xref: /openbmc/linux/net/sunrpc/xprtrdma/verbs.c (revision 3a83e4e6)
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 void rpcrdma_ep_get(struct rpcrdma_ep *ep);
88 static int rpcrdma_ep_put(struct rpcrdma_ep *ep);
89 static struct rpcrdma_regbuf *
90 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction,
91 		     gfp_t flags);
92 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb);
93 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb);
94 
95 /* Wait for outstanding transport work to finish. ib_drain_qp
96  * handles the drains in the wrong order for us, so open code
97  * them here.
98  */
99 static void rpcrdma_xprt_drain(struct rpcrdma_xprt *r_xprt)
100 {
101 	struct rpcrdma_ep *ep = r_xprt->rx_ep;
102 	struct rdma_cm_id *id = ep->re_id;
103 
104 	/* Flush Receives, then wait for deferred Reply work
105 	 * to complete.
106 	 */
107 	ib_drain_rq(id->qp);
108 
109 	/* Deferred Reply processing might have scheduled
110 	 * local invalidations.
111 	 */
112 	ib_drain_sq(id->qp);
113 
114 	rpcrdma_ep_put(ep);
115 }
116 
117 /**
118  * rpcrdma_qp_event_handler - Handle one QP event (error notification)
119  * @event: details of the event
120  * @context: ep that owns QP where event occurred
121  *
122  * Called from the RDMA provider (device driver) possibly in an interrupt
123  * context. The QP is always destroyed before the ID, so the ID will be
124  * reliably available when this handler is invoked.
125  */
126 static void rpcrdma_qp_event_handler(struct ib_event *event, void *context)
127 {
128 	struct rpcrdma_ep *ep = context;
129 
130 	trace_xprtrdma_qp_event(ep, event);
131 }
132 
133 /* Ensure xprt_force_disconnect() is invoked exactly once when a
134  * connection is closed or lost. (The important thing is it needs
135  * to be invoked "at least" once).
136  */
137 static void rpcrdma_force_disconnect(struct rpcrdma_ep *ep)
138 {
139 	if (atomic_add_unless(&ep->re_force_disconnect, 1, 1))
140 		xprt_force_disconnect(ep->re_xprt);
141 }
142 
143 /**
144  * rpcrdma_flush_disconnect - Disconnect on flushed completion
145  * @r_xprt: transport to disconnect
146  * @wc: work completion entry
147  *
148  * Must be called in process context.
149  */
150 void rpcrdma_flush_disconnect(struct rpcrdma_xprt *r_xprt, struct ib_wc *wc)
151 {
152 	if (wc->status != IB_WC_SUCCESS)
153 		rpcrdma_force_disconnect(r_xprt->rx_ep);
154 }
155 
156 /**
157  * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC
158  * @cq:	completion queue
159  * @wc:	WCE for a completed Send WR
160  *
161  */
162 static void rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
163 {
164 	struct ib_cqe *cqe = wc->wr_cqe;
165 	struct rpcrdma_sendctx *sc =
166 		container_of(cqe, struct rpcrdma_sendctx, sc_cqe);
167 	struct rpcrdma_xprt *r_xprt = cq->cq_context;
168 
169 	/* WARNING: Only wr_cqe and status are reliable at this point */
170 	trace_xprtrdma_wc_send(sc, wc);
171 	rpcrdma_sendctx_put_locked(r_xprt, sc);
172 	rpcrdma_flush_disconnect(r_xprt, wc);
173 }
174 
175 /**
176  * rpcrdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
177  * @cq:	completion queue
178  * @wc:	WCE for a completed Receive WR
179  *
180  */
181 static void rpcrdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
182 {
183 	struct ib_cqe *cqe = wc->wr_cqe;
184 	struct rpcrdma_rep *rep = container_of(cqe, struct rpcrdma_rep,
185 					       rr_cqe);
186 	struct rpcrdma_xprt *r_xprt = cq->cq_context;
187 
188 	/* WARNING: Only wr_cqe and status are reliable at this point */
189 	trace_xprtrdma_wc_receive(wc);
190 	--r_xprt->rx_ep->re_receive_count;
191 	if (wc->status != IB_WC_SUCCESS)
192 		goto out_flushed;
193 
194 	/* status == SUCCESS means all fields in wc are trustworthy */
195 	rpcrdma_set_xdrlen(&rep->rr_hdrbuf, wc->byte_len);
196 	rep->rr_wc_flags = wc->wc_flags;
197 	rep->rr_inv_rkey = wc->ex.invalidate_rkey;
198 
199 	ib_dma_sync_single_for_cpu(rdmab_device(rep->rr_rdmabuf),
200 				   rdmab_addr(rep->rr_rdmabuf),
201 				   wc->byte_len, DMA_FROM_DEVICE);
202 
203 	rpcrdma_reply_handler(rep);
204 	return;
205 
206 out_flushed:
207 	rpcrdma_flush_disconnect(r_xprt, wc);
208 	rpcrdma_rep_destroy(rep);
209 }
210 
211 static void rpcrdma_update_cm_private(struct rpcrdma_ep *ep,
212 				      struct rdma_conn_param *param)
213 {
214 	const struct rpcrdma_connect_private *pmsg = param->private_data;
215 	unsigned int rsize, wsize;
216 
217 	/* Default settings for RPC-over-RDMA Version One */
218 	ep->re_implicit_roundup = xprt_rdma_pad_optimize;
219 	rsize = RPCRDMA_V1_DEF_INLINE_SIZE;
220 	wsize = RPCRDMA_V1_DEF_INLINE_SIZE;
221 
222 	if (pmsg &&
223 	    pmsg->cp_magic == rpcrdma_cmp_magic &&
224 	    pmsg->cp_version == RPCRDMA_CMP_VERSION) {
225 		ep->re_implicit_roundup = true;
226 		rsize = rpcrdma_decode_buffer_size(pmsg->cp_send_size);
227 		wsize = rpcrdma_decode_buffer_size(pmsg->cp_recv_size);
228 	}
229 
230 	if (rsize < ep->re_inline_recv)
231 		ep->re_inline_recv = rsize;
232 	if (wsize < ep->re_inline_send)
233 		ep->re_inline_send = wsize;
234 
235 	rpcrdma_set_max_header_sizes(ep);
236 }
237 
238 /**
239  * rpcrdma_cm_event_handler - Handle RDMA CM events
240  * @id: rdma_cm_id on which an event has occurred
241  * @event: details of the event
242  *
243  * Called with @id's mutex held. Returns 1 if caller should
244  * destroy @id, otherwise 0.
