xref: /openbmc/linux/net/sunrpc/xprtrdma/verbs.c (revision 0a94608f)
1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
2 /*
3  * Copyright (c) 2014-2017 Oracle.  All rights reserved.
4  * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
5  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the BSD-type
10  * license below:
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  *
16  *      Redistributions of source code must retain the above copyright
17  *      notice, this list of conditions and the following disclaimer.
18  *
19  *      Redistributions in binary form must reproduce the above
20  *      copyright notice, this list of conditions and the following
21  *      disclaimer in the documentation and/or other materials provided
22  *      with the distribution.
23  *
24  *      Neither the name of the Network Appliance, Inc. nor the names of
25  *      its contributors may be used to endorse or promote products
26  *      derived from this software without specific prior written
27  *      permission.
28  *
29  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
30  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
31  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
32  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
33  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
34  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
35  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
36  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
37  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
38  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
39  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
40  */
41 
42 /*
43  * verbs.c
44  *
45  * Encapsulates the major functions managing:
46  *  o adapters
47  *  o endpoints
48  *  o connections
49  *  o buffer memory
50  */
51 
52 #include <linux/interrupt.h>
53 #include <linux/slab.h>
54 #include <linux/sunrpc/addr.h>
55 #include <linux/sunrpc/svc_rdma.h>
56 #include <linux/log2.h>
57 
58 #include <asm-generic/barrier.h>
59 #include <asm/bitops.h>
60 
61 #include <rdma/ib_cm.h>
62 
63 #include "xprt_rdma.h"
64 #include <trace/events/rpcrdma.h>
65 
66 static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt);
67 static void rpcrdma_sendctxs_destroy(struct rpcrdma_xprt *r_xprt);
68 static void rpcrdma_sendctx_put_locked(struct rpcrdma_xprt *r_xprt,
69 				       struct rpcrdma_sendctx *sc);
70 static int rpcrdma_reqs_setup(struct rpcrdma_xprt *r_xprt);
71 static void rpcrdma_reqs_reset(struct rpcrdma_xprt *r_xprt);
72 static void rpcrdma_rep_destroy(struct rpcrdma_rep *rep);
73 static void rpcrdma_reps_unmap(struct rpcrdma_xprt *r_xprt);
74 static void rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt);
75 static void rpcrdma_mrs_destroy(struct rpcrdma_xprt *r_xprt);
76 static void rpcrdma_ep_get(struct rpcrdma_ep *ep);
77 static int rpcrdma_ep_put(struct rpcrdma_ep *ep);
78 static struct rpcrdma_regbuf *
79 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction,
80 		     gfp_t flags);
81 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb);
82 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb);
83 
84 /* Wait for outstanding transport work to finish. ib_drain_qp
85  * handles the drains in the wrong order for us, so open code
86  * them here.
87  */
88 static void rpcrdma_xprt_drain(struct rpcrdma_xprt *r_xprt)
89 {
90 	struct rpcrdma_ep *ep = r_xprt->rx_ep;
91 	struct rdma_cm_id *id = ep->re_id;
92 
93 	/* Wait for rpcrdma_post_recvs() to leave its critical
94 	 * section.
95 	 */
96 	if (atomic_inc_return(&ep->re_receiving) > 1)
97 		wait_for_completion(&ep->re_done);
98 
99 	/* Flush Receives, then wait for deferred Reply work
100 	 * to complete.
101 	 */
102 	ib_drain_rq(id->qp);
103 
104 	/* Deferred Reply processing might have scheduled
105 	 * local invalidations.
106 	 */
107 	ib_drain_sq(id->qp);
108 
109 	rpcrdma_ep_put(ep);
110 }
111 
112 /* Ensure xprt_force_disconnect() is invoked exactly once when a
113  * connection is closed or lost. (The important thing is it needs
114  * to be invoked "at least" once).
115  */
116 void rpcrdma_force_disconnect(struct rpcrdma_ep *ep)
117 {
118 	if (atomic_add_unless(&ep->re_force_disconnect, 1, 1))
119 		xprt_force_disconnect(ep->re_xprt);
120 }
121 
122 /**
123  * rpcrdma_flush_disconnect - Disconnect on flushed completion
124  * @r_xprt: transport to disconnect
125  * @wc: work completion entry
126  *
127  * Must be called in process context.
128  */
129 void rpcrdma_flush_disconnect(struct rpcrdma_xprt *r_xprt, struct ib_wc *wc)
130 {
131 	if (wc->status != IB_WC_SUCCESS)
132 		rpcrdma_force_disconnect(r_xprt->rx_ep);
133 }
134 
135 /**
136  * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC
137  * @cq:	completion queue
138  * @wc:	WCE for a completed Send WR
139  *
140  */
141 static void rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
142 {
143 	struct ib_cqe *cqe = wc->wr_cqe;
144 	struct rpcrdma_sendctx *sc =
145 		container_of(cqe, struct rpcrdma_sendctx, sc_cqe);
146 	struct rpcrdma_xprt *r_xprt = cq->cq_context;
147 
148 	/* WARNING: Only wr_cqe and status are reliable at this point */
149 	trace_xprtrdma_wc_send(wc, &sc->sc_cid);
150 	rpcrdma_sendctx_put_locked(r_xprt, sc);
151 	rpcrdma_flush_disconnect(r_xprt, wc);
152 }
153 
154 /**
155  * rpcrdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
156  * @cq:	completion queue
157  * @wc:	WCE for a completed Receive WR
158  *
159  */
160 static void rpcrdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
161 {
162 	struct ib_cqe *cqe = wc->wr_cqe;
163 	struct rpcrdma_rep *rep = container_of(cqe, struct rpcrdma_rep,
164 					       rr_cqe);
165 	struct rpcrdma_xprt *r_xprt = cq->cq_context;
166 
167 	/* WARNING: Only wr_cqe and status are reliable at this point */
168 	trace_xprtrdma_wc_receive(wc, &rep->rr_cid);
169 	--r_xprt->rx_ep->re_receive_count;
170 	if (wc->status != IB_WC_SUCCESS)
171 		goto out_flushed;
172 
173 	/* status == SUCCESS means all fields in wc are trustworthy */
174 	rpcrdma_set_xdrlen(&rep->rr_hdrbuf, wc->byte_len);
175 	rep->rr_wc_flags = wc->wc_flags;
176 	rep->rr_inv_rkey = wc->ex.invalidate_rkey;
177 
178 	ib_dma_sync_single_for_cpu(rdmab_device(rep->rr_rdmabuf),
179 				   rdmab_addr(rep->rr_rdmabuf),
180 				   wc->byte_len, DMA_FROM_DEVICE);
181 
182 	rpcrdma_reply_handler(rep);
183 	return;
184 
185 out_flushed:
186 	rpcrdma_flush_disconnect(r_xprt, wc);
187 	rpcrdma_rep_put(&r_xprt->rx_buf, rep);
188 }
189 
190 static void rpcrdma_update_cm_private(struct rpcrdma_ep *ep,
191 				      struct rdma_conn_param *param)
192 {
193 	const struct rpcrdma_connect_private *pmsg = param->private_data;
194 	unsigned int rsize, wsize;
195 
196 	/* Default settings for RPC-over-RDMA Version One */
197 	rsize = RPCRDMA_V1_DEF_INLINE_SIZE;
198 	wsize = RPCRDMA_V1_DEF_INLINE_SIZE;
199 
200 	if (pmsg &&
201 	    pmsg->cp_magic == rpcrdma_cmp_magic &&
202 	    pmsg->cp_version == RPCRDMA_CMP_VERSION) {
203 		rsize = rpcrdma_decode_buffer_size(pmsg->cp_send_size);
204 		wsize = rpcrdma_decode_buffer_size(pmsg->cp_recv_size);
205 	}
206 
207 	if (rsize < ep->re_inline_recv)
208 		ep->re_inline_recv = rsize;
209 	if (wsize < ep->re_inline_send)
210 		ep->re_inline_send = wsize;
211 
212 	rpcrdma_set_max_header_sizes(ep);
213 }
214 
215 /**
216  * rpcrdma_cm_event_handler - Handle RDMA CM events
217  * @id: rdma_cm_id on which an event has occurred
218  * @event: details of the event
219  *
220  * Called with @id's mutex held. Returns 1 if caller should
221  * destroy @id, otherwise 0.
