xref: /linux/net/sunrpc/xprtrdma/verbs.c (revision b85d45947951d23cb22d90caecf4c1eb81342c96)
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 <asm/bitops.h>
55 #include <linux/module.h> /* try_module_get()/module_put() */
56 
57 #include "xprt_rdma.h"
58 
59 /*
60  * Globals/Macros
61  */
62 
63 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
64 # define RPCDBG_FACILITY	RPCDBG_TRANS
65 #endif
66 
67 /*
68  * internal functions
69  */
70 
71 /*
72  * handle replies in tasklet context, using a single, global list
73  * rdma tasklet function -- just turn around and call the func
74  * for all replies on the list
75  */
76 
77 static DEFINE_SPINLOCK(rpcrdma_tk_lock_g);
78 static LIST_HEAD(rpcrdma_tasklets_g);
79 
80 static void
81 rpcrdma_run_tasklet(unsigned long data)
82 {
83 	struct rpcrdma_rep *rep;
84 	unsigned long flags;
85 
86 	data = data;
87 	spin_lock_irqsave(&rpcrdma_tk_lock_g, flags);
88 	while (!list_empty(&rpcrdma_tasklets_g)) {
89 		rep = list_entry(rpcrdma_tasklets_g.next,
90 				 struct rpcrdma_rep, rr_list);
91 		list_del(&rep->rr_list);
92 		spin_unlock_irqrestore(&rpcrdma_tk_lock_g, flags);
93 
94 		rpcrdma_reply_handler(rep);
95 
96 		spin_lock_irqsave(&rpcrdma_tk_lock_g, flags);
97 	}
98 	spin_unlock_irqrestore(&rpcrdma_tk_lock_g, flags);
99 }
100 
101 static DECLARE_TASKLET(rpcrdma_tasklet_g, rpcrdma_run_tasklet, 0UL);
102 
103 static void
104 rpcrdma_schedule_tasklet(struct list_head *sched_list)
105 {
106 	unsigned long flags;
107 
108 	spin_lock_irqsave(&rpcrdma_tk_lock_g, flags);
109 	list_splice_tail(sched_list, &rpcrdma_tasklets_g);
110 	spin_unlock_irqrestore(&rpcrdma_tk_lock_g, flags);
111 	tasklet_schedule(&rpcrdma_tasklet_g);
112 }
113 
114 static void
115 rpcrdma_qp_async_error_upcall(struct ib_event *event, void *context)
116 {
117 	struct rpcrdma_ep *ep = context;
118 
119 	pr_err("RPC:       %s: %s on device %s ep %p\n",
120 	       __func__, ib_event_msg(event->event),
121 		event->device->name, context);
122 	if (ep->rep_connected == 1) {
123 		ep->rep_connected = -EIO;
124 		rpcrdma_conn_func(ep);
125 		wake_up_all(&ep->rep_connect_wait);
126 	}
127 }
128 
129 static void
130 rpcrdma_cq_async_error_upcall(struct ib_event *event, void *context)
131 {
132 	struct rpcrdma_ep *ep = context;
133 
134 	pr_err("RPC:       %s: %s on device %s ep %p\n",
135 	       __func__, ib_event_msg(event->event),
136 		event->device->name, context);
137 	if (ep->rep_connected == 1) {
138 		ep->rep_connected = -EIO;
139 		rpcrdma_conn_func(ep);
140 		wake_up_all(&ep->rep_connect_wait);
141 	}
142 }
143 
144 static void
145 rpcrdma_sendcq_process_wc(struct ib_wc *wc)
146 {
147 	/* WARNING: Only wr_id and status are reliable at this point */
148 	if (wc->wr_id == RPCRDMA_IGNORE_COMPLETION) {
149 		if (wc->status != IB_WC_SUCCESS &&
150 		    wc->status != IB_WC_WR_FLUSH_ERR)
151 			pr_err("RPC:       %s: SEND: %s\n",
152 			       __func__, ib_wc_status_msg(wc->status));
153 	} else {
154 		struct rpcrdma_mw *r;
155 
156 		r = (struct rpcrdma_mw *)(unsigned long)wc->wr_id;
157 		r->mw_sendcompletion(wc);
158 	}
159 }
160 
161 static int
162 rpcrdma_sendcq_poll(struct ib_cq *cq, struct rpcrdma_ep *ep)
163 {
164 	struct ib_wc *wcs;
165 	int budget, count, rc;
166 
167 	budget = RPCRDMA_WC_BUDGET / RPCRDMA_POLLSIZE;
168 	do {
169 		wcs = ep->rep_send_wcs;
170 
171 		rc = ib_poll_cq(cq, RPCRDMA_POLLSIZE, wcs);
172 		if (rc <= 0)
173 			return rc;
174 
175 		count = rc;
176 		while (count-- > 0)
177 			rpcrdma_sendcq_process_wc(wcs++);
178 	} while (rc == RPCRDMA_POLLSIZE && --budget);
179 	return 0;
180 }
181 
182 /*
183  * Handle send, fast_reg_mr, and local_inv completions.
184  *
185  * Send events are typically suppressed and thus do not result
186  * in an upcall. Occasionally one is signaled, however. This
187  * prevents the provider's completion queue from wrapping and
188  * losing a completion.
