summaryrefslogtreecommitdiff
path: root/net/rds/rdma.c
blob: 916f5ec373d8eac066b871f56901d041d0cb189b (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
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
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
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
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
/*
 * Copyright (c) 2007, 2017 Oracle and/or its affiliates. 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
 * OpenIB.org BSD 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.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 *
 */
#include <linux/pagemap.h>
#include <linux/slab.h>
#include <linux/rbtree.h>
#include <linux/dma-mapping.h> /* for DMA_*_DEVICE */

#include "rds.h"

/*
 * XXX
 *  - build with sparse
 *  - should we detect duplicate keys on a socket?  hmm.
 *  - an rdma is an mlock, apply rlimit?
 */

/*
 * get the number of pages by looking at the page indices that the start and
 * end addresses fall in.
 *
 * Returns 0 if the vec is invalid.  It is invalid if the number of bytes
 * causes the address to wrap or overflows an unsigned int.  This comes
 * from being stored in the 'length' member of 'struct scatterlist'.
 */
static unsigned int rds_pages_in_vec(struct rds_iovec *vec)
{
	if ((vec->addr + vec->bytes <= vec->addr) ||
	    (vec->bytes > (u64)UINT_MAX))
		return 0;

	return ((vec->addr + vec->bytes + PAGE_SIZE - 1) >> PAGE_SHIFT) -
		(vec->addr >> PAGE_SHIFT);
}

static struct rds_mr *rds_mr_tree_walk(struct rb_root *root, u64 key,
				       struct rds_mr *insert)
{
	struct rb_node **p = &root->rb_node;
	struct rb_node *parent = NULL;
	struct rds_mr *mr;

	while (*p) {
		parent = *p;
		mr = rb_entry(parent, struct rds_mr, r_rb_node);

		if (key < mr->r_key)
			p = &(*p)->rb_left;
		else if (key > mr->r_key)
			p = &(*p)->rb_right;
		else
			return mr;
	}

	if (insert) {
		rb_link_node(&insert->r_rb_node, parent, p);
		rb_insert_color(&insert->r_rb_node, root);
		refcount_inc(&insert->r_refcount);
	}
	return NULL;
}

/*
 * Destroy the transport-specific part of a MR.
 */
static void rds_destroy_mr(struct rds_mr *mr)
{
	struct rds_sock *rs = mr->r_sock;
	void *trans_private = NULL;
	unsigned long flags;

	rdsdebug("RDS: destroy mr key is %x refcnt %u\n",
			mr->r_key, refcount_read(&mr->r_refcount));

	if (test_and_set_bit(RDS_MR_DEAD, &mr->r_state))
		return;

	spin_lock_irqsave(&rs->rs_rdma_lock, flags);
	if (!RB_EMPTY_NODE(&mr->r_rb_node))
		rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
	trans_private = mr->r_trans_private;
	mr->r_trans_private = NULL;
	spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);

	if (trans_private)
		mr->r_trans->free_mr(trans_private, mr->r_invalidate);
}

void __rds_put_mr_final(struct rds_mr *mr)
{
	rds_destroy_mr(mr);
	kfree(mr);
}

/*
 * By the time this is called we can't have any more ioctls called on
 * the socket so we don't need to worry about racing with others.
 */
void rds_rdma_drop_keys(struct rds_sock *rs)
{
	struct rds_mr *mr;
	struct rb_node *node;
	unsigned long flags;

	/* Release any MRs associated with this socket */
	spin_lock_irqsave(&rs->rs_rdma_lock, flags);
	while ((node = rb_first(&rs->rs_rdma_keys))) {
		mr = rb_entry(node, struct rds_mr, r_rb_node);
		if (mr->r_trans == rs->rs_transport)
			mr->r_invalidate = 0;
		rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
		RB_CLEAR_NODE(&mr->r_rb_node);
		spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
		rds_destroy_mr(mr);
		rds_mr_put(mr);
		spin_lock_irqsave(&rs->rs_rdma_lock, flags);
	}
	spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);

	if (rs->rs_transport && rs->rs_transport->flush_mrs)
		rs->rs_transport->flush_mrs();
}

