR_ERR(req);
968 			goto out;
969 		}
970 		req->rl_backchannel = false;
971 		list_add(&req->rl_free, &buf->rb_send_bufs);
972 	}
973 
974 	INIT_LIST_HEAD(&buf->rb_recv_bufs);
975 	for (i = 0; i < buf->rb_max_requests + RPCRDMA_MAX_BC_REQUESTS; i++) {
976 		struct rpcrdma_rep *rep;
977 
978 		rep = rpcrdma_create_rep(r_xprt);
979 		if (IS_ERR(rep)) {
980 			dprintk("RPC:       %s: reply buffer %d alloc failed\n",
981 				__func__, i);
982 			rc = PTR_ERR(rep);
983 			goto out;
984 		}
985 		list_add(&rep->rr_list, &buf->rb_recv_bufs);
986 	}
987 
988 	return 0;
989 out:
990 	rpcrdma_buffer_destroy(buf);
991 	return rc;
992 }
993 
994 static struct rpcrdma_req *
995 rpcrdma_buffer_get_req_locked(struct rpcrdma_buffer *buf)
996 {
997 	struct rpcrdma_req *req;
998 
999 	req = list_first_entry(&buf->rb_send_bufs,
1000 			       struct rpcrdma_req, rl_free);
1001 	list_del(&req->rl_free);
1002 	return req;
1003 }
1004 
1005 static struct rpcrdma_rep *
1006 rpcrdma_buffer_get_rep_locked(struct rpcrdma_buffer *buf)
1007 {
1008 	struct rpcrdma_rep *rep;
1009 
1010 	rep = list_first_entry(&buf->rb_recv_bufs,
1011 			       struct rpcrdma_rep, rr_list);
1012 	list_del(&rep->rr_list);
1013 	return rep;
1014 }
1015 
1016 static void
1017 rpcrdma_destroy_rep(struct rpcrdma_rep *rep)
1018 {
1019 	rpcrdma_free_regbuf(rep->rr_rdmabuf);
1020 	kfree(rep);
1021 }
1022 
1023 void
1024 rpcrdma_destroy_req(struct rpcrdma_req *req)
1025 {
1026 	rpcrdma_free_regbuf(req->rl_recvbuf);
1027 	rpcrdma_free_regbuf(req->rl_sendbuf);
1028 	rpcrdma_free_regbuf(req->rl_rdmabuf);
1029 	kfree(req);
1030 }
1031 
1032 static void
1033 rpcrdma_destroy_mrs(struct rpcrdma_buffer *buf)
1034 {
1035 	struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
1036 						   rx_buf);
1037 	struct rpcrdma_ia *ia = rdmab_to_ia(buf);
1038 	struct rpcrdma_mw *mw;
1039 	unsigned int count;
1040 
1041 	count = 0;
1042 	spin_lock(&buf->rb_mwlock);
1043 	while (!list_empty(&buf->rb_all)) {
1044 		mw = list_entry(buf->rb_all.next, struct rpcrdma_mw, mw_all);
1045 		list_del(&mw->mw_all);
1046 
1047 		spin_unlock(&buf->rb_mwlock);
1048 		ia->ri_ops->ro_release_mr(mw);
1049 		count++;
1050 		spin_lock(&buf->rb_mwlock);
1051 	}
1052 	spin_unlock(&buf->rb_mwlock);
1053 	r_xprt->rx_stats.mrs_allocated = 0;
1054 
1055 	dprintk("RPC:       %s: released %u MRs\n", __func__, count);
1056 }
1057 
1058 void
1059 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1060 {
1061 	cancel_delayed_work_sync(&buf->rb_recovery_worker);
1062 
1063 	while (!list_empty(&buf->rb_recv_bufs)) {
1064 		struct rpcrdma_rep *rep;
1065 
1066 		rep = rpcrdma_buffer_get_rep_locked(buf);
1067 		rpcrdma_destroy_rep(rep);
1068 	}
1069 	buf->rb_send_count = 0;
1070 
1071 	spin_lock(&buf->rb_reqslock);
1072 	while (!list_empty(&buf->rb_allreqs)) {
1073 		struct rpcrdma_req *req;
1074 
