xref: /openbmc/linux/net/sunrpc/xprtrdma/verbs.c (revision 0e17c50f)
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 struct rpcrdma_regbuf *
88 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction,
89 		     gfp_t flags);
90 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb);
91 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb);
92 
93 /* Wait for outstanding transport work to finish. ib_drain_qp
94  * handles the drains in the wrong order for us, so open code
95  * them here.
96  */
97 static void rpcrdma_xprt_drain(struct rpcrdma_xprt *r_xprt)
98 {
99 	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
100 
101 	/* Flush Receives, then wait for deferred Reply work
102 	 * to complete.
103 	 */
104 	ib_drain_rq(ia->ri_id->qp);
105 
106 	/* Deferred Reply processing might have scheduled
107 	 * local invalidations.
108 	 */
109 	ib_drain_sq(ia->ri_id->qp);
110 }
111 
112 /**
113  * rpcrdma_qp_event_handler - Handle one QP event (error notification)
114  * @event: details of the event
115  * @context: ep that owns QP where event occurred
116  *
117  * Called from the RDMA provider (device driver) possibly in an interrupt
118  * context.
119  */
120 static void
121 rpcrdma_qp_event_handler(struct ib_event *event, void *context)
122 {
123 	struct rpcrdma_ep *ep = context;
124 	struct rpcrdma_xprt *r_xprt = container_of(ep, struct rpcrdma_xprt,
125 						   rx_ep);
126 
127 	trace_xprtrdma_qp_event(r_xprt, event);
128 }
129 
130 /**
131  * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC
132  * @cq:	completion queue
133  * @wc:	completed WR
134  *
135  */
136 static void
137 rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
138 {
139 	struct ib_cqe *cqe = wc->wr_cqe;
140 	struct rpcrdma_sendctx *sc =
141 		container_of(cqe, struct rpcrdma_sendctx, sc_cqe);
142 
143 	/* WARNING: Only wr_cqe and status are reliable at this point */
144 	trace_xprtrdma_wc_send(sc, wc);
145 	rpcrdma_sendctx_put_locked((struct rpcrdma_xprt *)cq->cq_context, sc);
146 }
147 
148 /**
149  * rpcrdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
150  * @cq:	completion queue (ignored)
151  * @wc:	completed WR
152  *
153  */
154 static void
155 rpcrdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
156 {
157 	struct ib_cqe *cqe = wc->wr_cqe;
158 	struct rpcrdma_rep *rep = container_of(cqe, struct rpcrdma_rep,
159 					       rr_cqe);
160 	struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
161 
162 	/* WARNING: Only wr_cqe and status are reliable at this point */
163 	trace_xprtrdma_wc_receive(wc);
164 	--r_xprt->rx_ep.rep_receive_count;
165 	if (wc->status != IB_WC_SUCCESS)
166 		goto out_flushed;
167 
168 	/* status == SUCCESS means all fields in wc are trustworthy */
169 	rpcrdma_set_xdrlen(&rep->rr_hdrbuf, wc->byte_len);
170 	rep->rr_wc_flags = wc->wc_flags;
171 	rep->rr_inv_rkey = wc->ex.invalidate_rkey;
172 
173 	ib_dma_sync_single_for_cpu(rdmab_device(rep->rr_rdmabuf),
174 				   rdmab_addr(rep->rr_rdmabuf),
175 				   wc->byte_len, DMA_FROM_DEVICE);
176 
177 	rpcrdma_reply_handler(rep);
178 	return;
179 
180 out_flushed:
181 	rpcrdma_rep_destroy(rep);
182 }
183 
184 static void rpcrdma_update_cm_private(struct rpcrdma_xprt *r_xprt,
185 				      struct rdma_conn_param *param)
186 {
187 	const struct rpcrdma_connect_private *pmsg = param->private_data;
188 	struct rpcrdma_ep *ep = &r_xprt->rx_ep;
189 	unsigned int rsize, wsize;
190 
191 	/* Default settings for RPC-over-RDMA Version One */
192 	r_xprt->rx_ia.ri_implicit_roundup = xprt_rdma_pad_optimize;
193 	rsize = RPCRDMA_V1_DEF_INLINE_SIZE;
194 	wsize = RPCRDMA_V1_DEF_INLINE_SIZE;
195 
196 	if (pmsg &&
197 	    pmsg->cp_magic == rpcrdma_cmp_magic &&
198 	    pmsg->cp_version == RPCRDMA_CMP_VERSION) {
199 		r_xprt->rx_ia.ri_implicit_roundup = true;
200 		rsize = rpcrdma_decode_buffer_size(pmsg->cp_send_size);
201 		wsize = rpcrdma_decode_buffer_size(pmsg->cp_recv_size);
202 	}
203 
204 	if (rsize < ep->rep_inline_recv)
205 		ep->rep_inline_recv = rsize;
206 	if (wsize < ep->rep_inline_send)
207 		ep->rep_inline_send = wsize;
208 
209 	rpcrdma_set_max_header_sizes(r_xprt);
210 }
211 
212 /**
213  * rpcrdma_cm_event_handler - Handle RDMA CM events
214  * @id: rdma_cm_id on which an event has occurred
215  * @event: details of the event
216  *
217  * Called with @id's mutex held. Returns 1 if caller should
218  * destroy @id, otherwise 0.