/*
 * Helper function to pin user pages.
 */
static int rds_pin_pages(unsigned long user_addr, unsigned int nr_pages,
			struct page **pages, int write)
{
	int ret;

	ret = get_user_pages_fast(user_addr, nr_pages, write ? FOLL_WRITE : 0,
				  pages);

	if (ret >= 0 && ret < nr_pages) {
		while (ret--)
			put_page(pages[ret]);
		ret = -EFAULT;
	}

	return ret;
}

static int __rds_rdma_map(struct rds_sock *rs, struct rds_get_mr_args *args,
			  u64 *cookie_ret, struct rds_mr **mr_ret,
			  struct rds_conn_path *cp)
{
	struct rds_mr *mr = NULL, *found;
	unsigned int nr_pages;
	struct page **pages = NULL;
	struct scatterlist *sg;
	void *trans_private;
	unsigned long flags;
	rds_rdma_cookie_t cookie;
	unsigned int nents;
	long i;
	int ret;

	if (ipv6_addr_any(&rs->rs_bound_addr) || !rs->rs_transport) {
		ret = -ENOTCONN; /* XXX not a great errno */
		goto out;
	}

	if (!rs->rs_transport->get_mr) {
		ret = -EOPNOTSUPP;
		goto out;
	}

	nr_pages = rds_pages_in_vec(&args->vec);
	if (nr_pages == 0) {
		ret = -EINVAL;
		goto out;
	}

	/* Restrict the size of mr irrespective of underlying transport
	 * To account for unaligned mr regions, subtract one from nr_pages
	 */
	if ((nr_pages - 1) > (RDS_MAX_MSG_SIZE >> PAGE_SHIFT)) {
		ret = -EMSGSIZE;
		goto out;
	}

	rdsdebug("RDS: get_mr addr %llx len %llu nr_pages %u\n",
		args->vec.addr, args->vec.bytes, nr_pages);

	/* XXX clamp nr_pages to limit the size of this alloc? */
	pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
	if (!pages) {
		ret = -ENOMEM;
		goto out;
	}

	mr = kzalloc(sizeof(struct rds_mr), GFP_KERNEL);
	if (!mr) {
		ret = -ENOMEM;
		goto out;
	}

	refcount_set(&mr->r_refcount, 1);
	RB_CLEAR_NODE(&mr->r_rb_node);
	mr->r_trans = rs->rs_transport;
	mr->r_sock = rs;

	if (args->flags & RDS_RDMA_USE_ONCE)
		mr->r_use_once = 1;
	if (args->flags & RDS_RDMA_INVALIDATE)
		mr->r_invalidate = 1;
	if (args->flags & RDS_RDMA_READWRITE)
		mr->r_write = 1;

	/*
	 * Pin the pages that make up the user buffer and transfer the page
	 * pointers to the mr's sg array.  We check to see if we've mapped
	 * the whole region after transferring the partial page references
	 * to the sg array so that we can have one page ref cleanup path.
	 *
	 * For now we have no flag that tells us whether the mapping is
	 * r/o or r/w. We need to assume r/w, or we'll do a lot of RDMA to
	 * the zero page.
	 */
	ret = rds_pin_pages(args->vec.addr, nr_pages, pages, 1);
	if (ret < 0)
		goto out;

	nents = ret;
	sg = kcalloc(nents, sizeof(*sg), GFP_KERNEL);
	if (!sg) {
		ret = -ENOMEM;
		goto out;
	}
	WARN_ON(!nents);
	sg_init_table(sg, nents);

	/* Stick all pages into the scatterlist */
	for (i = 0 ; i < nents; i++)
		sg_set_page(&sg[i], pages[i], PAGE_SIZE, 0);

	rdsdebug("RDS: trans_private nents is %u\n", nents);

