verbs.c 40.9 KB
Newer Older
1
/*
C
Chuck Lever 已提交
2
 * Copyright (c) 2014-2017 Oracle.  All rights reserved.
3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38
 * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
 *
 * This software is available to you under a choice of one of two
 * licenses.  You may choose to be licensed under the terms of the GNU
 * General Public License (GPL) Version 2, available from the file
 * COPYING in the main directory of this source tree, or the BSD-type
 * license below:
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions
 * are met:
 *
 *      Redistributions of source code must retain the above copyright
 *      notice, this list of conditions and the following disclaimer.
 *
 *      Redistributions in binary form must reproduce the above
 *      copyright notice, this list of conditions and the following
 *      disclaimer in the documentation and/or other materials provided
 *      with the distribution.
 *
 *      Neither the name of the Network Appliance, Inc. nor the names of
 *      its contributors may be used to endorse or promote products
 *      derived from this software without specific prior written
 *      permission.
 *
 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
39 40
 */

41 42 43 44 45 46 47 48 49 50
/*
 * verbs.c
 *
 * Encapsulates the major functions managing:
 *  o adapters
 *  o endpoints
 *  o connections
 *  o buffer memory
 */

51
#include <linux/interrupt.h>
52
#include <linux/slab.h>
53
#include <linux/sunrpc/addr.h>
54
#include <linux/sunrpc/svc_rdma.h>
55 56

#include <asm-generic/barrier.h>
57
#include <asm/bitops.h>
58

59
#include <rdma/ib_cm.h>
60

61 62
#include "xprt_rdma.h"

63 64 65 66
/*
 * Globals/Macros
 */

J
Jeff Layton 已提交
67
#if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
68 69 70 71 72 73
# define RPCDBG_FACILITY	RPCDBG_TRANS
#endif

/*
 * internal functions
 */
C
Chuck Lever 已提交
74 75
static void rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt);
static void rpcrdma_mrs_destroy(struct rpcrdma_buffer *buf);
76
static void rpcrdma_dma_unmap_regbuf(struct rpcrdma_regbuf *rb);
77

78
struct workqueue_struct *rpcrdma_receive_wq __read_mostly;
79

80 81
int
rpcrdma_alloc_wq(void)
82
{
83
	struct workqueue_struct *recv_wq;
84

85
	recv_wq = alloc_workqueue("xprtrdma_receive",
86
				  WQ_MEM_RECLAIM | WQ_HIGHPRI,
87 88 89
				  0);
	if (!recv_wq)
		return -ENOMEM;
90

91 92
	rpcrdma_receive_wq = recv_wq;
	return 0;
93 94
}

95 96
void
rpcrdma_destroy_wq(void)
97
{
98
	struct workqueue_struct *wq;
99

100 101 102 103 104
	if (rpcrdma_receive_wq) {
		wq = rpcrdma_receive_wq;
		rpcrdma_receive_wq = NULL;
		destroy_workqueue(wq);
	}
105 106
}

107 108 109 110 111
static void
rpcrdma_qp_async_error_upcall(struct ib_event *event, void *context)
{
	struct rpcrdma_ep *ep = context;

112 113 114
	pr_err("rpcrdma: %s on device %s ep %p\n",
	       ib_event_msg(event->event), event->device->name, context);

115 116
	if (ep->rep_connected == 1) {
		ep->rep_connected = -EIO;
117
		rpcrdma_conn_func(ep);
118 119 120 121
		wake_up_all(&ep->rep_connect_wait);
	}
}

122 123 124 125 126
/**
 * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC
 * @cq:	completion queue (ignored)
 * @wc:	completed WR
 *
127 128
 */
static void
129
rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
130
{
131 132 133 134
	struct ib_cqe *cqe = wc->wr_cqe;
	struct rpcrdma_sendctx *sc =
		container_of(cqe, struct rpcrdma_sendctx, sc_cqe);

135 136 137 138 139
	/* WARNING: Only wr_cqe and status are reliable at this point */
	if (wc->status != IB_WC_SUCCESS && wc->status != IB_WC_WR_FLUSH_ERR)
		pr_err("rpcrdma: Send: %s (%u/0x%x)\n",
		       ib_wc_status_msg(wc->status),
		       wc->status, wc->vendor_err);
140 141

	rpcrdma_sendctx_put_locked(sc);
142
}
143

144
/**
145
 * rpcrdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
146 147 148 149
 * @cq:	completion queue (ignored)
 * @wc:	completed WR
 *
 */
150
static void
151
rpcrdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
152
{
153 154 155
	struct ib_cqe *cqe = wc->wr_cqe;
	struct rpcrdma_rep *rep = container_of(cqe, struct rpcrdma_rep,
					       rr_cqe);
156

157 158 159
	/* WARNING: Only wr_id and status are reliable at this point */
	if (wc->status != IB_WC_SUCCESS)
		goto out_fail;
160

161 162 163 164
	/* status == SUCCESS means all fields in wc are trustworthy */
	dprintk("RPC:       %s: rep %p opcode 'recv', length %u: success\n",
		__func__, rep, wc->byte_len);

165
	rpcrdma_set_xdrlen(&rep->rr_hdrbuf, wc->byte_len);
166 167 168
	rep->rr_wc_flags = wc->wc_flags;
	rep->rr_inv_rkey = wc->ex.invalidate_rkey;

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
out_schedule:
174
	rpcrdma_reply_handler(rep);
175
	return;
176

177 178
out_fail:
	if (wc->status != IB_WC_WR_FLUSH_ERR)
179 180 181
		pr_err("rpcrdma: Recv: %s (%u/0x%x)\n",
		       ib_wc_status_msg(wc->status),
		       wc->status, wc->vendor_err);
182
	rpcrdma_set_xdrlen(&rep->rr_hdrbuf, 0);
183
	goto out_schedule;
184 185
}

186 187 188 189 190 191 192 193
static void
rpcrdma_update_connect_private(struct rpcrdma_xprt *r_xprt,
			       struct rdma_conn_param *param)
{
	struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
	const struct rpcrdma_connect_private *pmsg = param->private_data;
	unsigned int rsize, wsize;

194
	/* Default settings for RPC-over-RDMA Version One */
195
	r_xprt->rx_ia.ri_implicit_roundup = xprt_rdma_pad_optimize;
196 197 198 199 200 201
	rsize = RPCRDMA_V1_DEF_INLINE_SIZE;
	wsize = RPCRDMA_V1_DEF_INLINE_SIZE;

	if (pmsg &&
	    pmsg->cp_magic == rpcrdma_cmp_magic &&
	    pmsg->cp_version == RPCRDMA_CMP_VERSION) {
202
		r_xprt->rx_ia.ri_implicit_roundup = true;
203 204 205 206 207 208 209 210
		rsize = rpcrdma_decode_buffer_size(pmsg->cp_send_size);
		wsize = rpcrdma_decode_buffer_size(pmsg->cp_recv_size);
	}

	if (rsize < cdata->inline_rsize)
		cdata->inline_rsize = rsize;
	if (wsize < cdata->inline_wsize)
		cdata->inline_wsize = wsize;
211 212
	dprintk("RPC:       %s: max send %u, max recv %u\n",
		__func__, cdata->inline_wsize, cdata->inline_rsize);
213 214 215
	rpcrdma_set_max_header_sizes(r_xprt);
}

