summaryrefslogtreecommitdiffstats
path: root/net/sunrpc/xprtrdma/frwr_ops.c
blob: 6aea36a38bfdcbf95bd617089221060138e5ecaa (plain)
1
2
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
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
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
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
/*
 * Copyright (c) 2015 Oracle.  All rights reserved.
 * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
 */

/* Lightweight memory registration using Fast Registration Work
 * Requests (FRWR). Also referred to sometimes as FRMR mode.
 *
 * FRWR features ordered asynchronous registration and deregistration
 * of arbitrarily sized memory regions. This is the fastest and safest
 * but most complex memory registration mode.
 */

/* Normal operation
 *
 * A Memory Region is prepared for RDMA READ or WRITE using a FAST_REG
 * Work Request (frmr_op_map). When the RDMA operation is finished, this
 * Memory Region is invalidated using a LOCAL_INV Work Request
 * (frmr_op_unmap).
 *
 * Typically these Work Requests are not signaled, and neither are RDMA
 * SEND Work Requests (with the exception of signaling occasionally to
 * prevent provider work queue overflows). This greatly reduces HCA
 * interrupt workload.
 *
 * As an optimization, frwr_op_unmap marks MRs INVALID before the
 * LOCAL_INV WR is posted. If posting succeeds, the MR is placed on
 * rb_mws immediately so that no work (like managing a linked list
 * under a spinlock) is needed in the completion upcall.
 *
 * But this means that frwr_op_map() can occasionally encounter an MR
 * that is INVALID but the LOCAL_INV WR has not completed. Work Queue
 * ordering prevents a subsequent FAST_REG WR from executing against
 * that MR while it is still being invalidated.
 */

/* Transport recovery
 *
 * ->op_map and the transport connect worker cannot run at the same
 * time, but ->op_unmap can fire while the transport connect worker
 * is running. Thus MR recovery is handled in ->op_map, to guarantee
 * that recovered MRs are owned by a sending RPC, and not one where
 * ->op_unmap could fire at the same time transport reconnect is
 * being done.
 *
 * When the underlying transport disconnects, MRs are left in one of
 * four states:
 *
 * INVALID:	The MR was not in use before the QP entered ERROR state.
 *
 * VALID:	The MR was registered before the QP entered ERROR state.
 *
 * FLUSHED_FR:	The MR was being registered when the QP entered ERROR
 *		state, and the pending WR was flushed.
 *
 * FLUSHED_LI:	The MR was being invalidated when the QP entered ERROR
 *		state, and the pending WR was flushed.
 *
 * When frwr_op_map encounters FLUSHED and VALID MRs, they are recovered
 * with ib_dereg_mr and then are re-initialized. Because MR recovery
 * allocates fresh resources, it is deferred to a workqueue, and the
 * recovered MRs are placed back on the rb_mws list when recovery is
 * complete. frwr_op_map allocates another MR for the current RPC while
 * the broken MR is reset.
 *
 * To ensure that frwr_op_map doesn't encounter an MR that is marked
 * INVALID but that is about to be flushed due to a previous transport
 * disconnect, the transport connect worker attempts to drain all
 * pending send queue WRs before the transport is reconnected.
 */

#include <linux/sunrpc/rpc_rdma.h>

#include "xprt_rdma.h"

#if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
# define RPCDBG_FACILITY	RPCDBG_TRANS
#endif

bool
frwr_is_supported(struct rpcrdma_ia *ia)
{
	struct ib_device_attr *attrs = &ia->ri_device->attrs;

	if (!(attrs->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS))
		goto out_not_supported;
	if (attrs->max_fast_reg_page_list_len == 0)
		goto out_not_supported;
	return true;

out_not_supported:
	pr_info("rpcrdma: 'frwr' mode is not supported by device %s\n",
		ia->ri_device->name);
	return false;
}

static int
frwr_op_init_mr(struct rpcrdma_ia *ia, struct rpcrdma_mw *r)
{
	unsigned int depth = ia->ri_max_frmr_depth;
	struct rpcrdma_frmr *f = &r->frmr;
	int rc;

	f->fr_mr = ib_alloc_mr(ia->ri_pd, ia->ri_mrtype, depth);
	if (IS_ERR(f->fr_mr))
		goto out_mr_err;

	r->mw_sg = kcalloc(depth, sizeof(*r->mw_sg), GFP_KERNEL);
	if (!r->mw_sg)
		goto out_list_err;

	sg_init_table(r->mw_sg, depth);
	init_completion(&f->fr_linv_done);
	return 0;

out_mr_err:
	rc = PTR_ERR(f->fr_mr);
	dprintk("RPC:       %s: ib_alloc_mr status %i\n",
		__func__, rc);
	return rc;

out_list_err:
	rc = -ENOMEM;
	dprintk("RPC:       %s: sg allocation failure\n",
		__func__);
	ib_dereg_mr(f->fr_mr);
	return rc;
}

static void
frwr_op_release_mr(struct rpcrdma_mw *r)
{
	int rc;