	/* Obtain a transport specific MR. If this succeeds, the
	 * s/g list is now owned by the MR.
	 * Note that dma_map() implies that pending writes are
	 * flushed to RAM, so no dma_sync is needed here. */
	trans_private = rs->rs_transport->get_mr(sg, nents, rs,
						 &mr->r_key,
						 cp ? cp->cp_conn : NULL);

	if (IS_ERR(trans_private)) {
		for (i = 0 ; i < nents; i++)
			put_page(sg_page(&sg[i]));
		kfree(sg);
		ret = PTR_ERR(trans_private);
		goto out;
	}

	mr->r_trans_private = trans_private;

	rdsdebug("RDS: get_mr put_user key is %x cookie_addr %p\n",
	       mr->r_key, (void *)(unsigned long) args->cookie_addr);

	/* The user may pass us an unaligned address, but we can only
	 * map page aligned regions. So we keep the offset, and build
	 * a 64bit cookie containing <R_Key, offset> and pass that
	 * around. */
	cookie = rds_rdma_make_cookie(mr->r_key, args->vec.addr & ~PAGE_MASK);
	if (cookie_ret)
		*cookie_ret = cookie;

	if (args->cookie_addr && put_user(cookie, (u64 __user *)(unsigned long) args->cookie_addr)) {
		ret = -EFAULT;
		goto out;
	}

	/* Inserting the new MR into the rbtree bumps its
	 * reference count. */
	spin_lock_irqsave(&rs->rs_rdma_lock, flags);
	found = rds_mr_tree_walk(&rs->rs_rdma_keys, mr->r_key, mr);
	spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);

	BUG_ON(found && found != mr);

	rdsdebug("RDS: get_mr key is %x\n", mr->r_key);
	if (mr_ret) {
		refcount_inc(&mr->r_refcount);
		*mr_ret = mr;
	}

	ret = 0;
out:
	kfree(pages);
	if (mr)
		rds_mr_put(mr);
	return ret;
}

int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen)
{
	struct rds_get_mr_args args;

	if (optlen != sizeof(struct rds_get_mr_args))
		return -EINVAL;

	if (copy_from_user(&args, (struct rds_get_mr_args __user *)optval,
			   sizeof(struct rds_get_mr_args)))
		return -EFAULT;

	return __rds_rdma_map(rs, &args, NULL, NULL, NULL);
}

int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen)
{
	struct rds_get_mr_for_dest_args args;
	struct rds_get_mr_args new_args;

	if (optlen != sizeof(struct rds_get_mr_for_dest_args))
		return -EINVAL;

	if (copy_from_user(&args, (struct rds_get_mr_for_dest_args __user *)optval,
			   sizeof(struct rds_get_mr_for_dest_args)))
		return -EFAULT;

	/*
	 * Initially, just behave like get_mr().
	 * TODO: Implement get_mr as wrapper around this
	 *	 and deprecate it.
	 */
	new_args.vec = args.vec;
	new_args.cookie_addr = args.cookie_addr;
	new_args.flags = args.flags;

	return __rds_rdma_map(rs, &new_args, NULL, NULL, NULL);
}

/*
 * Free the MR indicated by the given R_Key
 */
int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen)
{
	struct rds_free_mr_args args;
	struct rds_mr *mr;
	unsigned long flags;

	if (optlen != sizeof(struct rds_free_mr_args))
		return -EINVAL;

	if (copy_from_user(&args, (struct rds_free_mr_args __user *)optval,
			   sizeof(struct rds_free_mr_args)))
		return -EFAULT;

	/* Special case - a null cookie means flush all unused MRs */
	if (args.cookie == 0) {
		if (!rs->rs_transport || !rs->rs_transport->flush_mrs)
			return -EINVAL;
		rs->rs_transport->flush_mrs();
		return 0;
	}