216 217 218 219 220 221 222 223 224 225 226
static int
rpcrdma_conn_upcall(struct rdma_cm_id *id, struct rdma_cm_event *event)
{
	struct rpcrdma_xprt *xprt = id->context;
	struct rpcrdma_ia *ia = &xprt->rx_ia;
	struct rpcrdma_ep *ep = &xprt->rx_ep;
	int connstate = 0;

	switch (event->event) {
	case RDMA_CM_EVENT_ADDR_RESOLVED:
	case RDMA_CM_EVENT_ROUTE_RESOLVED:
227
		ia->ri_async_rc = 0;
228 229 230 231 232 233 234 235 236 237 238 239 240 241
		complete(&ia->ri_done);
		break;
	case RDMA_CM_EVENT_ADDR_ERROR:
		ia->ri_async_rc = -EHOSTUNREACH;
		dprintk("RPC:       %s: CM address resolution error, ep 0x%p\n",
			__func__, ep);
		complete(&ia->ri_done);
		break;
	case RDMA_CM_EVENT_ROUTE_ERROR:
		ia->ri_async_rc = -ENETUNREACH;
		dprintk("RPC:       %s: CM route resolution error, ep 0x%p\n",
			__func__, ep);
		complete(&ia->ri_done);
		break;
242 243
	case RDMA_CM_EVENT_DEVICE_REMOVAL:
#if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
244
		pr_info("rpcrdma: removing device %s for %s:%s\n",
245
			ia->ri_device->name,
246
			rpcrdma_addrstr(xprt), rpcrdma_portstr(xprt));
247 248 249 250 251 252 253 254 255 256 257
#endif
		set_bit(RPCRDMA_IAF_REMOVING, &ia->ri_flags);
		ep->rep_connected = -ENODEV;
		xprt_force_disconnect(&xprt->rx_xprt);
		wait_for_completion(&ia->ri_remove_done);

		ia->ri_id = NULL;
		ia->ri_pd = NULL;
		ia->ri_device = NULL;
		/* Return 1 to ensure the core destroys the id. */
		return 1;
258 259
	case RDMA_CM_EVENT_ESTABLISHED:
		connstate = 1;
260
		rpcrdma_update_connect_private(xprt, &event->param.conn);
261 262 263 264 265 266 267 268
		goto connected;
	case RDMA_CM_EVENT_CONNECT_ERROR:
		connstate = -ENOTCONN;
		goto connected;
	case RDMA_CM_EVENT_UNREACHABLE:
		connstate = -ENETDOWN;
		goto connected;
	case RDMA_CM_EVENT_REJECTED:
269 270
		dprintk("rpcrdma: connection to %s:%s rejected: %s\n",
			rpcrdma_addrstr(xprt), rpcrdma_portstr(xprt),
271
			rdma_reject_msg(id, event->status));
272
		connstate = -ECONNREFUSED;
273 274
		if (event->status == IB_CM_REJ_STALE_CONN)
			connstate = -EAGAIN;
275 276 277 278
		goto connected;
	case RDMA_CM_EVENT_DISCONNECTED:
		connstate = -ECONNABORTED;
connected:
279
		xprt->rx_buf.rb_credits = 1;
280
		ep->rep_connected = connstate;
281
		rpcrdma_conn_func(ep);
282
		wake_up_all(&ep->rep_connect_wait);
283
		/*FALLTHROUGH*/
284
	default:
285 286 287
		dprintk("RPC:       %s: %s:%s on %s/%s (ep 0x%p): %s\n",
			__func__,
			rpcrdma_addrstr(xprt), rpcrdma_portstr(xprt),
288 289
			ia->ri_device->name, ia->ri_ops->ro_displayname,
			ep, rdma_event_msg(event->event));
290 291 292 293 294 295 296
		break;
	}

	return 0;
}

static struct rdma_cm_id *
297
rpcrdma_create_id(struct rpcrdma_xprt *xprt, struct rpcrdma_ia *ia)
298
{
299
	unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1;
300 301 302
	struct rdma_cm_id *id;
	int rc;

303
	init_completion(&ia->ri_done);
304
	init_completion(&ia->ri_remove_done);
305

306 307
	id = rdma_create_id(&init_net, rpcrdma_conn_upcall, xprt, RDMA_PS_TCP,
			    IB_QPT_RC);
308 309 310 311 312 313 314
	if (IS_ERR(id)) {
		rc = PTR_ERR(id);
		dprintk("RPC:       %s: rdma_create_id() failed %i\n",
			__func__, rc);
		return id;
	}

315
	ia->ri_async_rc = -ETIMEDOUT;
316 317 318
	rc = rdma_resolve_addr(id, NULL,
			       (struct sockaddr *)&xprt->rx_xprt.addr,
			       RDMA_RESOLVE_TIMEOUT);
319 320 321 322 323
	if (rc) {
		dprintk("RPC:       %s: rdma_resolve_addr() failed %i\n",
			__func__, rc);
		goto out;
	}
324 325 326 327 328 329
	rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
	if (rc < 0) {
		dprintk("RPC:       %s: wait() exited: %i\n",
			__func__, rc);
		goto out;
	}
330

331 332 333 334
	rc = ia->ri_async_rc;
	if (rc)
		goto out;

335
	ia->ri_async_rc = -ETIMEDOUT;
336 337 338 339
	rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
	if (rc) {
		dprintk("RPC:       %s: rdma_resolve_route() failed %i\n",
			__func__, rc);
340
		goto out;
341
	}
342 343 344 345
	rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
	if (rc < 0) {
		dprintk("RPC:       %s: wait() exited: %i\n",
			__func__, rc);
346
		goto out;
347
	}
348 349
	rc = ia->ri_async_rc;
	if (rc)
350
		goto out;
351 352

	return id;
353

354 355 356 357 358 359 360 361 362
out:
	rdma_destroy_id(id);
	return ERR_PTR(rc);
}

/*
 * Exported functions.
 */

363 364
/**
 * rpcrdma_ia_open - Open and initialize an Interface Adapter.
365
 * @xprt: transport with IA to (re)initialize
366 367 368
 *
 * Returns 0 on success, negative errno if an appropriate
 * Interface Adapter could not be found and opened.
369 370
 */
int
371
rpcrdma_ia_open(struct rpcrdma_xprt *xprt)
372 373
{
	struct rpcrdma_ia *ia = &xprt->rx_ia;
374 375
	int rc;

376
	ia->ri_id = rpcrdma_create_id(xprt, ia);
377 378
	if (IS_ERR(ia->ri_id)) {
		rc = PTR_ERR(ia->ri_id);
379
		goto out_err;
380
	}
381
	ia->ri_device = ia->ri_id->device;
382

383
	ia->ri_pd = ib_alloc_pd(ia->ri_device, 0);
384 385
	if (IS_ERR(ia->ri_pd)) {
		rc = PTR_ERR(ia->ri_pd);
386
		pr_err("rpcrdma: ib_alloc_pd() returned %d\n", rc);
387
		goto out_err;
388 389
	}

390
	switch (xprt_rdma_memreg_strategy) {
391
	case RPCRDMA_FRWR:
392 393 394 395 396
		if (frwr_is_supported(ia)) {
			ia->ri_ops = &rpcrdma_frwr_memreg_ops;
			break;
		}
		/*FALLTHROUGH*/
397
	case RPCRDMA_MTHCAFMR:
398 399 400 401 402
		if (fmr_is_supported(ia)) {
			ia->ri_ops = &rpcrdma_fmr_memreg_ops;
			break;
		}
		/*FALLTHROUGH*/
403
	default:
404 405
		pr_err("rpcrdma: Device %s does not support memreg mode %d\n",
		       ia->ri_device->name, xprt_rdma_memreg_strategy);
406
		rc = -EINVAL;
407
		goto out_err;
408 409 410
	}

	return 0;
411

412 413
out_err:
	rpcrdma_ia_close(ia);
414 415 416
	return rc;
}

417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460
/**
 * rpcrdma_ia_remove - Handle device driver unload
 * @ia: interface adapter being removed
 *
 * Divest transport H/W resources associated with this adapter,
 * but allow it to be restored later.
 */
void
rpcrdma_ia_remove(struct rpcrdma_ia *ia)
{
	struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
						   rx_ia);
	struct rpcrdma_ep *ep = &r_xprt->rx_ep;
	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
	struct rpcrdma_req *req;
	struct rpcrdma_rep *rep;

	cancel_delayed_work_sync(&buf->rb_refresh_worker);

