xref: /openbmc/linux/net/sunrpc/xprtrdma/verbs.c (revision 5388b581)
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 void rpcrdma_sendctx_put_locked(struct rpcrdma_xprt *r_xprt,
78 				       struct rpcrdma_sendctx *sc);
79 static void rpcrdma_reqs_reset(struct rpcrdma_xprt *r_xprt);
80 static void rpcrdma_reps_destroy(struct rpcrdma_buffer *buf);
81 static void rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt);
82 static void rpcrdma_mrs_destroy(struct rpcrdma_xprt *r_xprt);
83 static struct rpcrdma_regbuf *
84 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction,
85 		     gfp_t flags);
86 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb);
87 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb);
88 
89 /* Wait for outstanding transport work to finish. ib_drain_qp
90  * handles the drains in the wrong order for us, so open code
91  * them here.
92  */
93 static void rpcrdma_xprt_drain(struct rpcrdma_xprt *r_xprt)
94 {
95 	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
96 
97 	/* Flush Receives, then wait for deferred Reply work
98 	 * to complete.
99 	 */
100 	ib_drain_rq(ia->ri_id->qp);
101 
102 	/* Deferred Reply processing might have scheduled
103 	 * local invalidations.
104 	 */
105 	ib_drain_sq(ia->ri_id->qp);
106 }
107 
108 /**
109  * rpcrdma_qp_event_handler - Handle one QP event (error notification)
110  * @event: details of the event
111  * @context: ep that owns QP where event occurred
112  *
113  * Called from the RDMA provider (device driver) possibly in an interrupt
114  * context.
115  */
116 static void
117 rpcrdma_qp_event_handler(struct ib_event *event, void *context)
118 {
119 	struct rpcrdma_ep *ep = context;
120 	struct rpcrdma_xprt *r_xprt = container_of(ep, struct rpcrdma_xprt,
121 						   rx_ep);
122 
123 	trace_xprtrdma_qp_event(r_xprt, event);
124 }
125 
126 /**
127  * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC
128  * @cq:	completion queue
129  * @wc:	completed WR
130  *
131  */
132 static void
133 rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
134 {
135 	struct ib_cqe *cqe = wc->wr_cqe;
136 	struct rpcrdma_sendctx *sc =
137 		container_of(cqe, struct rpcrdma_sendctx, sc_cqe);
138 
139 	/* WARNING: Only wr_cqe and status are reliable at this point */
140 	trace_xprtrdma_wc_send(sc, wc);
141 	rpcrdma_sendctx_put_locked((struct rpcrdma_xprt *)cq->cq_context, sc);
142 }
143 
144 /**
145  * rpcrdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
146  * @cq:	completion queue (ignored)
147  * @wc:	completed WR
148  *
149  */
150 static void
151 rpcrdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
152 {
153 	struct ib_cqe *cqe = wc->wr_cqe;
154 	struct rpcrdma_rep *rep = container_of(cqe, struct rpcrdma_rep,
155 					       rr_cqe);
156 	struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
157 
158 	/* WARNING: Only wr_cqe and status are reliable at this point */
159 	trace_xprtrdma_wc_receive(wc);
160 	--r_xprt->rx_ep.rep_receive_count;
161 	if (wc->status != IB_WC_SUCCESS)
162 		goto out_flushed;
163 
164 	/* status == SUCCESS means all fields in wc are trustworthy */
165 	rpcrdma_set_xdrlen(&rep->rr_hdrbuf, wc->byte_len);
166 	rep->rr_wc_flags = wc->wc_flags;
167 	rep->rr_inv_rkey = wc->ex.invalidate_rkey;
168 
169 	ib_dma_sync_single_for_cpu(rdmab_device(rep->rr_rdmabuf),
170 				   rdmab_addr(rep->rr_rdmabuf),
171 				   wc->byte_len, DMA_FROM_DEVICE);
172 
173 	rpcrdma_reply_handler(rep);
174 	return;
175 
176 out_flushed:
177 	rpcrdma_recv_buffer_put(rep);
178 }
179 
180 static void rpcrdma_update_cm_private(struct rpcrdma_xprt *r_xprt,
181 				      struct rdma_conn_param *param)
182 {
183 	const struct rpcrdma_connect_private *pmsg = param->private_data;
184 	struct rpcrdma_ep *ep = &r_xprt->rx_ep;
185 	unsigned int rsize, wsize;
186 
187 	/* Default settings for RPC-over-RDMA Version One */
188 	r_xprt->rx_ia.ri_implicit_roundup = xprt_rdma_pad_optimize;
189 	rsize = RPCRDMA_V1_DEF_INLINE_SIZE;
190 	wsize = RPCRDMA_V1_DEF_INLINE_SIZE;
191 
192 	if (pmsg &&
193 	    pmsg->cp_magic == rpcrdma_cmp_magic &&
194 	    pmsg->cp_version == RPCRDMA_CMP_VERSION) {
195 		r_xprt->rx_ia.ri_implicit_roundup = true;
196 		rsize = rpcrdma_decode_buffer_size(pmsg->cp_send_size);
197 		wsize = rpcrdma_decode_buffer_size(pmsg->cp_recv_size);
198 	}
199 
200 	if (rsize < ep->rep_inline_recv)
201 		ep->rep_inline_recv = rsize;
202 	if (wsize < ep->rep_inline_send)
203 		ep->rep_inline_send = wsize;
204 
205 	rpcrdma_set_max_header_sizes(r_xprt);
206 }
207 
208 /**
209  * rpcrdma_cm_event_handler - Handle RDMA CM events
210  * @id: rdma_cm_id on which an event has occurred
211  * @event: details of the event
212  *
213  * Called with @id's mutex held. Returns 1 if caller should
214  * destroy @id, otherwise 0.