	/* Ensure MW is not on any rl_registered list */
	if (!list_empty(&r->mw_list))
		list_del(&r->mw_list);

	rc = ib_dereg_mr(r->frmr.fr_mr);
	if (rc)
		pr_err("rpcrdma: final ib_dereg_mr for %p returned %i\n",
		       r, rc);
	kfree(r->mw_sg);
	kfree(r);
}

static int
__frwr_reset_mr(struct rpcrdma_ia *ia, struct rpcrdma_mw *r)
{
	struct rpcrdma_frmr *f = &r->frmr;
	int rc;

	rc = ib_dereg_mr(f->fr_mr);
	if (rc) {
		pr_warn("rpcrdma: ib_dereg_mr status %d, frwr %p orphaned\n",
			rc, r);
		return rc;
	}

	f->fr_mr = ib_alloc_mr(ia->ri_pd, ia->ri_mrtype,
			       ia->ri_max_frmr_depth);
	if (IS_ERR(f->fr_mr)) {
		pr_warn("rpcrdma: ib_alloc_mr status %ld, frwr %p orphaned\n",
			PTR_ERR(f->fr_mr), r);
		return PTR_ERR(f->fr_mr);
	}

	dprintk("RPC:       %s: recovered FRMR %p\n", __func__, f);
	f->fr_state = FRMR_IS_INVALID;
	return 0;
}

/* Reset of a single FRMR. Generate a fresh rkey by replacing the MR.
 */
static void
frwr_op_recover_mr(struct rpcrdma_mw *mw)
{
	enum rpcrdma_frmr_state state = mw->frmr.fr_state;
	struct rpcrdma_xprt *r_xprt = mw->mw_xprt;
	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
	int rc;

	rc = __frwr_reset_mr(ia, mw);
	if (state != FRMR_FLUSHED_LI)
		ib_dma_unmap_sg(ia->ri_device,
				mw->mw_sg, mw->mw_nents, mw->mw_dir);
	if (rc)
		goto out_release;

	rpcrdma_put_mw(r_xprt, mw);
	r_xprt->rx_stats.mrs_recovered++;
	return;

out_release:
	pr_err("rpcrdma: FRMR reset failed %d, %p release\n", rc, mw);
	r_xprt->rx_stats.mrs_orphaned++;

	spin_lock(&r_xprt->rx_buf.rb_mwlock);
	list_del(&mw->mw_all);
	spin_unlock(&r_xprt->rx_buf.rb_mwlock);

	frwr_op_release_mr(mw);
}

static int
frwr_op_open(struct rpcrdma_ia *ia, struct rpcrdma_ep *ep,
	     struct rpcrdma_create_data_internal *cdata)
{
	struct ib_device_attr *attrs = &ia->ri_device->attrs;
	int depth, delta;

	ia->ri_mrtype = IB_MR_TYPE_MEM_REG;
	if (attrs->device_cap_flags & IB_DEVICE_SG_GAPS_REG)
		ia->ri_mrtype = IB_MR_TYPE_SG_GAPS;

	ia->ri_max_frmr_depth =
			min_t(unsigned int, RPCRDMA_MAX_DATA_SEGS,
			      attrs->max_fast_reg_page_list_len);
	dprintk("RPC:       %s: device's max FR page list len = %u\n",
		__func__, ia->ri_max_frmr_depth);

	/* Add room for frmr register and invalidate WRs.
	 * 1. FRMR reg WR for head
	 * 2. FRMR invalidate WR for head
	 * 3. N FRMR reg WRs for pagelist
	 * 4. N FRMR invalidate WRs for pagelist
	 * 5. FRMR reg WR for tail
	 * 6. FRMR invalidate WR for tail
	 * 7. The RDMA_SEND WR
	 */
	depth = 7;

