xref: /illumos-gate/usr/src/uts/common/io/ib/clients/rdsv3/rdma.c (revision 5328fc53d11d7151861fa272e4fb0248b8f0e145)
1 /*
2  * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
3  */
4 
5 /*
6  * This file contains code imported from the OFED rds source file rdma.c
7  * Oracle elects to have and use the contents of rdma.c under and governed
8  * by the OpenIB.org BSD license (see below for full license text). However,
9  * the following notice accompanied the original version of this file:
10  */
11 
12 /*
13  * Copyright (c) 2007 Oracle.  All rights reserved.
14  *
15  * This software is available to you under a choice of one of two
16  * licenses.  You may choose to be licensed under the terms of the GNU
17  * General Public License (GPL) Version 2, available from the file
18  * COPYING in the main directory of this source tree, or the
19  * OpenIB.org BSD license below:
20  *
21  *     Redistribution and use in source and binary forms, with or
22  *     without modification, are permitted provided that the following
23  *     conditions are met:
24  *
25  *      - Redistributions of source code must retain the above
26  *        copyright notice, this list of conditions and the following
27  *        disclaimer.
28  *
29  *      - Redistributions in binary form must reproduce the above
30  *        copyright notice, this list of conditions and the following
31  *        disclaimer in the documentation and/or other materials
32  *        provided with the distribution.
33  *
34  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
35  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
36  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
37  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
38  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
39  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
40  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
41  * SOFTWARE.
42  *
43  */
44 #include <sys/ib/clients/of/rdma/ib_verbs.h>
45 #include <sys/ib/clients/of/rdma/ib_addr.h>
46 #include <sys/ib/clients/of/rdma/rdma_cm.h>
47 
48 #include <sys/ib/clients/rdsv3/ib.h>
49 #include <sys/ib/clients/rdsv3/rdma.h>
50 #include <sys/ib/clients/rdsv3/rdsv3_debug.h>
51 #include <sys/containerof.h>
52 
53 #define	DMA_TO_DEVICE 0
54 #define	DMA_FROM_DEVICE 1
55 #define	RB_CLEAR_NODE(nodep) AVL_SETPARENT(nodep, nodep);
56 
57 /*
58  * XXX
59  *  - build with sparse
60  *  - should we limit the size of a mr region?  let transport return failure?
61  *  - should we detect duplicate keys on a socket?  hmm.
62  *  - an rdma is an mlock, apply rlimit?
63  */
64 
65 /*
66  * get the number of pages by looking at the page indices that the start and
67  * end addresses fall in.
68  *
69  * Returns 0 if the vec is invalid.  It is invalid if the number of bytes
70  * causes the address to wrap or overflows an unsigned int.  This comes
71  * from being stored in the 'length' member of 'struct rdsv3_scatterlist'.
72  */
73 static unsigned int
74 rdsv3_pages_in_vec(struct rds_iovec *vec)
75 {
76 	if ((vec->addr + vec->bytes <= vec->addr) ||
77 	    (vec->bytes > (uint64_t)UINT_MAX)) {
78 		return (0);
79 	}
80 
81 	return (((vec->addr + vec->bytes + PAGESIZE - 1) >>
82 	    PAGESHIFT) - (vec->addr >> PAGESHIFT));
83 }
84 
85 static struct rdsv3_mr *
86 rdsv3_mr_tree_walk(struct avl_tree *root, uint32_t key,
87     struct rdsv3_mr *insert)
88 {
89 	struct rdsv3_mr *mr;
90 	avl_index_t where;
91 
92 	mr = avl_find(root, &key, &where);
93 	if ((mr == NULL) && (insert != NULL)) {
94 		avl_insert(root, (void *)insert, where);
95 		atomic_inc_32(&insert->r_refcount);
96 		return (NULL);
97 	}
98 
99 	return (mr);
100 }
101 
102 /*
103  * Destroy the transport-specific part of a MR.