	/* Look up the MR given its R_key and remove it from the rbtree
	 * so nobody else finds it.
	 * This should also prevent races with rds_rdma_unuse.
	 */
	spin_lock_irqsave(&rs->rs_rdma_lock, flags);
	mr = rds_mr_tree_walk(&rs->rs_rdma_keys, rds_rdma_cookie_key(args.cookie), NULL);
	if (mr) {
		rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
		RB_CLEAR_NODE(&mr->r_rb_node);
		if (args.flags & RDS_RDMA_INVALIDATE)
			mr->r_invalidate = 1;
	}
	spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);

	if (!mr)
		return -EINVAL;

	/*
	 * call rds_destroy_mr() ourselves so that we're sure it's done by the time
	 * we return.  If we let rds_mr_put() do it it might not happen until
	 * someone else drops their ref.
	 */
	rds_destroy_mr(mr);
	rds_mr_put(mr);
	return 0;
}

/*
 * This is called when we receive an extension header that
 * tells us this MR was used. It allows us to implement
 * use_once semantics
 */
void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force)
{
	struct rds_mr *mr;
	unsigned long flags;
	int zot_me = 0;

	spin_lock_irqsave(&rs->rs_rdma_lock, flags);
	mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL);
	if (!mr) {
		pr_debug("rds: trying to unuse MR with unknown r_key %u!\n",
			 r_key);
		spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
		return;
	}

	if (mr->r_use_once || force) {
		rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
		RB_CLEAR_NODE(&mr->r_rb_node);
		zot_me = 1;
	}
	spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);

	/* May have to issue a dma_sync on this memory region.
	 * Note we could avoid this if the operation was a RDMA READ,
	 * but at this point we can't tell. */
	if (mr->r_trans->sync_mr)
		mr->r_trans->sync_mr(mr->r_trans_private, DMA_FROM_DEVICE);

	/* If the MR was marked as invalidate, this will
	 * trigger an async flush. */
	if (zot_me) {
		rds_destroy_mr(mr);
		rds_mr_put(mr);
	}
}

void rds_rdma_free_op(struct rm_rdma_op *ro)
{
	unsigned int i;

	for (i = 0; i < ro->op_nents; i++) {
		struct page *page = sg_page(&ro->op_sg[i]);

		/* Mark page dirty if it was possibly modified, which
		 * is the case for a RDMA_READ which copies from remote
		 * to local memory */
		if (!ro->op_write) {
			WARN_ON(!page->mapping && irqs_disabled());
			set_page_dirty(page);
		}
		put_page(page);
	}

	kfree(ro->op_notifier);
	ro->op_notifier = NULL;
	ro->op_active = 0;
}

void rds_atomic_free_op(struct rm_atomic_op *ao)
{
	struct page *page = sg_page(ao->op_sg);

	/* Mark page dirty if it was possibly modified, which
	 * is the case for a RDMA_READ which copies from remote
	 * to local memory */
	set_page_dirty(page);
	put_page(page);

	kfree(ao->op_notifier);
	ao->op_notifier = NULL;
	ao->op_active = 0;
}


/*
 * Count the number of pages needed to describe an incoming iovec array.
 */
static int rds_rdma_pages(struct rds_iovec iov[], int nr_iovecs)
{
	int tot_pages = 0;
	unsigned int nr_pages;
	unsigned int i;

	/* figure out the number of pages in the vector */
	for (i = 0; i < nr_iovecs; i++) {
		nr_pages = rds_pages_in_vec(&iov[i]);
		if (nr_pages == 0)
			return -EINVAL;

		tot_pages += nr_pages;

		/*
		 * nr_pages for one entry is limited to (UINT_MAX>>PAGE_SHIFT)+1,
		 * so tot_pages cannot overflow without first going negative.
		 */
		if (tot_pages < 0)
			return -EINVAL;
	}

	return tot_pages;
}

int rds_rdma_extra_size(struct rds_rdma_args *args,
			struct rds_iov_vector *iov)
{
	struct rds_iovec *vec;
	struct rds_iovec __user *local_vec;
	int tot_pages = 0;
	unsigned int nr_pages;
	unsigned int i;

	local_vec = (struct rds_iovec __user *)(unsigned long) args->local_vec_addr;

	if (args->nr_local == 0)
		return -EINVAL;

	iov->iov = kcalloc(args->nr_local,
			   sizeof(struct rds_iovec),
			   GFP_KERNEL);
	if (!iov->iov)
		return -ENOMEM;

	vec = &iov->iov[0];

	if (copy_from_user(vec, local_vec, args->nr_local *
			   sizeof(struct rds_iovec)))
		return -EFAULT;
	iov->len = args->nr_local;

