xref: /linux/net/sunrpc/xprtrdma/svc_rdma_transport.c (revision 2ba9268dd603d23e17643437b2246acb6844953b)
1 /*
2  * Copyright (c) 2014 Open Grid Computing, Inc. All rights reserved.
3  * Copyright (c) 2005-2007 Network Appliance, Inc. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the BSD-type
9  * license below:
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  *
15  *      Redistributions of source code must retain the above copyright
16  *      notice, this list of conditions and the following disclaimer.
17  *
18  *      Redistributions in binary form must reproduce the above
19  *      copyright notice, this list of conditions and the following
20  *      disclaimer in the documentation and/or other materials provided
21  *      with the distribution.
22  *
23  *      Neither the name of the Network Appliance, Inc. nor the names of
24  *      its contributors may be used to endorse or promote products
25  *      derived from this software without specific prior written
26  *      permission.
27  *
28  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
29  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
30  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
31  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
32  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
33  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
34  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
35  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
36  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
37  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
38  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
39  *
40  * Author: Tom Tucker <tom@opengridcomputing.com>
41  */
42 
43 #include <linux/sunrpc/svc_xprt.h>
44 #include <linux/sunrpc/debug.h>
45 #include <linux/sunrpc/rpc_rdma.h>
46 #include <linux/interrupt.h>
47 #include <linux/sched.h>
48 #include <linux/slab.h>
49 #include <linux/spinlock.h>
50 #include <linux/workqueue.h>
51 #include <rdma/ib_verbs.h>
52 #include <rdma/rdma_cm.h>
53 #include <linux/sunrpc/svc_rdma.h>
54 #include <linux/export.h>
55 #include "xprt_rdma.h"
56 
57 #define RPCDBG_FACILITY	RPCDBG_SVCXPRT
58 
59 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
60 					struct net *net,
61 					struct sockaddr *sa, int salen,
62 					int flags);
63 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt);
64 static void svc_rdma_release_rqst(struct svc_rqst *);
65 static void dto_tasklet_func(unsigned long data);
66 static void svc_rdma_detach(struct svc_xprt *xprt);
67 static void svc_rdma_free(struct svc_xprt *xprt);
68 static int svc_rdma_has_wspace(struct svc_xprt *xprt);
69 static int svc_rdma_secure_port(struct svc_rqst *);
70 static void rq_cq_reap(struct svcxprt_rdma *xprt);
71 static void sq_cq_reap(struct svcxprt_rdma *xprt);
72 
73 static DECLARE_TASKLET(dto_tasklet, dto_tasklet_func, 0UL);
74 static DEFINE_SPINLOCK(dto_lock);
75 static LIST_HEAD(dto_xprt_q);
76 
77 static struct svc_xprt_ops svc_rdma_ops = {
78 	.xpo_create = svc_rdma_create,
79 	.xpo_recvfrom = svc_rdma_recvfrom,
80 	.xpo_sendto = svc_rdma_sendto,
81 	.xpo_release_rqst = svc_rdma_release_rqst,
82 	.xpo_detach = svc_rdma_detach,
83 	.xpo_free = svc_rdma_free,
84 	.xpo_prep_reply_hdr = svc_rdma_prep_reply_hdr,
85 	.xpo_has_wspace = svc_rdma_has_wspace,
86 	.xpo_accept = svc_rdma_accept,
87 	.xpo_secure_port = svc_rdma_secure_port,
88 };
89 
90 struct svc_xprt_class svc_rdma_class = {
91 	.xcl_name = "rdma",
92 	.xcl_owner = THIS_MODULE,
93 	.xcl_ops = &svc_rdma_ops,
94 	.xcl_max_payload = RPCSVC_MAXPAYLOAD_RDMA,
95 	.xcl_ident = XPRT_TRANSPORT_RDMA,
96 };
97 
98 struct svc_rdma_op_ctxt *svc_rdma_get_context(struct svcxprt_rdma *xprt)
99 {
100 	struct svc_rdma_op_ctxt *ctxt;
101 
102 	while (1) {
103 		ctxt = kmem_cache_alloc(svc_rdma_ctxt_cachep, GFP_KERNEL);
104 		if (ctxt)
105 			break;
106 		schedule_timeout_uninterruptible(msecs_to_jiffies(500));
107 	}
108 	ctxt->xprt = xprt;
109 	INIT_LIST_HEAD(&ctxt->dto_q);
110 	ctxt->count = 0;
111 	ctxt->frmr = NULL;
112 	atomic_inc(&xprt->sc_ctxt_used);
113 	return ctxt;
114 }
115 
116 void svc_rdma_unmap_dma(struct svc_rdma_op_ctxt *ctxt)
117 {
118 	struct svcxprt_rdma *xprt = ctxt->xprt;
119 	int i;
120 	for (i = 0; i < ctxt->count && ctxt->sge[i].length; i++) {
121 		/*
122 		 * Unmap the DMA addr in the SGE if the lkey matches
123 		 * the sc_dma_lkey, otherwise, ignore it since it is
124 		 * an FRMR lkey and will be unmapped later when the
125 		 * last WR that uses it completes.
126 		 */
127 		if (ctxt->sge[i].lkey == xprt->sc_dma_lkey) {
128 			atomic_dec(&xprt->sc_dma_used);
129 			ib_dma_unmap_page(xprt->sc_cm_id->device,
130 					    ctxt->sge[i].addr,
131 					    ctxt->sge[i].length,
132 					    ctxt->direction);
133 		}
134 	}
135 }
136 
137 void svc_rdma_put_context(struct svc_rdma_op_ctxt *ctxt, int free_pages)
138 {
139 	struct svcxprt_rdma *xprt;
140 	int i;
141 
142 	xprt = ctxt->xprt;
143 	if (free_pages)
144 		for (i = 0; i < ctxt->count; i++)
145 			put_page(ctxt->pages[i]);
146 
147 	kmem_cache_free(svc_rdma_ctxt_cachep, ctxt);
148 	atomic_dec(&xprt->sc_ctxt_used);
149 }
150 
151 /*
152  * Temporary NFS req mappings are shared across all transport
153  * instances. These are short lived and should be bounded by the number
154  * of concurrent server threads * depth of the SQ.