1075 		req = list_first_entry(&buf->rb_allreqs,
1076 				       struct rpcrdma_req, rl_all);
1077 		list_del(&req->rl_all);
1078 
1079 		spin_unlock(&buf->rb_reqslock);
1080 		rpcrdma_destroy_req(req);
1081 		spin_lock(&buf->rb_reqslock);
1082 	}
1083 	spin_unlock(&buf->rb_reqslock);
1084 	buf->rb_recv_count = 0;
1085 
1086 	rpcrdma_destroy_mrs(buf);
1087 }
1088 
1089 struct rpcrdma_mw *
1090 rpcrdma_get_mw(struct rpcrdma_xprt *r_xprt)
1091 {
1092 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1093 	struct rpcrdma_mw *mw = NULL;
1094 
1095 	spin_lock(&buf->rb_mwlock);
1096 	if (!list_empty(&buf->rb_mws)) {
1097 		mw = list_first_entry(&buf->rb_mws,
1098 				      struct rpcrdma_mw, mw_list);
1099 		list_del_init(&mw->mw_list);
1100 	}
1101 	spin_unlock(&buf->rb_mwlock);
1102 
1103 	if (!mw)
1104 		goto out_nomws;
1105 	return mw;
1106 
1107 out_nomws:
1108 	dprintk("RPC:       %s: no MWs available\n", __func__);
1109 	schedule_delayed_work(&buf->rb_refresh_worker, 0);
1110 
1111 	/* Allow the reply handler and refresh worker to run */
1112 	cond_resched();
1113 
1114 	return NULL;
1115 }
1116 
1117 void
1118 rpcrdma_put_mw(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mw *mw)
1119 {
1120 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1121 
1122 	spin_lock(&buf->rb_mwlock);
1123 	list_add_tail(&mw->mw_list, &buf->rb_mws);
1124 	spin_unlock(&buf->rb_mwlock);
1125 }
1126 
1127 static struct rpcrdma_rep *
1128 rpcrdma_buffer_get_rep(struct rpcrdma_buffer *buffers)
1129 {
1130 	/* If an RPC previously completed without a reply (say, a
1131 	 * credential problem or a soft timeout occurs) then hold off
1132 	 * on supplying more Receive buffers until the number of new
1133 	 * pending RPCs catches up to the number of posted Receives.
1134 	 */
1135 	if (unlikely(buffers->rb_send_count < buffers->rb_recv_count))
1136 		return NULL;
1137 
1138 	if (unlikely(list_empty(&buffers->rb_recv_bufs)))
1139 		return NULL;
1140 	buffers->rb_recv_count++;
1141 	return rpcrdma_buffer_get_rep_locked(buffers);
1142 }
1143 
1144 /*
1145  * Get a set of request/reply buffers.
1146  *
1147  * Reply buffer (if available) is attached to send buffer upon return.
1148  */
1149 struct rpcrdma_req *
1150 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1151 {
1152 	struct rpcrdma_req *req;
1153 
1154 	spin_lock(&buffers->rb_lock);
1155 	if (list_empty(&buffers->rb_send_bufs))
1156 		goto out_reqbuf;
1157 	buffers->rb_send_count++;
1158 	req = rpcrdma_buffer_get_req_locked(buffers);
1159 	req->rl_reply = rpcrdma_buffer_get_rep(buffers);
1160 	spin_unlock(&buffers->rb_lock);
1161 	return req;
1162 
1163 out_reqbuf:
1164 	spin_unlock(&buffers->rb_lock);
1165 	pr_warn("RPC:       %s: out of request buffers\n", __func__);
1166 	return NULL;
1167 }
1168 
1169 /*
1170  * Put request/reply buffers back into pool.
1171  * Pre-decrement counter/array index.