245  */
246 static int
247 rpcrdma_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
248 {
249 	struct sockaddr *sap = (struct sockaddr *)&id->route.addr.dst_addr;
250 	struct rpcrdma_ep *ep = id->context;
251 
252 	might_sleep();
253 
254 	switch (event->event) {
255 	case RDMA_CM_EVENT_ADDR_RESOLVED:
256 	case RDMA_CM_EVENT_ROUTE_RESOLVED:
257 		ep->re_async_rc = 0;
258 		complete(&ep->re_done);
259 		return 0;
260 	case RDMA_CM_EVENT_ADDR_ERROR:
261 		ep->re_async_rc = -EPROTO;
262 		complete(&ep->re_done);
263 		return 0;
264 	case RDMA_CM_EVENT_ROUTE_ERROR:
265 		ep->re_async_rc = -ENETUNREACH;
266 		complete(&ep->re_done);
267 		return 0;
268 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
269 		pr_info("rpcrdma: removing device %s for %pISpc\n",
270 			ep->re_id->device->name, sap);
271 		fallthrough;
272 	case RDMA_CM_EVENT_ADDR_CHANGE:
273 		ep->re_connect_status = -ENODEV;
274 		goto disconnected;
275 	case RDMA_CM_EVENT_ESTABLISHED:
276 		rpcrdma_ep_get(ep);
277 		ep->re_connect_status = 1;
278 		rpcrdma_update_cm_private(ep, &event->param.conn);
279 		trace_xprtrdma_inline_thresh(ep);
280 		wake_up_all(&ep->re_connect_wait);
281 		break;
282 	case RDMA_CM_EVENT_CONNECT_ERROR:
283 		ep->re_connect_status = -ENOTCONN;
284 		goto wake_connect_worker;
285 	case RDMA_CM_EVENT_UNREACHABLE:
286 		ep->re_connect_status = -ENETUNREACH;
287 		goto wake_connect_worker;
288 	case RDMA_CM_EVENT_REJECTED:
289 		dprintk("rpcrdma: connection to %pISpc rejected: %s\n",
290 			sap, rdma_reject_msg(id, event->status));
291 		ep->re_connect_status = -ECONNREFUSED;
292 		if (event->status == IB_CM_REJ_STALE_CONN)
293 			ep->re_connect_status = -ENOTCONN;
294 wake_connect_worker:
295 		wake_up_all(&ep->re_connect_wait);
296 		return 0;
297 	case RDMA_CM_EVENT_DISCONNECTED:
298 		ep->re_connect_status = -ECONNABORTED;
299 disconnected:
300 		rpcrdma_force_disconnect(ep);
301 		return rpcrdma_ep_put(ep);
302 	default:
303 		break;
304 	}
305 
306 	dprintk("RPC:       %s: %pISpc on %s/frwr: %s\n", __func__, sap,
307 		ep->re_id->device->name, rdma_event_msg(event->event));
308 	return 0;
309 }
310 
311 static struct rdma_cm_id *rpcrdma_create_id(struct rpcrdma_xprt *r_xprt,
312 					    struct rpcrdma_ep *ep)
313 {
314 	unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1;
315 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
316 	struct rdma_cm_id *id;
317 	int rc;
318 
319 	init_completion(&ep->re_done);
320 
321 	id = rdma_create_id(xprt->xprt_net, rpcrdma_cm_event_handler, ep,
322 			    RDMA_PS_TCP, IB_QPT_RC);
323 	if (IS_ERR(id))
324 		return id;
325 
326 	ep->re_async_rc = -ETIMEDOUT;
327 	rc = rdma_resolve_addr(id, NULL, (struct sockaddr *)&xprt->addr,
328 			       RDMA_RESOLVE_TIMEOUT);
329 	if (rc)
330 		goto out;
331 	rc = wait_for_completion_interruptible_timeout(&ep->re_done, wtimeout);
332 	if (rc < 0)
333 		goto out;
334 
335 	rc = ep->re_async_rc;
336 	if (rc)
337 		goto out;
338 
339 	ep->re_async_rc = -ETIMEDOUT;
340 	rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
341 	if (rc)
342 		goto out;
343 	rc = wait_for_completion_interruptible_timeout(&ep->re_done, wtimeout);
344 	if (rc < 0)
345 		goto out;
346 	rc = ep->re_async_rc;
347 	if (rc)
348 		goto out;
349 
350 	return id;
351 
352 out:
353 	rdma_destroy_id(id);
354 	return ERR_PTR(rc);
355 }
356 
357 static void rpcrdma_ep_destroy(struct kref *kref)
358 {
359 	struct rpcrdma_ep *ep = container_of(kref, struct rpcrdma_ep, re_kref);
360 
361 	if (ep->re_id->qp) {
362 		rdma_destroy_qp(ep->re_id);
363 		ep->re_id->qp = NULL;
364 	}
365 
366 	if (ep->re_attr.recv_cq)
367 		ib_free_cq(ep->re_attr.recv_cq);
368 	ep->re_attr.recv_cq = NULL;
369 	if (ep->re_attr.send_cq)
370 		ib_free_cq(ep->re_attr.send_cq);
371 	ep->re_attr.send_cq = NULL;
372 
373 	if (ep->re_pd)
374 		ib_dealloc_pd(ep->re_pd);
375 	ep->re_pd = NULL;
376 
377 	kfree(ep);
378 	module_put(THIS_MODULE);
379 }
380 
381 static noinline void rpcrdma_ep_get(struct rpcrdma_ep *ep)
382 {
383 	kref_get(&ep->re_kref);
384 }
385 
386 /* Returns:
387  *     %0 if @ep still has a positive kref count, or
388  *     %1 if @ep was destroyed successfully.
