xref: /openbmc/linux/net/sunrpc/xprtrdma/verbs.c (revision 4f3ebd6f)
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 /*
67  * Globals/Macros
68  */
69 
70 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
71 # define RPCDBG_FACILITY	RPCDBG_TRANS
72 #endif
73 
74 /*
75  * internal functions
76  */
77 static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt);
78 static void rpcrdma_sendctxs_destroy(struct rpcrdma_xprt *r_xprt);
79 static void rpcrdma_sendctx_put_locked(struct rpcrdma_xprt *r_xprt,
80 				       struct rpcrdma_sendctx *sc);
81 static int rpcrdma_reqs_setup(struct rpcrdma_xprt *r_xprt);
82 static void rpcrdma_reqs_reset(struct rpcrdma_xprt *r_xprt);
83 static void rpcrdma_rep_destroy(struct rpcrdma_rep *rep);
84 static void rpcrdma_reps_unmap(struct rpcrdma_xprt *r_xprt);
85 static void rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt);
86 static void rpcrdma_mrs_destroy(struct rpcrdma_xprt *r_xprt);
87 static int rpcrdma_ep_destroy(struct rpcrdma_ep *ep);
88 static struct rpcrdma_regbuf *
89 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction,
90 		     gfp_t flags);
91 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb);
92 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb);
93 
94 /* Wait for outstanding transport work to finish. ib_drain_qp
95  * handles the drains in the wrong order for us, so open code
96  * them here.
97  */
98 static void rpcrdma_xprt_drain(struct rpcrdma_xprt *r_xprt)
99 {
100 	struct rdma_cm_id *id = r_xprt->rx_ep->re_id;
101 
102 	/* Flush Receives, then wait for deferred Reply work
103 	 * to complete.
104 	 */
105 	ib_drain_rq(id->qp);
106 
107 	/* Deferred Reply processing might have scheduled
108 	 * local invalidations.
109 	 */
110 	ib_drain_sq(id->qp);
111 }
112 
113 /**
114  * rpcrdma_qp_event_handler - Handle one QP event (error notification)
115  * @event: details of the event
116  * @context: ep that owns QP where event occurred
117  *
118  * Called from the RDMA provider (device driver) possibly in an interrupt
119  * context. The QP is always destroyed before the ID, so the ID will be
120  * reliably available when this handler is invoked.
121  */
122 static void rpcrdma_qp_event_handler(struct ib_event *event, void *context)
123 {
124 	struct rpcrdma_ep *ep = context;
125 
126 	trace_xprtrdma_qp_event(ep, event);
127 }
128 
129 /**
130  * rpcrdma_flush_disconnect - Disconnect on flushed completion
131  * @cq: completion queue
132  * @wc: work completion entry
133  *
134  * Must be called in process context.
135  */
136 void rpcrdma_flush_disconnect(struct ib_cq *cq, struct ib_wc *wc)
137 {
138 	struct rpcrdma_xprt *r_xprt = cq->cq_context;
139 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
140 
141 	if (wc->status != IB_WC_SUCCESS &&
142 	    r_xprt->rx_ep->re_connect_status == 1) {
143 		r_xprt->rx_ep->re_connect_status = -ECONNABORTED;
144 		xprt_force_disconnect(xprt);
145 	}
146 }
147 
148 /**
149  * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC
150  * @cq:	completion queue
151  * @wc:	WCE for a completed Send WR
152  *
153  */
154 static void rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
155 {
156 	struct ib_cqe *cqe = wc->wr_cqe;
157 	struct rpcrdma_sendctx *sc =
158 		container_of(cqe, struct rpcrdma_sendctx, sc_cqe);
159 
160 	/* WARNING: Only wr_cqe and status are reliable at this point */
161 	trace_xprtrdma_wc_send(sc, wc);
162 	rpcrdma_sendctx_put_locked((struct rpcrdma_xprt *)cq->cq_context, sc);
163 	rpcrdma_flush_disconnect(cq, wc);
164 }
165 
166 /**
167  * rpcrdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
168  * @cq:	completion queue
169  * @wc:	WCE for a completed Receive WR
170  *
171  */
172 static void rpcrdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
173 {
174 	struct ib_cqe *cqe = wc->wr_cqe;
175 	struct rpcrdma_rep *rep = container_of(cqe, struct rpcrdma_rep,
176 					       rr_cqe);
177 	struct rpcrdma_xprt *r_xprt = cq->cq_context;
178 
179 	/* WARNING: Only wr_cqe and status are reliable at this point */
180 	trace_xprtrdma_wc_receive(wc);
181 	--r_xprt->rx_ep->re_receive_count;
182 	if (wc->status != IB_WC_SUCCESS)
183 		goto out_flushed;
184 
185 	/* status == SUCCESS means all fields in wc are trustworthy */
186 	rpcrdma_set_xdrlen(&rep->rr_hdrbuf, wc->byte_len);
187 	rep->rr_wc_flags = wc->wc_flags;
188 	rep->rr_inv_rkey = wc->ex.invalidate_rkey;
189 
190 	ib_dma_sync_single_for_cpu(rdmab_device(rep->rr_rdmabuf),
191 				   rdmab_addr(rep->rr_rdmabuf),
192 				   wc->byte_len, DMA_FROM_DEVICE);
193 
194 	rpcrdma_reply_handler(rep);
195 	return;
196 
197 out_flushed:
198 	rpcrdma_flush_disconnect(cq, wc);
199 	rpcrdma_rep_destroy(rep);
200 }
201 
202 static void rpcrdma_update_cm_private(struct rpcrdma_ep *ep,
203 				      struct rdma_conn_param *param)
204 {
205 	const struct rpcrdma_connect_private *pmsg = param->private_data;
206 	unsigned int rsize, wsize;
207 
208 	/* Default settings for RPC-over-RDMA Version One */
209 	ep->re_implicit_roundup = xprt_rdma_pad_optimize;
210 	rsize = RPCRDMA_V1_DEF_INLINE_SIZE;
211 	wsize = RPCRDMA_V1_DEF_INLINE_SIZE;
212 
213 	if (pmsg &&
214 	    pmsg->cp_magic == rpcrdma_cmp_magic &&
215 	    pmsg->cp_version == RPCRDMA_CMP_VERSION) {
216 		ep->re_implicit_roundup = true;
217 		rsize = rpcrdma_decode_buffer_size(pmsg->cp_send_size);
218 		wsize = rpcrdma_decode_buffer_size(pmsg->cp_recv_size);
219 	}
220 
221 	if (rsize < ep->re_inline_recv)
222 		ep->re_inline_recv = rsize;
223 	if (wsize < ep->re_inline_send)
224 		ep->re_inline_send = wsize;
225 
226 	rpcrdma_set_max_header_sizes(ep);
227 }
228 
229 /**
230  * rpcrdma_cm_event_handler - Handle RDMA CM events
231  * @id: rdma_cm_id on which an event has occurred
232  * @event: details of the event
233  *
234  * Called with @id's mutex held. Returns 1 if caller should
235  * destroy @id, otherwise 0.
