xref: /linux/net/sunrpc/xprtrdma/frwr_ops.c (revision 8795a739e5c72abeec51caf36b6df2b37e5720c5)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (c) 2015, 2017 Oracle.  All rights reserved.
4  * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
5  */
6 
7 /* Lightweight memory registration using Fast Registration Work
8  * Requests (FRWR).
9  *
10  * FRWR features ordered asynchronous registration and invalidation
11  * of arbitrarily-sized memory regions. This is the fastest and safest
12  * but most complex memory registration mode.
13  */
14 
15 /* Normal operation
16  *
17  * A Memory Region is prepared for RDMA Read or Write using a FAST_REG
18  * Work Request (frwr_map). When the RDMA operation is finished, this
19  * Memory Region is invalidated using a LOCAL_INV Work Request
20  * (frwr_unmap_async and frwr_unmap_sync).
21  *
22  * Typically FAST_REG Work Requests are not signaled, and neither are
23  * RDMA Send Work Requests (with the exception of signaling occasionally
24  * to prevent provider work queue overflows). This greatly reduces HCA
25  * interrupt workload.
26  */
27 
28 /* Transport recovery
29  *
30  * frwr_map and frwr_unmap_* cannot run at the same time the transport
31  * connect worker is running. The connect worker holds the transport
32  * send lock, just as ->send_request does. This prevents frwr_map and
33  * the connect worker from running concurrently. When a connection is
34  * closed, the Receive completion queue is drained before the allowing
35  * the connect worker to get control. This prevents frwr_unmap and the
36  * connect worker from running concurrently.
37  *
38  * When the underlying transport disconnects, MRs that are in flight
39  * are flushed and are likely unusable. Thus all flushed MRs are
40  * destroyed. New MRs are created on demand.
41  */
42 
43 #include <linux/sunrpc/rpc_rdma.h>
44 #include <linux/sunrpc/svc_rdma.h>
45 
46 #include "xprt_rdma.h"
47 #include <trace/events/rpcrdma.h>
48 
49 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
50 # define RPCDBG_FACILITY	RPCDBG_TRANS
51 #endif
52 
53 /**
54  * frwr_is_supported - Check if device supports FRWR
55  * @device: interface adapter to check
56  *
57  * Returns true if device supports FRWR, otherwise false
58  */
59 bool frwr_is_supported(struct ib_device *device)
60 {
61 	struct ib_device_attr *attrs = &device->attrs;
62 
63 	if (!(attrs->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS))
64 		goto out_not_supported;
65 	if (attrs->max_fast_reg_page_list_len == 0)
66 		goto out_not_supported;
67 	return true;
68 
69 out_not_supported:
70 	pr_info("rpcrdma: 'frwr' mode is not supported by device %s\n",
71 		device->name);
72 	return false;
73 }
74 
75 /**
76  * frwr_release_mr - Destroy one MR
77  * @mr: MR allocated by frwr_init_mr
78  *
79  */
80 void frwr_release_mr(struct rpcrdma_mr *mr)
81 {
82 	int rc;
83 
84 	rc = ib_dereg_mr(mr->frwr.fr_mr);
85 	if (rc)
86 		trace_xprtrdma_frwr_dereg(mr, rc);
87 	kfree(mr->mr_sg);
88 	kfree(mr);
89 }
90 
91 static void frwr_mr_recycle(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mr *mr)
92 {
93 	trace_xprtrdma_mr_recycle(mr);
94 
95 	if (mr->mr_dir != DMA_NONE) {
96 		trace_xprtrdma_mr_unmap(mr);
97 		ib_dma_unmap_sg(r_xprt->rx_ia.ri_id->device,
98 				mr->mr_sg, mr->mr_nents, mr->mr_dir);
99 		mr->mr_dir = DMA_NONE;
100 	}
101 
102 	spin_lock(&r_xprt->rx_buf.rb_lock);
103 	list_del(&mr->mr_all);
104 	r_xprt->rx_stats.mrs_recycled++;
105 	spin_unlock(&r_xprt->rx_buf.rb_lock);
106 
107 	frwr_release_mr(mr);
108 }
109 
110 /* MRs are dynamically allocated, so simply clean up and release the MR.
111  * A replacement MR will subsequently be allocated on demand.