	/* This is similar to rpcrdma_ep_destroy, but:
	 * - Don't cancel the connect worker.
	 * - Don't call rpcrdma_ep_disconnect, which waits
	 *   for another conn upcall, which will deadlock.
	 * - rdma_disconnect is unneeded, the underlying
	 *   connection is already gone.
	 */
	if (ia->ri_id->qp) {
		ib_drain_qp(ia->ri_id->qp);
		rdma_destroy_qp(ia->ri_id);
		ia->ri_id->qp = NULL;
	}
	ib_free_cq(ep->rep_attr.recv_cq);
	ib_free_cq(ep->rep_attr.send_cq);

	/* The ULP is responsible for ensuring all DMA
	 * mappings and MRs are gone.
	 */
	list_for_each_entry(rep, &buf->rb_recv_bufs, rr_list)
		rpcrdma_dma_unmap_regbuf(rep->rr_rdmabuf);
	list_for_each_entry(req, &buf->rb_allreqs, rl_all) {
		rpcrdma_dma_unmap_regbuf(req->rl_rdmabuf);
		rpcrdma_dma_unmap_regbuf(req->rl_sendbuf);
		rpcrdma_dma_unmap_regbuf(req->rl_recvbuf);
	}
C
Chuck Lever 已提交
461
	rpcrdma_mrs_destroy(buf);
462 463 464 465 466

	/* Allow waiters to continue */
	complete(&ia->ri_remove_done);
}

467 468 469 470
/**
 * rpcrdma_ia_close - Clean up/close an IA.
 * @ia: interface adapter to close
 *
471 472 473 474 475
 */
void
rpcrdma_ia_close(struct rpcrdma_ia *ia)
{
	dprintk("RPC:       %s: entering\n", __func__);
476 477 478
	if (ia->ri_id != NULL && !IS_ERR(ia->ri_id)) {
		if (ia->ri_id->qp)
			rdma_destroy_qp(ia->ri_id);
479
		rdma_destroy_id(ia->ri_id);
480
	}
481 482
	ia->ri_id = NULL;
	ia->ri_device = NULL;
483 484 485

	/* If the pd is still busy, xprtrdma missed freeing a resource */
	if (ia->ri_pd && !IS_ERR(ia->ri_pd))
486
		ib_dealloc_pd(ia->ri_pd);
487
	ia->ri_pd = NULL;
488 489 490 491 492 493 494
}

/*
 * Create unconnected endpoint.
 */
int
rpcrdma_ep_create(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia,
495
		  struct rpcrdma_create_data_internal *cdata)
496
{
497
	struct rpcrdma_connect_private *pmsg = &ep->rep_cm_private;
498
	unsigned int max_qp_wr, max_sge;
499
	struct ib_cq *sendcq, *recvcq;
500
	int rc;
501

502 503
	max_sge = min_t(unsigned int, ia->ri_device->attrs.max_sge,
			RPCRDMA_MAX_SEND_SGES);
504 505
	if (max_sge < RPCRDMA_MIN_SEND_SGES) {
		pr_warn("rpcrdma: HCA provides only %d send SGEs\n", max_sge);
506 507
		return -ENOMEM;
	}
508
	ia->ri_max_send_sges = max_sge - RPCRDMA_MIN_SEND_SGES;
509

510
	if (ia->ri_device->attrs.max_qp_wr <= RPCRDMA_BACKWARD_WRS) {
511 512 513 514
		dprintk("RPC:       %s: insufficient wqe's available\n",
			__func__);
		return -ENOMEM;
	}
515
	max_qp_wr = ia->ri_device->attrs.max_qp_wr - RPCRDMA_BACKWARD_WRS - 1;
516

517
	/* check provider's send/recv wr limits */
518 519
	if (cdata->max_requests > max_qp_wr)
		cdata->max_requests = max_qp_wr;
520 521 522 523 524

	ep->rep_attr.event_handler = rpcrdma_qp_async_error_upcall;
	ep->rep_attr.qp_context = ep;
	ep->rep_attr.srq = NULL;
	ep->rep_attr.cap.max_send_wr = cdata->max_requests;
525
	ep->rep_attr.cap.max_send_wr += RPCRDMA_BACKWARD_WRS;
526
	ep->rep_attr.cap.max_send_wr += 1;	/* drain cqe */
C
Chuck Lever 已提交
527 528 529
	rc = ia->ri_ops->ro_open(ia, ep, cdata);
	if (rc)
		return rc;
530
	ep->rep_attr.cap.max_recv_wr = cdata->max_requests;
531
	ep->rep_attr.cap.max_recv_wr += RPCRDMA_BACKWARD_WRS;
532
	ep->rep_attr.cap.max_recv_wr += 1;	/* drain cqe */
533
	ep->rep_attr.cap.max_send_sge = max_sge;
534 535 536 537 538 539 540 541 542 543 544 545 546 547 548
	ep->rep_attr.cap.max_recv_sge = 1;
	ep->rep_attr.cap.max_inline_data = 0;
	ep->rep_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
	ep->rep_attr.qp_type = IB_QPT_RC;
	ep->rep_attr.port_num = ~0;

	dprintk("RPC:       %s: requested max: dtos: send %d recv %d; "
		"iovs: send %d recv %d\n",
		__func__,
		ep->rep_attr.cap.max_send_wr,
		ep->rep_attr.cap.max_recv_wr,
		ep->rep_attr.cap.max_send_sge,
		ep->rep_attr.cap.max_recv_sge);

	/* set trigger for requesting send completion */
549 550 551
	ep->rep_send_batch = min_t(unsigned int, RPCRDMA_MAX_SEND_BATCH,
				   cdata->max_requests >> 2);
	ep->rep_send_count = ep->rep_send_batch;
552
	init_waitqueue_head(&ep->rep_connect_wait);
553
	INIT_DELAYED_WORK(&ep->rep_connect_worker, rpcrdma_connect_worker);
554

555 556
	sendcq = ib_alloc_cq(ia->ri_device, NULL,
			     ep->rep_attr.cap.max_send_wr + 1,
557
			     1, IB_POLL_WORKQUEUE);
558 559 560
	if (IS_ERR(sendcq)) {
		rc = PTR_ERR(sendcq);
		dprintk("RPC:       %s: failed to create send CQ: %i\n",
561 562 563 564
			__func__, rc);
		goto out1;
	}

565 566
	recvcq = ib_alloc_cq(ia->ri_device, NULL,
			     ep->rep_attr.cap.max_recv_wr + 1,
567
			     0, IB_POLL_WORKQUEUE);
568 569 570 571 572 573 574 575 576
	if (IS_ERR(recvcq)) {
		rc = PTR_ERR(recvcq);
		dprintk("RPC:       %s: failed to create recv CQ: %i\n",
			__func__, rc);
		goto out2;
	}

	ep->rep_attr.send_cq = sendcq;
	ep->rep_attr.recv_cq = recvcq;
577 578

	/* Initialize cma parameters */
579
	memset(&ep->rep_remote_cma, 0, sizeof(ep->rep_remote_cma));
580

581 582 583
	/* Prepare RDMA-CM private message */
	pmsg->cp_magic = rpcrdma_cmp_magic;
	pmsg->cp_version = RPCRDMA_CMP_VERSION;
584
	pmsg->cp_flags |= ia->ri_ops->ro_send_w_inv_ok;
585 586 587 588
	pmsg->cp_send_size = rpcrdma_encode_buffer_size(cdata->inline_wsize);
	pmsg->cp_recv_size = rpcrdma_encode_buffer_size(cdata->inline_rsize);
	ep->rep_remote_cma.private_data = pmsg;
	ep->rep_remote_cma.private_data_len = sizeof(*pmsg);
589 590

	/* Client offers RDMA Read but does not initiate */
591
	ep->rep_remote_cma.initiator_depth = 0;
592
	if (ia->ri_device->attrs.max_qp_rd_atom > 32)	/* arbitrary but <= 255 */
593 594
		ep->rep_remote_cma.responder_resources = 32;
	else
595
		ep->rep_remote_cma.responder_resources =
596
						ia->ri_device->attrs.max_qp_rd_atom;
597