236  */
237 static int
238 rpcrdma_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
239 {
240 	struct sockaddr *sap = (struct sockaddr *)&id->route.addr.dst_addr;
241 	struct rpcrdma_ep *ep = id->context;
242 	struct rpc_xprt *xprt = ep->re_xprt;
243 
244 	might_sleep();
245 
246 	switch (event->event) {
247 	case RDMA_CM_EVENT_ADDR_RESOLVED:
248 	case RDMA_CM_EVENT_ROUTE_RESOLVED:
249 		ep->re_async_rc = 0;
250 		complete(&ep->re_done);
251 		return 0;
252 	case RDMA_CM_EVENT_ADDR_ERROR:
253 		ep->re_async_rc = -EPROTO;
254 		complete(&ep->re_done);
255 		return 0;
256 	case RDMA_CM_EVENT_ROUTE_ERROR:
257 		ep->re_async_rc = -ENETUNREACH;
258 		complete(&ep->re_done);
259 		return 0;
260 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
261 		pr_info("rpcrdma: removing device %s for %pISpc\n",
262 			ep->re_id->device->name, sap);
263 		/* fall through */
264 	case RDMA_CM_EVENT_ADDR_CHANGE:
265 		ep->re_connect_status = -ENODEV;
266 		xprt_force_disconnect(xprt);
267 		goto disconnected;
268 	case RDMA_CM_EVENT_ESTABLISHED:
269 		kref_get(&ep->re_kref);
270 		ep->re_connect_status = 1;
271 		rpcrdma_update_cm_private(ep, &event->param.conn);
272 		trace_xprtrdma_inline_thresh(ep);
273 		wake_up_all(&ep->re_connect_wait);
274 		break;
275 	case RDMA_CM_EVENT_CONNECT_ERROR:
276 		ep->re_connect_status = -ENOTCONN;
277 		goto disconnected;
278 	case RDMA_CM_EVENT_UNREACHABLE:
279 		ep->re_connect_status = -ENETUNREACH;
280 		goto disconnected;
281 	case RDMA_CM_EVENT_REJECTED:
282 		dprintk("rpcrdma: connection to %pISpc rejected: %s\n",
283 			sap, rdma_reject_msg(id, event->status));
284 		ep->re_connect_status = -ECONNREFUSED;
285 		if (event->status == IB_CM_REJ_STALE_CONN)
286 			ep->re_connect_status = -EAGAIN;
287 		goto disconnected;
288 	case RDMA_CM_EVENT_DISCONNECTED:
289 		ep->re_connect_status = -ECONNABORTED;
290 disconnected:
291 		xprt_force_disconnect(xprt);
292 		return rpcrdma_ep_destroy(ep);
293 	default:
294 		break;
295 	}
296 
297 	dprintk("RPC:       %s: %pISpc on %s/frwr: %s\n", __func__, sap,
298 		ep->re_id->device->name, rdma_event_msg(event->event));
299 	return 0;
300 }
301 
302 static struct rdma_cm_id *rpcrdma_create_id(struct rpcrdma_xprt *r_xprt,
303 					    struct rpcrdma_ep *ep)
304 {
305 	unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1;
306 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
307 	struct rdma_cm_id *id;
308 	int rc;
309 
310 	init_completion(&ep->re_done);
311 
312 	id = rdma_create_id(xprt->xprt_net, rpcrdma_cm_event_handler, ep,
313 			    RDMA_PS_TCP, IB_QPT_RC);
314 	if (IS_ERR(id))
315 		return id;
316 
317 	ep->re_async_rc = -ETIMEDOUT;
318 	rc = rdma_resolve_addr(id, NULL, (struct sockaddr *)&xprt->addr,
319 			       RDMA_RESOLVE_TIMEOUT);
320 	if (rc)
321 		goto out;
322 	rc = wait_for_completion_interruptible_timeout(&ep->re_done, wtimeout);
323 	if (rc < 0)
324 		goto out;
325 
326 	rc = ep->re_async_rc;
327 	if (rc)
328 		goto out;
329 
330 	ep->re_async_rc = -ETIMEDOUT;
331 	rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
332 	if (rc)
333 		goto out;
334 	rc = wait_for_completion_interruptible_timeout(&ep->re_done, wtimeout);
335 	if (rc < 0)
336 		goto out;
337 	rc = ep->re_async_rc;
338 	if (rc)
339 		goto out;
340 
341 	return id;
342 
343 out:
344 	rdma_destroy_id(id);
345 	return ERR_PTR(rc);
346 }
347 
348 static void rpcrdma_ep_put(struct kref *kref)
349 {
350 	struct rpcrdma_ep *ep = container_of(kref, struct rpcrdma_ep, re_kref);
351 
352 	if (ep->re_id->qp) {
353 		rdma_destroy_qp(ep->re_id);
354 		ep->re_id->qp = NULL;
355 	}
356 
357 	if (ep->re_attr.recv_cq)
358 		ib_free_cq(ep->re_attr.recv_cq);
359 	ep->re_attr.recv_cq = NULL;
360 	if (ep->re_attr.send_cq)
361 		ib_free_cq(ep->re_attr.send_cq);
362 	ep->re_attr.send_cq = NULL;
363 
364 	if (ep->re_pd)
365 		ib_dealloc_pd(ep->re_pd);
366 	ep->re_pd = NULL;
367 
368 	kfree(ep);
369 	module_put(THIS_MODULE);
370 }
371 
372 /* Returns:
373  *     %0 if @ep still has a positive kref count, or
374  *     %1 if @ep was destroyed successfully.