112  */
113 static void
114 frwr_mr_recycle_worker(struct work_struct *work)
115 {
116 	struct rpcrdma_mr *mr = container_of(work, struct rpcrdma_mr,
117 					     mr_recycle);
118 
119 	frwr_mr_recycle(mr->mr_xprt, mr);
120 }
121 
122 /* frwr_recycle - Discard MRs
123  * @req: request to reset
124  *
125  * Used after a reconnect. These MRs could be in flight, we can't
126  * tell. Safe thing to do is release them.
127  */
128 void frwr_recycle(struct rpcrdma_req *req)
129 {
130 	struct rpcrdma_mr *mr;
131 
132 	while ((mr = rpcrdma_mr_pop(&req->rl_registered)))
133 		frwr_mr_recycle(mr->mr_xprt, mr);
134 }
135 
136 /* frwr_reset - Place MRs back on the free list
137  * @req: request to reset
138  *
139  * Used after a failed marshal. For FRWR, this means the MRs
140  * don't have to be fully released and recreated.
141  *
142  * NB: This is safe only as long as none of @req's MRs are
143  * involved with an ongoing asynchronous FAST_REG or LOCAL_INV
144  * Work Request.
145  */
146 void frwr_reset(struct rpcrdma_req *req)
147 {
148 	struct rpcrdma_mr *mr;
149 
150 	while ((mr = rpcrdma_mr_pop(&req->rl_registered)))
151 		rpcrdma_mr_put(mr);
152 }
153 
154 /**
155  * frwr_init_mr - Initialize one MR
156  * @ia: interface adapter
157  * @mr: generic MR to prepare for FRWR
158  *
159  * Returns zero if successful. Otherwise a negative errno
160  * is returned.
161  */
162 int frwr_init_mr(struct rpcrdma_ia *ia, struct rpcrdma_mr *mr)
163 {
164 	unsigned int depth = ia->ri_max_frwr_depth;
165 	struct scatterlist *sg;
166 	struct ib_mr *frmr;
167 	int rc;
168 
169 	/* NB: ib_alloc_mr and device drivers typically allocate
170 	 *     memory with GFP_KERNEL.
171 	 */
172 	frmr = ib_alloc_mr(ia->ri_pd, ia->ri_mrtype, depth);
173 	if (IS_ERR(frmr))
174 		goto out_mr_err;
175 
176 	sg = kcalloc(depth, sizeof(*sg), GFP_NOFS);
177 	if (!sg)
178 		goto out_list_err;
179 
180 	mr->frwr.fr_mr = frmr;
181 	mr->mr_dir = DMA_NONE;
182 	INIT_LIST_HEAD(&mr->mr_list);
183 	INIT_WORK(&mr->mr_recycle, frwr_mr_recycle_worker);
184 	init_completion(&mr->frwr.fr_linv_done);
185 
186 	sg_init_table(sg, depth);
187 	mr->mr_sg = sg;
188 	return 0;
189 
190 out_mr_err:
191 	rc = PTR_ERR(frmr);
192 	trace_xprtrdma_frwr_alloc(mr, rc);
193 	return rc;
194 
195 out_list_err:
196 	ib_dereg_mr(frmr);
197 	return -ENOMEM;
198 }
199 
200 /**
201  * frwr_open - Prepare an endpoint for use with FRWR
202  * @ia: interface adapter this endpoint will use
203  * @ep: endpoint to prepare
204  *
205  * On success, sets:
206  *	ep->rep_attr.cap.max_send_wr
207  *	ep->rep_attr.cap.max_recv_wr
208  *	ep->rep_max_requests
209  *	ia->ri_max_segs
210  *
211  * And these FRWR-related fields:
212  *	ia->ri_max_frwr_depth
213  *	ia->ri_mrtype
214  *
215  * On failure, a negative errno is returned.
216  */
217 int frwr_open(struct rpcrdma_ia *ia, struct rpcrdma_ep *ep)
218 {
219 	struct ib_device_attr *attrs = &ia->ri_id->device->attrs;
220 	int max_qp_wr, depth, delta;
221 
222 	ia->ri_mrtype = IB_MR_TYPE_MEM_REG;
223 	if (attrs->device_cap_flags & IB_DEVICE_SG_GAPS_REG)
224 		ia->ri_mrtype = IB_MR_TYPE_SG_GAPS;
225 
226 	/* Quirk: Some devices advertise a large max_fast_reg_page_list_len
227 	 * capability, but perform optimally when the MRs are not larger
228 	 * than a page.