222  */
223 static int
224 rpcrdma_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
225 {
226 	struct sockaddr *sap = (struct sockaddr *)&id->route.addr.dst_addr;
227 	struct rpcrdma_ep *ep = id->context;
228 
229 	might_sleep();
230 
231 	switch (event->event) {
232 	case RDMA_CM_EVENT_ADDR_RESOLVED:
233 	case RDMA_CM_EVENT_ROUTE_RESOLVED:
234 		ep->re_async_rc = 0;
235 		complete(&ep->re_done);
236 		return 0;
237 	case RDMA_CM_EVENT_ADDR_ERROR:
238 		ep->re_async_rc = -EPROTO;
239 		complete(&ep->re_done);
240 		return 0;
241 	case RDMA_CM_EVENT_ROUTE_ERROR:
242 		ep->re_async_rc = -ENETUNREACH;
243 		complete(&ep->re_done);
244 		return 0;
245 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
246 		pr_info("rpcrdma: removing device %s for %pISpc\n",
247 			ep->re_id->device->name, sap);
248 		fallthrough;
249 	case RDMA_CM_EVENT_ADDR_CHANGE:
250 		ep->re_connect_status = -ENODEV;
251 		goto disconnected;
252 	case RDMA_CM_EVENT_ESTABLISHED:
253 		rpcrdma_ep_get(ep);
254 		ep->re_connect_status = 1;
255 		rpcrdma_update_cm_private(ep, &event->param.conn);
256 		trace_xprtrdma_inline_thresh(ep);
257 		wake_up_all(&ep->re_connect_wait);
258 		break;
259 	case RDMA_CM_EVENT_CONNECT_ERROR:
260 		ep->re_connect_status = -ENOTCONN;
261 		goto wake_connect_worker;
262 	case RDMA_CM_EVENT_UNREACHABLE:
263 		ep->re_connect_status = -ENETUNREACH;
264 		goto wake_connect_worker;
265 	case RDMA_CM_EVENT_REJECTED:
266 		ep->re_connect_status = -ECONNREFUSED;
267 		if (event->status == IB_CM_REJ_STALE_CONN)
268 			ep->re_connect_status = -ENOTCONN;
269 wake_connect_worker:
270 		wake_up_all(&ep->re_connect_wait);
271 		return 0;
272 	case RDMA_CM_EVENT_DISCONNECTED:
273 		ep->re_connect_status = -ECONNABORTED;
274 disconnected:
275 		rpcrdma_force_disconnect(ep);
276 		return rpcrdma_ep_put(ep);
277 	default:
278 		break;
279 	}
280 
281 	return 0;
282 }
283 
284 static struct rdma_cm_id *rpcrdma_create_id(struct rpcrdma_xprt *r_xprt,
285 					    struct rpcrdma_ep *ep)
286 {
287 	unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1;
288 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
289 	struct rdma_cm_id *id;
290 	int rc;
291 
292 	init_completion(&ep->re_done);
293 
294 	id = rdma_create_id(xprt->xprt_net, rpcrdma_cm_event_handler, ep,
295 			    RDMA_PS_TCP, IB_QPT_RC);
296 	if (IS_ERR(id))
297 		return id;
298 
299 	ep->re_async_rc = -ETIMEDOUT;
300 	rc = rdma_resolve_addr(id, NULL, (struct sockaddr *)&xprt->addr,
301 			       RDMA_RESOLVE_TIMEOUT);
302 	if (rc)
303 		goto out;
304 	rc = wait_for_completion_interruptible_timeout(&ep->re_done, wtimeout);
305 	if (rc < 0)
306 		goto out;
307 
308 	rc = ep->re_async_rc;
309 	if (rc)
310 		goto out;
311 
312 	ep->re_async_rc = -ETIMEDOUT;
313 	rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
314 	if (rc)
315 		goto out;
316 	rc = wait_for_completion_interruptible_timeout(&ep->re_done, wtimeout);
317 	if (rc < 0)
318 		goto out;
319 	rc = ep->re_async_rc;
320 	if (rc)
321 		goto out;
322 
323 	return id;
324 
325 out:
326 	rdma_destroy_id(id);
327 	return ERR_PTR(rc);
328 }
329 
330 static void rpcrdma_ep_destroy(struct kref *kref)
331 {
332 	struct rpcrdma_ep *ep = container_of(kref, struct rpcrdma_ep, re_kref);
333 
334 	if (ep->re_id->qp) {
335 		rdma_destroy_qp(ep->re_id);
336 		ep->re_id->qp = NULL;
337 	}
338 
339 	if (ep->re_attr.recv_cq)
340 		ib_free_cq(ep->re_attr.recv_cq);
341 	ep->re_attr.recv_cq = NULL;
342 	if (ep->re_attr.send_cq)
343 		ib_free_cq(ep->re_attr.send_cq);
344 	ep->re_attr.send_cq = NULL;
345 
346 	if (ep->re_pd)
347 		ib_dealloc_pd(ep->re_pd);
348 	ep->re_pd = NULL;
349 
350 	kfree(ep);
351 	module_put(THIS_MODULE);
352 }
353 
354 static noinline void rpcrdma_ep_get(struct rpcrdma_ep *ep)
355 {
356 	kref_get(&ep->re_kref);
357 }
358 
359 /* Returns:
360  *     %0 if @ep still has a positive kref count, or
361  *     %1 if @ep was destroyed successfully.
