1 /*
2  * Copyright (c) 2005-2007 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  * Author: Tom Tucker <tom@opengridcomputing.com>
40  */
41 
42 #include <linux/sunrpc/svc_xprt.h>
43 #include <linux/sunrpc/debug.h>
44 #include <linux/sunrpc/rpc_rdma.h>
45 #include <linux/spinlock.h>
46 #include <rdma/ib_verbs.h>
47 #include <rdma/rdma_cm.h>
48 #include <linux/sunrpc/svc_rdma.h>
49 
50 #define RPCDBG_FACILITY	RPCDBG_SVCXPRT
51 
52 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
53 					struct sockaddr *sa, int salen,
54 					int flags);
55 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt);
56 static void svc_rdma_release_rqst(struct svc_rqst *);
57 static void dto_tasklet_func(unsigned long data);
58 static void svc_rdma_detach(struct svc_xprt *xprt);
59 static void svc_rdma_free(struct svc_xprt *xprt);
60 static int svc_rdma_has_wspace(struct svc_xprt *xprt);
61 static void rq_cq_reap(struct svcxprt_rdma *xprt);
62 static void sq_cq_reap(struct svcxprt_rdma *xprt);
63 
64 DECLARE_TASKLET(dto_tasklet, dto_tasklet_func, 0UL);
65 static DEFINE_SPINLOCK(dto_lock);
66 static LIST_HEAD(dto_xprt_q);
67 
68 static struct svc_xprt_ops svc_rdma_ops = {
69 	.xpo_create = svc_rdma_create,
70 	.xpo_recvfrom = svc_rdma_recvfrom,
71 	.xpo_sendto = svc_rdma_sendto,
72 	.xpo_release_rqst = svc_rdma_release_rqst,
73 	.xpo_detach = svc_rdma_detach,
74 	.xpo_free = svc_rdma_free,
75 	.xpo_prep_reply_hdr = svc_rdma_prep_reply_hdr,
76 	.xpo_has_wspace = svc_rdma_has_wspace,
77 	.xpo_accept = svc_rdma_accept,
78 };
79 
80 struct svc_xprt_class svc_rdma_class = {
81 	.xcl_name = "rdma",
82 	.xcl_owner = THIS_MODULE,
83 	.xcl_ops = &svc_rdma_ops,
84 	.xcl_max_payload = RPCSVC_MAXPAYLOAD_TCP,
85 };
86 
87 static int rdma_bump_context_cache(struct svcxprt_rdma *xprt)
88 {
89 	int target;
90 	int at_least_one = 0;
91 	struct svc_rdma_op_ctxt *ctxt;
92 
93 	target = min(xprt->sc_ctxt_cnt + xprt->sc_ctxt_bump,
94 		     xprt->sc_ctxt_max);
95 
96 	spin_lock_bh(&xprt->sc_ctxt_lock);
97 	while (xprt->sc_ctxt_cnt < target) {
98 		xprt->sc_ctxt_cnt++;
99 		spin_unlock_bh(&xprt->sc_ctxt_lock);
100 
101 		ctxt = kmalloc(sizeof(*ctxt), GFP_KERNEL);
102 
103 		spin_lock_bh(&xprt->sc_ctxt_lock);
104 		if (ctxt) {
105 			at_least_one = 1;
106 			ctxt->next = xprt->sc_ctxt_head;
107 			xprt->sc_ctxt_head = ctxt;
108 		} else {
109 			/* kmalloc failed...give up for now */
110 			xprt->sc_ctxt_cnt--;
111 			break;
112 		}
113 	}
114 	spin_unlock_bh(&xprt->sc_ctxt_lock);
115 	dprintk("svcrdma: sc_ctxt_max=%d, sc_ctxt_cnt=%d\n",
116 		xprt->sc_ctxt_max, xprt->sc_ctxt_cnt);
117 	return at_least_one;
118 }
119 
120 struct svc_rdma_op_ctxt *svc_rdma_get_context(struct svcxprt_rdma *xprt)
121 {
122 	struct svc_rdma_op_ctxt *ctxt;
123 
124 	while (1) {
125 		spin_lock_bh(&xprt->sc_ctxt_lock);
126 		if (unlikely(xprt->sc_ctxt_head == NULL)) {
127 			/* Try to bump my cache. */
128 			spin_unlock_bh(&xprt->sc_ctxt_lock);
129 
130 			if (rdma_bump_context_cache(xprt))
131 				continue;
132 
133 			printk(KERN_INFO "svcrdma: sleeping waiting for "
134 			       "context memory on xprt=%p\n",
135 			       xprt);
136 			schedule_timeout_uninterruptible(msecs_to_jiffies(500));
137 			continue;
138 		}
139 		ctxt = xprt->sc_ctxt_head;
140 		xprt->sc_ctxt_head = ctxt->next;
141 		spin_unlock_bh(&xprt->sc_ctxt_lock);
142 		ctxt->xprt = xprt;
143 		INIT_LIST_HEAD(&ctxt->dto_q);
144 		ctxt->count = 0;
145 		break;
146 	}
147 	return ctxt;
148 }
149 
150 void svc_rdma_put_context(struct svc_rdma_op_ctxt *ctxt, int free_pages)
151 {
152 	struct svcxprt_rdma *xprt;
153 	int i;
154 
155 	BUG_ON(!ctxt);
156 	xprt = ctxt->xprt;
157 	if (free_pages)
158 		for (i = 0; i < ctxt->count; i++)
159 			put_page(ctxt->pages[i]);
160 
161 	for (i = 0; i < ctxt->count; i++)
162 		dma_unmap_single(xprt->sc_cm_id->device->dma_device,
163 				 ctxt->sge[i].addr,
164 				 ctxt->sge[i].length,
165 				 ctxt->direction);
166 	spin_lock_bh(&xprt->sc_ctxt_lock);
167 	ctxt->next = xprt->sc_ctxt_head;
168 	xprt->sc_ctxt_head = ctxt;