189  */
190 static void
191 rpcrdma_sendcq_upcall(struct ib_cq *cq, void *cq_context)
192 {
193 	struct rpcrdma_ep *ep = (struct rpcrdma_ep *)cq_context;
194 	int rc;
195 
196 	rc = rpcrdma_sendcq_poll(cq, ep);
197 	if (rc) {
198 		dprintk("RPC:       %s: ib_poll_cq failed: %i\n",
199 			__func__, rc);
200 		return;
201 	}
202 
203 	rc = ib_req_notify_cq(cq,
204 			IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS);
205 	if (rc == 0)
206 		return;
207 	if (rc < 0) {
208 		dprintk("RPC:       %s: ib_req_notify_cq failed: %i\n",
209 			__func__, rc);
210 		return;
211 	}
212 
213 	rpcrdma_sendcq_poll(cq, ep);
214 }
215 
216 static void
217 rpcrdma_recvcq_process_wc(struct ib_wc *wc, struct list_head *sched_list)
218 {
219 	struct rpcrdma_rep *rep =
220 			(struct rpcrdma_rep *)(unsigned long)wc->wr_id;
221 
222 	/* WARNING: Only wr_id and status are reliable at this point */
223 	if (wc->status != IB_WC_SUCCESS)
224 		goto out_fail;
225 
226 	/* status == SUCCESS means all fields in wc are trustworthy */
227 	if (wc->opcode != IB_WC_RECV)
228 		return;
229 
230 	dprintk("RPC:       %s: rep %p opcode 'recv', length %u: success\n",
231 		__func__, rep, wc->byte_len);
232 
233 	rep->rr_len = wc->byte_len;
234 	ib_dma_sync_single_for_cpu(rep->rr_device,
235 				   rdmab_addr(rep->rr_rdmabuf),
236 				   rep->rr_len, DMA_FROM_DEVICE);
237 	prefetch(rdmab_to_msg(rep->rr_rdmabuf));
238 
239 out_schedule:
240 	list_add_tail(&rep->rr_list, sched_list);
241 	return;
242 out_fail:
243 	if (wc->status != IB_WC_WR_FLUSH_ERR)
244 		pr_err("RPC:       %s: rep %p: %s\n",
245 		       __func__, rep, ib_wc_status_msg(wc->status));
246 	rep->rr_len = ~0U;
247 	goto out_schedule;
248 }
249 
250 static int
251 rpcrdma_recvcq_poll(struct ib_cq *cq, struct rpcrdma_ep *ep)
252 {
253 	struct list_head sched_list;
254 	struct ib_wc *wcs;
255 	int budget, count, rc;
256 
257 	INIT_LIST_HEAD(&sched_list);
258 	budget = RPCRDMA_WC_BUDGET / RPCRDMA_POLLSIZE;
259 	do {
260 		wcs = ep->rep_recv_wcs;
261 
262 		rc = ib_poll_cq(cq, RPCRDMA_POLLSIZE, wcs);
263 		if (rc <= 0)
264 			goto out_schedule;
265 
266 		count = rc;
267 		while (count-- > 0)
268 			rpcrdma_recvcq_process_wc(wcs++, &sched_list);
269 	} while (rc == RPCRDMA_POLLSIZE && --budget);
270 	rc = 0;
271 
272 out_schedule:
273 	rpcrdma_schedule_tasklet(&sched_list);
274 	return rc;
275 }
276 
277 /*
278  * Handle receive completions.
279  *
280  * It is reentrant but processes single events in order to maintain
281  * ordering of receives to keep server credits.
282  *
283  * It is the responsibility of the scheduled tasklet to return
284  * recv buffers to the pool. NOTE: this affects synchronization of
285  * connection shutdown. That is, the structures required for
286  * the completion of the reply handler must remain intact until
287  * all memory has been reclaimed.
288  */
289 static void
290 rpcrdma_recvcq_upcall(struct ib_cq *cq, void *cq_context)
291 {
292 	struct rpcrdma_ep *ep = (struct rpcrdma_ep *)cq_context;
293 	int rc;
294 
295 	rc = rpcrdma_recvcq_poll(cq, ep);
296 	if (rc) {
297 		dprintk("RPC:       %s: ib_poll_cq failed: %i\n",
298 			__func__, rc);
299 		return;
300 	}
301 
302 	rc = ib_req_notify_cq(cq,
303 			IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS);
304 	if (rc == 0)
305 		return;
306 	if (rc < 0) {
307 		dprintk("RPC:       %s: ib_req_notify_cq failed: %i\n",
308 			__func__, rc);
309 		return;
310 	}
311 
312 	rpcrdma_recvcq_poll(cq, ep);
313 }
314 
315 static void
316 rpcrdma_flush_cqs(struct rpcrdma_ep *ep)
317 {
318 	struct ib_wc wc;
319 	LIST_HEAD(sched_list);
320 
321 	while (ib_poll_cq(ep->rep_attr.recv_cq, 1, &wc) > 0)
322 		rpcrdma_recvcq_process_wc(&wc, &sched_list);
323 	if (!list_empty(&sched_list))
324 		rpcrdma_schedule_tasklet(&sched_list);
325 	while (ib_poll_cq(ep->rep_attr.send_cq, 1, &wc) > 0)
326 		rpcrdma_sendcq_process_wc(&wc);
327 }
328 
329 static int
330 rpcrdma_conn_upcall(struct rdma_cm_id *id, struct rdma_cm_event *event)
331 {
332 	struct rpcrdma_xprt *xprt = id->context;
333 	struct rpcrdma_ia *ia = &xprt->rx_ia;
334 	struct rpcrdma_ep *ep = &xprt->rx_ep;
335 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
336 	struct sockaddr *sap = (struct sockaddr *)&ep->rep_remote_addr;
337 #endif
338 	struct ib_qp_attr *attr = &ia->ri_qp_attr;
339 	struct ib_qp_init_attr *iattr = &ia->ri_qp_init_attr;
340 	int connstate = 0;
341 
342 	switch (event->event) {
343 	case RDMA_CM_EVENT_ADDR_RESOLVED:
344 	case RDMA_CM_EVENT_ROUTE_RESOLVED:
345 		ia->ri_async_rc = 0;
346 		complete(&ia->ri_done);
347 		break;
348 	case RDMA_CM_EVENT_ADDR_ERROR:
349 		ia->ri_async_rc = -EHOSTUNREACH;
350 		dprintk("RPC:       %s: CM address resolution error, ep 0x%p\n",
351 			__func__, ep);
352 		complete(&ia->ri_done);
353 		break;
354 	case RDMA_CM_EVENT_ROUTE_ERROR:
355 		ia->ri_async_rc = -ENETUNREACH;
356 		dprintk("RPC:       %s: CM route resolution error, ep 0x%p\n",
357 			__func__, ep);
358 		complete(&ia->ri_done);
359 		break;
360 	case RDMA_CM_EVENT_ESTABLISHED:
361 		connstate = 1;
362 		ib_query_qp(ia->ri_id->qp, attr,
363 			    IB_QP_MAX_QP_RD_ATOMIC | IB_QP_MAX_DEST_RD_ATOMIC,
364 			    iattr);
365 		dprintk("RPC:       %s: %d responder resources"
366 			" (%d initiator)\n",
367 			__func__, attr->max_dest_rd_atomic,
368 			attr->max_rd_atomic);
369 		goto connected;
370 	case RDMA_CM_EVENT_CONNECT_ERROR:
371 		connstate = -ENOTCONN;
372 		goto connected;
373 	case RDMA_CM_EVENT_UNREACHABLE:
374 		connstate = -ENETDOWN;
375 		goto connected;
376 	case RDMA_CM_EVENT_REJECTED:
377 		connstate = -ECONNREFUSED;
378 		goto connected;
379 	case RDMA_CM_EVENT_DISCONNECTED:
380 		connstate = -ECONNABORTED;
381 		goto connected;
382 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
383 		connstate = -ENODEV;