215  */
216 static int
217 rpcrdma_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
218 {
219 	struct rpcrdma_xprt *r_xprt = id->context;
220 	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
221 	struct rpcrdma_ep *ep = &r_xprt->rx_ep;
222 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
223 
224 	might_sleep();
225 
226 	trace_xprtrdma_cm_event(r_xprt, event);
227 	switch (event->event) {
228 	case RDMA_CM_EVENT_ADDR_RESOLVED:
229 	case RDMA_CM_EVENT_ROUTE_RESOLVED:
230 		ia->ri_async_rc = 0;
231 		complete(&ia->ri_done);
232 		return 0;
233 	case RDMA_CM_EVENT_ADDR_ERROR:
234 		ia->ri_async_rc = -EPROTO;
235 		complete(&ia->ri_done);
236 		return 0;
237 	case RDMA_CM_EVENT_ROUTE_ERROR:
238 		ia->ri_async_rc = -ENETUNREACH;
239 		complete(&ia->ri_done);
240 		return 0;
241 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
242 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
243 		pr_info("rpcrdma: removing device %s for %s:%s\n",
244 			ia->ri_id->device->name,
245 			rpcrdma_addrstr(r_xprt), rpcrdma_portstr(r_xprt));
246 #endif
247 		set_bit(RPCRDMA_IAF_REMOVING, &ia->ri_flags);
248 		ep->rep_connected = -ENODEV;
249 		xprt_force_disconnect(xprt);
250 		wait_for_completion(&ia->ri_remove_done);
251 
252 		ia->ri_id = NULL;
253 		/* Return 1 to ensure the core destroys the id. */
254 		return 1;
255 	case RDMA_CM_EVENT_ESTABLISHED:
256 		++xprt->connect_cookie;
257 		ep->rep_connected = 1;
258 		rpcrdma_update_cm_private(r_xprt, &event->param.conn);
259 		trace_xprtrdma_inline_thresh(r_xprt);
260 		wake_up_all(&ep->rep_connect_wait);
261 		break;
262 	case RDMA_CM_EVENT_CONNECT_ERROR:
263 		ep->rep_connected = -ENOTCONN;
264 		goto disconnected;
265 	case RDMA_CM_EVENT_UNREACHABLE:
266 		ep->rep_connected = -ENETUNREACH;
267 		goto disconnected;
268 	case RDMA_CM_EVENT_REJECTED:
269 		dprintk("rpcrdma: connection to %s:%s rejected: %s\n",
270 			rpcrdma_addrstr(r_xprt), rpcrdma_portstr(r_xprt),
271 			rdma_reject_msg(id, event->status));
272 		ep->rep_connected = -ECONNREFUSED;
273 		if (event->status == IB_CM_REJ_STALE_CONN)
274 			ep->rep_connected = -EAGAIN;
275 		goto disconnected;
276 	case RDMA_CM_EVENT_DISCONNECTED:
277 		ep->rep_connected = -ECONNABORTED;
278 disconnected:
279 		xprt_force_disconnect(xprt);
280 		wake_up_all(&ep->rep_connect_wait);
281 		break;
282 	default:
283 		break;
284 	}
285 
286 	dprintk("RPC:       %s: %s:%s on %s/frwr: %s\n", __func__,
287 		rpcrdma_addrstr(r_xprt), rpcrdma_portstr(r_xprt),
288 		ia->ri_id->device->name, rdma_event_msg(event->event));
289 	return 0;
290 }
291 
292 static struct rdma_cm_id *
293 rpcrdma_create_id(struct rpcrdma_xprt *xprt, struct rpcrdma_ia *ia)
294 {
295 	unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1;
296 	struct rdma_cm_id *id;
297 	int rc;
298 
299 	init_completion(&ia->ri_done);
300 	init_completion(&ia->ri_remove_done);
301 
302 	id = rdma_create_id(xprt->rx_xprt.xprt_net, rpcrdma_cm_event_handler,
303 			    xprt, RDMA_PS_TCP, IB_QPT_RC);
304 	if (IS_ERR(id))
305 		return id;
306 
307 	ia->ri_async_rc = -ETIMEDOUT;
308 	rc = rdma_resolve_addr(id, NULL,
309 			       (struct sockaddr *)&xprt->rx_xprt.addr,
310 			       RDMA_RESOLVE_TIMEOUT);
311 	if (rc)
312 		goto out;
313 	rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
314 	if (rc < 0)
315 		goto out;
316 
317 	rc = ia->ri_async_rc;
318 	if (rc)
319 		goto out;
320 
321 	ia->ri_async_rc = -ETIMEDOUT;
322 	rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
323 	if (rc)
324 		goto out;
325 	rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
326 	if (rc < 0)
327 		goto out;
328 	rc = ia->ri_async_rc;
329 	if (rc)
330 		goto out;
331 
332 	return id;
333 
334 out:
335 	rdma_destroy_id(id);
336 	return ERR_PTR(rc);
337 }
338 
339 /*
340  * Exported functions.
341  */
342 
343 /**
344  * rpcrdma_ia_open - Open and initialize an Interface Adapter.
345  * @xprt: transport with IA to (re)initialize
346  *
347  * Returns 0 on success, negative errno if an appropriate
348  * Interface Adapter could not be found and opened.
349  */
350 int
351 rpcrdma_ia_open(struct rpcrdma_xprt *xprt)
352 {
353 	struct rpcrdma_ia *ia = &xprt->rx_ia;
354 	int rc;
355 
356 	ia->ri_id = rpcrdma_create_id(xprt, ia);
357 	if (IS_ERR(ia->ri_id)) {
358 		rc = PTR_ERR(ia->ri_id);
359 		goto out_err;
360 	}
361 
362 	ia->ri_pd = ib_alloc_pd(ia->ri_id->device, 0);
363 	if (IS_ERR(ia->ri_pd)) {
364 		rc = PTR_ERR(ia->ri_pd);
365 		pr_err("rpcrdma: ib_alloc_pd() returned %d\n", rc);
366 		goto out_err;
367 	}
368 
369 	switch (xprt_rdma_memreg_strategy) {
370 	case RPCRDMA_FRWR:
371 		if (frwr_is_supported(ia->ri_id->device))
372 			break;
373 		/*FALLTHROUGH*/
374 	default:
375 		pr_err("rpcrdma: Device %s does not support memreg mode %d\n",
376 		       ia->ri_id->device->name, xprt_rdma_memreg_strategy);
377 		rc = -EINVAL;
378 		goto out_err;
379 	}
380 
381 	return 0;
382 
383 out_err:
384 	rpcrdma_ia_close(ia);
385 	return rc;
386 }
387 
388 /**
389  * rpcrdma_ia_remove - Handle device driver unload
390  * @ia: interface adapter being removed
391  *
392  * Divest transport H/W resources associated with this adapter,
393  * but allow it to be restored later.
394  */
395 void
396 rpcrdma_ia_remove(struct rpcrdma_ia *ia)
397 {
398 	struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
399 						   rx_ia);
400 	struct rpcrdma_ep *ep = &r_xprt->rx_ep;
401 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
402 	struct rpcrdma_req *req;
403 
404 	/* This is similar to rpcrdma_ep_destroy, but:
405 	 * - Don't cancel the connect worker.
406 	 * - Don't call rpcrdma_ep_disconnect, which waits
407 	 *   for another conn upcall, which will deadlock.
408 	 * - rdma_disconnect is unneeded, the underlying
409 	 *   connection is already gone.
410 	 */
411 	if (ia->ri_id->qp) {
412 		rpcrdma_xprt_drain(r_xprt);
413 		rdma_destroy_qp(ia->ri_id);
414 		ia->ri_id->qp = NULL;
415 	}
416 	ib_free_cq(ep->rep_attr.recv_cq);
417 	ep->rep_attr.recv_cq = NULL;
418 	ib_free_cq(ep->rep_attr.send_cq);
419 	ep->rep_attr.send_cq = NULL;
420 
421 	/* The ULP is responsible for ensuring all DMA
422 	 * mappings and MRs are gone.
423 	 */
424 	rpcrdma_reps_destroy(buf);
425 	list_for_each_entry(req, &buf->rb_allreqs, rl_all) {
426 		rpcrdma_regbuf_dma_unmap(req->rl_rdmabuf);
427 		rpcrdma_regbuf_dma_unmap(req->rl_sendbuf);
428 		rpcrdma_regbuf_dma_unmap(req->rl_recvbuf);
429 	}
430 	rpcrdma_mrs_destroy(r_xprt);
431 	ib_dealloc_pd(ia->ri_pd);
432 	ia->ri_pd = NULL;
433 
434 	/* Allow waiters to continue */
435 	complete(&ia->ri_remove_done);
436 
437 	trace_xprtrdma_remove(r_xprt);
438 }
439 
440 /**
441  * rpcrdma_ia_close - Clean up/close an IA.
