xref: /linux/net/sunrpc/xprtrdma/verbs.c (revision 3e44c471a2dab210f7e9b1e5f7d4d54d52df59eb)
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 
56 #include "xprt_rdma.h"
57 
58 /*
59  * Globals/Macros
60  */
61 
62 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
63 # define RPCDBG_FACILITY	RPCDBG_TRANS
64 #endif
65 
66 /*
67  * internal functions
68  */
69 
70 /*
71  * handle replies in tasklet context, using a single, global list
72  * rdma tasklet function -- just turn around and call the func
73  * for all replies on the list
74  */
75 
76 static DEFINE_SPINLOCK(rpcrdma_tk_lock_g);
77 static LIST_HEAD(rpcrdma_tasklets_g);
78 
79 static void
80 rpcrdma_run_tasklet(unsigned long data)
81 {
82 	struct rpcrdma_rep *rep;
83 	void (*func)(struct rpcrdma_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 		func = rep->rr_func;
93 		rep->rr_func = NULL;
94 		spin_unlock_irqrestore(&rpcrdma_tk_lock_g, flags);
95 
96 		if (func)
97 			func(rep);
98 		else
99 			rpcrdma_recv_buffer_put(rep);
100 
101 		spin_lock_irqsave(&rpcrdma_tk_lock_g, flags);
102 	}
103 	spin_unlock_irqrestore(&rpcrdma_tk_lock_g, flags);
104 }
105 
106 static DECLARE_TASKLET(rpcrdma_tasklet_g, rpcrdma_run_tasklet, 0UL);
107 
108 static void
109 rpcrdma_schedule_tasklet(struct list_head *sched_list)
110 {
111 	unsigned long flags;
112 
113 	spin_lock_irqsave(&rpcrdma_tk_lock_g, flags);
114 	list_splice_tail(sched_list, &rpcrdma_tasklets_g);
115 	spin_unlock_irqrestore(&rpcrdma_tk_lock_g, flags);
116 	tasklet_schedule(&rpcrdma_tasklet_g);
117 }
118 
119 static void
120 rpcrdma_qp_async_error_upcall(struct ib_event *event, void *context)
121 {
122 	struct rpcrdma_ep *ep = context;
123 
124 	pr_err("RPC:       %s: %s on device %s ep %p\n",
125 	       __func__, ib_event_msg(event->event),
126 		event->device->name, context);
127 	if (ep->rep_connected == 1) {
128 		ep->rep_connected = -EIO;
129 		rpcrdma_conn_func(ep);
130 		wake_up_all(&ep->rep_connect_wait);
131 	}
132 }
133 
134 static void
135 rpcrdma_cq_async_error_upcall(struct ib_event *event, void *context)
136 {
137 	struct rpcrdma_ep *ep = context;
138 
139 	pr_err("RPC:       %s: %s on device %s ep %p\n",
140 	       __func__, ib_event_msg(event->event),
141 		event->device->name, context);
142 	if (ep->rep_connected == 1) {
143 		ep->rep_connected = -EIO;
144 		rpcrdma_conn_func(ep);
145 		wake_up_all(&ep->rep_connect_wait);
146 	}
147 }
148 
149 static void
150 rpcrdma_sendcq_process_wc(struct ib_wc *wc)
151 {
152 	/* WARNING: Only wr_id and status are reliable at this point */
153 	if (wc->wr_id == RPCRDMA_IGNORE_COMPLETION) {
154 		if (wc->status != IB_WC_SUCCESS &&
155 		    wc->status != IB_WC_WR_FLUSH_ERR)
156 			pr_err("RPC:       %s: SEND: %s\n",
157 			       __func__, ib_wc_status_msg(wc->status));
158 	} else {
159 		struct rpcrdma_mw *r;
160 
161 		r = (struct rpcrdma_mw *)(unsigned long)wc->wr_id;
162 		r->mw_sendcompletion(wc);
163 	}
164 }
165 
166 static int
167 rpcrdma_sendcq_poll(struct ib_cq *cq, struct rpcrdma_ep *ep)
168 {
169 	struct ib_wc *wcs;
170 	int budget, count, rc;
171 
172 	budget = RPCRDMA_WC_BUDGET / RPCRDMA_POLLSIZE;
173 	do {
174 		wcs = ep->rep_send_wcs;
175 
176 		rc = ib_poll_cq(cq, RPCRDMA_POLLSIZE, wcs);
177 		if (rc <= 0)
178 			return rc;
179 
180 		count = rc;
181 		while (count-- > 0)
182 			rpcrdma_sendcq_process_wc(wcs++);
183 	} while (rc == RPCRDMA_POLLSIZE && --budget);
184 	return 0;
185 }
186 
187 /*
188  * Handle send, fast_reg_mr, and local_inv completions.
189  *
190  * Send events are typically suppressed and thus do not result
191  * in an upcall. Occasionally one is signaled, however. This
192  * prevents the provider's completion queue from wrapping and
193  * losing a completion.
194  */
195 static void
196 rpcrdma_sendcq_upcall(struct ib_cq *cq, void *cq_context)
197 {
198 	struct rpcrdma_ep *ep = (struct rpcrdma_ep *)cq_context;
199 	int rc;
200 
201 	rc = rpcrdma_sendcq_poll(cq, ep);
202 	if (rc) {
203 		dprintk("RPC:       %s: ib_poll_cq failed: %i\n",
204 			__func__, rc);
205 		return;
206 	}
207 
208 	rc = ib_req_notify_cq(cq,
209 			IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS);
210 	if (rc == 0)
211 		return;
212 	if (rc < 0) {
213 		dprintk("RPC:       %s: ib_req_notify_cq failed: %i\n",
214 			__func__, rc);
215 		return;
216 	}
217 
218 	rpcrdma_sendcq_poll(cq, ep);
219 }
220 
221 static void
222 rpcrdma_recvcq_process_wc(struct ib_wc *wc, struct list_head *sched_list)
223 {
224 	struct rpcrdma_rep *rep =
225 			(struct rpcrdma_rep *)(unsigned long)wc->wr_id;
226 
227 	/* WARNING: Only wr_id and status are reliable at this point */
228 	if (wc->status != IB_WC_SUCCESS)
229 		goto out_fail;
230 
231 	/* status == SUCCESS means all fields in wc are trustworthy */
232 	if (wc->opcode != IB_WC_RECV)
233 		return;
234 
235 	dprintk("RPC:       %s: rep %p opcode 'recv', length %u: success\n",
236 		__func__, rep, wc->byte_len);
237 
238 	rep->rr_len = wc->byte_len;
239 	ib_dma_sync_single_for_cpu(rdmab_to_ia(rep->rr_buffer)->ri_id->device,
240 				   rdmab_addr(rep->rr_rdmabuf),
241 				   rep->rr_len, DMA_FROM_DEVICE);
242 	prefetch(rdmab_to_msg(rep->rr_rdmabuf));
243 
244 out_schedule:
245 	list_add_tail(&rep->rr_list, sched_list);
246 	return;
247 out_fail:
248 	if (wc->status != IB_WC_WR_FLUSH_ERR)
249 		pr_err("RPC:       %s: rep %p: %s\n",
250 		       __func__, rep, ib_wc_status_msg(wc->status));
251 	rep->rr_len = ~0U;
252 	goto out_schedule;
253 }
254 
255 static int
256 rpcrdma_recvcq_poll(struct ib_cq *cq, struct rpcrdma_ep *ep)
257 {
258 	struct list_head sched_list;
259 	struct ib_wc *wcs;
260 	int budget, count, rc;
261 
262 	INIT_LIST_HEAD(&sched_list);
263 	budget = RPCRDMA_WC_BUDGET / RPCRDMA_POLLSIZE;
264 	do {
265 		wcs = ep->rep_recv_wcs;
266 
267 		rc = ib_poll_cq(cq, RPCRDMA_POLLSIZE, wcs);
268 		if (rc <= 0)
269 			goto out_schedule;
270 
271 		count = rc;
272 		while (count-- > 0)
273 			rpcrdma_recvcq_process_wc(wcs++, &sched_list);
274 	} while (rc == RPCRDMA_POLLSIZE && --budget);
275 	rc = 0;
276 
277 out_schedule:
278 	rpcrdma_schedule_tasklet(&sched_list);
279 	return rc;
280 }
281 
282 /*
283  * Handle receive completions.