384 connected:
385 		dprintk("RPC:       %s: %sconnected\n",
386 					__func__, connstate > 0 ? "" : "dis");
387 		ep->rep_connected = connstate;
388 		rpcrdma_conn_func(ep);
389 		wake_up_all(&ep->rep_connect_wait);
390 		/*FALLTHROUGH*/
391 	default:
392 		dprintk("RPC:       %s: %pIS:%u (ep 0x%p): %s\n",
393 			__func__, sap, rpc_get_port(sap), ep,
394 			rdma_event_msg(event->event));
395 		break;
396 	}
397 
398 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
399 	if (connstate == 1) {
400 		int ird = attr->max_dest_rd_atomic;
401 		int tird = ep->rep_remote_cma.responder_resources;
402 
403 		pr_info("rpcrdma: connection to %pIS:%u on %s, memreg '%s', %d credits, %d responders%s\n",
404 			sap, rpc_get_port(sap),
405 			ia->ri_device->name,
406 			ia->ri_ops->ro_displayname,
407 			xprt->rx_buf.rb_max_requests,
408 			ird, ird < 4 && ird < tird / 2 ? " (low!)" : "");
409 	} else if (connstate < 0) {
410 		pr_info("rpcrdma: connection to %pIS:%u closed (%d)\n",
411 			sap, rpc_get_port(sap), connstate);
412 	}
413 #endif
414 
415 	return 0;
416 }
417 
418 static void rpcrdma_destroy_id(struct rdma_cm_id *id)
419 {
420 	if (id) {
421 		module_put(id->device->owner);
422 		rdma_destroy_id(id);
423 	}
424 }
425 
426 static struct rdma_cm_id *
427 rpcrdma_create_id(struct rpcrdma_xprt *xprt,
428 			struct rpcrdma_ia *ia, struct sockaddr *addr)
429 {
430 	struct rdma_cm_id *id;
431 	int rc;
432 
433 	init_completion(&ia->ri_done);
434 
435 	id = rdma_create_id(rpcrdma_conn_upcall, xprt, RDMA_PS_TCP, IB_QPT_RC);
436 	if (IS_ERR(id)) {
437 		rc = PTR_ERR(id);
438 		dprintk("RPC:       %s: rdma_create_id() failed %i\n",
439 			__func__, rc);
440 		return id;
441 	}
442 
443 	ia->ri_async_rc = -ETIMEDOUT;
444 	rc = rdma_resolve_addr(id, NULL, addr, RDMA_RESOLVE_TIMEOUT);
445 	if (rc) {
446 		dprintk("RPC:       %s: rdma_resolve_addr() failed %i\n",
447 			__func__, rc);
448 		goto out;
449 	}
450 	wait_for_completion_interruptible_timeout(&ia->ri_done,
451 				msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1);
452 
453 	/* FIXME:
454 	 * Until xprtrdma supports DEVICE_REMOVAL, the provider must
455 	 * be pinned while there are active NFS/RDMA mounts to prevent
456 	 * hangs and crashes at umount time.
457 	 */
458 	if (!ia->ri_async_rc && !try_module_get(id->device->owner)) {
459 		dprintk("RPC:       %s: Failed to get device module\n",
460 			__func__);
461 		ia->ri_async_rc = -ENODEV;
462 	}
463 	rc = ia->ri_async_rc;
464 	if (rc)
465 		goto out;
466 
467 	ia->ri_async_rc = -ETIMEDOUT;
468 	rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
469 	if (rc) {
470 		dprintk("RPC:       %s: rdma_resolve_route() failed %i\n",
471 			__func__, rc);
472 		goto put;
473 	}
474 	wait_for_completion_interruptible_timeout(&ia->ri_done,
475 				msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1);
476 	rc = ia->ri_async_rc;
477 	if (rc)
478 		goto put;
479 
480 	return id;
481 put:
482 	module_put(id->device->owner);
483 out:
484 	rdma_destroy_id(id);
485 	return ERR_PTR(rc);
486 }
487 
488 /*
489  * Drain any cq, prior to teardown.
490  */
491 static void
492 rpcrdma_clean_cq(struct ib_cq *cq)
493 {
494 	struct ib_wc wc;
495 	int count = 0;
496 
497 	while (1 == ib_poll_cq(cq, 1, &wc))
498 		++count;
499 
500 	if (count)
501 		dprintk("RPC:       %s: flushed %d events (last 0x%x)\n",
502 			__func__, count, wc.opcode);
503 }
504 
505 /*
506  * Exported functions.
507  */
508 
509 /*
510  * Open and initialize an Interface Adapter.
511  *  o initializes fields of struct rpcrdma_ia, including
512  *    interface and provider attributes and protection zone.
513  */
514 int
515 rpcrdma_ia_open(struct rpcrdma_xprt *xprt, struct sockaddr *addr, int memreg)
516 {
517 	struct rpcrdma_ia *ia = &xprt->rx_ia;
518 	struct ib_device_attr *devattr = &ia->ri_devattr;
519 	int rc;
520 
521 	ia->ri_dma_mr = NULL;
522 
523 	ia->ri_id = rpcrdma_create_id(xprt, ia, addr);
524 	if (IS_ERR(ia->ri_id)) {
525 		rc = PTR_ERR(ia->ri_id);
526 		goto out1;
527 	}
528 	ia->ri_device = ia->ri_id->device;
529 
530 	ia->ri_pd = ib_alloc_pd(ia->ri_device);
531 	if (IS_ERR(ia->ri_pd)) {
532 		rc = PTR_ERR(ia->ri_pd);
533 		dprintk("RPC:       %s: ib_alloc_pd() failed %i\n",
534 			__func__, rc);
535 		goto out2;
536 	}
537 
538 	rc = ib_query_device(ia->ri_device, devattr);
539 	if (rc) {
540 		dprintk("RPC:       %s: ib_query_device failed %d\n",
541 			__func__, rc);
542 		goto out3;
543 	}
544 
545 	if (memreg == RPCRDMA_FRMR) {
546 		if (!(devattr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) ||
547 		    (devattr->max_fast_reg_page_list_len == 0)) {
548 			dprintk("RPC:       %s: FRMR registration "
549 				"not supported by HCA\n", __func__);
550 			memreg = RPCRDMA_MTHCAFMR;
551 		}
552 	}
553 	if (memreg == RPCRDMA_MTHCAFMR) {
554 		if (!ia->ri_device->alloc_fmr) {
555 			dprintk("RPC:       %s: MTHCAFMR registration "
556 				"not supported by HCA\n", __func__);
557 			rc = -EINVAL;
558 			goto out3;
559 		}
560 	}
561 
562 	switch (memreg) {
563 	case RPCRDMA_FRMR:
564 		ia->ri_ops = &rpcrdma_frwr_memreg_ops;
565 		break;
566 	case RPCRDMA_ALLPHYSICAL:
567 		ia->ri_ops = &rpcrdma_physical_memreg_ops;
568 		break;
569 	case RPCRDMA_MTHCAFMR:
570 		ia->ri_ops = &rpcrdma_fmr_memreg_ops;
571 		break;
572 	default:
573 		printk(KERN_ERR "RPC: Unsupported memory "
574 				"registration mode: %d\n", memreg);
575 		rc = -ENOMEM;
576 		goto out3;
577 	}
578 	dprintk("RPC:       %s: memory registration strategy is '%s'\n",
579 		__func__, ia->ri_ops->ro_displayname);
580 
581 	rwlock_init(&ia->ri_qplock);
582 	return 0;
583 
584 out3:
585 	ib_dealloc_pd(ia->ri_pd);
586 	ia->ri_pd = NULL;
587 out2:
588 	rpcrdma_destroy_id(ia->ri_id);
589 	ia->ri_id = NULL;
590 out1:
591 	return rc;
592 }
593 
594 /*
595  * Clean up/close an IA.