375  */
376 static int rpcrdma_ep_destroy(struct rpcrdma_ep *ep)
377 {
378 	return kref_put(&ep->re_kref, rpcrdma_ep_put);
379 }
380 
381 static int rpcrdma_ep_create(struct rpcrdma_xprt *r_xprt)
382 {
383 	struct rpcrdma_connect_private *pmsg;
384 	struct ib_device *device;
385 	struct rdma_cm_id *id;
386 	struct rpcrdma_ep *ep;
387 	int rc;
388 
389 	ep = kzalloc(sizeof(*ep), GFP_NOFS);
390 	if (!ep)
391 		return -EAGAIN;
392 	ep->re_xprt = &r_xprt->rx_xprt;
393 	kref_init(&ep->re_kref);
394 
395 	id = rpcrdma_create_id(r_xprt, ep);
396 	if (IS_ERR(id)) {
397 		rc = PTR_ERR(id);
398 		goto out_free;
399 	}
400 	__module_get(THIS_MODULE);
401 	device = id->device;
402 	ep->re_id = id;
403 
404 	ep->re_max_requests = r_xprt->rx_xprt.max_reqs;
405 	ep->re_inline_send = xprt_rdma_max_inline_write;
406 	ep->re_inline_recv = xprt_rdma_max_inline_read;
407 	rc = frwr_query_device(ep, device);
408 	if (rc)
409 		goto out_destroy;
410 
411 	r_xprt->rx_buf.rb_max_requests = cpu_to_be32(ep->re_max_requests);
412 
413 	ep->re_attr.event_handler = rpcrdma_qp_event_handler;
414 	ep->re_attr.qp_context = ep;
415 	ep->re_attr.srq = NULL;
416 	ep->re_attr.cap.max_inline_data = 0;
417 	ep->re_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
418 	ep->re_attr.qp_type = IB_QPT_RC;
419 	ep->re_attr.port_num = ~0;
420 
421 	dprintk("RPC:       %s: requested max: dtos: send %d recv %d; "
422 		"iovs: send %d recv %d\n",
423 		__func__,
424 		ep->re_attr.cap.max_send_wr,
425 		ep->re_attr.cap.max_recv_wr,
426 		ep->re_attr.cap.max_send_sge,
427 		ep->re_attr.cap.max_recv_sge);
428 
429 	ep->re_send_batch = ep->re_max_requests >> 3;
430 	ep->re_send_count = ep->re_send_batch;
431 	init_waitqueue_head(&ep->re_connect_wait);
432 
433 	ep->re_attr.send_cq = ib_alloc_cq_any(device, r_xprt,
434 					      ep->re_attr.cap.max_send_wr,
435 					      IB_POLL_WORKQUEUE);
436 	if (IS_ERR(ep->re_attr.send_cq)) {
437 		rc = PTR_ERR(ep->re_attr.send_cq);
438 		goto out_destroy;
439 	}
440 
441 	ep->re_attr.recv_cq = ib_alloc_cq_any(device, r_xprt,
442 					      ep->re_attr.cap.max_recv_wr,
443 					      IB_POLL_WORKQUEUE);
444 	if (IS_ERR(ep->re_attr.recv_cq)) {
445 		rc = PTR_ERR(ep->re_attr.recv_cq);
446 		goto out_destroy;
447 	}
448 	ep->re_receive_count = 0;
449 
450 	/* Initialize cma parameters */
451 	memset(&ep->re_remote_cma, 0, sizeof(ep->re_remote_cma));
452 
453 	/* Prepare RDMA-CM private message */
454 	pmsg = &ep->re_cm_private;
455 	pmsg->cp_magic = rpcrdma_cmp_magic;
456 	pmsg->cp_version = RPCRDMA_CMP_VERSION;
457 	pmsg->cp_flags |= RPCRDMA_CMP_F_SND_W_INV_OK;
458 	pmsg->cp_send_size = rpcrdma_encode_buffer_size(ep->re_inline_send);
459 	pmsg->cp_recv_size = rpcrdma_encode_buffer_size(ep->re_inline_recv);
460 	ep->re_remote_cma.private_data = pmsg;
461 	ep->re_remote_cma.private_data_len = sizeof(*pmsg);
462 
463 	/* Client offers RDMA Read but does not initiate */
464 	ep->re_remote_cma.initiator_depth = 0;
465 	ep->re_remote_cma.responder_resources =
466 		min_t(int, U8_MAX, device->attrs.max_qp_rd_atom);
467 
468 	/* Limit transport retries so client can detect server
469 	 * GID changes quickly. RPC layer handles re-establishing
470 	 * transport connection and retransmission.
471 	 */
472 	ep->re_remote_cma.retry_count = 6;
473 
474 	/* RPC-over-RDMA handles its own flow control. In addition,
475 	 * make all RNR NAKs visible so we know that RPC-over-RDMA
476 	 * flow control is working correctly (no NAKs should be seen).
477 	 */
478 	ep->re_remote_cma.flow_control = 0;
479 	ep->re_remote_cma.rnr_retry_count = 0;
480 
481 	ep->re_pd = ib_alloc_pd(device, 0);
482 	if (IS_ERR(ep->re_pd)) {
483 		rc = PTR_ERR(ep->re_pd);
484 		goto out_destroy;
485 	}
486 
487 	rc = rdma_create_qp(id, ep->re_pd, &ep->re_attr);
488 	if (rc)
489 		goto out_destroy;
490 
491 	r_xprt->rx_ep = ep;
492 	return 0;
493 
494 out_destroy:
495 	rpcrdma_ep_destroy(ep);
496 	rdma_destroy_id(id);
497 out_free:
498 	kfree(ep);
499 	r_xprt->rx_ep = NULL;
500 	return rc;
501 }
502 
503 /**
504  * rpcrdma_xprt_connect - Connect an unconnected transport
505  * @r_xprt: controlling transport instance
506  *
507  * Returns 0 on success or a negative errno.
508  */
509 int rpcrdma_xprt_connect(struct rpcrdma_xprt *r_xprt)
510 {
511 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
512 	struct rpcrdma_ep *ep;
513 	int rc;
514 
515 retry:
516 	rpcrdma_xprt_disconnect(r_xprt);
517 	rc = rpcrdma_ep_create(r_xprt);
518 	if (rc)
519 		return rc;
520 	ep = r_xprt->rx_ep;
521 
522 	ep->re_connect_status = 0;
523 	xprt_clear_connected(xprt);
524 
525 	rpcrdma_reset_cwnd(r_xprt);
526 	rpcrdma_post_recvs(r_xprt, true);
527 
528 	rc = rpcrdma_sendctxs_create(r_xprt);
529 	if (rc)
530 		goto out;
531 
532 	rc = rdma_connect(ep->re_id, &ep->re_remote_cma);
533 	if (rc)
534 		goto out;
535 
536 	if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
537 		xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
538 	wait_event_interruptible(ep->re_connect_wait,
539 				 ep->re_connect_status != 0);
540 	if (ep->re_connect_status <= 0) {
541 		if (ep->re_connect_status == -EAGAIN)
542 			goto retry;
543 		rc = ep->re_connect_status;
544 		goto out;
545 	}
546 
547 	rc = rpcrdma_reqs_setup(r_xprt);
548 	if (rc) {
549 		rpcrdma_xprt_disconnect(r_xprt);
550 		goto out;
551 	}
552 	rpcrdma_mrs_create(r_xprt);
553 
554 out:
555 	if (rc)
556 		ep->re_connect_status = rc;
557 	trace_xprtrdma_connect(r_xprt, rc);
558 	return rc;
559 }
560 
561 /**
562  * rpcrdma_xprt_disconnect - Disconnect underlying transport
563  * @r_xprt: controlling transport instance
564  *
565  * Caller serializes. Either the transport send lock is held,
566  * or we're being called to destroy the transport.