284  *
285  * It is reentrant but processes single events in order to maintain
286  * ordering of receives to keep server credits.
287  *
288  * It is the responsibility of the scheduled tasklet to return
289  * recv buffers to the pool. NOTE: this affects synchronization of
290  * connection shutdown. That is, the structures required for
291  * the completion of the reply handler must remain intact until
292  * all memory has been reclaimed.
293  */
294 static void
295 rpcrdma_recvcq_upcall(struct ib_cq *cq, void *cq_context)
296 {
297 	struct rpcrdma_ep *ep = (struct rpcrdma_ep *)cq_context;
298 	int rc;
299 
300 	rc = rpcrdma_recvcq_poll(cq, ep);
301 	if (rc) {
302 		dprintk("RPC:       %s: ib_poll_cq failed: %i\n",
303 			__func__, rc);
304 		return;
305 	}
306 
307 	rc = ib_req_notify_cq(cq,
308 			IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS);
309 	if (rc == 0)
310 		return;
311 	if (rc < 0) {
312 		dprintk("RPC:       %s: ib_req_notify_cq failed: %i\n",
313 			__func__, rc);
314 		return;
315 	}
316 
317 	rpcrdma_recvcq_poll(cq, ep);
318 }
319 
320 static void
321 rpcrdma_flush_cqs(struct rpcrdma_ep *ep)
322 {
323 	struct ib_wc wc;
324 	LIST_HEAD(sched_list);
325 
326 	while (ib_poll_cq(ep->rep_attr.recv_cq, 1, &wc) > 0)
327 		rpcrdma_recvcq_process_wc(&wc, &sched_list);
328 	if (!list_empty(&sched_list))
329 		rpcrdma_schedule_tasklet(&sched_list);
330 	while (ib_poll_cq(ep->rep_attr.send_cq, 1, &wc) > 0)
331 		rpcrdma_sendcq_process_wc(&wc);
332 }
333 
334 static int
335 rpcrdma_conn_upcall(struct rdma_cm_id *id, struct rdma_cm_event *event)
336 {
337 	struct rpcrdma_xprt *xprt = id->context;
338 	struct rpcrdma_ia *ia = &xprt->rx_ia;
339 	struct rpcrdma_ep *ep = &xprt->rx_ep;
340 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
341 	struct sockaddr *sap = (struct sockaddr *)&ep->rep_remote_addr;
342 #endif
343 	struct ib_qp_attr *attr = &ia->ri_qp_attr;
344 	struct ib_qp_init_attr *iattr = &ia->ri_qp_init_attr;
345 	int connstate = 0;
346 
347 	switch (event->event) {
348 	case RDMA_CM_EVENT_ADDR_RESOLVED:
349 	case RDMA_CM_EVENT_ROUTE_RESOLVED:
350 		ia->ri_async_rc = 0;
351 		complete(&ia->ri_done);
352 		break;
353 	case RDMA_CM_EVENT_ADDR_ERROR:
354 		ia->ri_async_rc = -EHOSTUNREACH;
355 		dprintk("RPC:       %s: CM address resolution error, ep 0x%p\n",
356 			__func__, ep);
357 		complete(&ia->ri_done);
358 		break;
359 	case RDMA_CM_EVENT_ROUTE_ERROR:
360 		ia->ri_async_rc = -ENETUNREACH;
361 		dprintk("RPC:       %s: CM route resolution error, ep 0x%p\n",
362 			__func__, ep);
363 		complete(&ia->ri_done);
364 		break;
365 	case RDMA_CM_EVENT_ESTABLISHED:
366 		connstate = 1;
367 		ib_query_qp(ia->ri_id->qp, attr,
368 			    IB_QP_MAX_QP_RD_ATOMIC | IB_QP_MAX_DEST_RD_ATOMIC,
369 			    iattr);
370 		dprintk("RPC:       %s: %d responder resources"
371 			" (%d initiator)\n",
372 			__func__, attr->max_dest_rd_atomic,
373 			attr->max_rd_atomic);
374 		goto connected;
375 	case RDMA_CM_EVENT_CONNECT_ERROR:
376 		connstate = -ENOTCONN;
377 		goto connected;
378 	case RDMA_CM_EVENT_UNREACHABLE:
379 		connstate = -ENETDOWN;
380 		goto connected;
381 	case RDMA_CM_EVENT_REJECTED:
382 		connstate = -ECONNREFUSED;
383 		goto connected;
384 	case RDMA_CM_EVENT_DISCONNECTED:
385 		connstate = -ECONNABORTED;
386 		goto connected;
387 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
388 		connstate = -ENODEV;
389 connected:
390 		dprintk("RPC:       %s: %sconnected\n",
391 					__func__, connstate > 0 ? "" : "dis");
392 		ep->rep_connected = connstate;
393 		rpcrdma_conn_func(ep);
394 		wake_up_all(&ep->rep_connect_wait);
395 		/*FALLTHROUGH*/
396 	default:
397 		dprintk("RPC:       %s: %pIS:%u (ep 0x%p): %s\n",
398 			__func__, sap, rpc_get_port(sap), ep,
399 			rdma_event_msg(event->event));
400 		break;
401 	}
402 
403 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
404 	if (connstate == 1) {
405 		int ird = attr->max_dest_rd_atomic;
406 		int tird = ep->rep_remote_cma.responder_resources;
407 
408 		pr_info("rpcrdma: connection to %pIS:%u on %s, memreg '%s', %d credits, %d responders%s\n",
409 			sap, rpc_get_port(sap),
410 			ia->ri_id->device->name,
411 			ia->ri_ops->ro_displayname,
412 			xprt->rx_buf.rb_max_requests,
413 			ird, ird < 4 && ird < tird / 2 ? " (low!)" : "");
414 	} else if (connstate < 0) {
415 		pr_info("rpcrdma: connection to %pIS:%u closed (%d)\n",
416 			sap, rpc_get_port(sap), connstate);
417 	}
418 #endif
419 
420 	return 0;
421 }
422 
423 static struct rdma_cm_id *
424 rpcrdma_create_id(struct rpcrdma_xprt *xprt,
425 			struct rpcrdma_ia *ia, struct sockaddr *addr)
426 {
427 	struct rdma_cm_id *id;
428 	int rc;
429 
430 	init_completion(&ia->ri_done);
431 
432 	id = rdma_create_id(rpcrdma_conn_upcall, xprt, RDMA_PS_TCP, IB_QPT_RC);
433 	if (IS_ERR(id)) {
434 		rc = PTR_ERR(id);
435 		dprintk("RPC:       %s: rdma_create_id() failed %i\n",
436 			__func__, rc);
437 		return id;
438 	}
439 
440 	ia->ri_async_rc = -ETIMEDOUT;
441 	rc = rdma_resolve_addr(id, NULL, addr, RDMA_RESOLVE_TIMEOUT);
442 	if (rc) {
443 		dprintk("RPC:       %s: rdma_resolve_addr() failed %i\n",
444 			__func__, rc);
445 		goto out;
446 	}
447 	wait_for_completion_interruptible_timeout(&ia->ri_done,
448 				msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1);
449 	rc = ia->ri_async_rc;
450 	if (rc)
451 		goto out;
452 
453 	ia->ri_async_rc = -ETIMEDOUT;
454 	rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
455 	if (rc) {
456 		dprintk("RPC:       %s: rdma_resolve_route() failed %i\n",
457 			__func__, rc);
458 		goto out;
459 	}
460 	wait_for_completion_interruptible_timeout(&ia->ri_done,
461 				msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1);
462 	rc = ia->ri_async_rc;
463 	if (rc)
464 		goto out;
465 
466 	return id;
467 
468 out:
469 	rdma_destroy_id(id);
470 	return ERR_PTR(rc);
471 }
472 
473 /*
474  * Drain any cq, prior to teardown.