	/* figure out the number of pages in the vector */
	for (i = 0; i < args->nr_local; i++, vec++) {

		nr_pages = rds_pages_in_vec(vec);
		if (nr_pages == 0)
			return -EINVAL;

		tot_pages += nr_pages;

		/*
		 * nr_pages for one entry is limited to (UINT_MAX>>PAGE_SHIFT)+1,
		 * so tot_pages cannot overflow without first going negative.
		 */
		if (tot_pages < 0)
			return -EINVAL;
	}

	return tot_pages * sizeof(struct scatterlist);
}

/*
 * The application asks for a RDMA transfer.
 * Extract all arguments and set up the rdma_op
 */
int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
		       struct cmsghdr *cmsg,
		       struct rds_iov_vector *vec)
{
	struct rds_rdma_args *args;
	struct rm_rdma_op *op = &rm->rdma;
	int nr_pages;
	unsigned int nr_bytes;
	struct page **pages = NULL;
	struct rds_iovec *iovs;
	unsigned int i, j;
	int ret = 0;

	if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_rdma_args))
	    || rm->rdma.op_active)
		return -EINVAL;

	args = CMSG_DATA(cmsg);

	if (ipv6_addr_any(&rs->rs_bound_addr)) {
		ret = -ENOTCONN; /* XXX not a great errno */
		goto out_ret;
	}

	if (args->nr_local > UIO_MAXIOV) {
		ret = -EMSGSIZE;
		goto out_ret;
	}

	if (vec->len != args->nr_local) {
		ret = -EINVAL;
		goto out_ret;
	}

	iovs = vec->iov;

	nr_pages = rds_rdma_pages(iovs, args->nr_local);
	if (nr_pages < 0) {
		ret = -EINVAL;
		goto out_ret;
	}

	pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
	if (!pages) {
		ret = -ENOMEM;
		goto out_ret;
	}

	op->op_write = !!(args->flags & RDS_RDMA_READWRITE);
	op->op_fence = !!(args->flags & RDS_RDMA_FENCE);
	op->op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME);
	op->op_silent = !!(args->flags & RDS_RDMA_SILENT);
	op->op_active = 1;
	op->op_recverr = rs->rs_recverr;
	WARN_ON(!nr_pages);
	op->op_sg = rds_message_alloc_sgs(rm, nr_pages, &ret);
	if (!op->op_sg)
		goto out_pages;

	if (op->op_notify || op->op_recverr) {
		/* We allocate an uninitialized notifier here, because
		 * we don't want to do that in the completion handler. We
		 * would have to use GFP_ATOMIC there, and don't want to deal
		 * with failed allocations.
		 */
		op->op_notifier = kmalloc(sizeof(struct rds_notifier), GFP_KERNEL);
		if (!op->op_notifier) {
			ret = -ENOMEM;
			goto out_pages;
		}
		op->op_notifier->n_user_token = args->user_token;
		op->op_notifier->n_status = RDS_RDMA_SUCCESS;
	}

	/* The cookie contains the R_Key of the remote memory region, and
	 * optionally an offset into it. This is how we implement RDMA into
	 * unaligned memory.
	 * When setting up the RDMA, we need to add that offset to the
	 * destination address (which is really an offset into the MR)
	 * FIXME: We may want to move this into ib_rdma.c
	 */
	op->op_rkey = rds_rdma_cookie_key(args->cookie);
	op->op_remote_addr = args->remote_vec.addr + rds_rdma_cookie_offset(args->cookie);

	nr_bytes = 0;

	rdsdebug("RDS: rdma prepare nr_local %llu rva %llx rkey %x\n",
	       (unsigned long long)args->nr_local,
	       (unsigned long long)args->remote_vec.addr,
	       op->op_rkey);

	for (i = 0; i < args->nr_local; i++) {
		struct rds_iovec *iov = &iovs[i];
		/* don't need to check, rds_rdma_pages() verified nr will be +nonzero */
		unsigned int nr = rds_pages_in_vec(iov);

		rs->rs_user_addr = iov->addr;
		rs->rs_user_bytes = iov->bytes;