229 	 */
230 	if (attrs->max_sge_rd > 1)
231 		ia->ri_max_frwr_depth = attrs->max_sge_rd;
232 	else
233 		ia->ri_max_frwr_depth = attrs->max_fast_reg_page_list_len;
234 	if (ia->ri_max_frwr_depth > RPCRDMA_MAX_DATA_SEGS)
235 		ia->ri_max_frwr_depth = RPCRDMA_MAX_DATA_SEGS;
236 	dprintk("RPC:       %s: max FR page list depth = %u\n",
237 		__func__, ia->ri_max_frwr_depth);
238 
239 	/* Add room for frwr register and invalidate WRs.
240 	 * 1. FRWR reg WR for head
241 	 * 2. FRWR invalidate WR for head
242 	 * 3. N FRWR reg WRs for pagelist
243 	 * 4. N FRWR invalidate WRs for pagelist
244 	 * 5. FRWR reg WR for tail
245 	 * 6. FRWR invalidate WR for tail
246 	 * 7. The RDMA_SEND WR
247 	 */
248 	depth = 7;
249 
250 	/* Calculate N if the device max FRWR depth is smaller than
251 	 * RPCRDMA_MAX_DATA_SEGS.
252 	 */
253 	if (ia->ri_max_frwr_depth < RPCRDMA_MAX_DATA_SEGS) {
254 		delta = RPCRDMA_MAX_DATA_SEGS - ia->ri_max_frwr_depth;
255 		do {
256 			depth += 2; /* FRWR reg + invalidate */
257 			delta -= ia->ri_max_frwr_depth;
258 		} while (delta > 0);
259 	}
260 
261 	max_qp_wr = ia->ri_id->device->attrs.max_qp_wr;
262 	max_qp_wr -= RPCRDMA_BACKWARD_WRS;
263 	max_qp_wr -= 1;
264 	if (max_qp_wr < RPCRDMA_MIN_SLOT_TABLE)
265 		return -ENOMEM;
266 	if (ep->rep_max_requests > max_qp_wr)
267 		ep->rep_max_requests = max_qp_wr;
268 	ep->rep_attr.cap.max_send_wr = ep->rep_max_requests * depth;
269 	if (ep->rep_attr.cap.max_send_wr > max_qp_wr) {
270 		ep->rep_max_requests = max_qp_wr / depth;
271 		if (!ep->rep_max_requests)
272 			return -EINVAL;
273 		ep->rep_attr.cap.max_send_wr = ep->rep_max_requests * depth;
274 	}
275 	ep->rep_attr.cap.max_send_wr += RPCRDMA_BACKWARD_WRS;
276 	ep->rep_attr.cap.max_send_wr += 1; /* for ib_drain_sq */
277 	ep->rep_attr.cap.max_recv_wr = ep->rep_max_requests;
278 	ep->rep_attr.cap.max_recv_wr += RPCRDMA_BACKWARD_WRS;
279 	ep->rep_attr.cap.max_recv_wr += 1; /* for ib_drain_rq */
280 
281 	ia->ri_max_segs =
282 		DIV_ROUND_UP(RPCRDMA_MAX_DATA_SEGS, ia->ri_max_frwr_depth);
283 	/* Reply chunks require segments for head and tail buffers */
284 	ia->ri_max_segs += 2;
285 	if (ia->ri_max_segs > RPCRDMA_MAX_HDR_SEGS)
286 		ia->ri_max_segs = RPCRDMA_MAX_HDR_SEGS;
287 	return 0;
288 }
289 
290 /**
291  * frwr_maxpages - Compute size of largest payload
292  * @r_xprt: transport
293  *
294  * Returns maximum size of an RPC message, in pages.
295  *
296  * FRWR mode conveys a list of pages per chunk segment. The
297  * maximum length of that list is the FRWR page list depth.