1172  */
1173 void
1174 rpcrdma_buffer_put(struct rpcrdma_req *req)
1175 {
1176 	struct rpcrdma_buffer *buffers = req->rl_buffer;
1177 	struct rpcrdma_rep *rep = req->rl_reply;
1178 
1179 	req->rl_send_wr.num_sge = 0;
1180 	req->rl_reply = NULL;
1181 
1182 	spin_lock(&buffers->rb_lock);
1183 	buffers->rb_send_count--;
1184 	list_add_tail(&req->rl_free, &buffers->rb_send_bufs);
1185 	if (rep) {
1186 		buffers->rb_recv_count--;
1187 		list_add_tail(&rep->rr_list, &buffers->rb_recv_bufs);
1188 	}
1189 	spin_unlock(&buffers->rb_lock);
1190 }
1191 
1192 /*
1193  * Recover reply buffers from pool.
1194  * This happens when recovering from disconnect.
1195  */
1196 void
1197 rpcrdma_recv_buffer_get(struct rpcrdma_req *req)
1198 {
1199 	struct rpcrdma_buffer *buffers = req->rl_buffer;
1200 
1201 	spin_lock(&buffers->rb_lock);
1202 	req->rl_reply = rpcrdma_buffer_get_rep(buffers);
1203 	spin_unlock(&buffers->rb_lock);
1204 }
1205 
1206 /*
1207  * Put reply buffers back into pool when not attached to
1208  * request. This happens in error conditions.
1209  */
1210 void
1211 rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1212 {
1213 	struct rpcrdma_buffer *buffers = &rep->rr_rxprt->rx_buf;
1214 
1215 	spin_lock(&buffers->rb_lock);
1216 	buffers->rb_recv_count--;
1217 	list_add_tail(&rep->rr_list, &buffers->rb_recv_bufs);
1218 	spin_unlock(&buffers->rb_lock);
1219 }
1220 
1221 /**
1222  * rpcrdma_alloc_regbuf - allocate and DMA-map memory for SEND/RECV buffers
1223  * @size: size of buffer to be allocated, in bytes
1224  * @direction: direction of data movement
1225  * @flags: GFP flags
1226  *
1227  * Returns an ERR_PTR, or a pointer to a regbuf, a buffer that
1228  * can be persistently DMA-mapped for I/O.
1229  *
1230  * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1231  * receiving the payload of RDMA RECV operations. During Long Calls
1232  * or Replies they may be registered externally via ro_map.
1233  */
1234 struct rpcrdma_regbuf *
1235 rpcrdma_alloc_regbuf(size_t size, enum dma_data_direction direction,
1236 		     gfp_t flags)
1237 {
1238 	struct rpcrdma_regbuf *rb;
1239 
1240 	rb = kmalloc(sizeof(*rb) + size, flags);
1241 	if (rb == NULL)
1242 		return ERR_PTR(-ENOMEM);
1243 
1244 	rb->rg_device = NULL;
1245 	rb->rg_direction = direction;
1246 	rb->rg_iov.length = size;
1247 
1248 	return rb;
1249 }
1250 
1251 /**
1252  * __rpcrdma_map_regbuf - DMA-map a regbuf
1253  * @ia: controlling rpcrdma_ia
1254  * @rb: regbuf to be mapped
1255  */
1256 bool
1257 __rpcrdma_dma_map_regbuf(struct rpcrdma_ia *ia, struct rpcrdma_regbuf *rb)
1258 {
1259 	if (rb->rg_direction == DMA_NONE)
1260 		return false;
1261 
1262 	rb->rg_iov.addr = ib_dma_map_single(ia->ri_device,
1263 					    (void *)rb->rg_base,
1264 					    rdmab_length(rb),
1265 					    rb->rg_direction);
1266 	if (ib_dma_mapping_error(ia->ri_device, rdmab_addr(rb)))
1267 		return false;
1268 
1269 	rb->rg_device = ia->ri_device;
1270 	rb->rg_iov.lkey = ia->ri_pd->local_dma_lkey;
1271 	return true;
1272 }
1273 
1274 static void
1275 rpcrdma_dma_unmap_regbuf(struct rpcrdma_regbuf *rb)
1276 {
1277 	if (!rpcrdma_regbuf_is_mapped(rb))
1278 		return;
1279 
1280 	ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb),
1281 			    rdmab_length(rb), rb->rg_direction);
1282 	rb->rg_device = NULL;
1283 }
1284 
1285 /**
1286  * rpcrdma_free_regbuf - deregister and free registered buffer
1287  * @rb: regbuf to be deregistered and freed
1288  */
1289 void
1290 rpcrdma_free_regbuf(struct rpcrdma_regbuf *rb)
1291 {
1292 	if (!rb)
1293 		return;
1294 
1295 	rpcrdma_dma_unmap_regbuf(rb);
1296 	kfree(rb);
1297 }
1298 
1299 /*
1300  * Prepost any receive buffer, then post send.