442  * @ia: interface adapter to close
443  *
444  */
445 void
446 rpcrdma_ia_close(struct rpcrdma_ia *ia)
447 {
448 	if (ia->ri_id != NULL && !IS_ERR(ia->ri_id)) {
449 		if (ia->ri_id->qp)
450 			rdma_destroy_qp(ia->ri_id);
451 		rdma_destroy_id(ia->ri_id);
452 	}
453 	ia->ri_id = NULL;
454 
455 	/* If the pd is still busy, xprtrdma missed freeing a resource */
456 	if (ia->ri_pd && !IS_ERR(ia->ri_pd))
457 		ib_dealloc_pd(ia->ri_pd);
458 	ia->ri_pd = NULL;
459 }
460 
461 /**
462  * rpcrdma_ep_create - Create unconnected endpoint
463  * @r_xprt: transport to instantiate
464  *
465  * Returns zero on success, or a negative errno.
466  */
467 int rpcrdma_ep_create(struct rpcrdma_xprt *r_xprt)
468 {
469 	struct rpcrdma_ep *ep = &r_xprt->rx_ep;
470 	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
471 	struct rpcrdma_connect_private *pmsg = &ep->rep_cm_private;
472 	struct ib_cq *sendcq, *recvcq;
473 	unsigned int max_sge;
474 	int rc;
475 
476 	ep->rep_max_requests = xprt_rdma_slot_table_entries;
477 	ep->rep_inline_send = xprt_rdma_max_inline_write;
478 	ep->rep_inline_recv = xprt_rdma_max_inline_read;
479 
480 	max_sge = min_t(unsigned int, ia->ri_id->device->attrs.max_send_sge,
481 			RPCRDMA_MAX_SEND_SGES);
482 	if (max_sge < RPCRDMA_MIN_SEND_SGES) {
483 		pr_warn("rpcrdma: HCA provides only %d send SGEs\n", max_sge);
484 		return -ENOMEM;
485 	}
486 	ia->ri_max_send_sges = max_sge;
487 
488 	rc = frwr_open(ia, ep);
489 	if (rc)
490 		return rc;
491 
492 	ep->rep_attr.event_handler = rpcrdma_qp_event_handler;
493 	ep->rep_attr.qp_context = ep;
494 	ep->rep_attr.srq = NULL;
495 	ep->rep_attr.cap.max_send_sge = max_sge;
496 	ep->rep_attr.cap.max_recv_sge = 1;
497 	ep->rep_attr.cap.max_inline_data = 0;
498 	ep->rep_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
499 	ep->rep_attr.qp_type = IB_QPT_RC;
500 	ep->rep_attr.port_num = ~0;
501 
502 	dprintk("RPC:       %s: requested max: dtos: send %d recv %d; "
503 		"iovs: send %d recv %d\n",
504 		__func__,
505 		ep->rep_attr.cap.max_send_wr,
506 		ep->rep_attr.cap.max_recv_wr,
507 		ep->rep_attr.cap.max_send_sge,
508 		ep->rep_attr.cap.max_recv_sge);
509 
510 	ep->rep_send_batch = ep->rep_max_requests >> 3;
511 	ep->rep_send_count = ep->rep_send_batch;
512 	init_waitqueue_head(&ep->rep_connect_wait);
513 	ep->rep_receive_count = 0;
514 
515 	sendcq = ib_alloc_cq_any(ia->ri_id->device, r_xprt,
516 				 ep->rep_attr.cap.max_send_wr + 1,
517 				 IB_POLL_WORKQUEUE);
518 	if (IS_ERR(sendcq)) {
519 		rc = PTR_ERR(sendcq);
520 		goto out1;
521 	}
522 
523 	recvcq = ib_alloc_cq_any(ia->ri_id->device, NULL,
524 				 ep->rep_attr.cap.max_recv_wr + 1,
525 				 IB_POLL_WORKQUEUE);
526 	if (IS_ERR(recvcq)) {
527 		rc = PTR_ERR(recvcq);
528 		goto out2;
529 	}
530 
531 	ep->rep_attr.send_cq = sendcq;
532 	ep->rep_attr.recv_cq = recvcq;
533 
534 	/* Initialize cma parameters */
535 	memset(&ep->rep_remote_cma, 0, sizeof(ep->rep_remote_cma));
536 
537 	/* Prepare RDMA-CM private message */
538 	pmsg->cp_magic = rpcrdma_cmp_magic;
539 	pmsg->cp_version = RPCRDMA_CMP_VERSION;
540 	pmsg->cp_flags |= RPCRDMA_CMP_F_SND_W_INV_OK;
541 	pmsg->cp_send_size = rpcrdma_encode_buffer_size(ep->rep_inline_send);
542 	pmsg->cp_recv_size = rpcrdma_encode_buffer_size(ep->rep_inline_recv);
543 	ep->rep_remote_cma.private_data = pmsg;
544 	ep->rep_remote_cma.private_data_len = sizeof(*pmsg);
545 
546 	/* Client offers RDMA Read but does not initiate */
547 	ep->rep_remote_cma.initiator_depth = 0;
548 	ep->rep_remote_cma.responder_resources =
549 		min_t(int, U8_MAX, ia->ri_id->device->attrs.max_qp_rd_atom);
550 
551 	/* Limit transport retries so client can detect server
552 	 * GID changes quickly. RPC layer handles re-establishing
553 	 * transport connection and retransmission.
554 	 */
555 	ep->rep_remote_cma.retry_count = 6;
556 
557 	/* RPC-over-RDMA handles its own flow control. In addition,
558 	 * make all RNR NAKs visible so we know that RPC-over-RDMA
559 	 * flow control is working correctly (no NAKs should be seen).
560 	 */
561 	ep->rep_remote_cma.flow_control = 0;
562 	ep->rep_remote_cma.rnr_retry_count = 0;
563 
564 	return 0;
565 
566 out2:
567 	ib_free_cq(sendcq);
568 out1:
569 	return rc;
570 }
571 
572 /**
573  * rpcrdma_ep_destroy - Disconnect and destroy endpoint.
574  * @r_xprt: transport instance to shut down
575  *
576  */
577 void rpcrdma_ep_destroy(struct rpcrdma_xprt *r_xprt)
578 {
579 	struct rpcrdma_ep *ep = &r_xprt->rx_ep;
580 	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
581 
582 	if (ia->ri_id && ia->ri_id->qp) {
583 		rpcrdma_ep_disconnect(ep, ia);
584 		rdma_destroy_qp(ia->ri_id);
585 		ia->ri_id->qp = NULL;
586 	}
587 
588 	if (ep->rep_attr.recv_cq)
589 		ib_free_cq(ep->rep_attr.recv_cq);
590 	if (ep->rep_attr.send_cq)
591 		ib_free_cq(ep->rep_attr.send_cq);
592 }
593 
594 /* Re-establish a connection after a device removal event.