104  */
105 static void
106 rdsv3_destroy_mr(struct rdsv3_mr *mr)
107 {
108 	struct rdsv3_sock *rs = mr->r_sock;
109 	void *trans_private = NULL;
110 	avl_node_t *np;
111 
112 	RDSV3_DPRINTF5("rdsv3_destroy_mr",
113 	    "RDS: destroy mr key is %x refcnt %u",
114 	    mr->r_key, atomic_get(&mr->r_refcount));
115 
116 	if (test_and_set_bit(RDSV3_MR_DEAD, &mr->r_state))
117 		return;
118 
119 	mutex_enter(&rs->rs_rdma_lock);
120 	np = &mr->r_rb_node;
121 	if (AVL_XPARENT(np) != np)
122 		avl_remove(&rs->rs_rdma_keys, mr);
123 	trans_private = mr->r_trans_private;
124 	mr->r_trans_private = NULL;
125 	mutex_exit(&rs->rs_rdma_lock);
126 
127 	if (trans_private)
128 		mr->r_trans->free_mr(trans_private, mr->r_invalidate);
129 }
130 
131 void
132 __rdsv3_put_mr_final(struct rdsv3_mr *mr)
133 {
134 	rdsv3_destroy_mr(mr);
135 	kmem_free(mr, sizeof (*mr));
136 }
137 
138 /*
139  * By the time this is called we can't have any more ioctls called on
140  * the socket so we don't need to worry about racing with others.
141  */
142 void
143 rdsv3_rdma_drop_keys(struct rdsv3_sock *rs)
144 {
145 	struct rdsv3_mr *mr;
146 	struct avl_node *node;
147 
148 	/* Release any MRs associated with this socket */
149 	mutex_enter(&rs->rs_rdma_lock);
150 	while ((node = avl_first(&rs->rs_rdma_keys))) {
151 		mr = __containerof(node, struct rdsv3_mr, r_rb_node);
152 		if (mr->r_trans == rs->rs_transport)
153 			mr->r_invalidate = 0;
154 		avl_remove(&rs->rs_rdma_keys, &mr->r_rb_node);
155 		RB_CLEAR_NODE(&mr->r_rb_node)
156 		mutex_exit(&rs->rs_rdma_lock);
157 		rdsv3_destroy_mr(mr);
158 		rdsv3_mr_put(mr);
159 		mutex_enter(&rs->rs_rdma_lock);
160 	}
161 	mutex_exit(&rs->rs_rdma_lock);
162 
163 	if (rs->rs_transport && rs->rs_transport->flush_mrs)
164 		rs->rs_transport->flush_mrs();
165 }
166 
167 static int
168 __rdsv3_rdma_map(struct rdsv3_sock *rs, struct rds_get_mr_args *args,
169     uint64_t *cookie_ret, struct rdsv3_mr **mr_ret)
170 {
171 	struct rdsv3_mr *mr = NULL, *found;
172 	void *trans_private;
173 	rds_rdma_cookie_t cookie;
174 	unsigned int nents = 0;
175 	int ret;
176 
177 	if (rs->rs_bound_addr == 0) {
178 		ret = -ENOTCONN; /* XXX not a great errno */
179 		goto out;
180 	}
181 
182 	if (!rs->rs_transport->get_mr) {
183 		ret = -EOPNOTSUPP;
184 		goto out;
185 	}
186 
187 	mr = kmem_zalloc(sizeof (struct rdsv3_mr), KM_NOSLEEP);
188 	if (!mr) {
189 		ret = -ENOMEM;
190 		goto out;
191 	}
192 
193 	mr->r_refcount = 1;
194 	RB_CLEAR_NODE(&mr->r_rb_node);
195 	mr->r_trans = rs->rs_transport;
196 	mr->r_sock = rs;
197 
198 	if (args->flags & RDS_RDMA_USE_ONCE)
199 		mr->r_use_once = 1;
200 	if (args->flags & RDS_RDMA_INVALIDATE)
201 		mr->r_invalidate = 1;
202 	if (args->flags & RDS_RDMA_READWRITE)
203 		mr->r_write = 1;
204 
205 	/*
206 	 * Obtain a transport specific MR. If this succeeds, the
207 	 * s/g list is now owned by the MR.
208 	 * Note that dma_map() implies that pending writes are
209 	 * flushed to RAM, so no dma_sync is needed here.