596  *   o if event handles and PD have been initialized, free them.
597  *   o close the IA
598  */
599 void
600 rpcrdma_ia_close(struct rpcrdma_ia *ia)
601 {
602 	dprintk("RPC:       %s: entering\n", __func__);
603 	if (ia->ri_id != NULL && !IS_ERR(ia->ri_id)) {
604 		if (ia->ri_id->qp)
605 			rdma_destroy_qp(ia->ri_id);
606 		rpcrdma_destroy_id(ia->ri_id);
607 		ia->ri_id = NULL;
608 	}
609 
610 	/* If the pd is still busy, xprtrdma missed freeing a resource */
611 	if (ia->ri_pd && !IS_ERR(ia->ri_pd))
612 		ib_dealloc_pd(ia->ri_pd);
613 }
614 
615 /*
616  * Create unconnected endpoint.
617  */
618 int
619 rpcrdma_ep_create(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia,
620 				struct rpcrdma_create_data_internal *cdata)
621 {
622 	struct ib_device_attr *devattr = &ia->ri_devattr;
623 	struct ib_cq *sendcq, *recvcq;
624 	struct ib_cq_init_attr cq_attr = {};
625 	int rc, err;
626 
627 	if (devattr->max_sge < RPCRDMA_MAX_IOVS) {
628 		dprintk("RPC:       %s: insufficient sge's available\n",
629 			__func__);
630 		return -ENOMEM;
631 	}
632 
633 	/* check provider's send/recv wr limits */
634 	if (cdata->max_requests > devattr->max_qp_wr)
635 		cdata->max_requests = devattr->max_qp_wr;
636 
637 	ep->rep_attr.event_handler = rpcrdma_qp_async_error_upcall;
638 	ep->rep_attr.qp_context = ep;
639 	ep->rep_attr.srq = NULL;
640 	ep->rep_attr.cap.max_send_wr = cdata->max_requests;
641 	rc = ia->ri_ops->ro_open(ia, ep, cdata);
642 	if (rc)
643 		return rc;
644 	ep->rep_attr.cap.max_recv_wr = cdata->max_requests;
645 	ep->rep_attr.cap.max_send_sge = RPCRDMA_MAX_IOVS;
646 	ep->rep_attr.cap.max_recv_sge = 1;
647 	ep->rep_attr.cap.max_inline_data = 0;
648 	ep->rep_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
649 	ep->rep_attr.qp_type = IB_QPT_RC;
650 	ep->rep_attr.port_num = ~0;
651 
652 	dprintk("RPC:       %s: requested max: dtos: send %d recv %d; "
653 		"iovs: send %d recv %d\n",
654 		__func__,
655 		ep->rep_attr.cap.max_send_wr,
656 		ep->rep_attr.cap.max_recv_wr,
657 		ep->rep_attr.cap.max_send_sge,
658 		ep->rep_attr.cap.max_recv_sge);
659 
660 	/* set trigger for requesting send completion */
661 	ep->rep_cqinit = ep->rep_attr.cap.max_send_wr/2 - 1;
662 	if (ep->rep_cqinit > RPCRDMA_MAX_UNSIGNALED_SENDS)
663 		ep->rep_cqinit = RPCRDMA_MAX_UNSIGNALED_SENDS;
664 	else if (ep->rep_cqinit <= 2)
665 		ep->rep_cqinit = 0;
666 	INIT_CQCOUNT(ep);
667 	init_waitqueue_head(&ep->rep_connect_wait);
668 	INIT_DELAYED_WORK(&ep->rep_connect_worker, rpcrdma_connect_worker);
669 
670 	cq_attr.cqe = ep->rep_attr.cap.max_send_wr + 1;
671 	sendcq = ib_create_cq(ia->ri_device, rpcrdma_sendcq_upcall,
672 			      rpcrdma_cq_async_error_upcall, ep, &cq_attr);
673 	if (IS_ERR(sendcq)) {
674 		rc = PTR_ERR(sendcq);
675 		dprintk("RPC:       %s: failed to create send CQ: %i\n",
676 			__func__, rc);
677 		goto out1;
678 	}
679 
680 	rc = ib_req_notify_cq(sendcq, IB_CQ_NEXT_COMP);
681 	if (rc) {
682 		dprintk("RPC:       %s: ib_req_notify_cq failed: %i\n",
683 			__func__, rc);
684 		goto out2;
685 	}
686 
687 	cq_attr.cqe = ep->rep_attr.cap.max_recv_wr + 1;
688 	recvcq = ib_create_cq(ia->ri_device, rpcrdma_recvcq_upcall,
689 			      rpcrdma_cq_async_error_upcall, ep, &cq_attr);
690 	if (IS_ERR(recvcq)) {
691 		rc = PTR_ERR(recvcq);
692 		dprintk("RPC:       %s: failed to create recv CQ: %i\n",
693 			__func__, rc);
694 		goto out2;
695 	}
696 
697 	rc = ib_req_notify_cq(recvcq, IB_CQ_NEXT_COMP);
698 	if (rc) {
699 		dprintk("RPC:       %s: ib_req_notify_cq failed: %i\n",
700 			__func__, rc);
701 		ib_destroy_cq(recvcq);
702 		goto out2;
703 	}
704 
705 	ep->rep_attr.send_cq = sendcq;
706 	ep->rep_attr.recv_cq = recvcq;
707 
708 	/* Initialize cma parameters */
709 
710 	/* RPC/RDMA does not use private data */
711 	ep->rep_remote_cma.private_data = NULL;
712 	ep->rep_remote_cma.private_data_len = 0;
713 
714 	/* Client offers RDMA Read but does not initiate */
715 	ep->rep_remote_cma.initiator_depth = 0;
716 	if (devattr->max_qp_rd_atom > 32)	/* arbitrary but <= 255 */
717 		ep->rep_remote_cma.responder_resources = 32;
718 	else
719 		ep->rep_remote_cma.responder_resources =
720 						devattr->max_qp_rd_atom;
721 
722 	ep->rep_remote_cma.retry_count = 7;
723 	ep->rep_remote_cma.flow_control = 0;
724 	ep->rep_remote_cma.rnr_retry_count = 0;
725 
726 	return 0;
727 
728 out2:
729 	err = ib_destroy_cq(sendcq);
730 	if (err)
731 		dprintk("RPC:       %s: ib_destroy_cq returned %i\n",
732 			__func__, err);
733 out1:
734 	if (ia->ri_dma_mr)
735 		ib_dereg_mr(ia->ri_dma_mr);
736 	return rc;
737 }
738 
739 /*
740  * rpcrdma_ep_destroy
741  *
742  * Disconnect and destroy endpoint. After this, the only
743  * valid operations on the ep are to free it (if dynamically
744  * allocated) or re-create it.