598 599 600 601 602 603 604 605 606 607
	/* Limit transport retries so client can detect server
	 * GID changes quickly. RPC layer handles re-establishing
	 * transport connection and retransmission.
	 */
	ep->rep_remote_cma.retry_count = 6;

	/* RPC-over-RDMA handles its own flow control. In addition,
	 * make all RNR NAKs visible so we know that RPC-over-RDMA
	 * flow control is working correctly (no NAKs should be seen).
	 */
608 609 610 611 612 613
	ep->rep_remote_cma.flow_control = 0;
	ep->rep_remote_cma.rnr_retry_count = 0;

	return 0;

out2:
614
	ib_free_cq(sendcq);
615 616 617 618 619 620 621 622 623 624 625
out1:
	return rc;
}

/*
 * rpcrdma_ep_destroy
 *
 * Disconnect and destroy endpoint. After this, the only
 * valid operations on the ep are to free it (if dynamically
 * allocated) or re-create it.
 */
626
void
627 628 629 630 631
rpcrdma_ep_destroy(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
{
	dprintk("RPC:       %s: entering, connected is %d\n",
		__func__, ep->rep_connected);

632 633
	cancel_delayed_work_sync(&ep->rep_connect_worker);

634
	if (ia->ri_id->qp) {
635
		rpcrdma_ep_disconnect(ep, ia);
636 637
		rdma_destroy_qp(ia->ri_id);
		ia->ri_id->qp = NULL;
638 639
	}

640
	ib_free_cq(ep->rep_attr.recv_cq);
641
	ib_free_cq(ep->rep_attr.send_cq);
642 643
}

644 645 646 647 648 649 650 651 652 653 654 655 656
/* Re-establish a connection after a device removal event.
 * Unlike a normal reconnection, a fresh PD and a new set
 * of MRs and buffers is needed.
 */
static int
rpcrdma_ep_recreate_xprt(struct rpcrdma_xprt *r_xprt,
			 struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
{
	int rc, err;

	pr_info("%s: r_xprt = %p\n", __func__, r_xprt);

	rc = -EHOSTUNREACH;
657
	if (rpcrdma_ia_open(r_xprt))
658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673
		goto out1;

	rc = -ENOMEM;
	err = rpcrdma_ep_create(ep, ia, &r_xprt->rx_data);
	if (err) {
		pr_err("rpcrdma: rpcrdma_ep_create returned %d\n", err);
		goto out2;
	}

	rc = -ENETUNREACH;
	err = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
	if (err) {
		pr_err("rpcrdma: rdma_create_qp returned %d\n", err);
		goto out3;
	}

C
Chuck Lever 已提交
674
	rpcrdma_mrs_create(r_xprt);
675 676 677 678 679 680 681 682 683 684
	return 0;

out3:
	rpcrdma_ep_destroy(ep, ia);
out2:
	rpcrdma_ia_close(ia);
out1:
	return rc;
}

685 686 687 688 689 690 691 692 693 694 695 696
static int
rpcrdma_ep_reconnect(struct rpcrdma_xprt *r_xprt, struct rpcrdma_ep *ep,
		     struct rpcrdma_ia *ia)
{
	struct rdma_cm_id *id, *old;
	int err, rc;

	dprintk("RPC:       %s: reconnecting...\n", __func__);

	rpcrdma_ep_disconnect(ep, ia);

	rc = -EHOSTUNREACH;
697
	id = rpcrdma_create_id(r_xprt, ia);
698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729
	if (IS_ERR(id))
		goto out;

	/* As long as the new ID points to the same device as the
	 * old ID, we can reuse the transport's existing PD and all
	 * previously allocated MRs. Also, the same device means
	 * the transport's previous DMA mappings are still valid.
	 *
	 * This is a sanity check only. There should be no way these
	 * point to two different devices here.
	 */
	old = id;
	rc = -ENETUNREACH;
	if (ia->ri_device != id->device) {
		pr_err("rpcrdma: can't reconnect on different device!\n");
		goto out_destroy;
	}

	err = rdma_create_qp(id, ia->ri_pd, &ep->rep_attr);
	if (err) {
		dprintk("RPC:       %s: rdma_create_qp returned %d\n",
			__func__, err);
		goto out_destroy;
	}

	/* Atomically replace the transport's ID and QP. */
	rc = 0;
	old = ia->ri_id;
	ia->ri_id = id;
	rdma_destroy_qp(old);

out_destroy:
730
	rdma_destroy_id(old);
731 732 733 734
out:
	return rc;
}

735 736 737 738 739 740
/*
 * Connect unconnected endpoint.
 */
int
rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
{
741 742 743
	struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
						   rx_ia);
	unsigned int extras;
744
	int rc;
745 746

retry:
747 748
	switch (ep->rep_connected) {
	case 0:
749 750 751 752 753
		dprintk("RPC:       %s: connecting...\n", __func__);
		rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
		if (rc) {
			dprintk("RPC:       %s: rdma_create_qp failed %i\n",
				__func__, rc);
754 755
			rc = -ENETUNREACH;
			goto out_noupdate;
756
		}
757
		break;
758 759 760 761 762
	case -ENODEV:
		rc = rpcrdma_ep_recreate_xprt(r_xprt, ep, ia);
		if (rc)
			goto out_noupdate;
		break;
763 764 765 766
	default:
		rc = rpcrdma_ep_reconnect(r_xprt, ep, ia);
		if (rc)
			goto out;
767 768 769 770 771 772 773 774 775 776 777 778 779
	}

	ep->rep_connected = 0;

	rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma);
	if (rc) {
		dprintk("RPC:       %s: rdma_connect() failed with %i\n",
				__func__, rc);
		goto out;
	}

	wait_event_interruptible(ep->rep_connect_wait, ep->rep_connected != 0);
	if (ep->rep_connected <= 0) {
780
		if (ep->rep_connected == -EAGAIN)
781 782
			goto retry;
		rc = ep->rep_connected;
783
		goto out;
784 785
	}

786 787 788 789 790
	dprintk("RPC:       %s: connected\n", __func__);
	extras = r_xprt->rx_buf.rb_bc_srv_max_requests;
	if (extras)
		rpcrdma_ep_post_extra_recv(r_xprt, extras);

791 792 793
out:
	if (rc)
		ep->rep_connected = rc;
794 795

out_noupdate:
796 797 798 799 800 801 802 803 804 805 806 807
	return rc;
}

/*
 * rpcrdma_ep_disconnect
 *
 * This is separate from destroy to facilitate the ability
 * to reconnect without recreating the endpoint.
 *
 * This call is not reentrant, and must not be made in parallel
 * on the same endpoint.
 */
808
void
809 810 811 812 813 814 815 816 817 818 819 820 821 822 823
rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
{
	int rc;

	rc = rdma_disconnect(ia->ri_id);
	if (!rc) {
		/* returns without wait if not connected */
		wait_event_interruptible(ep->rep_connect_wait,
							ep->rep_connected != 1);
		dprintk("RPC:       %s: after wait, %sconnected\n", __func__,
			(ep->rep_connected == 1) ? "still " : "dis");
	} else {
		dprintk("RPC:       %s: rdma_disconnect %i\n", __func__, rc);
		ep->rep_connected = rc;
	}
824 825

	ib_drain_qp(ia->ri_id->qp);
826 827
}

828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989
/* Fixed-size circular FIFO queue. This implementation is wait-free and
 * lock-free.
 *
 * Consumer is the code path that posts Sends. This path dequeues a
 * sendctx for use by a Send operation. Multiple consumer threads
 * are serialized by the RPC transport lock, which allows only one
 * ->send_request call at a time.
 *
 * Producer is the code path that handles Send completions. This path
 * enqueues a sendctx that has been completed. Multiple producer
 * threads are serialized by the ib_poll_cq() function.
 */