210 	 */
211 	trans_private = rs->rs_transport->get_mr(&args->vec, nents, rs,
212 	    &mr->r_key);
213 
214 	if (IS_ERR(trans_private)) {
215 		ret = PTR_ERR(trans_private);
216 		goto out;
217 	}
218 
219 	mr->r_trans_private = trans_private;
220 
221 	/*
222 	 * The user may pass us an unaligned address, but we can only
223 	 * map page aligned regions. So we keep the offset, and build
224 	 * a 64bit cookie containing <R_Key, offset> and pass that
225 	 * around.
226 	 */
227 	cookie = rdsv3_rdma_make_cookie(mr->r_key, args->vec.addr & ~PAGEMASK);
228 	if (cookie_ret)
229 		*cookie_ret = cookie;
230 
231 	/*
232 	 * copy value of cookie to user address at args->cookie_addr
233 	 */
234 	if (args->cookie_addr) {
235 		ret = ddi_copyout((void *)&cookie,
236 		    (void *)((intptr_t)args->cookie_addr),
237 		    sizeof (rds_rdma_cookie_t), 0);
238 		if (ret != 0) {
239 			ret = -EFAULT;
240 			goto out;
241 		}
242 	}
243 
244 	RDSV3_DPRINTF5("__rdsv3_rdma_map",
245 	    "RDS: get_mr mr 0x%p addr 0x%llx key 0x%x",
246 	    mr, args->vec.addr, mr->r_key);
247 	/*
248 	 * Inserting the new MR into the rbtree bumps its
249 	 * reference count.
250 	 */
251 	mutex_enter(&rs->rs_rdma_lock);
252 	found = rdsv3_mr_tree_walk(&rs->rs_rdma_keys, mr->r_key, mr);
253 	mutex_exit(&rs->rs_rdma_lock);
254 
255 	ASSERT(!(found && found != mr));
256 
257 	if (mr_ret) {
258 		atomic_inc_32(&mr->r_refcount);
259 		*mr_ret = mr;
260 	}
261 
262 	ret = 0;
263 out:
264 	if (mr)
265 		rdsv3_mr_put(mr);
266 	return (ret);
267 }
268 
269 int
270 rdsv3_get_mr(struct rdsv3_sock *rs, const void *optval, int optlen)
271 {
272 	struct rds_get_mr_args args;
273 
274 	if (optlen != sizeof (struct rds_get_mr_args))
275 		return (-EINVAL);
276 
277 #if 1
278 	bcopy((struct rds_get_mr_args *)optval, &args,
279 	    sizeof (struct rds_get_mr_args));
280 #else
281 	if (ddi_copyin(optval, &args, optlen, 0))
282 		return (-EFAULT);
283 #endif
284 
285 	return (__rdsv3_rdma_map(rs, &args, NULL, NULL));
286 }
287 
288 int
289 rdsv3_get_mr_for_dest(struct rdsv3_sock *rs, const void *optval,
290     int optlen)
291 {
292 	struct rds_get_mr_for_dest_args args;
293 	struct rds_get_mr_args new_args;
294 
295 	if (optlen != sizeof (struct rds_get_mr_for_dest_args))
296 		return (-EINVAL);
297 
298 #if 1
299 	bcopy((struct rds_get_mr_for_dest_args *)optval, &args,
300 	    sizeof (struct rds_get_mr_for_dest_args));
301 #else
302 	if (ddi_copyin(optval, &args, optlen, 0))
303 		return (-EFAULT);
304 #endif
305 
306 	/*
307 	 * Initially, just behave like get_mr().
308 	 * TODO: Implement get_mr as wrapper around this
309 	 *	 and deprecate it.