169 	spin_unlock_bh(&xprt->sc_ctxt_lock);
170 }
171 
172 /* ib_cq event handler */
173 static void cq_event_handler(struct ib_event *event, void *context)
174 {
175 	struct svc_xprt *xprt = context;
176 	dprintk("svcrdma: received CQ event id=%d, context=%p\n",
177 		event->event, context);
178 	set_bit(XPT_CLOSE, &xprt->xpt_flags);
179 }
180 
181 /* QP event handler */
182 static void qp_event_handler(struct ib_event *event, void *context)
183 {
184 	struct svc_xprt *xprt = context;
185 
186 	switch (event->event) {
187 	/* These are considered benign events */
188 	case IB_EVENT_PATH_MIG:
189 	case IB_EVENT_COMM_EST:
190 	case IB_EVENT_SQ_DRAINED:
191 	case IB_EVENT_QP_LAST_WQE_REACHED:
192 		dprintk("svcrdma: QP event %d received for QP=%p\n",
193 			event->event, event->element.qp);
194 		break;
195 	/* These are considered fatal events */
196 	case IB_EVENT_PATH_MIG_ERR:
197 	case IB_EVENT_QP_FATAL:
198 	case IB_EVENT_QP_REQ_ERR:
199 	case IB_EVENT_QP_ACCESS_ERR:
200 	case IB_EVENT_DEVICE_FATAL:
201 	default:
202 		dprintk("svcrdma: QP ERROR event %d received for QP=%p, "
203 			"closing transport\n",
204 			event->event, event->element.qp);
205 		set_bit(XPT_CLOSE, &xprt->xpt_flags);
206 		break;
207 	}
208 }
209 
210 /*
211  * Data Transfer Operation Tasklet
212  *
213  * Walks a list of transports with I/O pending, removing entries as
214  * they are added to the server's I/O pending list. Two bits indicate
215  * if SQ, RQ, or both have I/O pending. The dto_lock is an irqsave
216  * spinlock that serializes access to the transport list with the RQ
217  * and SQ interrupt handlers.
218  */
219 static void dto_tasklet_func(unsigned long data)
220 {
221 	struct svcxprt_rdma *xprt;
222 	unsigned long flags;
223 
224 	spin_lock_irqsave(&dto_lock, flags);
225 	while (!list_empty(&dto_xprt_q)) {
226 		xprt = list_entry(dto_xprt_q.next,
227 				  struct svcxprt_rdma, sc_dto_q);
228 		list_del_init(&xprt->sc_dto_q);
229 		spin_unlock_irqrestore(&dto_lock, flags);
230 
231 		if (test_and_clear_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags)) {
232 			ib_req_notify_cq(xprt->sc_rq_cq, IB_CQ_NEXT_COMP);
233 			rq_cq_reap(xprt);
234 			set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
235 			/*
236 			 * If data arrived before established event,
237 			 * don't enqueue. This defers RPC I/O until the
238 			 * RDMA connection is complete.
239 			 */
240 			if (!test_bit(RDMAXPRT_CONN_PENDING, &xprt->sc_flags))
241 				svc_xprt_enqueue(&xprt->sc_xprt);
242 		}
243 
244 		if (test_and_clear_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags)) {
245 			ib_req_notify_cq(xprt->sc_sq_cq, IB_CQ_NEXT_COMP);
246 			sq_cq_reap(xprt);
247 		}
248 
249 		svc_xprt_put(&xprt->sc_xprt);
250 		spin_lock_irqsave(&dto_lock, flags);
251 	}
252 	spin_unlock_irqrestore(&dto_lock, flags);
253 }
254 
255 /*
256  * Receive Queue Completion Handler
257  *
258  * Since an RQ completion handler is called on interrupt context, we
259  * need to defer the handling of the I/O to a tasklet
260  */
261 static void rq_comp_handler(struct ib_cq *cq, void *cq_context)
262 {
263 	struct svcxprt_rdma *xprt = cq_context;
264 	unsigned long flags;
265 
266 	/*
267 	 * Set the bit regardless of whether or not it's on the list
268 	 * because it may be on the list already due to an SQ
269 	 * completion.
270 	*/
271 	set_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags);
272 
273 	/*
274 	 * If this transport is not already on the DTO transport queue,
275 	 * add it
276 	 */
277 	spin_lock_irqsave(&dto_lock, flags);
278 	if (list_empty(&xprt->sc_dto_q)) {
279 		svc_xprt_get(&xprt->sc_xprt);
280 		list_add_tail(&xprt->sc_dto_q, &dto_xprt_q);
281 	}
282 	spin_unlock_irqrestore(&dto_lock, flags);
283 
284 	/* Tasklet does all the work to avoid irqsave locks. */
285 	tasklet_schedule(&dto_tasklet);
286 }
287 
288 /*
289  * rq_cq_reap - Process the RQ CQ.
290  *
291  * Take all completing WC off the CQE and enqueue the associated DTO
292  * context on the dto_q for the transport.