/* rpcrdma_sendctxs_destroy() assumes caller has already quiesced
 * queue activity, and ib_drain_qp has flushed all remaining Send
 * requests.
 */
static void rpcrdma_sendctxs_destroy(struct rpcrdma_buffer *buf)
{
	unsigned long i;

	for (i = 0; i <= buf->rb_sc_last; i++)
		kfree(buf->rb_sc_ctxs[i]);
	kfree(buf->rb_sc_ctxs);
}

static struct rpcrdma_sendctx *rpcrdma_sendctx_create(struct rpcrdma_ia *ia)
{
	struct rpcrdma_sendctx *sc;

	sc = kzalloc(sizeof(*sc) +
		     ia->ri_max_send_sges * sizeof(struct ib_sge),
		     GFP_KERNEL);
	if (!sc)
		return NULL;

	sc->sc_wr.wr_cqe = &sc->sc_cqe;
	sc->sc_wr.sg_list = sc->sc_sges;
	sc->sc_wr.opcode = IB_WR_SEND;
	sc->sc_cqe.done = rpcrdma_wc_send;
	return sc;
}

static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt)
{
	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
	struct rpcrdma_sendctx *sc;
	unsigned long i;

	/* Maximum number of concurrent outstanding Send WRs. Capping
	 * the circular queue size stops Send Queue overflow by causing
	 * the ->send_request call to fail temporarily before too many
	 * Sends are posted.
	 */
	i = buf->rb_max_requests + RPCRDMA_MAX_BC_REQUESTS;
	dprintk("RPC:       %s: allocating %lu send_ctxs\n", __func__, i);
	buf->rb_sc_ctxs = kcalloc(i, sizeof(sc), GFP_KERNEL);
	if (!buf->rb_sc_ctxs)
		return -ENOMEM;

	buf->rb_sc_last = i - 1;
	for (i = 0; i <= buf->rb_sc_last; i++) {
		sc = rpcrdma_sendctx_create(&r_xprt->rx_ia);
		if (!sc)
			goto out_destroy;

		sc->sc_xprt = r_xprt;
		buf->rb_sc_ctxs[i] = sc;
	}

	return 0;

out_destroy:
	rpcrdma_sendctxs_destroy(buf);
	return -ENOMEM;
}

/* The sendctx queue is not guaranteed to have a size that is a
 * power of two, thus the helpers in circ_buf.h cannot be used.
 * The other option is to use modulus (%), which can be expensive.
 */
static unsigned long rpcrdma_sendctx_next(struct rpcrdma_buffer *buf,
					  unsigned long item)
{
	return likely(item < buf->rb_sc_last) ? item + 1 : 0;
}

/**
 * rpcrdma_sendctx_get_locked - Acquire a send context
 * @buf: transport buffers from which to acquire an unused context
 *
 * Returns pointer to a free send completion context; or NULL if
 * the queue is empty.
 *
 * Usage: Called to acquire an SGE array before preparing a Send WR.
 *
 * The caller serializes calls to this function (per rpcrdma_buffer),
 * and provides an effective memory barrier that flushes the new value
 * of rb_sc_head.
 */
struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_buffer *buf)
{
	struct rpcrdma_xprt *r_xprt;
	struct rpcrdma_sendctx *sc;
	unsigned long next_head;

	next_head = rpcrdma_sendctx_next(buf, buf->rb_sc_head);

	if (next_head == READ_ONCE(buf->rb_sc_tail))
		goto out_emptyq;

	/* ORDER: item must be accessed _before_ head is updated */
	sc = buf->rb_sc_ctxs[next_head];

	/* Releasing the lock in the caller acts as a memory
	 * barrier that flushes rb_sc_head.
	 */
	buf->rb_sc_head = next_head;

	return sc;

out_emptyq:
	/* The queue is "empty" if there have not been enough Send
	 * completions recently. This is a sign the Send Queue is
	 * backing up. Cause the caller to pause and try again.
	 */
	dprintk("RPC:       %s: empty sendctx queue\n", __func__);
	r_xprt = container_of(buf, struct rpcrdma_xprt, rx_buf);
	r_xprt->rx_stats.empty_sendctx_q++;
	return NULL;
}

/**
 * rpcrdma_sendctx_put_locked - Release a send context
 * @sc: send context to release
 *
 * Usage: Called from Send completion to return a sendctxt
 * to the queue.
 *
 * The caller serializes calls to this function (per rpcrdma_buffer).
 */
void rpcrdma_sendctx_put_locked(struct rpcrdma_sendctx *sc)
{
	struct rpcrdma_buffer *buf = &sc->sc_xprt->rx_buf;
	unsigned long next_tail;

	/* Unmap SGEs of previously completed by unsignaled
	 * Sends by walking up the queue until @sc is found.
	 */
	next_tail = buf->rb_sc_tail;
	do {
		next_tail = rpcrdma_sendctx_next(buf, next_tail);

		/* ORDER: item must be accessed _before_ tail is updated */
		rpcrdma_unmap_sendctx(buf->rb_sc_ctxs[next_tail]);

	} while (buf->rb_sc_ctxs[next_tail] != sc);

	/* Paired with READ_ONCE */
	smp_store_release(&buf->rb_sc_tail, next_tail);
}

990 991 992 993 994
static void
rpcrdma_mr_recovery_worker(struct work_struct *work)
{
	struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
						  rb_recovery_worker.work);
C
Chuck Lever 已提交
995
	struct rpcrdma_mr *mr;
996 997 998

	spin_lock(&buf->rb_recovery_lock);
	while (!list_empty(&buf->rb_stale_mrs)) {
C
Chuck Lever 已提交
999
		mr = rpcrdma_mr_pop(&buf->rb_stale_mrs);
1000 1001
		spin_unlock(&buf->rb_recovery_lock);

C
Chuck Lever 已提交
1002 1003
		dprintk("RPC:       %s: recovering MR %p\n", __func__, mr);
		mr->mr_xprt->rx_ia.ri_ops->ro_recover_mr(mr);
1004 1005

		spin_lock(&buf->rb_recovery_lock);
1006
	}
1007 1008 1009 1010
	spin_unlock(&buf->rb_recovery_lock);
}

void
C
Chuck Lever 已提交
1011
rpcrdma_mr_defer_recovery(struct rpcrdma_mr *mr)
1012
{
C
Chuck Lever 已提交
1013
	struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
1014 1015 1016
	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;

	spin_lock(&buf->rb_recovery_lock);
C
Chuck Lever 已提交
1017
	rpcrdma_mr_push(mr, &buf->rb_stale_mrs);
1018 1019 1020 1021 1022
	spin_unlock(&buf->rb_recovery_lock);

	schedule_delayed_work(&buf->rb_recovery_worker, 0);
}

C
Chuck Lever 已提交
1023
static void
C
Chuck Lever 已提交
1024
rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt)
C
Chuck Lever 已提交
1025 1026 1027 1028 1029 1030 1031 1032
{
	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
	unsigned int count;
	LIST_HEAD(free);
	LIST_HEAD(all);

	for (count = 0; count < 32; count++) {
C
Chuck Lever 已提交
1033
		struct rpcrdma_mr *mr;
C
Chuck Lever 已提交
1034 1035
		int rc;

C
Chuck Lever 已提交
1036 1037
		mr = kzalloc(sizeof(*mr), GFP_KERNEL);
		if (!mr)
C
Chuck Lever 已提交
1038 1039
			break;

C
Chuck Lever 已提交
1040
		rc = ia->ri_ops->ro_init_mr(ia, mr);
C
Chuck Lever 已提交
1041
		if (rc) {
C
Chuck Lever 已提交
1042
			kfree(mr);
C
Chuck Lever 已提交
1043 1044 1045
			break;
		}

C
Chuck Lever 已提交
1046
		mr->mr_xprt = r_xprt;
C
Chuck Lever 已提交
1047

C
Chuck Lever 已提交
1048 1049
		list_add(&mr->mr_list, &free);
		list_add(&mr->mr_all, &all);
C
Chuck Lever 已提交
1050 1051
	}