595  * Unlike a normal reconnection, a fresh PD and a new set
596  * of MRs and buffers is needed.
597  */
598 static int rpcrdma_ep_recreate_xprt(struct rpcrdma_xprt *r_xprt,
599 				    struct ib_qp_init_attr *qp_init_attr)
600 {
601 	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
602 	int rc, err;
603 
604 	trace_xprtrdma_reinsert(r_xprt);
605 
606 	rc = -EHOSTUNREACH;
607 	if (rpcrdma_ia_open(r_xprt))
608 		goto out1;
609 
610 	rc = -ENOMEM;
611 	err = rpcrdma_ep_create(r_xprt);
612 	if (err) {
613 		pr_err("rpcrdma: rpcrdma_ep_create returned %d\n", err);
614 		goto out2;
615 	}
616 
617 	rc = -ENETUNREACH;
618 	err = rdma_create_qp(ia->ri_id, ia->ri_pd, qp_init_attr);
619 	if (err) {
620 		pr_err("rpcrdma: rdma_create_qp returned %d\n", err);
621 		goto out3;
622 	}
623 	return 0;
624 
625 out3:
626 	rpcrdma_ep_destroy(r_xprt);
627 out2:
628 	rpcrdma_ia_close(ia);
629 out1:
630 	return rc;
631 }
632 
633 static int rpcrdma_ep_reconnect(struct rpcrdma_xprt *r_xprt,
634 				struct ib_qp_init_attr *qp_init_attr)
635 {
636 	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
637 	struct rdma_cm_id *id, *old;
638 	int err, rc;
639 
640 	rpcrdma_ep_disconnect(&r_xprt->rx_ep, ia);
641 
642 	rc = -EHOSTUNREACH;
643 	id = rpcrdma_create_id(r_xprt, ia);
644 	if (IS_ERR(id))
645 		goto out;
646 
647 	/* As long as the new ID points to the same device as the
648 	 * old ID, we can reuse the transport's existing PD and all
649 	 * previously allocated MRs. Also, the same device means
650 	 * the transport's previous DMA mappings are still valid.
651 	 *
652 	 * This is a sanity check only. There should be no way these
653 	 * point to two different devices here.
654 	 */
655 	old = id;
656 	rc = -ENETUNREACH;
657 	if (ia->ri_id->device != id->device) {
658 		pr_err("rpcrdma: can't reconnect on different device!\n");
659 		goto out_destroy;
660 	}
661 
662 	err = rdma_create_qp(id, ia->ri_pd, qp_init_attr);
663 	if (err)
664 		goto out_destroy;
665 
666 	/* Atomically replace the transport's ID and QP. */
667 	rc = 0;
668 	old = ia->ri_id;
669 	ia->ri_id = id;
670 	rdma_destroy_qp(old);
671 
672 out_destroy:
673 	rdma_destroy_id(old);
674 out:
675 	return rc;
676 }
677 
678 /*
679  * Connect unconnected endpoint.
680  */
681 int
682 rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
683 {
684 	struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
685 						   rx_ia);
686 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
687 	struct ib_qp_init_attr qp_init_attr;
688 	int rc;
689 
690 retry:
691 	memcpy(&qp_init_attr, &ep->rep_attr, sizeof(qp_init_attr));
692 	switch (ep->rep_connected) {
693 	case 0:
694 		rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &qp_init_attr);
695 		if (rc) {
696 			rc = -ENETUNREACH;
697 			goto out_noupdate;
698 		}
699 		break;
700 	case -ENODEV:
701 		rc = rpcrdma_ep_recreate_xprt(r_xprt, &qp_init_attr);
702 		if (rc)
703 			goto out_noupdate;
704 		break;
705 	default:
706 		rc = rpcrdma_ep_reconnect(r_xprt, &qp_init_attr);
707 		if (rc)
708 			goto out;
709 	}
710 
711 	ep->rep_connected = 0;
712 	xprt_clear_connected(xprt);
713 
714 	rpcrdma_reset_cwnd(r_xprt);
715 	rpcrdma_post_recvs(r_xprt, true);
716 
717 	rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma);
718 	if (rc)
719 		goto out;
720 
721 	if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
722 		xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
723 	wait_event_interruptible(ep->rep_connect_wait, ep->rep_connected != 0);
724 	if (ep->rep_connected <= 0) {
725 		if (ep->rep_connected == -EAGAIN)
726 			goto retry;
727 		rc = ep->rep_connected;
728 		goto out;
729 	}
730 
731 	rpcrdma_mrs_create(r_xprt);
732 
733 out:
734 	if (rc)
735 		ep->rep_connected = rc;
736 
737 out_noupdate:
738 	trace_xprtrdma_connect(r_xprt, rc);
739 	return rc;
740 }
741 
742 /**
743  * rpcrdma_ep_disconnect - Disconnect underlying transport
744  * @ep: endpoint to disconnect
745  * @ia: associated interface adapter
746  *
747  * Caller serializes. Either the transport send lock is held,
748  * or we're being called to destroy the transport.
749  */
750 void
751 rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
752 {
753 	struct rpcrdma_xprt *r_xprt = container_of(ep, struct rpcrdma_xprt,
754 						   rx_ep);
755 	int rc;
756 
757 	/* returns without wait if ID is not connected */
758 	rc = rdma_disconnect(ia->ri_id);
759 	if (!rc)
760 		wait_event_interruptible(ep->rep_connect_wait,
761 							ep->rep_connected != 1);
762 	else
763 		ep->rep_connected = rc;
764 	trace_xprtrdma_disconnect(r_xprt, rc);
765 
766 	rpcrdma_xprt_drain(r_xprt);
767 	rpcrdma_reqs_reset(r_xprt);
768 	rpcrdma_mrs_destroy(r_xprt);
769 }
770 
771 /* Fixed-size circular FIFO queue. This implementation is wait-free and
772  * lock-free.
773  *
774  * Consumer is the code path that posts Sends. This path dequeues a
775  * sendctx for use by a Send operation. Multiple consumer threads
776  * are serialized by the RPC transport lock, which allows only one
777  * ->send_request call at a time.
778  *
779  * Producer is the code path that handles Send completions. This path
780  * enqueues a sendctx that has been completed. Multiple producer
781  * threads are serialized by the ib_poll_cq() function.
782  */
783 
784 /* rpcrdma_sendctxs_destroy() assumes caller has already quiesced
785  * queue activity, and rpcrdma_xprt_drain has flushed all remaining
786  * Send requests.
787  */
788 static void rpcrdma_sendctxs_destroy(struct rpcrdma_buffer *buf)
789 {
790 	unsigned long i;
791 
792 	for (i = 0; i <= buf->rb_sc_last; i++)
793 		kfree(buf->rb_sc_ctxs[i]);
794 	kfree(buf->rb_sc_ctxs);
795 }
796 
797 static struct rpcrdma_sendctx *rpcrdma_sendctx_create(struct rpcrdma_ia *ia)
798 {
799 	struct rpcrdma_sendctx *sc;
800 
801 	sc = kzalloc(struct_size(sc, sc_sges, ia->ri_max_send_sges),
802 		     GFP_KERNEL);
803 	if (!sc)
804 		return NULL;
805 
806 	sc->sc_cqe.done = rpcrdma_wc_send;
807 	return sc;
808 }
809 
810 static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt)
811 {
812 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
813 	struct rpcrdma_sendctx *sc;
814 	unsigned long i;
815 
816 	/* Maximum number of concurrent outstanding Send WRs. Capping
817 	 * the circular queue size stops Send Queue overflow by causing
818 	 * the ->send_request call to fail temporarily before too many
819 	 * Sends are posted.