362  */
363 static noinline int rpcrdma_ep_put(struct rpcrdma_ep *ep)
364 {
365 	return kref_put(&ep->re_kref, rpcrdma_ep_destroy);
366 }
367 
368 static int rpcrdma_ep_create(struct rpcrdma_xprt *r_xprt)
369 {
370 	struct rpcrdma_connect_private *pmsg;
371 	struct ib_device *device;
372 	struct rdma_cm_id *id;
373 	struct rpcrdma_ep *ep;
374 	int rc;
375 
376 	ep = kzalloc(sizeof(*ep), GFP_KERNEL);
377 	if (!ep)
378 		return -ENOTCONN;
379 	ep->re_xprt = &r_xprt->rx_xprt;
380 	kref_init(&ep->re_kref);
381 
382 	id = rpcrdma_create_id(r_xprt, ep);
383 	if (IS_ERR(id)) {
384 		kfree(ep);
385 		return PTR_ERR(id);
386 	}
387 	__module_get(THIS_MODULE);
388 	device = id->device;
389 	ep->re_id = id;
390 	reinit_completion(&ep->re_done);
391 
392 	ep->re_max_requests = r_xprt->rx_xprt.max_reqs;
393 	ep->re_inline_send = xprt_rdma_max_inline_write;
394 	ep->re_inline_recv = xprt_rdma_max_inline_read;
395 	rc = frwr_query_device(ep, device);
396 	if (rc)
397 		goto out_destroy;
398 
399 	r_xprt->rx_buf.rb_max_requests = cpu_to_be32(ep->re_max_requests);
400 
401 	ep->re_attr.srq = NULL;
402 	ep->re_attr.cap.max_inline_data = 0;
403 	ep->re_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
404 	ep->re_attr.qp_type = IB_QPT_RC;
405 	ep->re_attr.port_num = ~0;
406 
407 	ep->re_send_batch = ep->re_max_requests >> 3;
408 	ep->re_send_count = ep->re_send_batch;
409 	init_waitqueue_head(&ep->re_connect_wait);
410 
411 	ep->re_attr.send_cq = ib_alloc_cq_any(device, r_xprt,
412 					      ep->re_attr.cap.max_send_wr,
413 					      IB_POLL_WORKQUEUE);
414 	if (IS_ERR(ep->re_attr.send_cq)) {
415 		rc = PTR_ERR(ep->re_attr.send_cq);
416 		ep->re_attr.send_cq = NULL;
417 		goto out_destroy;
418 	}
419 
420 	ep->re_attr.recv_cq = ib_alloc_cq_any(device, r_xprt,
421 					      ep->re_attr.cap.max_recv_wr,
422 					      IB_POLL_WORKQUEUE);
423 	if (IS_ERR(ep->re_attr.recv_cq)) {
424 		rc = PTR_ERR(ep->re_attr.recv_cq);
425 		ep->re_attr.recv_cq = NULL;
426 		goto out_destroy;
427 	}
428 	ep->re_receive_count = 0;
429 
430 	/* Initialize cma parameters */
431 	memset(&ep->re_remote_cma, 0, sizeof(ep->re_remote_cma));
432 
433 	/* Prepare RDMA-CM private message */
434 	pmsg = &ep->re_cm_private;
435 	pmsg->cp_magic = rpcrdma_cmp_magic;
436 	pmsg->cp_version = RPCRDMA_CMP_VERSION;
437 	pmsg->cp_flags |= RPCRDMA_CMP_F_SND_W_INV_OK;
438 	pmsg->cp_send_size = rpcrdma_encode_buffer_size(ep->re_inline_send);
439 	pmsg->cp_recv_size = rpcrdma_encode_buffer_size(ep->re_inline_recv);
440 	ep->re_remote_cma.private_data = pmsg;
441 	ep->re_remote_cma.private_data_len = sizeof(*pmsg);
442 
443 	/* Client offers RDMA Read but does not initiate */
444 	ep->re_remote_cma.initiator_depth = 0;
445 	ep->re_remote_cma.responder_resources =
446 		min_t(int, U8_MAX, device->attrs.max_qp_rd_atom);
447 
448 	/* Limit transport retries so client can detect server
449 	 * GID changes quickly. RPC layer handles re-establishing
450 	 * transport connection and retransmission.
451 	 */
452 	ep->re_remote_cma.retry_count = 6;
453 
454 	/* RPC-over-RDMA handles its own flow control. In addition,
455 	 * make all RNR NAKs visible so we know that RPC-over-RDMA
456 	 * flow control is working correctly (no NAKs should be seen).
457 	 */
458 	ep->re_remote_cma.flow_control = 0;
459 	ep->re_remote_cma.rnr_retry_count = 0;
460 
461 	ep->re_pd = ib_alloc_pd(device, 0);
462 	if (IS_ERR(ep->re_pd)) {
463 		rc = PTR_ERR(ep->re_pd);
464 		ep->re_pd = NULL;
465 		goto out_destroy;
466 	}
467 
468 	rc = rdma_create_qp(id, ep->re_pd, &ep->re_attr);
469 	if (rc)
470 		goto out_destroy;
471 
472 	r_xprt->rx_ep = ep;
473 	return 0;
474 
475 out_destroy:
476 	rpcrdma_ep_put(ep);
477 	rdma_destroy_id(id);
478 	return rc;
479 }
480 
481 /**
482  * rpcrdma_xprt_connect - Connect an unconnected transport
483  * @r_xprt: controlling transport instance
484  *
485  * Returns 0 on success or a negative errno.
486  */
487 int rpcrdma_xprt_connect(struct rpcrdma_xprt *r_xprt)
488 {
489 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
490 	struct rpcrdma_ep *ep;
491 	int rc;
492 
493 	rc = rpcrdma_ep_create(r_xprt);
494 	if (rc)
495 		return rc;
496 	ep = r_xprt->rx_ep;
497 
498 	xprt_clear_connected(xprt);
499 	rpcrdma_reset_cwnd(r_xprt);
500 
501 	/* Bump the ep's reference count while there are
502 	 * outstanding Receives.