298  */
299 size_t frwr_maxpages(struct rpcrdma_xprt *r_xprt)
300 {
301 	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
302 
303 	return min_t(unsigned int, RPCRDMA_MAX_DATA_SEGS,
304 		     (ia->ri_max_segs - 2) * ia->ri_max_frwr_depth);
305 }
306 
307 /**
308  * frwr_map - Register a memory region
309  * @r_xprt: controlling transport
310  * @seg: memory region co-ordinates
311  * @nsegs: number of segments remaining
312  * @writing: true when RDMA Write will be used
313  * @xid: XID of RPC using the registered memory
314  * @mr: MR to fill in
315  *
316  * Prepare a REG_MR Work Request to register a memory region
317  * for remote access via RDMA READ or RDMA WRITE.
318  *
319  * Returns the next segment or a negative errno pointer.
320  * On success, @mr is filled in.
321  */
322 struct rpcrdma_mr_seg *frwr_map(struct rpcrdma_xprt *r_xprt,
323 				struct rpcrdma_mr_seg *seg,
324 				int nsegs, bool writing, __be32 xid,
325 				struct rpcrdma_mr *mr)
326 {
327 	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
328 	struct ib_reg_wr *reg_wr;
329 	struct ib_mr *ibmr;
330 	int i, n;
331 	u8 key;
332 
333 	if (nsegs > ia->ri_max_frwr_depth)
334 		nsegs = ia->ri_max_frwr_depth;
335 	for (i = 0; i < nsegs;) {
336 		if (seg->mr_page)
337 			sg_set_page(&mr->mr_sg[i],
338 				    seg->mr_page,
339 				    seg->mr_len,
340 				    offset_in_page(seg->mr_offset));
341 		else
342 			sg_set_buf(&mr->mr_sg[i], seg->mr_offset,
343 				   seg->mr_len);
344 
345 		++seg;
346 		++i;
347 		if (ia->ri_mrtype == IB_MR_TYPE_SG_GAPS)
348 			continue;
349 		if ((i < nsegs && offset_in_page(seg->mr_offset)) ||
350 		    offset_in_page((seg-1)->mr_offset + (seg-1)->mr_len))
351 			break;
352 	}
353 	mr->mr_dir = rpcrdma_data_dir(writing);
354 
355 	mr->mr_nents =
356 		ib_dma_map_sg(ia->ri_id->device, mr->mr_sg, i, mr->mr_dir);
357 	if (!mr->mr_nents)
358 		goto out_dmamap_err;
359 
360 	ibmr = mr->frwr.fr_mr;
361 	n = ib_map_mr_sg(ibmr, mr->mr_sg, mr->mr_nents, NULL, PAGE_SIZE);
362 	if (unlikely(n != mr->mr_nents))
363 		goto out_mapmr_err;
364 
365 	ibmr->iova &= 0x00000000ffffffff;
366 	ibmr->iova |= ((u64)be32_to_cpu(xid)) << 32;
367 	key = (u8)(ibmr->rkey & 0x000000FF);
368 	ib_update_fast_reg_key(ibmr, ++key);
369 
370 	reg_wr = &mr->frwr.fr_regwr;
371 	reg_wr->mr = ibmr;
372 	reg_wr->key = ibmr->rkey;
373 	reg_wr->access = writing ?
374 			 IB_ACCESS_REMOTE_WRITE | IB_ACCESS_LOCAL_WRITE :
375 			 IB_ACCESS_REMOTE_READ;
376 
377 	mr->mr_handle = ibmr->rkey;
378 	mr->mr_length = ibmr->length;
379 	mr->mr_offset = ibmr->iova;
380 	trace_xprtrdma_mr_map(mr);
381 
382 	return seg;
383 
384 out_dmamap_err:
385 	mr->mr_dir = DMA_NONE;
386 	trace_xprtrdma_frwr_sgerr(mr, i);
387 	return ERR_PTR(-EIO);
388 
389 out_mapmr_err:
390 	trace_xprtrdma_frwr_maperr(mr, n);
391 	return ERR_PTR(-EIO);
392 }
393 
394 /**
395  * frwr_wc_fastreg - Invoked by RDMA provider for a flushed FastReg WC
396  * @cq:	completion queue (ignored)
397  * @wc:	completed WR
398  *
399  */
400 static void frwr_wc_fastreg(struct ib_cq *cq, struct ib_wc *wc)
401 {
402 	struct ib_cqe *cqe = wc->wr_cqe;
403 	struct rpcrdma_frwr *frwr =
404 		container_of(cqe, struct rpcrdma_frwr, fr_cqe);
405 
406 	/* WARNING: Only wr_cqe and status are reliable at this point */
407 	trace_xprtrdma_wc_fastreg(wc, frwr);
408 	/* The MR will get recycled when the associated req is retransmitted */
409 }
410 
411 /**
412  * frwr_send - post Send WR containing the RPC Call message
413  * @ia: interface adapter
414  * @req: Prepared RPC Call
415  *
416  * For FRWR, chain any FastReg WRs to the Send WR. Only a
417  * single ib_post_send call is needed to register memory
418  * and then post the Send WR.