155  */
156 struct svc_rdma_req_map *svc_rdma_get_req_map(void)
157 {
158 	struct svc_rdma_req_map *map;
159 	while (1) {
160 		map = kmem_cache_alloc(svc_rdma_map_cachep, GFP_KERNEL);
161 		if (map)
162 			break;
163 		schedule_timeout_uninterruptible(msecs_to_jiffies(500));
164 	}
165 	map->count = 0;
166 	return map;
167 }
168 
169 void svc_rdma_put_req_map(struct svc_rdma_req_map *map)
170 {
171 	kmem_cache_free(svc_rdma_map_cachep, map);
172 }
173 
174 /* ib_cq event handler */
175 static void cq_event_handler(struct ib_event *event, void *context)
176 {
177 	struct svc_xprt *xprt = context;
178 	dprintk("svcrdma: received CQ event id=%d, context=%p\n",
179 		event->event, context);
180 	set_bit(XPT_CLOSE, &xprt->xpt_flags);
181 }
182 
183 /* QP event handler */
184 static void qp_event_handler(struct ib_event *event, void *context)
185 {
186 	struct svc_xprt *xprt = context;
187 
188 	switch (event->event) {
189 	/* These are considered benign events */
190 	case IB_EVENT_PATH_MIG:
191 	case IB_EVENT_COMM_EST:
192 	case IB_EVENT_SQ_DRAINED:
193 	case IB_EVENT_QP_LAST_WQE_REACHED:
194 		dprintk("svcrdma: QP event %d received for QP=%p\n",
195 			event->event, event->element.qp);
196 		break;
197 	/* These are considered fatal events */
198 	case IB_EVENT_PATH_MIG_ERR:
199 	case IB_EVENT_QP_FATAL:
200 	case IB_EVENT_QP_REQ_ERR:
201 	case IB_EVENT_QP_ACCESS_ERR:
202 	case IB_EVENT_DEVICE_FATAL:
203 	default:
204 		dprintk("svcrdma: QP ERROR event %d received for QP=%p, "
205 			"closing transport\n",
206 			event->event, event->element.qp);
207 		set_bit(XPT_CLOSE, &xprt->xpt_flags);
208 		break;
209 	}
210 }
211 
212 /*
213  * Data Transfer Operation Tasklet
214  *
215  * Walks a list of transports with I/O pending, removing entries as
216  * they are added to the server's I/O pending list. Two bits indicate
217  * if SQ, RQ, or both have I/O pending. The dto_lock is an irqsave
218  * spinlock that serializes access to the transport list with the RQ
219  * and SQ interrupt handlers.
220  */
221 static void dto_tasklet_func(unsigned long data)
222 {
223 	struct svcxprt_rdma *xprt;
224 	unsigned long flags;
225 
226 	spin_lock_irqsave(&dto_lock, flags);
227 	while (!list_empty(&dto_xprt_q)) {
228 		xprt = list_entry(dto_xprt_q.next,
229 				  struct svcxprt_rdma, sc_dto_q);
230 		list_del_init(&xprt->sc_dto_q);
231 		spin_unlock_irqrestore(&dto_lock, flags);
232 
233 		rq_cq_reap(xprt);
234 		sq_cq_reap(xprt);
235 
236 		svc_xprt_put(&xprt->sc_xprt);
237 		spin_lock_irqsave(&dto_lock, flags);
238 	}
239 	spin_unlock_irqrestore(&dto_lock, flags);
240 }
241 
242 /*
243  * Receive Queue Completion Handler
244  *
245  * Since an RQ completion handler is called on interrupt context, we
246  * need to defer the handling of the I/O to a tasklet
247  */
248 static void rq_comp_handler(struct ib_cq *cq, void *cq_context)
249 {
250 	struct svcxprt_rdma *xprt = cq_context;
251 	unsigned long flags;
252 
253 	/* Guard against unconditional flush call for destroyed QP */
254 	if (atomic_read(&xprt->sc_xprt.xpt_ref.refcount)==0)
255 		return;
256 
257 	/*
258 	 * Set the bit regardless of whether or not it's on the list
259 	 * because it may be on the list already due to an SQ
260 	 * completion.
261 	 */
262 	set_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags);
263 
264 	/*
265 	 * If this transport is not already on the DTO transport queue,
266 	 * add it
267 	 */
268 	spin_lock_irqsave(&dto_lock, flags);
269 	if (list_empty(&xprt->sc_dto_q)) {
270 		svc_xprt_get(&xprt->sc_xprt);
271 		list_add_tail(&xprt->sc_dto_q, &dto_xprt_q);
272 	}
273 	spin_unlock_irqrestore(&dto_lock, flags);
274 
275 	/* Tasklet does all the work to avoid irqsave locks. */
276 	tasklet_schedule(&dto_tasklet);
277 }
278 
279 /*
280  * rq_cq_reap - Process the RQ CQ.
281  *
282  * Take all completing WC off the CQE and enqueue the associated DTO
283  * context on the dto_q for the transport.
284  *
285  * Note that caller must hold a transport reference.
286  */
287 static void rq_cq_reap(struct svcxprt_rdma *xprt)
288 {
289 	int ret;
290 	struct ib_wc wc;
291 	struct svc_rdma_op_ctxt *ctxt = NULL;
292 
293 	if (!test_and_clear_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags))
294 		return;
295 
296 	ib_req_notify_cq(xprt->sc_rq_cq, IB_CQ_NEXT_COMP);
297 	atomic_inc(&rdma_stat_rq_poll);
298 
299 	while ((ret = ib_poll_cq(xprt->sc_rq_cq, 1, &wc)) > 0) {
300 		ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id;
301 		ctxt->wc_status = wc.status;
302 		ctxt->byte_len = wc.byte_len;
303 		svc_rdma_unmap_dma(ctxt);
304 		if (wc.status != IB_WC_SUCCESS) {
305 			/* Close the transport */
306 			dprintk("svcrdma: transport closing putting ctxt %p\n", ctxt);
307 			set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
308 			svc_rdma_put_context(ctxt, 1);
309 			svc_xprt_put(&xprt->sc_xprt);
310 			continue;
311 		}
312 		spin_lock_bh(&xprt->sc_rq_dto_lock);
313 		list_add_tail(&ctxt->dto_q, &xprt->sc_rq_dto_q);
314 		spin_unlock_bh(&xprt->sc_rq_dto_lock);
315 		svc_xprt_put(&xprt->sc_xprt);
316 	}
317 
318 	if (ctxt)
319 		atomic_inc(&rdma_stat_rq_prod);
320 
321 	set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
322 	/*
323 	 * If data arrived before established event,
324 	 * don't enqueue. This defers RPC I/O until the
325 	 * RDMA connection is complete.
326 	 */
327 	if (!test_bit(RDMAXPRT_CONN_PENDING, &xprt->sc_flags))
328 		svc_xprt_enqueue(&xprt->sc_xprt);
329 }
330 
331 /*
332  * Process a completion context
333  */
334 static void process_context(struct svcxprt_rdma *xprt,
335 			    struct svc_rdma_op_ctxt *ctxt)
336 {
337 	svc_rdma_unmap_dma(ctxt);
338 
339 	switch (ctxt->wr_op) {
340 	case IB_WR_SEND:
341 		if (ctxt->frmr)
342 			pr_err("svcrdma: SEND: ctxt->frmr != NULL\n");
343 		svc_rdma_put_context(ctxt, 1);
344 		break;
345 
346 	case IB_WR_RDMA_WRITE:
347 		if (ctxt->frmr)
348 			pr_err("svcrdma: WRITE: ctxt->frmr != NULL\n");
349 		svc_rdma_put_context(ctxt, 0);
350 		break;
351 
352 	case IB_WR_RDMA_READ:
353 	case IB_WR_RDMA_READ_WITH_INV:
354 		svc_rdma_put_frmr(xprt, ctxt->frmr);
355 		if (test_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags)) {
356 			struct svc_rdma_op_ctxt *read_hdr = ctxt->read_hdr;
357 			if (read_hdr) {
358 				spin_lock_bh(&xprt->sc_rq_dto_lock);
359 				set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
360 				list_add_tail(&read_hdr->dto_q,
361 					      &xprt->sc_read_complete_q);
362 				spin_unlock_bh(&xprt->sc_rq_dto_lock);
363 			} else {
364 				pr_err("svcrdma: ctxt->read_hdr == NULL\n");
365 			}
366 			svc_xprt_enqueue(&xprt->sc_xprt);
367 		}
368 		svc_rdma_put_context(ctxt, 0);
369 		break;
370 
371 	default:
372 		printk(KERN_ERR "svcrdma: unexpected completion type, "
373 		       "opcode=%d\n",
374 		       ctxt->wr_op);
375 		break;
376 	}
377 }
378 
379 /*
380  * Send Queue Completion Handler - potentially called on interrupt context.