C
Chuck Lever 已提交
1052 1053
	spin_lock(&buf->rb_mrlock);
	list_splice(&free, &buf->rb_mrs);
C
Chuck Lever 已提交
1054 1055
	list_splice(&all, &buf->rb_all);
	r_xprt->rx_stats.mrs_allocated += count;
C
Chuck Lever 已提交
1056
	spin_unlock(&buf->rb_mrlock);
C
Chuck Lever 已提交
1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068

	dprintk("RPC:       %s: created %u MRs\n", __func__, count);
}

static void
rpcrdma_mr_refresh_worker(struct work_struct *work)
{
	struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
						  rb_refresh_worker.work);
	struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
						   rx_buf);

C
Chuck Lever 已提交
1069
	rpcrdma_mrs_create(r_xprt);
C
Chuck Lever 已提交
1070 1071
}

1072
struct rpcrdma_req *
1073 1074
rpcrdma_create_req(struct rpcrdma_xprt *r_xprt)
{
1075
	struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
1076 1077
	struct rpcrdma_req *req;

1078
	req = kzalloc(sizeof(*req), GFP_KERNEL);
1079
	if (req == NULL)
1080
		return ERR_PTR(-ENOMEM);
1081

1082 1083 1084
	spin_lock(&buffer->rb_reqslock);
	list_add(&req->rl_all, &buffer->rb_allreqs);
	spin_unlock(&buffer->rb_reqslock);
1085
	req->rl_buffer = &r_xprt->rx_buf;
1086
	INIT_LIST_HEAD(&req->rl_registered);
1087 1088 1089
	return req;
}

1090 1091 1092 1093 1094 1095 1096
/**
 * rpcrdma_create_rep - Allocate an rpcrdma_rep object
 * @r_xprt: controlling transport
 *
 * Returns 0 on success or a negative errno on failure.
 */
int
1097 1098 1099
rpcrdma_create_rep(struct rpcrdma_xprt *r_xprt)
{
	struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
1100
	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1101 1102 1103 1104
	struct rpcrdma_rep *rep;
	int rc;

	rc = -ENOMEM;
1105
	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1106 1107 1108
	if (rep == NULL)
		goto out;

1109
	rep->rr_rdmabuf = rpcrdma_alloc_regbuf(cdata->inline_rsize,
1110
					       DMA_FROM_DEVICE, GFP_KERNEL);
1111 1112
	if (IS_ERR(rep->rr_rdmabuf)) {
		rc = PTR_ERR(rep->rr_rdmabuf);
1113
		goto out_free;
1114
	}
1115 1116
	xdr_buf_init(&rep->rr_hdrbuf, rep->rr_rdmabuf->rg_base,
		     rdmab_length(rep->rr_rdmabuf));
1117

1118
	rep->rr_cqe.done = rpcrdma_wc_receive;
1119
	rep->rr_rxprt = r_xprt;
1120
	INIT_WORK(&rep->rr_work, rpcrdma_deferred_completion);
1121 1122 1123 1124
	rep->rr_recv_wr.next = NULL;
	rep->rr_recv_wr.wr_cqe = &rep->rr_cqe;
	rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
	rep->rr_recv_wr.num_sge = 1;
1125 1126 1127 1128 1129

	spin_lock(&buf->rb_lock);
	list_add(&rep->rr_list, &buf->rb_recv_bufs);
	spin_unlock(&buf->rb_lock);
	return 0;
1130 1131 1132 1133

out_free:
	kfree(rep);
out:
1134 1135 1136
	dprintk("RPC:       %s: reply buffer %d alloc failed\n",
		__func__, rc);
	return rc;
1137 1138
}

1139
int
1140
rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
1141
{
1142
	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1143 1144
	int i, rc;

1145
	buf->rb_max_requests = r_xprt->rx_data.max_requests;
1146
	buf->rb_bc_srv_max_requests = 0;
C
Chuck Lever 已提交
1147
	spin_lock_init(&buf->rb_mrlock);
1148 1149
	spin_lock_init(&buf->rb_lock);
	spin_lock_init(&buf->rb_recovery_lock);
C
Chuck Lever 已提交
1150
	INIT_LIST_HEAD(&buf->rb_mrs);
C
Chuck Lever 已提交
1151
	INIT_LIST_HEAD(&buf->rb_all);
1152
	INIT_LIST_HEAD(&buf->rb_stale_mrs);
C
Chuck Lever 已提交
1153 1154
	INIT_DELAYED_WORK(&buf->rb_refresh_worker,
			  rpcrdma_mr_refresh_worker);
1155 1156
	INIT_DELAYED_WORK(&buf->rb_recovery_worker,
			  rpcrdma_mr_recovery_worker);
1157

C
Chuck Lever 已提交
1158
	rpcrdma_mrs_create(r_xprt);
1159

1160
	INIT_LIST_HEAD(&buf->rb_send_bufs);
1161 1162
	INIT_LIST_HEAD(&buf->rb_allreqs);
	spin_lock_init(&buf->rb_reqslock);
1163 1164 1165
	for (i = 0; i < buf->rb_max_requests; i++) {
		struct rpcrdma_req *req;

1166 1167
		req = rpcrdma_create_req(r_xprt);
		if (IS_ERR(req)) {
1168 1169
			dprintk("RPC:       %s: request buffer %d alloc"
				" failed\n", __func__, i);
1170
			rc = PTR_ERR(req);
1171 1172
			goto out;
		}
1173
		list_add(&req->rl_list, &buf->rb_send_bufs);
1174 1175 1176
	}

	INIT_LIST_HEAD(&buf->rb_recv_bufs);
1177 1178 1179
	for (i = 0; i <= buf->rb_max_requests; i++) {
		rc = rpcrdma_create_rep(r_xprt);
		if (rc)
1180 1181
			goto out;
	}
1182

1183 1184 1185 1186
	rc = rpcrdma_sendctxs_create(r_xprt);
	if (rc)
		goto out;

1187 1188 1189 1190 1191 1192
	return 0;
out:
	rpcrdma_buffer_destroy(buf);
	return rc;
}

1193 1194 1195 1196 1197 1198
static struct rpcrdma_req *
rpcrdma_buffer_get_req_locked(struct rpcrdma_buffer *buf)
{
	struct rpcrdma_req *req;

	req = list_first_entry(&buf->rb_send_bufs,
1199
			       struct rpcrdma_req, rl_list);
1200
	list_del_init(&req->rl_list);
1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214
	return req;
}

static struct rpcrdma_rep *
rpcrdma_buffer_get_rep_locked(struct rpcrdma_buffer *buf)
{
	struct rpcrdma_rep *rep;

	rep = list_first_entry(&buf->rb_recv_bufs,
			       struct rpcrdma_rep, rr_list);
	list_del(&rep->rr_list);
	return rep;
}

1215
static void
1216
rpcrdma_destroy_rep(struct rpcrdma_rep *rep)
1217
{
1218
	rpcrdma_free_regbuf(rep->rr_rdmabuf);
1219 1220 1221
	kfree(rep);
}

1222
void
1223
rpcrdma_destroy_req(struct rpcrdma_req *req)
1224
{
1225 1226 1227
	rpcrdma_free_regbuf(req->rl_recvbuf);
	rpcrdma_free_regbuf(req->rl_sendbuf);
	rpcrdma_free_regbuf(req->rl_rdmabuf);
1228 1229 1230
	kfree(req);
}

