• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
2 /*
3  * Copyright (c) 2014-2017 Oracle.  All rights reserved.
4  * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
5  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the BSD-type
10  * license below:
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  *
16  *      Redistributions of source code must retain the above copyright
17  *      notice, this list of conditions and the following disclaimer.
18  *
19  *      Redistributions in binary form must reproduce the above
20  *      copyright notice, this list of conditions and the following
21  *      disclaimer in the documentation and/or other materials provided
22  *      with the distribution.
23  *
24  *      Neither the name of the Network Appliance, Inc. nor the names of
25  *      its contributors may be used to endorse or promote products
26  *      derived from this software without specific prior written
27  *      permission.
28  *
29  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
30  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
31  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
32  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
33  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
34  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
35  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
36  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
37  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
38  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
39  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
40  */
41 
42 /*
43  * verbs.c
44  *
45  * Encapsulates the major functions managing:
46  *  o adapters
47  *  o endpoints
48  *  o connections
49  *  o buffer memory
50  */
51 
52 #include <linux/interrupt.h>
53 #include <linux/slab.h>
54 #include <linux/sunrpc/addr.h>
55 #include <linux/sunrpc/svc_rdma.h>
56 #include <linux/log2.h>
57 
58 #include <asm-generic/barrier.h>
59 #include <asm/bitops.h>
60 
61 #include <rdma/ib_cm.h>
62 
63 #include "xprt_rdma.h"
64 #include <trace/events/rpcrdma.h>
65 
66 /*
67  * Globals/Macros
68  */
69 
70 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
71 # define RPCDBG_FACILITY	RPCDBG_TRANS
72 #endif
73 
74 /*
75  * internal functions
76  */
77 static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt);
78 static void rpcrdma_sendctxs_destroy(struct rpcrdma_xprt *r_xprt);
79 static void rpcrdma_sendctx_put_locked(struct rpcrdma_xprt *r_xprt,
80 				       struct rpcrdma_sendctx *sc);
81 static int rpcrdma_reqs_setup(struct rpcrdma_xprt *r_xprt);
82 static void rpcrdma_reqs_reset(struct rpcrdma_xprt *r_xprt);
83 static void rpcrdma_rep_destroy(struct rpcrdma_rep *rep);
84 static void rpcrdma_reps_unmap(struct rpcrdma_xprt *r_xprt);
85 static void rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt);
86 static void rpcrdma_mrs_destroy(struct rpcrdma_xprt *r_xprt);
87 static void rpcrdma_ep_get(struct rpcrdma_ep *ep);
88 static int rpcrdma_ep_put(struct rpcrdma_ep *ep);
89 static struct rpcrdma_regbuf *
90 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction,
91 		     gfp_t flags);
92 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb);
93 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb);
94 
95 /* Wait for outstanding transport work to finish. ib_drain_qp
96  * handles the drains in the wrong order for us, so open code
97  * them here.
98  */
rpcrdma_xprt_drain(struct rpcrdma_xprt * r_xprt)99 static void rpcrdma_xprt_drain(struct rpcrdma_xprt *r_xprt)
100 {
101 	struct rpcrdma_ep *ep = r_xprt->rx_ep;
102 	struct rdma_cm_id *id = ep->re_id;
103 
104 	/* Flush Receives, then wait for deferred Reply work
105 	 * to complete.
106 	 */
107 	ib_drain_rq(id->qp);
108 
109 	/* Deferred Reply processing might have scheduled
110 	 * local invalidations.
111 	 */
112 	ib_drain_sq(id->qp);
113 
114 	rpcrdma_ep_put(ep);
115 }
116 
117 /**
118  * rpcrdma_qp_event_handler - Handle one QP event (error notification)
119  * @event: details of the event
120  * @context: ep that owns QP where event occurred
121  *
122  * Called from the RDMA provider (device driver) possibly in an interrupt
123  * context. The QP is always destroyed before the ID, so the ID will be
124  * reliably available when this handler is invoked.
125  */
rpcrdma_qp_event_handler(struct ib_event * event,void * context)126 static void rpcrdma_qp_event_handler(struct ib_event *event, void *context)
127 {
128 	struct rpcrdma_ep *ep = context;
129 
130 	trace_xprtrdma_qp_event(ep, event);
131 }
132 
133 /* Ensure xprt_force_disconnect() is invoked exactly once when a
134  * connection is closed or lost. (The important thing is it needs
135  * to be invoked "at least" once).
136  */
rpcrdma_force_disconnect(struct rpcrdma_ep * ep)137 static void rpcrdma_force_disconnect(struct rpcrdma_ep *ep)
138 {
139 	if (atomic_add_unless(&ep->re_force_disconnect, 1, 1))
140 		xprt_force_disconnect(ep->re_xprt);
141 }
142 
143 /**
144  * rpcrdma_flush_disconnect - Disconnect on flushed completion
145  * @r_xprt: transport to disconnect
146  * @wc: work completion entry
147  *
148  * Must be called in process context.
149  */
rpcrdma_flush_disconnect(struct rpcrdma_xprt * r_xprt,struct ib_wc * wc)150 void rpcrdma_flush_disconnect(struct rpcrdma_xprt *r_xprt, struct ib_wc *wc)
151 {
152 	if (wc->status != IB_WC_SUCCESS)
153 		rpcrdma_force_disconnect(r_xprt->rx_ep);
154 }
155 
156 /**
157  * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC
158  * @cq:	completion queue
159  * @wc:	WCE for a completed Send WR
160  *
161  */
rpcrdma_wc_send(struct ib_cq * cq,struct ib_wc * wc)162 static void rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
163 {
164 	struct ib_cqe *cqe = wc->wr_cqe;
165 	struct rpcrdma_sendctx *sc =
166 		container_of(cqe, struct rpcrdma_sendctx, sc_cqe);
167 	struct rpcrdma_xprt *r_xprt = cq->cq_context;
168 
169 	/* WARNING: Only wr_cqe and status are reliable at this point */
170 	trace_xprtrdma_wc_send(sc, wc);
171 	rpcrdma_sendctx_put_locked(r_xprt, sc);
172 	rpcrdma_flush_disconnect(r_xprt, wc);
173 }
174 
175 /**
176  * rpcrdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
177  * @cq:	completion queue
178  * @wc:	WCE for a completed Receive WR
179  *
180  */
rpcrdma_wc_receive(struct ib_cq * cq,struct ib_wc * wc)181 static void rpcrdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
182 {
183 	struct ib_cqe *cqe = wc->wr_cqe;
184 	struct rpcrdma_rep *rep = container_of(cqe, struct rpcrdma_rep,
185 					       rr_cqe);
186 	struct rpcrdma_xprt *r_xprt = cq->cq_context;
187 
188 	/* WARNING: Only wr_cqe and status are reliable at this point */
189 	trace_xprtrdma_wc_receive(wc);
190 	--r_xprt->rx_ep->re_receive_count;
191 	if (wc->status != IB_WC_SUCCESS)
192 		goto out_flushed;
193 
194 	/* status == SUCCESS means all fields in wc are trustworthy */
195 	rpcrdma_set_xdrlen(&rep->rr_hdrbuf, wc->byte_len);
196 	rep->rr_wc_flags = wc->wc_flags;
197 	rep->rr_inv_rkey = wc->ex.invalidate_rkey;
198 
199 	ib_dma_sync_single_for_cpu(rdmab_device(rep->rr_rdmabuf),
200 				   rdmab_addr(rep->rr_rdmabuf),
201 				   wc->byte_len, DMA_FROM_DEVICE);
202 
203 	rpcrdma_reply_handler(rep);
204 	return;
205 
206 out_flushed:
207 	rpcrdma_flush_disconnect(r_xprt, wc);
208 	rpcrdma_rep_destroy(rep);
209 }
210 
rpcrdma_update_cm_private(struct rpcrdma_ep * ep,struct rdma_conn_param * param)211 static void rpcrdma_update_cm_private(struct rpcrdma_ep *ep,
212 				      struct rdma_conn_param *param)
213 {
214 	const struct rpcrdma_connect_private *pmsg = param->private_data;
215 	unsigned int rsize, wsize;
216 
217 	/* Default settings for RPC-over-RDMA Version One */
218 	ep->re_implicit_roundup = xprt_rdma_pad_optimize;
219 	rsize = RPCRDMA_V1_DEF_INLINE_SIZE;
220 	wsize = RPCRDMA_V1_DEF_INLINE_SIZE;
221 
222 	if (pmsg &&
223 	    pmsg->cp_magic == rpcrdma_cmp_magic &&
224 	    pmsg->cp_version == RPCRDMA_CMP_VERSION) {
225 		ep->re_implicit_roundup = true;
226 		rsize = rpcrdma_decode_buffer_size(pmsg->cp_send_size);
227 		wsize = rpcrdma_decode_buffer_size(pmsg->cp_recv_size);
228 	}
229 
230 	if (rsize < ep->re_inline_recv)
231 		ep->re_inline_recv = rsize;
232 	if (wsize < ep->re_inline_send)
233 		ep->re_inline_send = wsize;
234 
235 	rpcrdma_set_max_header_sizes(ep);
236 }
237 
238 /**
239  * rpcrdma_cm_event_handler - Handle RDMA CM events
240  * @id: rdma_cm_id on which an event has occurred
241  * @event: details of the event
242  *
243  * Called with @id's mutex held. Returns 1 if caller should
244  * destroy @id, otherwise 0.