567  *
568  * On return, @r_xprt is completely divested of all hardware
569  * resources and prepared for the next ->connect operation.
570  */
571 void rpcrdma_xprt_disconnect(struct rpcrdma_xprt *r_xprt)
572 {
573 	struct rpcrdma_ep *ep = r_xprt->rx_ep;
574 	struct rdma_cm_id *id;
575 	int rc;
576 
577 	if (!ep)
578 		return;
579 
580 	id = ep->re_id;
581 	rc = rdma_disconnect(id);
582 	trace_xprtrdma_disconnect(r_xprt, rc);
583 
584 	rpcrdma_xprt_drain(r_xprt);
585 	rpcrdma_reps_unmap(r_xprt);
586 	rpcrdma_reqs_reset(r_xprt);
587 	rpcrdma_mrs_destroy(r_xprt);
588 	rpcrdma_sendctxs_destroy(r_xprt);
589 
590 	if (rpcrdma_ep_destroy(ep))
591 		rdma_destroy_id(id);
592 
593 	r_xprt->rx_ep = NULL;
594 }
595 
596 /* Fixed-size circular FIFO queue. This implementation is wait-free and
597  * lock-free.
598  *
599  * Consumer is the code path that posts Sends. This path dequeues a
600  * sendctx for use by a Send operation. Multiple consumer threads
601  * are serialized by the RPC transport lock, which allows only one
602  * ->send_request call at a time.
603  *
604  * Producer is the code path that handles Send completions. This path
605  * enqueues a sendctx that has been completed. Multiple producer
606  * threads are serialized by the ib_poll_cq() function.
607  */
608 
609 /* rpcrdma_sendctxs_destroy() assumes caller has already quiesced
610  * queue activity, and rpcrdma_xprt_drain has flushed all remaining
611  * Send requests.
612  */
613 static void rpcrdma_sendctxs_destroy(struct rpcrdma_xprt *r_xprt)
614 {
615 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
616 	unsigned long i;
617 
618 	if (!buf->rb_sc_ctxs)
619 		return;
620 	for (i = 0; i <= buf->rb_sc_last; i++)
621 		kfree(buf->rb_sc_ctxs[i]);
622 	kfree(buf->rb_sc_ctxs);
623 	buf->rb_sc_ctxs = NULL;
624 }
625 
626 static struct rpcrdma_sendctx *rpcrdma_sendctx_create(struct rpcrdma_ep *ep)
627 {
628 	struct rpcrdma_sendctx *sc;
629 
630 	sc = kzalloc(struct_size(sc, sc_sges, ep->re_attr.cap.max_send_sge),
631 		     GFP_KERNEL);
632 	if (!sc)
633 		return NULL;
634 
635 	sc->sc_cqe.done = rpcrdma_wc_send;
636 	return sc;
637 }
638 
639 static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt)
640 {
641 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
642 	struct rpcrdma_sendctx *sc;
643 	unsigned long i;
644 
645 	/* Maximum number of concurrent outstanding Send WRs. Capping
646 	 * the circular queue size stops Send Queue overflow by causing
647 	 * the ->send_request call to fail temporarily before too many
648 	 * Sends are posted.
649 	 */
650 	i = r_xprt->rx_ep->re_max_requests + RPCRDMA_MAX_BC_REQUESTS;
651 	buf->rb_sc_ctxs = kcalloc(i, sizeof(sc), GFP_KERNEL);
652 	if (!buf->rb_sc_ctxs)
653 		return -ENOMEM;
654 
655 	buf->rb_sc_last = i - 1;
656 	for (i = 0; i <= buf->rb_sc_last; i++) {
657 		sc = rpcrdma_sendctx_create(r_xprt->rx_ep);
658 		if (!sc)
659 			return -ENOMEM;
660 
661 		buf->rb_sc_ctxs[i] = sc;
662 	}
663 
664 	buf->rb_sc_head = 0;
665 	buf->rb_sc_tail = 0;
666 	return 0;
667 }
668 
669 /* The sendctx queue is not guaranteed to have a size that is a
670  * power of two, thus the helpers in circ_buf.h cannot be used.
671  * The other option is to use modulus (%), which can be expensive.
672  */
673 static unsigned long rpcrdma_sendctx_next(struct rpcrdma_buffer *buf,
674 					  unsigned long item)
675 {
676 	return likely(item < buf->rb_sc_last) ? item + 1 : 0;
677 }
678 
679 /**
680  * rpcrdma_sendctx_get_locked - Acquire a send context
681  * @r_xprt: controlling transport instance
682  *
683  * Returns pointer to a free send completion context; or NULL if
684  * the queue is empty.
685  *
686  * Usage: Called to acquire an SGE array before preparing a Send WR.
687  *
688  * The caller serializes calls to this function (per transport), and
689  * provides an effective memory barrier that flushes the new value
690  * of rb_sc_head.
691  */
692 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_xprt *r_xprt)
693 {
694 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
695 	struct rpcrdma_sendctx *sc;
696 	unsigned long next_head;
697 
698 	next_head = rpcrdma_sendctx_next(buf, buf->rb_sc_head);
699 
700 	if (next_head == READ_ONCE(buf->rb_sc_tail))
701 		goto out_emptyq;
702 
703 	/* ORDER: item must be accessed _before_ head is updated */
704 	sc = buf->rb_sc_ctxs[next_head];
705 
706 	/* Releasing the lock in the caller acts as a memory
707 	 * barrier that flushes rb_sc_head.
708 	 */
709 	buf->rb_sc_head = next_head;
710 
711 	return sc;
712 
713 out_emptyq:
714 	/* The queue is "empty" if there have not been enough Send
715 	 * completions recently. This is a sign the Send Queue is
716 	 * backing up. Cause the caller to pause and try again.
717 	 */
718 	xprt_wait_for_buffer_space(&r_xprt->rx_xprt);
719 	r_xprt->rx_stats.empty_sendctx_q++;
720 	return NULL;
721 }
722 
723 /**
724  * rpcrdma_sendctx_put_locked - Release a send context
725  * @r_xprt: controlling transport instance
726  * @sc: send context to release
727  *
728  * Usage: Called from Send completion to return a sendctxt
729  * to the queue.