419  *
420  * Returns the result of ib_post_send.
421  */
422 int frwr_send(struct rpcrdma_ia *ia, struct rpcrdma_req *req)
423 {
424 	struct ib_send_wr *post_wr;
425 	struct rpcrdma_mr *mr;
426 
427 	post_wr = &req->rl_sendctx->sc_wr;
428 	list_for_each_entry(mr, &req->rl_registered, mr_list) {
429 		struct rpcrdma_frwr *frwr;
430 
431 		frwr = &mr->frwr;
432 
433 		frwr->fr_cqe.done = frwr_wc_fastreg;
434 		frwr->fr_regwr.wr.next = post_wr;
435 		frwr->fr_regwr.wr.wr_cqe = &frwr->fr_cqe;
436 		frwr->fr_regwr.wr.num_sge = 0;
437 		frwr->fr_regwr.wr.opcode = IB_WR_REG_MR;
438 		frwr->fr_regwr.wr.send_flags = 0;
439 
440 		post_wr = &frwr->fr_regwr.wr;
441 	}
442 
443 	/* If ib_post_send fails, the next ->send_request for
444 	 * @req will queue these MRs for recovery.
445 	 */
446 	return ib_post_send(ia->ri_id->qp, post_wr, NULL);
447 }
448 
449 /**
450  * frwr_reminv - handle a remotely invalidated mr on the @mrs list
451  * @rep: Received reply
452  * @mrs: list of MRs to check
453  *
454  */
455 void frwr_reminv(struct rpcrdma_rep *rep, struct list_head *mrs)
456 {
457 	struct rpcrdma_mr *mr;
458 
459 	list_for_each_entry(mr, mrs, mr_list)
460 		if (mr->mr_handle == rep->rr_inv_rkey) {
461 			list_del_init(&mr->mr_list);
462 			trace_xprtrdma_mr_remoteinv(mr);
463 			rpcrdma_mr_put(mr);
464 			break;	/* only one invalidated MR per RPC */
465 		}
466 }
467 
468 static void __frwr_release_mr(struct ib_wc *wc, struct rpcrdma_mr *mr)
469 {
470 	if (wc->status != IB_WC_SUCCESS)
471 		rpcrdma_mr_recycle(mr);
472 	else
473 		rpcrdma_mr_put(mr);
474 }
475 
476 /**
477  * frwr_wc_localinv - Invoked by RDMA provider for a LOCAL_INV WC
478  * @cq:	completion queue (ignored)
479  * @wc:	completed WR
480  *
481  */
482 static void frwr_wc_localinv(struct ib_cq *cq, struct ib_wc *wc)
483 {
484 	struct ib_cqe *cqe = wc->wr_cqe;
485 	struct rpcrdma_frwr *frwr =
486 		container_of(cqe, struct rpcrdma_frwr, fr_cqe);
487 	struct rpcrdma_mr *mr = container_of(frwr, struct rpcrdma_mr, frwr);
488 
489 	/* WARNING: Only wr_cqe and status are reliable at this point */
490 	trace_xprtrdma_wc_li(wc, frwr);
491 	__frwr_release_mr(wc, mr);
492 }
493 
494 /**
495  * frwr_wc_localinv_wake - Invoked by RDMA provider for a LOCAL_INV WC
496  * @cq:	completion queue (ignored)
497  * @wc:	completed WR
498  *
499  * Awaken anyone waiting for an MR to finish being fenced.