381  *
382  * Note that caller must hold a transport reference.
383  */
384 static void sq_cq_reap(struct svcxprt_rdma *xprt)
385 {
386 	struct svc_rdma_op_ctxt *ctxt = NULL;
387 	struct ib_wc wc_a[6];
388 	struct ib_wc *wc;
389 	struct ib_cq *cq = xprt->sc_sq_cq;
390 	int ret;
391 
392 	memset(wc_a, 0, sizeof(wc_a));
393 
394 	if (!test_and_clear_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags))
395 		return;
396 
397 	ib_req_notify_cq(xprt->sc_sq_cq, IB_CQ_NEXT_COMP);
398 	atomic_inc(&rdma_stat_sq_poll);
399 	while ((ret = ib_poll_cq(cq, ARRAY_SIZE(wc_a), wc_a)) > 0) {
400 		int i;
401 
402 		for (i = 0; i < ret; i++) {
403 			wc = &wc_a[i];
404 			if (wc->status != IB_WC_SUCCESS) {
405 				dprintk("svcrdma: sq wc err status %d\n",
406 					wc->status);
407 
408 				/* Close the transport */
409 				set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
410 			}
411 
412 			/* Decrement used SQ WR count */
413 			atomic_dec(&xprt->sc_sq_count);
414 			wake_up(&xprt->sc_send_wait);
415 
416 			ctxt = (struct svc_rdma_op_ctxt *)
417 				(unsigned long)wc->wr_id;
418 			if (ctxt)
419 				process_context(xprt, ctxt);
420 
421 			svc_xprt_put(&xprt->sc_xprt);
422 		}
423 	}
424 
425 	if (ctxt)
426 		atomic_inc(&rdma_stat_sq_prod);
427 }
428 
429 static void sq_comp_handler(struct ib_cq *cq, void *cq_context)
430 {
431 	struct svcxprt_rdma *xprt = cq_context;
432 	unsigned long flags;
433 
434 	/* Guard against unconditional flush call for destroyed QP */
435 	if (atomic_read(&xprt->sc_xprt.xpt_ref.refcount)==0)
436 		return;
437 
438 	/*
439 	 * Set the bit regardless of whether or not it's on the list
440 	 * because it may be on the list already due to an RQ
441 	 * completion.
442 	 */
443 	set_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags);
444 
445 	/*
446 	 * If this transport is not already on the DTO transport queue,
447 	 * add it
448 	 */
449 	spin_lock_irqsave(&dto_lock, flags);
450 	if (list_empty(&xprt->sc_dto_q)) {
451 		svc_xprt_get(&xprt->sc_xprt);
452 		list_add_tail(&xprt->sc_dto_q, &dto_xprt_q);
453 	}
454 	spin_unlock_irqrestore(&dto_lock, flags);
455 
456 	/* Tasklet does all the work to avoid irqsave locks. */
457 	tasklet_schedule(&dto_tasklet);
458 }
459 
460 static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *serv,
461 					     int listener)
462 {
463 	struct svcxprt_rdma *cma_xprt = kzalloc(sizeof *cma_xprt, GFP_KERNEL);
464 
465 	if (!cma_xprt)
466 		return NULL;
467 	svc_xprt_init(&init_net, &svc_rdma_class, &cma_xprt->sc_xprt, serv);
468 	INIT_LIST_HEAD(&cma_xprt->sc_accept_q);
469 	INIT_LIST_HEAD(&cma_xprt->sc_dto_q);
470 	INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q);
471 	INIT_LIST_HEAD(&cma_xprt->sc_read_complete_q);
472 	INIT_LIST_HEAD(&cma_xprt->sc_frmr_q);
473 	init_waitqueue_head(&cma_xprt->sc_send_wait);
474 
475 	spin_lock_init(&cma_xprt->sc_lock);
476 	spin_lock_init(&cma_xprt->sc_rq_dto_lock);
477 	spin_lock_init(&cma_xprt->sc_frmr_q_lock);
478 
479 	cma_xprt->sc_ord = svcrdma_ord;
480 
481 	cma_xprt->sc_max_req_size = svcrdma_max_req_size;
482 	cma_xprt->sc_max_requests = svcrdma_max_requests;
483 	cma_xprt->sc_sq_depth = svcrdma_max_requests * RPCRDMA_SQ_DEPTH_MULT;
484 	atomic_set(&cma_xprt->sc_sq_count, 0);
485 	atomic_set(&cma_xprt->sc_ctxt_used, 0);
486 
487 	if (listener)
488 		set_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
489 
490 	return cma_xprt;
491 }
492 
493 struct page *svc_rdma_get_page(void)
494 {
495 	struct page *page;
496 
497 	while ((page = alloc_page(GFP_KERNEL)) == NULL) {
498 		/* If we can't get memory, wait a bit and try again */
499 		printk(KERN_INFO "svcrdma: out of memory...retrying in 1s\n");
500 		schedule_timeout_uninterruptible(msecs_to_jiffies(1000));
501 	}
502 	return page;
503 }
504 
505 int svc_rdma_post_recv(struct svcxprt_rdma *xprt)
506 {
507 	struct ib_recv_wr recv_wr, *bad_recv_wr;
508 	struct svc_rdma_op_ctxt *ctxt;
509 	struct page *page;
510 	dma_addr_t pa;
511 	int sge_no;
512 	int buflen;
513 	int ret;
514 
515 	ctxt = svc_rdma_get_context(xprt);
516 	buflen = 0;
517 	ctxt->direction = DMA_FROM_DEVICE;
518 	for (sge_no = 0; buflen < xprt->sc_max_req_size; sge_no++) {
519 		if (sge_no >= xprt->sc_max_sge) {
520 			pr_err("svcrdma: Too many sges (%d)\n", sge_no);
521 			goto err_put_ctxt;
522 		}
523 		page = svc_rdma_get_page();
524 		ctxt->pages[sge_no] = page;
525 		pa = ib_dma_map_page(xprt->sc_cm_id->device,
526 				     page, 0, PAGE_SIZE,
527 				     DMA_FROM_DEVICE);
528 		if (ib_dma_mapping_error(xprt->sc_cm_id->device, pa))
529 			goto err_put_ctxt;
530 		atomic_inc(&xprt->sc_dma_used);
531 		ctxt->sge[sge_no].addr = pa;
532 		ctxt->sge[sge_no].length = PAGE_SIZE;
533 		ctxt->sge[sge_no].lkey = xprt->sc_dma_lkey;
534 		ctxt->count = sge_no + 1;
535 		buflen += PAGE_SIZE;
536 	}
537 	recv_wr.next = NULL;
538 	recv_wr.sg_list = &ctxt->sge[0];
539 	recv_wr.num_sge = ctxt->count;
540 	recv_wr.wr_id = (u64)(unsigned long)ctxt;
541 
542 	svc_xprt_get(&xprt->sc_xprt);
543 	ret = ib_post_recv(xprt->sc_qp, &recv_wr, &bad_recv_wr);
544 	if (ret) {
545 		svc_rdma_unmap_dma(ctxt);
546 		svc_rdma_put_context(ctxt, 1);
547 		svc_xprt_put(&xprt->sc_xprt);
548 	}
549 	return ret;
550 
551  err_put_ctxt:
552 	svc_rdma_unmap_dma(ctxt);
553 	svc_rdma_put_context(ctxt, 1);
554 	return -ENOMEM;
555 }
556 
557 /*
558  * This function handles the CONNECT_REQUEST event on a listening
559  * endpoint. It is passed the cma_id for the _new_ connection. The context in
560  * this cma_id is inherited from the listening cma_id and is the svc_xprt
561  * structure for the listening endpoint.