245  */
246 static int
rpcrdma_cm_event_handler(struct rdma_cm_id * id,struct rdma_cm_event * event)247 rpcrdma_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
248 {
249 	struct sockaddr *sap = (struct sockaddr *)&id->route.addr.dst_addr;
250 	struct rpcrdma_ep *ep = id->context;
251 
252 	might_sleep();
253 
254 	switch (event->event) {
255 	case RDMA_CM_EVENT_ADDR_RESOLVED:
256 	case RDMA_CM_EVENT_ROUTE_RESOLVED:
257 		ep->re_async_rc = 0;
258 		complete(&ep->re_done);
259 		return 0;
260 	case RDMA_CM_EVENT_ADDR_ERROR:
261 		ep->re_async_rc = -EPROTO;
262 		complete(&ep->re_done);
263 		return 0;
264 	case RDMA_CM_EVENT_ROUTE_ERROR:
265 		ep->re_async_rc = -ENETUNREACH;
266 		complete(&ep->re_done);
267 		return 0;
268 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
269 		pr_info("rpcrdma: removing device %s for %pISpc\n",
270 			ep->re_id->device->name, sap);
271 		fallthrough;
272 	case RDMA_CM_EVENT_ADDR_CHANGE:
273 		ep->re_connect_status = -ENODEV;
274 		goto disconnected;
275 	case RDMA_CM_EVENT_ESTABLISHED:
276 		rpcrdma_ep_get(ep);
277 		ep->re_connect_status = 1;
278 		rpcrdma_update_cm_private(ep, &event->param.conn);
279 		trace_xprtrdma_inline_thresh(ep);
280 		wake_up_all(&ep->re_connect_wait);
281 		break;
282 	case RDMA_CM_EVENT_CONNECT_ERROR:
283 		ep->re_connect_status = -ENOTCONN;
284 		goto wake_connect_worker;
285 	case RDMA_CM_EVENT_UNREACHABLE:
286 		ep->re_connect_status = -ENETUNREACH;
287 		goto wake_connect_worker;
288 	case RDMA_CM_EVENT_REJECTED:
289 		dprintk("rpcrdma: connection to %pISpc rejected: %s\n",
290 			sap, rdma_reject_msg(id, event->status));
291 		ep->re_connect_status = -ECONNREFUSED;
292 		if (event->status == IB_CM_REJ_STALE_CONN)
293 			ep->re_connect_status = -ENOTCONN;
294 wake_connect_worker:
295 		wake_up_all(&ep->re_connect_wait);
296 		return 0;
297 	case RDMA_CM_EVENT_DISCONNECTED:
298 		ep->re_connect_status = -ECONNABORTED;
299 disconnected:
300 		rpcrdma_force_disconnect(ep);
301 		return rpcrdma_ep_put(ep);
302 	default:
303 		break;
304 	}
305 
306 	dprintk("RPC:       %s: %pISpc on %s/frwr: %s\n", __func__, sap,
307 		ep->re_id->device->name, rdma_event_msg(event->event));
308 	return 0;
309 }
310 
rpcrdma_create_id(struct rpcrdma_xprt * r_xprt,struct rpcrdma_ep * ep)311 static struct rdma_cm_id *rpcrdma_create_id(struct rpcrdma_xprt *r_xprt,
312 					    struct rpcrdma_ep *ep)
313 {
314 	unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1;
315 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
316 	struct rdma_cm_id *id;
317 	int rc;
318 
319 	init_completion(&ep->re_done);
320 
321 	id = rdma_create_id(xprt->xprt_net, rpcrdma_cm_event_handler, ep,
322 			    RDMA_PS_TCP, IB_QPT_RC);
323 	if (IS_ERR(id))
324 		return id;
325 
326 	ep->re_async_rc = -ETIMEDOUT;
327 	rc = rdma_resolve_addr(id, NULL, (struct sockaddr *)&xprt->addr,
328 			       RDMA_RESOLVE_TIMEOUT);
329 	if (rc)
330 		goto out;
331 	rc = wait_for_completion_interruptible_timeout(&ep->re_done, wtimeout);
332 	if (rc < 0)
333 		goto out;
334 
335 	rc = ep->re_async_rc;
336 	if (rc)
337 		goto out;
338 
339 	ep->re_async_rc = -ETIMEDOUT;
340 	rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
341 	if (rc)
342 		goto out;
343 	rc = wait_for_completion_interruptible_timeout(&ep->re_done, wtimeout);
344 	if (rc < 0)
345 		goto out;
346 	rc = ep->re_async_rc;
347 	if (rc)
348 		goto out;
349 
350 	return id;
351 
352 out:
353 	rdma_destroy_id(id);
354 	return ERR_PTR(rc);
355 }
356 
rpcrdma_ep_destroy(struct kref * kref)357 static void rpcrdma_ep_destroy(struct kref *kref)
358 {
359 	struct rpcrdma_ep *ep = container_of(kref, struct rpcrdma_ep, re_kref);
360 
361 	if (ep->re_id->qp) {
362 		rdma_destroy_qp(ep->re_id);
363 		ep->re_id->qp = NULL;
364 	}
365 
366 	if (ep->re_attr.recv_cq)
367 		ib_free_cq(ep->re_attr.recv_cq);
368 	ep->re_attr.recv_cq = NULL;
369 	if (ep->re_attr.send_cq)
370 		ib_free_cq(ep->re_attr.send_cq);
371 	ep->re_attr.send_cq = NULL;
372 
373 	if (ep->re_pd)
374 		ib_dealloc_pd(ep->re_pd);
375 	ep->re_pd = NULL;
376 
377 	kfree(ep);
378 	module_put(THIS_MODULE);
379 }
380 
rpcrdma_ep_get(struct rpcrdma_ep * ep)381 static noinline void rpcrdma_ep_get(struct rpcrdma_ep *ep)
382 {
383 	kref_get(&ep->re_kref);
384 }
385 
386 /* Returns:
387  *     %0 if @ep still has a positive kref count, or
388  *     %1 if @ep was destroyed successfully.