500  */
501 static void frwr_wc_localinv_wake(struct ib_cq *cq, struct ib_wc *wc)
502 {
503 	struct ib_cqe *cqe = wc->wr_cqe;
504 	struct rpcrdma_frwr *frwr =
505 		container_of(cqe, struct rpcrdma_frwr, fr_cqe);
506 	struct rpcrdma_mr *mr = container_of(frwr, struct rpcrdma_mr, frwr);
507 
508 	/* WARNING: Only wr_cqe and status are reliable at this point */
509 	trace_xprtrdma_wc_li_wake(wc, frwr);
510 	__frwr_release_mr(wc, mr);
511 	complete(&frwr->fr_linv_done);
512 }
513 
514 /**
515  * frwr_unmap_sync - invalidate memory regions that were registered for @req
516  * @r_xprt: controlling transport instance
517  * @req: rpcrdma_req with a non-empty list of MRs to process
518  *
519  * Sleeps until it is safe for the host CPU to access the previously mapped
520  * memory regions. This guarantees that registered MRs are properly fenced
521  * from the server before the RPC consumer accesses the data in them. It
522  * also ensures proper Send flow control: waking the next RPC waits until
523  * this RPC has relinquished all its Send Queue entries.
524  */
525 void frwr_unmap_sync(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
526 {
527 	struct ib_send_wr *first, **prev, *last;
528 	const struct ib_send_wr *bad_wr;
529 	struct rpcrdma_frwr *frwr;
530 	struct rpcrdma_mr *mr;
531 	int rc;
532 
533 	/* ORDER: Invalidate all of the MRs first
534 	 *
535 	 * Chain the LOCAL_INV Work Requests and post them with
536 	 * a single ib_post_send() call.
537 	 */
538 	frwr = NULL;
539 	prev = &first;
540 	while ((mr = rpcrdma_mr_pop(&req->rl_registered))) {
541 
542 		trace_xprtrdma_mr_localinv(mr);
543 		r_xprt->rx_stats.local_inv_needed++;
544 
545 		frwr = &mr->frwr;
546 		frwr->fr_cqe.done = frwr_wc_localinv;
547 		last = &frwr->fr_invwr;
548 		last->next = NULL;
549 		last->wr_cqe = &frwr->fr_cqe;
550 		last->sg_list = NULL;
551 		last->num_sge = 0;
552 		last->opcode = IB_WR_LOCAL_INV;
553 		last->send_flags = IB_SEND_SIGNALED;
554 		last->ex.invalidate_rkey = mr->mr_handle;
555 
556 		*prev = last;
557 		prev = &last->next;
558 	}
559 
560 	/* Strong send queue ordering guarantees that when the
561 	 * last WR in the chain completes, all WRs in the chain
562 	 * are complete.
563 	 */
564 	frwr->fr_cqe.done = frwr_wc_localinv_wake;
565 	reinit_completion(&frwr->fr_linv_done);
566 
567 	/* Transport disconnect drains the receive CQ before it
568 	 * replaces the QP. The RPC reply handler won't call us
569 	 * unless ri_id->qp is a valid pointer.
570 	 */
571 	bad_wr = NULL;
572 	rc = ib_post_send(r_xprt->rx_ia.ri_id->qp, first, &bad_wr);
573 	trace_xprtrdma_post_send(req, rc);
574 
575 	/* The final LOCAL_INV WR in the chain is supposed to
576 	 * do the wake. If it was never posted, the wake will
577 	 * not happen, so don't wait in that case.
578 	 */
579 	if (bad_wr != first)
580 		wait_for_completion(&frwr->fr_linv_done);
581 	if (!rc)
582 		return;
583 
584 	/* Recycle MRs in the LOCAL_INV chain that did not get posted.