219  */
220 static int
221 rpcrdma_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
222 {
223 	struct rpcrdma_xprt *r_xprt = id->context;
224 	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
225 	struct rpcrdma_ep *ep = &r_xprt->rx_ep;
226 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
227 
228 	might_sleep();
229 
230 	trace_xprtrdma_cm_event(r_xprt, event);
231 	switch (event->event) {
232 	case RDMA_CM_EVENT_ADDR_RESOLVED:
233 	case RDMA_CM_EVENT_ROUTE_RESOLVED:
234 		ia->ri_async_rc = 0;
235 		complete(&ia->ri_done);
236 		return 0;
237 	case RDMA_CM_EVENT_ADDR_ERROR:
238 		ia->ri_async_rc = -EPROTO;
239 		complete(&ia->ri_done);
240 		return 0;
241 	case RDMA_CM_EVENT_ROUTE_ERROR:
242 		ia->ri_async_rc = -ENETUNREACH;
243 		complete(&ia->ri_done);
244 		return 0;
245 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
246 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
247 		pr_info("rpcrdma: removing device %s for %s:%s\n",
248 			ia->ri_id->device->name,
249 			rpcrdma_addrstr(r_xprt), rpcrdma_portstr(r_xprt));
250 #endif
251 		init_completion(&ia->ri_remove_done);
252 		set_bit(RPCRDMA_IAF_REMOVING, &ia->ri_flags);
253 		ep->rep_connected = -ENODEV;
254 		xprt_force_disconnect(xprt);
255 		wait_for_completion(&ia->ri_remove_done);
256 
257 		ia->ri_id = NULL;
258 		/* Return 1 to ensure the core destroys the id. */
259 		return 1;
260 	case RDMA_CM_EVENT_ESTABLISHED:
261 		++xprt->connect_cookie;
262 		ep->rep_connected = 1;
263 		rpcrdma_update_cm_private(r_xprt, &event->param.conn);
264 		trace_xprtrdma_inline_thresh(r_xprt);
265 		wake_up_all(&ep->rep_connect_wait);
266 		break;
267 	case RDMA_CM_EVENT_CONNECT_ERROR:
268 		ep->rep_connected = -ENOTCONN;
269 		goto disconnected;
270 	case RDMA_CM_EVENT_UNREACHABLE:
271 		ep->rep_connected = -ENETUNREACH;
272 		goto disconnected;
273 	case RDMA_CM_EVENT_REJECTED:
274 		dprintk("rpcrdma: connection to %s:%s rejected: %s\n",
275 			rpcrdma_addrstr(r_xprt), rpcrdma_portstr(r_xprt),
276 			rdma_reject_msg(id, event->status));
277 		ep->rep_connected = -ECONNREFUSED;
278 		if (event->status == IB_CM_REJ_STALE_CONN)
279 			ep->rep_connected = -EAGAIN;
280 		goto disconnected;
281 	case RDMA_CM_EVENT_DISCONNECTED:
282 		ep->rep_connected = -ECONNABORTED;
283 disconnected:
284 		xprt_force_disconnect(xprt);
285 		wake_up_all(&ep->rep_connect_wait);
286 		break;
287 	default:
288 		break;
289 	}
290 
291 	dprintk("RPC:       %s: %s:%s on %s/frwr: %s\n", __func__,
292 		rpcrdma_addrstr(r_xprt), rpcrdma_portstr(r_xprt),
293 		ia->ri_id->device->name, rdma_event_msg(event->event));
294 	return 0;
295 }
296 
297 static struct rdma_cm_id *
298 rpcrdma_create_id(struct rpcrdma_xprt *xprt, struct rpcrdma_ia *ia)
299 {
300 	unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1;
301 	struct rdma_cm_id *id;
302 	int rc;
303 
304 	init_completion(&ia->ri_done);
305 
306 	id = rdma_create_id(xprt->rx_xprt.xprt_net, rpcrdma_cm_event_handler,
307 			    xprt, RDMA_PS_TCP, IB_QPT_RC);
308 	if (IS_ERR(id))
309 		return id;
310 
311 	ia->ri_async_rc = -ETIMEDOUT;
312 	rc = rdma_resolve_addr(id, NULL,
313 			       (struct sockaddr *)&xprt->rx_xprt.addr,
314 			       RDMA_RESOLVE_TIMEOUT);
315 	if (rc)
316 		goto out;
317 	rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
318 	if (rc < 0)
319 		goto out;
320 
321 	rc = ia->ri_async_rc;
322 	if (rc)
323 		goto out;
324 
325 	ia->ri_async_rc = -ETIMEDOUT;
326 	rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
327 	if (rc)
328 		goto out;
329 	rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
330 	if (rc < 0)
331 		goto out;
332 	rc = ia->ri_async_rc;
333 	if (rc)
334 		goto out;
335 
336 	return id;
337 
338 out:
339 	rdma_destroy_id(id);
340 	return ERR_PTR(rc);
341 }
342 
343 /*
344  * Exported functions.
345  */
346 
347 /**
348  * rpcrdma_ia_open - Open and initialize an Interface Adapter.
349  * @xprt: transport with IA to (re)initialize
350  *
351  * Returns 0 on success, negative errno if an appropriate
352  * Interface Adapter could not be found and opened.
353  */
354 int
355 rpcrdma_ia_open(struct rpcrdma_xprt *xprt)
356 {
357 	struct rpcrdma_ia *ia = &xprt->rx_ia;
358 	int rc;
359 
360 	ia->ri_id = rpcrdma_create_id(xprt, ia);
361 	if (IS_ERR(ia->ri_id)) {
362 		rc = PTR_ERR(ia->ri_id);
363 		goto out_err;
364 	}
365 
366 	ia->ri_pd = ib_alloc_pd(ia->ri_id->device, 0);
367 	if (IS_ERR(ia->ri_pd)) {
368 		rc = PTR_ERR(ia->ri_pd);
369 		pr_err("rpcrdma: ib_alloc_pd() returned %d\n", rc);
370 		goto out_err;
371 	}
372 
373 	return 0;
374 
375 out_err:
376 	rpcrdma_ia_close(ia);
377 	return rc;
378 }
379 
380 /**
381  * rpcrdma_ia_remove - Handle device driver unload
382  * @ia: interface adapter being removed
383  *
384  * Divest transport H/W resources associated with this adapter,
385  * but allow it to be restored later.
386  *
387  * Caller must hold the transport send lock.
388  */
389 void
390 rpcrdma_ia_remove(struct rpcrdma_ia *ia)
391 {
392 	struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
393 						   rx_ia);
394 	struct rpcrdma_ep *ep = &r_xprt->rx_ep;
395 
396 	/* This is similar to rpcrdma_ep_destroy, but:
397 	 * - Don't cancel the connect worker.
398 	 * - Don't call rpcrdma_ep_disconnect, which waits
399 	 *   for another conn upcall, which will deadlock.
400 	 * - rdma_disconnect is unneeded, the underlying
401 	 *   connection is already gone.
402 	 */
403 	if (ia->ri_id->qp) {
404 		rpcrdma_xprt_drain(r_xprt);
405 		rdma_destroy_qp(ia->ri_id);
406 		ia->ri_id->qp = NULL;
407 	}
408 	ib_free_cq(ep->rep_attr.recv_cq);
409 	ep->rep_attr.recv_cq = NULL;
410 	ib_free_cq(ep->rep_attr.send_cq);
411 	ep->rep_attr.send_cq = NULL;
412 
413 	/* The ULP is responsible for ensuring all DMA
414 	 * mappings and MRs are gone.
415 	 */
416 	rpcrdma_reps_unmap(r_xprt);
417 	rpcrdma_reqs_reset(r_xprt);
418 	rpcrdma_mrs_destroy(r_xprt);
419 	rpcrdma_sendctxs_destroy(r_xprt);
420 	ib_dealloc_pd(ia->ri_pd);
421 	ia->ri_pd = NULL;
422 
423 	/* Allow waiters to continue */
424 	complete(&ia->ri_remove_done);
425 
426 	trace_xprtrdma_remove(r_xprt);
427 }
428 
429 /**
430  * rpcrdma_ia_close - Clean up/close an IA.
431  * @ia: interface adapter to close
432  *
433  */
434 void
435 rpcrdma_ia_close(struct rpcrdma_ia *ia)
436 {
437 	if (ia->ri_id != NULL && !IS_ERR(ia->ri_id)) {
438 		if (ia->ri_id->qp)
439 			rdma_destroy_qp(ia->ri_id);
440 		rdma_destroy_id(ia->ri_id);
441 	}
442 	ia->ri_id = NULL;
443 
444 	/* If the pd is still busy, xprtrdma missed freeing a resource */
445 	if (ia->ri_pd && !IS_ERR(ia->ri_pd))
446 		ib_dealloc_pd(ia->ri_pd);
447 	ia->ri_pd = NULL;
448 }
449 
450 /**
451  * rpcrdma_ep_create - Create unconnected endpoint
452  * @r_xprt: transport to instantiate
453  *
454  * Returns zero on success, or a negative errno.