293  */
294 static void rq_cq_reap(struct svcxprt_rdma *xprt)
295 {
296 	int ret;
297 	struct ib_wc wc;
298 	struct svc_rdma_op_ctxt *ctxt = NULL;
299 
300 	atomic_inc(&rdma_stat_rq_poll);
301 
302 	spin_lock_bh(&xprt->sc_rq_dto_lock);
303 	while ((ret = ib_poll_cq(xprt->sc_rq_cq, 1, &wc)) > 0) {
304 		ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id;
305 		ctxt->wc_status = wc.status;
306 		ctxt->byte_len = wc.byte_len;
307 		if (wc.status != IB_WC_SUCCESS) {
308 			/* Close the transport */
309 			set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
310 			svc_rdma_put_context(ctxt, 1);
311 			continue;
312 		}
313 		list_add_tail(&ctxt->dto_q, &xprt->sc_rq_dto_q);
314 	}
315 	spin_unlock_bh(&xprt->sc_rq_dto_lock);
316 
317 	if (ctxt)
318 		atomic_inc(&rdma_stat_rq_prod);
319 }
320 
321 /*
322  * Send Queue Completion Handler - potentially called on interrupt context.
323  */
324 static void sq_cq_reap(struct svcxprt_rdma *xprt)
325 {
326 	struct svc_rdma_op_ctxt *ctxt = NULL;
327 	struct ib_wc wc;
328 	struct ib_cq *cq = xprt->sc_sq_cq;
329 	int ret;
330 
331 	atomic_inc(&rdma_stat_sq_poll);
332 	while ((ret = ib_poll_cq(cq, 1, &wc)) > 0) {
333 		ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id;
334 		xprt = ctxt->xprt;
335 
336 		if (wc.status != IB_WC_SUCCESS)
337 			/* Close the transport */
338 			set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
339 
340 		/* Decrement used SQ WR count */
341 		atomic_dec(&xprt->sc_sq_count);
342 		wake_up(&xprt->sc_send_wait);
343 
344 		switch (ctxt->wr_op) {
345 		case IB_WR_SEND:
346 		case IB_WR_RDMA_WRITE:
347 			svc_rdma_put_context(ctxt, 1);
348 			break;
349 
350 		case IB_WR_RDMA_READ:
351 			if (test_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags)) {
352 				set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
353 				set_bit(RDMACTXT_F_READ_DONE, &ctxt->flags);
354 				spin_lock_bh(&xprt->sc_read_complete_lock);
355 				list_add_tail(&ctxt->dto_q,
356 					      &xprt->sc_read_complete_q);
357 				spin_unlock_bh(&xprt->sc_read_complete_lock);
358 				svc_xprt_enqueue(&xprt->sc_xprt);
359 			}
360 			break;
361 
362 		default:
363 			printk(KERN_ERR "svcrdma: unexpected completion type, "
364 			       "opcode=%d, status=%d\n",
365 			       wc.opcode, wc.status);
366 			break;
367 		}
368 	}
369 
370 	if (ctxt)
371 		atomic_inc(&rdma_stat_sq_prod);
372 }
373 
374 static void sq_comp_handler(struct ib_cq *cq, void *cq_context)
375 {
376 	struct svcxprt_rdma *xprt = cq_context;
377 	unsigned long flags;
378 
379 	/*
380 	 * Set the bit regardless of whether or not it's on the list
381 	 * because it may be on the list already due to an RQ
382 	 * completion.
383 	*/
384 	set_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags);
385 
386 	/*
387 	 * If this transport is not already on the DTO transport queue,
388 	 * add it
389 	 */
390 	spin_lock_irqsave(&dto_lock, flags);
391 	if (list_empty(&xprt->sc_dto_q)) {
392 		svc_xprt_get(&xprt->sc_xprt);
393 		list_add_tail(&xprt->sc_dto_q, &dto_xprt_q);
394 	}
395 	spin_unlock_irqrestore(&dto_lock, flags);
396 
397 	/* Tasklet does all the work to avoid irqsave locks. */
398 	tasklet_schedule(&dto_tasklet);
399 }
400 
401 static void create_context_cache(struct svcxprt_rdma *xprt,
402 				 int ctxt_count, int ctxt_bump, int ctxt_max)
403 {
404 	struct svc_rdma_op_ctxt *ctxt;
405 	int i;
406 
407 	xprt->sc_ctxt_max = ctxt_max;
408 	xprt->sc_ctxt_bump = ctxt_bump;
409 	xprt->sc_ctxt_cnt = 0;
410 	xprt->sc_ctxt_head = NULL;
411 	for (i = 0; i < ctxt_count; i++) {
412 		ctxt = kmalloc(sizeof(*ctxt), GFP_KERNEL);
413 		if (ctxt) {
414 			ctxt->next = xprt->sc_ctxt_head;
415 			xprt->sc_ctxt_head = ctxt;
416 			xprt->sc_ctxt_cnt++;
417 		}
418 	}
419 }
420 
421 static void destroy_context_cache(struct svc_rdma_op_ctxt *ctxt)
422 {
423 	struct svc_rdma_op_ctxt *next;
424 	if (!ctxt)
425 		return;
426 
427 	do {
428 		next = ctxt->next;
429 		kfree(ctxt);
430 		ctxt = next;
431 	} while (next);
432 }
433 
434 static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *serv,
435 					     int listener)
436 {
437 	struct svcxprt_rdma *cma_xprt = kzalloc(sizeof *cma_xprt, GFP_KERNEL);
438 
439 	if (!cma_xprt)
440 		return NULL;
441 	svc_xprt_init(&svc_rdma_class, &cma_xprt->sc_xprt, serv);
442 	INIT_LIST_HEAD(&cma_xprt->sc_accept_q);
443 	INIT_LIST_HEAD(&cma_xprt->sc_dto_q);