389  */
rpcrdma_ep_put(struct rpcrdma_ep * ep)390 static noinline int rpcrdma_ep_put(struct rpcrdma_ep *ep)
391 {
392 	return kref_put(&ep->re_kref, rpcrdma_ep_destroy);
393 }
394 
rpcrdma_ep_create(struct rpcrdma_xprt * r_xprt)395 static int rpcrdma_ep_create(struct rpcrdma_xprt *r_xprt)
396 {
397 	struct rpcrdma_connect_private *pmsg;
398 	struct ib_device *device;
399 	struct rdma_cm_id *id;
400 	struct rpcrdma_ep *ep;
401 	int rc;
402 
403 	ep = kzalloc(sizeof(*ep), GFP_NOFS);
404 	if (!ep)
405 		return -ENOTCONN;
406 	ep->re_xprt = &r_xprt->rx_xprt;
407 	kref_init(&ep->re_kref);
408 
409 	id = rpcrdma_create_id(r_xprt, ep);
410 	if (IS_ERR(id)) {
411 		kfree(ep);
412 		return PTR_ERR(id);
413 	}
414 	__module_get(THIS_MODULE);
415 	device = id->device;
416 	ep->re_id = id;
417 
418 	ep->re_max_requests = r_xprt->rx_xprt.max_reqs;
419 	ep->re_inline_send = xprt_rdma_max_inline_write;
420 	ep->re_inline_recv = xprt_rdma_max_inline_read;
421 	rc = frwr_query_device(ep, device);
422 	if (rc)
423 		goto out_destroy;
424 
425 	r_xprt->rx_buf.rb_max_requests = cpu_to_be32(ep->re_max_requests);
426 
427 	ep->re_attr.event_handler = rpcrdma_qp_event_handler;
428 	ep->re_attr.qp_context = ep;
429 	ep->re_attr.srq = NULL;
430 	ep->re_attr.cap.max_inline_data = 0;
431 	ep->re_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
432 	ep->re_attr.qp_type = IB_QPT_RC;
433 	ep->re_attr.port_num = ~0;
434 
435 	dprintk("RPC:       %s: requested max: dtos: send %d recv %d; "
436 		"iovs: send %d recv %d\n",
437 		__func__,
438 		ep->re_attr.cap.max_send_wr,
439 		ep->re_attr.cap.max_recv_wr,
440 		ep->re_attr.cap.max_send_sge,
441 		ep->re_attr.cap.max_recv_sge);
442 
443 	ep->re_send_batch = ep->re_max_requests >> 3;
444 	ep->re_send_count = ep->re_send_batch;
445 	init_waitqueue_head(&ep->re_connect_wait);
446 
447 	ep->re_attr.send_cq = ib_alloc_cq_any(device, r_xprt,
448 					      ep->re_attr.cap.max_send_wr,
449 					      IB_POLL_WORKQUEUE);
450 	if (IS_ERR(ep->re_attr.send_cq)) {
451 		rc = PTR_ERR(ep->re_attr.send_cq);
452 		ep->re_attr.send_cq = NULL;
453 		goto out_destroy;
454 	}
455 
456 	ep->re_attr.recv_cq = ib_alloc_cq_any(device, r_xprt,
457 					      ep->re_attr.cap.max_recv_wr,
458 					      IB_POLL_WORKQUEUE);
459 	if (IS_ERR(ep->re_attr.recv_cq)) {
460 		rc = PTR_ERR(ep->re_attr.recv_cq);
461 		ep->re_attr.recv_cq = NULL;
462 		goto out_destroy;
463 	}
464 	ep->re_receive_count = 0;
465 
466 	/* Initialize cma parameters */
467 	memset(&ep->re_remote_cma, 0, sizeof(ep->re_remote_cma));
468 
469 	/* Prepare RDMA-CM private message */
470 	pmsg = &ep->re_cm_private;
471 	pmsg->cp_magic = rpcrdma_cmp_magic;
472 	pmsg->cp_version = RPCRDMA_CMP_VERSION;
473 	pmsg->cp_flags |= RPCRDMA_CMP_F_SND_W_INV_OK;
474 	pmsg->cp_send_size = rpcrdma_encode_buffer_size(ep->re_inline_send);
475 	pmsg->cp_recv_size = rpcrdma_encode_buffer_size(ep->re_inline_recv);
476 	ep->re_remote_cma.private_data = pmsg;
477 	ep->re_remote_cma.private_data_len = sizeof(*pmsg);
478 
479 	/* Client offers RDMA Read but does not initiate */
480 	ep->re_remote_cma.initiator_depth = 0;
481 	ep->re_remote_cma.responder_resources =
482 		min_t(int, U8_MAX, device->attrs.max_qp_rd_atom);
483 
484 	/* Limit transport retries so client can detect server
485 	 * GID changes quickly. RPC layer handles re-establishing
486 	 * transport connection and retransmission.
487 	 */
488 	ep->re_remote_cma.retry_count = 6;
489 
490 	/* RPC-over-RDMA handles its own flow control. In addition,
491 	 * make all RNR NAKs visible so we know that RPC-over-RDMA
492 	 * flow control is working correctly (no NAKs should be seen).
493 	 */
494 	ep->re_remote_cma.flow_control = 0;
495 	ep->re_remote_cma.rnr_retry_count = 0;
496 
497 	ep->re_pd = ib_alloc_pd(device, 0);
498 	if (IS_ERR(ep->re_pd)) {
499 		rc = PTR_ERR(ep->re_pd);
500 		ep->re_pd = NULL;
501 		goto out_destroy;
502 	}
503 
504 	rc = rdma_create_qp(id, ep->re_pd, &ep->re_attr);
505 	if (rc)
506 		goto out_destroy;
507 
508 	r_xprt->rx_ep = ep;
509 	return 0;
510 
511 out_destroy:
512 	rpcrdma_ep_put(ep);
513 	rdma_destroy_id(id);
514 	return rc;
515 }
516 
517 /**
518  * rpcrdma_xprt_connect - Connect an unconnected transport
519  * @r_xprt: controlling transport instance
520  *
521  * Returns 0 on success or a negative errno.
522  */
rpcrdma_xprt_connect(struct rpcrdma_xprt * r_xprt)523 int rpcrdma_xprt_connect(struct rpcrdma_xprt *r_xprt)
524 {
525 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
526 	struct rpcrdma_ep *ep;
527 	int rc;
528 
529 	rc = rpcrdma_ep_create(r_xprt);
530 	if (rc)
531 		return rc;
532 	ep = r_xprt->rx_ep;
533 
534 	xprt_clear_connected(xprt);
535 	rpcrdma_reset_cwnd(r_xprt);
536 
537 	/* Bump the ep's reference count while there are
538 	 * outstanding Receives.