455  */
456 int rpcrdma_ep_create(struct rpcrdma_xprt *r_xprt)
457 {
458 	struct rpcrdma_ep *ep = &r_xprt->rx_ep;
459 	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
460 	struct rpcrdma_connect_private *pmsg = &ep->rep_cm_private;
461 	struct ib_cq *sendcq, *recvcq;
462 	int rc;
463 
464 	ep->rep_max_requests = r_xprt->rx_xprt.max_reqs;
465 	ep->rep_inline_send = xprt_rdma_max_inline_write;
466 	ep->rep_inline_recv = xprt_rdma_max_inline_read;
467 
468 	rc = frwr_query_device(r_xprt, ia->ri_id->device);
469 	if (rc)
470 		return rc;
471 	r_xprt->rx_buf.rb_max_requests = cpu_to_be32(ep->rep_max_requests);
472 
473 	ep->rep_attr.event_handler = rpcrdma_qp_event_handler;
474 	ep->rep_attr.qp_context = ep;
475 	ep->rep_attr.srq = NULL;
476 	ep->rep_attr.cap.max_inline_data = 0;
477 	ep->rep_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
478 	ep->rep_attr.qp_type = IB_QPT_RC;
479 	ep->rep_attr.port_num = ~0;
480 
481 	dprintk("RPC:       %s: requested max: dtos: send %d recv %d; "
482 		"iovs: send %d recv %d\n",
483 		__func__,
484 		ep->rep_attr.cap.max_send_wr,
485 		ep->rep_attr.cap.max_recv_wr,
486 		ep->rep_attr.cap.max_send_sge,
487 		ep->rep_attr.cap.max_recv_sge);
488 
489 	ep->rep_send_batch = ep->rep_max_requests >> 3;
490 	ep->rep_send_count = ep->rep_send_batch;
491 	init_waitqueue_head(&ep->rep_connect_wait);
492 	ep->rep_receive_count = 0;
493 
494 	sendcq = ib_alloc_cq_any(ia->ri_id->device, r_xprt,
495 				 ep->rep_attr.cap.max_send_wr + 1,
496 				 IB_POLL_WORKQUEUE);
497 	if (IS_ERR(sendcq)) {
498 		rc = PTR_ERR(sendcq);
499 		goto out1;
500 	}
501 
502 	recvcq = ib_alloc_cq_any(ia->ri_id->device, NULL,
503 				 ep->rep_attr.cap.max_recv_wr + 1,
504 				 IB_POLL_WORKQUEUE);
505 	if (IS_ERR(recvcq)) {
506 		rc = PTR_ERR(recvcq);
507 		goto out2;
508 	}
509 
510 	ep->rep_attr.send_cq = sendcq;
511 	ep->rep_attr.recv_cq = recvcq;
512 
513 	/* Initialize cma parameters */
514 	memset(&ep->rep_remote_cma, 0, sizeof(ep->rep_remote_cma));
515 
516 	/* Prepare RDMA-CM private message */
517 	pmsg->cp_magic = rpcrdma_cmp_magic;
518 	pmsg->cp_version = RPCRDMA_CMP_VERSION;
519 	pmsg->cp_flags |= RPCRDMA_CMP_F_SND_W_INV_OK;
520 	pmsg->cp_send_size = rpcrdma_encode_buffer_size(ep->rep_inline_send);
521 	pmsg->cp_recv_size = rpcrdma_encode_buffer_size(ep->rep_inline_recv);
522 	ep->rep_remote_cma.private_data = pmsg;
523 	ep->rep_remote_cma.private_data_len = sizeof(*pmsg);
524 
525 	/* Client offers RDMA Read but does not initiate */
526 	ep->rep_remote_cma.initiator_depth = 0;
527 	ep->rep_remote_cma.responder_resources =
528 		min_t(int, U8_MAX, ia->ri_id->device->attrs.max_qp_rd_atom);
529 
530 	/* Limit transport retries so client can detect server
531 	 * GID changes quickly. RPC layer handles re-establishing
532 	 * transport connection and retransmission.
533 	 */
534 	ep->rep_remote_cma.retry_count = 6;
535 
536 	/* RPC-over-RDMA handles its own flow control. In addition,
537 	 * make all RNR NAKs visible so we know that RPC-over-RDMA
538 	 * flow control is working correctly (no NAKs should be seen).
539 	 */
540 	ep->rep_remote_cma.flow_control = 0;
541 	ep->rep_remote_cma.rnr_retry_count = 0;
542 
543 	return 0;
544 
545 out2:
546 	ib_free_cq(sendcq);
547 out1:
548 	return rc;
549 }
550 
551 /**
552  * rpcrdma_ep_destroy - Disconnect and destroy endpoint.
553  * @r_xprt: transport instance to shut down
554  *
555  */
556 void rpcrdma_ep_destroy(struct rpcrdma_xprt *r_xprt)
557 {
558 	struct rpcrdma_ep *ep = &r_xprt->rx_ep;
559 	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
560 
561 	if (ia->ri_id && ia->ri_id->qp) {
562 		rpcrdma_ep_disconnect(ep, ia);
563 		rdma_destroy_qp(ia->ri_id);
564 		ia->ri_id->qp = NULL;
565 	}
566 
567 	if (ep->rep_attr.recv_cq)
568 		ib_free_cq(ep->rep_attr.recv_cq);
569 	if (ep->rep_attr.send_cq)
570 		ib_free_cq(ep->rep_attr.send_cq);
571 }
572 
573 /* Re-establish a connection after a device removal event.
574  * Unlike a normal reconnection, a fresh PD and a new set
575  * of MRs and buffers is needed.
576  */
577 static int rpcrdma_ep_recreate_xprt(struct rpcrdma_xprt *r_xprt,
578 				    struct ib_qp_init_attr *qp_init_attr)
579 {
580 	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
581 	struct rpcrdma_ep *ep = &r_xprt->rx_ep;
582 	int rc, err;
583 
584 	trace_xprtrdma_reinsert(r_xprt);
585 
586 	rc = -EHOSTUNREACH;
587 	if (rpcrdma_ia_open(r_xprt))
588 		goto out1;
589 
590 	rc = -ENOMEM;
591 	err = rpcrdma_ep_create(r_xprt);
592 	if (err) {
593 		pr_err("rpcrdma: rpcrdma_ep_create returned %d\n", err);
594 		goto out2;
595 	}
596 	memcpy(qp_init_attr, &ep->rep_attr, sizeof(*qp_init_attr));
597 
598 	rc = -ENETUNREACH;
599 	err = rdma_create_qp(ia->ri_id, ia->ri_pd, qp_init_attr);
600 	if (err) {
601 		pr_err("rpcrdma: rdma_create_qp returned %d\n", err);
602 		goto out3;
603 	}
604 	return 0;
605 
606 out3:
607 	rpcrdma_ep_destroy(r_xprt);
608 out2:
609 	rpcrdma_ia_close(ia);
610 out1:
611 	return rc;
612 }
613 
614 static int rpcrdma_ep_reconnect(struct rpcrdma_xprt *r_xprt,
615 				struct ib_qp_init_attr *qp_init_attr)
616 {
617 	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
618 	struct rdma_cm_id *id, *old;
619 	int err, rc;
620 
621 	rpcrdma_ep_disconnect(&r_xprt->rx_ep, ia);
622 
623 	rc = -EHOSTUNREACH;
624 	id = rpcrdma_create_id(r_xprt, ia);
625 	if (IS_ERR(id))
626 		goto out;
627 
628 	/* As long as the new ID points to the same device as the
629 	 * old ID, we can reuse the transport's existing PD and all
630 	 * previously allocated MRs. Also, the same device means
631 	 * the transport's previous DMA mappings are still valid.