475  */
476 static void
477 rpcrdma_clean_cq(struct ib_cq *cq)
478 {
479 	struct ib_wc wc;
480 	int count = 0;
481 
482 	while (1 == ib_poll_cq(cq, 1, &wc))
483 		++count;
484 
485 	if (count)
486 		dprintk("RPC:       %s: flushed %d events (last 0x%x)\n",
487 			__func__, count, wc.opcode);
488 }
489 
490 /*
491  * Exported functions.
492  */
493 
494 /*
495  * Open and initialize an Interface Adapter.
496  *  o initializes fields of struct rpcrdma_ia, including
497  *    interface and provider attributes and protection zone.
498  */
499 int
500 rpcrdma_ia_open(struct rpcrdma_xprt *xprt, struct sockaddr *addr, int memreg)
501 {
502 	int rc, mem_priv;
503 	struct rpcrdma_ia *ia = &xprt->rx_ia;
504 	struct ib_device_attr *devattr = &ia->ri_devattr;
505 
506 	ia->ri_id = rpcrdma_create_id(xprt, ia, addr);
507 	if (IS_ERR(ia->ri_id)) {
508 		rc = PTR_ERR(ia->ri_id);
509 		goto out1;
510 	}
511 
512 	ia->ri_pd = ib_alloc_pd(ia->ri_id->device);
513 	if (IS_ERR(ia->ri_pd)) {
514 		rc = PTR_ERR(ia->ri_pd);
515 		dprintk("RPC:       %s: ib_alloc_pd() failed %i\n",
516 			__func__, rc);
517 		goto out2;
518 	}
519 
520 	rc = ib_query_device(ia->ri_id->device, devattr);
521 	if (rc) {
522 		dprintk("RPC:       %s: ib_query_device failed %d\n",
523 			__func__, rc);
524 		goto out3;
525 	}
526 
527 	if (devattr->device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY) {
528 		ia->ri_have_dma_lkey = 1;
529 		ia->ri_dma_lkey = ia->ri_id->device->local_dma_lkey;
530 	}
531 
532 	if (memreg == RPCRDMA_FRMR) {
533 		/* Requires both frmr reg and local dma lkey */
534 		if (((devattr->device_cap_flags &
535 		     (IB_DEVICE_MEM_MGT_EXTENSIONS|IB_DEVICE_LOCAL_DMA_LKEY)) !=
536 		    (IB_DEVICE_MEM_MGT_EXTENSIONS|IB_DEVICE_LOCAL_DMA_LKEY)) ||
537 		      (devattr->max_fast_reg_page_list_len == 0)) {
538 			dprintk("RPC:       %s: FRMR registration "
539 				"not supported by HCA\n", __func__);
540 			memreg = RPCRDMA_MTHCAFMR;
541 		}
542 	}
543 	if (memreg == RPCRDMA_MTHCAFMR) {
544 		if (!ia->ri_id->device->alloc_fmr) {
545 			dprintk("RPC:       %s: MTHCAFMR registration "
546 				"not supported by HCA\n", __func__);
547 			memreg = RPCRDMA_ALLPHYSICAL;
548 		}
549 	}
550 
551 	/*
552 	 * Optionally obtain an underlying physical identity mapping in
553 	 * order to do a memory window-based bind. This base registration
554 	 * is protected from remote access - that is enabled only by binding
555 	 * for the specific bytes targeted during each RPC operation, and
556 	 * revoked after the corresponding completion similar to a storage
557 	 * adapter.
558 	 */
559 	switch (memreg) {
560 	case RPCRDMA_FRMR:
561 		ia->ri_ops = &rpcrdma_frwr_memreg_ops;
562 		break;
563 	case RPCRDMA_ALLPHYSICAL:
564 		ia->ri_ops = &rpcrdma_physical_memreg_ops;
565 		mem_priv = IB_ACCESS_LOCAL_WRITE |
566 				IB_ACCESS_REMOTE_WRITE |
567 				IB_ACCESS_REMOTE_READ;
568 		goto register_setup;
569 	case RPCRDMA_MTHCAFMR:
570 		ia->ri_ops = &rpcrdma_fmr_memreg_ops;
571 		if (ia->ri_have_dma_lkey)
572 			break;
573 		mem_priv = IB_ACCESS_LOCAL_WRITE;
574 	register_setup:
575 		ia->ri_bind_mem = ib_get_dma_mr(ia->ri_pd, mem_priv);
576 		if (IS_ERR(ia->ri_bind_mem)) {
577 			printk(KERN_ALERT "%s: ib_get_dma_mr for "
578 				"phys register failed with %lX\n",
579 				__func__, PTR_ERR(ia->ri_bind_mem));
580 			rc = -ENOMEM;
581 			goto out3;
582 		}
583 		break;
584 	default:
585 		printk(KERN_ERR "RPC: Unsupported memory "
586 				"registration mode: %d\n", memreg);
587 		rc = -ENOMEM;
588 		goto out3;
589 	}
590 	dprintk("RPC:       %s: memory registration strategy is '%s'\n",
591 		__func__, ia->ri_ops->ro_displayname);
592 
593 	/* Else will do memory reg/dereg for each chunk */
594 	ia->ri_memreg_strategy = memreg;
595 
596 	rwlock_init(&ia->ri_qplock);
597 	return 0;
598 
599 out3:
600 	ib_dealloc_pd(ia->ri_pd);
601 	ia->ri_pd = NULL;
602 out2:
603 	rdma_destroy_id(ia->ri_id);
604 	ia->ri_id = NULL;
605 out1:
606 	return rc;
607 }
608 
609 /*
610  * Clean up/close an IA.
611  *   o if event handles and PD have been initialized, free them.
612  *   o close the IA
613  */
614 void
615 rpcrdma_ia_close(struct rpcrdma_ia *ia)
616 {
617 	int rc;
618 
619 	dprintk("RPC:       %s: entering\n", __func__);
620 	if (ia->ri_bind_mem != NULL) {
621 		rc = ib_dereg_mr(ia->ri_bind_mem);
622 		dprintk("RPC:       %s: ib_dereg_mr returned %i\n",
623 			__func__, rc);
624 	}
625 	if (ia->ri_id != NULL && !IS_ERR(ia->ri_id)) {
626 		if (ia->ri_id->qp)
627 			rdma_destroy_qp(ia->ri_id);
628 		rdma_destroy_id(ia->ri_id);
629 		ia->ri_id = NULL;
630 	}
631 	if (ia->ri_pd != NULL && !IS_ERR(ia->ri_pd)) {
632 		rc = ib_dealloc_pd(ia->ri_pd);
633 		dprintk("RPC:       %s: ib_dealloc_pd returned %i\n",
634 			__func__, rc);
635 	}
636 }
637 
638 /*
639  * Create unconnected endpoint.