389  */
390 static noinline int rpcrdma_ep_put(struct rpcrdma_ep *ep)
391 {
392 	return kref_put(&ep->re_kref, rpcrdma_ep_destroy);
393 }
394 
395 static int rpcrdma_ep_create(struct rpcrdma_xprt *r_xprt)
396 {
397 	struct rpcrdma_connect_private *pmsg;
398 	struct ib_device *device;
399 	struct rdma_cm_id *id;
400 	struct rpcrdma_ep *ep;
401 	int rc;
402 
403 	ep = kzalloc(sizeof(*ep), GFP_NOFS);
404 	if (!ep)
405 		return -ENOTCONN;
406 	ep->re_xprt = &r_xprt->rx_xprt;
407 	kref_init(&ep->re_kref);
408 
409 	id = rpcrdma_create_id(r_xprt, ep);
410 	if (IS_ERR(id)) {
411 		kfree(ep);
412 		return PTR_ERR(id);
413 	}
414 	__module_get(THIS_MODULE);
415 	device = id->device;
416 	ep->re_id = id;
417 
418 	ep->re_max_requests = r_xprt->rx_xprt.max_reqs;
419 	ep->re_inline_send = xprt_rdma_max_inline_write;
420 	ep->re_inline_recv = xprt_rdma_max_inline_read;
421 	rc = frwr_query_device(ep, device);
422 	if (rc)
423 		goto out_destroy;
424 
425 	r_xprt->rx_buf.rb_max_requests = cpu_to_be32(ep->re_max_requests);
426 
427 	ep->re_attr.event_handler = rpcrdma_qp_event_handler;
428 	ep->re_attr.qp_context = ep;
429 	ep->re_attr.srq = NULL;
430 	ep->re_attr.cap.max_inline_data = 0;
431 	ep->re_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
432 	ep->re_attr.qp_type = IB_QPT_RC;
433 	ep->re_attr.port_num = ~0;
434 
435 	dprintk("RPC:       %s: requested max: dtos: send %d recv %d; "
436 		"iovs: send %d recv %d\n",
437 		__func__,
438 		ep->re_attr.cap.max_send_wr,
439 		ep->re_attr.cap.max_recv_wr,
440 		ep->re_attr.cap.max_send_sge,
441 		ep->re_attr.cap.max_recv_sge);
442 
443 	ep->re_send_batch = ep->re_max_requests >> 3;
444 	ep->re_send_count = ep->re_send_batch;
445 	init_waitqueue_head(&ep->re_connect_wait);
446 
447 	ep->re_attr.send_cq = ib_alloc_cq_any(device, r_xprt,
448 					      ep->re_attr.cap.max_send_wr,
449 					      IB_POLL_WORKQUEUE);
450 	if (IS_ERR(ep->re_attr.send_cq)) {
451 		rc = PTR_ERR(ep->re_attr.send_cq);
452 		goto out_destroy;
453 	}
454 
455 	ep->re_attr.recv_cq = ib_alloc_cq_any(device, r_xprt,
456 					      ep->re_attr.cap.max_recv_wr,
457 					      IB_POLL_WORKQUEUE);
458 	if (IS_ERR(ep->re_attr.recv_cq)) {
459 		rc = PTR_ERR(ep->re_attr.recv_cq);
460 		goto out_destroy;
461 	}
462 	ep->re_receive_count = 0;
463 
464 	/* Initialize cma parameters */
465 	memset(&ep->re_remote_cma, 0, sizeof(ep->re_remote_cma));
466 
467 	/* Prepare RDMA-CM private message */
468 	pmsg = &ep->re_cm_private;
469 	pmsg->cp_magic = rpcrdma_cmp_magic;
470 	pmsg->cp_version = RPCRDMA_CMP_VERSION;
471 	pmsg->cp_flags |= RPCRDMA_CMP_F_SND_W_INV_OK;
472 	pmsg->cp_send_size = rpcrdma_encode_buffer_size(ep->re_inline_send);
473 	pmsg->cp_recv_size = rpcrdma_encode_buffer_size(ep->re_inline_recv);
474 	ep->re_remote_cma.private_data = pmsg;
475 	ep->re_remote_cma.private_data_len = sizeof(*pmsg);
476 
477 	/* Client offers RDMA Read but does not initiate */
478 	ep->re_remote_cma.initiator_depth = 0;
479 	ep->re_remote_cma.responder_resources =
480 		min_t(int, U8_MAX, device->attrs.max_qp_rd_atom);
481 
482 	/* Limit transport retries so client can detect server
483 	 * GID changes quickly. RPC layer handles re-establishing
484 	 * transport connection and retransmission.
485 	 */
486 	ep->re_remote_cma.retry_count = 6;
487 
488 	/* RPC-over-RDMA handles its own flow control. In addition,
489 	 * make all RNR NAKs visible so we know that RPC-over-RDMA
490 	 * flow control is working correctly (no NAKs should be seen).
491 	 */
492 	ep->re_remote_cma.flow_control = 0;
493 	ep->re_remote_cma.rnr_retry_count = 0;
494 
495 	ep->re_pd = ib_alloc_pd(device, 0);
496 	if (IS_ERR(ep->re_pd)) {
497 		rc = PTR_ERR(ep->re_pd);
498 		goto out_destroy;
499 	}
500 
501 	rc = rdma_create_qp(id, ep->re_pd, &ep->re_attr);
502 	if (rc)
503 		goto out_destroy;
504 
505 	r_xprt->rx_ep = ep;
506 	return 0;
507 
508 out_destroy:
509 	rpcrdma_ep_put(ep);
510 	rdma_destroy_id(id);
511 	return rc;
512 }
513 
514 /**
515  * rpcrdma_xprt_connect - Connect an unconnected transport
516  * @r_xprt: controlling transport instance
517  *
518  * Returns 0 on success or a negative errno.
519  */
520 int rpcrdma_xprt_connect(struct rpcrdma_xprt *r_xprt)
521 {
522 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
523 	struct rpcrdma_ep *ep;
524 	int rc;
525 
526 	rc = rpcrdma_ep_create(r_xprt);
527 	if (rc)
528 		return rc;
529 	ep = r_xprt->rx_ep;
530 
531 	xprt_clear_connected(xprt);
532 	rpcrdma_reset_cwnd(r_xprt);
533 
534 	/* Bump the ep's reference count while there are
535 	 * outstanding Receives.
536 	 */
537 	rpcrdma_ep_get(ep);
538 	rpcrdma_post_recvs(r_xprt, true);
539 
540 	rc = rdma_connect(ep->re_id, &ep->re_remote_cma);
541 	if (rc)
542 		goto out;
543 
544 	if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
545 		xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
546 	wait_event_interruptible(ep->re_connect_wait,
547 				 ep->re_connect_status != 0);
548 	if (ep->re_connect_status <= 0) {
549 		rc = ep->re_connect_status;
550 		goto out;
551 	}
552 
553 	rc = rpcrdma_sendctxs_create(r_xprt);
554 	if (rc) {
555 		rc = -ENOTCONN;
556 		goto out;
557 	}
558 
559 	rc = rpcrdma_reqs_setup(r_xprt);
560 	if (rc) {
561 		rc = -ENOTCONN;
562 		goto out;
563 	}
564 	rpcrdma_mrs_create(r_xprt);
565 
566 out:
567 	trace_xprtrdma_connect(r_xprt, rc);
568 	return rc;
569 }
570 
571 /**
572  * rpcrdma_xprt_disconnect - Disconnect underlying transport
573  * @r_xprt: controlling transport instance
574  *
575  * Caller serializes. Either the transport send lock is held,
576  * or we're being called to destroy the transport.