632 	 *
633 	 * This is a sanity check only. There should be no way these
634 	 * point to two different devices here.
635 	 */
636 	old = id;
637 	rc = -ENETUNREACH;
638 	if (ia->ri_id->device != id->device) {
639 		pr_err("rpcrdma: can't reconnect on different device!\n");
640 		goto out_destroy;
641 	}
642 
643 	err = rdma_create_qp(id, ia->ri_pd, qp_init_attr);
644 	if (err)
645 		goto out_destroy;
646 
647 	/* Atomically replace the transport's ID and QP. */
648 	rc = 0;
649 	old = ia->ri_id;
650 	ia->ri_id = id;
651 	rdma_destroy_qp(old);
652 
653 out_destroy:
654 	rdma_destroy_id(old);
655 out:
656 	return rc;
657 }
658 
659 /*
660  * Connect unconnected endpoint.
661  */
662 int
663 rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
664 {
665 	struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
666 						   rx_ia);
667 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
668 	struct ib_qp_init_attr qp_init_attr;
669 	int rc;
670 
671 retry:
672 	memcpy(&qp_init_attr, &ep->rep_attr, sizeof(qp_init_attr));
673 	switch (ep->rep_connected) {
674 	case 0:
675 		rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &qp_init_attr);
676 		if (rc) {
677 			rc = -ENETUNREACH;
678 			goto out_noupdate;
679 		}
680 		break;
681 	case -ENODEV:
682 		rc = rpcrdma_ep_recreate_xprt(r_xprt, &qp_init_attr);
683 		if (rc)
684 			goto out_noupdate;
685 		break;
686 	default:
687 		rc = rpcrdma_ep_reconnect(r_xprt, &qp_init_attr);
688 		if (rc)
689 			goto out;
690 	}
691 
692 	ep->rep_connected = 0;
693 	xprt_clear_connected(xprt);
694 
695 	rpcrdma_reset_cwnd(r_xprt);
696 	rpcrdma_post_recvs(r_xprt, true);
697 
698 	rc = rpcrdma_sendctxs_create(r_xprt);
699 	if (rc)
700 		goto out;
701 
702 	rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma);
703 	if (rc)
704 		goto out;
705 
706 	if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
707 		xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
708 	wait_event_interruptible(ep->rep_connect_wait, ep->rep_connected != 0);
709 	if (ep->rep_connected <= 0) {
710 		if (ep->rep_connected == -EAGAIN)
711 			goto retry;
712 		rc = ep->rep_connected;
713 		goto out;
714 	}
715 
716 	rc = rpcrdma_reqs_setup(r_xprt);
717 	if (rc) {
718 		rpcrdma_ep_disconnect(ep, ia);
719 		goto out;
720 	}
721 	rpcrdma_mrs_create(r_xprt);
722 
723 out:
724 	if (rc)
725 		ep->rep_connected = rc;
726 
727 out_noupdate:
728 	trace_xprtrdma_connect(r_xprt, rc);
729 	return rc;
730 }
731 
732 /**
733  * rpcrdma_ep_disconnect - Disconnect underlying transport
734  * @ep: endpoint to disconnect
735  * @ia: associated interface adapter
736  *
737  * Caller serializes. Either the transport send lock is held,
738  * or we're being called to destroy the transport.
739  */
740 void
741 rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
742 {
743 	struct rpcrdma_xprt *r_xprt = container_of(ep, struct rpcrdma_xprt,
744 						   rx_ep);
745 	int rc;
746 
747 	/* returns without wait if ID is not connected */
748 	rc = rdma_disconnect(ia->ri_id);
749 	if (!rc)
750 		wait_event_interruptible(ep->rep_connect_wait,
751 							ep->rep_connected != 1);
752 	else
753 		ep->rep_connected = rc;
754 	trace_xprtrdma_disconnect(r_xprt, rc);
755 
756 	rpcrdma_xprt_drain(r_xprt);
757 	rpcrdma_reqs_reset(r_xprt);
758 	rpcrdma_mrs_destroy(r_xprt);
759 	rpcrdma_sendctxs_destroy(r_xprt);
760 }
761 
762 /* Fixed-size circular FIFO queue. This implementation is wait-free and
763  * lock-free.
764  *
765  * Consumer is the code path that posts Sends. This path dequeues a
766  * sendctx for use by a Send operation. Multiple consumer threads
767  * are serialized by the RPC transport lock, which allows only one
768  * ->send_request call at a time.
769  *
770  * Producer is the code path that handles Send completions. This path
771  * enqueues a sendctx that has been completed. Multiple producer
772  * threads are serialized by the ib_poll_cq() function.
773  */
774 
775 /* rpcrdma_sendctxs_destroy() assumes caller has already quiesced
776  * queue activity, and rpcrdma_xprt_drain has flushed all remaining
777  * Send requests.
778  */
779 static void rpcrdma_sendctxs_destroy(struct rpcrdma_xprt *r_xprt)
780 {
781 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
782 	unsigned long i;
783 
784 	if (!buf->rb_sc_ctxs)
785 		return;
786 	for (i = 0; i <= buf->rb_sc_last; i++)
787 		kfree(buf->rb_sc_ctxs[i]);
788 	kfree(buf->rb_sc_ctxs);
789 	buf->rb_sc_ctxs = NULL;
790 }
791 
792 static struct rpcrdma_sendctx *rpcrdma_sendctx_create(struct rpcrdma_ep *ep)
793 {
794 	struct rpcrdma_sendctx *sc;
795 
796 	sc = kzalloc(struct_size(sc, sc_sges, ep->rep_attr.cap.max_send_sge),
797 		     GFP_KERNEL);
798 	if (!sc)
799 		return NULL;
800 
801 	sc->sc_cqe.done = rpcrdma_wc_send;
802 	return sc;
803 }
804 
805 static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt)
806 {
807 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
808 	struct rpcrdma_sendctx *sc;
809 	unsigned long i;
810 
811 	/* Maximum number of concurrent outstanding Send WRs. Capping
812 	 * the circular queue size stops Send Queue overflow by causing
813 	 * the ->send_request call to fail temporarily before too many
814 	 * Sends are posted.
815 	 */
816 	i = r_xprt->rx_ep.rep_max_requests + RPCRDMA_MAX_BC_REQUESTS;
817 	buf->rb_sc_ctxs = kcalloc(i, sizeof(sc), GFP_KERNEL);
818 	if (!buf->rb_sc_ctxs)
819 		return -ENOMEM;
820 
821 	buf->rb_sc_last = i - 1;
822 	for (i = 0; i <= buf->rb_sc_last; i++) {
823 		sc = rpcrdma_sendctx_create(&r_xprt->rx_ep);
824 		if (!sc)
825 			return -ENOMEM;
826 
827 		buf->rb_sc_ctxs[i] = sc;
828 	}
829 
830 	buf->rb_sc_head = 0;
831 	buf->rb_sc_tail = 0;
832 	return 0;
833 }
834 
835 /* The sendctx queue is not guaranteed to have a size that is a
836  * power of two, thus the helpers in circ_buf.h cannot be used.