562  *
563  * This function creates a new xprt for the new connection and enqueues it on
564  * the accept queue for the listent xprt. When the listen thread is kicked, it
565  * will call the recvfrom method on the listen xprt which will accept the new
566  * connection.
567  */
568 static void handle_connect_req(struct rdma_cm_id *new_cma_id, size_t client_ird)
569 {
570 	struct svcxprt_rdma *listen_xprt = new_cma_id->context;
571 	struct svcxprt_rdma *newxprt;
572 	struct sockaddr *sa;
573 
574 	/* Create a new transport */
575 	newxprt = rdma_create_xprt(listen_xprt->sc_xprt.xpt_server, 0);
576 	if (!newxprt) {
577 		dprintk("svcrdma: failed to create new transport\n");
578 		return;
579 	}
580 	newxprt->sc_cm_id = new_cma_id;
581 	new_cma_id->context = newxprt;
582 	dprintk("svcrdma: Creating newxprt=%p, cm_id=%p, listenxprt=%p\n",
583 		newxprt, newxprt->sc_cm_id, listen_xprt);
584 
585 	/* Save client advertised inbound read limit for use later in accept. */
586 	newxprt->sc_ord = client_ird;
587 
588 	/* Set the local and remote addresses in the transport */
589 	sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
590 	svc_xprt_set_remote(&newxprt->sc_xprt, sa, svc_addr_len(sa));
591 	sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
592 	svc_xprt_set_local(&newxprt->sc_xprt, sa, svc_addr_len(sa));
593 
594 	/*
595 	 * Enqueue the new transport on the accept queue of the listening
596 	 * transport
597 	 */
598 	spin_lock_bh(&listen_xprt->sc_lock);
599 	list_add_tail(&newxprt->sc_accept_q, &listen_xprt->sc_accept_q);
600 	spin_unlock_bh(&listen_xprt->sc_lock);
601 
602 	set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags);
603 	svc_xprt_enqueue(&listen_xprt->sc_xprt);
604 }
605 
606 /*
607  * Handles events generated on the listening endpoint. These events will be
608  * either be incoming connect requests or adapter removal  events.
609  */
610 static int rdma_listen_handler(struct rdma_cm_id *cma_id,
611 			       struct rdma_cm_event *event)
612 {
613 	struct svcxprt_rdma *xprt = cma_id->context;
614 	int ret = 0;
615 
616 	switch (event->event) {
617 	case RDMA_CM_EVENT_CONNECT_REQUEST:
618 		dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, "
619 			"event=%d\n", cma_id, cma_id->context, event->event);
620 		handle_connect_req(cma_id,
621 				   event->param.conn.initiator_depth);
622 		break;
623 
624 	case RDMA_CM_EVENT_ESTABLISHED:
625 		/* Accept complete */
626 		dprintk("svcrdma: Connection completed on LISTEN xprt=%p, "
627 			"cm_id=%p\n", xprt, cma_id);
628 		break;
629 
630 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
631 		dprintk("svcrdma: Device removal xprt=%p, cm_id=%p\n",
632 			xprt, cma_id);
633 		if (xprt)
634 			set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
635 		break;
636 
637 	default:
638 		dprintk("svcrdma: Unexpected event on listening endpoint %p, "
639 			"event=%d\n", cma_id, event->event);
640 		break;
641 	}
642 
643 	return ret;
644 }
645 
646 static int rdma_cma_handler(struct rdma_cm_id *cma_id,
647 			    struct rdma_cm_event *event)
648 {
649 	struct svc_xprt *xprt = cma_id->context;
650 	struct svcxprt_rdma *rdma =
651 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
652 	switch (event->event) {
653 	case RDMA_CM_EVENT_ESTABLISHED:
654 		/* Accept complete */
655 		svc_xprt_get(xprt);
656 		dprintk("svcrdma: Connection completed on DTO xprt=%p, "
657 			"cm_id=%p\n", xprt, cma_id);
658 		clear_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags);
659 		svc_xprt_enqueue(xprt);
660 		break;
661 	case RDMA_CM_EVENT_DISCONNECTED:
662 		dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n",
663 			xprt, cma_id);
664 		if (xprt) {
665 			set_bit(XPT_CLOSE, &xprt->xpt_flags);
666 			svc_xprt_enqueue(xprt);
667 			svc_xprt_put(xprt);
668 		}
669 		break;
670 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
671 		dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, "
672 			"event=%d\n", cma_id, xprt, event->event);
673 		if (xprt) {
674 			set_bit(XPT_CLOSE, &xprt->xpt_flags);
675 			svc_xprt_enqueue(xprt);
676 		}
677 		break;
678 	default:
679 		dprintk("svcrdma: Unexpected event on DTO endpoint %p, "
680 			"event=%d\n", cma_id, event->event);
681 		break;
682 	}
683 	return 0;
684 }
685 
686 /*
687  * Create a listening RDMA service endpoint.
688  */
689 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
690 					struct net *net,
691 					struct sockaddr *sa, int salen,
692 					int flags)
693 {
694 	struct rdma_cm_id *listen_id;
695 	struct svcxprt_rdma *cma_xprt;
696 	int ret;
697 
698 	dprintk("svcrdma: Creating RDMA socket\n");
699 	if (sa->sa_family != AF_INET) {
700 		dprintk("svcrdma: Address family %d is not supported.\n", sa->sa_family);
701 		return ERR_PTR(-EAFNOSUPPORT);
702 	}
703 	cma_xprt = rdma_create_xprt(serv, 1);
704 	if (!cma_xprt)
705 		return ERR_PTR(-ENOMEM);
706 
707 	listen_id = rdma_create_id(rdma_listen_handler, cma_xprt, RDMA_PS_TCP,
708 				   IB_QPT_RC);
709 	if (IS_ERR(listen_id)) {
710 		ret = PTR_ERR(listen_id);
711 		dprintk("svcrdma: rdma_create_id failed = %d\n", ret);
712 		goto err0;
713 	}
714 
715 	ret = rdma_bind_addr(listen_id, sa);
716 	if (ret) {
717 		dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret);
718 		goto err1;
719 	}
720 	cma_xprt->sc_cm_id = listen_id;
721 
722 	ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG);
723 	if (ret) {
724 		dprintk("svcrdma: rdma_listen failed = %d\n", ret);
725 		goto err1;
726 	}
727 
728 	/*
729 	 * We need to use the address from the cm_id in case the
730 	 * caller specified 0 for the port number.