730  *
731  * The caller serializes calls to this function (per transport).
732  */
733 static void rpcrdma_sendctx_put_locked(struct rpcrdma_xprt *r_xprt,
734 				       struct rpcrdma_sendctx *sc)
735 {
736 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
737 	unsigned long next_tail;
738 
739 	/* Unmap SGEs of previously completed but unsignaled
740 	 * Sends by walking up the queue until @sc is found.
741 	 */
742 	next_tail = buf->rb_sc_tail;
743 	do {
744 		next_tail = rpcrdma_sendctx_next(buf, next_tail);
745 
746 		/* ORDER: item must be accessed _before_ tail is updated */
747 		rpcrdma_sendctx_unmap(buf->rb_sc_ctxs[next_tail]);
748 
749 	} while (buf->rb_sc_ctxs[next_tail] != sc);
750 
751 	/* Paired with READ_ONCE */
752 	smp_store_release(&buf->rb_sc_tail, next_tail);
753 
754 	xprt_write_space(&r_xprt->rx_xprt);
755 }
756 
757 static void
758 rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt)
759 {
760 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
761 	struct rpcrdma_ep *ep = r_xprt->rx_ep;
762 	unsigned int count;
763 
764 	for (count = 0; count < ep->re_max_rdma_segs; count++) {
765 		struct rpcrdma_mr *mr;
766 		int rc;
767 
768 		mr = kzalloc(sizeof(*mr), GFP_NOFS);
769 		if (!mr)
770 			break;
771 
772 		rc = frwr_mr_init(r_xprt, mr);
773 		if (rc) {
774 			kfree(mr);
775 			break;
776 		}
777 
778 		spin_lock(&buf->rb_lock);
779 		rpcrdma_mr_push(mr, &buf->rb_mrs);
780 		list_add(&mr->mr_all, &buf->rb_all_mrs);
781 		spin_unlock(&buf->rb_lock);
782 	}
783 
784 	r_xprt->rx_stats.mrs_allocated += count;
785 	trace_xprtrdma_createmrs(r_xprt, count);
786 }
787 
788 static void
789 rpcrdma_mr_refresh_worker(struct work_struct *work)
790 {
791 	struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
792 						  rb_refresh_worker);
793 	struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
794 						   rx_buf);
795 
796 	rpcrdma_mrs_create(r_xprt);
797 	xprt_write_space(&r_xprt->rx_xprt);
798 }
799 
800 /**
801  * rpcrdma_mrs_refresh - Wake the MR refresh worker
802  * @r_xprt: controlling transport instance
803  *
804  */
805 void rpcrdma_mrs_refresh(struct rpcrdma_xprt *r_xprt)
806 {
807 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
808 	struct rpcrdma_ep *ep = r_xprt->rx_ep;
809 
810 	/* If there is no underlying connection, it's no use
811 	 * to wake the refresh worker.
812 	 */
813 	if (ep->re_connect_status == 1) {
814 		/* The work is scheduled on a WQ_MEM_RECLAIM
815 		 * workqueue in order to prevent MR allocation
816 		 * from recursing into NFS during direct reclaim.
817 		 */
818 		queue_work(xprtiod_workqueue, &buf->rb_refresh_worker);
819 	}
820 }
821 
822 /**
823  * rpcrdma_req_create - Allocate an rpcrdma_req object
824  * @r_xprt: controlling r_xprt
825  * @size: initial size, in bytes, of send and receive buffers
826  * @flags: GFP flags passed to memory allocators
827  *
828  * Returns an allocated and fully initialized rpcrdma_req or NULL.
829  */
830 struct rpcrdma_req *rpcrdma_req_create(struct rpcrdma_xprt *r_xprt, size_t size,
831 				       gfp_t flags)
832 {
833 	struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
834 	struct rpcrdma_req *req;
835 
836 	req = kzalloc(sizeof(*req), flags);
837 	if (req == NULL)
838 		goto out1;
839 
840 	req->rl_sendbuf = rpcrdma_regbuf_alloc(size, DMA_TO_DEVICE, flags);
841 	if (!req->rl_sendbuf)
842 		goto out2;
843 
844 	req->rl_recvbuf = rpcrdma_regbuf_alloc(size, DMA_NONE, flags);
845 	if (!req->rl_recvbuf)
846 		goto out3;
847 
848 	INIT_LIST_HEAD(&req->rl_free_mrs);
849 	INIT_LIST_HEAD(&req->rl_registered);
850 	spin_lock(&buffer->rb_lock);
851 	list_add(&req->rl_all, &buffer->rb_allreqs);
852 	spin_unlock(&buffer->rb_lock);
853 	return req;
854 
855 out3:
856 	kfree(req->rl_sendbuf);
857 out2:
858 	kfree(req);
859 out1:
860 	return NULL;
861 }
862 
863 /**
864  * rpcrdma_req_setup - Per-connection instance setup of an rpcrdma_req object
865  * @r_xprt: controlling transport instance
866  * @req: rpcrdma_req object to set up
867  *
868  * Returns zero on success, and a negative errno on failure.
869  */
870 int rpcrdma_req_setup(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
871 {
872 	struct rpcrdma_regbuf *rb;
873 	size_t maxhdrsize;
874 
875 	/* Compute maximum header buffer size in bytes */
876 	maxhdrsize = rpcrdma_fixed_maxsz + 3 +
877 		     r_xprt->rx_ep->re_max_rdma_segs * rpcrdma_readchunk_maxsz;
878 	maxhdrsize *= sizeof(__be32);
879 	rb = rpcrdma_regbuf_alloc(__roundup_pow_of_two(maxhdrsize),
880 				  DMA_TO_DEVICE, GFP_KERNEL);
881 	if (!rb)
882 		goto out;
883 
884 	if (!__rpcrdma_regbuf_dma_map(r_xprt, rb))
885 		goto out_free;
886 
887 	req->rl_rdmabuf = rb;
888 	xdr_buf_init(&req->rl_hdrbuf, rdmab_data(rb), rdmab_length(rb));
889 	return 0;
890 
891 out_free:
892 	rpcrdma_regbuf_free(rb);
893 out:
894 	return -ENOMEM;
895 }
896 
897 /* ASSUMPTION: the rb_allreqs list is stable for the duration,
898  * and thus can be walked without holding rb_lock. Eg. the
899  * caller is holding the transport send lock to exclude
900  * device removal or disconnection.