837  * The other option is to use modulus (%), which can be expensive.
838  */
839 static unsigned long rpcrdma_sendctx_next(struct rpcrdma_buffer *buf,
840 					  unsigned long item)
841 {
842 	return likely(item < buf->rb_sc_last) ? item + 1 : 0;
843 }
844 
845 /**
846  * rpcrdma_sendctx_get_locked - Acquire a send context
847  * @r_xprt: controlling transport instance
848  *
849  * Returns pointer to a free send completion context; or NULL if
850  * the queue is empty.
851  *
852  * Usage: Called to acquire an SGE array before preparing a Send WR.
853  *
854  * The caller serializes calls to this function (per transport), and
855  * provides an effective memory barrier that flushes the new value
856  * of rb_sc_head.
857  */
858 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_xprt *r_xprt)
859 {
860 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
861 	struct rpcrdma_sendctx *sc;
862 	unsigned long next_head;
863 
864 	next_head = rpcrdma_sendctx_next(buf, buf->rb_sc_head);
865 
866 	if (next_head == READ_ONCE(buf->rb_sc_tail))
867 		goto out_emptyq;
868 
869 	/* ORDER: item must be accessed _before_ head is updated */
870 	sc = buf->rb_sc_ctxs[next_head];
871 
872 	/* Releasing the lock in the caller acts as a memory
873 	 * barrier that flushes rb_sc_head.
874 	 */
875 	buf->rb_sc_head = next_head;
876 
877 	return sc;
878 
879 out_emptyq:
880 	/* The queue is "empty" if there have not been enough Send
881 	 * completions recently. This is a sign the Send Queue is
882 	 * backing up. Cause the caller to pause and try again.
883 	 */
884 	xprt_wait_for_buffer_space(&r_xprt->rx_xprt);
885 	r_xprt->rx_stats.empty_sendctx_q++;
886 	return NULL;
887 }
888 
889 /**
890  * rpcrdma_sendctx_put_locked - Release a send context
891  * @r_xprt: controlling transport instance
892  * @sc: send context to release
893  *
894  * Usage: Called from Send completion to return a sendctxt
895  * to the queue.
896  *
897  * The caller serializes calls to this function (per transport).
898  */
899 static void rpcrdma_sendctx_put_locked(struct rpcrdma_xprt *r_xprt,
900 				       struct rpcrdma_sendctx *sc)
901 {
902 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
903 	unsigned long next_tail;
904 
905 	/* Unmap SGEs of previously completed but unsignaled
906 	 * Sends by walking up the queue until @sc is found.
907 	 */
908 	next_tail = buf->rb_sc_tail;
909 	do {
910 		next_tail = rpcrdma_sendctx_next(buf, next_tail);
911 
912 		/* ORDER: item must be accessed _before_ tail is updated */
913 		rpcrdma_sendctx_unmap(buf->rb_sc_ctxs[next_tail]);
914 
915 	} while (buf->rb_sc_ctxs[next_tail] != sc);
916 
917 	/* Paired with READ_ONCE */
918 	smp_store_release(&buf->rb_sc_tail, next_tail);
919 
920 	xprt_write_space(&r_xprt->rx_xprt);
921 }
922 
923 static void
924 rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt)
925 {
926 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
927 	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
928 	unsigned int count;
929 
930 	for (count = 0; count < ia->ri_max_rdma_segs; count++) {
931 		struct rpcrdma_mr *mr;
932 		int rc;
933 
934 		mr = kzalloc(sizeof(*mr), GFP_NOFS);
935 		if (!mr)
936 			break;
937 
938 		rc = frwr_init_mr(ia, mr);
939 		if (rc) {
940 			kfree(mr);
941 			break;
942 		}
943 
944 		mr->mr_xprt = r_xprt;
945 
946 		spin_lock(&buf->rb_lock);
947 		rpcrdma_mr_push(mr, &buf->rb_mrs);
948 		list_add(&mr->mr_all, &buf->rb_all_mrs);
949 		spin_unlock(&buf->rb_lock);
950 	}
951 
952 	r_xprt->rx_stats.mrs_allocated += count;
953 	trace_xprtrdma_createmrs(r_xprt, count);
954 }
955 
956 static void
957 rpcrdma_mr_refresh_worker(struct work_struct *work)
958 {
959 	struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
960 						  rb_refresh_worker);
961 	struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
962 						   rx_buf);
963 
964 	rpcrdma_mrs_create(r_xprt);
965 	xprt_write_space(&r_xprt->rx_xprt);
966 }
967 
968 /**
969  * rpcrdma_mrs_refresh - Wake the MR refresh worker
970  * @r_xprt: controlling transport instance
971  *
972  */
973 void rpcrdma_mrs_refresh(struct rpcrdma_xprt *r_xprt)
974 {
975 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
976 	struct rpcrdma_ep *ep = &r_xprt->rx_ep;
977 
978 	/* If there is no underlying device, it's no use to
979 	 * wake the refresh worker.
980 	 */
981 	if (ep->rep_connected != -ENODEV) {
982 		/* The work is scheduled on a WQ_MEM_RECLAIM
983 		 * workqueue in order to prevent MR allocation
984 		 * from recursing into NFS during direct reclaim.
985 		 */
986 		queue_work(xprtiod_workqueue, &buf->rb_refresh_worker);
987 	}
988 }
989 
990 /**
991  * rpcrdma_req_create - Allocate an rpcrdma_req object
992  * @r_xprt: controlling r_xprt
993  * @size: initial size, in bytes, of send and receive buffers
994  * @flags: GFP flags passed to memory allocators
995  *
996  * Returns an allocated and fully initialized rpcrdma_req or NULL.
997  */
998 struct rpcrdma_req *rpcrdma_req_create(struct rpcrdma_xprt *r_xprt, size_t size,
999 				       gfp_t flags)
1000 {
1001 	struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
1002 	struct rpcrdma_req *req;
1003 
1004 	req = kzalloc(sizeof(*req), flags);
1005 	if (req == NULL)
1006 		goto out1;
1007 
1008 	req->rl_sendbuf = rpcrdma_regbuf_alloc(size, DMA_TO_DEVICE, flags);
1009 	if (!req->rl_sendbuf)
1010 		goto out2;
1011 
1012 	req->rl_recvbuf = rpcrdma_regbuf_alloc(size, DMA_NONE, flags);
1013 	if (!req->rl_recvbuf)
1014 		goto out3;
1015 
1016 	INIT_LIST_HEAD(&req->rl_free_mrs);
1017 	INIT_LIST_HEAD(&req->rl_registered);
1018 	spin_lock(&buffer->rb_lock);
1019 	list_add(&req->rl_all, &buffer->rb_allreqs);
1020 	spin_unlock(&buffer->rb_lock);
1021 	return req;
1022 
1023 out3:
1024 	kfree(req->rl_sendbuf);
1025 out2:
1026 	kfree(req);
1027 out1:
1028 	return NULL;
1029 }
1030 
1031 /**
1032  * rpcrdma_req_setup - Per-connection instance setup of an rpcrdma_req object
1033  * @r_xprt: controlling transport instance
1034  * @req: rpcrdma_req object to set up
1035  *
1036  * Returns zero on success, and a negative errno on failure.