503 	 */
504 	rpcrdma_ep_get(ep);
505 	rpcrdma_post_recvs(r_xprt, 1, true);
506 
507 	rc = rdma_connect(ep->re_id, &ep->re_remote_cma);
508 	if (rc)
509 		goto out;
510 
511 	if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
512 		xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
513 	wait_event_interruptible(ep->re_connect_wait,
514 				 ep->re_connect_status != 0);
515 	if (ep->re_connect_status <= 0) {
516 		rc = ep->re_connect_status;
517 		goto out;
518 	}
519 
520 	rc = rpcrdma_sendctxs_create(r_xprt);
521 	if (rc) {
522 		rc = -ENOTCONN;
523 		goto out;
524 	}
525 
526 	rc = rpcrdma_reqs_setup(r_xprt);
527 	if (rc) {
528 		rc = -ENOTCONN;
529 		goto out;
530 	}
531 	rpcrdma_mrs_create(r_xprt);
532 	frwr_wp_create(r_xprt);
533 
534 out:
535 	trace_xprtrdma_connect(r_xprt, rc);
536 	return rc;
537 }
538 
539 /**
540  * rpcrdma_xprt_disconnect - Disconnect underlying transport
541  * @r_xprt: controlling transport instance
542  *
543  * Caller serializes. Either the transport send lock is held,
544  * or we're being called to destroy the transport.
545  *
546  * On return, @r_xprt is completely divested of all hardware
547  * resources and prepared for the next ->connect operation.
548  */
549 void rpcrdma_xprt_disconnect(struct rpcrdma_xprt *r_xprt)
550 {
551 	struct rpcrdma_ep *ep = r_xprt->rx_ep;
552 	struct rdma_cm_id *id;
553 	int rc;
554 
555 	if (!ep)
556 		return;
557 
558 	id = ep->re_id;
559 	rc = rdma_disconnect(id);
560 	trace_xprtrdma_disconnect(r_xprt, rc);
561 
562 	rpcrdma_xprt_drain(r_xprt);
563 	rpcrdma_reps_unmap(r_xprt);
564 	rpcrdma_reqs_reset(r_xprt);
565 	rpcrdma_mrs_destroy(r_xprt);
566 	rpcrdma_sendctxs_destroy(r_xprt);
567 
568 	if (rpcrdma_ep_put(ep))
569 		rdma_destroy_id(id);
570 
571 	r_xprt->rx_ep = NULL;
572 }
573 
574 /* Fixed-size circular FIFO queue. This implementation is wait-free and
575  * lock-free.
576  *
577  * Consumer is the code path that posts Sends. This path dequeues a
578  * sendctx for use by a Send operation. Multiple consumer threads
579  * are serialized by the RPC transport lock, which allows only one
580  * ->send_request call at a time.
581  *
582  * Producer is the code path that handles Send completions. This path
583  * enqueues a sendctx that has been completed. Multiple producer
584  * threads are serialized by the ib_poll_cq() function.
585  */
586 
587 /* rpcrdma_sendctxs_destroy() assumes caller has already quiesced
588  * queue activity, and rpcrdma_xprt_drain has flushed all remaining
589  * Send requests.
590  */
591 static void rpcrdma_sendctxs_destroy(struct rpcrdma_xprt *r_xprt)
592 {
593 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
594 	unsigned long i;
595 
596 	if (!buf->rb_sc_ctxs)
597 		return;
598 	for (i = 0; i <= buf->rb_sc_last; i++)
599 		kfree(buf->rb_sc_ctxs[i]);
600 	kfree(buf->rb_sc_ctxs);
601 	buf->rb_sc_ctxs = NULL;
602 }
603 
604 static struct rpcrdma_sendctx *rpcrdma_sendctx_create(struct rpcrdma_ep *ep)
605 {
606 	struct rpcrdma_sendctx *sc;
607 
608 	sc = kzalloc(struct_size(sc, sc_sges, ep->re_attr.cap.max_send_sge),
609 		     GFP_KERNEL);
610 	if (!sc)
611 		return NULL;
612 
613 	sc->sc_cqe.done = rpcrdma_wc_send;
614 	sc->sc_cid.ci_queue_id = ep->re_attr.send_cq->res.id;
615 	sc->sc_cid.ci_completion_id =
616 		atomic_inc_return(&ep->re_completion_ids);
617 	return sc;
618 }
619 
620 static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt)
621 {
622 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
623 	struct rpcrdma_sendctx *sc;
624 	unsigned long i;
625 
626 	/* Maximum number of concurrent outstanding Send WRs. Capping
627 	 * the circular queue size stops Send Queue overflow by causing
628 	 * the ->send_request call to fail temporarily before too many
629 	 * Sends are posted.
630 	 */
631 	i = r_xprt->rx_ep->re_max_requests + RPCRDMA_MAX_BC_REQUESTS;
632 	buf->rb_sc_ctxs = kcalloc(i, sizeof(sc), GFP_KERNEL);
633 	if (!buf->rb_sc_ctxs)
634 		return -ENOMEM;
635 
636 	buf->rb_sc_last = i - 1;
637 	for (i = 0; i <= buf->rb_sc_last; i++) {
638 		sc = rpcrdma_sendctx_create(r_xprt->rx_ep);
639 		if (!sc)
640 			return -ENOMEM;
641 
642 		buf->rb_sc_ctxs[i] = sc;
643 	}
644 
645 	buf->rb_sc_head = 0;
646 	buf->rb_sc_tail = 0;
647 	return 0;
648 }
649 
650 /* The sendctx queue is not guaranteed to have a size that is a
651  * power of two, thus the helpers in circ_buf.h cannot be used.
652  * The other option is to use modulus (%), which can be expensive.
653  */
654 static unsigned long rpcrdma_sendctx_next(struct rpcrdma_buffer *buf,
655 					  unsigned long item)
656 {
657 	return likely(item < buf->rb_sc_last) ? item + 1 : 0;
658 }
659 
660 /**
661  * rpcrdma_sendctx_get_locked - Acquire a send context
662  * @r_xprt: controlling transport instance
663  *
664  * Returns pointer to a free send completion context; or NULL if
665  * the queue is empty.
666  *
667  * Usage: Called to acquire an SGE array before preparing a Send WR.
668  *
669  * The caller serializes calls to this function (per transport), and
670  * provides an effective memory barrier that flushes the new value
671  * of rb_sc_head.
672  */
673 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_xprt *r_xprt)
674 {
675 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
676 	struct rpcrdma_sendctx *sc;
677 	unsigned long next_head;
678 
679 	next_head = rpcrdma_sendctx_next(buf, buf->rb_sc_head);
680 
681 	if (next_head == READ_ONCE(buf->rb_sc_tail))
682 		goto out_emptyq;
683 
684 	/* ORDER: item must be accessed _before_ head is updated */
685 	sc = buf->rb_sc_ctxs[next_head];
686 
687 	/* Releasing the lock in the caller acts as a memory
688 	 * barrier that flushes rb_sc_head.