577  *
578  * On return, @r_xprt is completely divested of all hardware
579  * resources and prepared for the next ->connect operation.
580  */
581 void rpcrdma_xprt_disconnect(struct rpcrdma_xprt *r_xprt)
582 {
583 	struct rpcrdma_ep *ep = r_xprt->rx_ep;
584 	struct rdma_cm_id *id;
585 	int rc;
586 
587 	if (!ep)
588 		return;
589 
590 	id = ep->re_id;
591 	rc = rdma_disconnect(id);
592 	trace_xprtrdma_disconnect(r_xprt, rc);
593 
594 	rpcrdma_xprt_drain(r_xprt);
595 	rpcrdma_reps_unmap(r_xprt);
596 	rpcrdma_reqs_reset(r_xprt);
597 	rpcrdma_mrs_destroy(r_xprt);
598 	rpcrdma_sendctxs_destroy(r_xprt);
599 
600 	if (rpcrdma_ep_put(ep))
601 		rdma_destroy_id(id);
602 
603 	r_xprt->rx_ep = NULL;
604 }
605 
606 /* Fixed-size circular FIFO queue. This implementation is wait-free and
607  * lock-free.
608  *
609  * Consumer is the code path that posts Sends. This path dequeues a
610  * sendctx for use by a Send operation. Multiple consumer threads
611  * are serialized by the RPC transport lock, which allows only one
612  * ->send_request call at a time.
613  *
614  * Producer is the code path that handles Send completions. This path
615  * enqueues a sendctx that has been completed. Multiple producer
616  * threads are serialized by the ib_poll_cq() function.
617  */
618 
619 /* rpcrdma_sendctxs_destroy() assumes caller has already quiesced
620  * queue activity, and rpcrdma_xprt_drain has flushed all remaining
621  * Send requests.
622  */
623 static void rpcrdma_sendctxs_destroy(struct rpcrdma_xprt *r_xprt)
624 {
625 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
626 	unsigned long i;
627 
628 	if (!buf->rb_sc_ctxs)
629 		return;
630 	for (i = 0; i <= buf->rb_sc_last; i++)
631 		kfree(buf->rb_sc_ctxs[i]);
632 	kfree(buf->rb_sc_ctxs);
633 	buf->rb_sc_ctxs = NULL;
634 }
635 
636 static struct rpcrdma_sendctx *rpcrdma_sendctx_create(struct rpcrdma_ep *ep)
637 {
638 	struct rpcrdma_sendctx *sc;
639 
640 	sc = kzalloc(struct_size(sc, sc_sges, ep->re_attr.cap.max_send_sge),
641 		     GFP_KERNEL);
642 	if (!sc)
643 		return NULL;
644 
645 	sc->sc_cqe.done = rpcrdma_wc_send;
646 	return sc;
647 }
648 
649 static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt)
650 {
651 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
652 	struct rpcrdma_sendctx *sc;
653 	unsigned long i;
654 
655 	/* Maximum number of concurrent outstanding Send WRs. Capping
656 	 * the circular queue size stops Send Queue overflow by causing
657 	 * the ->send_request call to fail temporarily before too many
658 	 * Sends are posted.
659 	 */
660 	i = r_xprt->rx_ep->re_max_requests + RPCRDMA_MAX_BC_REQUESTS;
661 	buf->rb_sc_ctxs = kcalloc(i, sizeof(sc), GFP_KERNEL);
662 	if (!buf->rb_sc_ctxs)
663 		return -ENOMEM;
664 
665 	buf->rb_sc_last = i - 1;
666 	for (i = 0; i <= buf->rb_sc_last; i++) {
667 		sc = rpcrdma_sendctx_create(r_xprt->rx_ep);
668 		if (!sc)
669 			return -ENOMEM;
670 
671 		buf->rb_sc_ctxs[i] = sc;
672 	}
673 
674 	buf->rb_sc_head = 0;
675 	buf->rb_sc_tail = 0;
676 	return 0;
677 }
678 
679 /* The sendctx queue is not guaranteed to have a size that is a
680  * power of two, thus the helpers in circ_buf.h cannot be used.
681  * The other option is to use modulus (%), which can be expensive.
682  */
683 static unsigned long rpcrdma_sendctx_next(struct rpcrdma_buffer *buf,
684 					  unsigned long item)
685 {
686 	return likely(item < buf->rb_sc_last) ? item + 1 : 0;
687 }
688 
689 /**
690  * rpcrdma_sendctx_get_locked - Acquire a send context
691  * @r_xprt: controlling transport instance
692  *
693  * Returns pointer to a free send completion context; or NULL if
694  * the queue is empty.
695  *
696  * Usage: Called to acquire an SGE array before preparing a Send WR.
697  *
698  * The caller serializes calls to this function (per transport), and
699  * provides an effective memory barrier that flushes the new value
700  * of rb_sc_head.
701  */
702 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_xprt *r_xprt)
703 {
704 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
705 	struct rpcrdma_sendctx *sc;
706 	unsigned long next_head;
707 
708 	next_head = rpcrdma_sendctx_next(buf, buf->rb_sc_head);
709 
710 	if (next_head == READ_ONCE(buf->rb_sc_tail))
711 		goto out_emptyq;
712 
713 	/* ORDER: item must be accessed _before_ head is updated */
714 	sc = buf->rb_sc_ctxs[next_head];
715 
716 	/* Releasing the lock in the caller acts as a memory
717 	 * barrier that flushes rb_sc_head.
718 	 */
719 	buf->rb_sc_head = next_head;
720 
721 	return sc;
722 
723 out_emptyq:
724 	/* The queue is "empty" if there have not been enough Send
725 	 * completions recently. This is a sign the Send Queue is
726 	 * backing up. Cause the caller to pause and try again.
727 	 */
728 	xprt_wait_for_buffer_space(&r_xprt->rx_xprt);
729 	r_xprt->rx_stats.empty_sendctx_q++;
730 	return NULL;
731 }
732 
733 /**
734  * rpcrdma_sendctx_put_locked - Release a send context
735  * @r_xprt: controlling transport instance
736  * @sc: send context to release
737  *
738  * Usage: Called from Send completion to return a sendctxt
739  * to the queue.
740  *
741  * The caller serializes calls to this function (per transport).