640  */
641 int
642 rpcrdma_ep_create(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia,
643 				struct rpcrdma_create_data_internal *cdata)
644 {
645 	struct ib_device_attr *devattr = &ia->ri_devattr;
646 	struct ib_cq *sendcq, *recvcq;
647 	struct ib_cq_init_attr cq_attr = {};
648 	int rc, err;
649 
650 	/* check provider's send/recv wr limits */
651 	if (cdata->max_requests > devattr->max_qp_wr)
652 		cdata->max_requests = devattr->max_qp_wr;
653 
654 	ep->rep_attr.event_handler = rpcrdma_qp_async_error_upcall;
655 	ep->rep_attr.qp_context = ep;
656 	ep->rep_attr.srq = NULL;
657 	ep->rep_attr.cap.max_send_wr = cdata->max_requests;
658 	rc = ia->ri_ops->ro_open(ia, ep, cdata);
659 	if (rc)
660 		return rc;
661 	ep->rep_attr.cap.max_recv_wr = cdata->max_requests;
662 	ep->rep_attr.cap.max_send_sge = (cdata->padding ? 4 : 2);
663 	ep->rep_attr.cap.max_recv_sge = 1;
664 	ep->rep_attr.cap.max_inline_data = 0;
665 	ep->rep_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
666 	ep->rep_attr.qp_type = IB_QPT_RC;
667 	ep->rep_attr.port_num = ~0;
668 
669 	if (cdata->padding) {
670 		ep->rep_padbuf = rpcrdma_alloc_regbuf(ia, cdata->padding,
671 						      GFP_KERNEL);
672 		if (IS_ERR(ep->rep_padbuf))
673 			return PTR_ERR(ep->rep_padbuf);
674 	} else
675 		ep->rep_padbuf = NULL;
676 
677 	dprintk("RPC:       %s: requested max: dtos: send %d recv %d; "
678 		"iovs: send %d recv %d\n",
679 		__func__,
680 		ep->rep_attr.cap.max_send_wr,
681 		ep->rep_attr.cap.max_recv_wr,
682 		ep->rep_attr.cap.max_send_sge,
683 		ep->rep_attr.cap.max_recv_sge);
684 
685 	/* set trigger for requesting send completion */
686 	ep->rep_cqinit = ep->rep_attr.cap.max_send_wr/2 - 1;
687 	if (ep->rep_cqinit > RPCRDMA_MAX_UNSIGNALED_SENDS)
688 		ep->rep_cqinit = RPCRDMA_MAX_UNSIGNALED_SENDS;
689 	else if (ep->rep_cqinit <= 2)
690 		ep->rep_cqinit = 0;
691 	INIT_CQCOUNT(ep);
692 	init_waitqueue_head(&ep->rep_connect_wait);
693 	INIT_DELAYED_WORK(&ep->rep_connect_worker, rpcrdma_connect_worker);
694 
695 	cq_attr.cqe = ep->rep_attr.cap.max_send_wr + 1;
696 	sendcq = ib_create_cq(ia->ri_id->device, rpcrdma_sendcq_upcall,
697 				  rpcrdma_cq_async_error_upcall, ep, &cq_attr);
698 	if (IS_ERR(sendcq)) {
699 		rc = PTR_ERR(sendcq);
700 		dprintk("RPC:       %s: failed to create send CQ: %i\n",
701 			__func__, rc);
702 		goto out1;
703 	}
704 
705 	rc = ib_req_notify_cq(sendcq, IB_CQ_NEXT_COMP);
706 	if (rc) {
707 		dprintk("RPC:       %s: ib_req_notify_cq failed: %i\n",
708 			__func__, rc);
709 		goto out2;
710 	}
711 
712 	cq_attr.cqe = ep->rep_attr.cap.max_recv_wr + 1;
713 	recvcq = ib_create_cq(ia->ri_id->device, rpcrdma_recvcq_upcall,
714 				  rpcrdma_cq_async_error_upcall, ep, &cq_attr);
715 	if (IS_ERR(recvcq)) {
716 		rc = PTR_ERR(recvcq);
717 		dprintk("RPC:       %s: failed to create recv CQ: %i\n",
718 			__func__, rc);
719 		goto out2;
720 	}
721 
722 	rc = ib_req_notify_cq(recvcq, IB_CQ_NEXT_COMP);
723 	if (rc) {
724 		dprintk("RPC:       %s: ib_req_notify_cq failed: %i\n",
725 			__func__, rc);
726 		ib_destroy_cq(recvcq);
727 		goto out2;
728 	}
729 
730 	ep->rep_attr.send_cq = sendcq;
731 	ep->rep_attr.recv_cq = recvcq;
732 
733 	/* Initialize cma parameters */
734 
735 	/* RPC/RDMA does not use private data */
736 	ep->rep_remote_cma.private_data = NULL;
737 	ep->rep_remote_cma.private_data_len = 0;
738 
739 	/* Client offers RDMA Read but does not initiate */
740 	ep->rep_remote_cma.initiator_depth = 0;
741 	if (devattr->max_qp_rd_atom > 32)	/* arbitrary but <= 255 */
742 		ep->rep_remote_cma.responder_resources = 32;
743 	else
744 		ep->rep_remote_cma.responder_resources =
745 						devattr->max_qp_rd_atom;
746 
747 	ep->rep_remote_cma.retry_count = 7;
748 	ep->rep_remote_cma.flow_control = 0;
749 	ep->rep_remote_cma.rnr_retry_count = 0;
750 
751 	return 0;
752 
753 out2:
754 	err = ib_destroy_cq(sendcq);
755 	if (err)
756 		dprintk("RPC:       %s: ib_destroy_cq returned %i\n",
757 			__func__, err);
758 out1:
759 	rpcrdma_free_regbuf(ia, ep->rep_padbuf);
760 	return rc;
761 }
762 
763 /*
764  * rpcrdma_ep_destroy
765  *
766  * Disconnect and destroy endpoint. After this, the only
767  * valid operations on the ep are to free it (if dynamically
768  * allocated) or re-create it.
769  */
770 void
771 rpcrdma_ep_destroy(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
772 {
773 	int rc;
774 
775 	dprintk("RPC:       %s: entering, connected is %d\n",
776 		__func__, ep->rep_connected);
777 
778 	cancel_delayed_work_sync(&ep->rep_connect_worker);
779 
780 	if (ia->ri_id->qp) {
781 		rpcrdma_ep_disconnect(ep, ia);
782 		rdma_destroy_qp(ia->ri_id);
783 		ia->ri_id->qp = NULL;
784 	}
785 
786 	rpcrdma_free_regbuf(ia, ep->rep_padbuf);
787 
788 	rpcrdma_clean_cq(ep->rep_attr.recv_cq);
789 	rc = ib_destroy_cq(ep->rep_attr.recv_cq);
790 	if (rc)
791 		dprintk("RPC:       %s: ib_destroy_cq returned %i\n",
792 			__func__, rc);
793 
794 	rpcrdma_clean_cq(ep->rep_attr.send_cq);
795 	rc = ib_destroy_cq(ep->rep_attr.send_cq);
796 	if (rc)
797 		dprintk("RPC:       %s: ib_destroy_cq returned %i\n",
798 			__func__, rc);
799 }
800 
801 /*
802  * Connect unconnected endpoint.
803  */
804 int
805 rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
806 {
807 	struct rdma_cm_id *id, *old;
808 	int rc = 0;
809 	int retry_count = 0;
810 
811 	if (ep->rep_connected != 0) {
812 		struct rpcrdma_xprt *xprt;
813 retry:
814 		dprintk("RPC:       %s: reconnecting...\n", __func__);
815 
816 		rpcrdma_ep_disconnect(ep, ia);
817 		rpcrdma_flush_cqs(ep);
818 
819 		xprt = container_of(ia, struct rpcrdma_xprt, rx_ia);
820 		ia->ri_ops->ro_reset(xprt);
821 
822 		id = rpcrdma_create_id(xprt, ia,
823 				(struct sockaddr *)&xprt->rx_data.addr);
824 		if (IS_ERR(id)) {
825 			rc = -EHOSTUNREACH;
826 			goto out;
827 		}
828 		/* TEMP TEMP TEMP - fail if new device:
829 		 * Deregister/remarshal *all* requests!