731 	 */
732 	sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr;
733 	svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen);
734 
735 	return &cma_xprt->sc_xprt;
736 
737  err1:
738 	rdma_destroy_id(listen_id);
739  err0:
740 	kfree(cma_xprt);
741 	return ERR_PTR(ret);
742 }
743 
744 static struct svc_rdma_fastreg_mr *rdma_alloc_frmr(struct svcxprt_rdma *xprt)
745 {
746 	struct ib_mr *mr;
747 	struct ib_fast_reg_page_list *pl;
748 	struct svc_rdma_fastreg_mr *frmr;
749 
750 	frmr = kmalloc(sizeof(*frmr), GFP_KERNEL);
751 	if (!frmr)
752 		goto err;
753 
754 	mr = ib_alloc_fast_reg_mr(xprt->sc_pd, RPCSVC_MAXPAGES);
755 	if (IS_ERR(mr))
756 		goto err_free_frmr;
757 
758 	pl = ib_alloc_fast_reg_page_list(xprt->sc_cm_id->device,
759 					 RPCSVC_MAXPAGES);
760 	if (IS_ERR(pl))
761 		goto err_free_mr;
762 
763 	frmr->mr = mr;
764 	frmr->page_list = pl;
765 	INIT_LIST_HEAD(&frmr->frmr_list);
766 	return frmr;
767 
768  err_free_mr:
769 	ib_dereg_mr(mr);
770  err_free_frmr:
771 	kfree(frmr);
772  err:
773 	return ERR_PTR(-ENOMEM);
774 }
775 
776 static void rdma_dealloc_frmr_q(struct svcxprt_rdma *xprt)
777 {
778 	struct svc_rdma_fastreg_mr *frmr;
779 
780 	while (!list_empty(&xprt->sc_frmr_q)) {
781 		frmr = list_entry(xprt->sc_frmr_q.next,
782 				  struct svc_rdma_fastreg_mr, frmr_list);
783 		list_del_init(&frmr->frmr_list);
784 		ib_dereg_mr(frmr->mr);
785 		ib_free_fast_reg_page_list(frmr->page_list);
786 		kfree(frmr);
787 	}
788 }
789 
790 struct svc_rdma_fastreg_mr *svc_rdma_get_frmr(struct svcxprt_rdma *rdma)
791 {
792 	struct svc_rdma_fastreg_mr *frmr = NULL;
793 
794 	spin_lock_bh(&rdma->sc_frmr_q_lock);
795 	if (!list_empty(&rdma->sc_frmr_q)) {
796 		frmr = list_entry(rdma->sc_frmr_q.next,
797 				  struct svc_rdma_fastreg_mr, frmr_list);
798 		list_del_init(&frmr->frmr_list);
799 		frmr->map_len = 0;
800 		frmr->page_list_len = 0;
801 	}
802 	spin_unlock_bh(&rdma->sc_frmr_q_lock);
803 	if (frmr)
804 		return frmr;
805 
806 	return rdma_alloc_frmr(rdma);
807 }
808 
809 static void frmr_unmap_dma(struct svcxprt_rdma *xprt,
810 			   struct svc_rdma_fastreg_mr *frmr)
811 {
812 	int page_no;
813 	for (page_no = 0; page_no < frmr->page_list_len; page_no++) {
814 		dma_addr_t addr = frmr->page_list->page_list[page_no];
815 		if (ib_dma_mapping_error(frmr->mr->device, addr))
816 			continue;
817 		atomic_dec(&xprt->sc_dma_used);
818 		ib_dma_unmap_page(frmr->mr->device, addr, PAGE_SIZE,
819 				  frmr->direction);
820 	}
821 }
822 
823 void svc_rdma_put_frmr(struct svcxprt_rdma *rdma,
824 		       struct svc_rdma_fastreg_mr *frmr)
825 {
826 	if (frmr) {
827 		frmr_unmap_dma(rdma, frmr);
828 		spin_lock_bh(&rdma->sc_frmr_q_lock);
829 		WARN_ON_ONCE(!list_empty(&frmr->frmr_list));
830 		list_add(&frmr->frmr_list, &rdma->sc_frmr_q);
831 		spin_unlock_bh(&rdma->sc_frmr_q_lock);
832 	}
833 }
834 
835 /*
836  * This is the xpo_recvfrom function for listening endpoints. Its
837  * purpose is to accept incoming connections. The CMA callback handler
838  * has already created a new transport and attached it to the new CMA
839  * ID.
840  *
841  * There is a queue of pending connections hung on the listening
842  * transport. This queue contains the new svc_xprt structure. This
843  * function takes svc_xprt structures off the accept_q and completes
844  * the connection.
845  */
846 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
847 {
848 	struct svcxprt_rdma *listen_rdma;
849 	struct svcxprt_rdma *newxprt = NULL;
850 	struct rdma_conn_param conn_param;
851 	struct ib_qp_init_attr qp_attr;
852 	struct ib_device_attr devattr;
853 	int uninitialized_var(dma_mr_acc);
854 	int need_dma_mr;
855 	int ret;
856 	int i;
857 
858 	listen_rdma = container_of(xprt, struct svcxprt_rdma, sc_xprt);
859 	clear_bit(XPT_CONN, &xprt->xpt_flags);
860 	/* Get the next entry off the accept list */
861 	spin_lock_bh(&listen_rdma->sc_lock);
862 	if (!list_empty(&listen_rdma->sc_accept_q)) {
863 		newxprt = list_entry(listen_rdma->sc_accept_q.next,
864 				     struct svcxprt_rdma, sc_accept_q);
865 		list_del_init(&newxprt->sc_accept_q);
866 	}
867 	if (!list_empty(&listen_rdma->sc_accept_q))
868 		set_bit(XPT_CONN, &listen_rdma->sc_xprt.xpt_flags);
869 	spin_unlock_bh(&listen_rdma->sc_lock);
870 	if (!newxprt)
871 		return NULL;
872 
873 	dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n",
874 		newxprt, newxprt->sc_cm_id);
875 
876 	ret = ib_query_device(newxprt->sc_cm_id->device, &devattr);
877 	if (ret) {
878 		dprintk("svcrdma: could not query device attributes on "
879 			"device %p, rc=%d\n", newxprt->sc_cm_id->device, ret);
880 		goto errout;
881 	}
882 
883 	/* Qualify the transport resource defaults with the
884 	 * capabilities of this particular device */
885 	newxprt->sc_max_sge = min((size_t)devattr.max_sge,
886 				  (size_t)RPCSVC_MAXPAGES);
887 	newxprt->sc_max_requests = min((size_t)devattr.max_qp_wr,
888 				   (size_t)svcrdma_max_requests);
889 	newxprt->sc_sq_depth = RPCRDMA_SQ_DEPTH_MULT * newxprt->sc_max_requests;
890 
891 	/*
892 	 * Limit ORD based on client limit, local device limit, and
893 	 * configured svcrdma limit.