901  */
902 static int rpcrdma_reqs_setup(struct rpcrdma_xprt *r_xprt)
903 {
904 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
905 	struct rpcrdma_req *req;
906 	int rc;
907 
908 	list_for_each_entry(req, &buf->rb_allreqs, rl_all) {
909 		rc = rpcrdma_req_setup(r_xprt, req);
910 		if (rc)
911 			return rc;
912 	}
913 	return 0;
914 }
915 
916 static void rpcrdma_req_reset(struct rpcrdma_req *req)
917 {
918 	/* Credits are valid for only one connection */
919 	req->rl_slot.rq_cong = 0;
920 
921 	rpcrdma_regbuf_free(req->rl_rdmabuf);
922 	req->rl_rdmabuf = NULL;
923 
924 	rpcrdma_regbuf_dma_unmap(req->rl_sendbuf);
925 	rpcrdma_regbuf_dma_unmap(req->rl_recvbuf);
926 }
927 
928 /* ASSUMPTION: the rb_allreqs list is stable for the duration,
929  * and thus can be walked without holding rb_lock. Eg. the
930  * caller is holding the transport send lock to exclude
931  * device removal or disconnection.
932  */
933 static void rpcrdma_reqs_reset(struct rpcrdma_xprt *r_xprt)
934 {
935 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
936 	struct rpcrdma_req *req;
937 
938 	list_for_each_entry(req, &buf->rb_allreqs, rl_all)
939 		rpcrdma_req_reset(req);
940 }
941 
942 /* No locking needed here. This function is called only by the
943  * Receive completion handler.
944  */
945 static noinline
946 struct rpcrdma_rep *rpcrdma_rep_create(struct rpcrdma_xprt *r_xprt,
947 				       bool temp)
948 {
949 	struct rpcrdma_rep *rep;
950 
951 	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
952 	if (rep == NULL)
953 		goto out;
954 
955 	rep->rr_rdmabuf = rpcrdma_regbuf_alloc(r_xprt->rx_ep->re_inline_recv,
956 					       DMA_FROM_DEVICE, GFP_KERNEL);
957 	if (!rep->rr_rdmabuf)
958 		goto out_free;
959 
960 	if (!rpcrdma_regbuf_dma_map(r_xprt, rep->rr_rdmabuf))
961 		goto out_free_regbuf;
962 
963 	xdr_buf_init(&rep->rr_hdrbuf, rdmab_data(rep->rr_rdmabuf),
964 		     rdmab_length(rep->rr_rdmabuf));
965 	rep->rr_cqe.done = rpcrdma_wc_receive;
966 	rep->rr_rxprt = r_xprt;
967 	rep->rr_recv_wr.next = NULL;
968 	rep->rr_recv_wr.wr_cqe = &rep->rr_cqe;
969 	rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
970 	rep->rr_recv_wr.num_sge = 1;
971 	rep->rr_temp = temp;
972 	list_add(&rep->rr_all, &r_xprt->rx_buf.rb_all_reps);
973 	return rep;
974 
975 out_free_regbuf:
976 	rpcrdma_regbuf_free(rep->rr_rdmabuf);
977 out_free:
978 	kfree(rep);
979 out:
980 	return NULL;
981 }
982 
983 /* No locking needed here. This function is invoked only by the
984  * Receive completion handler, or during transport shutdown.
985  */
986 static void rpcrdma_rep_destroy(struct rpcrdma_rep *rep)
987 {
988 	list_del(&rep->rr_all);
989 	rpcrdma_regbuf_free(rep->rr_rdmabuf);
990 	kfree(rep);
991 }
992 
993 static struct rpcrdma_rep *rpcrdma_rep_get_locked(struct rpcrdma_buffer *buf)
994 {
995 	struct llist_node *node;
996 
997 	/* Calls to llist_del_first are required to be serialized */
998 	node = llist_del_first(&buf->rb_free_reps);
999 	if (!node)
1000 		return NULL;
1001 	return llist_entry(node, struct rpcrdma_rep, rr_node);
1002 }
1003 
1004 static void rpcrdma_rep_put(struct rpcrdma_buffer *buf,
1005 			    struct rpcrdma_rep *rep)
1006 {
1007 	llist_add(&rep->rr_node, &buf->rb_free_reps);
1008 }
1009 
1010 static void rpcrdma_reps_unmap(struct rpcrdma_xprt *r_xprt)
1011 {
1012 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1013 	struct rpcrdma_rep *rep;
1014 
1015 	list_for_each_entry(rep, &buf->rb_all_reps, rr_all) {
1016 		rpcrdma_regbuf_dma_unmap(rep->rr_rdmabuf);
1017 		rep->rr_temp = true;
1018 	}
1019 }
1020 
1021 static void rpcrdma_reps_destroy(struct rpcrdma_buffer *buf)
1022 {
1023 	struct rpcrdma_rep *rep;
1024 
1025 	while ((rep = rpcrdma_rep_get_locked(buf)) != NULL)
1026 		rpcrdma_rep_destroy(rep);
1027 }
1028 
1029 /**
1030  * rpcrdma_buffer_create - Create initial set of req/rep objects
1031  * @r_xprt: transport instance to (re)initialize
1032  *
1033  * Returns zero on success, otherwise a negative errno.
1034  */
1035 int rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
1036 {
1037 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1038 	int i, rc;
1039 
1040 	buf->rb_bc_srv_max_requests = 0;
1041 	spin_lock_init(&buf->rb_lock);
1042 	INIT_LIST_HEAD(&buf->rb_mrs);
1043 	INIT_LIST_HEAD(&buf->rb_all_mrs);
1044 	INIT_WORK(&buf->rb_refresh_worker, rpcrdma_mr_refresh_worker);
1045 
1046 	INIT_LIST_HEAD(&buf->rb_send_bufs);
1047 	INIT_LIST_HEAD(&buf->rb_allreqs);
1048 	INIT_LIST_HEAD(&buf->rb_all_reps);
1049 
1050 	rc = -ENOMEM;
1051 	for (i = 0; i < r_xprt->rx_xprt.max_reqs; i++) {
1052 		struct rpcrdma_req *req;
1053 
1054 		req = rpcrdma_req_create(r_xprt, RPCRDMA_V1_DEF_INLINE_SIZE * 2,
1055 					 GFP_KERNEL);
1056 		if (!req)
1057 			goto out;
1058 		list_add(&req->rl_list, &buf->rb_send_bufs);
1059 	}
1060 
1061 	init_llist_head(&buf->rb_free_reps);
1062 
1063 	return 0;
1064 out:
1065 	rpcrdma_buffer_destroy(buf);
1066 	return rc;
1067 }
1068 
1069 /**
1070  * rpcrdma_req_destroy - Destroy an rpcrdma_req object
1071  * @req: unused object to be destroyed
1072  *
1073  * Relies on caller holding the transport send lock to protect
1074  * removing req->rl_all from buf->rb_all_reqs safely.
