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