444 	INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q);
445 	INIT_LIST_HEAD(&cma_xprt->sc_read_complete_q);
446 	init_waitqueue_head(&cma_xprt->sc_send_wait);
447 
448 	spin_lock_init(&cma_xprt->sc_lock);
449 	spin_lock_init(&cma_xprt->sc_read_complete_lock);
450 	spin_lock_init(&cma_xprt->sc_ctxt_lock);
451 	spin_lock_init(&cma_xprt->sc_rq_dto_lock);
452 
453 	cma_xprt->sc_ord = svcrdma_ord;
454 
455 	cma_xprt->sc_max_req_size = svcrdma_max_req_size;
456 	cma_xprt->sc_max_requests = svcrdma_max_requests;
457 	cma_xprt->sc_sq_depth = svcrdma_max_requests * RPCRDMA_SQ_DEPTH_MULT;
458 	atomic_set(&cma_xprt->sc_sq_count, 0);
459 
460 	if (!listener) {
461 		int reqs = cma_xprt->sc_max_requests;
462 		create_context_cache(cma_xprt,
463 				     reqs << 1, /* starting size */
464 				     reqs,	/* bump amount */
465 				     reqs +
466 				     cma_xprt->sc_sq_depth +
467 				     RPCRDMA_MAX_THREADS + 1); /* max */
468 		if (!cma_xprt->sc_ctxt_head) {
469 			kfree(cma_xprt);
470 			return NULL;
471 		}
472 		clear_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
473 	} else
474 		set_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
475 
476 	return cma_xprt;
477 }
478 
479 struct page *svc_rdma_get_page(void)
480 {
481 	struct page *page;
482 
483 	while ((page = alloc_page(GFP_KERNEL)) == NULL) {
484 		/* If we can't get memory, wait a bit and try again */
485 		printk(KERN_INFO "svcrdma: out of memory...retrying in 1000 "
486 		       "jiffies.\n");
487 		schedule_timeout_uninterruptible(msecs_to_jiffies(1000));
488 	}
489 	return page;
490 }
491 
492 int svc_rdma_post_recv(struct svcxprt_rdma *xprt)
493 {
494 	struct ib_recv_wr recv_wr, *bad_recv_wr;
495 	struct svc_rdma_op_ctxt *ctxt;
496 	struct page *page;
497 	unsigned long pa;
498 	int sge_no;
499 	int buflen;
500 	int ret;
501 
502 	ctxt = svc_rdma_get_context(xprt);
503 	buflen = 0;
504 	ctxt->direction = DMA_FROM_DEVICE;
505 	for (sge_no = 0; buflen < xprt->sc_max_req_size; sge_no++) {
506 		BUG_ON(sge_no >= xprt->sc_max_sge);
507 		page = svc_rdma_get_page();
508 		ctxt->pages[sge_no] = page;
509 		pa = ib_dma_map_page(xprt->sc_cm_id->device,
510 				     page, 0, PAGE_SIZE,
511 				     DMA_FROM_DEVICE);
512 		ctxt->sge[sge_no].addr = pa;
513 		ctxt->sge[sge_no].length = PAGE_SIZE;
514 		ctxt->sge[sge_no].lkey = xprt->sc_phys_mr->lkey;
515 		buflen += PAGE_SIZE;
516 	}
517 	ctxt->count = sge_no;
518 	recv_wr.next = NULL;
519 	recv_wr.sg_list = &ctxt->sge[0];
520 	recv_wr.num_sge = ctxt->count;
521 	recv_wr.wr_id = (u64)(unsigned long)ctxt;
522 
523 	ret = ib_post_recv(xprt->sc_qp, &recv_wr, &bad_recv_wr);
524 	return ret;
525 }
526 
527 /*
528  * This function handles the CONNECT_REQUEST event on a listening
529  * endpoint. It is passed the cma_id for the _new_ connection. The context in
530  * this cma_id is inherited from the listening cma_id and is the svc_xprt
531  * structure for the listening endpoint.
532  *
533  * This function creates a new xprt for the new connection and enqueues it on
534  * the accept queue for the listent xprt. When the listen thread is kicked, it
535  * will call the recvfrom method on the listen xprt which will accept the new
536  * connection.
537  */
538 static void handle_connect_req(struct rdma_cm_id *new_cma_id)
539 {
540 	struct svcxprt_rdma *listen_xprt = new_cma_id->context;
541 	struct svcxprt_rdma *newxprt;
542 
543 	/* Create a new transport */
544 	newxprt = rdma_create_xprt(listen_xprt->sc_xprt.xpt_server, 0);
545 	if (!newxprt) {
546 		dprintk("svcrdma: failed to create new transport\n");
547 		return;
548 	}
549 	newxprt->sc_cm_id = new_cma_id;
550 	new_cma_id->context = newxprt;
551 	dprintk("svcrdma: Creating newxprt=%p, cm_id=%p, listenxprt=%p\n",
552 		newxprt, newxprt->sc_cm_id, listen_xprt);
553 
554 	/*
555 	 * Enqueue the new transport on the accept queue of the listening
556 	 * transport
557 	 */
558 	spin_lock_bh(&listen_xprt->sc_lock);
559 	list_add_tail(&newxprt->sc_accept_q, &listen_xprt->sc_accept_q);
560 	spin_unlock_bh(&listen_xprt->sc_lock);
561 
562 	/*
563 	 * Can't use svc_xprt_received here because we are not on a
564 	 * rqstp thread
565 	*/
566 	set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags);
567 	svc_xprt_enqueue(&listen_xprt->sc_xprt);
568 }
569 
570 /*
571  * Handles events generated on the listening endpoint. These events will be
572  * either be incoming connect requests or adapter removal  events.