539 	 */
540 	rpcrdma_ep_get(ep);
541 	rpcrdma_post_recvs(r_xprt, 1, true);
542 
543 	rc = rdma_connect(ep->re_id, &ep->re_remote_cma);
544 	if (rc)
545 		goto out;
546 
547 	if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
548 		xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
549 	wait_event_interruptible(ep->re_connect_wait,
550 				 ep->re_connect_status != 0);
551 	if (ep->re_connect_status <= 0) {
552 		rc = ep->re_connect_status;
553 		goto out;
554 	}
555 
556 	rc = rpcrdma_sendctxs_create(r_xprt);
557 	if (rc) {
558 		rc = -ENOTCONN;
559 		goto out;
560 	}
561 
562 	rc = rpcrdma_reqs_setup(r_xprt);
563 	if (rc) {
564 		rc = -ENOTCONN;
565 		goto out;
566 	}
567 	rpcrdma_mrs_create(r_xprt);
568 
569 out:
570 	trace_xprtrdma_connect(r_xprt, rc);
571 	return rc;
572 }
573 
574 /**
575  * rpcrdma_xprt_disconnect - Disconnect underlying transport
576  * @r_xprt: controlling transport instance
577  *
578  * Caller serializes. Either the transport send lock is held,
579  * or we're being called to destroy the transport.
580  *
581  * On return, @r_xprt is completely divested of all hardware
582  * resources and prepared for the next ->connect operation.
583  */
rpcrdma_xprt_disconnect(struct rpcrdma_xprt * r_xprt)584 void rpcrdma_xprt_disconnect(struct rpcrdma_xprt *r_xprt)
585 {
586 	struct rpcrdma_ep *ep = r_xprt->rx_ep;
587 	struct rdma_cm_id *id;
588 	int rc;
589 
590 	if (!ep)
591 		return;
592 
593 	id = ep->re_id;
594 	rc = rdma_disconnect(id);
595 	trace_xprtrdma_disconnect(r_xprt, rc);
596 
597 	rpcrdma_xprt_drain(r_xprt);
598 	rpcrdma_reps_unmap(r_xprt);
599 	rpcrdma_reqs_reset(r_xprt);
600 	rpcrdma_mrs_destroy(r_xprt);
601 	rpcrdma_sendctxs_destroy(r_xprt);
602 
603 	if (rpcrdma_ep_put(ep))
604 		rdma_destroy_id(id);
605 
606 	r_xprt->rx_ep = NULL;
607 }
608 
609 /* Fixed-size circular FIFO queue. This implementation is wait-free and
610  * lock-free.
611  *
612  * Consumer is the code path that posts Sends. This path dequeues a
613  * sendctx for use by a Send operation. Multiple consumer threads
614  * are serialized by the RPC transport lock, which allows only one
615  * ->send_request call at a time.
616  *
617  * Producer is the code path that handles Send completions. This path
618  * enqueues a sendctx that has been completed. Multiple producer
619  * threads are serialized by the ib_poll_cq() function.
620  */
621 
622 /* rpcrdma_sendctxs_destroy() assumes caller has already quiesced
623  * queue activity, and rpcrdma_xprt_drain has flushed all remaining
624  * Send requests.
625  */
rpcrdma_sendctxs_destroy(struct rpcrdma_xprt * r_xprt)626 static void rpcrdma_sendctxs_destroy(struct rpcrdma_xprt *r_xprt)
627 {
628 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
629 	unsigned long i;
630 
631 	if (!buf->rb_sc_ctxs)
632 		return;
633 	for (i = 0; i <= buf->rb_sc_last; i++)
634 		kfree(buf->rb_sc_ctxs[i]);
635 	kfree(buf->rb_sc_ctxs);
636 	buf->rb_sc_ctxs = NULL;
637 }
638 
rpcrdma_sendctx_create(struct rpcrdma_ep * ep)639 static struct rpcrdma_sendctx *rpcrdma_sendctx_create(struct rpcrdma_ep *ep)
640 {
641 	struct rpcrdma_sendctx *sc;
642 
643 	sc = kzalloc(struct_size(sc, sc_sges, ep->re_attr.cap.max_send_sge),
644 		     GFP_KERNEL);
645 	if (!sc)
646 		return NULL;
647 
648 	sc->sc_cqe.done = rpcrdma_wc_send;
649 	return sc;
650 }
651 
rpcrdma_sendctxs_create(struct rpcrdma_xprt * r_xprt)652 static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt)
653 {
654 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
655 	struct rpcrdma_sendctx *sc;
656 	unsigned long i;
657 
658 	/* Maximum number of concurrent outstanding Send WRs. Capping
659 	 * the circular queue size stops Send Queue overflow by causing
660 	 * the ->send_request call to fail temporarily before too many
661 	 * Sends are posted.
662 	 */
663 	i = r_xprt->rx_ep->re_max_requests + RPCRDMA_MAX_BC_REQUESTS;
664 	buf->rb_sc_ctxs = kcalloc(i, sizeof(sc), GFP_KERNEL);
665 	if (!buf->rb_sc_ctxs)
666 		return -ENOMEM;
667 
668 	buf->rb_sc_last = i - 1;
669 	for (i = 0; i <= buf->rb_sc_last; i++) {
670 		sc = rpcrdma_sendctx_create(r_xprt->rx_ep);
671 		if (!sc)
672 			return -ENOMEM;
673 
674 		buf->rb_sc_ctxs[i] = sc;
675 	}
676 
677 	buf->rb_sc_head = 0;
678 	buf->rb_sc_tail = 0;
679 	return 0;
680 }
681 
682 /* The sendctx queue is not guaranteed to have a size that is a
683  * power of two, thus the helpers in circ_buf.h cannot be used.
684  * The other option is to use modulus (%), which can be expensive.
685  */
rpcrdma_sendctx_next(struct rpcrdma_buffer * buf,unsigned long item)686 static unsigned long rpcrdma_sendctx_next(struct rpcrdma_buffer *buf,
687 					  unsigned long item)
688 {
689 	return likely(item < buf->rb_sc_last) ? item + 1 : 0;
690 }
691 
692 /**
693  * rpcrdma_sendctx_get_locked - Acquire a send context
694  * @r_xprt: controlling transport instance
695  *
696  * Returns pointer to a free send completion context; or NULL if
697  * the queue is empty.
698  *
699  * Usage: Called to acquire an SGE array before preparing a Send WR.
700  *
701  * The caller serializes calls to this function (per transport), and
702  * provides an effective memory barrier that flushes the new value
703  * of rb_sc_head.
704  */
rpcrdma_sendctx_get_locked(struct rpcrdma_xprt * r_xprt)705 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_xprt *r_xprt)
706 {
707 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
708 	struct rpcrdma_sendctx *sc;
709 	unsigned long next_head;
710 
711 	next_head = rpcrdma_sendctx_next(buf, buf->rb_sc_head);
712 
713 	if (next_head == READ_ONCE(buf->rb_sc_tail))
714 		goto out_emptyq;
715 
716 	/* ORDER: item must be accessed _before_ head is updated */
717 	sc = buf->rb_sc_ctxs[next_head];
718 
719 	/* Releasing the lock in the caller acts as a memory
720 	 * barrier that flushes rb_sc_head.
721 	 */
722 	buf->rb_sc_head = next_head;
723 
724 	return sc;
725 
726 out_emptyq:
727 	/* The queue is "empty" if there have not been enough Send
728 	 * completions recently. This is a sign the Send Queue is
729 	 * backing up. Cause the caller to pause and try again.