742  */
743 static void rpcrdma_sendctx_put_locked(struct rpcrdma_xprt *r_xprt,
744 				       struct rpcrdma_sendctx *sc)
745 {
746 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
747 	unsigned long next_tail;
748 
749 	/* Unmap SGEs of previously completed but unsignaled
750 	 * Sends by walking up the queue until @sc is found.
751 	 */
752 	next_tail = buf->rb_sc_tail;
753 	do {
754 		next_tail = rpcrdma_sendctx_next(buf, next_tail);
755 
756 		/* ORDER: item must be accessed _before_ tail is updated */
757 		rpcrdma_sendctx_unmap(buf->rb_sc_ctxs[next_tail]);
758 
759 	} while (buf->rb_sc_ctxs[next_tail] != sc);
760 
761 	/* Paired with READ_ONCE */
762 	smp_store_release(&buf->rb_sc_tail, next_tail);
763 
764 	xprt_write_space(&r_xprt->rx_xprt);
765 }
766 
767 static void
768 rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt)
769 {
770 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
771 	struct rpcrdma_ep *ep = r_xprt->rx_ep;
772 	unsigned int count;
773 
774 	for (count = 0; count < ep->re_max_rdma_segs; count++) {
775 		struct rpcrdma_mr *mr;
776 		int rc;
777 
778 		mr = kzalloc(sizeof(*mr), GFP_NOFS);
779 		if (!mr)
780 			break;
781 
782 		rc = frwr_mr_init(r_xprt, mr);
783 		if (rc) {
784 			kfree(mr);
785 			break;
786 		}
787 
788 		spin_lock(&buf->rb_lock);
789 		rpcrdma_mr_push(mr, &buf->rb_mrs);
790 		list_add(&mr->mr_all, &buf->rb_all_mrs);
791 		spin_unlock(&buf->rb_lock);
792 	}
793 
794 	r_xprt->rx_stats.mrs_allocated += count;
795 	trace_xprtrdma_createmrs(r_xprt, count);
796 }
797 
798 static void
799 rpcrdma_mr_refresh_worker(struct work_struct *work)
800 {
801 	struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
802 						  rb_refresh_worker);
803 	struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
804 						   rx_buf);
805 
806 	rpcrdma_mrs_create(r_xprt);
807 	xprt_write_space(&r_xprt->rx_xprt);
808 }
809 
810 /**
811  * rpcrdma_mrs_refresh - Wake the MR refresh worker
812  * @r_xprt: controlling transport instance
813  *
814  */
815 void rpcrdma_mrs_refresh(struct rpcrdma_xprt *r_xprt)
816 {
817 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
818 	struct rpcrdma_ep *ep = r_xprt->rx_ep;
819 
820 	/* If there is no underlying connection, it's no use
821 	 * to wake the refresh worker.
822 	 */
823 	if (ep->re_connect_status == 1) {
824 		/* The work is scheduled on a WQ_MEM_RECLAIM
825 		 * workqueue in order to prevent MR allocation
826 		 * from recursing into NFS during direct reclaim.
827 		 */
828 		queue_work(xprtiod_workqueue, &buf->rb_refresh_worker);
829 	}
830 }
831 
832 /**
833  * rpcrdma_req_create - Allocate an rpcrdma_req object
834  * @r_xprt: controlling r_xprt
835  * @size: initial size, in bytes, of send and receive buffers
836  * @flags: GFP flags passed to memory allocators
837  *
838  * Returns an allocated and fully initialized rpcrdma_req or NULL.
839  */
840 struct rpcrdma_req *rpcrdma_req_create(struct rpcrdma_xprt *r_xprt, size_t size,
841 				       gfp_t flags)
842 {
843 	struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
844 	struct rpcrdma_req *req;
845 
846 	req = kzalloc(sizeof(*req), flags);
847 	if (req == NULL)
848 		goto out1;
849 
850 	req->rl_sendbuf = rpcrdma_regbuf_alloc(size, DMA_TO_DEVICE, flags);
851 	if (!req->rl_sendbuf)
852 		goto out2;
853 
854 	req->rl_recvbuf = rpcrdma_regbuf_alloc(size, DMA_NONE, flags);
855 	if (!req->rl_recvbuf)
856 		goto out3;
857 
858 	INIT_LIST_HEAD(&req->rl_free_mrs);
859 	INIT_LIST_HEAD(&req->rl_registered);
860 	spin_lock(&buffer->rb_lock);
861 	list_add(&req->rl_all, &buffer->rb_allreqs);
862 	spin_unlock(&buffer->rb_lock);
863 	return req;
864 
865 out3:
866 	kfree(req->rl_sendbuf);
867 out2:
868 	kfree(req);
869 out1:
870 	return NULL;
871 }
872 
873 /**
874  * rpcrdma_req_setup - Per-connection instance setup of an rpcrdma_req object
875  * @r_xprt: controlling transport instance
876  * @req: rpcrdma_req object to set up
877  *
878  * Returns zero on success, and a negative errno on failure.
879  */
880 int rpcrdma_req_setup(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
881 {
882 	struct rpcrdma_regbuf *rb;
883 	size_t maxhdrsize;
884 
885 	/* Compute maximum header buffer size in bytes */
886 	maxhdrsize = rpcrdma_fixed_maxsz + 3 +
887 		     r_xprt->rx_ep->re_max_rdma_segs * rpcrdma_readchunk_maxsz;
888 	maxhdrsize *= sizeof(__be32);
889 	rb = rpcrdma_regbuf_alloc(__roundup_pow_of_two(maxhdrsize),
890 				  DMA_TO_DEVICE, GFP_KERNEL);
891 	if (!rb)
892 		goto out;
893 
894 	if (!__rpcrdma_regbuf_dma_map(r_xprt, rb))
895 		goto out_free;
896 
897 	req->rl_rdmabuf = rb;
898 	xdr_buf_init(&req->rl_hdrbuf, rdmab_data(rb), rdmab_length(rb));
899 	return 0;
900 
901 out_free:
902 	rpcrdma_regbuf_free(rb);
903 out:
904 	return -ENOMEM;
905 }
906 
907 /* ASSUMPTION: the rb_allreqs list is stable for the duration,
908  * and thus can be walked without holding rb_lock. Eg. the
909  * caller is holding the transport send lock to exclude
910  * device removal or disconnection.