894 	 */
895 	newxprt->sc_ord = min_t(size_t, devattr.max_qp_rd_atom, newxprt->sc_ord);
896 	newxprt->sc_ord = min_t(size_t,	svcrdma_ord, newxprt->sc_ord);
897 
898 	newxprt->sc_pd = ib_alloc_pd(newxprt->sc_cm_id->device);
899 	if (IS_ERR(newxprt->sc_pd)) {
900 		dprintk("svcrdma: error creating PD for connect request\n");
901 		goto errout;
902 	}
903 	newxprt->sc_sq_cq = ib_create_cq(newxprt->sc_cm_id->device,
904 					 sq_comp_handler,
905 					 cq_event_handler,
906 					 newxprt,
907 					 newxprt->sc_sq_depth,
908 					 0);
909 	if (IS_ERR(newxprt->sc_sq_cq)) {
910 		dprintk("svcrdma: error creating SQ CQ for connect request\n");
911 		goto errout;
912 	}
913 	newxprt->sc_rq_cq = ib_create_cq(newxprt->sc_cm_id->device,
914 					 rq_comp_handler,
915 					 cq_event_handler,
916 					 newxprt,
917 					 newxprt->sc_max_requests,
918 					 0);
919 	if (IS_ERR(newxprt->sc_rq_cq)) {
920 		dprintk("svcrdma: error creating RQ CQ for connect request\n");
921 		goto errout;
922 	}
923 
924 	memset(&qp_attr, 0, sizeof qp_attr);
925 	qp_attr.event_handler = qp_event_handler;
926 	qp_attr.qp_context = &newxprt->sc_xprt;
927 	qp_attr.cap.max_send_wr = newxprt->sc_sq_depth;
928 	qp_attr.cap.max_recv_wr = newxprt->sc_max_requests;
929 	qp_attr.cap.max_send_sge = newxprt->sc_max_sge;
930 	qp_attr.cap.max_recv_sge = newxprt->sc_max_sge;
931 	qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
932 	qp_attr.qp_type = IB_QPT_RC;
933 	qp_attr.send_cq = newxprt->sc_sq_cq;
934 	qp_attr.recv_cq = newxprt->sc_rq_cq;
935 	dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n"
936 		"    cm_id->device=%p, sc_pd->device=%p\n"
937 		"    cap.max_send_wr = %d\n"
938 		"    cap.max_recv_wr = %d\n"
939 		"    cap.max_send_sge = %d\n"
940 		"    cap.max_recv_sge = %d\n",
941 		newxprt->sc_cm_id, newxprt->sc_pd,
942 		newxprt->sc_cm_id->device, newxprt->sc_pd->device,
943 		qp_attr.cap.max_send_wr,
944 		qp_attr.cap.max_recv_wr,
945 		qp_attr.cap.max_send_sge,
946 		qp_attr.cap.max_recv_sge);
947 
948 	ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr);
949 	if (ret) {
950 		dprintk("svcrdma: failed to create QP, ret=%d\n", ret);
951 		goto errout;
952 	}
953 	newxprt->sc_qp = newxprt->sc_cm_id->qp;
954 
955 	/*
956 	 * Use the most secure set of MR resources based on the
957 	 * transport type and available memory management features in
958 	 * the device. Here's the table implemented below:
959 	 *
960 	 *		Fast	Global	DMA	Remote WR
961 	 *		Reg	LKEY	MR	Access
962 	 *		Sup'd	Sup'd	Needed	Needed
963 	 *
964 	 * IWARP	N	N	Y	Y
965 	 *		N	Y	Y	Y
966 	 *		Y	N	Y	N
967 	 *		Y	Y	N	-
968 	 *
969 	 * IB		N	N	Y	N
970 	 *		N	Y	N	-
971 	 *		Y	N	Y	N
972 	 *		Y	Y	N	-
973 	 *
974 	 * NB:	iWARP requires remote write access for the data sink
975 	 *	of an RDMA_READ. IB does not.
976 	 */
977 	newxprt->sc_reader = rdma_read_chunk_lcl;
978 	if (devattr.device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
979 		newxprt->sc_frmr_pg_list_len =
980 			devattr.max_fast_reg_page_list_len;
981 		newxprt->sc_dev_caps |= SVCRDMA_DEVCAP_FAST_REG;
982 		newxprt->sc_reader = rdma_read_chunk_frmr;
983 	}
984 
985 	/*
986 	 * Determine if a DMA MR is required and if so, what privs are required
987 	 */
988 	switch (rdma_node_get_transport(newxprt->sc_cm_id->device->node_type)) {
989 	case RDMA_TRANSPORT_IWARP:
990 		newxprt->sc_dev_caps |= SVCRDMA_DEVCAP_READ_W_INV;
991 		if (!(newxprt->sc_dev_caps & SVCRDMA_DEVCAP_FAST_REG)) {
992 			need_dma_mr = 1;
993 			dma_mr_acc =
994 				(IB_ACCESS_LOCAL_WRITE |
995 				 IB_ACCESS_REMOTE_WRITE);
996 		} else if (!(devattr.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)) {
997 			need_dma_mr = 1;
998 			dma_mr_acc = IB_ACCESS_LOCAL_WRITE;
999 		} else
1000 			need_dma_mr = 0;
1001 		break;
1002 	case RDMA_TRANSPORT_IB:
1003 		if (!(newxprt->sc_dev_caps & SVCRDMA_DEVCAP_FAST_REG)) {
1004 			need_dma_mr = 1;
1005 			dma_mr_acc = IB_ACCESS_LOCAL_WRITE;
1006 		} else if (!(devattr.device_cap_flags &
1007 			     IB_DEVICE_LOCAL_DMA_LKEY)) {
1008 			need_dma_mr = 1;
1009 			dma_mr_acc = IB_ACCESS_LOCAL_WRITE;
1010 		} else
1011 			need_dma_mr = 0;
1012 		break;
1013 	default:
1014 		goto errout;
1015 	}
1016 
1017 	/* Create the DMA MR if needed, otherwise, use the DMA LKEY */
1018 	if (need_dma_mr) {
1019 		/* Register all of physical memory */
1020 		newxprt->sc_phys_mr =
1021 			ib_get_dma_mr(newxprt->sc_pd, dma_mr_acc);
1022 		if (IS_ERR(newxprt->sc_phys_mr)) {
1023 			dprintk("svcrdma: Failed to create DMA MR ret=%d\n",
1024 				ret);
1025 			goto errout;
1026 		}
1027 		newxprt->sc_dma_lkey = newxprt->sc_phys_mr->lkey;
1028 	} else
1029 		newxprt->sc_dma_lkey =
1030 			newxprt->sc_cm_id->device->local_dma_lkey;
1031 
1032 	/* Post receive buffers */
1033 	for (i = 0; i < newxprt->sc_max_requests; i++) {
1034 		ret = svc_rdma_post_recv(newxprt);
1035 		if (ret) {
1036 			dprintk("svcrdma: failure posting receive buffers\n");
1037 			goto errout;
1038 		}
1039 	}
1040 
1041 	/* Swap out the handler */
1042 	newxprt->sc_cm_id->event_handler = rdma_cma_handler;
1043 
1044 	/*
1045 	 * Arm the CQs for the SQ and RQ before accepting so we can't
1046 	 * miss the first message
1047 	 */
1048 	ib_req_notify_cq(newxprt->sc_sq_cq, IB_CQ_NEXT_COMP);
1049 	ib_req_notify_cq(newxprt->sc_rq_cq, IB_CQ_NEXT_COMP);
1050 
1051 	/* Accept Connection */
1052 	set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags);
1053 	memset(&conn_param, 0, sizeof conn_param);