820 	 */
821 	i = buf->rb_max_requests + RPCRDMA_MAX_BC_REQUESTS;
822 	dprintk("RPC:       %s: allocating %lu send_ctxs\n", __func__, i);
823 	buf->rb_sc_ctxs = kcalloc(i, sizeof(sc), GFP_KERNEL);
824 	if (!buf->rb_sc_ctxs)
825 		return -ENOMEM;
826 
827 	buf->rb_sc_last = i - 1;
828 	for (i = 0; i <= buf->rb_sc_last; i++) {
829 		sc = rpcrdma_sendctx_create(&r_xprt->rx_ia);
830 		if (!sc)
831 			return -ENOMEM;
832 
833 		buf->rb_sc_ctxs[i] = sc;
834 	}
835 
836 	return 0;
837 }
838 
839 /* The sendctx queue is not guaranteed to have a size that is a
840  * power of two, thus the helpers in circ_buf.h cannot be used.
841  * The other option is to use modulus (%), which can be expensive.
842  */
843 static unsigned long rpcrdma_sendctx_next(struct rpcrdma_buffer *buf,
844 					  unsigned long item)
845 {
846 	return likely(item < buf->rb_sc_last) ? item + 1 : 0;
847 }
848 
849 /**
850  * rpcrdma_sendctx_get_locked - Acquire a send context
851  * @r_xprt: controlling transport instance
852  *
853  * Returns pointer to a free send completion context; or NULL if
854  * the queue is empty.
855  *
856  * Usage: Called to acquire an SGE array before preparing a Send WR.
857  *
858  * The caller serializes calls to this function (per transport), and
859  * provides an effective memory barrier that flushes the new value
860  * of rb_sc_head.
861  */
862 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_xprt *r_xprt)
863 {
864 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
865 	struct rpcrdma_sendctx *sc;
866 	unsigned long next_head;
867 
868 	next_head = rpcrdma_sendctx_next(buf, buf->rb_sc_head);
869 
870 	if (next_head == READ_ONCE(buf->rb_sc_tail))
871 		goto out_emptyq;
872 
873 	/* ORDER: item must be accessed _before_ head is updated */
874 	sc = buf->rb_sc_ctxs[next_head];
875 
876 	/* Releasing the lock in the caller acts as a memory
877 	 * barrier that flushes rb_sc_head.
878 	 */
879 	buf->rb_sc_head = next_head;
880 
881 	return sc;
882 
883 out_emptyq:
884 	/* The queue is "empty" if there have not been enough Send
885 	 * completions recently. This is a sign the Send Queue is
886 	 * backing up. Cause the caller to pause and try again.
887 	 */
888 	xprt_wait_for_buffer_space(&r_xprt->rx_xprt);
889 	r_xprt->rx_stats.empty_sendctx_q++;
890 	return NULL;
891 }
892 
893 /**
894  * rpcrdma_sendctx_put_locked - Release a send context
895  * @r_xprt: controlling transport instance
896  * @sc: send context to release
897  *
898  * Usage: Called from Send completion to return a sendctxt
899  * to the queue.
900  *
901  * The caller serializes calls to this function (per transport).
902  */
903 static void rpcrdma_sendctx_put_locked(struct rpcrdma_xprt *r_xprt,
904 				       struct rpcrdma_sendctx *sc)
905 {
906 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
907 	unsigned long next_tail;
908 
909 	/* Unmap SGEs of previously completed but unsignaled
910 	 * Sends by walking up the queue until @sc is found.
911 	 */
912 	next_tail = buf->rb_sc_tail;
913 	do {
914 		next_tail = rpcrdma_sendctx_next(buf, next_tail);
915 
916 		/* ORDER: item must be accessed _before_ tail is updated */
917 		rpcrdma_sendctx_unmap(buf->rb_sc_ctxs[next_tail]);
918 
919 	} while (buf->rb_sc_ctxs[next_tail] != sc);
920 
921 	/* Paired with READ_ONCE */
922 	smp_store_release(&buf->rb_sc_tail, next_tail);
923 
924 	xprt_write_space(&r_xprt->rx_xprt);
925 }
926 
927 static void
928 rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt)
929 {
930 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
931 	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
932 	unsigned int count;
933 
934 	for (count = 0; count < ia->ri_max_segs; count++) {
935 		struct rpcrdma_mr *mr;
936 		int rc;
937 
938 		mr = kzalloc(sizeof(*mr), GFP_NOFS);
939 		if (!mr)
940 			break;
941 
942 		rc = frwr_init_mr(ia, mr);
943 		if (rc) {
944 			kfree(mr);
945 			break;
946 		}
947 
948 		mr->mr_xprt = r_xprt;
949 
950 		spin_lock(&buf->rb_lock);
951 		rpcrdma_mr_push(mr, &buf->rb_mrs);
952 		list_add(&mr->mr_all, &buf->rb_all_mrs);
953 		spin_unlock(&buf->rb_lock);
954 	}
955 
956 	r_xprt->rx_stats.mrs_allocated += count;
957 	trace_xprtrdma_createmrs(r_xprt, count);
958 }
959 
960 static void
961 rpcrdma_mr_refresh_worker(struct work_struct *work)
962 {
963 	struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
964 						  rb_refresh_worker);
965 	struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
966 						   rx_buf);
967 
968 	rpcrdma_mrs_create(r_xprt);
969 	xprt_write_space(&r_xprt->rx_xprt);
970 }
971 
972 /**
973  * rpcrdma_mrs_refresh - Wake the MR refresh worker
974  * @r_xprt: controlling transport instance
975  *
976  */
977 void rpcrdma_mrs_refresh(struct rpcrdma_xprt *r_xprt)
978 {
979 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
980 	struct rpcrdma_ep *ep = &r_xprt->rx_ep;
981 
982 	/* If there is no underlying device, it's no use to
983 	 * wake the refresh worker.
984 	 */
985 	if (ep->rep_connected != -ENODEV) {
986 		/* The work is scheduled on a WQ_MEM_RECLAIM
987 		 * workqueue in order to prevent MR allocation
988 		 * from recursing into NFS during direct reclaim.
989 		 */
990 		queue_work(xprtiod_workqueue, &buf->rb_refresh_worker);
991 	}
992 }
993 
994 /**
995  * rpcrdma_req_create - Allocate an rpcrdma_req object
996  * @r_xprt: controlling r_xprt
997  * @size: initial size, in bytes, of send and receive buffers
998  * @flags: GFP flags passed to memory allocators
999  *
1000  * Returns an allocated and fully initialized rpcrdma_req or NULL.