689 	 */
690 	buf->rb_sc_head = next_head;
691 
692 	return sc;
693 
694 out_emptyq:
695 	/* The queue is "empty" if there have not been enough Send
696 	 * completions recently. This is a sign the Send Queue is
697 	 * backing up. Cause the caller to pause and try again.
698 	 */
699 	xprt_wait_for_buffer_space(&r_xprt->rx_xprt);
700 	r_xprt->rx_stats.empty_sendctx_q++;
701 	return NULL;
702 }
703 
704 /**
705  * rpcrdma_sendctx_put_locked - Release a send context
706  * @r_xprt: controlling transport instance
707  * @sc: send context to release
708  *
709  * Usage: Called from Send completion to return a sendctxt
710  * to the queue.
711  *
712  * The caller serializes calls to this function (per transport).
713  */
714 static void rpcrdma_sendctx_put_locked(struct rpcrdma_xprt *r_xprt,
715 				       struct rpcrdma_sendctx *sc)
716 {
717 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
718 	unsigned long next_tail;
719 
720 	/* Unmap SGEs of previously completed but unsignaled
721 	 * Sends by walking up the queue until @sc is found.
722 	 */
723 	next_tail = buf->rb_sc_tail;
724 	do {
725 		next_tail = rpcrdma_sendctx_next(buf, next_tail);
726 
727 		/* ORDER: item must be accessed _before_ tail is updated */
728 		rpcrdma_sendctx_unmap(buf->rb_sc_ctxs[next_tail]);
729 
730 	} while (buf->rb_sc_ctxs[next_tail] != sc);
731 
732 	/* Paired with READ_ONCE */
733 	smp_store_release(&buf->rb_sc_tail, next_tail);
734 
735 	xprt_write_space(&r_xprt->rx_xprt);
736 }
737 
738 static void
739 rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt)
740 {
741 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
742 	struct rpcrdma_ep *ep = r_xprt->rx_ep;
743 	unsigned int count;
744 
745 	for (count = 0; count < ep->re_max_rdma_segs; count++) {
746 		struct rpcrdma_mr *mr;
747 		int rc;
748 
749 		mr = kzalloc(sizeof(*mr), GFP_KERNEL);
750 		if (!mr)
751 			break;
752 
753 		rc = frwr_mr_init(r_xprt, mr);
754 		if (rc) {
755 			kfree(mr);
756 			break;
757 		}
758 
759 		spin_lock(&buf->rb_lock);
760 		rpcrdma_mr_push(mr, &buf->rb_mrs);
761 		list_add(&mr->mr_all, &buf->rb_all_mrs);
762 		spin_unlock(&buf->rb_lock);
763 	}
764 
765 	r_xprt->rx_stats.mrs_allocated += count;
766 	trace_xprtrdma_createmrs(r_xprt, count);
767 }
768 
769 static void
770 rpcrdma_mr_refresh_worker(struct work_struct *work)
771 {
772 	struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
773 						  rb_refresh_worker);
774 	struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
775 						   rx_buf);
776 
777 	rpcrdma_mrs_create(r_xprt);
778 	xprt_write_space(&r_xprt->rx_xprt);
779 }
780 
781 /**
782  * rpcrdma_mrs_refresh - Wake the MR refresh worker
783  * @r_xprt: controlling transport instance
784  *
785  */
786 void rpcrdma_mrs_refresh(struct rpcrdma_xprt *r_xprt)
787 {
788 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
789 	struct rpcrdma_ep *ep = r_xprt->rx_ep;
790 
791 	/* If there is no underlying connection, it's no use
792 	 * to wake the refresh worker.
793 	 */
794 	if (ep->re_connect_status == 1) {
795 		/* The work is scheduled on a WQ_MEM_RECLAIM
796 		 * workqueue in order to prevent MR allocation
797 		 * from recursing into NFS during direct reclaim.
798 		 */
799 		queue_work(xprtiod_workqueue, &buf->rb_refresh_worker);
800 	}
801 }
802 
803 /**
804  * rpcrdma_req_create - Allocate an rpcrdma_req object
805  * @r_xprt: controlling r_xprt
806  * @size: initial size, in bytes, of send and receive buffers
807  * @flags: GFP flags passed to memory allocators
808  *
809  * Returns an allocated and fully initialized rpcrdma_req or NULL.
810  */
811 struct rpcrdma_req *rpcrdma_req_create(struct rpcrdma_xprt *r_xprt, size_t size,
812 				       gfp_t flags)
813 {
814 	struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
815 	struct rpcrdma_req *req;
816 
817 	req = kzalloc(sizeof(*req), flags);
818 	if (req == NULL)
819 		goto out1;
820 
821 	req->rl_sendbuf = rpcrdma_regbuf_alloc(size, DMA_TO_DEVICE, flags);
822 	if (!req->rl_sendbuf)
823 		goto out2;
824 
825 	req->rl_recvbuf = rpcrdma_regbuf_alloc(size, DMA_NONE, flags);
826 	if (!req->rl_recvbuf)
827 		goto out3;
828 
829 	INIT_LIST_HEAD(&req->rl_free_mrs);
830 	INIT_LIST_HEAD(&req->rl_registered);
831 	spin_lock(&buffer->rb_lock);
832 	list_add(&req->rl_all, &buffer->rb_allreqs);
833 	spin_unlock(&buffer->rb_lock);
834 	return req;
835 
836 out3:
837 	kfree(req->rl_sendbuf);
838 out2:
839 	kfree(req);
840 out1:
841 	return NULL;
842 }
843 
844 /**
845  * rpcrdma_req_setup - Per-connection instance setup of an rpcrdma_req object
846  * @r_xprt: controlling transport instance
847  * @req: rpcrdma_req object to set up
848  *
849  * Returns zero on success, and a negative errno on failure.
850  */
851 int rpcrdma_req_setup(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
852 {
853 	struct rpcrdma_regbuf *rb;
854 	size_t maxhdrsize;
855 
856 	/* Compute maximum header buffer size in bytes */
857 	maxhdrsize = rpcrdma_fixed_maxsz + 3 +
858 		     r_xprt->rx_ep->re_max_rdma_segs * rpcrdma_readchunk_maxsz;
859 	maxhdrsize *= sizeof(__be32);
860 	rb = rpcrdma_regbuf_alloc(__roundup_pow_of_two(maxhdrsize),
861 				  DMA_TO_DEVICE, GFP_KERNEL);
862 	if (!rb)
863 		goto out;
864 
865 	if (!__rpcrdma_regbuf_dma_map(r_xprt, rb))
866 		goto out_free;
867 
868 	req->rl_rdmabuf = rb;
869 	xdr_buf_init(&req->rl_hdrbuf, rdmab_data(rb), rdmab_length(rb));
870 	return 0;
871 
872 out_free:
873 	rpcrdma_regbuf_free(rb);
874 out:
875 	return -ENOMEM;
876 }
877 
878 /* ASSUMPTION: the rb_allreqs list is stable for the duration,
879  * and thus can be walked without holding rb_lock. Eg. the
880  * caller is holding the transport send lock to exclude
881  * device removal or disconnection.