745  */
746 void
747 rpcrdma_ep_destroy(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
748 {
749 	int rc;
750 
751 	dprintk("RPC:       %s: entering, connected is %d\n",
752 		__func__, ep->rep_connected);
753 
754 	cancel_delayed_work_sync(&ep->rep_connect_worker);
755 
756 	if (ia->ri_id->qp)
757 		rpcrdma_ep_disconnect(ep, ia);
758 
759 	rpcrdma_clean_cq(ep->rep_attr.recv_cq);
760 	rpcrdma_clean_cq(ep->rep_attr.send_cq);
761 
762 	if (ia->ri_id->qp) {
763 		rdma_destroy_qp(ia->ri_id);
764 		ia->ri_id->qp = NULL;
765 	}
766 
767 	rc = ib_destroy_cq(ep->rep_attr.recv_cq);
768 	if (rc)
769 		dprintk("RPC:       %s: ib_destroy_cq returned %i\n",
770 			__func__, rc);
771 
772 	rc = ib_destroy_cq(ep->rep_attr.send_cq);
773 	if (rc)
774 		dprintk("RPC:       %s: ib_destroy_cq returned %i\n",
775 			__func__, rc);
776 
777 	if (ia->ri_dma_mr) {
778 		rc = ib_dereg_mr(ia->ri_dma_mr);
779 		dprintk("RPC:       %s: ib_dereg_mr returned %i\n",
780 			__func__, rc);
781 	}
782 }
783 
784 /*
785  * Connect unconnected endpoint.
786  */
787 int
788 rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
789 {
790 	struct rdma_cm_id *id, *old;
791 	int rc = 0;
792 	int retry_count = 0;
793 
794 	if (ep->rep_connected != 0) {
795 		struct rpcrdma_xprt *xprt;
796 retry:
797 		dprintk("RPC:       %s: reconnecting...\n", __func__);
798 
799 		rpcrdma_ep_disconnect(ep, ia);
800 		rpcrdma_flush_cqs(ep);
801 
802 		xprt = container_of(ia, struct rpcrdma_xprt, rx_ia);
803 		id = rpcrdma_create_id(xprt, ia,
804 				(struct sockaddr *)&xprt->rx_data.addr);
805 		if (IS_ERR(id)) {
806 			rc = -EHOSTUNREACH;
807 			goto out;
808 		}
809 		/* TEMP TEMP TEMP - fail if new device:
810 		 * Deregister/remarshal *all* requests!
811 		 * Close and recreate adapter, pd, etc!
812 		 * Re-determine all attributes still sane!
813 		 * More stuff I haven't thought of!
814 		 * Rrrgh!
815 		 */
816 		if (ia->ri_device != id->device) {
817 			printk("RPC:       %s: can't reconnect on "
818 				"different device!\n", __func__);
819 			rpcrdma_destroy_id(id);
820 			rc = -ENETUNREACH;
821 			goto out;
822 		}
823 		/* END TEMP */
824 		rc = rdma_create_qp(id, ia->ri_pd, &ep->rep_attr);
825 		if (rc) {
826 			dprintk("RPC:       %s: rdma_create_qp failed %i\n",
827 				__func__, rc);
828 			rpcrdma_destroy_id(id);
829 			rc = -ENETUNREACH;
830 			goto out;
831 		}
832 
833 		write_lock(&ia->ri_qplock);
834 		old = ia->ri_id;
835 		ia->ri_id = id;
836 		write_unlock(&ia->ri_qplock);
837 
838 		rdma_destroy_qp(old);
839 		rpcrdma_destroy_id(old);
840 	} else {
841 		dprintk("RPC:       %s: connecting...\n", __func__);
842 		rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
843 		if (rc) {
844 			dprintk("RPC:       %s: rdma_create_qp failed %i\n",
845 				__func__, rc);
846 			/* do not update ep->rep_connected */
847 			return -ENETUNREACH;
848 		}
849 	}
850 
851 	ep->rep_connected = 0;
852 
853 	rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma);
854 	if (rc) {
855 		dprintk("RPC:       %s: rdma_connect() failed with %i\n",
856 				__func__, rc);
857 		goto out;
858 	}
859 
860 	wait_event_interruptible(ep->rep_connect_wait, ep->rep_connected != 0);
861 
862 	/*
863 	 * Check state. A non-peer reject indicates no listener
864 	 * (ECONNREFUSED), which may be a transient state. All
865 	 * others indicate a transport condition which has already
866 	 * undergone a best-effort.