1037  */
1038 int rpcrdma_req_setup(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
1039 {
1040 	struct rpcrdma_regbuf *rb;
1041 	size_t maxhdrsize;
1042 
1043 	/* Compute maximum header buffer size in bytes */
1044 	maxhdrsize = rpcrdma_fixed_maxsz + 3 +
1045 		     r_xprt->rx_ia.ri_max_rdma_segs * rpcrdma_readchunk_maxsz;
1046 	maxhdrsize *= sizeof(__be32);
1047 	rb = rpcrdma_regbuf_alloc(__roundup_pow_of_two(maxhdrsize),
1048 				  DMA_TO_DEVICE, GFP_KERNEL);
1049 	if (!rb)
1050 		goto out;
1051 
1052 	if (!__rpcrdma_regbuf_dma_map(r_xprt, rb))
1053 		goto out_free;
1054 
1055 	req->rl_rdmabuf = rb;
1056 	xdr_buf_init(&req->rl_hdrbuf, rdmab_data(rb), rdmab_length(rb));
1057 	return 0;
1058 
1059 out_free:
1060 	rpcrdma_regbuf_free(rb);
1061 out:
1062 	return -ENOMEM;
1063 }
1064 
1065 /* ASSUMPTION: the rb_allreqs list is stable for the duration,
1066  * and thus can be walked without holding rb_lock. Eg. the
1067  * caller is holding the transport send lock to exclude
1068  * device removal or disconnection.
1069  */
1070 static int rpcrdma_reqs_setup(struct rpcrdma_xprt *r_xprt)
1071 {
1072 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1073 	struct rpcrdma_req *req;
1074 	int rc;
1075 
1076 	list_for_each_entry(req, &buf->rb_allreqs, rl_all) {
1077 		rc = rpcrdma_req_setup(r_xprt, req);
1078 		if (rc)
1079 			return rc;
1080 	}
1081 	return 0;
1082 }
1083 
1084 static void rpcrdma_req_reset(struct rpcrdma_req *req)
1085 {
1086 	/* Credits are valid for only one connection */
1087 	req->rl_slot.rq_cong = 0;
1088 
1089 	rpcrdma_regbuf_free(req->rl_rdmabuf);
1090 	req->rl_rdmabuf = NULL;
1091 
1092 	rpcrdma_regbuf_dma_unmap(req->rl_sendbuf);
1093 	rpcrdma_regbuf_dma_unmap(req->rl_recvbuf);
1094 }
1095 
1096 /* ASSUMPTION: the rb_allreqs list is stable for the duration,
1097  * and thus can be walked without holding rb_lock. Eg. the
1098  * caller is holding the transport send lock to exclude
1099  * device removal or disconnection.
1100  */
1101 static void rpcrdma_reqs_reset(struct rpcrdma_xprt *r_xprt)
1102 {
1103 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1104 	struct rpcrdma_req *req;
1105 
1106 	list_for_each_entry(req, &buf->rb_allreqs, rl_all)
1107 		rpcrdma_req_reset(req);
1108 }
1109 
1110 /* No locking needed here. This function is called only by the
1111  * Receive completion handler.
1112  */
1113 static noinline
1114 struct rpcrdma_rep *rpcrdma_rep_create(struct rpcrdma_xprt *r_xprt,
1115 				       bool temp)
1116 {
1117 	struct rpcrdma_rep *rep;
1118 
1119 	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1120 	if (rep == NULL)
1121 		goto out;
1122 
1123 	rep->rr_rdmabuf = rpcrdma_regbuf_alloc(r_xprt->rx_ep.rep_inline_recv,
1124 					       DMA_FROM_DEVICE, GFP_KERNEL);
1125 	if (!rep->rr_rdmabuf)
1126 		goto out_free;
1127 
1128 	if (!rpcrdma_regbuf_dma_map(r_xprt, rep->rr_rdmabuf))
1129 		goto out_free_regbuf;
1130 
1131 	xdr_buf_init(&rep->rr_hdrbuf, rdmab_data(rep->rr_rdmabuf),
1132 		     rdmab_length(rep->rr_rdmabuf));
1133 	rep->rr_cqe.done = rpcrdma_wc_receive;
1134 	rep->rr_rxprt = r_xprt;
1135 	rep->rr_recv_wr.next = NULL;
1136 	rep->rr_recv_wr.wr_cqe = &rep->rr_cqe;
1137 	rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
1138 	rep->rr_recv_wr.num_sge = 1;
1139 	rep->rr_temp = temp;
1140 	list_add(&rep->rr_all, &r_xprt->rx_buf.rb_all_reps);
1141 	return rep;
1142 
1143 out_free_regbuf:
1144 	rpcrdma_regbuf_free(rep->rr_rdmabuf);
1145 out_free:
1146 	kfree(rep);
1147 out:
1148 	return NULL;
1149 }
1150 
1151 /* No locking needed here. This function is invoked only by the
1152  * Receive completion handler, or during transport shutdown.
1153  */
1154 static void rpcrdma_rep_destroy(struct rpcrdma_rep *rep)
1155 {
1156 	list_del(&rep->rr_all);
1157 	rpcrdma_regbuf_free(rep->rr_rdmabuf);
1158 	kfree(rep);
1159 }
1160 
1161 static struct rpcrdma_rep *rpcrdma_rep_get_locked(struct rpcrdma_buffer *buf)
1162 {
1163 	struct llist_node *node;
1164 
1165 	/* Calls to llist_del_first are required to be serialized */
1166 	node = llist_del_first(&buf->rb_free_reps);
1167 	if (!node)
1168 		return NULL;
1169 	return llist_entry(node, struct rpcrdma_rep, rr_node);
1170 }
1171 
1172 static void rpcrdma_rep_put(struct rpcrdma_buffer *buf,
1173 			    struct rpcrdma_rep *rep)
1174 {
1175 	llist_add(&rep->rr_node, &buf->rb_free_reps);
1176 }
1177 
1178 static void rpcrdma_reps_unmap(struct rpcrdma_xprt *r_xprt)
1179 {
1180 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1181 	struct rpcrdma_rep *rep;
1182 
1183 	list_for_each_entry(rep, &buf->rb_all_reps, rr_all) {
1184 		rpcrdma_regbuf_dma_unmap(rep->rr_rdmabuf);
1185 		rep->rr_temp = true;
1186 	}
1187 }
1188 
1189 static void rpcrdma_reps_destroy(struct rpcrdma_buffer *buf)
1190 {
1191 	struct rpcrdma_rep *rep;
1192 
1193 	while ((rep = rpcrdma_rep_get_locked(buf)) != NULL)
1194 		rpcrdma_rep_destroy(rep);
1195 }
1196 
1197 /**
1198  * rpcrdma_buffer_create - Create initial set of req/rep objects
1199  * @r_xprt: transport instance to (re)initialize
1200  *
1201  * Returns zero on success, otherwise a negative errno.