573  */
574 static int rdma_listen_handler(struct rdma_cm_id *cma_id,
575 			       struct rdma_cm_event *event)
576 {
577 	struct svcxprt_rdma *xprt = cma_id->context;
578 	int ret = 0;
579 
580 	switch (event->event) {
581 	case RDMA_CM_EVENT_CONNECT_REQUEST:
582 		dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, "
583 			"event=%d\n", cma_id, cma_id->context, event->event);
584 		handle_connect_req(cma_id);
585 		break;
586 
587 	case RDMA_CM_EVENT_ESTABLISHED:
588 		/* Accept complete */
589 		dprintk("svcrdma: Connection completed on LISTEN xprt=%p, "
590 			"cm_id=%p\n", xprt, cma_id);
591 		break;
592 
593 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
594 		dprintk("svcrdma: Device removal xprt=%p, cm_id=%p\n",
595 			xprt, cma_id);
596 		if (xprt)
597 			set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
598 		break;
599 
600 	default:
601 		dprintk("svcrdma: Unexpected event on listening endpoint %p, "
602 			"event=%d\n", cma_id, event->event);
603 		break;
604 	}
605 
606 	return ret;
607 }
608 
609 static int rdma_cma_handler(struct rdma_cm_id *cma_id,
610 			    struct rdma_cm_event *event)
611 {
612 	struct svc_xprt *xprt = cma_id->context;
613 	struct svcxprt_rdma *rdma =
614 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
615 	switch (event->event) {
616 	case RDMA_CM_EVENT_ESTABLISHED:
617 		/* Accept complete */
618 		svc_xprt_get(xprt);
619 		dprintk("svcrdma: Connection completed on DTO xprt=%p, "
620 			"cm_id=%p\n", xprt, cma_id);
621 		clear_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags);
622 		svc_xprt_enqueue(xprt);
623 		break;
624 	case RDMA_CM_EVENT_DISCONNECTED:
625 		dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n",
626 			xprt, cma_id);
627 		if (xprt) {
628 			set_bit(XPT_CLOSE, &xprt->xpt_flags);
629 			svc_xprt_enqueue(xprt);
630 		}
631 		break;
632 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
633 		dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, "
634 			"event=%d\n", cma_id, xprt, event->event);
635 		if (xprt) {
636 			set_bit(XPT_CLOSE, &xprt->xpt_flags);
637 			svc_xprt_enqueue(xprt);
638 		}
639 		break;
640 	default:
641 		dprintk("svcrdma: Unexpected event on DTO endpoint %p, "
642 			"event=%d\n", cma_id, event->event);
643 		break;
644 	}
645 	return 0;
646 }
647 
648 /*
649  * Create a listening RDMA service endpoint.
650  */
651 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
652 					struct sockaddr *sa, int salen,
653 					int flags)
654 {
655 	struct rdma_cm_id *listen_id;
656 	struct svcxprt_rdma *cma_xprt;
657 	struct svc_xprt *xprt;
658 	int ret;
659 
660 	dprintk("svcrdma: Creating RDMA socket\n");
661 
662 	cma_xprt = rdma_create_xprt(serv, 1);
663 	if (!cma_xprt)
664 		return ERR_PTR(ENOMEM);
665 	xprt = &cma_xprt->sc_xprt;
666 
667 	listen_id = rdma_create_id(rdma_listen_handler, cma_xprt, RDMA_PS_TCP);
668 	if (IS_ERR(listen_id)) {
669 		svc_xprt_put(&cma_xprt->sc_xprt);
670 		dprintk("svcrdma: rdma_create_id failed = %ld\n",
671 			PTR_ERR(listen_id));
672 		return (void *)listen_id;
673 	}
674 	ret = rdma_bind_addr(listen_id, sa);
675 	if (ret) {
676 		rdma_destroy_id(listen_id);
677 		svc_xprt_put(&cma_xprt->sc_xprt);
678 		dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret);
679 		return ERR_PTR(ret);
680 	}
681 	cma_xprt->sc_cm_id = listen_id;
682 
683 	ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG);
684 	if (ret) {
685 		rdma_destroy_id(listen_id);
686 		svc_xprt_put(&cma_xprt->sc_xprt);
687 		dprintk("svcrdma: rdma_listen failed = %d\n", ret);
688 		return ERR_PTR(ret);
689 	}
690 
691 	/*
692 	 * We need to use the address from the cm_id in case the
693 	 * caller specified 0 for the port number.
694 	 */
695 	sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr;
696 	svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen);
697 
698 	return &cma_xprt->sc_xprt;
699 }
700 
701 /*
702  * This is the xpo_recvfrom function for listening endpoints. Its
703  * purpose is to accept incoming connections. The CMA callback handler
704  * has already created a new transport and attached it to the new CMA
705  * ID.
706  *
707  * There is a queue of pending connections hung on the listening
708  * transport. This queue contains the new svc_xprt structure. This
709  * function takes svc_xprt structures off the accept_q and completes
710  * the connection.