	/* Calculate N if the device max FRMR depth is smaller than
	 * RPCRDMA_MAX_DATA_SEGS.
	 */
	if (ia->ri_max_frmr_depth < RPCRDMA_MAX_DATA_SEGS) {
		delta = RPCRDMA_MAX_DATA_SEGS - ia->ri_max_frmr_depth;
		do {
			depth += 2; /* FRMR reg + invalidate */
			delta -= ia->ri_max_frmr_depth;
		} while (delta > 0);
	}

	ep->rep_attr.cap.max_send_wr *= depth;
	if (ep->rep_attr.cap.max_send_wr > attrs->max_qp_wr) {
		cdata->max_requests = attrs->max_qp_wr / depth;
		if (!cdata->max_requests)
			return -EINVAL;
		ep->rep_attr.cap.max_send_wr = cdata->max_requests *
					       depth;
	}

	ia->ri_max_segs = max_t(unsigned int, 1, RPCRDMA_MAX_DATA_SEGS /
				ia->ri_max_frmr_depth);
	return 0;
}

/* FRWR mode conveys a list of pages per chunk segment. The
 * maximum length of that list is the FRWR page list depth.
 */
static size_t
frwr_op_maxpages(struct rpcrdma_xprt *r_xprt)
{
	struct rpcrdma_ia *ia = &r_xprt->rx_ia;

	return min_t(unsigned int, RPCRDMA_MAX_DATA_SEGS,
		     RPCRDMA_MAX_HDR_SEGS * ia->ri_max_frmr_depth);
}

static void
__frwr_sendcompletion_flush(struct ib_wc *wc, const char *wr)
{
	if (wc->status != IB_WC_WR_FLUSH_ERR)
		pr_err("rpcrdma: %s: %s (%u/0x%x)\n",
		       wr, ib_wc_status_msg(wc->status),
		       wc->status, wc->vendor_err);
}

/**
 * frwr_wc_fastreg - Invoked by RDMA provider for a flushed FastReg WC
 * @cq:	completion queue (ignored)
 * @wc:	completed WR
 *
 */
static void
frwr_wc_fastreg(struct ib_cq *cq, struct ib_wc *wc)
{
	struct rpcrdma_frmr *frmr;
	struct ib_cqe *cqe;

	/* WARNING: Only wr_cqe and status are reliable at this point */
	if (wc->status != IB_WC_SUCCESS) {
		cqe = wc->wr_cqe;
		frmr = container_of(cqe, struct rpcrdma_frmr, fr_cqe);
		frmr->fr_state = FRMR_FLUSHED_FR;
		__frwr_sendcompletion_flush(wc, "fastreg");
	}
}

/**
 * frwr_wc_localinv - Invoked by RDMA provider for a flushed LocalInv WC
 * @cq:	completion queue (ignored)
 * @wc:	completed WR
 *
 */
static void
frwr_wc_localinv(struct ib_cq *cq, struct ib_wc *wc)
{
	struct rpcrdma_frmr *frmr;
	struct ib_cqe *cqe;

	/* WARNING: Only wr_cqe and status are reliable at this point */
	if (wc->status != IB_WC_SUCCESS) {
		cqe = wc->wr_cqe;
		frmr = container_of(cqe, struct rpcrdma_frmr, fr_cqe);
		frmr->fr_state = FRMR_FLUSHED_LI;
		__frwr_sendcompletion_flush(wc, "localinv");
	}
}

/**
 * frwr_wc_localinv_wake - Invoked by RDMA provider for a signaled LocalInv WC
 * @cq:	completion queue (ignored)
 * @wc:	completed WR
 *
 * Awaken anyone waiting for an MR to finish being fenced.
 */
static void
frwr_wc_localinv_wake(struct ib_cq *cq, struct ib_wc *wc)
{
	struct rpcrdma_frmr *frmr;
	struct ib_cqe *cqe;

	/* WARNING: Only wr_cqe and status are reliable at this point */
	cqe = wc->wr_cqe;
	frmr = container_of(cqe, struct rpcrdma_frmr, fr_cqe);
	if (wc->status != IB_WC_SUCCESS) {
		frmr->fr_state = FRMR_FLUSHED_LI;
		__frwr_sendcompletion_flush(wc, "localinv");
	}
	complete(&frmr->fr_linv_done);
}