585 	 */
586 	while (bad_wr) {
587 		frwr = container_of(bad_wr, struct rpcrdma_frwr,
588 				    fr_invwr);
589 		mr = container_of(frwr, struct rpcrdma_mr, frwr);
590 		bad_wr = bad_wr->next;
591 
592 		list_del_init(&mr->mr_list);
593 		rpcrdma_mr_recycle(mr);
594 	}
595 }
596 
597 /**
598  * frwr_wc_localinv_done - Invoked by RDMA provider for a signaled LOCAL_INV WC
599  * @cq:	completion queue (ignored)
600  * @wc:	completed WR
601  *
602  */
603 static void frwr_wc_localinv_done(struct ib_cq *cq, struct ib_wc *wc)
604 {
605 	struct ib_cqe *cqe = wc->wr_cqe;
606 	struct rpcrdma_frwr *frwr =
607 		container_of(cqe, struct rpcrdma_frwr, fr_cqe);
608 	struct rpcrdma_mr *mr = container_of(frwr, struct rpcrdma_mr, frwr);
609 	struct rpcrdma_rep *rep = mr->mr_req->rl_reply;
610 
611 	/* WARNING: Only wr_cqe and status are reliable at this point */
612 	trace_xprtrdma_wc_li_done(wc, frwr);
613 	__frwr_release_mr(wc, mr);
614 
615 	/* Ensure @rep is generated before __frwr_release_mr */
616 	smp_rmb();
617 	rpcrdma_complete_rqst(rep);
618 }
619 
620 /**
621  * frwr_unmap_async - invalidate memory regions that were registered for @req
622  * @r_xprt: controlling transport instance
623  * @req: rpcrdma_req with a non-empty list of MRs to process
624  *
625  * This guarantees that registered MRs are properly fenced from the
626  * server before the RPC consumer accesses the data in them. It also
627  * ensures proper Send flow control: waking the next RPC waits until
628  * this RPC has relinquished all its Send Queue entries.
629  */
630 void frwr_unmap_async(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
631 {
632 	struct ib_send_wr *first, *last, **prev;
633 	const struct ib_send_wr *bad_wr;
634 	struct rpcrdma_frwr *frwr;
635 	struct rpcrdma_mr *mr;
636 	int rc;
637 
638 	/* Chain the LOCAL_INV Work Requests and post them with
639 	 * a single ib_post_send() call.
640 	 */
641 	frwr = NULL;
642 	prev = &first;
643 	while ((mr = rpcrdma_mr_pop(&req->rl_registered))) {
644 
645 		trace_xprtrdma_mr_localinv(mr);
646 		r_xprt->rx_stats.local_inv_needed++;
647 
648 		frwr = &mr->frwr;
649 		frwr->fr_cqe.done = frwr_wc_localinv;
650 		last = &frwr->fr_invwr;
651 		last->next = NULL;
652 		last->wr_cqe = &frwr->fr_cqe;
653 		last->sg_list = NULL;
654 		last->num_sge = 0;
655 		last->opcode = IB_WR_LOCAL_INV;
656 		last->send_flags = IB_SEND_SIGNALED;
657 		last->ex.invalidate_rkey = mr->mr_handle;
658 
659 		*prev = last;
660 		prev = &last->next;
661 	}
662 
663 	/* Strong send queue ordering guarantees that when the
664 	 * last WR in the chain completes, all WRs in the chain
665 	 * are complete. The last completion will wake up the
666 	 * RPC waiter.
667 	 */
668 	frwr->fr_cqe.done = frwr_wc_localinv_done;
669 
670 	/* Transport disconnect drains the receive CQ before it
671 	 * replaces the QP. The RPC reply handler won't call us
672 	 * unless ri_id->qp is a valid pointer.
673 	 */
674 	bad_wr = NULL;
675 	rc = ib_post_send(r_xprt->rx_ia.ri_id->qp, first, &bad_wr);
676 	trace_xprtrdma_post_send(req, rc);
677 	if (!rc)
678 		return;
679 
680 	/* Recycle MRs in the LOCAL_INV chain that did not get posted.
681 	 */
682 	while (bad_wr) {
683 		frwr = container_of(bad_wr, struct rpcrdma_frwr, fr_invwr);
684 		mr = container_of(frwr, struct rpcrdma_mr, frwr);
685 		bad_wr = bad_wr->next;
686 
687 		rpcrdma_mr_recycle(mr);
688 	}
689 
690 	/* The final LOCAL_INV WR in the chain is supposed to
691 	 * do the wake. If it was never posted, the wake will
692 	 * not happen, so wake here in that case.
693 	 */
694 	rpcrdma_complete_rqst(req->rl_reply);
695 }
696