1202  */
1203 int rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
1204 {
1205 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1206 	int i, rc;
1207 
1208 	buf->rb_bc_srv_max_requests = 0;
1209 	spin_lock_init(&buf->rb_lock);
1210 	INIT_LIST_HEAD(&buf->rb_mrs);
1211 	INIT_LIST_HEAD(&buf->rb_all_mrs);
1212 	INIT_WORK(&buf->rb_refresh_worker, rpcrdma_mr_refresh_worker);
1213 
1214 	INIT_LIST_HEAD(&buf->rb_send_bufs);
1215 	INIT_LIST_HEAD(&buf->rb_allreqs);
1216 	INIT_LIST_HEAD(&buf->rb_all_reps);
1217 
1218 	rc = -ENOMEM;
1219 	for (i = 0; i < r_xprt->rx_xprt.max_reqs; i++) {
1220 		struct rpcrdma_req *req;
1221 
1222 		req = rpcrdma_req_create(r_xprt, RPCRDMA_V1_DEF_INLINE_SIZE * 2,
1223 					 GFP_KERNEL);
1224 		if (!req)
1225 			goto out;
1226 		list_add(&req->rl_list, &buf->rb_send_bufs);
1227 	}
1228 
1229 	init_llist_head(&buf->rb_free_reps);
1230 
1231 	return 0;
1232 out:
1233 	rpcrdma_buffer_destroy(buf);
1234 	return rc;
1235 }
1236 
1237 /**
1238  * rpcrdma_req_destroy - Destroy an rpcrdma_req object
1239  * @req: unused object to be destroyed
1240  *
1241  * Relies on caller holding the transport send lock to protect
1242  * removing req->rl_all from buf->rb_all_reqs safely.
1243  */
1244 void rpcrdma_req_destroy(struct rpcrdma_req *req)
1245 {
1246 	struct rpcrdma_mr *mr;
1247 
1248 	list_del(&req->rl_all);
1249 
1250 	while ((mr = rpcrdma_mr_pop(&req->rl_free_mrs))) {
1251 		struct rpcrdma_buffer *buf = &mr->mr_xprt->rx_buf;
1252 
1253 		spin_lock(&buf->rb_lock);
1254 		list_del(&mr->mr_all);
1255 		spin_unlock(&buf->rb_lock);
1256 
1257 		frwr_release_mr(mr);
1258 	}
1259 
1260 	rpcrdma_regbuf_free(req->rl_recvbuf);
1261 	rpcrdma_regbuf_free(req->rl_sendbuf);
1262 	rpcrdma_regbuf_free(req->rl_rdmabuf);
1263 	kfree(req);
1264 }
1265 
1266 /**
1267  * rpcrdma_mrs_destroy - Release all of a transport's MRs
1268  * @r_xprt: controlling transport instance
1269  *
1270  * Relies on caller holding the transport send lock to protect
1271  * removing mr->mr_list from req->rl_free_mrs safely.
1272  */
1273 static void rpcrdma_mrs_destroy(struct rpcrdma_xprt *r_xprt)
1274 {
1275 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1276 	struct rpcrdma_mr *mr;
1277 
1278 	cancel_work_sync(&buf->rb_refresh_worker);
1279 
1280 	spin_lock(&buf->rb_lock);
1281 	while ((mr = list_first_entry_or_null(&buf->rb_all_mrs,
1282 					      struct rpcrdma_mr,
1283 					      mr_all)) != NULL) {
1284 		list_del(&mr->mr_list);
1285 		list_del(&mr->mr_all);
1286 		spin_unlock(&buf->rb_lock);
1287 
1288 		frwr_release_mr(mr);
1289 
1290 		spin_lock(&buf->rb_lock);
1291 	}
1292 	spin_unlock(&buf->rb_lock);
1293 }
1294 
1295 /**
1296  * rpcrdma_buffer_destroy - Release all hw resources
1297  * @buf: root control block for resources
1298  *
1299  * ORDERING: relies on a prior rpcrdma_xprt_drain :
1300  * - No more Send or Receive completions can occur
1301  * - All MRs, reps, and reqs are returned to their free lists
1302  */
1303 void
1304 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1305 {
1306 	rpcrdma_reps_destroy(buf);
1307 
1308 	while (!list_empty(&buf->rb_send_bufs)) {
1309 		struct rpcrdma_req *req;
1310 
1311 		req = list_first_entry(&buf->rb_send_bufs,
1312 				       struct rpcrdma_req, rl_list);
1313 		list_del(&req->rl_list);
1314 		rpcrdma_req_destroy(req);
1315 	}
1316 }
1317 
1318 /**
1319  * rpcrdma_mr_get - Allocate an rpcrdma_mr object
1320  * @r_xprt: controlling transport
1321  *
1322  * Returns an initialized rpcrdma_mr or NULL if no free
1323  * rpcrdma_mr objects are available.
1324  */
1325 struct rpcrdma_mr *
1326 rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt)
1327 {
1328 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1329 	struct rpcrdma_mr *mr;
1330 
1331 	spin_lock(&buf->rb_lock);
1332 	mr = rpcrdma_mr_pop(&buf->rb_mrs);
1333 	spin_unlock(&buf->rb_lock);
1334 	return mr;
1335 }
1336 
1337 /**
1338  * rpcrdma_mr_put - DMA unmap an MR and release it
1339  * @mr: MR to release
1340  *
1341  */
1342 void rpcrdma_mr_put(struct rpcrdma_mr *mr)
1343 {
1344 	struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
1345 
1346 	if (mr->mr_dir != DMA_NONE) {
1347 		trace_xprtrdma_mr_unmap(mr);
1348 		ib_dma_unmap_sg(r_xprt->rx_ia.ri_id->device,
1349 				mr->mr_sg, mr->mr_nents, mr->mr_dir);
1350 		mr->mr_dir = DMA_NONE;
1351 	}
1352 
1353 	rpcrdma_mr_push(mr, &mr->mr_req->rl_free_mrs);
1354 }
1355 
1356 /**
1357  * rpcrdma_buffer_get - Get a request buffer
1358  * @buffers: Buffer pool from which to obtain a buffer
1359  *
1360  * Returns a fresh rpcrdma_req, or NULL if none are available.
1361  */
1362 struct rpcrdma_req *
1363 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1364 {
1365 	struct rpcrdma_req *req;
1366 
1367 	spin_lock(&buffers->rb_lock);
1368 	req = list_first_entry_or_null(&buffers->rb_send_bufs,
1369 				       struct rpcrdma_req, rl_list);
1370 	if (req)
1371 		list_del_init(&req->rl_list);
1372 	spin_unlock(&buffers->rb_lock);
1373 	return req;
1374 }
1375 
1376 /**
1377  * rpcrdma_buffer_put - Put request/reply buffers back into pool
1378  * @buffers: buffer pool
1379  * @req: object to return
1380  *
1381  */
1382 void rpcrdma_buffer_put(struct rpcrdma_buffer *buffers, struct rpcrdma_req *req)
1383 {
1384 	if (req->rl_reply)
1385 		rpcrdma_rep_put(buffers, req->rl_reply);
1386 	req->rl_reply = NULL;
1387 
1388 	spin_lock(&buffers->rb_lock);
1389 	list_add(&req->rl_list, &buffers->rb_send_bufs);
1390 	spin_unlock(&buffers->rb_lock);
1391 }
1392 
1393 /**
1394  * rpcrdma_recv_buffer_put - Release rpcrdma_rep back to free list
1395  * @rep: rep to release
1396  *
1397  * Used after error conditions.