1301  *
1302  * Receive buffer is donated to hardware, reclaimed upon recv completion.
1303  */
1304 int
1305 rpcrdma_ep_post(struct rpcrdma_ia *ia,
1306 		struct rpcrdma_ep *ep,
1307 		struct rpcrdma_req *req)
1308 {
1309 	struct ib_send_wr *send_wr = &req->rl_send_wr;
1310 	struct ib_send_wr *send_wr_fail;
1311 	int rc;
1312 
1313 	if (req->rl_reply) {
1314 		rc = rpcrdma_ep_post_recv(ia, req->rl_reply);
1315 		if (rc)
1316 			return rc;
1317 		req->rl_reply = NULL;
1318 	}
1319 
1320 	dprintk("RPC:       %s: posting %d s/g entries\n",
1321 		__func__, send_wr->num_sge);
1322 
1323 	rpcrdma_set_signaled(ep, send_wr);
1324 	rc = ib_post_send(ia->ri_id->qp, send_wr, &send_wr_fail);
1325 	if (rc)
1326 		goto out_postsend_err;
1327 	return 0;
1328 
1329 out_postsend_err:
1330 	pr_err("rpcrdma: RDMA Send ib_post_send returned %i\n", rc);
1331 	return -ENOTCONN;
1332 }
1333 
1334 int
1335 rpcrdma_ep_post_recv(struct rpcrdma_ia *ia,
1336 		     struct rpcrdma_rep *rep)
1337 {
1338 	struct ib_recv_wr *recv_wr_fail;
1339 	int rc;
1340 
1341 	if (!rpcrdma_dma_map_regbuf(ia, rep->rr_rdmabuf))
1342 		goto out_map;
1343 	rc = ib_post_recv(ia->ri_id->qp, &rep->rr_recv_wr, &recv_wr_fail);
1344 	if (rc)
1345 		goto out_postrecv;
1346 	return 0;
1347 
1348 out_map:
1349 	pr_err("rpcrdma: failed to DMA map the Receive buffer\n");
1350 	return -EIO;
1351 
1352 out_postrecv:
1353 	pr_err("rpcrdma: ib_post_recv returned %i\n", rc);
1354 	return -ENOTCONN;
1355 }
1356 
1357 /**
1358  * rpcrdma_ep_post_extra_recv - Post buffers for incoming backchannel requests
1359  * @r_xprt: transport associated with these backchannel resources
1360  * @min_reqs: minimum number of incoming requests expected
1361  *
1362  * Returns zero if all requested buffers were posted, or a negative errno.
1363  */
1364 int
1365 rpcrdma_ep_post_extra_recv(struct rpcrdma_xprt *r_xprt, unsigned int count)
1366 {
1367 	struct rpcrdma_buffer *buffers = &r_xprt->rx_buf;
1368 	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
1369 	struct rpcrdma_rep *rep;
1370 	int rc;
1371 
1372 	while (count--) {
1373 		spin_lock(&buffers->rb_lock);
1374 		if (list_empty(&buffers->rb_recv_bufs))
1375 			goto out_reqbuf;
1376 		rep = rpcrdma_buffer_get_rep_locked(buffers);
1377 		spin_unlock(&buffers->rb_lock);
1378 
1379 		rc = rpcrdma_ep_post_recv(ia, rep);
1380 		if (rc)
1381 			goto out_rc;
1382 	}
1383 
1384 	return 0;
1385 
1386 out_reqbuf:
1387 	spin_unlock(&buffers->rb_lock);
1388 	pr_warn("%s: no extra receive buffers\n", __func__);
1389 	return -ENOMEM;
1390 
1391 out_rc:
1392 	rpcrdma_recv_buffer_put(rep);
1393 	return rc;
1394 }
1395