1001  */
1002 struct rpcrdma_req *rpcrdma_req_create(struct rpcrdma_xprt *r_xprt, size_t size,
1003 				       gfp_t flags)
1004 {
1005 	struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
1006 	struct rpcrdma_regbuf *rb;
1007 	struct rpcrdma_req *req;
1008 	size_t maxhdrsize;
1009 
1010 	req = kzalloc(sizeof(*req), flags);
1011 	if (req == NULL)
1012 		goto out1;
1013 
1014 	/* Compute maximum header buffer size in bytes */
1015 	maxhdrsize = rpcrdma_fixed_maxsz + 3 +
1016 		     r_xprt->rx_ia.ri_max_segs * rpcrdma_readchunk_maxsz;
1017 	maxhdrsize *= sizeof(__be32);
1018 	rb = rpcrdma_regbuf_alloc(__roundup_pow_of_two(maxhdrsize),
1019 				  DMA_TO_DEVICE, flags);
1020 	if (!rb)
1021 		goto out2;
1022 	req->rl_rdmabuf = rb;
1023 	xdr_buf_init(&req->rl_hdrbuf, rdmab_data(rb), rdmab_length(rb));
1024 
1025 	req->rl_sendbuf = rpcrdma_regbuf_alloc(size, DMA_TO_DEVICE, flags);
1026 	if (!req->rl_sendbuf)
1027 		goto out3;
1028 
1029 	req->rl_recvbuf = rpcrdma_regbuf_alloc(size, DMA_NONE, flags);
1030 	if (!req->rl_recvbuf)
1031 		goto out4;
1032 
1033 	INIT_LIST_HEAD(&req->rl_free_mrs);
1034 	INIT_LIST_HEAD(&req->rl_registered);
1035 	spin_lock(&buffer->rb_lock);
1036 	list_add(&req->rl_all, &buffer->rb_allreqs);
1037 	spin_unlock(&buffer->rb_lock);
1038 	return req;
1039 
1040 out4:
1041 	kfree(req->rl_sendbuf);
1042 out3:
1043 	kfree(req->rl_rdmabuf);
1044 out2:
1045 	kfree(req);
1046 out1:
1047 	return NULL;
1048 }
1049 
1050 /**
1051  * rpcrdma_reqs_reset - Reset all reqs owned by a transport
1052  * @r_xprt: controlling transport instance
1053  *
1054  * ASSUMPTION: the rb_allreqs list is stable for the duration,
1055  * and thus can be walked without holding rb_lock. Eg. the
1056  * caller is holding the transport send lock to exclude
1057  * device removal or disconnection.
1058  */
1059 static void rpcrdma_reqs_reset(struct rpcrdma_xprt *r_xprt)
1060 {
1061 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1062 	struct rpcrdma_req *req;
1063 
1064 	list_for_each_entry(req, &buf->rb_allreqs, rl_all) {
1065 		/* Credits are valid only for one connection */
1066 		req->rl_slot.rq_cong = 0;
1067 	}
1068 }
1069 
1070 static struct rpcrdma_rep *rpcrdma_rep_create(struct rpcrdma_xprt *r_xprt,
1071 					      bool temp)
1072 {
1073 	struct rpcrdma_rep *rep;
1074 
1075 	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1076 	if (rep == NULL)
1077 		goto out;
1078 
1079 	rep->rr_rdmabuf = rpcrdma_regbuf_alloc(r_xprt->rx_ep.rep_inline_recv,
1080 					       DMA_FROM_DEVICE, GFP_KERNEL);
1081 	if (!rep->rr_rdmabuf)
1082 		goto out_free;
1083 
1084 	xdr_buf_init(&rep->rr_hdrbuf, rdmab_data(rep->rr_rdmabuf),
1085 		     rdmab_length(rep->rr_rdmabuf));
1086 	rep->rr_cqe.done = rpcrdma_wc_receive;
1087 	rep->rr_rxprt = r_xprt;
1088 	rep->rr_recv_wr.next = NULL;
1089 	rep->rr_recv_wr.wr_cqe = &rep->rr_cqe;
1090 	rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
1091 	rep->rr_recv_wr.num_sge = 1;
1092 	rep->rr_temp = temp;
1093 	return rep;
1094 
1095 out_free:
1096 	kfree(rep);
1097 out:
1098 	return NULL;
1099 }
1100 
1101 static void rpcrdma_rep_destroy(struct rpcrdma_rep *rep)
1102 {
1103 	rpcrdma_regbuf_free(rep->rr_rdmabuf);
1104 	kfree(rep);
1105 }
1106 
1107 static struct rpcrdma_rep *rpcrdma_rep_get_locked(struct rpcrdma_buffer *buf)
1108 {
1109 	struct llist_node *node;
1110 
1111 	/* Calls to llist_del_first are required to be serialized */
1112 	node = llist_del_first(&buf->rb_free_reps);
1113 	if (!node)
1114 		return NULL;
1115 	return llist_entry(node, struct rpcrdma_rep, rr_node);
1116 }
1117 
1118 static void rpcrdma_rep_put(struct rpcrdma_buffer *buf,
1119 			    struct rpcrdma_rep *rep)
1120 {
1121 	if (!rep->rr_temp)
1122 		llist_add(&rep->rr_node, &buf->rb_free_reps);
1123 	else
1124 		rpcrdma_rep_destroy(rep);
1125 }
1126 
1127 static void rpcrdma_reps_destroy(struct rpcrdma_buffer *buf)
1128 {
1129 	struct rpcrdma_rep *rep;
1130 
1131 	while ((rep = rpcrdma_rep_get_locked(buf)) != NULL)
1132 		rpcrdma_rep_destroy(rep);
1133 }
1134 
1135 /**
1136  * rpcrdma_buffer_create - Create initial set of req/rep objects
1137  * @r_xprt: transport instance to (re)initialize
1138  *
1139  * Returns zero on success, otherwise a negative errno.
1140  */
1141 int rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
1142 {
1143 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1144 	int i, rc;
1145 
1146 	buf->rb_max_requests = r_xprt->rx_ep.rep_max_requests;
1147 	buf->rb_bc_srv_max_requests = 0;
1148 	spin_lock_init(&buf->rb_lock);
1149 	INIT_LIST_HEAD(&buf->rb_mrs);
1150 	INIT_LIST_HEAD(&buf->rb_all_mrs);
1151 	INIT_WORK(&buf->rb_refresh_worker, rpcrdma_mr_refresh_worker);
1152 
1153 	INIT_LIST_HEAD(&buf->rb_send_bufs);
1154 	INIT_LIST_HEAD(&buf->rb_allreqs);
1155 
1156 	rc = -ENOMEM;
1157 	for (i = 0; i < buf->rb_max_requests; i++) {
1158 		struct rpcrdma_req *req;
1159 
1160 		req = rpcrdma_req_create(r_xprt, RPCRDMA_V1_DEF_INLINE_SIZE * 2,
1161 					 GFP_KERNEL);
1162 		if (!req)
1163 			goto out;
1164 		list_add(&req->rl_list, &buf->rb_send_bufs);
1165 	}
1166 
1167 	init_llist_head(&buf->rb_free_reps);
1168 
1169 	rc = rpcrdma_sendctxs_create(r_xprt);
1170 	if (rc)
1171 		goto out;
1172 
1173 	return 0;
1174 out:
1175 	rpcrdma_buffer_destroy(buf);
1176 	return rc;
1177 }
1178 
1179 /**
1180  * rpcrdma_req_destroy - Destroy an rpcrdma_req object
1181  * @req: unused object to be destroyed
1182  *
1183  * Relies on caller holding the transport send lock to protect
1184  * removing req->rl_all from buf->rb_all_reqs safely.