911  */
912 static int rpcrdma_reqs_setup(struct rpcrdma_xprt *r_xprt)
913 {
914 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
915 	struct rpcrdma_req *req;
916 	int rc;
917 
918 	list_for_each_entry(req, &buf->rb_allreqs, rl_all) {
919 		rc = rpcrdma_req_setup(r_xprt, req);
920 		if (rc)
921 			return rc;
922 	}
923 	return 0;
924 }
925 
926 static void rpcrdma_req_reset(struct rpcrdma_req *req)
927 {
928 	/* Credits are valid for only one connection */
929 	req->rl_slot.rq_cong = 0;
930 
931 	rpcrdma_regbuf_free(req->rl_rdmabuf);
932 	req->rl_rdmabuf = NULL;
933 
934 	rpcrdma_regbuf_dma_unmap(req->rl_sendbuf);
935 	rpcrdma_regbuf_dma_unmap(req->rl_recvbuf);
936 
937 	frwr_reset(req);
938 }
939 
940 /* ASSUMPTION: the rb_allreqs list is stable for the duration,
941  * and thus can be walked without holding rb_lock. Eg. the
942  * caller is holding the transport send lock to exclude
943  * device removal or disconnection.
944  */
945 static void rpcrdma_reqs_reset(struct rpcrdma_xprt *r_xprt)
946 {
947 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
948 	struct rpcrdma_req *req;
949 
950 	list_for_each_entry(req, &buf->rb_allreqs, rl_all)
951 		rpcrdma_req_reset(req);
952 }
953 
954 /* No locking needed here. This function is called only by the
955  * Receive completion handler.
956  */
957 static noinline
958 struct rpcrdma_rep *rpcrdma_rep_create(struct rpcrdma_xprt *r_xprt,
959 				       bool temp)
960 {
961 	struct rpcrdma_rep *rep;
962 
963 	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
964 	if (rep == NULL)
965 		goto out;
966 
967 	rep->rr_rdmabuf = rpcrdma_regbuf_alloc(r_xprt->rx_ep->re_inline_recv,
968 					       DMA_FROM_DEVICE, GFP_KERNEL);
969 	if (!rep->rr_rdmabuf)
970 		goto out_free;
971 
972 	if (!rpcrdma_regbuf_dma_map(r_xprt, rep->rr_rdmabuf))
973 		goto out_free_regbuf;
974 
975 	xdr_buf_init(&rep->rr_hdrbuf, rdmab_data(rep->rr_rdmabuf),
976 		     rdmab_length(rep->rr_rdmabuf));
977 	rep->rr_cqe.done = rpcrdma_wc_receive;
978 	rep->rr_rxprt = r_xprt;
979 	rep->rr_recv_wr.next = NULL;
980 	rep->rr_recv_wr.wr_cqe = &rep->rr_cqe;
981 	rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
982 	rep->rr_recv_wr.num_sge = 1;
983 	rep->rr_temp = temp;
984 	list_add(&rep->rr_all, &r_xprt->rx_buf.rb_all_reps);
985 	return rep;
986 
987 out_free_regbuf:
988 	rpcrdma_regbuf_free(rep->rr_rdmabuf);
989 out_free:
990 	kfree(rep);
991 out:
992 	return NULL;
993 }
994 
995 /* No locking needed here. This function is invoked only by the
996  * Receive completion handler, or during transport shutdown.
997  */
998 static void rpcrdma_rep_destroy(struct rpcrdma_rep *rep)
999 {
1000 	list_del(&rep->rr_all);
1001 	rpcrdma_regbuf_free(rep->rr_rdmabuf);
1002 	kfree(rep);
1003 }
1004 
1005 static struct rpcrdma_rep *rpcrdma_rep_get_locked(struct rpcrdma_buffer *buf)
1006 {
1007 	struct llist_node *node;
1008 
1009 	/* Calls to llist_del_first are required to be serialized */
1010 	node = llist_del_first(&buf->rb_free_reps);
1011 	if (!node)
1012 		return NULL;
1013 	return llist_entry(node, struct rpcrdma_rep, rr_node);
1014 }
1015 
1016 static void rpcrdma_rep_put(struct rpcrdma_buffer *buf,
1017 			    struct rpcrdma_rep *rep)
1018 {
1019 	llist_add(&rep->rr_node, &buf->rb_free_reps);
1020 }
1021 
1022 static void rpcrdma_reps_unmap(struct rpcrdma_xprt *r_xprt)
1023 {
1024 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1025 	struct rpcrdma_rep *rep;
1026 
1027 	list_for_each_entry(rep, &buf->rb_all_reps, rr_all) {
1028 		rpcrdma_regbuf_dma_unmap(rep->rr_rdmabuf);
1029 		rep->rr_temp = true;
1030 	}
1031 }
1032 
1033 static void rpcrdma_reps_destroy(struct rpcrdma_buffer *buf)
1034 {
1035 	struct rpcrdma_rep *rep;
1036 
1037 	while ((rep = rpcrdma_rep_get_locked(buf)) != NULL)
1038 		rpcrdma_rep_destroy(rep);
1039 }
1040 
1041 /**
1042  * rpcrdma_buffer_create - Create initial set of req/rep objects
1043  * @r_xprt: transport instance to (re)initialize
1044  *
1045  * Returns zero on success, otherwise a negative errno.
1046  */
1047 int rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
1048 {
1049 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1050 	int i, rc;
1051 
1052 	buf->rb_bc_srv_max_requests = 0;
1053 	spin_lock_init(&buf->rb_lock);
1054 	INIT_LIST_HEAD(&buf->rb_mrs);
1055 	INIT_LIST_HEAD(&buf->rb_all_mrs);
1056 	INIT_WORK(&buf->rb_refresh_worker, rpcrdma_mr_refresh_worker);
1057 
1058 	INIT_LIST_HEAD(&buf->rb_send_bufs);
1059 	INIT_LIST_HEAD(&buf->rb_allreqs);
1060 	INIT_LIST_HEAD(&buf->rb_all_reps);
1061 
1062 	rc = -ENOMEM;
1063 	for (i = 0; i < r_xprt->rx_xprt.max_reqs; i++) {
1064 		struct rpcrdma_req *req;
1065 
1066 		req = rpcrdma_req_create(r_xprt, RPCRDMA_V1_DEF_INLINE_SIZE * 2,
1067 					 GFP_KERNEL);
1068 		if (!req)
1069 			goto out;
1070 		list_add(&req->rl_list, &buf->rb_send_bufs);
1071 	}
1072 
1073 	init_llist_head(&buf->rb_free_reps);
1074 
1075 	return 0;
1076 out:
1077 	rpcrdma_buffer_destroy(buf);
1078 	return rc;
1079 }
1080 
1081 /**
1082  * rpcrdma_req_destroy - Destroy an rpcrdma_req object
1083  * @req: unused object to be destroyed
1084  *
1085  * Relies on caller holding the transport send lock to protect
1086  * removing req->rl_all from buf->rb_all_reqs safely.