310 	 */
311 	new_args.vec = args.vec;
312 	new_args.cookie_addr = args.cookie_addr;
313 	new_args.flags = args.flags;
314 
315 	return (__rdsv3_rdma_map(rs, &new_args, NULL, NULL));
316 }
317 
318 /*
319  * Free the MR indicated by the given R_Key
320  */
321 int
322 rdsv3_free_mr(struct rdsv3_sock *rs, const void *optval, int optlen)
323 {
324 	struct rds_free_mr_args args;
325 	struct rdsv3_mr *mr;
326 
327 	if (optlen != sizeof (struct rds_free_mr_args))
328 		return (-EINVAL);
329 
330 #if 1
331 	bcopy((struct rds_free_mr_args *)optval, &args,
332 	    sizeof (struct rds_free_mr_args));
333 #else
334 	if (ddi_copyin((struct rds_free_mr_args *)optval, &args,
335 	    sizeof (struct rds_free_mr_args), 0))
336 		return (-EFAULT);
337 #endif
338 
339 	/* Special case - a null cookie means flush all unused MRs */
340 	if (args.cookie == 0) {
341 		if (!rs->rs_transport || !rs->rs_transport->flush_mrs)
342 			return (-EINVAL);
343 		rs->rs_transport->flush_mrs();
344 		return (0);
345 	}
346 
347 	/*
348 	 * Look up the MR given its R_key and remove it from the rbtree
349 	 * so nobody else finds it.
350 	 * This should also prevent races with rdsv3_rdma_unuse.
351 	 */
352 	mutex_enter(&rs->rs_rdma_lock);
353 	mr = rdsv3_mr_tree_walk(&rs->rs_rdma_keys,
354 	    rdsv3_rdma_cookie_key(args.cookie), NULL);
355 	if (mr) {
356 		avl_remove(&rs->rs_rdma_keys, &mr->r_rb_node);
357 		RB_CLEAR_NODE(&mr->r_rb_node);
358 		if (args.flags & RDS_RDMA_INVALIDATE)
359 			mr->r_invalidate = 1;
360 	}
361 	mutex_exit(&rs->rs_rdma_lock);
362 
363 	if (!mr)
364 		return (-EINVAL);
365 
366 	/*
367 	 * call rdsv3_destroy_mr() ourselves so that we're sure it's done
368 	 * by time we return.  If we let rdsv3_mr_put() do it it might not
369 	 * happen until someone else drops their ref.
370 	 */
371 	rdsv3_destroy_mr(mr);
372 	rdsv3_mr_put(mr);
373 	return (0);
374 }
375 
376 /*
377  * This is called when we receive an extension header that
378  * tells us this MR was used. It allows us to implement
379  * use_once semantics
380  */
381 void
382 rdsv3_rdma_unuse(struct rdsv3_sock *rs, uint32_t r_key, int force)
383 {
384 	struct rdsv3_mr *mr;
385 	int zot_me = 0;
386 
387 	RDSV3_DPRINTF4("rdsv3_rdma_unuse", "Enter rkey: 0x%x", r_key);
388 
389 	mutex_enter(&rs->rs_rdma_lock);
390 	mr = rdsv3_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL);
391 	if (!mr) {
392 		RDSV3_DPRINTF4("rdsv3_rdma_unuse",
393 		    "rdsv3: trying to unuse MR with unknown r_key %u!", r_key);
394 		mutex_exit(&rs->rs_rdma_lock);
395 		return;
396 	}
397 
398 	if (mr->r_use_once || force) {
399 		avl_remove(&rs->rs_rdma_keys, &mr->r_rb_node);
400 		RB_CLEAR_NODE(&mr->r_rb_node);
401 		zot_me = 1;
402 	} else {
403 		atomic_inc_32(&mr->r_refcount);
404 	}
405 	mutex_exit(&rs->rs_rdma_lock);
406 
407 	/*
408 	 * May have to issue a dma_sync on this memory region.
409 	 * Note we could avoid this if the operation was a RDMA READ,
410 	 * but at this point we can't tell.
411 	 */
412 	if (mr->r_trans->sync_mr)
413 		mr->r_trans->sync_mr(mr->r_trans_private, DMA_FROM_DEVICE);
414 
415 	/*
416 	 * If the MR was marked as invalidate, this will
417 	 * trigger an async flush.
418 	 */
419 	if (zot_me)
420 		rdsv3_destroy_mr(mr);
421 	rdsv3_mr_put(mr);
422 	RDSV3_DPRINTF4("rdsv3_rdma_unuse", "Return");
423 }
424 
425 void
426 rdsv3_rdma_free_op(struct rdsv3_rdma_op *ro)
427 {
428 	unsigned int i;
429 
430 	/* deallocate RDMA resources on rdsv3_message */
431 	for (i = 0; i < ro->r_nents; i++) {
432 		ddi_umem_unlock(ro->r_rdma_sg[i].umem_cookie);
433 	}
434 
435 	if (ro->r_notifier)
436 		kmem_free(ro->r_notifier, sizeof (*ro->r_notifier));
437 	kmem_free(ro, sizeof (*ro));
438 }
439 
440 /*
441  * args is a pointer to an in-kernel copy in the sendmsg cmsg.