1185  */
1186 void rpcrdma_req_destroy(struct rpcrdma_req *req)
1187 {
1188 	struct rpcrdma_mr *mr;
1189 
1190 	list_del(&req->rl_all);
1191 
1192 	while ((mr = rpcrdma_mr_pop(&req->rl_free_mrs))) {
1193 		struct rpcrdma_buffer *buf = &mr->mr_xprt->rx_buf;
1194 
1195 		spin_lock(&buf->rb_lock);
1196 		list_del(&mr->mr_all);
1197 		spin_unlock(&buf->rb_lock);
1198 
1199 		frwr_release_mr(mr);
1200 	}
1201 
1202 	rpcrdma_regbuf_free(req->rl_recvbuf);
1203 	rpcrdma_regbuf_free(req->rl_sendbuf);
1204 	rpcrdma_regbuf_free(req->rl_rdmabuf);
1205 	kfree(req);
1206 }
1207 
1208 /**
1209  * rpcrdma_mrs_destroy - Release all of a transport's MRs
1210  * @r_xprt: controlling transport instance
1211  *
1212  * Relies on caller holding the transport send lock to protect
1213  * removing mr->mr_list from req->rl_free_mrs safely.
1214  */
1215 static void rpcrdma_mrs_destroy(struct rpcrdma_xprt *r_xprt)
1216 {
1217 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1218 	struct rpcrdma_mr *mr;
1219 
1220 	cancel_work_sync(&buf->rb_refresh_worker);
1221 
1222 	spin_lock(&buf->rb_lock);
1223 	while ((mr = list_first_entry_or_null(&buf->rb_all_mrs,
1224 					      struct rpcrdma_mr,
1225 					      mr_all)) != NULL) {
1226 		list_del(&mr->mr_list);
1227 		list_del(&mr->mr_all);
1228 		spin_unlock(&buf->rb_lock);
1229 
1230 		frwr_release_mr(mr);
1231 
1232 		spin_lock(&buf->rb_lock);
1233 	}
1234 	spin_unlock(&buf->rb_lock);
1235 }
1236 
1237 /**
1238  * rpcrdma_buffer_destroy - Release all hw resources
1239  * @buf: root control block for resources
1240  *
1241  * ORDERING: relies on a prior rpcrdma_xprt_drain :
1242  * - No more Send or Receive completions can occur
1243  * - All MRs, reps, and reqs are returned to their free lists
1244  */
1245 void
1246 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1247 {
1248 	rpcrdma_sendctxs_destroy(buf);
1249 	rpcrdma_reps_destroy(buf);
1250 
1251 	while (!list_empty(&buf->rb_send_bufs)) {
1252 		struct rpcrdma_req *req;
1253 
1254 		req = list_first_entry(&buf->rb_send_bufs,
1255 				       struct rpcrdma_req, rl_list);
1256 		list_del(&req->rl_list);
1257 		rpcrdma_req_destroy(req);
1258 	}
1259 }
1260 
1261 /**
1262  * rpcrdma_mr_get - Allocate an rpcrdma_mr object
1263  * @r_xprt: controlling transport
1264  *
1265  * Returns an initialized rpcrdma_mr or NULL if no free
1266  * rpcrdma_mr objects are available.
1267  */
1268 struct rpcrdma_mr *
1269 rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt)
1270 {
1271 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1272 	struct rpcrdma_mr *mr;
1273 
1274 	spin_lock(&buf->rb_lock);
1275 	mr = rpcrdma_mr_pop(&buf->rb_mrs);
1276 	spin_unlock(&buf->rb_lock);
1277 	return mr;
1278 }
1279 
1280 /**
1281  * rpcrdma_mr_put - DMA unmap an MR and release it
1282  * @mr: MR to release
1283  *
1284  */
1285 void rpcrdma_mr_put(struct rpcrdma_mr *mr)
1286 {
1287 	struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
1288 
1289 	if (mr->mr_dir != DMA_NONE) {
1290 		trace_xprtrdma_mr_unmap(mr);
1291 		ib_dma_unmap_sg(r_xprt->rx_ia.ri_id->device,
1292 				mr->mr_sg, mr->mr_nents, mr->mr_dir);
1293 		mr->mr_dir = DMA_NONE;
1294 	}
1295 
1296 	rpcrdma_mr_push(mr, &mr->mr_req->rl_free_mrs);
1297 }
1298 
1299 /**
1300  * rpcrdma_buffer_get - Get a request buffer
1301  * @buffers: Buffer pool from which to obtain a buffer
1302  *
1303  * Returns a fresh rpcrdma_req, or NULL if none are available.
1304  */
1305 struct rpcrdma_req *
1306 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1307 {
1308 	struct rpcrdma_req *req;
1309 
1310 	spin_lock(&buffers->rb_lock);
1311 	req = list_first_entry_or_null(&buffers->rb_send_bufs,
1312 				       struct rpcrdma_req, rl_list);
1313 	if (req)
1314 		list_del_init(&req->rl_list);
1315 	spin_unlock(&buffers->rb_lock);
1316 	return req;
1317 }
1318 
1319 /**
1320  * rpcrdma_buffer_put - Put request/reply buffers back into pool
1321  * @buffers: buffer pool
1322  * @req: object to return
1323  *
1324  */
1325 void rpcrdma_buffer_put(struct rpcrdma_buffer *buffers, struct rpcrdma_req *req)
1326 {
1327 	if (req->rl_reply)
1328 		rpcrdma_rep_put(buffers, req->rl_reply);
1329 	req->rl_reply = NULL;
1330 
1331 	spin_lock(&buffers->rb_lock);
1332 	list_add(&req->rl_list, &buffers->rb_send_bufs);
1333 	spin_unlock(&buffers->rb_lock);
1334 }
1335 
1336 /**
1337  * rpcrdma_recv_buffer_put - Release rpcrdma_rep back to free list
1338  * @rep: rep to release
1339  *
1340  * Used after error conditions.
1341  */
1342 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1343 {
1344 	rpcrdma_rep_put(&rep->rr_rxprt->rx_buf, rep);
1345 }
1346 
1347 /* Returns a pointer to a rpcrdma_regbuf object, or NULL.
1348  *
1349  * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1350  * receiving the payload of RDMA RECV operations. During Long Calls
1351  * or Replies they may be registered externally via frwr_map.