1087  */
1088 void rpcrdma_req_destroy(struct rpcrdma_req *req)
1089 {
1090 	struct rpcrdma_mr *mr;
1091 
1092 	list_del(&req->rl_all);
1093 
1094 	while ((mr = rpcrdma_mr_pop(&req->rl_free_mrs))) {
1095 		struct rpcrdma_buffer *buf = &mr->mr_xprt->rx_buf;
1096 
1097 		spin_lock(&buf->rb_lock);
1098 		list_del(&mr->mr_all);
1099 		spin_unlock(&buf->rb_lock);
1100 
1101 		frwr_release_mr(mr);
1102 	}
1103 
1104 	rpcrdma_regbuf_free(req->rl_recvbuf);
1105 	rpcrdma_regbuf_free(req->rl_sendbuf);
1106 	rpcrdma_regbuf_free(req->rl_rdmabuf);
1107 	kfree(req);
1108 }
1109 
1110 /**
1111  * rpcrdma_mrs_destroy - Release all of a transport's MRs
1112  * @r_xprt: controlling transport instance
1113  *
1114  * Relies on caller holding the transport send lock to protect
1115  * removing mr->mr_list from req->rl_free_mrs safely.
1116  */
1117 static void rpcrdma_mrs_destroy(struct rpcrdma_xprt *r_xprt)
1118 {
1119 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1120 	struct rpcrdma_mr *mr;
1121 
1122 	cancel_work_sync(&buf->rb_refresh_worker);
1123 
1124 	spin_lock(&buf->rb_lock);
1125 	while ((mr = list_first_entry_or_null(&buf->rb_all_mrs,
1126 					      struct rpcrdma_mr,
1127 					      mr_all)) != NULL) {
1128 		list_del(&mr->mr_list);
1129 		list_del(&mr->mr_all);
1130 		spin_unlock(&buf->rb_lock);
1131 
1132 		frwr_release_mr(mr);
1133 
1134 		spin_lock(&buf->rb_lock);
1135 	}
1136 	spin_unlock(&buf->rb_lock);
1137 }
1138 
1139 /**
1140  * rpcrdma_buffer_destroy - Release all hw resources
1141  * @buf: root control block for resources
1142  *
1143  * ORDERING: relies on a prior rpcrdma_xprt_drain :
1144  * - No more Send or Receive completions can occur
1145  * - All MRs, reps, and reqs are returned to their free lists
1146  */
1147 void
1148 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1149 {
1150 	rpcrdma_reps_destroy(buf);
1151 
1152 	while (!list_empty(&buf->rb_send_bufs)) {
1153 		struct rpcrdma_req *req;
1154 
1155 		req = list_first_entry(&buf->rb_send_bufs,
1156 				       struct rpcrdma_req, rl_list);
1157 		list_del(&req->rl_list);
1158 		rpcrdma_req_destroy(req);
1159 	}
1160 }
1161 
1162 /**
1163  * rpcrdma_mr_get - Allocate an rpcrdma_mr object
1164  * @r_xprt: controlling transport
1165  *
1166  * Returns an initialized rpcrdma_mr or NULL if no free
1167  * rpcrdma_mr objects are available.
1168  */
1169 struct rpcrdma_mr *
1170 rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt)
1171 {
1172 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1173 	struct rpcrdma_mr *mr;
1174 
1175 	spin_lock(&buf->rb_lock);
1176 	mr = rpcrdma_mr_pop(&buf->rb_mrs);
1177 	spin_unlock(&buf->rb_lock);
1178 	return mr;
1179 }
1180 
1181 /**
1182  * rpcrdma_mr_put - DMA unmap an MR and release it
1183  * @mr: MR to release
1184  *
1185  */
1186 void rpcrdma_mr_put(struct rpcrdma_mr *mr)
1187 {
1188 	struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
1189 
1190 	if (mr->mr_dir != DMA_NONE) {
1191 		trace_xprtrdma_mr_unmap(mr);
1192 		ib_dma_unmap_sg(r_xprt->rx_ep->re_id->device,
1193 				mr->mr_sg, mr->mr_nents, mr->mr_dir);
1194 		mr->mr_dir = DMA_NONE;
1195 	}
1196 
1197 	rpcrdma_mr_push(mr, &mr->mr_req->rl_free_mrs);
1198 }
1199 
1200 /**
1201  * rpcrdma_buffer_get - Get a request buffer
1202  * @buffers: Buffer pool from which to obtain a buffer
1203  *
1204  * Returns a fresh rpcrdma_req, or NULL if none are available.
1205  */
1206 struct rpcrdma_req *
1207 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1208 {
1209 	struct rpcrdma_req *req;
1210 
1211 	spin_lock(&buffers->rb_lock);
1212 	req = list_first_entry_or_null(&buffers->rb_send_bufs,
1213 				       struct rpcrdma_req, rl_list);
1214 	if (req)
1215 		list_del_init(&req->rl_list);
1216 	spin_unlock(&buffers->rb_lock);
1217 	return req;
1218 }
1219 
1220 /**
1221  * rpcrdma_buffer_put - Put request/reply buffers back into pool
1222  * @buffers: buffer pool
1223  * @req: object to return
1224  *
1225  */
1226 void rpcrdma_buffer_put(struct rpcrdma_buffer *buffers, struct rpcrdma_req *req)
1227 {
1228 	if (req->rl_reply)
1229 		rpcrdma_rep_put(buffers, req->rl_reply);
1230 	req->rl_reply = NULL;
1231 
1232 	spin_lock(&buffers->rb_lock);
1233 	list_add(&req->rl_list, &buffers->rb_send_bufs);
1234 	spin_unlock(&buffers->rb_lock);
1235 }
1236 
1237 /**
1238  * rpcrdma_recv_buffer_put - Release rpcrdma_rep back to free list
1239  * @rep: rep to release
1240  *
1241  * Used after error conditions.
1242  */
1243 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1244 {
1245 	rpcrdma_rep_put(&rep->rr_rxprt->rx_buf, rep);
1246 }
1247 
1248 /* Returns a pointer to a rpcrdma_regbuf object, or NULL.
1249  *
1250  * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1251  * receiving the payload of RDMA RECV operations. During Long Calls
1252  * or Replies they may be registered externally via frwr_map.