/* Post a REG_MR Work Request to register a memory region
 * for remote access via RDMA READ or RDMA WRITE.
 */
static int
frwr_op_map(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mr_seg *seg,
	    int nsegs, bool writing, struct rpcrdma_mw **out)
{
	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
	bool holes_ok = ia->ri_mrtype == IB_MR_TYPE_SG_GAPS;
	struct rpcrdma_mw *mw;
	struct rpcrdma_frmr *frmr;
	struct ib_mr *mr;
	struct ib_reg_wr *reg_wr;
	struct ib_send_wr *bad_wr;
	int rc, i, n;
	u8 key;

	mw = NULL;
	do {
		if (mw)
			rpcrdma_defer_mr_recovery(mw);
		mw = rpcrdma_get_mw(r_xprt);
		if (!mw)
			return -ENOBUFS;
	} while (mw->frmr.fr_state != FRMR_IS_INVALID);
	frmr = &mw->frmr;
	frmr->fr_state = FRMR_IS_VALID;
	mr = frmr->fr_mr;
	reg_wr = &frmr->fr_regwr;

	if (nsegs > ia->ri_max_frmr_depth)
		nsegs = ia->ri_max_frmr_depth;
	for (i = 0; i < nsegs;) {
		if (seg->mr_page)
			sg_set_page(&mw->mw_sg[i],
				    seg->mr_page,
				    seg->mr_len,
				    offset_in_page(seg->mr_offset));
		else
			sg_set_buf(&mw->mw_sg[i], seg->mr_offset,
				   seg->mr_len);

		++seg;
		++i;
		if (holes_ok)
			continue;
		if ((i < nsegs && offset_in_page(seg->mr_offset)) ||
		    offset_in_page((seg-1)->mr_offset + (seg-1)->mr_len))
			break;
	}
	mw->mw_dir = rpcrdma_data_dir(writing);

	mw->mw_nents = ib_dma_map_sg(ia->ri_device, mw->mw_sg, i, mw->mw_dir);
	if (!mw->mw_nents)
		goto out_dmamap_err;

	n = ib_map_mr_sg(mr, mw->mw_sg, mw->mw_nents, NULL, PAGE_SIZE);
	if (unlikely(n != mw->mw_nents))
		goto out_mapmr_err;

	dprintk("RPC:       %s: Using frmr %p to map %u segments (%u bytes)\n",
		__func__, frmr, mw->mw_nents, mr->length);

	key = (u8)(mr->rkey & 0x000000FF);
	ib_update_fast_reg_key(mr, ++key);

	reg_wr->wr.next = NULL;
	reg_wr->wr.opcode = IB_WR_REG_MR;
	frmr->fr_cqe.done = frwr_wc_fastreg;
	reg_wr->wr.wr_cqe = &frmr->fr_cqe;
	reg_wr->wr.num_sge = 0;
	reg_wr->wr.send_flags = 0;
	reg_wr->mr = mr;
	reg_wr->key = mr->rkey;
	reg_wr->access = writing ?
			 IB_ACCESS_REMOTE_WRITE | IB_ACCESS_LOCAL_WRITE :
			 IB_ACCESS_REMOTE_READ;

	rpcrdma_set_signaled(&r_xprt->rx_ep, &reg_wr->wr);
	rc = ib_post_send(ia->ri_id->qp, &reg_wr->wr, &bad_wr);
	if (rc)
		goto out_senderr;

	mw->mw_handle = mr->rkey;
	mw->mw_length = mr->length;
	mw->mw_offset = mr->iova;

	*out = mw;
	return mw->mw_nents;

out_dmamap_err:
	pr_err("rpcrdma: failed to DMA map sg %p sg_nents %d\n",
	       mw->mw_sg, i);
	frmr->fr_state = FRMR_IS_INVALID;
	rpcrdma_put_mw(r_xprt, mw);
	return -EIO;

out_mapmr_err:
	pr_err("rpcrdma: failed to map mr %p (%d/%d)\n",
	       frmr->fr_mr, n, mw->mw_nents);
	rpcrdma_defer_mr_recovery(mw);
	return -EIO;

out_senderr:
	pr_err("rpcrdma: FRMR registration ib_post_send returned %i\n", rc);
	rpcrdma_defer_mr_recovery(mw);
	return -ENOTCONN;
}

/* Invalidate all memory regions that were registered for "req".
 *
 * Sleeps until it is safe for the host CPU to access the
 * previously mapped memory regions.
 *
 * Caller ensures that @mws is not empty before the call. This
 * function empties the list.
 */
static void
frwr_op_unmap_sync(struct rpcrdma_xprt *r_xprt, struct list_head *mws)
{
	struct ib_send_wr *first, **prev, *last, *bad_wr;
	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
	struct rpcrdma_frmr *f;
	struct rpcrdma_mw *mw;
	int count, rc;