882  */
883 static int rpcrdma_reqs_setup(struct rpcrdma_xprt *r_xprt)
884 {
885 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
886 	struct rpcrdma_req *req;
887 	int rc;
888 
889 	list_for_each_entry(req, &buf->rb_allreqs, rl_all) {
890 		rc = rpcrdma_req_setup(r_xprt, req);
891 		if (rc)
892 			return rc;
893 	}
894 	return 0;
895 }
896 
897 static void rpcrdma_req_reset(struct rpcrdma_req *req)
898 {
899 	/* Credits are valid for only one connection */
900 	req->rl_slot.rq_cong = 0;
901 
902 	rpcrdma_regbuf_free(req->rl_rdmabuf);
903 	req->rl_rdmabuf = NULL;
904 
905 	rpcrdma_regbuf_dma_unmap(req->rl_sendbuf);
906 	rpcrdma_regbuf_dma_unmap(req->rl_recvbuf);
907 
908 	frwr_reset(req);
909 }
910 
911 /* ASSUMPTION: the rb_allreqs list is stable for the duration,
912  * and thus can be walked without holding rb_lock. Eg. the
913  * caller is holding the transport send lock to exclude
914  * device removal or disconnection.
915  */
916 static void rpcrdma_reqs_reset(struct rpcrdma_xprt *r_xprt)
917 {
918 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
919 	struct rpcrdma_req *req;
920 
921 	list_for_each_entry(req, &buf->rb_allreqs, rl_all)
922 		rpcrdma_req_reset(req);
923 }
924 
925 static noinline
926 struct rpcrdma_rep *rpcrdma_rep_create(struct rpcrdma_xprt *r_xprt,
927 				       bool temp)
928 {
929 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
930 	struct rpcrdma_rep *rep;
931 
932 	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
933 	if (rep == NULL)
934 		goto out;
935 
936 	rep->rr_rdmabuf = rpcrdma_regbuf_alloc(r_xprt->rx_ep->re_inline_recv,
937 					       DMA_FROM_DEVICE, GFP_KERNEL);
938 	if (!rep->rr_rdmabuf)
939 		goto out_free;
940 
941 	if (!rpcrdma_regbuf_dma_map(r_xprt, rep->rr_rdmabuf))
942 		goto out_free_regbuf;
943 
944 	rep->rr_cid.ci_completion_id =
945 		atomic_inc_return(&r_xprt->rx_ep->re_completion_ids);
946 
947 	xdr_buf_init(&rep->rr_hdrbuf, rdmab_data(rep->rr_rdmabuf),
948 		     rdmab_length(rep->rr_rdmabuf));
949 	rep->rr_cqe.done = rpcrdma_wc_receive;
950 	rep->rr_rxprt = r_xprt;
951 	rep->rr_recv_wr.next = NULL;
952 	rep->rr_recv_wr.wr_cqe = &rep->rr_cqe;
953 	rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
954 	rep->rr_recv_wr.num_sge = 1;
955 	rep->rr_temp = temp;
956 
957 	spin_lock(&buf->rb_lock);
958 	list_add(&rep->rr_all, &buf->rb_all_reps);
959 	spin_unlock(&buf->rb_lock);
960 	return rep;
961 
962 out_free_regbuf:
963 	rpcrdma_regbuf_free(rep->rr_rdmabuf);
964 out_free:
965 	kfree(rep);
966 out:
967 	return NULL;
968 }
969 
970 static void rpcrdma_rep_free(struct rpcrdma_rep *rep)
971 {
972 	rpcrdma_regbuf_free(rep->rr_rdmabuf);
973 	kfree(rep);
974 }
975 
976 static void rpcrdma_rep_destroy(struct rpcrdma_rep *rep)
977 {
978 	struct rpcrdma_buffer *buf = &rep->rr_rxprt->rx_buf;
979 
980 	spin_lock(&buf->rb_lock);
981 	list_del(&rep->rr_all);
982 	spin_unlock(&buf->rb_lock);
983 
984 	rpcrdma_rep_free(rep);
985 }
986 
987 static struct rpcrdma_rep *rpcrdma_rep_get_locked(struct rpcrdma_buffer *buf)
988 {
989 	struct llist_node *node;
990 
991 	/* Calls to llist_del_first are required to be serialized */
992 	node = llist_del_first(&buf->rb_free_reps);
993 	if (!node)
994 		return NULL;
995 	return llist_entry(node, struct rpcrdma_rep, rr_node);
996 }
997 
998 /**
999  * rpcrdma_rep_put - Release rpcrdma_rep back to free list
1000  * @buf: buffer pool
1001  * @rep: rep to release
1002  *
1003  */
1004 void rpcrdma_rep_put(struct rpcrdma_buffer *buf, struct rpcrdma_rep *rep)
1005 {
1006 	llist_add(&rep->rr_node, &buf->rb_free_reps);
1007 }
1008 
1009 /* Caller must ensure the QP is quiescent (RQ is drained) before
1010  * invoking this function, to guarantee rb_all_reps is not
1011  * changing.
1012  */
1013 static void rpcrdma_reps_unmap(struct rpcrdma_xprt *r_xprt)
1014 {
1015 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1016 	struct rpcrdma_rep *rep;
1017 
1018 	list_for_each_entry(rep, &buf->rb_all_reps, rr_all) {
1019 		rpcrdma_regbuf_dma_unmap(rep->rr_rdmabuf);
1020 		rep->rr_temp = true;	/* Mark this rep for destruction */
1021 	}
1022 }
1023 
1024 static void rpcrdma_reps_destroy(struct rpcrdma_buffer *buf)
1025 {
1026 	struct rpcrdma_rep *rep;
1027 
1028 	spin_lock(&buf->rb_lock);
1029 	while ((rep = list_first_entry_or_null(&buf->rb_all_reps,
1030 					       struct rpcrdma_rep,
1031 					       rr_all)) != NULL) {
1032 		list_del(&rep->rr_all);
1033 		spin_unlock(&buf->rb_lock);
1034 
1035 		rpcrdma_rep_free(rep);
1036 
1037 		spin_lock(&buf->rb_lock);
1038 	}
1039 	spin_unlock(&buf->rb_lock);
1040 }
1041 
1042 /**
1043  * rpcrdma_buffer_create - Create initial set of req/rep objects
1044  * @r_xprt: transport instance to (re)initialize
1045  *
1046  * Returns zero on success, otherwise a negative errno.