711  */
712 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
713 {
714 	struct svcxprt_rdma *listen_rdma;
715 	struct svcxprt_rdma *newxprt = NULL;
716 	struct rdma_conn_param conn_param;
717 	struct ib_qp_init_attr qp_attr;
718 	struct ib_device_attr devattr;
719 	struct sockaddr *sa;
720 	int ret;
721 	int i;
722 
723 	listen_rdma = container_of(xprt, struct svcxprt_rdma, sc_xprt);
724 	clear_bit(XPT_CONN, &xprt->xpt_flags);
725 	/* Get the next entry off the accept list */
726 	spin_lock_bh(&listen_rdma->sc_lock);
727 	if (!list_empty(&listen_rdma->sc_accept_q)) {
728 		newxprt = list_entry(listen_rdma->sc_accept_q.next,
729 				     struct svcxprt_rdma, sc_accept_q);
730 		list_del_init(&newxprt->sc_accept_q);
731 	}
732 	if (!list_empty(&listen_rdma->sc_accept_q))
733 		set_bit(XPT_CONN, &listen_rdma->sc_xprt.xpt_flags);
734 	spin_unlock_bh(&listen_rdma->sc_lock);
735 	if (!newxprt)
736 		return NULL;
737 
738 	dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n",
739 		newxprt, newxprt->sc_cm_id);
740 
741 	ret = ib_query_device(newxprt->sc_cm_id->device, &devattr);
742 	if (ret) {
743 		dprintk("svcrdma: could not query device attributes on "
744 			"device %p, rc=%d\n", newxprt->sc_cm_id->device, ret);
745 		goto errout;
746 	}
747 
748 	/* Qualify the transport resource defaults with the
749 	 * capabilities of this particular device */
750 	newxprt->sc_max_sge = min((size_t)devattr.max_sge,
751 				  (size_t)RPCSVC_MAXPAGES);
752 	newxprt->sc_max_requests = min((size_t)devattr.max_qp_wr,
753 				   (size_t)svcrdma_max_requests);
754 	newxprt->sc_sq_depth = RPCRDMA_SQ_DEPTH_MULT * newxprt->sc_max_requests;
755 
756 	newxprt->sc_ord =  min((size_t)devattr.max_qp_rd_atom,
757 			       (size_t)svcrdma_ord);
758 
759 	newxprt->sc_pd = ib_alloc_pd(newxprt->sc_cm_id->device);
760 	if (IS_ERR(newxprt->sc_pd)) {
761 		dprintk("svcrdma: error creating PD for connect request\n");
762 		goto errout;
763 	}
764 	newxprt->sc_sq_cq = ib_create_cq(newxprt->sc_cm_id->device,
765 					 sq_comp_handler,
766 					 cq_event_handler,
767 					 newxprt,
768 					 newxprt->sc_sq_depth,
769 					 0);
770 	if (IS_ERR(newxprt->sc_sq_cq)) {
771 		dprintk("svcrdma: error creating SQ CQ for connect request\n");
772 		goto errout;
773 	}
774 	newxprt->sc_rq_cq = ib_create_cq(newxprt->sc_cm_id->device,
775 					 rq_comp_handler,
776 					 cq_event_handler,
777 					 newxprt,
778 					 newxprt->sc_max_requests,
779 					 0);
780 	if (IS_ERR(newxprt->sc_rq_cq)) {
781 		dprintk("svcrdma: error creating RQ CQ for connect request\n");
782 		goto errout;
783 	}
784 
785 	memset(&qp_attr, 0, sizeof qp_attr);
786 	qp_attr.event_handler = qp_event_handler;
787 	qp_attr.qp_context = &newxprt->sc_xprt;
788 	qp_attr.cap.max_send_wr = newxprt->sc_sq_depth;
789 	qp_attr.cap.max_recv_wr = newxprt->sc_max_requests;
790 	qp_attr.cap.max_send_sge = newxprt->sc_max_sge;
791 	qp_attr.cap.max_recv_sge = newxprt->sc_max_sge;
792 	qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
793 	qp_attr.qp_type = IB_QPT_RC;
794 	qp_attr.send_cq = newxprt->sc_sq_cq;
795 	qp_attr.recv_cq = newxprt->sc_rq_cq;
796 	dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n"
797 		"    cm_id->device=%p, sc_pd->device=%p\n"
798 		"    cap.max_send_wr = %d\n"
799 		"    cap.max_recv_wr = %d\n"
800 		"    cap.max_send_sge = %d\n"
801 		"    cap.max_recv_sge = %d\n",
802 		newxprt->sc_cm_id, newxprt->sc_pd,
803 		newxprt->sc_cm_id->device, newxprt->sc_pd->device,
804 		qp_attr.cap.max_send_wr,
805 		qp_attr.cap.max_recv_wr,
806 		qp_attr.cap.max_send_sge,
807 		qp_attr.cap.max_recv_sge);
808 
809 	ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr);
810 	if (ret) {
811 		/*
812 		 * XXX: This is a hack. We need a xx_request_qp interface
813 		 * that will adjust the qp_attr's with a best-effort
814 		 * number
815 		 */
816 		qp_attr.cap.max_send_sge -= 2;
817 		qp_attr.cap.max_recv_sge -= 2;
818 		ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd,
819 				     &qp_attr);
820 		if (ret) {
821 			dprintk("svcrdma: failed to create QP, ret=%d\n", ret);
822 			goto errout;
823 		}
824 		newxprt->sc_max_sge = qp_attr.cap.max_send_sge;
825 		newxprt->sc_max_sge = qp_attr.cap.max_recv_sge;
826 		newxprt->sc_sq_depth = qp_attr.cap.max_send_wr;
827 		newxprt->sc_max_requests = qp_attr.cap.max_recv_wr;
828 	}
829 	svc_xprt_get(&newxprt->sc_xprt);
830 	newxprt->sc_qp = newxprt->sc_cm_id->qp;