1054 	conn_param.responder_resources = 0;
1055 	conn_param.initiator_depth = newxprt->sc_ord;
1056 	ret = rdma_accept(newxprt->sc_cm_id, &conn_param);
1057 	if (ret) {
1058 		dprintk("svcrdma: failed to accept new connection, ret=%d\n",
1059 		       ret);
1060 		goto errout;
1061 	}
1062 
1063 	dprintk("svcrdma: new connection %p accepted with the following "
1064 		"attributes:\n"
1065 		"    local_ip        : %pI4\n"
1066 		"    local_port	     : %d\n"
1067 		"    remote_ip       : %pI4\n"
1068 		"    remote_port     : %d\n"
1069 		"    max_sge         : %d\n"
1070 		"    sq_depth        : %d\n"
1071 		"    max_requests    : %d\n"
1072 		"    ord             : %d\n",
1073 		newxprt,
1074 		&((struct sockaddr_in *)&newxprt->sc_cm_id->
1075 			 route.addr.src_addr)->sin_addr.s_addr,
1076 		ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id->
1077 		       route.addr.src_addr)->sin_port),
1078 		&((struct sockaddr_in *)&newxprt->sc_cm_id->
1079 			 route.addr.dst_addr)->sin_addr.s_addr,
1080 		ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id->
1081 		       route.addr.dst_addr)->sin_port),
1082 		newxprt->sc_max_sge,
1083 		newxprt->sc_sq_depth,
1084 		newxprt->sc_max_requests,
1085 		newxprt->sc_ord);
1086 
1087 	return &newxprt->sc_xprt;
1088 
1089  errout:
1090 	dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret);
1091 	/* Take a reference in case the DTO handler runs */
1092 	svc_xprt_get(&newxprt->sc_xprt);
1093 	if (newxprt->sc_qp && !IS_ERR(newxprt->sc_qp))
1094 		ib_destroy_qp(newxprt->sc_qp);
1095 	rdma_destroy_id(newxprt->sc_cm_id);
1096 	/* This call to put will destroy the transport */
1097 	svc_xprt_put(&newxprt->sc_xprt);
1098 	return NULL;
1099 }
1100 
1101 static void svc_rdma_release_rqst(struct svc_rqst *rqstp)
1102 {
1103 }
1104 
1105 /*
1106  * When connected, an svc_xprt has at least two references:
1107  *
1108  * - A reference held by the cm_id between the ESTABLISHED and
1109  *   DISCONNECTED events. If the remote peer disconnected first, this
1110  *   reference could be gone.
1111  *
1112  * - A reference held by the svc_recv code that called this function
1113  *   as part of close processing.
1114  *
1115  * At a minimum one references should still be held.
1116  */
1117 static void svc_rdma_detach(struct svc_xprt *xprt)
1118 {
1119 	struct svcxprt_rdma *rdma =
1120 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
1121 	dprintk("svc: svc_rdma_detach(%p)\n", xprt);
1122 
1123 	/* Disconnect and flush posted WQE */
1124 	rdma_disconnect(rdma->sc_cm_id);
1125 }
1126 
1127 static void __svc_rdma_free(struct work_struct *work)
1128 {
1129 	struct svcxprt_rdma *rdma =
1130 		container_of(work, struct svcxprt_rdma, sc_work);
1131 	dprintk("svcrdma: svc_rdma_free(%p)\n", rdma);
1132 
1133 	/* We should only be called from kref_put */
1134 	if (atomic_read(&rdma->sc_xprt.xpt_ref.refcount) != 0)
1135 		pr_err("svcrdma: sc_xprt still in use? (%d)\n",
1136 		       atomic_read(&rdma->sc_xprt.xpt_ref.refcount));
1137 
1138 	/*
1139 	 * Destroy queued, but not processed read completions. Note
1140 	 * that this cleanup has to be done before destroying the
1141 	 * cm_id because the device ptr is needed to unmap the dma in
1142 	 * svc_rdma_put_context.
1143 	 */
1144 	while (!list_empty(&rdma->sc_read_complete_q)) {
1145 		struct svc_rdma_op_ctxt *ctxt;
1146 		ctxt = list_entry(rdma->sc_read_complete_q.next,
1147 				  struct svc_rdma_op_ctxt,
1148 				  dto_q);
1149 		list_del_init(&ctxt->dto_q);
1150 		svc_rdma_put_context(ctxt, 1);
1151 	}
1152 
1153 	/* Destroy queued, but not processed recv completions */
1154 	while (!list_empty(&rdma->sc_rq_dto_q)) {
1155 		struct svc_rdma_op_ctxt *ctxt;
1156 		ctxt = list_entry(rdma->sc_rq_dto_q.next,
1157 				  struct svc_rdma_op_ctxt,
1158 				  dto_q);
1159 		list_del_init(&ctxt->dto_q);
1160 		svc_rdma_put_context(ctxt, 1);
1161 	}
1162 
1163 	/* Warn if we leaked a resource or under-referenced */
1164 	if (atomic_read(&rdma->sc_ctxt_used) != 0)
1165 		pr_err("svcrdma: ctxt still in use? (%d)\n",
1166 		       atomic_read(&rdma->sc_ctxt_used));
1167 	if (atomic_read(&rdma->sc_dma_used) != 0)
1168 		pr_err("svcrdma: dma still in use? (%d)\n",
1169 		       atomic_read(&rdma->sc_dma_used));
1170 
1171 	/* De-allocate fastreg mr */
1172 	rdma_dealloc_frmr_q(rdma);
1173 
1174 	/* Destroy the QP if present (not a listener) */
1175 	if (rdma->sc_qp && !IS_ERR(rdma->sc_qp))
1176 		ib_destroy_qp(rdma->sc_qp);
1177 
1178 	if (rdma->sc_sq_cq && !IS_ERR(rdma->sc_sq_cq))
1179 		ib_destroy_cq(rdma->sc_sq_cq);
1180 
1181 	if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq))
1182 		ib_destroy_cq(rdma->sc_rq_cq);
1183 
1184 	if (rdma->sc_phys_mr && !IS_ERR(rdma->sc_phys_mr))
1185 		ib_dereg_mr(rdma->sc_phys_mr);
1186 
1187 	if (rdma->sc_pd && !IS_ERR(rdma->sc_pd))
1188 		ib_dealloc_pd(rdma->sc_pd);
1189 
1190 	/* Destroy the CM ID */
1191 	rdma_destroy_id(rdma->sc_cm_id);
1192 
1193 	kfree(rdma);
1194 }
1195 
1196 static void svc_rdma_free(struct svc_xprt *xprt)
1197 {
1198 	struct svcxprt_rdma *rdma =
1199 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
1200 	INIT_WORK(&rdma->sc_work, __svc_rdma_free);
1201 	queue_work(svc_rdma_wq, &rdma->sc_work);
1202 }
1203 
1204 static int svc_rdma_has_wspace(struct svc_xprt *xprt)
1205 {
1206 	struct svcxprt_rdma *rdma =
1207 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
1208 
1209 	/*
1210 	 * If there are already waiters on the SQ,
1211 	 * return false.