442  */
443 static struct rdsv3_rdma_op *
444 rdsv3_rdma_prepare(struct rdsv3_sock *rs, struct rds_rdma_args *args)
445 {
446 	struct rds_iovec vec;
447 	struct rdsv3_rdma_op *op = NULL;
448 	unsigned int nr_bytes;
449 	struct rds_iovec *local_vec;
450 	unsigned int nr;
451 	unsigned int i;
452 	ddi_umem_cookie_t umem_cookie;
453 	size_t umem_len;
454 	caddr_t umem_addr;
455 	int ret;
456 
457 	if (rs->rs_bound_addr == 0) {
458 		ret = -ENOTCONN; /* XXX not a great errno */
459 		goto out;
460 	}
461 
462 	if (args->nr_local > (uint64_t)UINT_MAX) {
463 		ret = -EMSGSIZE;
464 		goto out;
465 	}
466 
467 	op = kmem_zalloc(offsetof(struct rdsv3_rdma_op,
468 	    r_rdma_sg[args->nr_local]), KM_NOSLEEP);
469 	if (op == NULL) {
470 		ret = -ENOMEM;
471 		goto out;
472 	}
473 
474 	op->r_write = !!(args->flags & RDS_RDMA_READWRITE);
475 	op->r_fence = !!(args->flags & RDS_RDMA_FENCE);
476 	op->r_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME);
477 	op->r_recverr = rs->rs_recverr;
478 
479 	if (op->r_notify || op->r_recverr) {
480 		/*
481 		 * We allocate an uninitialized notifier here, because
482 		 * we don't want to do that in the completion handler. We
483 		 * would have to use GFP_ATOMIC there, and don't want to deal
484 		 * with failed allocations.
485 		 */
486 		op->r_notifier = kmem_alloc(sizeof (struct rdsv3_notifier),
487 		    KM_NOSLEEP);
488 		if (!op->r_notifier) {
489 			ret = -ENOMEM;
490 			goto out;
491 		}
492 		op->r_notifier->n_user_token = args->user_token;
493 		op->r_notifier->n_status = RDS_RDMA_SUCCESS;
494 	}
495 
496 	/*
497 	 * The cookie contains the R_Key of the remote memory region, and
498 	 * optionally an offset into it. This is how we implement RDMA into
499 	 * unaligned memory.
500 	 * When setting up the RDMA, we need to add that offset to the
501 	 * destination address (which is really an offset into the MR)
502 	 * FIXME: We may want to move this into ib_rdma.c
503 	 */
504 	op->r_key = rdsv3_rdma_cookie_key(args->cookie);
505 	op->r_remote_addr = args->remote_vec.addr +
506 	    rdsv3_rdma_cookie_offset(args->cookie);
507 
508 	nr_bytes = 0;
509 
510 	RDSV3_DPRINTF5("rdsv3_rdma_prepare",
511 	    "RDS: rdma prepare nr_local %llu rva %llx rkey %x",
512 	    (unsigned long long)args->nr_local,
513 	    (unsigned long long)args->remote_vec.addr,
514 	    op->r_key);
515 
516 	local_vec = (struct rds_iovec *)(unsigned long) args->local_vec_addr;
517 
518 	/* pin the scatter list of user buffers */
519 	for (i = 0; i < args->nr_local; i++) {
520 		if (ddi_copyin(&local_vec[i], &vec,
521 		    sizeof (struct rds_iovec), 0)) {
522 			ret = -EFAULT;
523 			goto out;
524 		}
525 
526 		nr = rdsv3_pages_in_vec(&vec);
527 		if (nr == 0) {
528 			RDSV3_DPRINTF2("rdsv3_rdma_prepare",
529 			    "rdsv3_pages_in_vec returned 0");