730 	 */
731 	xprt_wait_for_buffer_space(&r_xprt->rx_xprt);
732 	r_xprt->rx_stats.empty_sendctx_q++;
733 	return NULL;
734 }
735 
736 /**
737  * rpcrdma_sendctx_put_locked - Release a send context
738  * @r_xprt: controlling transport instance
739  * @sc: send context to release
740  *
741  * Usage: Called from Send completion to return a sendctxt
742  * to the queue.
743  *
744  * The caller serializes calls to this function (per transport).
745  */
rpcrdma_sendctx_put_locked(struct rpcrdma_xprt * r_xprt,struct rpcrdma_sendctx * sc)746 static void rpcrdma_sendctx_put_locked(struct rpcrdma_xprt *r_xprt,
747 				       struct rpcrdma_sendctx *sc)
748 {
749 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
750 	unsigned long next_tail;
751 
752 	/* Unmap SGEs of previously completed but unsignaled
753 	 * Sends by walking up the queue until @sc is found.
754 	 */
755 	next_tail = buf->rb_sc_tail;
756 	do {
757 		next_tail = rpcrdma_sendctx_next(buf, next_tail);
758 
759 		/* ORDER: item must be accessed _before_ tail is updated */
760 		rpcrdma_sendctx_unmap(buf->rb_sc_ctxs[next_tail]);
761 
762 	} while (buf->rb_sc_ctxs[next_tail] != sc);
763 
764 	/* Paired with READ_ONCE */
765 	smp_store_release(&buf->rb_sc_tail, next_tail);
766 
767 	xprt_write_space(&r_xprt->rx_xprt);
768 }
769 
770 static void
rpcrdma_mrs_create(struct rpcrdma_xprt * r_xprt)771 rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt)
772 {
773 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
774 	struct rpcrdma_ep *ep = r_xprt->rx_ep;
775 	unsigned int count;
776 
777 	for (count = 0; count < ep->re_max_rdma_segs; count++) {
778 		struct rpcrdma_mr *mr;
779 		int rc;
780 
781 		mr = kzalloc(sizeof(*mr), GFP_NOFS);
782 		if (!mr)
783 			break;
784 
785 		rc = frwr_mr_init(r_xprt, mr);
786 		if (rc) {
787 			kfree(mr);
788 			break;
789 		}
790 
791 		spin_lock(&buf->rb_lock);
792 		rpcrdma_mr_push(mr, &buf->rb_mrs);
793 		list_add(&mr->mr_all, &buf->rb_all_mrs);
794 		spin_unlock(&buf->rb_lock);
795 	}
796 
797 	r_xprt->rx_stats.mrs_allocated += count;
798 	trace_xprtrdma_createmrs(r_xprt, count);
799 }
800 
801 static void
rpcrdma_mr_refresh_worker(struct work_struct * work)802 rpcrdma_mr_refresh_worker(struct work_struct *work)
803 {
804 	struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
805 						  rb_refresh_worker);
806 	struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
807 						   rx_buf);
808 
809 	rpcrdma_mrs_create(r_xprt);
810 	xprt_write_space(&r_xprt->rx_xprt);
811 }
812 
813 /**
814  * rpcrdma_mrs_refresh - Wake the MR refresh worker
815  * @r_xprt: controlling transport instance
816  *
817  */
rpcrdma_mrs_refresh(struct rpcrdma_xprt * r_xprt)818 void rpcrdma_mrs_refresh(struct rpcrdma_xprt *r_xprt)
819 {
820 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
821 	struct rpcrdma_ep *ep = r_xprt->rx_ep;
822 
823 	/* If there is no underlying connection, it's no use
824 	 * to wake the refresh worker.
825 	 */
826 	if (ep->re_connect_status == 1) {
827 		/* The work is scheduled on a WQ_MEM_RECLAIM
828 		 * workqueue in order to prevent MR allocation
829 		 * from recursing into NFS during direct reclaim.
830 		 */
831 		queue_work(xprtiod_workqueue, &buf->rb_refresh_worker);
832 	}
833 }
834 
835 /**
836  * rpcrdma_req_create - Allocate an rpcrdma_req object
837  * @r_xprt: controlling r_xprt
838  * @size: initial size, in bytes, of send and receive buffers
839  * @flags: GFP flags passed to memory allocators
840  *
841  * Returns an allocated and fully initialized rpcrdma_req or NULL.
842  */
rpcrdma_req_create(struct rpcrdma_xprt * r_xprt,size_t size,gfp_t flags)843 struct rpcrdma_req *rpcrdma_req_create(struct rpcrdma_xprt *r_xprt, size_t size,
844 				       gfp_t flags)
845 {
846 	struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
847 	struct rpcrdma_req *req;
848 
849 	req = kzalloc(sizeof(*req), flags);
850 	if (req == NULL)
851 		goto out1;
852 
853 	req->rl_sendbuf = rpcrdma_regbuf_alloc(size, DMA_TO_DEVICE, flags);
854 	if (!req->rl_sendbuf)
855 		goto out2;
856 
857 	req->rl_recvbuf = rpcrdma_regbuf_alloc(size, DMA_NONE, flags);
858 	if (!req->rl_recvbuf)
859 		goto out3;
860 
861 	INIT_LIST_HEAD(&req->rl_free_mrs);
862 	INIT_LIST_HEAD(&req->rl_registered);
863 	spin_lock(&buffer->rb_lock);
864 	list_add(&req->rl_all, &buffer->rb_allreqs);
865 	spin_unlock(&buffer->rb_lock);
866 	return req;
867 
868 out3:
869 	rpcrdma_regbuf_free(req->rl_sendbuf);
870 out2:
871 	kfree(req);
872 out1:
873 	return NULL;
874 }
875 
876 /**
877  * rpcrdma_req_setup - Per-connection instance setup of an rpcrdma_req object
878  * @r_xprt: controlling transport instance
879  * @req: rpcrdma_req object to set up
880  *
881  * Returns zero on success, and a negative errno on failure.
882  */
rpcrdma_req_setup(struct rpcrdma_xprt * r_xprt,struct rpcrdma_req * req)883 int rpcrdma_req_setup(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
884 {
885 	struct rpcrdma_regbuf *rb;
886 	size_t maxhdrsize;
887 
888 	/* Compute maximum header buffer size in bytes */
889 	maxhdrsize = rpcrdma_fixed_maxsz + 3 +
890 		     r_xprt->rx_ep->re_max_rdma_segs * rpcrdma_readchunk_maxsz;
891 	maxhdrsize *= sizeof(__be32);
892 	rb = rpcrdma_regbuf_alloc(__roundup_pow_of_two(maxhdrsize),
893 				  DMA_TO_DEVICE, GFP_KERNEL);
894 	if (!rb)
895 		goto out;
896 
897 	if (!__rpcrdma_regbuf_dma_map(r_xprt, rb))
898 		goto out_free;
899 
900 	req->rl_rdmabuf = rb;
901 	xdr_buf_init(&req->rl_hdrbuf, rdmab_data(rb), rdmab_length(rb));
902 	return 0;
903 
904 out_free:
905 	rpcrdma_regbuf_free(rb);
906 out:
907 	return -ENOMEM;
908 }
909 
910 /* ASSUMPTION: the rb_allreqs list is stable for the duration,
911  * and thus can be walked without holding rb_lock. Eg. the
912  * caller is holding the transport send lock to exclude
913  * device removal or disconnection.