1047  */
1048 int rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
1049 {
1050 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1051 	int i, rc;
1052 
1053 	buf->rb_bc_srv_max_requests = 0;
1054 	spin_lock_init(&buf->rb_lock);
1055 	INIT_LIST_HEAD(&buf->rb_mrs);
1056 	INIT_LIST_HEAD(&buf->rb_all_mrs);
1057 	INIT_WORK(&buf->rb_refresh_worker, rpcrdma_mr_refresh_worker);
1058 
1059 	INIT_LIST_HEAD(&buf->rb_send_bufs);
1060 	INIT_LIST_HEAD(&buf->rb_allreqs);
1061 	INIT_LIST_HEAD(&buf->rb_all_reps);
1062 
1063 	rc = -ENOMEM;
1064 	for (i = 0; i < r_xprt->rx_xprt.max_reqs; i++) {
1065 		struct rpcrdma_req *req;
1066 
1067 		req = rpcrdma_req_create(r_xprt, RPCRDMA_V1_DEF_INLINE_SIZE * 2,
1068 					 GFP_KERNEL);
1069 		if (!req)
1070 			goto out;
1071 		list_add(&req->rl_list, &buf->rb_send_bufs);
1072 	}
1073 
1074 	init_llist_head(&buf->rb_free_reps);
1075 
1076 	return 0;
1077 out:
1078 	rpcrdma_buffer_destroy(buf);
1079 	return rc;
1080 }
1081 
1082 /**
1083  * rpcrdma_req_destroy - Destroy an rpcrdma_req object
1084  * @req: unused object to be destroyed
1085  *
1086  * Relies on caller holding the transport send lock to protect
1087  * removing req->rl_all from buf->rb_all_reqs safely.
1088  */
1089 void rpcrdma_req_destroy(struct rpcrdma_req *req)
1090 {
1091 	struct rpcrdma_mr *mr;
1092 
1093 	list_del(&req->rl_all);
1094 
1095 	while ((mr = rpcrdma_mr_pop(&req->rl_free_mrs))) {
1096 		struct rpcrdma_buffer *buf = &mr->mr_xprt->rx_buf;
1097 
1098 		spin_lock(&buf->rb_lock);
1099 		list_del(&mr->mr_all);
1100 		spin_unlock(&buf->rb_lock);
1101 
1102 		frwr_mr_release(mr);
1103 	}
1104 
1105 	rpcrdma_regbuf_free(req->rl_recvbuf);
1106 	rpcrdma_regbuf_free(req->rl_sendbuf);
1107 	rpcrdma_regbuf_free(req->rl_rdmabuf);
1108 	kfree(req);
1109 }
1110 
1111 /**
1112  * rpcrdma_mrs_destroy - Release all of a transport's MRs
1113  * @r_xprt: controlling transport instance
1114  *
1115  * Relies on caller holding the transport send lock to protect
1116  * removing mr->mr_list from req->rl_free_mrs safely.
1117  */
1118 static void rpcrdma_mrs_destroy(struct rpcrdma_xprt *r_xprt)
1119 {
1120 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1121 	struct rpcrdma_mr *mr;
1122 
1123 	cancel_work_sync(&buf->rb_refresh_worker);
1124 
1125 	spin_lock(&buf->rb_lock);
1126 	while ((mr = list_first_entry_or_null(&buf->rb_all_mrs,
1127 					      struct rpcrdma_mr,
1128 					      mr_all)) != NULL) {
1129 		list_del(&mr->mr_list);
1130 		list_del(&mr->mr_all);
1131 		spin_unlock(&buf->rb_lock);
1132 
1133 		frwr_mr_release(mr);
1134 
1135 		spin_lock(&buf->rb_lock);
1136 	}
1137 	spin_unlock(&buf->rb_lock);
1138 }
1139 
1140 /**
1141  * rpcrdma_buffer_destroy - Release all hw resources
1142  * @buf: root control block for resources
1143  *
1144  * ORDERING: relies on a prior rpcrdma_xprt_drain :
1145  * - No more Send or Receive completions can occur
1146  * - All MRs, reps, and reqs are returned to their free lists
1147  */
1148 void
1149 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1150 {
1151 	rpcrdma_reps_destroy(buf);
1152 
1153 	while (!list_empty(&buf->rb_send_bufs)) {
1154 		struct rpcrdma_req *req;
1155 
1156 		req = list_first_entry(&buf->rb_send_bufs,
1157 				       struct rpcrdma_req, rl_list);
1158 		list_del(&req->rl_list);
1159 		rpcrdma_req_destroy(req);
1160 	}
1161 }
1162 
1163 /**
1164  * rpcrdma_mr_get - Allocate an rpcrdma_mr object
1165  * @r_xprt: controlling transport
1166  *
1167  * Returns an initialized rpcrdma_mr or NULL if no free
1168  * rpcrdma_mr objects are available.
1169  */
1170 struct rpcrdma_mr *
1171 rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt)
1172 {
1173 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1174 	struct rpcrdma_mr *mr;
1175 
1176 	spin_lock(&buf->rb_lock);
1177 	mr = rpcrdma_mr_pop(&buf->rb_mrs);
1178 	spin_unlock(&buf->rb_lock);
1179 	return mr;
1180 }
1181 
1182 /**
1183  * rpcrdma_reply_put - Put reply buffers back into pool
1184  * @buffers: buffer pool
1185  * @req: object to return
1186  *
1187  */
1188 void rpcrdma_reply_put(struct rpcrdma_buffer *buffers, struct rpcrdma_req *req)
1189 {
1190 	if (req->rl_reply) {
1191 		rpcrdma_rep_put(buffers, req->rl_reply);
1192 		req->rl_reply = NULL;
1193 	}
1194 }
1195 
1196 /**
1197  * rpcrdma_buffer_get - Get a request buffer
1198  * @buffers: Buffer pool from which to obtain a buffer
1199  *
1200  * Returns a fresh rpcrdma_req, or NULL if none are available.
1201  */
1202 struct rpcrdma_req *
1203 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1204 {
1205 	struct rpcrdma_req *req;
1206 
1207 	spin_lock(&buffers->rb_lock);
1208 	req = list_first_entry_or_null(&buffers->rb_send_bufs,
1209 				       struct rpcrdma_req, rl_list);
1210 	if (req)
1211 		list_del_init(&req->rl_list);
1212 	spin_unlock(&buffers->rb_lock);
1213 	return req;
1214 }
1215 
1216 /**
1217  * rpcrdma_buffer_put - Put request/reply buffers back into pool
1218  * @buffers: buffer pool
1219  * @req: object to return
1220  *
1221  */
1222 void rpcrdma_buffer_put(struct rpcrdma_buffer *buffers, struct rpcrdma_req *req)
1223 {
1224 	rpcrdma_reply_put(buffers, req);
1225 
1226 	spin_lock(&buffers->rb_lock);
1227 	list_add(&req->rl_list, &buffers->rb_send_bufs);
1228 	spin_unlock(&buffers->rb_lock);
1229 }
1230 
1231 /* Returns a pointer to a rpcrdma_regbuf object, or NULL.