1253  */
1254 static struct rpcrdma_regbuf *
1255 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction,
1256 		     gfp_t flags)
1257 {
1258 	struct rpcrdma_regbuf *rb;
1259 
1260 	rb = kmalloc(sizeof(*rb), flags);
1261 	if (!rb)
1262 		return NULL;
1263 	rb->rg_data = kmalloc(size, flags);
1264 	if (!rb->rg_data) {
1265 		kfree(rb);
1266 		return NULL;
1267 	}
1268 
1269 	rb->rg_device = NULL;
1270 	rb->rg_direction = direction;
1271 	rb->rg_iov.length = size;
1272 	return rb;
1273 }
1274 
1275 /**
1276  * rpcrdma_regbuf_realloc - re-allocate a SEND/RECV buffer
1277  * @rb: regbuf to reallocate
1278  * @size: size of buffer to be allocated, in bytes
1279  * @flags: GFP flags
1280  *
1281  * Returns true if reallocation was successful. If false is
1282  * returned, @rb is left untouched.
1283  */
1284 bool rpcrdma_regbuf_realloc(struct rpcrdma_regbuf *rb, size_t size, gfp_t flags)
1285 {
1286 	void *buf;
1287 
1288 	buf = kmalloc(size, flags);
1289 	if (!buf)
1290 		return false;
1291 
1292 	rpcrdma_regbuf_dma_unmap(rb);
1293 	kfree(rb->rg_data);
1294 
1295 	rb->rg_data = buf;
1296 	rb->rg_iov.length = size;
1297 	return true;
1298 }
1299 
1300 /**
1301  * __rpcrdma_regbuf_dma_map - DMA-map a regbuf
1302  * @r_xprt: controlling transport instance
1303  * @rb: regbuf to be mapped
1304  *
1305  * Returns true if the buffer is now DMA mapped to @r_xprt's device
1306  */
1307 bool __rpcrdma_regbuf_dma_map(struct rpcrdma_xprt *r_xprt,
1308 			      struct rpcrdma_regbuf *rb)
1309 {
1310 	struct ib_device *device = r_xprt->rx_ep->re_id->device;
1311 
1312 	if (rb->rg_direction == DMA_NONE)
1313 		return false;
1314 
1315 	rb->rg_iov.addr = ib_dma_map_single(device, rdmab_data(rb),
1316 					    rdmab_length(rb), rb->rg_direction);
1317 	if (ib_dma_mapping_error(device, rdmab_addr(rb))) {
1318 		trace_xprtrdma_dma_maperr(rdmab_addr(rb));
1319 		return false;
1320 	}
1321 
1322 	rb->rg_device = device;
1323 	rb->rg_iov.lkey = r_xprt->rx_ep->re_pd->local_dma_lkey;
1324 	return true;
1325 }
1326 
1327 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb)
1328 {
1329 	if (!rb)
1330 		return;
1331 
1332 	if (!rpcrdma_regbuf_is_mapped(rb))
1333 		return;
1334 
1335 	ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb), rdmab_length(rb),
1336 			    rb->rg_direction);
1337 	rb->rg_device = NULL;
1338 }
1339 
1340 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb)
1341 {
1342 	rpcrdma_regbuf_dma_unmap(rb);
1343 	if (rb)
1344 		kfree(rb->rg_data);
1345 	kfree(rb);
1346 }
1347 
1348 /**
1349  * rpcrdma_post_sends - Post WRs to a transport's Send Queue
1350  * @r_xprt: controlling transport instance
1351  * @req: rpcrdma_req containing the Send WR to post
1352  *
1353  * Returns 0 if the post was successful, otherwise -ENOTCONN
1354  * is returned.
1355  */
1356 int rpcrdma_post_sends(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
1357 {
1358 	struct ib_send_wr *send_wr = &req->rl_wr;
1359 	struct rpcrdma_ep *ep = r_xprt->rx_ep;
1360 	int rc;
1361 
1362 	if (!ep->re_send_count || kref_read(&req->rl_kref) > 1) {
1363 		send_wr->send_flags |= IB_SEND_SIGNALED;
1364 		ep->re_send_count = ep->re_send_batch;
1365 	} else {
1366 		send_wr->send_flags &= ~IB_SEND_SIGNALED;
1367 		--ep->re_send_count;
1368 	}
1369 
1370 	trace_xprtrdma_post_send(req);
1371 	rc = frwr_send(r_xprt, req);
1372 	if (rc)
1373 		return -ENOTCONN;
1374 	return 0;
1375 }
1376 
1377 /**
1378  * rpcrdma_post_recvs - Refill the Receive Queue
1379  * @r_xprt: controlling transport instance
1380  * @temp: mark Receive buffers to be deleted after use
1381  *
1382  */
1383 void rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, bool temp)
1384 {
1385 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1386 	struct rpcrdma_ep *ep = r_xprt->rx_ep;
1387 	struct ib_recv_wr *wr, *bad_wr;
1388 	struct rpcrdma_rep *rep;
1389 	int needed, count, rc;
1390 
1391 	rc = 0;
1392 	count = 0;
1393 
1394 	needed = buf->rb_credits + (buf->rb_bc_srv_max_requests << 1);
1395 	if (likely(ep->re_receive_count > needed))
1396 		goto out;
1397 	needed -= ep->re_receive_count;
1398 	if (!temp)
1399 		needed += RPCRDMA_MAX_RECV_BATCH;
1400 
1401 	/* fast path: all needed reps can be found on the free list */
1402 	wr = NULL;
1403 	while (needed) {
1404 		rep = rpcrdma_rep_get_locked(buf);
1405 		if (rep && rep->rr_temp) {
1406 			rpcrdma_rep_destroy(rep);
1407 			continue;
1408 		}
1409 		if (!rep)
1410 			rep = rpcrdma_rep_create(r_xprt, temp);
1411 		if (!rep)
1412 			break;
1413 
1414 		trace_xprtrdma_post_recv(rep);
1415 		rep->rr_recv_wr.next = wr;
1416 		wr = &rep->rr_recv_wr;
1417 		--needed;
1418 		++count;
1419 	}
1420 	if (!wr)
1421 		goto out;
1422 
1423 	rc = ib_post_recv(ep->re_id->qp, wr,
1424 			  (const struct ib_recv_wr **)&bad_wr);
1425 out:
1426 	trace_xprtrdma_post_recvs(r_xprt, count, rc);
1427 	if (rc) {
1428 		for (wr = bad_wr; wr;) {
1429 			struct rpcrdma_rep *rep;
1430 
1431 			rep = container_of(wr, struct rpcrdma_rep, rr_recv_wr);
1432 			wr = wr->next;
1433 			rpcrdma_recv_buffer_put(rep);
1434 			--count;
1435 		}
1436 	}
1437 	ep->re_receive_count += count;
1438 	return;
1439 }
1440