914  */
rpcrdma_reqs_setup(struct rpcrdma_xprt * r_xprt)915 static int rpcrdma_reqs_setup(struct rpcrdma_xprt *r_xprt)
916 {
917 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
918 	struct rpcrdma_req *req;
919 	int rc;
920 
921 	list_for_each_entry(req, &buf->rb_allreqs, rl_all) {
922 		rc = rpcrdma_req_setup(r_xprt, req);
923 		if (rc)
924 			return rc;
925 	}
926 	return 0;
927 }
928 
rpcrdma_req_reset(struct rpcrdma_req * req)929 static void rpcrdma_req_reset(struct rpcrdma_req *req)
930 {
931 	/* Credits are valid for only one connection */
932 	req->rl_slot.rq_cong = 0;
933 
934 	rpcrdma_regbuf_free(req->rl_rdmabuf);
935 	req->rl_rdmabuf = NULL;
936 
937 	rpcrdma_regbuf_dma_unmap(req->rl_sendbuf);
938 	rpcrdma_regbuf_dma_unmap(req->rl_recvbuf);
939 
940 	frwr_reset(req);
941 }
942 
943 /* ASSUMPTION: the rb_allreqs list is stable for the duration,
944  * and thus can be walked without holding rb_lock. Eg. the
945  * caller is holding the transport send lock to exclude
946  * device removal or disconnection.
947  */
rpcrdma_reqs_reset(struct rpcrdma_xprt * r_xprt)948 static void rpcrdma_reqs_reset(struct rpcrdma_xprt *r_xprt)
949 {
950 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
951 	struct rpcrdma_req *req;
952 
953 	list_for_each_entry(req, &buf->rb_allreqs, rl_all)
954 		rpcrdma_req_reset(req);
955 }
956 
957 /* No locking needed here. This function is called only by the
958  * Receive completion handler.
959  */
960 static noinline
rpcrdma_rep_create(struct rpcrdma_xprt * r_xprt,bool temp)961 struct rpcrdma_rep *rpcrdma_rep_create(struct rpcrdma_xprt *r_xprt,
962 				       bool temp)
963 {
964 	struct rpcrdma_rep *rep;
965 
966 	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
967 	if (rep == NULL)
968 		goto out;
969 
970 	rep->rr_rdmabuf = rpcrdma_regbuf_alloc(r_xprt->rx_ep->re_inline_recv,
971 					       DMA_FROM_DEVICE, GFP_KERNEL);
972 	if (!rep->rr_rdmabuf)
973 		goto out_free;
974 
975 	if (!rpcrdma_regbuf_dma_map(r_xprt, rep->rr_rdmabuf))
976 		goto out_free_regbuf;
977 
978 	xdr_buf_init(&rep->rr_hdrbuf, rdmab_data(rep->rr_rdmabuf),
979 		     rdmab_length(rep->rr_rdmabuf));
980 	rep->rr_cqe.done = rpcrdma_wc_receive;
981 	rep->rr_rxprt = r_xprt;
982 	rep->rr_recv_wr.next = NULL;
983 	rep->rr_recv_wr.wr_cqe = &rep->rr_cqe;
984 	rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
985 	rep->rr_recv_wr.num_sge = 1;
986 	rep->rr_temp = temp;
987 	list_add(&rep->rr_all, &r_xprt->rx_buf.rb_all_reps);
988 	return rep;
989 
990 out_free_regbuf:
991 	rpcrdma_regbuf_free(rep->rr_rdmabuf);
992 out_free:
993 	kfree(rep);
994 out:
995 	return NULL;
996 }
997 
998 /* No locking needed here. This function is invoked only by the
999  * Receive completion handler, or during transport shutdown.
1000  */
rpcrdma_rep_destroy(struct rpcrdma_rep * rep)1001 static void rpcrdma_rep_destroy(struct rpcrdma_rep *rep)
1002 {
1003 	list_del(&rep->rr_all);
1004 	rpcrdma_regbuf_free(rep->rr_rdmabuf);
1005 	kfree(rep);
1006 }
1007 
rpcrdma_rep_get_locked(struct rpcrdma_buffer * buf)1008 static struct rpcrdma_rep *rpcrdma_rep_get_locked(struct rpcrdma_buffer *buf)
1009 {
1010 	struct llist_node *node;
1011 
1012 	/* Calls to llist_del_first are required to be serialized */
1013 	node = llist_del_first(&buf->rb_free_reps);
1014 	if (!node)
1015 		return NULL;
1016 	return llist_entry(node, struct rpcrdma_rep, rr_node);
1017 }
1018 
rpcrdma_rep_put(struct rpcrdma_buffer * buf,struct rpcrdma_rep * rep)1019 static void rpcrdma_rep_put(struct rpcrdma_buffer *buf,
1020 			    struct rpcrdma_rep *rep)
1021 {
1022 	llist_add(&rep->rr_node, &buf->rb_free_reps);
1023 }
1024 
rpcrdma_reps_unmap(struct rpcrdma_xprt * r_xprt)1025 static void rpcrdma_reps_unmap(struct rpcrdma_xprt *r_xprt)
1026 {
1027 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1028 	struct rpcrdma_rep *rep;
1029 
1030 	list_for_each_entry(rep, &buf->rb_all_reps, rr_all) {
1031 		rpcrdma_regbuf_dma_unmap(rep->rr_rdmabuf);
1032 		rep->rr_temp = true;
1033 	}
1034 }
1035 
rpcrdma_reps_destroy(struct rpcrdma_buffer * buf)1036 static void rpcrdma_reps_destroy(struct rpcrdma_buffer *buf)
1037 {
1038 	struct rpcrdma_rep *rep;
1039 
1040 	while ((rep = rpcrdma_rep_get_locked(buf)) != NULL)
1041 		rpcrdma_rep_destroy(rep);
1042 }
1043 
1044 /**
1045  * rpcrdma_buffer_create - Create initial set of req/rep objects
1046  * @r_xprt: transport instance to (re)initialize
1047  *
1048  * Returns zero on success, otherwise a negative errno.
1049  */
rpcrdma_buffer_create(struct rpcrdma_xprt * r_xprt)1050 int rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
1051 {
1052 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1053 	int i, rc;
1054 
1055 	buf->rb_bc_srv_max_requests = 0;
1056 	spin_lock_init(&buf->rb_lock);
1057 	INIT_LIST_HEAD(&buf->rb_mrs);
1058 	INIT_LIST_HEAD(&buf->rb_all_mrs);
1059 	INIT_WORK(&buf->rb_refresh_worker, rpcrdma_mr_refresh_worker);
1060 
1061 	INIT_LIST_HEAD(&buf->rb_send_bufs);
1062 	INIT_LIST_HEAD(&buf->rb_allreqs);
1063 	INIT_LIST_HEAD(&buf->rb_all_reps);
1064 
1065 	rc = -ENOMEM;
1066 	for (i = 0; i < r_xprt->rx_xprt.max_reqs; i++) {
1067 		struct rpcrdma_req *req;
1068 
1069 		req = rpcrdma_req_create(r_xprt, RPCRDMA_V1_DEF_INLINE_SIZE * 2,
1070 					 GFP_KERNEL);
1071 		if (!req)
1072 			goto out;
1073 		list_add(&req->rl_list, &buf->rb_send_bufs);
1074 	}
1075 
1076 	init_llist_head(&buf->rb_free_reps);
1077 
1078 	return 0;
1079 out:
1080 	rpcrdma_buffer_destroy(buf);
1081 	return rc;
1082 }
1083 
1084 /**
1085  * rpcrdma_req_destroy - Destroy an rpcrdma_req object
1086  * @req: unused object to be destroyed
1087  *
1088  * Relies on caller holding the transport send lock to protect
1089  * removing req->rl_all from buf->rb_all_reqs safely.