830 		 * Close and recreate adapter, pd, etc!
831 		 * Re-determine all attributes still sane!
832 		 * More stuff I haven't thought of!
833 		 * Rrrgh!
834 		 */
835 		if (ia->ri_id->device != id->device) {
836 			printk("RPC:       %s: can't reconnect on "
837 				"different device!\n", __func__);
838 			rdma_destroy_id(id);
839 			rc = -ENETUNREACH;
840 			goto out;
841 		}
842 		/* END TEMP */
843 		rc = rdma_create_qp(id, ia->ri_pd, &ep->rep_attr);
844 		if (rc) {
845 			dprintk("RPC:       %s: rdma_create_qp failed %i\n",
846 				__func__, rc);
847 			rdma_destroy_id(id);
848 			rc = -ENETUNREACH;
849 			goto out;
850 		}
851 
852 		write_lock(&ia->ri_qplock);
853 		old = ia->ri_id;
854 		ia->ri_id = id;
855 		write_unlock(&ia->ri_qplock);
856 
857 		rdma_destroy_qp(old);
858 		rdma_destroy_id(old);
859 	} else {
860 		dprintk("RPC:       %s: connecting...\n", __func__);
861 		rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
862 		if (rc) {
863 			dprintk("RPC:       %s: rdma_create_qp failed %i\n",
864 				__func__, rc);
865 			/* do not update ep->rep_connected */
866 			return -ENETUNREACH;
867 		}
868 	}
869 
870 	ep->rep_connected = 0;
871 
872 	rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma);
873 	if (rc) {
874 		dprintk("RPC:       %s: rdma_connect() failed with %i\n",
875 				__func__, rc);
876 		goto out;
877 	}
878 
879 	wait_event_interruptible(ep->rep_connect_wait, ep->rep_connected != 0);
880 
881 	/*
882 	 * Check state. A non-peer reject indicates no listener
883 	 * (ECONNREFUSED), which may be a transient state. All
884 	 * others indicate a transport condition which has already
885 	 * undergone a best-effort.
886 	 */
887 	if (ep->rep_connected == -ECONNREFUSED &&
888 	    ++retry_count <= RDMA_CONNECT_RETRY_MAX) {
889 		dprintk("RPC:       %s: non-peer_reject, retry\n", __func__);
890 		goto retry;
891 	}
892 	if (ep->rep_connected <= 0) {
893 		/* Sometimes, the only way to reliably connect to remote
894 		 * CMs is to use same nonzero values for ORD and IRD. */
895 		if (retry_count++ <= RDMA_CONNECT_RETRY_MAX + 1 &&
896 		    (ep->rep_remote_cma.responder_resources == 0 ||
897 		     ep->rep_remote_cma.initiator_depth !=
898 				ep->rep_remote_cma.responder_resources)) {
899 			if (ep->rep_remote_cma.responder_resources == 0)
900 				ep->rep_remote_cma.responder_resources = 1;
901 			ep->rep_remote_cma.initiator_depth =
902 				ep->rep_remote_cma.responder_resources;
903 			goto retry;
904 		}
905 		rc = ep->rep_connected;
906 	} else {
907 		dprintk("RPC:       %s: connected\n", __func__);
908 	}
909 
910 out:
911 	if (rc)
912 		ep->rep_connected = rc;
913 	return rc;
914 }
915 
916 /*
917  * rpcrdma_ep_disconnect
918  *
919  * This is separate from destroy to facilitate the ability
920  * to reconnect without recreating the endpoint.
921  *
922  * This call is not reentrant, and must not be made in parallel
923  * on the same endpoint.
924  */
925 void
926 rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
927 {
928 	int rc;
929 
930 	rpcrdma_flush_cqs(ep);
931 	rc = rdma_disconnect(ia->ri_id);
932 	if (!rc) {
933 		/* returns without wait if not connected */
934 		wait_event_interruptible(ep->rep_connect_wait,
935 							ep->rep_connected != 1);
936 		dprintk("RPC:       %s: after wait, %sconnected\n", __func__,
937 			(ep->rep_connected == 1) ? "still " : "dis");
938 	} else {
939 		dprintk("RPC:       %s: rdma_disconnect %i\n", __func__, rc);
940 		ep->rep_connected = rc;
941 	}
942 }
943 
944 static struct rpcrdma_req *
945 rpcrdma_create_req(struct rpcrdma_xprt *r_xprt)
946 {
947 	struct rpcrdma_req *req;
948 
949 	req = kzalloc(sizeof(*req), GFP_KERNEL);
950 	if (req == NULL)
951 		return ERR_PTR(-ENOMEM);
952 
953 	req->rl_buffer = &r_xprt->rx_buf;
954 	return req;
955 }
956 
957 static struct rpcrdma_rep *
958 rpcrdma_create_rep(struct rpcrdma_xprt *r_xprt)
959 {
960 	struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
961 	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
962 	struct rpcrdma_rep *rep;
963 	int rc;
964 
965 	rc = -ENOMEM;
966 	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
967 	if (rep == NULL)
968 		goto out;
969 
970 	rep->rr_rdmabuf = rpcrdma_alloc_regbuf(ia, cdata->inline_rsize,
971 					       GFP_KERNEL);
972 	if (IS_ERR(rep->rr_rdmabuf)) {
973 		rc = PTR_ERR(rep->rr_rdmabuf);
974 		goto out_free;
975 	}
976 
977 	rep->rr_buffer = &r_xprt->rx_buf;
978 	return rep;
979 
980 out_free:
981 	kfree(rep);
982 out:
983 	return ERR_PTR(rc);
984 }
985 
986 int
987 rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
988 {
989 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
990 	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
991 	struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
992 	char *p;
993 	size_t len;
994 	int i, rc;
995 
996 	buf->rb_max_requests = cdata->max_requests;
997 	spin_lock_init(&buf->rb_lock);
998 
999 	/* Need to allocate:
1000 	 *   1.  arrays for send and recv pointers
1001 	 *   2.  arrays of struct rpcrdma_req to fill in pointers
1002 	 *   3.  array of struct rpcrdma_rep for replies
1003 	 * Send/recv buffers in req/rep need to be registered
1004 	 */
1005 	len = buf->rb_max_requests *
1006 		(sizeof(struct rpcrdma_req *) + sizeof(struct rpcrdma_rep *));
1007 
1008 	p = kzalloc(len, GFP_KERNEL);
1009 	if (p == NULL) {
1010 		dprintk("RPC:       %s: req_t/rep_t/pad kzalloc(%zd) failed\n",
1011 			__func__, len);
1012 		rc = -ENOMEM;
1013 		goto out;
1014 	}
1015 	buf->rb_pool = p;	/* for freeing it later */
1016 
1017 	buf->rb_send_bufs = (struct rpcrdma_req **) p;
1018 	p = (char *) &buf->rb_send_bufs[buf->rb_max_requests];
1019 	buf->rb_recv_bufs = (struct rpcrdma_rep **) p;
1020 	p = (char *) &buf->rb_recv_bufs[buf->rb_max_requests];
1021 
1022 	rc = ia->ri_ops->ro_init(r_xprt);
1023 	if (rc)
1024 		goto out;
1025 
1026 	for (i = 0; i < buf->rb_max_requests; i++) {
1027 		struct rpcrdma_req *req;
1028 		struct rpcrdma_rep *rep;
1029 
1030 		req = rpcrdma_create_req(r_xprt);
1031 		if (IS_ERR(req)) {
1032 			dprintk("RPC:       %s: request buffer %d alloc"
1033 				" failed\n", __func__, i);