1212 	 */
1213 	if (waitqueue_active(&rdma->sc_send_wait))
1214 		return 0;
1215 
1216 	/* Otherwise return true. */
1217 	return 1;
1218 }
1219 
1220 static int svc_rdma_secure_port(struct svc_rqst *rqstp)
1221 {
1222 	return 1;
1223 }
1224 
1225 /*
1226  * Attempt to register the kvec representing the RPC memory with the
1227  * device.
1228  *
1229  * Returns:
1230  *  NULL : The device does not support fastreg or there were no more
1231  *         fastreg mr.
1232  *  frmr : The kvec register request was successfully posted.
1233  *    <0 : An error was encountered attempting to register the kvec.
1234  */
1235 int svc_rdma_fastreg(struct svcxprt_rdma *xprt,
1236 		     struct svc_rdma_fastreg_mr *frmr)
1237 {
1238 	struct ib_send_wr fastreg_wr;
1239 	u8 key;
1240 
1241 	/* Bump the key */
1242 	key = (u8)(frmr->mr->lkey & 0x000000FF);
1243 	ib_update_fast_reg_key(frmr->mr, ++key);
1244 
1245 	/* Prepare FASTREG WR */
1246 	memset(&fastreg_wr, 0, sizeof fastreg_wr);
1247 	fastreg_wr.opcode = IB_WR_FAST_REG_MR;
1248 	fastreg_wr.send_flags = IB_SEND_SIGNALED;
1249 	fastreg_wr.wr.fast_reg.iova_start = (unsigned long)frmr->kva;
1250 	fastreg_wr.wr.fast_reg.page_list = frmr->page_list;
1251 	fastreg_wr.wr.fast_reg.page_list_len = frmr->page_list_len;
1252 	fastreg_wr.wr.fast_reg.page_shift = PAGE_SHIFT;
1253 	fastreg_wr.wr.fast_reg.length = frmr->map_len;
1254 	fastreg_wr.wr.fast_reg.access_flags = frmr->access_flags;
1255 	fastreg_wr.wr.fast_reg.rkey = frmr->mr->lkey;
1256 	return svc_rdma_send(xprt, &fastreg_wr);
1257 }
1258 
1259 int svc_rdma_send(struct svcxprt_rdma *xprt, struct ib_send_wr *wr)
1260 {
1261 	struct ib_send_wr *bad_wr, *n_wr;
1262 	int wr_count;
1263 	int i;
1264 	int ret;
1265 
1266 	if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1267 		return -ENOTCONN;
1268 
1269 	wr_count = 1;
1270 	for (n_wr = wr->next; n_wr; n_wr = n_wr->next)
1271 		wr_count++;
1272 
1273 	/* If the SQ is full, wait until an SQ entry is available */
1274 	while (1) {
1275 		spin_lock_bh(&xprt->sc_lock);
1276 		if (xprt->sc_sq_depth < atomic_read(&xprt->sc_sq_count) + wr_count) {
1277 			spin_unlock_bh(&xprt->sc_lock);
1278 			atomic_inc(&rdma_stat_sq_starve);
1279 
1280 			/* See if we can opportunistically reap SQ WR to make room */
1281 			sq_cq_reap(xprt);
1282 
1283 			/* Wait until SQ WR available if SQ still full */
1284 			wait_event(xprt->sc_send_wait,
1285 				   atomic_read(&xprt->sc_sq_count) <
1286 				   xprt->sc_sq_depth);
1287 			if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1288 				return -ENOTCONN;
1289 			continue;
1290 		}
1291 		/* Take a transport ref for each WR posted */
1292 		for (i = 0; i < wr_count; i++)
1293 			svc_xprt_get(&xprt->sc_xprt);
1294 
1295 		/* Bump used SQ WR count and post */
1296 		atomic_add(wr_count, &xprt->sc_sq_count);
1297 		ret = ib_post_send(xprt->sc_qp, wr, &bad_wr);
1298 		if (ret) {
1299 			set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
1300 			atomic_sub(wr_count, &xprt->sc_sq_count);
1301 			for (i = 0; i < wr_count; i ++)
1302 				svc_xprt_put(&xprt->sc_xprt);
1303 			dprintk("svcrdma: failed to post SQ WR rc=%d, "
1304 			       "sc_sq_count=%d, sc_sq_depth=%d\n",
1305 			       ret, atomic_read(&xprt->sc_sq_count),
1306 			       xprt->sc_sq_depth);
1307 		}
1308 		spin_unlock_bh(&xprt->sc_lock);
1309 		if (ret)
1310 			wake_up(&xprt->sc_send_wait);
1311 		break;
1312 	}
1313 	return ret;
1314 }
1315 
1316 void svc_rdma_send_error(struct svcxprt_rdma *xprt, struct rpcrdma_msg *rmsgp,
1317 			 enum rpcrdma_errcode err)
1318 {
1319 	struct ib_send_wr err_wr;
1320 	struct page *p;
1321 	struct svc_rdma_op_ctxt *ctxt;
1322 	u32 *va;
1323 	int length;
1324 	int ret;
1325 
1326 	p = svc_rdma_get_page();
1327 	va = page_address(p);
1328 
1329 	/* XDR encode error */
1330 	length = svc_rdma_xdr_encode_error(xprt, rmsgp, err, va);
1331 
1332 	ctxt = svc_rdma_get_context(xprt);
1333 	ctxt->direction = DMA_FROM_DEVICE;
1334 	ctxt->count = 1;
1335 	ctxt->pages[0] = p;
1336 
1337 	/* Prepare SGE for local address */
1338 	ctxt->sge[0].addr = ib_dma_map_page(xprt->sc_cm_id->device,
1339 					    p, 0, length, DMA_FROM_DEVICE);
1340 	if (ib_dma_mapping_error(xprt->sc_cm_id->device, ctxt->sge[0].addr)) {
1341 		put_page(p);
1342 		svc_rdma_put_context(ctxt, 1);
1343 		return;
1344 	}
1345 	atomic_inc(&xprt->sc_dma_used);
1346 	ctxt->sge[0].lkey = xprt->sc_dma_lkey;
1347 	ctxt->sge[0].length = length;
1348 
1349 	/* Prepare SEND WR */
1350 	memset(&err_wr, 0, sizeof err_wr);
1351 	ctxt->wr_op = IB_WR_SEND;
1352 	err_wr.wr_id = (unsigned long)ctxt;
1353 	err_wr.sg_list = ctxt->sge;
1354 	err_wr.num_sge = 1;
1355 	err_wr.opcode = IB_WR_SEND;
1356 	err_wr.send_flags = IB_SEND_SIGNALED;
1357 
1358 	/* Post It */
1359 	ret = svc_rdma_send(xprt, &err_wr);
1360 	if (ret) {
1361 		dprintk("svcrdma: Error %d posting send for protocol error\n",
1362 			ret);
1363 		svc_rdma_unmap_dma(ctxt);
1364 		svc_rdma_put_context(ctxt, 1);
1365 	}
1366 }
1367