1352  */
1353 static struct rpcrdma_regbuf *
1354 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction,
1355 		     gfp_t flags)
1356 {
1357 	struct rpcrdma_regbuf *rb;
1358 
1359 	rb = kmalloc(sizeof(*rb), flags);
1360 	if (!rb)
1361 		return NULL;
1362 	rb->rg_data = kmalloc(size, flags);
1363 	if (!rb->rg_data) {
1364 		kfree(rb);
1365 		return NULL;
1366 	}
1367 
1368 	rb->rg_device = NULL;
1369 	rb->rg_direction = direction;
1370 	rb->rg_iov.length = size;
1371 	return rb;
1372 }
1373 
1374 /**
1375  * rpcrdma_regbuf_realloc - re-allocate a SEND/RECV buffer
1376  * @rb: regbuf to reallocate
1377  * @size: size of buffer to be allocated, in bytes
1378  * @flags: GFP flags
1379  *
1380  * Returns true if reallocation was successful. If false is
1381  * returned, @rb is left untouched.
1382  */
1383 bool rpcrdma_regbuf_realloc(struct rpcrdma_regbuf *rb, size_t size, gfp_t flags)
1384 {
1385 	void *buf;
1386 
1387 	buf = kmalloc(size, flags);
1388 	if (!buf)
1389 		return false;
1390 
1391 	rpcrdma_regbuf_dma_unmap(rb);
1392 	kfree(rb->rg_data);
1393 
1394 	rb->rg_data = buf;
1395 	rb->rg_iov.length = size;
1396 	return true;
1397 }
1398 
1399 /**
1400  * __rpcrdma_regbuf_dma_map - DMA-map a regbuf
1401  * @r_xprt: controlling transport instance
1402  * @rb: regbuf to be mapped
1403  *
1404  * Returns true if the buffer is now DMA mapped to @r_xprt's device
1405  */
1406 bool __rpcrdma_regbuf_dma_map(struct rpcrdma_xprt *r_xprt,
1407 			      struct rpcrdma_regbuf *rb)
1408 {
1409 	struct ib_device *device = r_xprt->rx_ia.ri_id->device;
1410 
1411 	if (rb->rg_direction == DMA_NONE)
1412 		return false;
1413 
1414 	rb->rg_iov.addr = ib_dma_map_single(device, rdmab_data(rb),
1415 					    rdmab_length(rb), rb->rg_direction);
1416 	if (ib_dma_mapping_error(device, rdmab_addr(rb))) {
1417 		trace_xprtrdma_dma_maperr(rdmab_addr(rb));
1418 		return false;
1419 	}
1420 
1421 	rb->rg_device = device;
1422 	rb->rg_iov.lkey = r_xprt->rx_ia.ri_pd->local_dma_lkey;
1423 	return true;
1424 }
1425 
1426 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb)
1427 {
1428 	if (!rb)
1429 		return;
1430 
1431 	if (!rpcrdma_regbuf_is_mapped(rb))
1432 		return;
1433 
1434 	ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb), rdmab_length(rb),
1435 			    rb->rg_direction);
1436 	rb->rg_device = NULL;
1437 }
1438 
1439 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb)
1440 {
1441 	rpcrdma_regbuf_dma_unmap(rb);
1442 	if (rb)
1443 		kfree(rb->rg_data);
1444 	kfree(rb);
1445 }
1446 
1447 /**
1448  * rpcrdma_ep_post - Post WRs to a transport's Send Queue
1449  * @ia: transport's device information
1450  * @ep: transport's RDMA endpoint information
1451  * @req: rpcrdma_req containing the Send WR to post
1452  *
1453  * Returns 0 if the post was successful, otherwise -ENOTCONN
1454  * is returned.
1455  */
1456 int
1457 rpcrdma_ep_post(struct rpcrdma_ia *ia,
1458 		struct rpcrdma_ep *ep,
1459 		struct rpcrdma_req *req)
1460 {
1461 	struct ib_send_wr *send_wr = &req->rl_wr;
1462 	int rc;
1463 
1464 	if (!ep->rep_send_count || kref_read(&req->rl_kref) > 1) {
1465 		send_wr->send_flags |= IB_SEND_SIGNALED;
1466 		ep->rep_send_count = ep->rep_send_batch;
1467 	} else {
1468 		send_wr->send_flags &= ~IB_SEND_SIGNALED;
1469 		--ep->rep_send_count;
1470 	}
1471 
1472 	rc = frwr_send(ia, req);
1473 	trace_xprtrdma_post_send(req, rc);
1474 	if (rc)
1475 		return -ENOTCONN;
1476 	return 0;
1477 }
1478 
1479 /**
1480  * rpcrdma_post_recvs - Refill the Receive Queue
1481  * @r_xprt: controlling transport instance
1482  * @temp: mark Receive buffers to be deleted after use
1483  *
1484  */
1485 void rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, bool temp)
1486 {
1487 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1488 	struct rpcrdma_ep *ep = &r_xprt->rx_ep;
1489 	struct ib_recv_wr *i, *wr, *bad_wr;
1490 	struct rpcrdma_rep *rep;
1491 	int needed, count, rc;
1492 
1493 	rc = 0;
1494 	count = 0;
1495 
1496 	needed = buf->rb_credits + (buf->rb_bc_srv_max_requests << 1);
1497 	if (likely(ep->rep_receive_count > needed))
1498 		goto out;
1499 	needed -= ep->rep_receive_count;
1500 	if (!temp)
1501 		needed += RPCRDMA_MAX_RECV_BATCH;
1502 
1503 	/* fast path: all needed reps can be found on the free list */
1504 	wr = NULL;
1505 	while (needed) {
1506 		rep = rpcrdma_rep_get_locked(buf);
1507 		if (!rep)
1508 			rep = rpcrdma_rep_create(r_xprt, temp);
1509 		if (!rep)
1510 			break;
1511 
1512 		rep->rr_recv_wr.next = wr;
1513 		wr = &rep->rr_recv_wr;
1514 		--needed;
1515 	}
1516 	if (!wr)
1517 		goto out;
1518 
1519 	for (i = wr; i; i = i->next) {
1520 		rep = container_of(i, struct rpcrdma_rep, rr_recv_wr);
1521 
1522 		if (!rpcrdma_regbuf_dma_map(r_xprt, rep->rr_rdmabuf))
1523 			goto release_wrs;
1524 
1525 		trace_xprtrdma_post_recv(rep);
1526 		++count;
1527 	}
1528 
1529 	rc = ib_post_recv(r_xprt->rx_ia.ri_id->qp, wr,
1530 			  (const struct ib_recv_wr **)&bad_wr);
1531 out:
1532 	trace_xprtrdma_post_recvs(r_xprt, count, rc);
1533 	if (rc) {
1534 		for (wr = bad_wr; wr;) {
1535 			struct rpcrdma_rep *rep;
1536 
1537 			rep = container_of(wr, struct rpcrdma_rep, rr_recv_wr);
1538 			wr = wr->next;
1539 			rpcrdma_recv_buffer_put(rep);
1540 			--count;
1541 		}
1542 	}
1543 	ep->rep_receive_count += count;
1544 	return;
1545 
1546 release_wrs:
1547 	for (i = wr; i;) {
1548 		rep = container_of(i, struct rpcrdma_rep, rr_recv_wr);
1549 		i = i->next;
1550 		rpcrdma_recv_buffer_put(rep);
1551 	}
1552 }
1553