1034 			rc = PTR_ERR(req);
1035 			goto out;
1036 		}
1037 		buf->rb_send_bufs[i] = req;
1038 
1039 		rep = rpcrdma_create_rep(r_xprt);
1040 		if (IS_ERR(rep)) {
1041 			dprintk("RPC:       %s: reply buffer %d alloc failed\n",
1042 				__func__, i);
1043 			rc = PTR_ERR(rep);
1044 			goto out;
1045 		}
1046 		buf->rb_recv_bufs[i] = rep;
1047 	}
1048 
1049 	return 0;
1050 out:
1051 	rpcrdma_buffer_destroy(buf);
1052 	return rc;
1053 }
1054 
1055 static void
1056 rpcrdma_destroy_rep(struct rpcrdma_ia *ia, struct rpcrdma_rep *rep)
1057 {
1058 	if (!rep)
1059 		return;
1060 
1061 	rpcrdma_free_regbuf(ia, rep->rr_rdmabuf);
1062 	kfree(rep);
1063 }
1064 
1065 static void
1066 rpcrdma_destroy_req(struct rpcrdma_ia *ia, struct rpcrdma_req *req)
1067 {
1068 	if (!req)
1069 		return;
1070 
1071 	rpcrdma_free_regbuf(ia, req->rl_sendbuf);
1072 	rpcrdma_free_regbuf(ia, req->rl_rdmabuf);
1073 	kfree(req);
1074 }
1075 
1076 void
1077 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1078 {
1079 	struct rpcrdma_ia *ia = rdmab_to_ia(buf);
1080 	int i;
1081 
1082 	/* clean up in reverse order from create
1083 	 *   1.  recv mr memory (mr free, then kfree)
1084 	 *   2.  send mr memory (mr free, then kfree)
1085 	 *   3.  MWs
1086 	 */
1087 	dprintk("RPC:       %s: entering\n", __func__);
1088 
1089 	for (i = 0; i < buf->rb_max_requests; i++) {
1090 		if (buf->rb_recv_bufs)
1091 			rpcrdma_destroy_rep(ia, buf->rb_recv_bufs[i]);
1092 		if (buf->rb_send_bufs)
1093 			rpcrdma_destroy_req(ia, buf->rb_send_bufs[i]);
1094 	}
1095 
1096 	ia->ri_ops->ro_destroy(buf);
1097 
1098 	kfree(buf->rb_pool);
1099 }
1100 
1101 /* "*mw" can be NULL when rpcrdma_buffer_get_mrs() fails, leaving
1102  * some req segments uninitialized.
1103  */
1104 static void
1105 rpcrdma_buffer_put_mr(struct rpcrdma_mw **mw, struct rpcrdma_buffer *buf)
1106 {
1107 	if (*mw) {
1108 		list_add_tail(&(*mw)->mw_list, &buf->rb_mws);
1109 		*mw = NULL;
1110 	}
1111 }
1112 
1113 /* Cycle mw's back in reverse order, and "spin" them.
1114  * This delays and scrambles reuse as much as possible.
1115  */
1116 static void
1117 rpcrdma_buffer_put_mrs(struct rpcrdma_req *req, struct rpcrdma_buffer *buf)
1118 {
1119 	struct rpcrdma_mr_seg *seg = req->rl_segments;
1120 	struct rpcrdma_mr_seg *seg1 = seg;
1121 	int i;
1122 
1123 	for (i = 1, seg++; i < RPCRDMA_MAX_SEGS; seg++, i++)
1124 		rpcrdma_buffer_put_mr(&seg->rl_mw, buf);
1125 	rpcrdma_buffer_put_mr(&seg1->rl_mw, buf);
1126 }
1127 
1128 static void
1129 rpcrdma_buffer_put_sendbuf(struct rpcrdma_req *req, struct rpcrdma_buffer *buf)
1130 {
1131 	buf->rb_send_bufs[--buf->rb_send_index] = req;
1132 	req->rl_niovs = 0;
1133 	if (req->rl_reply) {
1134 		buf->rb_recv_bufs[--buf->rb_recv_index] = req->rl_reply;
1135 		req->rl_reply->rr_func = NULL;
1136 		req->rl_reply = NULL;
1137 	}
1138 }
1139 
1140 /* rpcrdma_unmap_one() was already done during deregistration.
1141  * Redo only the ib_post_send().
1142  */
1143 static void
1144 rpcrdma_retry_local_inv(struct rpcrdma_mw *r, struct rpcrdma_ia *ia)
1145 {
1146 	struct rpcrdma_xprt *r_xprt =
1147 				container_of(ia, struct rpcrdma_xprt, rx_ia);
1148 	struct ib_send_wr invalidate_wr, *bad_wr;
1149 	int rc;
1150 
1151 	dprintk("RPC:       %s: FRMR %p is stale\n", __func__, r);
1152 
1153 	/* When this FRMR is re-inserted into rb_mws, it is no longer stale */
1154 	r->r.frmr.fr_state = FRMR_IS_INVALID;
1155 
1156 	memset(&invalidate_wr, 0, sizeof(invalidate_wr));
1157 	invalidate_wr.wr_id = (unsigned long)(void *)r;
1158 	invalidate_wr.opcode = IB_WR_LOCAL_INV;
1159 	invalidate_wr.ex.invalidate_rkey = r->r.frmr.fr_mr->rkey;
1160 	DECR_CQCOUNT(&r_xprt->rx_ep);
1161 
1162 	dprintk("RPC:       %s: frmr %p invalidating rkey %08x\n",
1163 		__func__, r, r->r.frmr.fr_mr->rkey);
1164 
1165 	read_lock(&ia->ri_qplock);
1166 	rc = ib_post_send(ia->ri_id->qp, &invalidate_wr, &bad_wr);
1167 	read_unlock(&ia->ri_qplock);
1168 	if (rc) {
1169 		/* Force rpcrdma_buffer_get() to retry */
1170 		r->r.frmr.fr_state = FRMR_IS_STALE;
1171 		dprintk("RPC:       %s: ib_post_send failed, %i\n",
1172 			__func__, rc);
1173 	}
1174 }
1175 
1176 static void
1177 rpcrdma_retry_flushed_linv(struct list_head *stale,
1178 			   struct rpcrdma_buffer *buf)
1179 {
1180 	struct rpcrdma_ia *ia = rdmab_to_ia(buf);
1181 	struct list_head *pos;
1182 	struct rpcrdma_mw *r;
1183 	unsigned long flags;
1184 
1185 	list_for_each(pos, stale) {
1186 		r = list_entry(pos, struct rpcrdma_mw, mw_list);
1187 		rpcrdma_retry_local_inv(r, ia);
1188 	}
1189 
1190 	spin_lock_irqsave(&buf->rb_lock, flags);
1191 	list_splice_tail(stale, &buf->rb_mws);
1192 	spin_unlock_irqrestore(&buf->rb_lock, flags);
1193 }
1194 
1195 static struct rpcrdma_req *
1196 rpcrdma_buffer_get_frmrs(struct rpcrdma_req *req, struct rpcrdma_buffer *buf,
1197 			 struct list_head *stale)
1198 {
1199 	struct rpcrdma_mw *r;
1200 	int i;
1201 
1202 	i = RPCRDMA_MAX_SEGS - 1;
1203 	while (!list_empty(&buf->rb_mws)) {
1204 		r = list_entry(buf->rb_mws.next,
1205 			       struct rpcrdma_mw, mw_list);
1206 		list_del(&r->mw_list);
1207 		if (r->r.frmr.fr_state == FRMR_IS_STALE) {
1208 			list_add(&r->mw_list, stale);
1209 			continue;
1210 		}
1211 		req->rl_segments[i].rl_mw = r;
1212 		if (unlikely(i-- == 0))
1213 			return req;	/* Success */
1214 	}
1215 
1216 	/* Not enough entries on rb_mws for this req */
1217 	rpcrdma_buffer_put_sendbuf(req, buf);
1218 	rpcrdma_buffer_put_mrs(req, buf);
1219 	return NULL;
1220 }
1221 
1222 static struct rpcrdma_req *
1223 rpcrdma_buffer_get_fmrs(struct rpcrdma_req *req, struct rpcrdma_buffer *buf)
1224 {
1225 	struct rpcrdma_mw *r;
1226 	int i;
1227 
1228 	i = RPCRDMA_MAX_SEGS - 1;
1229 	while (!list_empty(&buf->rb_mws)) {
1230 		r = list_entry(buf->rb_mws.next,
1231 			       struct rpcrdma_mw, mw_list);
1232 		list_del(&r->mw_list);
1233 		req->rl_segments[i].rl_mw = r;
1234 		if (unlikely(i-- == 0))
1235 			return req;	/* Success */
1236 	}
1237 
1238 	/* Not enough entries on rb_mws for this req */
1239 	rpcrdma_buffer_put_sendbuf(req, buf);
1240 	rpcrdma_buffer_put_mrs(req, buf);
1241 	return NULL;
1242 }
1243 
1244 /*
1245  * Get a set of request/reply buffers.