867 	 */
868 	if (ep->rep_connected == -ECONNREFUSED &&
869 	    ++retry_count <= RDMA_CONNECT_RETRY_MAX) {
870 		dprintk("RPC:       %s: non-peer_reject, retry\n", __func__);
871 		goto retry;
872 	}
873 	if (ep->rep_connected <= 0) {
874 		/* Sometimes, the only way to reliably connect to remote
875 		 * CMs is to use same nonzero values for ORD and IRD. */
876 		if (retry_count++ <= RDMA_CONNECT_RETRY_MAX + 1 &&
877 		    (ep->rep_remote_cma.responder_resources == 0 ||
878 		     ep->rep_remote_cma.initiator_depth !=
879 				ep->rep_remote_cma.responder_resources)) {
880 			if (ep->rep_remote_cma.responder_resources == 0)
881 				ep->rep_remote_cma.responder_resources = 1;
882 			ep->rep_remote_cma.initiator_depth =
883 				ep->rep_remote_cma.responder_resources;
884 			goto retry;
885 		}
886 		rc = ep->rep_connected;
887 	} else {
888 		dprintk("RPC:       %s: connected\n", __func__);
889 	}
890 
891 out:
892 	if (rc)
893 		ep->rep_connected = rc;
894 	return rc;
895 }
896 
897 /*
898  * rpcrdma_ep_disconnect
899  *
900  * This is separate from destroy to facilitate the ability
901  * to reconnect without recreating the endpoint.
902  *
903  * This call is not reentrant, and must not be made in parallel
904  * on the same endpoint.
905  */
906 void
907 rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
908 {
909 	int rc;
910 
911 	rpcrdma_flush_cqs(ep);
912 	rc = rdma_disconnect(ia->ri_id);
913 	if (!rc) {
914 		/* returns without wait if not connected */
915 		wait_event_interruptible(ep->rep_connect_wait,
916 							ep->rep_connected != 1);
917 		dprintk("RPC:       %s: after wait, %sconnected\n", __func__,
918 			(ep->rep_connected == 1) ? "still " : "dis");
919 	} else {
920 		dprintk("RPC:       %s: rdma_disconnect %i\n", __func__, rc);
921 		ep->rep_connected = rc;
922 	}
923 }
924 
925 static struct rpcrdma_req *
926 rpcrdma_create_req(struct rpcrdma_xprt *r_xprt)
927 {
928 	struct rpcrdma_req *req;
929 
930 	req = kzalloc(sizeof(*req), GFP_KERNEL);
931 	if (req == NULL)
932 		return ERR_PTR(-ENOMEM);
933 
934 	req->rl_buffer = &r_xprt->rx_buf;
935 	return req;
936 }
937 
938 static struct rpcrdma_rep *
939 rpcrdma_create_rep(struct rpcrdma_xprt *r_xprt)
940 {
941 	struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
942 	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
943 	struct rpcrdma_rep *rep;
944 	int rc;
945 
946 	rc = -ENOMEM;
947 	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
948 	if (rep == NULL)
949 		goto out;
950 
951 	rep->rr_rdmabuf = rpcrdma_alloc_regbuf(ia, cdata->inline_rsize,
952 					       GFP_KERNEL);
953 	if (IS_ERR(rep->rr_rdmabuf)) {
954 		rc = PTR_ERR(rep->rr_rdmabuf);
955 		goto out_free;
956 	}
957 
958 	rep->rr_device = ia->ri_device;
959 	rep->rr_rxprt = r_xprt;
960 	return rep;
961 
962 out_free:
963 	kfree(rep);
964 out:
965 	return ERR_PTR(rc);
966 }
967 
968 int
969 rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
970 {
971 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
972 	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
973 	struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
974 	char *p;
975 	size_t len;
976 	int i, rc;
977 
978 	buf->rb_max_requests = cdata->max_requests;
979 	spin_lock_init(&buf->rb_lock);
980 
981 	/* Need to allocate:
982 	 *   1.  arrays for send and recv pointers
983 	 *   2.  arrays of struct rpcrdma_req to fill in pointers
984 	 *   3.  array of struct rpcrdma_rep for replies
985 	 * Send/recv buffers in req/rep need to be registered
986 	 */
987 	len = buf->rb_max_requests *
988 		(sizeof(struct rpcrdma_req *) + sizeof(struct rpcrdma_rep *));
989 
990 	p = kzalloc(len, GFP_KERNEL);
991 	if (p == NULL) {
992 		dprintk("RPC:       %s: req_t/rep_t/pad kzalloc(%zd) failed\n",
993 			__func__, len);
994 		rc = -ENOMEM;
995 		goto out;
996 	}
997 	buf->rb_pool = p;	/* for freeing it later */
998 
999 	buf->rb_send_bufs = (struct rpcrdma_req **) p;
1000 	p = (char *) &buf->rb_send_bufs[buf->rb_max_requests];
1001 	buf->rb_recv_bufs = (struct rpcrdma_rep **) p;
1002 	p = (char *) &buf->rb_recv_bufs[buf->rb_max_requests];
1003 
1004 	rc = ia->ri_ops->ro_init(r_xprt);
1005 	if (rc)
1006 		goto out;
1007 
1008 	for (i = 0; i < buf->rb_max_requests; i++) {
1009 		struct rpcrdma_req *req;
1010 		struct rpcrdma_rep *rep;
1011 
1012 		req = rpcrdma_create_req(r_xprt);
1013 		if (IS_ERR(req)) {
1014 			dprintk("RPC:       %s: request buffer %d alloc"
1015 				" failed\n", __func__, i);
1016 			rc = PTR_ERR(req);
1017 			goto out;
1018 		}
1019 		buf->rb_send_bufs[i] = req;
1020 
1021 		rep = rpcrdma_create_rep(r_xprt);
1022 		if (IS_ERR(rep)) {
1023 			dprintk("RPC:       %s: reply buffer %d alloc failed\n",
1024 				__func__, i);
1025 			rc = PTR_ERR(rep);
1026 			goto out;
1027 		}
1028 		buf->rb_recv_bufs[i] = rep;
1029 	}
1030 
1031 	return 0;
1032 out:
1033 	rpcrdma_buffer_destroy(buf);
1034 	return rc;
1035 }
1036 
1037 static void
1038 rpcrdma_destroy_rep(struct rpcrdma_ia *ia, struct rpcrdma_rep *rep)
1039 {
1040 	if (!rep)
1041 		return;
1042 
1043 	rpcrdma_free_regbuf(ia, rep->rr_rdmabuf);
1044 	kfree(rep);
1045 }
1046 
1047 static void
1048 rpcrdma_destroy_req(struct rpcrdma_ia *ia, struct rpcrdma_req *req)
1049 {
1050 	if (!req)
1051 		return;
1052 
1053 	rpcrdma_free_regbuf(ia, req->rl_sendbuf);
1054 	rpcrdma_free_regbuf(ia, req->rl_rdmabuf);
1055 	kfree(req);
1056 }
1057 
1058 void
1059 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1060 {