831 
832 	/* Register all of physical memory */
833 	newxprt->sc_phys_mr = ib_get_dma_mr(newxprt->sc_pd,
834 					    IB_ACCESS_LOCAL_WRITE |
835 					    IB_ACCESS_REMOTE_WRITE);
836 	if (IS_ERR(newxprt->sc_phys_mr)) {
837 		dprintk("svcrdma: Failed to create DMA MR ret=%d\n", ret);
838 		goto errout;
839 	}
840 
841 	/* Post receive buffers */
842 	for (i = 0; i < newxprt->sc_max_requests; i++) {
843 		ret = svc_rdma_post_recv(newxprt);
844 		if (ret) {
845 			dprintk("svcrdma: failure posting receive buffers\n");
846 			goto errout;
847 		}
848 	}
849 
850 	/* Swap out the handler */
851 	newxprt->sc_cm_id->event_handler = rdma_cma_handler;
852 
853 	/* Accept Connection */
854 	set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags);
855 	memset(&conn_param, 0, sizeof conn_param);
856 	conn_param.responder_resources = 0;
857 	conn_param.initiator_depth = newxprt->sc_ord;
858 	ret = rdma_accept(newxprt->sc_cm_id, &conn_param);
859 	if (ret) {
860 		dprintk("svcrdma: failed to accept new connection, ret=%d\n",
861 		       ret);
862 		goto errout;
863 	}
864 
865 	dprintk("svcrdma: new connection %p accepted with the following "
866 		"attributes:\n"
867 		"    local_ip        : %d.%d.%d.%d\n"
868 		"    local_port	     : %d\n"
869 		"    remote_ip       : %d.%d.%d.%d\n"
870 		"    remote_port     : %d\n"
871 		"    max_sge         : %d\n"
872 		"    sq_depth        : %d\n"
873 		"    max_requests    : %d\n"
874 		"    ord             : %d\n",
875 		newxprt,
876 		NIPQUAD(((struct sockaddr_in *)&newxprt->sc_cm_id->
877 			 route.addr.src_addr)->sin_addr.s_addr),
878 		ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id->
879 		       route.addr.src_addr)->sin_port),
880 		NIPQUAD(((struct sockaddr_in *)&newxprt->sc_cm_id->
881 			 route.addr.dst_addr)->sin_addr.s_addr),
882 		ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id->
883 		       route.addr.dst_addr)->sin_port),
884 		newxprt->sc_max_sge,
885 		newxprt->sc_sq_depth,
886 		newxprt->sc_max_requests,
887 		newxprt->sc_ord);
888 
889 	/* Set the local and remote addresses in the transport */
890 	sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
891 	svc_xprt_set_remote(&newxprt->sc_xprt, sa, svc_addr_len(sa));
892 	sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
893 	svc_xprt_set_local(&newxprt->sc_xprt, sa, svc_addr_len(sa));
894 
895 	ib_req_notify_cq(newxprt->sc_sq_cq, IB_CQ_NEXT_COMP);
896 	ib_req_notify_cq(newxprt->sc_rq_cq, IB_CQ_NEXT_COMP);
897 	return &newxprt->sc_xprt;
898 
899  errout:
900 	dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret);
901 	/* Take a reference in case the DTO handler runs */
902 	svc_xprt_get(&newxprt->sc_xprt);
903 	if (newxprt->sc_qp && !IS_ERR(newxprt->sc_qp)) {
904 		ib_destroy_qp(newxprt->sc_qp);
905 		svc_xprt_put(&newxprt->sc_xprt);
906 	}
907 	rdma_destroy_id(newxprt->sc_cm_id);
908 	/* This call to put will destroy the transport */
909 	svc_xprt_put(&newxprt->sc_xprt);
910 	return NULL;
911 }
912 
913 /*
914  * Post an RQ WQE to the RQ when the rqst is being released. This
915  * effectively returns an RQ credit to the client. The rq_xprt_ctxt
916  * will be null if the request is deferred due to an RDMA_READ or the
917  * transport had no data ready (EAGAIN). Note that an RPC deferred in
918  * svc_process will still return the credit, this is because the data
919  * is copied and no longer consume a WQE/WC.
920  */
921 static void svc_rdma_release_rqst(struct svc_rqst *rqstp)
922 {
923 	int err;
924 	struct svcxprt_rdma *rdma =
925 		container_of(rqstp->rq_xprt, struct svcxprt_rdma, sc_xprt);
926 	if (rqstp->rq_xprt_ctxt) {
927 		BUG_ON(rqstp->rq_xprt_ctxt != rdma);
928 		err = svc_rdma_post_recv(rdma);
929 		if (err)
930 			dprintk("svcrdma: failed to post an RQ WQE error=%d\n",
931 				err);
932 	}
933 	rqstp->rq_xprt_ctxt = NULL;
934 }
935 
936 /*
937  * When connected, an svc_xprt has at least three references:
938  *
939  * - A reference held by the QP. We still hold that here because this
940  *   code deletes the QP and puts the reference.
941  *
942  * - A reference held by the cm_id between the ESTABLISHED and
943  *   DISCONNECTED events. If the remote peer disconnected first, this
944  *   reference could be gone.
945  *
946  * - A reference held by the svc_recv code that called this function
947  *   as part of close processing.
948  *
949  * At a minimum two references should still be held.