	/* ORDER: Invalidate all of the MRs first
	 *
	 * Chain the LOCAL_INV Work Requests and post them with
	 * a single ib_post_send() call.
	 */
	f = NULL;
	count = 0;
	prev = &first;
	list_for_each_entry(mw, mws, mw_list) {
		mw->frmr.fr_state = FRMR_IS_INVALID;

		if (mw->mw_flags & RPCRDMA_MW_F_RI)
			continue;

		f = &mw->frmr;
		dprintk("RPC:       %s: invalidating frmr %p\n",
			__func__, f);

		f->fr_cqe.done = frwr_wc_localinv;
		last = &f->fr_invwr;
		memset(last, 0, sizeof(*last));
		last->wr_cqe = &f->fr_cqe;
		last->opcode = IB_WR_LOCAL_INV;
		last->ex.invalidate_rkey = mw->mw_handle;
		count++;

		*prev = last;
		prev = &last->next;
	}
	if (!f)
		goto unmap;

	/* Strong send queue ordering guarantees that when the
	 * last WR in the chain completes, all WRs in the chain
	 * are complete.
	 */
	last->send_flags = IB_SEND_SIGNALED;
	f->fr_cqe.done = frwr_wc_localinv_wake;
	reinit_completion(&f->fr_linv_done);

	/* Initialize CQ count, since there is always a signaled
	 * WR being posted here.  The new cqcount depends on how
	 * many SQEs are about to be consumed.
	 */
	rpcrdma_init_cqcount(&r_xprt->rx_ep, count);

	/* Transport disconnect drains the receive CQ before it
	 * replaces the QP. The RPC reply handler won't call us
	 * unless ri_id->qp is a valid pointer.
	 */
	r_xprt->rx_stats.local_inv_needed++;
	bad_wr = NULL;
	rc = ib_post_send(ia->ri_id->qp, first, &bad_wr);
	if (bad_wr != first)
		wait_for_completion(&f->fr_linv_done);
	if (rc)
		goto reset_mrs;

	/* ORDER: Now DMA unmap all of the MRs, and return
	 * them to the free MW list.
	 */
unmap:
	while (!list_empty(mws)) {
		mw = rpcrdma_pop_mw(mws);
		dprintk("RPC:       %s: DMA unmapping frmr %p\n",
			__func__, &mw->frmr);
		ib_dma_unmap_sg(ia->ri_device,
				mw->mw_sg, mw->mw_nents, mw->mw_dir);
		rpcrdma_put_mw(r_xprt, mw);
	}
	return;

reset_mrs:
	pr_err("rpcrdma: FRMR invalidate ib_post_send returned %i\n", rc);

	/* Find and reset the MRs in the LOCAL_INV WRs that did not
	 * get posted.
	 */
	rpcrdma_init_cqcount(&r_xprt->rx_ep, -count);
	while (bad_wr) {
		f = container_of(bad_wr, struct rpcrdma_frmr,
				 fr_invwr);
		mw = container_of(f, struct rpcrdma_mw, frmr);

		__frwr_reset_mr(ia, mw);

		bad_wr = bad_wr->next;
	}
	goto unmap;
}

/* Use a slow, safe mechanism to invalidate all memory regions
 * that were registered for "req".
 */
static void
frwr_op_unmap_safe(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
		   bool sync)
{
	struct rpcrdma_mw *mw;

	while (!list_empty(&req->rl_registered)) {
		mw = rpcrdma_pop_mw(&req->rl_registered);
		if (sync)
			frwr_op_recover_mr(mw);
		else
			rpcrdma_defer_mr_recovery(mw);
	}
}

const struct rpcrdma_memreg_ops rpcrdma_frwr_memreg_ops = {
	.ro_map				= frwr_op_map,
	.ro_unmap_sync			= frwr_op_unmap_sync,
	.ro_unmap_safe			= frwr_op_unmap_safe,
	.ro_recover_mr			= frwr_op_recover_mr,
	.ro_open			= frwr_op_open,
	.ro_maxpages			= frwr_op_maxpages,
	.ro_init_mr			= frwr_op_init_mr,
	.ro_release_mr			= frwr_op_release_mr,
	.ro_displayname			= "frwr",
	.ro_send_w_inv_ok		= RPCRDMA_CMP_F_SND_W_INV_OK,
};