C
Chuck Lever 已提交
1231
static void
C
Chuck Lever 已提交
1232
rpcrdma_mrs_destroy(struct rpcrdma_buffer *buf)
C
Chuck Lever 已提交
1233 1234 1235 1236
{
	struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
						   rx_buf);
	struct rpcrdma_ia *ia = rdmab_to_ia(buf);
C
Chuck Lever 已提交
1237
	struct rpcrdma_mr *mr;
C
Chuck Lever 已提交
1238 1239 1240
	unsigned int count;

	count = 0;
C
Chuck Lever 已提交
1241
	spin_lock(&buf->rb_mrlock);
C
Chuck Lever 已提交
1242
	while (!list_empty(&buf->rb_all)) {
C
Chuck Lever 已提交
1243 1244
		mr = list_entry(buf->rb_all.next, struct rpcrdma_mr, mr_all);
		list_del(&mr->mr_all);
C
Chuck Lever 已提交
1245

C
Chuck Lever 已提交
1246 1247
		spin_unlock(&buf->rb_mrlock);
		ia->ri_ops->ro_release_mr(mr);
C
Chuck Lever 已提交
1248
		count++;
C
Chuck Lever 已提交
1249
		spin_lock(&buf->rb_mrlock);
C
Chuck Lever 已提交
1250
	}
C
Chuck Lever 已提交
1251
	spin_unlock(&buf->rb_mrlock);
C
Chuck Lever 已提交
1252 1253 1254 1255 1256
	r_xprt->rx_stats.mrs_allocated = 0;

	dprintk("RPC:       %s: released %u MRs\n", __func__, count);
}

1257 1258 1259
void
rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
{
1260
	cancel_delayed_work_sync(&buf->rb_recovery_worker);
1261
	cancel_delayed_work_sync(&buf->rb_refresh_worker);
1262

1263 1264
	rpcrdma_sendctxs_destroy(buf);

1265 1266
	while (!list_empty(&buf->rb_recv_bufs)) {
		struct rpcrdma_rep *rep;
1267

1268
		rep = rpcrdma_buffer_get_rep_locked(buf);
1269
		rpcrdma_destroy_rep(rep);
1270
	}
1271
	buf->rb_send_count = 0;
1272

1273 1274
	spin_lock(&buf->rb_reqslock);
	while (!list_empty(&buf->rb_allreqs)) {
1275
		struct rpcrdma_req *req;
A
Allen Andrews 已提交
1276

1277 1278 1279 1280 1281
		req = list_first_entry(&buf->rb_allreqs,
				       struct rpcrdma_req, rl_all);
		list_del(&req->rl_all);

		spin_unlock(&buf->rb_reqslock);
1282
		rpcrdma_destroy_req(req);
1283
		spin_lock(&buf->rb_reqslock);
1284
	}
1285
	spin_unlock(&buf->rb_reqslock);
1286
	buf->rb_recv_count = 0;
A
Allen Andrews 已提交
1287

C
Chuck Lever 已提交
1288
	rpcrdma_mrs_destroy(buf);
1289 1290
}

C
Chuck Lever 已提交
1291 1292 1293 1294 1295 1296 1297 1298 1299
/**
 * rpcrdma_mr_get - Allocate an rpcrdma_mr object
 * @r_xprt: controlling transport
 *
 * Returns an initialized rpcrdma_mr or NULL if no free
 * rpcrdma_mr objects are available.
 */
struct rpcrdma_mr *
rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt)
1300
{
1301
	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
C
Chuck Lever 已提交
1302
	struct rpcrdma_mr *mr = NULL;
1303

C
Chuck Lever 已提交
1304 1305 1306 1307
	spin_lock(&buf->rb_mrlock);
	if (!list_empty(&buf->rb_mrs))
		mr = rpcrdma_mr_pop(&buf->rb_mrs);
	spin_unlock(&buf->rb_mrlock);
1308

C
Chuck Lever 已提交
1309 1310 1311
	if (!mr)
		goto out_nomrs;
	return mr;
C
Chuck Lever 已提交
1312

C
Chuck Lever 已提交
1313 1314
out_nomrs:
	dprintk("RPC:       %s: no MRs available\n", __func__);
1315 1316
	if (r_xprt->rx_ep.rep_connected != -ENODEV)
		schedule_delayed_work(&buf->rb_refresh_worker, 0);
C
Chuck Lever 已提交
1317 1318 1319 1320 1321

	/* Allow the reply handler and refresh worker to run */
	cond_resched();

	return NULL;
1322 1323
}

1324 1325 1326 1327 1328 1329 1330 1331
static void
__rpcrdma_mr_put(struct rpcrdma_buffer *buf, struct rpcrdma_mr *mr)
{
	spin_lock(&buf->rb_mrlock);
	rpcrdma_mr_push(mr, &buf->rb_mrs);
	spin_unlock(&buf->rb_mrlock);
}

C
Chuck Lever 已提交
1332 1333 1334 1335 1336
/**
 * rpcrdma_mr_put - Release an rpcrdma_mr object
 * @mr: object to release
 *
 */
1337
void
C
Chuck Lever 已提交
1338
rpcrdma_mr_put(struct rpcrdma_mr *mr)
1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349
{
	__rpcrdma_mr_put(&mr->mr_xprt->rx_buf, mr);
}

/**
 * rpcrdma_mr_unmap_and_put - DMA unmap an MR and release it
 * @mr: object to release
 *
 */
void
rpcrdma_mr_unmap_and_put(struct rpcrdma_mr *mr)
1350
{
C
Chuck Lever 已提交
1351
	struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
1352

1353 1354 1355
	ib_dma_unmap_sg(r_xprt->rx_ia.ri_device,
			mr->mr_sg, mr->mr_nents, mr->mr_dir);
	__rpcrdma_mr_put(&r_xprt->rx_buf, mr);
1356 1357
}

1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374
static struct rpcrdma_rep *
rpcrdma_buffer_get_rep(struct rpcrdma_buffer *buffers)
{
	/* If an RPC previously completed without a reply (say, a
	 * credential problem or a soft timeout occurs) then hold off
	 * on supplying more Receive buffers until the number of new
	 * pending RPCs catches up to the number of posted Receives.
	 */
	if (unlikely(buffers->rb_send_count < buffers->rb_recv_count))
		return NULL;

	if (unlikely(list_empty(&buffers->rb_recv_bufs)))
		return NULL;
	buffers->rb_recv_count++;
	return rpcrdma_buffer_get_rep_locked(buffers);
}

1375 1376
/*
 * Get a set of request/reply buffers.
1377 1378
 *
 * Reply buffer (if available) is attached to send buffer upon return.
1379 1380 1381 1382 1383
 */
struct rpcrdma_req *
rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
{
	struct rpcrdma_req *req;
1384

1385
	spin_lock(&buffers->rb_lock);
1386 1387
	if (list_empty(&buffers->rb_send_bufs))
		goto out_reqbuf;
1388
	buffers->rb_send_count++;
1389
	req = rpcrdma_buffer_get_req_locked(buffers);
1390
	req->rl_reply = rpcrdma_buffer_get_rep(buffers);
1391
	spin_unlock(&buffers->rb_lock);
1392
	return req;
1393

1394
out_reqbuf:
1395
	spin_unlock(&buffers->rb_lock);
1396
	pr_warn("RPC:       %s: out of request buffers\n", __func__);
1397
	return NULL;
1398 1399 1400 1401 1402 1403 1404 1405 1406 1407
}

/*
 * Put request/reply buffers back into pool.
 * Pre-decrement counter/array index.
 */
void
rpcrdma_buffer_put(struct rpcrdma_req *req)
{
	struct rpcrdma_buffer *buffers = req->rl_buffer;
1408
	struct rpcrdma_rep *rep = req->rl_reply;
1409

1410 1411
	req->rl_reply = NULL;

1412
	spin_lock(&buffers->rb_lock);
1413
	buffers->rb_send_count--;
1414
	list_add_tail(&req->rl_list, &buffers->rb_send_bufs);
1415 1416
	if (rep) {
		buffers->rb_recv_count--;
1417
		list_add_tail(&rep->rr_list, &buffers->rb_recv_bufs);
1418
	}
1419
	spin_unlock(&buffers->rb_lock);
1420 1421 1422 1423
}