950  */
951 static void svc_rdma_detach(struct svc_xprt *xprt)
952 {
953 	struct svcxprt_rdma *rdma =
954 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
955 	dprintk("svc: svc_rdma_detach(%p)\n", xprt);
956 
957 	/* Disconnect and flush posted WQE */
958 	rdma_disconnect(rdma->sc_cm_id);
959 
960 	/* Destroy the QP if present (not a listener) */
961 	if (rdma->sc_qp && !IS_ERR(rdma->sc_qp)) {
962 		ib_destroy_qp(rdma->sc_qp);
963 		svc_xprt_put(xprt);
964 	}
965 
966 	/* Destroy the CM ID */
967 	rdma_destroy_id(rdma->sc_cm_id);
968 }
969 
970 static void svc_rdma_free(struct svc_xprt *xprt)
971 {
972 	struct svcxprt_rdma *rdma = (struct svcxprt_rdma *)xprt;
973 	dprintk("svcrdma: svc_rdma_free(%p)\n", rdma);
974 	/* We should only be called from kref_put */
975 	BUG_ON(atomic_read(&xprt->xpt_ref.refcount) != 0);
976 	if (rdma->sc_sq_cq && !IS_ERR(rdma->sc_sq_cq))
977 		ib_destroy_cq(rdma->sc_sq_cq);
978 
979 	if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq))
980 		ib_destroy_cq(rdma->sc_rq_cq);
981 
982 	if (rdma->sc_phys_mr && !IS_ERR(rdma->sc_phys_mr))
983 		ib_dereg_mr(rdma->sc_phys_mr);
984 
985 	if (rdma->sc_pd && !IS_ERR(rdma->sc_pd))
986 		ib_dealloc_pd(rdma->sc_pd);
987 
988 	destroy_context_cache(rdma->sc_ctxt_head);
989 	kfree(rdma);
990 }
991 
992 static int svc_rdma_has_wspace(struct svc_xprt *xprt)
993 {
994 	struct svcxprt_rdma *rdma =
995 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
996 
997 	/*
998 	 * If there are fewer SQ WR available than required to send a
999 	 * simple response, return false.
1000 	 */
1001 	if ((rdma->sc_sq_depth - atomic_read(&rdma->sc_sq_count) < 3))
1002 		return 0;
1003 
1004 	/*
1005 	 * ...or there are already waiters on the SQ,
1006 	 * return false.
1007 	 */
1008 	if (waitqueue_active(&rdma->sc_send_wait))
1009 		return 0;
1010 
1011 	/* Otherwise return true. */
1012 	return 1;
1013 }
1014 
1015 int svc_rdma_send(struct svcxprt_rdma *xprt, struct ib_send_wr *wr)
1016 {
1017 	struct ib_send_wr *bad_wr;
1018 	int ret;
1019 
1020 	if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1021 		return 0;
1022 
1023 	BUG_ON(wr->send_flags != IB_SEND_SIGNALED);
1024 	BUG_ON(((struct svc_rdma_op_ctxt *)(unsigned long)wr->wr_id)->wr_op !=
1025 		wr->opcode);
1026 	/* If the SQ is full, wait until an SQ entry is available */
1027 	while (1) {
1028 		spin_lock_bh(&xprt->sc_lock);
1029 		if (xprt->sc_sq_depth == atomic_read(&xprt->sc_sq_count)) {
1030 			spin_unlock_bh(&xprt->sc_lock);
1031 			atomic_inc(&rdma_stat_sq_starve);
1032 			/* See if we can reap some SQ WR */
1033 			sq_cq_reap(xprt);
1034 
1035 			/* Wait until SQ WR available if SQ still full */
1036 			wait_event(xprt->sc_send_wait,
1037 				   atomic_read(&xprt->sc_sq_count) <
1038 				   xprt->sc_sq_depth);
1039 			continue;
1040 		}
1041 		/* Bumped used SQ WR count and post */
1042 		ret = ib_post_send(xprt->sc_qp, wr, &bad_wr);
1043 		if (!ret)
1044 			atomic_inc(&xprt->sc_sq_count);
1045 		else
1046 			dprintk("svcrdma: failed to post SQ WR rc=%d, "
1047 			       "sc_sq_count=%d, sc_sq_depth=%d\n",
1048 			       ret, atomic_read(&xprt->sc_sq_count),
1049 			       xprt->sc_sq_depth);
1050 		spin_unlock_bh(&xprt->sc_lock);
1051 		break;
1052 	}
1053 	return ret;
1054 }
1055 
1056 int svc_rdma_send_error(struct svcxprt_rdma *xprt, struct rpcrdma_msg *rmsgp,
1057 			enum rpcrdma_errcode err)
1058 {
1059 	struct ib_send_wr err_wr;
1060 	struct ib_sge sge;
1061 	struct page *p;
1062 	struct svc_rdma_op_ctxt *ctxt;
1063 	u32 *va;
1064 	int length;
1065 	int ret;
1066 
1067 	p = svc_rdma_get_page();
1068 	va = page_address(p);
1069 
1070 	/* XDR encode error */
1071 	length = svc_rdma_xdr_encode_error(xprt, rmsgp, err, va);
1072 
1073 	/* Prepare SGE for local address */
1074 	sge.addr = ib_dma_map_page(xprt->sc_cm_id->device,
1075 				   p, 0, PAGE_SIZE, DMA_FROM_DEVICE);
1076 	sge.lkey = xprt->sc_phys_mr->lkey;
1077 	sge.length = length;
1078 
1079 	ctxt = svc_rdma_get_context(xprt);
1080 	ctxt->count = 1;
1081 	ctxt->pages[0] = p;
1082 
1083 	/* Prepare SEND WR */
1084 	memset(&err_wr, 0, sizeof err_wr);
1085 	ctxt->wr_op = IB_WR_SEND;
1086 	err_wr.wr_id = (unsigned long)ctxt;
1087 	err_wr.sg_list = &sge;
1088 	err_wr.num_sge = 1;
1089 	err_wr.opcode = IB_WR_SEND;
1090 	err_wr.send_flags = IB_SEND_SIGNALED;
1091 
1092 	/* Post It */
1093 	ret = svc_rdma_send(xprt, &err_wr);
1094 	if (ret) {
1095 		dprintk("svcrdma: Error posting send = %d\n", ret);
1096 		svc_rdma_put_context(ctxt, 1);
1097 	}
1098 
1099 	return ret;
1100 }
1101