1090  */
rpcrdma_req_destroy(struct rpcrdma_req * req)1091 void rpcrdma_req_destroy(struct rpcrdma_req *req)
1092 {
1093 	struct rpcrdma_mr *mr;
1094 
1095 	list_del(&req->rl_all);
1096 
1097 	while ((mr = rpcrdma_mr_pop(&req->rl_free_mrs))) {
1098 		struct rpcrdma_buffer *buf = &mr->mr_xprt->rx_buf;
1099 
1100 		spin_lock(&buf->rb_lock);
1101 		list_del(&mr->mr_all);
1102 		spin_unlock(&buf->rb_lock);
1103 
1104 		frwr_release_mr(mr);
1105 	}
1106 
1107 	rpcrdma_regbuf_free(req->rl_recvbuf);
1108 	rpcrdma_regbuf_free(req->rl_sendbuf);
1109 	rpcrdma_regbuf_free(req->rl_rdmabuf);
1110 	kfree(req);
1111 }
1112 
1113 /**
1114  * rpcrdma_mrs_destroy - Release all of a transport's MRs
1115  * @r_xprt: controlling transport instance
1116  *
1117  * Relies on caller holding the transport send lock to protect
1118  * removing mr->mr_list from req->rl_free_mrs safely.
1119  */
rpcrdma_mrs_destroy(struct rpcrdma_xprt * r_xprt)1120 static void rpcrdma_mrs_destroy(struct rpcrdma_xprt *r_xprt)
1121 {
1122 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1123 	struct rpcrdma_mr *mr;
1124 
1125 	cancel_work_sync(&buf->rb_refresh_worker);
1126 
1127 	spin_lock(&buf->rb_lock);
1128 	while ((mr = list_first_entry_or_null(&buf->rb_all_mrs,
1129 					      struct rpcrdma_mr,
1130 					      mr_all)) != NULL) {
1131 		list_del(&mr->mr_list);
1132 		list_del(&mr->mr_all);
1133 		spin_unlock(&buf->rb_lock);
1134 
1135 		frwr_release_mr(mr);
1136 
1137 		spin_lock(&buf->rb_lock);
1138 	}
1139 	spin_unlock(&buf->rb_lock);
1140 }
1141 
1142 /**
1143  * rpcrdma_buffer_destroy - Release all hw resources
1144  * @buf: root control block for resources
1145  *
1146  * ORDERING: relies on a prior rpcrdma_xprt_drain :
1147  * - No more Send or Receive completions can occur
1148  * - All MRs, reps, and reqs are returned to their free lists
1149  */
1150 void
rpcrdma_buffer_destroy(struct rpcrdma_buffer * buf)1151 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1152 {
1153 	rpcrdma_reps_destroy(buf);
1154 
1155 	while (!list_empty(&buf->rb_send_bufs)) {
1156 		struct rpcrdma_req *req;
1157 
1158 		req = list_first_entry(&buf->rb_send_bufs,
1159 				       struct rpcrdma_req, rl_list);
1160 		list_del(&req->rl_list);
1161 		rpcrdma_req_destroy(req);
1162 	}
1163 }
1164 
1165 /**
1166  * rpcrdma_mr_get - Allocate an rpcrdma_mr object
1167  * @r_xprt: controlling transport
1168  *
1169  * Returns an initialized rpcrdma_mr or NULL if no free
1170  * rpcrdma_mr objects are available.
1171  */
1172 struct rpcrdma_mr *
rpcrdma_mr_get(struct rpcrdma_xprt * r_xprt)1173 rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt)
1174 {
1175 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1176 	struct rpcrdma_mr *mr;
1177 
1178 	spin_lock(&buf->rb_lock);
1179 	mr = rpcrdma_mr_pop(&buf->rb_mrs);
1180 	spin_unlock(&buf->rb_lock);
1181 	return mr;
1182 }
1183 
1184 /**
1185  * rpcrdma_mr_put - DMA unmap an MR and release it
1186  * @mr: MR to release
1187  *
1188  */
rpcrdma_mr_put(struct rpcrdma_mr * mr)1189 void rpcrdma_mr_put(struct rpcrdma_mr *mr)
1190 {
1191 	struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
1192 
1193 	if (mr->mr_dir != DMA_NONE) {
1194 		trace_xprtrdma_mr_unmap(mr);
1195 		ib_dma_unmap_sg(r_xprt->rx_ep->re_id->device,
1196 				mr->mr_sg, mr->mr_nents, mr->mr_dir);
1197 		mr->mr_dir = DMA_NONE;
1198 	}
1199 
1200 	rpcrdma_mr_push(mr, &mr->mr_req->rl_free_mrs);
1201 }
1202 
1203 /**
1204  * rpcrdma_reply_put - Put reply buffers back into pool
1205  * @buffers: buffer pool
1206  * @req: object to return
1207  *
1208  */
rpcrdma_reply_put(struct rpcrdma_buffer * buffers,struct rpcrdma_req * req)1209 void rpcrdma_reply_put(struct rpcrdma_buffer *buffers, struct rpcrdma_req *req)
1210 {
1211 	if (req->rl_reply) {
1212 		rpcrdma_rep_put(buffers, req->rl_reply);
1213 		req->rl_reply = NULL;
1214 	}
1215 }
1216 
1217 /**
1218  * rpcrdma_buffer_get - Get a request buffer
1219  * @buffers: Buffer pool from which to obtain a buffer
1220  *
1221  * Returns a fresh rpcrdma_req, or NULL if none are available.
1222  */
1223 struct rpcrdma_req *
rpcrdma_buffer_get(struct rpcrdma_buffer * buffers)1224 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1225 {
1226 	struct rpcrdma_req *req;
1227 
1228 	spin_lock(&buffers->rb_lock);
1229 	req = list_first_entry_or_null(&buffers->rb_send_bufs,
1230 				       struct rpcrdma_req, rl_list);
1231 	if (req)
1232 		list_del_init(&req->rl_list);
1233 	spin_unlock(&buffers->rb_lock);
1234 	return req;
1235 }
1236 
1237 /**
1238  * rpcrdma_buffer_put - Put request/reply buffers back into pool
1239  * @buffers: buffer pool
1240  * @req: object to return
1241  *
1242  */
rpcrdma_buffer_put(struct rpcrdma_buffer * buffers,struct rpcrdma_req * req)1243 void rpcrdma_buffer_put(struct rpcrdma_buffer *buffers, struct rpcrdma_req *req)
1244 {
1245 	rpcrdma_reply_put(buffers, req);
1246 
1247 	spin_lock(&buffers->rb_lock);
1248 	list_add(&req->rl_list, &buffers->rb_send_bufs);
1249 	spin_unlock(&buffers->rb_lock);
1250 }
1251 
1252 /**
1253  * rpcrdma_recv_buffer_put - Release rpcrdma_rep back to free list
1254  * @rep: rep to release
1255  *
1256  * Used after error conditions.
1257  */
rpcrdma_recv_buffer_put(struct rpcrdma_rep * rep)1258 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1259 {
1260 	rpcrdma_rep_put(&rep->rr_rxprt->rx_buf, rep);
1261 }
1262 
1263 /* Returns a pointer to a rpcrdma_regbuf object, or NULL.
1264  *
1265  * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1266  * receiving the payload of RDMA RECV operations. During Long Calls
1267  * or Replies they may be registered externally via frwr_map.