1246  *
1247  * Reply buffer (if needed) is attached to send buffer upon return.
1248  * Rule:
1249  *    rb_send_index and rb_recv_index MUST always be pointing to the
1250  *    *next* available buffer (non-NULL). They are incremented after
1251  *    removing buffers, and decremented *before* returning them.
1252  */
1253 struct rpcrdma_req *
1254 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1255 {
1256 	struct rpcrdma_ia *ia = rdmab_to_ia(buffers);
1257 	struct list_head stale;
1258 	struct rpcrdma_req *req;
1259 	unsigned long flags;
1260 
1261 	spin_lock_irqsave(&buffers->rb_lock, flags);
1262 	if (buffers->rb_send_index == buffers->rb_max_requests) {
1263 		spin_unlock_irqrestore(&buffers->rb_lock, flags);
1264 		dprintk("RPC:       %s: out of request buffers\n", __func__);
1265 		return ((struct rpcrdma_req *)NULL);
1266 	}
1267 
1268 	req = buffers->rb_send_bufs[buffers->rb_send_index];
1269 	if (buffers->rb_send_index < buffers->rb_recv_index) {
1270 		dprintk("RPC:       %s: %d extra receives outstanding (ok)\n",
1271 			__func__,
1272 			buffers->rb_recv_index - buffers->rb_send_index);
1273 		req->rl_reply = NULL;
1274 	} else {
1275 		req->rl_reply = buffers->rb_recv_bufs[buffers->rb_recv_index];
1276 		buffers->rb_recv_bufs[buffers->rb_recv_index++] = NULL;
1277 	}
1278 	buffers->rb_send_bufs[buffers->rb_send_index++] = NULL;
1279 
1280 	INIT_LIST_HEAD(&stale);
1281 	switch (ia->ri_memreg_strategy) {
1282 	case RPCRDMA_FRMR:
1283 		req = rpcrdma_buffer_get_frmrs(req, buffers, &stale);
1284 		break;
1285 	case RPCRDMA_MTHCAFMR:
1286 		req = rpcrdma_buffer_get_fmrs(req, buffers);
1287 		break;
1288 	default:
1289 		break;
1290 	}
1291 	spin_unlock_irqrestore(&buffers->rb_lock, flags);
1292 	if (!list_empty(&stale))
1293 		rpcrdma_retry_flushed_linv(&stale, buffers);
1294 	return req;
1295 }
1296 
1297 /*
1298  * Put request/reply buffers back into pool.
1299  * Pre-decrement counter/array index.
1300  */
1301 void
1302 rpcrdma_buffer_put(struct rpcrdma_req *req)
1303 {
1304 	struct rpcrdma_buffer *buffers = req->rl_buffer;
1305 	struct rpcrdma_ia *ia = rdmab_to_ia(buffers);
1306 	unsigned long flags;
1307 
1308 	spin_lock_irqsave(&buffers->rb_lock, flags);
1309 	rpcrdma_buffer_put_sendbuf(req, buffers);
1310 	switch (ia->ri_memreg_strategy) {
1311 	case RPCRDMA_FRMR:
1312 	case RPCRDMA_MTHCAFMR:
1313 		rpcrdma_buffer_put_mrs(req, buffers);
1314 		break;
1315 	default:
1316 		break;
1317 	}
1318 	spin_unlock_irqrestore(&buffers->rb_lock, flags);
1319 }
1320 
1321 /*
1322  * Recover reply buffers from pool.
1323  * This happens when recovering from error conditions.
1324  * Post-increment counter/array index.
1325  */
1326 void
1327 rpcrdma_recv_buffer_get(struct rpcrdma_req *req)
1328 {
1329 	struct rpcrdma_buffer *buffers = req->rl_buffer;
1330 	unsigned long flags;
1331 
1332 	spin_lock_irqsave(&buffers->rb_lock, flags);
1333 	if (buffers->rb_recv_index < buffers->rb_max_requests) {
1334 		req->rl_reply = buffers->rb_recv_bufs[buffers->rb_recv_index];
1335 		buffers->rb_recv_bufs[buffers->rb_recv_index++] = NULL;
1336 	}
1337 	spin_unlock_irqrestore(&buffers->rb_lock, flags);
1338 }
1339 
1340 /*
1341  * Put reply buffers back into pool when not attached to
1342  * request. This happens in error conditions.
1343  */
1344 void
1345 rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1346 {
1347 	struct rpcrdma_buffer *buffers = rep->rr_buffer;
1348 	unsigned long flags;
1349 
1350 	rep->rr_func = NULL;
1351 	spin_lock_irqsave(&buffers->rb_lock, flags);
1352 	buffers->rb_recv_bufs[--buffers->rb_recv_index] = rep;
1353 	spin_unlock_irqrestore(&buffers->rb_lock, flags);
1354 }
1355 
1356 /*
1357  * Wrappers for internal-use kmalloc memory registration, used by buffer code.
1358  */
1359 
1360 void
1361 rpcrdma_mapping_error(struct rpcrdma_mr_seg *seg)
1362 {
1363 	dprintk("RPC:       map_one: offset %p iova %llx len %zu\n",
1364 		seg->mr_offset,
1365 		(unsigned long long)seg->mr_dma, seg->mr_dmalen);
1366 }
1367 
1368 static int
1369 rpcrdma_register_internal(struct rpcrdma_ia *ia, void *va, int len,
1370 				struct ib_mr **mrp, struct ib_sge *iov)
1371 {
1372 	struct ib_phys_buf ipb;
1373 	struct ib_mr *mr;
1374 	int rc;
1375 
1376 	/*
1377 	 * All memory passed here was kmalloc'ed, therefore phys-contiguous.