1061 	struct rpcrdma_ia *ia = rdmab_to_ia(buf);
1062 	int i;
1063 
1064 	/* clean up in reverse order from create
1065 	 *   1.  recv mr memory (mr free, then kfree)
1066 	 *   2.  send mr memory (mr free, then kfree)
1067 	 *   3.  MWs
1068 	 */
1069 	dprintk("RPC:       %s: entering\n", __func__);
1070 
1071 	for (i = 0; i < buf->rb_max_requests; i++) {
1072 		if (buf->rb_recv_bufs)
1073 			rpcrdma_destroy_rep(ia, buf->rb_recv_bufs[i]);
1074 		if (buf->rb_send_bufs)
1075 			rpcrdma_destroy_req(ia, buf->rb_send_bufs[i]);
1076 	}
1077 
1078 	ia->ri_ops->ro_destroy(buf);
1079 
1080 	kfree(buf->rb_pool);
1081 }
1082 
1083 struct rpcrdma_mw *
1084 rpcrdma_get_mw(struct rpcrdma_xprt *r_xprt)
1085 {
1086 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1087 	struct rpcrdma_mw *mw = NULL;
1088 
1089 	spin_lock(&buf->rb_mwlock);
1090 	if (!list_empty(&buf->rb_mws)) {
1091 		mw = list_first_entry(&buf->rb_mws,
1092 				      struct rpcrdma_mw, mw_list);
1093 		list_del_init(&mw->mw_list);
1094 	}
1095 	spin_unlock(&buf->rb_mwlock);
1096 
1097 	if (!mw)
1098 		pr_err("RPC:       %s: no MWs available\n", __func__);
1099 	return mw;
1100 }
1101 
1102 void
1103 rpcrdma_put_mw(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mw *mw)
1104 {
1105 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1106 
1107 	spin_lock(&buf->rb_mwlock);
1108 	list_add_tail(&mw->mw_list, &buf->rb_mws);
1109 	spin_unlock(&buf->rb_mwlock);
1110 }
1111 
1112 static void
1113 rpcrdma_buffer_put_sendbuf(struct rpcrdma_req *req, struct rpcrdma_buffer *buf)
1114 {
1115 	buf->rb_send_bufs[--buf->rb_send_index] = req;
1116 	req->rl_niovs = 0;
1117 	if (req->rl_reply) {
1118 		buf->rb_recv_bufs[--buf->rb_recv_index] = req->rl_reply;
1119 		req->rl_reply = NULL;
1120 	}
1121 }
1122 
1123 /*
1124  * Get a set of request/reply buffers.
1125  *
1126  * Reply buffer (if needed) is attached to send buffer upon return.
1127  * Rule:
1128  *    rb_send_index and rb_recv_index MUST always be pointing to the
1129  *    *next* available buffer (non-NULL). They are incremented after
1130  *    removing buffers, and decremented *before* returning them.
1131  */
1132 struct rpcrdma_req *
1133 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1134 {
1135 	struct rpcrdma_req *req;
1136 	unsigned long flags;
1137 
1138 	spin_lock_irqsave(&buffers->rb_lock, flags);
1139 
1140 	if (buffers->rb_send_index == buffers->rb_max_requests) {
1141 		spin_unlock_irqrestore(&buffers->rb_lock, flags);
1142 		dprintk("RPC:       %s: out of request buffers\n", __func__);
1143 		return ((struct rpcrdma_req *)NULL);
1144 	}
1145 
1146 	req = buffers->rb_send_bufs[buffers->rb_send_index];
1147 	if (buffers->rb_send_index < buffers->rb_recv_index) {
1148 		dprintk("RPC:       %s: %d extra receives outstanding (ok)\n",
1149 			__func__,
1150 			buffers->rb_recv_index - buffers->rb_send_index);
1151 		req->rl_reply = NULL;
1152 	} else {
1153 		req->rl_reply = buffers->rb_recv_bufs[buffers->rb_recv_index];
1154 		buffers->rb_recv_bufs[buffers->rb_recv_index++] = NULL;
1155 	}
1156 	buffers->rb_send_bufs[buffers->rb_send_index++] = NULL;
1157 
1158 	spin_unlock_irqrestore(&buffers->rb_lock, flags);
1159 	return req;
1160 }
1161 
1162 /*
1163  * Put request/reply buffers back into pool.
1164  * Pre-decrement counter/array index.
1165  */
1166 void
1167 rpcrdma_buffer_put(struct rpcrdma_req *req)
1168 {
1169 	struct rpcrdma_buffer *buffers = req->rl_buffer;
1170 	unsigned long flags;
1171 
1172 	spin_lock_irqsave(&buffers->rb_lock, flags);
1173 	rpcrdma_buffer_put_sendbuf(req, buffers);
1174 	spin_unlock_irqrestore(&buffers->rb_lock, flags);
1175 }
1176 
1177 /*
1178  * Recover reply buffers from pool.
1179  * This happens when recovering from error conditions.
1180  * Post-increment counter/array index.
1181  */
1182 void
1183 rpcrdma_recv_buffer_get(struct rpcrdma_req *req)
1184 {
1185 	struct rpcrdma_buffer *buffers = req->rl_buffer;
1186 	unsigned long flags;
1187 
1188 	spin_lock_irqsave(&buffers->rb_lock, flags);
1189 	if (buffers->rb_recv_index < buffers->rb_max_requests) {
1190 		req->rl_reply = buffers->rb_recv_bufs[buffers->rb_recv_index];
1191 		buffers->rb_recv_bufs[buffers->rb_recv_index++] = NULL;
1192 	}
1193 	spin_unlock_irqrestore(&buffers->rb_lock, flags);
1194 }
1195 
1196 /*
1197  * Put reply buffers back into pool when not attached to
1198  * request. This happens in error conditions.
1199  */
1200 void
1201 rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1202 {
1203 	struct rpcrdma_buffer *buffers = &rep->rr_rxprt->rx_buf;
1204 	unsigned long flags;
1205 
1206 	spin_lock_irqsave(&buffers->rb_lock, flags);
1207 	buffers->rb_recv_bufs[--buffers->rb_recv_index] = rep;
1208 	spin_unlock_irqrestore(&buffers->rb_lock, flags);
1209 }
1210 
1211 /*
1212  * Wrappers for internal-use kmalloc memory registration, used by buffer code.
1213  */
1214 
1215 void
1216 rpcrdma_mapping_error(struct rpcrdma_mr_seg *seg)
1217 {
1218 	dprintk("RPC:       map_one: offset %p iova %llx len %zu\n",
1219 		seg->mr_offset,
1220 		(unsigned long long)seg->mr_dma, seg->mr_dmalen);
1221 }
1222 
1223 /**
1224  * rpcrdma_alloc_regbuf - kmalloc and register memory for SEND/RECV buffers
1225  * @ia: controlling rpcrdma_ia
1226  * @size: size of buffer to be allocated, in bytes
1227  * @flags: GFP flags
1228  *
1229  * Returns pointer to private header of an area of internally
1230  * registered memory, or an ERR_PTR. The registered buffer follows
1231  * the end of the private header.