1075  */
1076 void rpcrdma_req_destroy(struct rpcrdma_req *req)
1077 {
1078 	struct rpcrdma_mr *mr;
1079 
1080 	list_del(&req->rl_all);
1081 
1082 	while ((mr = rpcrdma_mr_pop(&req->rl_free_mrs))) {
1083 		struct rpcrdma_buffer *buf = &mr->mr_xprt->rx_buf;
1084 
1085 		spin_lock(&buf->rb_lock);
1086 		list_del(&mr->mr_all);
1087 		spin_unlock(&buf->rb_lock);
1088 
1089 		frwr_release_mr(mr);
1090 	}
1091 
1092 	rpcrdma_regbuf_free(req->rl_recvbuf);
1093 	rpcrdma_regbuf_free(req->rl_sendbuf);
1094 	rpcrdma_regbuf_free(req->rl_rdmabuf);
1095 	kfree(req);
1096 }
1097 
1098 /**
1099  * rpcrdma_mrs_destroy - Release all of a transport's MRs
1100  * @r_xprt: controlling transport instance
1101  *
1102  * Relies on caller holding the transport send lock to protect
1103  * removing mr->mr_list from req->rl_free_mrs safely.
1104  */
1105 static void rpcrdma_mrs_destroy(struct rpcrdma_xprt *r_xprt)
1106 {
1107 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1108 	struct rpcrdma_mr *mr;
1109 
1110 	cancel_work_sync(&buf->rb_refresh_worker);
1111 
1112 	spin_lock(&buf->rb_lock);
1113 	while ((mr = list_first_entry_or_null(&buf->rb_all_mrs,
1114 					      struct rpcrdma_mr,
1115 					      mr_all)) != NULL) {
1116 		list_del(&mr->mr_list);
1117 		list_del(&mr->mr_all);
1118 		spin_unlock(&buf->rb_lock);
1119 
1120 		frwr_release_mr(mr);
1121 
1122 		spin_lock(&buf->rb_lock);
1123 	}
1124 	spin_unlock(&buf->rb_lock);
1125 }
1126 
1127 /**
1128  * rpcrdma_buffer_destroy - Release all hw resources
1129  * @buf: root control block for resources
1130  *
1131  * ORDERING: relies on a prior rpcrdma_xprt_drain :
1132  * - No more Send or Receive completions can occur
1133  * - All MRs, reps, and reqs are returned to their free lists
1134  */
1135 void
1136 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1137 {
1138 	rpcrdma_reps_destroy(buf);
1139 
1140 	while (!list_empty(&buf->rb_send_bufs)) {
1141 		struct rpcrdma_req *req;
1142 
1143 		req = list_first_entry(&buf->rb_send_bufs,
1144 				       struct rpcrdma_req, rl_list);
1145 		list_del(&req->rl_list);
1146 		rpcrdma_req_destroy(req);
1147 	}
1148 }
1149 
1150 /**
1151  * rpcrdma_mr_get - Allocate an rpcrdma_mr object
1152  * @r_xprt: controlling transport
1153  *
1154  * Returns an initialized rpcrdma_mr or NULL if no free
1155  * rpcrdma_mr objects are available.
1156  */
1157 struct rpcrdma_mr *
1158 rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt)
1159 {
1160 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1161 	struct rpcrdma_mr *mr;
1162 
1163 	spin_lock(&buf->rb_lock);
1164 	mr = rpcrdma_mr_pop(&buf->rb_mrs);
1165 	spin_unlock(&buf->rb_lock);
1166 	return mr;
1167 }
1168 
1169 /**
1170  * rpcrdma_mr_put - DMA unmap an MR and release it
1171  * @mr: MR to release
1172  *
1173  */
1174 void rpcrdma_mr_put(struct rpcrdma_mr *mr)
1175 {
1176 	struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
1177 
1178 	if (mr->mr_dir != DMA_NONE) {
1179 		trace_xprtrdma_mr_unmap(mr);
1180 		ib_dma_unmap_sg(r_xprt->rx_ep->re_id->device,
1181 				mr->mr_sg, mr->mr_nents, mr->mr_dir);
1182 		mr->mr_dir = DMA_NONE;
1183 	}
1184 
1185 	rpcrdma_mr_push(mr, &mr->mr_req->rl_free_mrs);
1186 }
1187 
1188 /**
1189  * rpcrdma_buffer_get - Get a request buffer
1190  * @buffers: Buffer pool from which to obtain a buffer
1191  *
1192  * Returns a fresh rpcrdma_req, or NULL if none are available.
1193  */
1194 struct rpcrdma_req *
1195 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1196 {
1197 	struct rpcrdma_req *req;
1198 
1199 	spin_lock(&buffers->rb_lock);
1200 	req = list_first_entry_or_null(&buffers->rb_send_bufs,
1201 				       struct rpcrdma_req, rl_list);
1202 	if (req)
1203 		list_del_init(&req->rl_list);
1204 	spin_unlock(&buffers->rb_lock);
1205 	return req;
1206 }
1207 
1208 /**
1209  * rpcrdma_buffer_put - Put request/reply buffers back into pool
1210  * @buffers: buffer pool
1211  * @req: object to return
1212  *
1213  */
1214 void rpcrdma_buffer_put(struct rpcrdma_buffer *buffers, struct rpcrdma_req *req)
1215 {
1216 	if (req->rl_reply)
1217 		rpcrdma_rep_put(buffers, req->rl_reply);
1218 	req->rl_reply = NULL;
1219 
1220 	spin_lock(&buffers->rb_lock);
1221 	list_add(&req->rl_list, &buffers->rb_send_bufs);
1222 	spin_unlock(&buffers->rb_lock);
1223 }
1224 
1225 /**
1226  * rpcrdma_recv_buffer_put - Release rpcrdma_rep back to free list
1227  * @rep: rep to release
1228  *
1229  * Used after error conditions.
1230  */
1231 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1232 {
1233 	rpcrdma_rep_put(&rep->rr_rxprt->rx_buf, rep);
1234 }
1235 
1236 /* Returns a pointer to a rpcrdma_regbuf object, or NULL.
1237  *
1238  * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1239  * receiving the payload of RDMA RECV operations. During Long Calls
1240  * or Replies they may be registered externally via frwr_map.