/*
 * Recover reply buffers from pool.
1424
 * This happens when recovering from disconnect.
1425 1426 1427 1428 1429 1430
 */
void
rpcrdma_recv_buffer_get(struct rpcrdma_req *req)
{
	struct rpcrdma_buffer *buffers = req->rl_buffer;

1431
	spin_lock(&buffers->rb_lock);
1432
	req->rl_reply = rpcrdma_buffer_get_rep(buffers);
1433
	spin_unlock(&buffers->rb_lock);
1434 1435 1436 1437
}

/*
 * Put reply buffers back into pool when not attached to
1438
 * request. This happens in error conditions.
1439 1440 1441 1442
 */
void
rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
{
1443
	struct rpcrdma_buffer *buffers = &rep->rr_rxprt->rx_buf;
1444

1445
	spin_lock(&buffers->rb_lock);
1446
	buffers->rb_recv_count--;
1447
	list_add_tail(&rep->rr_list, &buffers->rb_recv_bufs);
1448
	spin_unlock(&buffers->rb_lock);
1449 1450
}

1451
/**
1452
 * rpcrdma_alloc_regbuf - allocate and DMA-map memory for SEND/RECV buffers
1453
 * @size: size of buffer to be allocated, in bytes
1454
 * @direction: direction of data movement
1455 1456
 * @flags: GFP flags
 *
1457 1458
 * Returns an ERR_PTR, or a pointer to a regbuf, a buffer that
 * can be persistently DMA-mapped for I/O.
1459 1460
 *
 * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1461 1462
 * receiving the payload of RDMA RECV operations. During Long Calls
 * or Replies they may be registered externally via ro_map.
1463 1464
 */
struct rpcrdma_regbuf *
1465 1466
rpcrdma_alloc_regbuf(size_t size, enum dma_data_direction direction,
		     gfp_t flags)
1467 1468 1469 1470 1471
{
	struct rpcrdma_regbuf *rb;

	rb = kmalloc(sizeof(*rb) + size, flags);
	if (rb == NULL)
1472
		return ERR_PTR(-ENOMEM);
1473

1474
	rb->rg_device = NULL;
1475
	rb->rg_direction = direction;
1476
	rb->rg_iov.length = size;
1477 1478

	return rb;
1479
}
1480

1481 1482 1483 1484 1485 1486 1487 1488
/**
 * __rpcrdma_map_regbuf - DMA-map a regbuf
 * @ia: controlling rpcrdma_ia
 * @rb: regbuf to be mapped
 */
bool
__rpcrdma_dma_map_regbuf(struct rpcrdma_ia *ia, struct rpcrdma_regbuf *rb)
{
1489 1490
	struct ib_device *device = ia->ri_device;

1491 1492 1493
	if (rb->rg_direction == DMA_NONE)
		return false;

1494
	rb->rg_iov.addr = ib_dma_map_single(device,
1495 1496 1497
					    (void *)rb->rg_base,
					    rdmab_length(rb),
					    rb->rg_direction);
1498
	if (ib_dma_mapping_error(device, rdmab_addr(rb)))
1499 1500
		return false;

1501
	rb->rg_device = device;
1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514
	rb->rg_iov.lkey = ia->ri_pd->local_dma_lkey;
	return true;
}

static void
rpcrdma_dma_unmap_regbuf(struct rpcrdma_regbuf *rb)
{
	if (!rpcrdma_regbuf_is_mapped(rb))
		return;

	ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb),
			    rdmab_length(rb), rb->rg_direction);
	rb->rg_device = NULL;
1515 1516 1517 1518 1519 1520 1521
}

/**
 * rpcrdma_free_regbuf - deregister and free registered buffer
 * @rb: regbuf to be deregistered and freed
 */
void
1522
rpcrdma_free_regbuf(struct rpcrdma_regbuf *rb)
1523
{
1524 1525 1526
	if (!rb)
		return;

1527
	rpcrdma_dma_unmap_regbuf(rb);
1528
	kfree(rb);
1529 1530
}

1531 1532 1533 1534 1535 1536 1537 1538 1539 1540
/*
 * Prepost any receive buffer, then post send.
 *
 * Receive buffer is donated to hardware, reclaimed upon recv completion.
 */
int
rpcrdma_ep_post(struct rpcrdma_ia *ia,
		struct rpcrdma_ep *ep,
		struct rpcrdma_req *req)
{
1541
	struct ib_send_wr *send_wr = &req->rl_sendctx->sc_wr;
1542
	struct ib_send_wr *send_wr_fail;
1543
	int rc;
1544

1545 1546
	if (req->rl_reply) {
		rc = rpcrdma_ep_post_recv(ia, req->rl_reply);
1547
		if (rc)
1548
			return rc;
1549 1550 1551
		req->rl_reply = NULL;
	}

1552
	dprintk("RPC:       %s: posting %d s/g entries\n",
1553
		__func__, send_wr->num_sge);
1554

1555 1556
	if (!ep->rep_send_count ||
	    test_bit(RPCRDMA_REQ_F_TX_RESOURCES, &req->rl_flags)) {
1557 1558 1559 1560 1561 1562
		send_wr->send_flags |= IB_SEND_SIGNALED;
		ep->rep_send_count = ep->rep_send_batch;
	} else {
		send_wr->send_flags &= ~IB_SEND_SIGNALED;
		--ep->rep_send_count;
	}
1563
	rc = ib_post_send(ia->ri_id->qp, send_wr, &send_wr_fail);
1564
	if (rc)
1565 1566 1567 1568 1569 1570
		goto out_postsend_err;
	return 0;

out_postsend_err:
	pr_err("rpcrdma: RDMA Send ib_post_send returned %i\n", rc);
	return -ENOTCONN;
1571 1572 1573 1574 1575 1576
}

int
rpcrdma_ep_post_recv(struct rpcrdma_ia *ia,
		     struct rpcrdma_rep *rep)
{
1577
	struct ib_recv_wr *recv_wr_fail;
1578 1579
	int rc;

1580 1581
	if (!rpcrdma_dma_map_regbuf(ia, rep->rr_rdmabuf))
		goto out_map;
1582
	rc = ib_post_recv(ia->ri_id->qp, &rep->rr_recv_wr, &recv_wr_fail);
1583
	if (rc)
1584 1585 1586
		goto out_postrecv;
	return 0;

1587 1588 1589 1590
out_map:
	pr_err("rpcrdma: failed to DMA map the Receive buffer\n");
	return -EIO;

1591 1592 1593
out_postrecv:
	pr_err("rpcrdma: ib_post_recv returned %i\n", rc);
	return -ENOTCONN;
1594
}
1595

1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611
/**
 * rpcrdma_ep_post_extra_recv - Post buffers for incoming backchannel requests
 * @r_xprt: transport associated with these backchannel resources
 * @min_reqs: minimum number of incoming requests expected
 *
 * Returns zero if all requested buffers were posted, or a negative errno.
 */
int
rpcrdma_ep_post_extra_recv(struct rpcrdma_xprt *r_xprt, unsigned int count)
{
	struct rpcrdma_buffer *buffers = &r_xprt->rx_buf;
	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
	struct rpcrdma_rep *rep;
	int rc;

	while (count--) {
1612
		spin_lock(&buffers->rb_lock);
1613 1614 1615
		if (list_empty(&buffers->rb_recv_bufs))
			goto out_reqbuf;
		rep = rpcrdma_buffer_get_rep_locked(buffers);
1616
		spin_unlock(&buffers->rb_lock);
1617

1618
		rc = rpcrdma_ep_post_recv(ia, rep);
1619 1620 1621 1622 1623 1624 1625
		if (rc)
			goto out_rc;
	}

	return 0;

out_reqbuf:
1626
	spin_unlock(&buffers->rb_lock);
1627 1628 1629 1630 1631 1632 1633
	pr_warn("%s: no extra receive buffers\n", __func__);
	return -ENOMEM;

out_rc:
	rpcrdma_recv_buffer_put(rep);
	return rc;
}