1232  *
1233  * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1234  * receiving the payload of RDMA RECV operations. regbufs are not
1235  * used for RDMA READ/WRITE operations, thus are registered only for
1236  * LOCAL access.
1237  */
1238 struct rpcrdma_regbuf *
1239 rpcrdma_alloc_regbuf(struct rpcrdma_ia *ia, size_t size, gfp_t flags)
1240 {
1241 	struct rpcrdma_regbuf *rb;
1242 	struct ib_sge *iov;
1243 
1244 	rb = kmalloc(sizeof(*rb) + size, flags);
1245 	if (rb == NULL)
1246 		goto out;
1247 
1248 	iov = &rb->rg_iov;
1249 	iov->addr = ib_dma_map_single(ia->ri_device,
1250 				      (void *)rb->rg_base, size,
1251 				      DMA_BIDIRECTIONAL);
1252 	if (ib_dma_mapping_error(ia->ri_device, iov->addr))
1253 		goto out_free;
1254 
1255 	iov->length = size;
1256 	iov->lkey = ia->ri_pd->local_dma_lkey;
1257 	rb->rg_size = size;
1258 	rb->rg_owner = NULL;
1259 	return rb;
1260 
1261 out_free:
1262 	kfree(rb);
1263 out:
1264 	return ERR_PTR(-ENOMEM);
1265 }
1266 
1267 /**
1268  * rpcrdma_free_regbuf - deregister and free registered buffer
1269  * @ia: controlling rpcrdma_ia
1270  * @rb: regbuf to be deregistered and freed
1271  */
1272 void
1273 rpcrdma_free_regbuf(struct rpcrdma_ia *ia, struct rpcrdma_regbuf *rb)
1274 {
1275 	struct ib_sge *iov;
1276 
1277 	if (!rb)
1278 		return;
1279 
1280 	iov = &rb->rg_iov;
1281 	ib_dma_unmap_single(ia->ri_device,
1282 			    iov->addr, iov->length, DMA_BIDIRECTIONAL);
1283 	kfree(rb);
1284 }
1285 
1286 /*
1287  * Prepost any receive buffer, then post send.
1288  *
1289  * Receive buffer is donated to hardware, reclaimed upon recv completion.
1290  */
1291 int
1292 rpcrdma_ep_post(struct rpcrdma_ia *ia,
1293 		struct rpcrdma_ep *ep,
1294 		struct rpcrdma_req *req)
1295 {
1296 	struct ib_device *device = ia->ri_device;
1297 	struct ib_send_wr send_wr, *send_wr_fail;
1298 	struct rpcrdma_rep *rep = req->rl_reply;
1299 	struct ib_sge *iov = req->rl_send_iov;
1300 	int i, rc;
1301 
1302 	if (rep) {
1303 		rc = rpcrdma_ep_post_recv(ia, ep, rep);
1304 		if (rc)
1305 			goto out;
1306 		req->rl_reply = NULL;
1307 	}
1308 
1309 	send_wr.next = NULL;
1310 	send_wr.wr_id = RPCRDMA_IGNORE_COMPLETION;
1311 	send_wr.sg_list = iov;
1312 	send_wr.num_sge = req->rl_niovs;
1313 	send_wr.opcode = IB_WR_SEND;
1314 
1315 	for (i = 0; i < send_wr.num_sge; i++)
1316 		ib_dma_sync_single_for_device(device, iov[i].addr,
1317 					      iov[i].length, DMA_TO_DEVICE);
1318 	dprintk("RPC:       %s: posting %d s/g entries\n",
1319 		__func__, send_wr.num_sge);
1320 
1321 	if (DECR_CQCOUNT(ep) > 0)
1322 		send_wr.send_flags = 0;
1323 	else { /* Provider must take a send completion every now and then */
1324 		INIT_CQCOUNT(ep);
1325 		send_wr.send_flags = IB_SEND_SIGNALED;
1326 	}
1327 
1328 	rc = ib_post_send(ia->ri_id->qp, &send_wr, &send_wr_fail);
1329 	if (rc)
1330 		dprintk("RPC:       %s: ib_post_send returned %i\n", __func__,
1331 			rc);
1332 out:
1333 	return rc;
1334 }
1335 
1336 /*
1337  * (Re)post a receive buffer.
1338  */
1339 int
1340 rpcrdma_ep_post_recv(struct rpcrdma_ia *ia,
1341 		     struct rpcrdma_ep *ep,
1342 		     struct rpcrdma_rep *rep)
1343 {
1344 	struct ib_recv_wr recv_wr, *recv_wr_fail;
1345 	int rc;
1346 
1347 	recv_wr.next = NULL;
1348 	recv_wr.wr_id = (u64) (unsigned long) rep;
1349 	recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
1350 	recv_wr.num_sge = 1;
1351 
1352 	ib_dma_sync_single_for_cpu(ia->ri_device,
1353 				   rdmab_addr(rep->rr_rdmabuf),
1354 				   rdmab_length(rep->rr_rdmabuf),
1355 				   DMA_BIDIRECTIONAL);
1356 
1357 	rc = ib_post_recv(ia->ri_id->qp, &recv_wr, &recv_wr_fail);
1358 
1359 	if (rc)
1360 		dprintk("RPC:       %s: ib_post_recv returned %i\n", __func__,
1361 			rc);
1362 	return rc;
1363 }
1364 
1365 /* How many chunk list items fit within our inline buffers?
1366  */
1367 unsigned int
1368 rpcrdma_max_segments(struct rpcrdma_xprt *r_xprt)
1369 {
1370 	struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
1371 	int bytes, segments;
1372 
1373 	bytes = min_t(unsigned int, cdata->inline_wsize, cdata->inline_rsize);
1374 	bytes -= RPCRDMA_HDRLEN_MIN;
1375 	if (bytes < sizeof(struct rpcrdma_segment) * 2) {
1376 		pr_warn("RPC:       %s: inline threshold too small\n",
1377 			__func__);
1378 		return 0;
1379 	}
1380 
1381 	segments = 1 << (fls(bytes / sizeof(struct rpcrdma_segment)) - 1);
1382 	dprintk("RPC:       %s: max chunk list size = %d segments\n",
1383 		__func__, segments);
1384 	return segments;
1385 }
1386