1232  *
1233  * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1234  * receiving the payload of RDMA RECV operations. During Long Calls
1235  * or Replies they may be registered externally via frwr_map.
1236  */
1237 static struct rpcrdma_regbuf *
1238 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction,
1239 		     gfp_t flags)
1240 {
1241 	struct rpcrdma_regbuf *rb;
1242 
1243 	rb = kmalloc(sizeof(*rb), flags);
1244 	if (!rb)
1245 		return NULL;
1246 	rb->rg_data = kmalloc(size, flags);
1247 	if (!rb->rg_data) {
1248 		kfree(rb);
1249 		return NULL;
1250 	}
1251 
1252 	rb->rg_device = NULL;
1253 	rb->rg_direction = direction;
1254 	rb->rg_iov.length = size;
1255 	return rb;
1256 }
1257 
1258 /**
1259  * rpcrdma_regbuf_realloc - re-allocate a SEND/RECV buffer
1260  * @rb: regbuf to reallocate
1261  * @size: size of buffer to be allocated, in bytes
1262  * @flags: GFP flags
1263  *
1264  * Returns true if reallocation was successful. If false is
1265  * returned, @rb is left untouched.
1266  */
1267 bool rpcrdma_regbuf_realloc(struct rpcrdma_regbuf *rb, size_t size, gfp_t flags)
1268 {
1269 	void *buf;
1270 
1271 	buf = kmalloc(size, flags);
1272 	if (!buf)
1273 		return false;
1274 
1275 	rpcrdma_regbuf_dma_unmap(rb);
1276 	kfree(rb->rg_data);
1277 
1278 	rb->rg_data = buf;
1279 	rb->rg_iov.length = size;
1280 	return true;
1281 }
1282 
1283 /**
1284  * __rpcrdma_regbuf_dma_map - DMA-map a regbuf
1285  * @r_xprt: controlling transport instance
1286  * @rb: regbuf to be mapped
1287  *
1288  * Returns true if the buffer is now DMA mapped to @r_xprt's device
1289  */
1290 bool __rpcrdma_regbuf_dma_map(struct rpcrdma_xprt *r_xprt,
1291 			      struct rpcrdma_regbuf *rb)
1292 {
1293 	struct ib_device *device = r_xprt->rx_ep->re_id->device;
1294 
1295 	if (rb->rg_direction == DMA_NONE)
1296 		return false;
1297 
1298 	rb->rg_iov.addr = ib_dma_map_single(device, rdmab_data(rb),
1299 					    rdmab_length(rb), rb->rg_direction);
1300 	if (ib_dma_mapping_error(device, rdmab_addr(rb))) {
1301 		trace_xprtrdma_dma_maperr(rdmab_addr(rb));
1302 		return false;
1303 	}
1304 
1305 	rb->rg_device = device;
1306 	rb->rg_iov.lkey = r_xprt->rx_ep->re_pd->local_dma_lkey;
1307 	return true;
1308 }
1309 
1310 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb)
1311 {
1312 	if (!rb)
1313 		return;
1314 
1315 	if (!rpcrdma_regbuf_is_mapped(rb))
1316 		return;
1317 
1318 	ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb), rdmab_length(rb),
1319 			    rb->rg_direction);
1320 	rb->rg_device = NULL;
1321 }
1322 
1323 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb)
1324 {
1325 	rpcrdma_regbuf_dma_unmap(rb);
1326 	if (rb)
1327 		kfree(rb->rg_data);
1328 	kfree(rb);
1329 }
1330 
1331 /**
1332  * rpcrdma_post_recvs - Refill the Receive Queue
1333  * @r_xprt: controlling transport instance
1334  * @needed: current credit grant
1335  * @temp: mark Receive buffers to be deleted after one use
1336  *
1337  */
1338 void rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, int needed, bool temp)
1339 {
1340 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1341 	struct rpcrdma_ep *ep = r_xprt->rx_ep;
1342 	struct ib_recv_wr *wr, *bad_wr;
1343 	struct rpcrdma_rep *rep;
1344 	int count, rc;
1345 
1346 	rc = 0;
1347 	count = 0;
1348 
1349 	if (likely(ep->re_receive_count > needed))
1350 		goto out;
1351 	needed -= ep->re_receive_count;
1352 	if (!temp)
1353 		needed += RPCRDMA_MAX_RECV_BATCH;
1354 
1355 	if (atomic_inc_return(&ep->re_receiving) > 1)
1356 		goto out;
1357 
1358 	/* fast path: all needed reps can be found on the free list */
1359 	wr = NULL;
1360 	while (needed) {
1361 		rep = rpcrdma_rep_get_locked(buf);
1362 		if (rep && rep->rr_temp) {
1363 			rpcrdma_rep_destroy(rep);
1364 			continue;
1365 		}
1366 		if (!rep)
1367 			rep = rpcrdma_rep_create(r_xprt, temp);
1368 		if (!rep)
1369 			break;
1370 
1371 		rep->rr_cid.ci_queue_id = ep->re_attr.recv_cq->res.id;
1372 		trace_xprtrdma_post_recv(rep);
1373 		rep->rr_recv_wr.next = wr;
1374 		wr = &rep->rr_recv_wr;
1375 		--needed;
1376 		++count;
1377 	}
1378 	if (!wr)
1379 		goto out;
1380 
1381 	rc = ib_post_recv(ep->re_id->qp, wr,
1382 			  (const struct ib_recv_wr **)&bad_wr);
1383 	if (rc) {
1384 		trace_xprtrdma_post_recvs_err(r_xprt, rc);
1385 		for (wr = bad_wr; wr;) {
1386 			struct rpcrdma_rep *rep;
1387 
1388 			rep = container_of(wr, struct rpcrdma_rep, rr_recv_wr);
1389 			wr = wr->next;
1390 			rpcrdma_rep_put(buf, rep);
1391 			--count;
1392 		}
1393 	}
1394 	if (atomic_dec_return(&ep->re_receiving) > 0)
1395 		complete(&ep->re_done);
1396 
1397 out:
1398 	trace_xprtrdma_post_recvs(r_xprt, count);
1399 	ep->re_receive_count += count;
1400 	return;
1401 }
1402