1398  */
1399 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1400 {
1401 	rpcrdma_rep_put(&rep->rr_rxprt->rx_buf, rep);
1402 }
1403 
1404 /* Returns a pointer to a rpcrdma_regbuf object, or NULL.
1405  *
1406  * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1407  * receiving the payload of RDMA RECV operations. During Long Calls
1408  * or Replies they may be registered externally via frwr_map.
1409  */
1410 static struct rpcrdma_regbuf *
1411 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction,
1412 		     gfp_t flags)
1413 {
1414 	struct rpcrdma_regbuf *rb;
1415 
1416 	rb = kmalloc(sizeof(*rb), flags);
1417 	if (!rb)
1418 		return NULL;
1419 	rb->rg_data = kmalloc(size, flags);
1420 	if (!rb->rg_data) {
1421 		kfree(rb);
1422 		return NULL;
1423 	}
1424 
1425 	rb->rg_device = NULL;
1426 	rb->rg_direction = direction;
1427 	rb->rg_iov.length = size;
1428 	return rb;
1429 }
1430 
1431 /**
1432  * rpcrdma_regbuf_realloc - re-allocate a SEND/RECV buffer
1433  * @rb: regbuf to reallocate
1434  * @size: size of buffer to be allocated, in bytes
1435  * @flags: GFP flags
1436  *
1437  * Returns true if reallocation was successful. If false is
1438  * returned, @rb is left untouched.
1439  */
1440 bool rpcrdma_regbuf_realloc(struct rpcrdma_regbuf *rb, size_t size, gfp_t flags)
1441 {
1442 	void *buf;
1443 
1444 	buf = kmalloc(size, flags);
1445 	if (!buf)
1446 		return false;
1447 
1448 	rpcrdma_regbuf_dma_unmap(rb);
1449 	kfree(rb->rg_data);
1450 
1451 	rb->rg_data = buf;
1452 	rb->rg_iov.length = size;
1453 	return true;
1454 }
1455 
1456 /**
1457  * __rpcrdma_regbuf_dma_map - DMA-map a regbuf
1458  * @r_xprt: controlling transport instance
1459  * @rb: regbuf to be mapped
1460  *
1461  * Returns true if the buffer is now DMA mapped to @r_xprt's device
1462  */
1463 bool __rpcrdma_regbuf_dma_map(struct rpcrdma_xprt *r_xprt,
1464 			      struct rpcrdma_regbuf *rb)
1465 {
1466 	struct ib_device *device = r_xprt->rx_ia.ri_id->device;
1467 
1468 	if (rb->rg_direction == DMA_NONE)
1469 		return false;
1470 
1471 	rb->rg_iov.addr = ib_dma_map_single(device, rdmab_data(rb),
1472 					    rdmab_length(rb), rb->rg_direction);
1473 	if (ib_dma_mapping_error(device, rdmab_addr(rb))) {
1474 		trace_xprtrdma_dma_maperr(rdmab_addr(rb));
1475 		return false;
1476 	}
1477 
1478 	rb->rg_device = device;
1479 	rb->rg_iov.lkey = r_xprt->rx_ia.ri_pd->local_dma_lkey;
1480 	return true;
1481 }
1482 
1483 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb)
1484 {
1485 	if (!rb)
1486 		return;
1487 
1488 	if (!rpcrdma_regbuf_is_mapped(rb))
1489 		return;
1490 
1491 	ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb), rdmab_length(rb),
1492 			    rb->rg_direction);
1493 	rb->rg_device = NULL;
1494 }
1495 
1496 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb)
1497 {
1498 	rpcrdma_regbuf_dma_unmap(rb);
1499 	if (rb)
1500 		kfree(rb->rg_data);
1501 	kfree(rb);
1502 }
1503 
1504 /**
1505  * rpcrdma_ep_post - Post WRs to a transport's Send Queue
1506  * @ia: transport's device information
1507  * @ep: transport's RDMA endpoint information
1508  * @req: rpcrdma_req containing the Send WR to post
1509  *
1510  * Returns 0 if the post was successful, otherwise -ENOTCONN
1511  * is returned.
1512  */
1513 int
1514 rpcrdma_ep_post(struct rpcrdma_ia *ia,
1515 		struct rpcrdma_ep *ep,
1516 		struct rpcrdma_req *req)
1517 {
1518 	struct ib_send_wr *send_wr = &req->rl_wr;
1519 	int rc;
1520 
1521 	if (!ep->rep_send_count || kref_read(&req->rl_kref) > 1) {
1522 		send_wr->send_flags |= IB_SEND_SIGNALED;
1523 		ep->rep_send_count = ep->rep_send_batch;
1524 	} else {
1525 		send_wr->send_flags &= ~IB_SEND_SIGNALED;
1526 		--ep->rep_send_count;
1527 	}
1528 
1529 	rc = frwr_send(ia, req);
1530 	trace_xprtrdma_post_send(req, rc);
1531 	if (rc)
1532 		return -ENOTCONN;
1533 	return 0;
1534 }
1535 
1536 /**
1537  * rpcrdma_post_recvs - Refill the Receive Queue
1538  * @r_xprt: controlling transport instance
1539  * @temp: mark Receive buffers to be deleted after use
1540  *
1541  */
1542 void rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, bool temp)
1543 {
1544 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1545 	struct rpcrdma_ep *ep = &r_xprt->rx_ep;
1546 	struct ib_recv_wr *wr, *bad_wr;
1547 	struct rpcrdma_rep *rep;
1548 	int needed, count, rc;
1549 
1550 	rc = 0;
1551 	count = 0;
1552 
1553 	needed = buf->rb_credits + (buf->rb_bc_srv_max_requests << 1);
1554 	if (likely(ep->rep_receive_count > needed))
1555 		goto out;
1556 	needed -= ep->rep_receive_count;
1557 	if (!temp)
1558 		needed += RPCRDMA_MAX_RECV_BATCH;
1559 
1560 	/* fast path: all needed reps can be found on the free list */
1561 	wr = NULL;
1562 	while (needed) {
1563 		rep = rpcrdma_rep_get_locked(buf);
1564 		if (rep && rep->rr_temp) {
1565 			rpcrdma_rep_destroy(rep);
1566 			continue;
1567 		}
1568 		if (!rep)
1569 			rep = rpcrdma_rep_create(r_xprt, temp);
1570 		if (!rep)
1571 			break;
1572 
1573 		trace_xprtrdma_post_recv(rep);
1574 		rep->rr_recv_wr.next = wr;
1575 		wr = &rep->rr_recv_wr;
1576 		--needed;
1577 		++count;
1578 	}
1579 	if (!wr)
1580 		goto out;
1581 
1582 	rc = ib_post_recv(r_xprt->rx_ia.ri_id->qp, wr,
1583 			  (const struct ib_recv_wr **)&bad_wr);
1584 out:
1585 	trace_xprtrdma_post_recvs(r_xprt, count, rc);
1586 	if (rc) {
1587 		for (wr = bad_wr; wr;) {
1588 			struct rpcrdma_rep *rep;
1589 
1590 			rep = container_of(wr, struct rpcrdma_rep, rr_recv_wr);
1591 			wr = wr->next;
1592 			rpcrdma_recv_buffer_put(rep);
1593 			--count;
1594 		}
1595 	}
1596 	ep->rep_receive_count += count;
1597 	return;
1598 }
1599