		/* If it's a WRITE operation, we want to pin the pages for reading.
		 * If it's a READ operation, we need to pin the pages for writing.
		 */
		ret = rds_pin_pages(iov->addr, nr, pages, !op->op_write);
		if (ret < 0)
			goto out_pages;
		else
			ret = 0;

		rdsdebug("RDS: nr_bytes %u nr %u iov->bytes %llu iov->addr %llx\n",
			 nr_bytes, nr, iov->bytes, iov->addr);

		nr_bytes += iov->bytes;

		for (j = 0; j < nr; j++) {
			unsigned int offset = iov->addr & ~PAGE_MASK;
			struct scatterlist *sg;

			sg = &op->op_sg[op->op_nents + j];
			sg_set_page(sg, pages[j],
					min_t(unsigned int, iov->bytes, PAGE_SIZE - offset),
					offset);

			rdsdebug("RDS: sg->offset %x sg->len %x iov->addr %llx iov->bytes %llu\n",
			       sg->offset, sg->length, iov->addr, iov->bytes);

			iov->addr += sg->length;
			iov->bytes -= sg->length;
		}

		op->op_nents += nr;
	}

	if (nr_bytes > args->remote_vec.bytes) {
		rdsdebug("RDS nr_bytes %u remote_bytes %u do not match\n",
				nr_bytes,
				(unsigned int) args->remote_vec.bytes);
		ret = -EINVAL;
		goto out_pages;
	}
	op->op_bytes = nr_bytes;

out_pages:
	kfree(pages);
out_ret:
	if (ret)
		rds_rdma_free_op(op);
	else
		rds_stats_inc(s_send_rdma);

	return ret;
}

/*
 * The application wants us to pass an RDMA destination (aka MR)
 * to the remote
 */
int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
			  struct cmsghdr *cmsg)
{
	unsigned long flags;
	struct rds_mr *mr;
	u32 r_key;
	int err = 0;

	if (cmsg->cmsg_len < CMSG_LEN(sizeof(rds_rdma_cookie_t)) ||
	    rm->m_rdma_cookie != 0)
		return -EINVAL;

	memcpy(&rm->m_rdma_cookie, CMSG_DATA(cmsg), sizeof(rm->m_rdma_cookie));

	/* We are reusing a previously mapped MR here. Most likely, the
	 * application has written to the buffer, so we need to explicitly
	 * flush those writes to RAM. Otherwise the HCA may not see them
	 * when doing a DMA from that buffer.
	 */
	r_key = rds_rdma_cookie_key(rm->m_rdma_cookie);

	spin_lock_irqsave(&rs->rs_rdma_lock, flags);
	mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL);
	if (!mr)
		err = -EINVAL;	/* invalid r_key */
	else
		refcount_inc(&mr->r_refcount);
	spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);

	if (mr) {
		mr->r_trans->sync_mr(mr->r_trans_private, DMA_TO_DEVICE);
		rm->rdma.op_rdma_mr = mr;
	}
	return err;
}

/*
 * The application passes us an address range it wants to enable RDMA
 * to/from. We map the area, and save the <R_Key,offset> pair
 * in rm->m_rdma_cookie. This causes it to be sent along to the peer
 * in an extension header.
 */
int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
			  struct cmsghdr *cmsg)
{
	if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_get_mr_args)) ||
	    rm->m_rdma_cookie != 0)
		return -EINVAL;

	return __rds_rdma_map(rs, CMSG_DATA(cmsg), &rm->m_rdma_cookie,
			      &rm->rdma.op_rdma_mr, rm->m_conn_path);
}