1378 	 */
1379 	iov->addr = ib_dma_map_single(ia->ri_id->device,
1380 			va, len, DMA_BIDIRECTIONAL);
1381 	if (ib_dma_mapping_error(ia->ri_id->device, iov->addr))
1382 		return -ENOMEM;
1383 
1384 	iov->length = len;
1385 
1386 	if (ia->ri_have_dma_lkey) {
1387 		*mrp = NULL;
1388 		iov->lkey = ia->ri_dma_lkey;
1389 		return 0;
1390 	} else if (ia->ri_bind_mem != NULL) {
1391 		*mrp = NULL;
1392 		iov->lkey = ia->ri_bind_mem->lkey;
1393 		return 0;
1394 	}
1395 
1396 	ipb.addr = iov->addr;
1397 	ipb.size = iov->length;
1398 	mr = ib_reg_phys_mr(ia->ri_pd, &ipb, 1,
1399 			IB_ACCESS_LOCAL_WRITE, &iov->addr);
1400 
1401 	dprintk("RPC:       %s: phys convert: 0x%llx "
1402 			"registered 0x%llx length %d\n",
1403 			__func__, (unsigned long long)ipb.addr,
1404 			(unsigned long long)iov->addr, len);
1405 
1406 	if (IS_ERR(mr)) {
1407 		*mrp = NULL;
1408 		rc = PTR_ERR(mr);
1409 		dprintk("RPC:       %s: failed with %i\n", __func__, rc);
1410 	} else {
1411 		*mrp = mr;
1412 		iov->lkey = mr->lkey;
1413 		rc = 0;
1414 	}
1415 
1416 	return rc;
1417 }
1418 
1419 static int
1420 rpcrdma_deregister_internal(struct rpcrdma_ia *ia,
1421 				struct ib_mr *mr, struct ib_sge *iov)
1422 {
1423 	int rc;
1424 
1425 	ib_dma_unmap_single(ia->ri_id->device,
1426 			iov->addr, iov->length, DMA_BIDIRECTIONAL);
1427 
1428 	if (NULL == mr)
1429 		return 0;
1430 
1431 	rc = ib_dereg_mr(mr);
1432 	if (rc)
1433 		dprintk("RPC:       %s: ib_dereg_mr failed %i\n", __func__, rc);
1434 	return rc;
1435 }
1436 
1437 /**
1438  * rpcrdma_alloc_regbuf - kmalloc and register memory for SEND/RECV buffers
1439  * @ia: controlling rpcrdma_ia
1440  * @size: size of buffer to be allocated, in bytes
1441  * @flags: GFP flags
1442  *
1443  * Returns pointer to private header of an area of internally
1444  * registered memory, or an ERR_PTR. The registered buffer follows
1445  * the end of the private header.
1446  *
1447  * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1448  * receiving the payload of RDMA RECV operations. regbufs are not
1449  * used for RDMA READ/WRITE operations, thus are registered only for
1450  * LOCAL access.
1451  */
1452 struct rpcrdma_regbuf *
1453 rpcrdma_alloc_regbuf(struct rpcrdma_ia *ia, size_t size, gfp_t flags)
1454 {
1455 	struct rpcrdma_regbuf *rb;
1456 	int rc;
1457 
1458 	rc = -ENOMEM;
1459 	rb = kmalloc(sizeof(*rb) + size, flags);
1460 	if (rb == NULL)
1461 		goto out;
1462 
1463 	rb->rg_size = size;
1464 	rb->rg_owner = NULL;
1465 	rc = rpcrdma_register_internal(ia, rb->rg_base, size,
1466 				       &rb->rg_mr, &rb->rg_iov);
1467 	if (rc)
1468 		goto out_free;
1469 
1470 	return rb;
1471 
1472 out_free:
1473 	kfree(rb);
1474 out:
1475 	return ERR_PTR(rc);
1476 }
1477 
1478 /**
1479  * rpcrdma_free_regbuf - deregister and free registered buffer
1480  * @ia: controlling rpcrdma_ia
1481  * @rb: regbuf to be deregistered and freed
1482  */
1483 void
1484 rpcrdma_free_regbuf(struct rpcrdma_ia *ia, struct rpcrdma_regbuf *rb)
1485 {
1486 	if (rb) {
1487 		rpcrdma_deregister_internal(ia, rb->rg_mr, &rb->rg_iov);
1488 		kfree(rb);
1489 	}
1490 }
1491 
1492 /*
1493  * Prepost any receive buffer, then post send.
1494  *
1495  * Receive buffer is donated to hardware, reclaimed upon recv completion.
1496  */
1497 int
1498 rpcrdma_ep_post(struct rpcrdma_ia *ia,
1499 		struct rpcrdma_ep *ep,
1500 		struct rpcrdma_req *req)
1501 {
1502 	struct ib_send_wr send_wr, *send_wr_fail;
1503 	struct rpcrdma_rep *rep = req->rl_reply;
1504 	int rc;
1505 
1506 	if (rep) {
1507 		rc = rpcrdma_ep_post_recv(ia, ep, rep);
1508 		if (rc)
1509 			goto out;
1510 		req->rl_reply = NULL;
1511 	}
1512 
1513 	send_wr.next = NULL;
1514 	send_wr.wr_id = RPCRDMA_IGNORE_COMPLETION;
1515 	send_wr.sg_list = req->rl_send_iov;
1516 	send_wr.num_sge = req->rl_niovs;
1517 	send_wr.opcode = IB_WR_SEND;
1518 	if (send_wr.num_sge == 4)	/* no need to sync any pad (constant) */
1519 		ib_dma_sync_single_for_device(ia->ri_id->device,
1520 			req->rl_send_iov[3].addr, req->rl_send_iov[3].length,
1521 			DMA_TO_DEVICE);
1522 	ib_dma_sync_single_for_device(ia->ri_id->device,
1523 		req->rl_send_iov[1].addr, req->rl_send_iov[1].length,
1524 		DMA_TO_DEVICE);
1525 	ib_dma_sync_single_for_device(ia->ri_id->device,
1526 		req->rl_send_iov[0].addr, req->rl_send_iov[0].length,
1527 		DMA_TO_DEVICE);
1528 
1529 	if (DECR_CQCOUNT(ep) > 0)
1530 		send_wr.send_flags = 0;
1531 	else { /* Provider must take a send completion every now and then */
1532 		INIT_CQCOUNT(ep);
1533 		send_wr.send_flags = IB_SEND_SIGNALED;
1534 	}
1535 
1536 	rc = ib_post_send(ia->ri_id->qp, &send_wr, &send_wr_fail);
1537 	if (rc)
1538 		dprintk("RPC:       %s: ib_post_send returned %i\n", __func__,
1539 			rc);
1540 out:
1541 	return rc;
1542 }
1543 
1544 /*
1545  * (Re)post a receive buffer.
1546  */
1547 int
1548 rpcrdma_ep_post_recv(struct rpcrdma_ia *ia,
1549 		     struct rpcrdma_ep *ep,
1550 		     struct rpcrdma_rep *rep)
1551 {
1552 	struct ib_recv_wr recv_wr, *recv_wr_fail;
1553 	int rc;
1554 
1555 	recv_wr.next = NULL;
1556 	recv_wr.wr_id = (u64) (unsigned long) rep;
1557 	recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
1558 	recv_wr.num_sge = 1;
1559 
1560 	ib_dma_sync_single_for_cpu(ia->ri_id->device,
1561 				   rdmab_addr(rep->rr_rdmabuf),
1562 				   rdmab_length(rep->rr_rdmabuf),
1563 				   DMA_BIDIRECTIONAL);
1564 
1565 	rc = ib_post_recv(ia->ri_id->qp, &recv_wr, &recv_wr_fail);
1566 
1567 	if (rc)
1568 		dprintk("RPC:       %s: ib_post_recv returned %i\n", __func__,
1569 			rc);
1570 	return rc;
1571 }
1572 
1573 /* How many chunk list items fit within our inline buffers?
1574  */
1575 unsigned int
1576 rpcrdma_max_segments(struct rpcrdma_xprt *r_xprt)
1577 {
1578 	struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
1579 	int bytes, segments;
1580 
1581 	bytes = min_t(unsigned int, cdata->inline_wsize, cdata->inline_rsize);
1582 	bytes -= RPCRDMA_HDRLEN_MIN;
1583 	if (bytes < sizeof(struct rpcrdma_segment) * 2) {
1584 		pr_warn("RPC:       %s: inline threshold too small\n",
1585 			__func__);
1586 		return 0;
1587 	}
1588 
1589 	segments = 1 << (fls(bytes / sizeof(struct rpcrdma_segment)) - 1);
1590 	dprintk("RPC:       %s: max chunk list size = %d segments\n",
1591 		__func__, segments);
1592 	return segments;
1593 }
1594