1241  */
1242 static struct rpcrdma_regbuf *
1243 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction,
1244 		     gfp_t flags)
1245 {
1246 	struct rpcrdma_regbuf *rb;
1247 
1248 	rb = kmalloc(sizeof(*rb), flags);
1249 	if (!rb)
1250 		return NULL;
1251 	rb->rg_data = kmalloc(size, flags);
1252 	if (!rb->rg_data) {
1253 		kfree(rb);
1254 		return NULL;
1255 	}
1256 
1257 	rb->rg_device = NULL;
1258 	rb->rg_direction = direction;
1259 	rb->rg_iov.length = size;
1260 	return rb;
1261 }
1262 
1263 /**
1264  * rpcrdma_regbuf_realloc - re-allocate a SEND/RECV buffer
1265  * @rb: regbuf to reallocate
1266  * @size: size of buffer to be allocated, in bytes
1267  * @flags: GFP flags
1268  *
1269  * Returns true if reallocation was successful. If false is
1270  * returned, @rb is left untouched.
1271  */
1272 bool rpcrdma_regbuf_realloc(struct rpcrdma_regbuf *rb, size_t size, gfp_t flags)
1273 {
1274 	void *buf;
1275 
1276 	buf = kmalloc(size, flags);
1277 	if (!buf)
1278 		return false;
1279 
1280 	rpcrdma_regbuf_dma_unmap(rb);
1281 	kfree(rb->rg_data);
1282 
1283 	rb->rg_data = buf;
1284 	rb->rg_iov.length = size;
1285 	return true;
1286 }
1287 
1288 /**
1289  * __rpcrdma_regbuf_dma_map - DMA-map a regbuf
1290  * @r_xprt: controlling transport instance
1291  * @rb: regbuf to be mapped
1292  *
1293  * Returns true if the buffer is now DMA mapped to @r_xprt's device
1294  */
1295 bool __rpcrdma_regbuf_dma_map(struct rpcrdma_xprt *r_xprt,
1296 			      struct rpcrdma_regbuf *rb)
1297 {
1298 	struct ib_device *device = r_xprt->rx_ep->re_id->device;
1299 
1300 	if (rb->rg_direction == DMA_NONE)
1301 		return false;
1302 
1303 	rb->rg_iov.addr = ib_dma_map_single(device, rdmab_data(rb),
1304 					    rdmab_length(rb), rb->rg_direction);
1305 	if (ib_dma_mapping_error(device, rdmab_addr(rb))) {
1306 		trace_xprtrdma_dma_maperr(rdmab_addr(rb));
1307 		return false;
1308 	}
1309 
1310 	rb->rg_device = device;
1311 	rb->rg_iov.lkey = r_xprt->rx_ep->re_pd->local_dma_lkey;
1312 	return true;
1313 }
1314 
1315 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb)
1316 {
1317 	if (!rb)
1318 		return;
1319 
1320 	if (!rpcrdma_regbuf_is_mapped(rb))
1321 		return;
1322 
1323 	ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb), rdmab_length(rb),
1324 			    rb->rg_direction);
1325 	rb->rg_device = NULL;
1326 }
1327 
1328 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb)
1329 {
1330 	rpcrdma_regbuf_dma_unmap(rb);
1331 	if (rb)
1332 		kfree(rb->rg_data);
1333 	kfree(rb);
1334 }
1335 
1336 /**
1337  * rpcrdma_post_sends - Post WRs to a transport's Send Queue
1338  * @r_xprt: controlling transport instance
1339  * @req: rpcrdma_req containing the Send WR to post
1340  *
1341  * Returns 0 if the post was successful, otherwise -ENOTCONN
1342  * is returned.
1343  */
1344 int rpcrdma_post_sends(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
1345 {
1346 	struct ib_send_wr *send_wr = &req->rl_wr;
1347 	struct rpcrdma_ep *ep = r_xprt->rx_ep;
1348 	int rc;
1349 
1350 	if (!ep->re_send_count || kref_read(&req->rl_kref) > 1) {
1351 		send_wr->send_flags |= IB_SEND_SIGNALED;
1352 		ep->re_send_count = ep->re_send_batch;
1353 	} else {
1354 		send_wr->send_flags &= ~IB_SEND_SIGNALED;
1355 		--ep->re_send_count;
1356 	}
1357 
1358 	trace_xprtrdma_post_send(req);
1359 	rc = frwr_send(r_xprt, req);
1360 	if (rc)
1361 		return -ENOTCONN;
1362 	return 0;
1363 }
1364 
1365 /**
1366  * rpcrdma_post_recvs - Refill the Receive Queue
1367  * @r_xprt: controlling transport instance
1368  * @temp: mark Receive buffers to be deleted after use
1369  *
1370  */
1371 void rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, bool temp)
1372 {
1373 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1374 	struct rpcrdma_ep *ep = r_xprt->rx_ep;
1375 	struct ib_recv_wr *wr, *bad_wr;
1376 	struct rpcrdma_rep *rep;
1377 	int needed, count, rc;
1378 
1379 	rc = 0;
1380 	count = 0;
1381 
1382 	needed = buf->rb_credits + (buf->rb_bc_srv_max_requests << 1);
1383 	if (likely(ep->re_receive_count > needed))
1384 		goto out;
1385 	needed -= ep->re_receive_count;
1386 	if (!temp)
1387 		needed += RPCRDMA_MAX_RECV_BATCH;
1388 
1389 	/* fast path: all needed reps can be found on the free list */
1390 	wr = NULL;
1391 	while (needed) {
1392 		rep = rpcrdma_rep_get_locked(buf);
1393 		if (rep && rep->rr_temp) {
1394 			rpcrdma_rep_destroy(rep);
1395 			continue;
1396 		}
1397 		if (!rep)
1398 			rep = rpcrdma_rep_create(r_xprt, temp);
1399 		if (!rep)
1400 			break;
1401 
1402 		trace_xprtrdma_post_recv(rep);
1403 		rep->rr_recv_wr.next = wr;
1404 		wr = &rep->rr_recv_wr;
1405 		--needed;
1406 		++count;
1407 	}
1408 	if (!wr)
1409 		goto out;
1410 
1411 	rc = ib_post_recv(ep->re_id->qp, wr,
1412 			  (const struct ib_recv_wr **)&bad_wr);
1413 out:
1414 	trace_xprtrdma_post_recvs(r_xprt, count, rc);
1415 	if (rc) {
1416 		for (wr = bad_wr; wr;) {
1417 			struct rpcrdma_rep *rep;
1418 
1419 			rep = container_of(wr, struct rpcrdma_rep, rr_recv_wr);
1420 			wr = wr->next;
1421 			rpcrdma_recv_buffer_put(rep);
1422 			--count;
1423 		}
1424 	}
1425 	ep->re_receive_count += count;
1426 	return;
1427 }
1428