/*
 * Fill in rds_message for an atomic request.
 */
int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
		    struct cmsghdr *cmsg)
{
	struct page *page = NULL;
	struct rds_atomic_args *args;
	int ret = 0;

	if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_atomic_args))
	 || rm->atomic.op_active)
		return -EINVAL;

	args = CMSG_DATA(cmsg);

	/* Nonmasked & masked cmsg ops converted to masked hw ops */
	switch (cmsg->cmsg_type) {
	case RDS_CMSG_ATOMIC_FADD:
		rm->atomic.op_type = RDS_ATOMIC_TYPE_FADD;
		rm->atomic.op_m_fadd.add = args->fadd.add;
		rm->atomic.op_m_fadd.nocarry_mask = 0;
		break;
	case RDS_CMSG_MASKED_ATOMIC_FADD:
		rm->atomic.op_type = RDS_ATOMIC_TYPE_FADD;
		rm->atomic.op_m_fadd.add = args->m_fadd.add;
		rm->atomic.op_m_fadd.nocarry_mask = args->m_fadd.nocarry_mask;
		break;
	case RDS_CMSG_ATOMIC_CSWP:
		rm->atomic.op_type = RDS_ATOMIC_TYPE_CSWP;
		rm->atomic.op_m_cswp.compare = args->cswp.compare;
		rm->atomic.op_m_cswp.swap = args->cswp.swap;
		rm->atomic.op_m_cswp.compare_mask = ~0;
		rm->atomic.op_m_cswp.swap_mask = ~0;
		break;
	case RDS_CMSG_MASKED_ATOMIC_CSWP:
		rm->atomic.op_type = RDS_ATOMIC_TYPE_CSWP;
		rm->atomic.op_m_cswp.compare = args->m_cswp.compare;
		rm->atomic.op_m_cswp.swap = args->m_cswp.swap;
		rm->atomic.op_m_cswp.compare_mask = args->m_cswp.compare_mask;
		rm->atomic.op_m_cswp.swap_mask = args->m_cswp.swap_mask;
		break;
	default:
		BUG(); /* should never happen */
	}

	rm->atomic.op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME);
	rm->atomic.op_silent = !!(args->flags & RDS_RDMA_SILENT);
	rm->atomic.op_active = 1;
	rm->atomic.op_recverr = rs->rs_recverr;
	rm->atomic.op_sg = rds_message_alloc_sgs(rm, 1, &ret);
	if (!rm->atomic.op_sg)
		goto err;

	/* verify 8 byte-aligned */
	if (args->local_addr & 0x7) {
		ret = -EFAULT;
		goto err;
	}

	ret = rds_pin_pages(args->local_addr, 1, &page, 1);
	if (ret != 1)
		goto err;
	ret = 0;

	sg_set_page(rm->atomic.op_sg, page, 8, offset_in_page(args->local_addr));

	if (rm->atomic.op_notify || rm->atomic.op_recverr) {
		/* We allocate an uninitialized notifier here, because
		 * we don't want to do that in the completion handler. We
		 * would have to use GFP_ATOMIC there, and don't want to deal
		 * with failed allocations.
		 */
		rm->atomic.op_notifier = kmalloc(sizeof(*rm->atomic.op_notifier), GFP_KERNEL);
		if (!rm->atomic.op_notifier) {
			ret = -ENOMEM;
			goto err;
		}

		rm->atomic.op_notifier->n_user_token = args->user_token;
		rm->atomic.op_notifier->n_status = RDS_RDMA_SUCCESS;
	}

	rm->atomic.op_rkey = rds_rdma_cookie_key(args->cookie);
	rm->atomic.op_remote_addr = args->remote_addr + rds_rdma_cookie_offset(args->cookie);

	return ret;
err:
	if (page)
		put_page(page);
	rm->atomic.op_active = 0;
	kfree(rm->atomic.op_notifier);

	return ret;
}