1268  */
1269 static struct rpcrdma_regbuf *
rpcrdma_regbuf_alloc(size_t size,enum dma_data_direction direction,gfp_t flags)1270 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction,
1271 		     gfp_t flags)
1272 {
1273 	struct rpcrdma_regbuf *rb;
1274 
1275 	rb = kmalloc(sizeof(*rb), flags);
1276 	if (!rb)
1277 		return NULL;
1278 	rb->rg_data = kmalloc(size, flags);
1279 	if (!rb->rg_data) {
1280 		kfree(rb);
1281 		return NULL;
1282 	}
1283 
1284 	rb->rg_device = NULL;
1285 	rb->rg_direction = direction;
1286 	rb->rg_iov.length = size;
1287 	return rb;
1288 }
1289 
1290 /**
1291  * rpcrdma_regbuf_realloc - re-allocate a SEND/RECV buffer
1292  * @rb: regbuf to reallocate
1293  * @size: size of buffer to be allocated, in bytes
1294  * @flags: GFP flags
1295  *
1296  * Returns true if reallocation was successful. If false is
1297  * returned, @rb is left untouched.
1298  */
rpcrdma_regbuf_realloc(struct rpcrdma_regbuf * rb,size_t size,gfp_t flags)1299 bool rpcrdma_regbuf_realloc(struct rpcrdma_regbuf *rb, size_t size, gfp_t flags)
1300 {
1301 	void *buf;
1302 
1303 	buf = kmalloc(size, flags);
1304 	if (!buf)
1305 		return false;
1306 
1307 	rpcrdma_regbuf_dma_unmap(rb);
1308 	kfree(rb->rg_data);
1309 
1310 	rb->rg_data = buf;
1311 	rb->rg_iov.length = size;
1312 	return true;
1313 }
1314 
1315 /**
1316  * __rpcrdma_regbuf_dma_map - DMA-map a regbuf
1317  * @r_xprt: controlling transport instance
1318  * @rb: regbuf to be mapped
1319  *
1320  * Returns true if the buffer is now DMA mapped to @r_xprt's device
1321  */
__rpcrdma_regbuf_dma_map(struct rpcrdma_xprt * r_xprt,struct rpcrdma_regbuf * rb)1322 bool __rpcrdma_regbuf_dma_map(struct rpcrdma_xprt *r_xprt,
1323 			      struct rpcrdma_regbuf *rb)
1324 {
1325 	struct ib_device *device = r_xprt->rx_ep->re_id->device;
1326 
1327 	if (rb->rg_direction == DMA_NONE)
1328 		return false;
1329 
1330 	rb->rg_iov.addr = ib_dma_map_single(device, rdmab_data(rb),
1331 					    rdmab_length(rb), rb->rg_direction);
1332 	if (ib_dma_mapping_error(device, rdmab_addr(rb))) {
1333 		trace_xprtrdma_dma_maperr(rdmab_addr(rb));
1334 		return false;
1335 	}
1336 
1337 	rb->rg_device = device;
1338 	rb->rg_iov.lkey = r_xprt->rx_ep->re_pd->local_dma_lkey;
1339 	return true;
1340 }
1341 
rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf * rb)1342 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb)
1343 {
1344 	if (!rb)
1345 		return;
1346 
1347 	if (!rpcrdma_regbuf_is_mapped(rb))
1348 		return;
1349 
1350 	ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb), rdmab_length(rb),
1351 			    rb->rg_direction);
1352 	rb->rg_device = NULL;
1353 }
1354 
rpcrdma_regbuf_free(struct rpcrdma_regbuf * rb)1355 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb)
1356 {
1357 	rpcrdma_regbuf_dma_unmap(rb);
1358 	if (rb)
1359 		kfree(rb->rg_data);
1360 	kfree(rb);
1361 }
1362 
1363 /**
1364  * rpcrdma_post_sends - Post WRs to a transport's Send Queue
1365  * @r_xprt: controlling transport instance
1366  * @req: rpcrdma_req containing the Send WR to post
1367  *
1368  * Returns 0 if the post was successful, otherwise -ENOTCONN
1369  * is returned.
1370  */
rpcrdma_post_sends(struct rpcrdma_xprt * r_xprt,struct rpcrdma_req * req)1371 int rpcrdma_post_sends(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
1372 {
1373 	struct ib_send_wr *send_wr = &req->rl_wr;
1374 	struct rpcrdma_ep *ep = r_xprt->rx_ep;
1375 	int rc;
1376 
1377 	if (!ep->re_send_count || kref_read(&req->rl_kref) > 1) {
1378 		send_wr->send_flags |= IB_SEND_SIGNALED;
1379 		ep->re_send_count = ep->re_send_batch;
1380 	} else {
1381 		send_wr->send_flags &= ~IB_SEND_SIGNALED;
1382 		--ep->re_send_count;
1383 	}
1384 
1385 	trace_xprtrdma_post_send(req);
1386 	rc = frwr_send(r_xprt, req);
1387 	if (rc)
1388 		return -ENOTCONN;
1389 	return 0;
1390 }
1391 
1392 /**
1393  * rpcrdma_post_recvs - Refill the Receive Queue
1394  * @r_xprt: controlling transport instance
1395  * @needed: current credit grant
1396  * @temp: mark Receive buffers to be deleted after one use
1397  *
1398  */
rpcrdma_post_recvs(struct rpcrdma_xprt * r_xprt,int needed,bool temp)1399 void rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, int needed, bool temp)
1400 {
1401 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1402 	struct rpcrdma_ep *ep = r_xprt->rx_ep;
1403 	struct ib_recv_wr *wr, *bad_wr;
1404 	struct rpcrdma_rep *rep;
1405 	int count, rc;
1406 
1407 	rc = 0;
1408 	count = 0;
1409 
1410 	if (likely(ep->re_receive_count > needed))
1411 		goto out;
1412 	needed -= ep->re_receive_count;
1413 	if (!temp)
1414 		needed += RPCRDMA_MAX_RECV_BATCH;
1415 
1416 	/* fast path: all needed reps can be found on the free list */
1417 	wr = NULL;
1418 	while (needed) {
1419 		rep = rpcrdma_rep_get_locked(buf);
1420 		if (rep && rep->rr_temp) {
1421 			rpcrdma_rep_destroy(rep);
1422 			continue;
1423 		}
1424 		if (!rep)
1425 			rep = rpcrdma_rep_create(r_xprt, temp);
1426 		if (!rep)
1427 			break;
1428 
1429 		trace_xprtrdma_post_recv(rep);
1430 		rep->rr_recv_wr.next = wr;
1431 		wr = &rep->rr_recv_wr;
1432 		--needed;
1433 		++count;
1434 	}
1435 	if (!wr)
1436 		goto out;
1437 
1438 	rc = ib_post_recv(ep->re_id->qp, wr,
1439 			  (const struct ib_recv_wr **)&bad_wr);
1440 out:
1441 	trace_xprtrdma_post_recvs(r_xprt, count, rc);
1442 	if (rc) {
1443 		for (wr = bad_wr; wr;) {
1444 			struct rpcrdma_rep *rep;
1445 
1446 			rep = container_of(wr, struct rpcrdma_rep, rr_recv_wr);
1447 			wr = wr->next;
1448 			rpcrdma_recv_buffer_put(rep);
1449 			--count;
1450 		}
1451 	}
1452 	ep->re_receive_count += count;
1453 	return;
1454 }
1455