530 			ret = -EINVAL;
531 			goto out;
532 		}
533 
534 		rs->rs_user_addr = vec.addr;
535 		rs->rs_user_bytes = vec.bytes;
536 
537 		/* pin user memory pages */
538 		umem_len = ptob(btopr(vec.bytes +
539 		    ((uintptr_t)vec.addr & PAGEOFFSET)));
540 		umem_addr = (caddr_t)((uintptr_t)vec.addr & ~PAGEOFFSET);
541 		ret = umem_lockmemory(umem_addr, umem_len,
542 		    DDI_UMEMLOCK_WRITE | DDI_UMEMLOCK_READ,
543 		    &umem_cookie, NULL, NULL);
544 		if (ret != 0) {
545 			RDSV3_DPRINTF2("rdsv3_rdma_prepare",
546 			    "umem_lockmemory() returned %d", ret);
547 			ret = -EFAULT;
548 			goto out;
549 		}
550 		op->r_rdma_sg[i].umem_cookie = umem_cookie;
551 		op->r_rdma_sg[i].iovec = vec;
552 		nr_bytes += vec.bytes;
553 
554 		RDSV3_DPRINTF5("rdsv3_rdma_prepare",
555 		    "RDS: nr_bytes %u nr %u vec.bytes %llu vec.addr %llx",
556 		    nr_bytes, nr, vec.bytes, vec.addr);
557 	}
558 	op->r_nents = i;
559 
560 	if (nr_bytes > args->remote_vec.bytes) {
561 		RDSV3_DPRINTF2("rdsv3_rdma_prepare",
562 		    "RDS nr_bytes %u remote_bytes %u do not match",
563 		    nr_bytes, (unsigned int) args->remote_vec.bytes);
564 		ret = -EINVAL;
565 		goto out;
566 	}
567 	op->r_bytes = nr_bytes;
568 
569 	ret = 0;
570 out:
571 	if (ret) {
572 		if (op)
573 			rdsv3_rdma_free_op(op);
574 		op = ERR_PTR(ret);
575 	}
576 	return (op);
577 }
578 
579 #define	CEIL(x, y)	(((x) + (y) - 1) / (y))
580 
581 /*
582  * The application asks for a RDMA transfer.
583  * Extract all arguments and set up the rdma_op
584  */
585 int
586 rdsv3_cmsg_rdma_args(struct rdsv3_sock *rs, struct rdsv3_message *rm,
587     struct cmsghdr *cmsg)
588 {
589 	struct rdsv3_rdma_op *op;
590 	/* uint64_t alignment on the buffer */
591 	uint64_t buf[CEIL(CMSG_LEN(sizeof (struct rds_rdma_args)),
592 	    sizeof (uint64_t))];
593 
594 	if (cmsg->cmsg_len != CMSG_LEN(sizeof (struct rds_rdma_args)) ||
595 	    rm->m_rdma_op != NULL)
596 		return (-EINVAL);
597 
598 	ASSERT(sizeof (buf) >= cmsg->cmsg_len && ((uintptr_t)buf & 0x7) == 0);
599 
600 	bcopy(CMSG_DATA(cmsg), (char *)buf, cmsg->cmsg_len);
601 	op = rdsv3_rdma_prepare(rs, (struct rds_rdma_args *)buf);
602 
603 	if (IS_ERR(op))
604 		return (PTR_ERR(op));
605 	rdsv3_stats_inc(s_send_rdma);
606 	rm->m_rdma_op = op;
607 	return (0);
608 }
609 
610 /*
611  * The application wants us to pass an RDMA destination (aka MR)
612  * to the remote
613  */
614 int
615 rdsv3_cmsg_rdma_dest(struct rdsv3_sock *rs, struct rdsv3_message *rm,
616     struct cmsghdr *cmsg)
617 {
618 	struct rdsv3_mr *mr;
619 	uint32_t r_key;
620 	int err = 0;
621 
622 	if (cmsg->cmsg_len != CMSG_LEN(sizeof (rds_rdma_cookie_t)) ||
623 	    rm->m_rdma_cookie != 0)
624 		return (-EINVAL);
625 
626 	(void) memcpy(&rm->m_rdma_cookie, CMSG_DATA(cmsg),
627 	    sizeof (rm->m_rdma_cookie));
628 
629 	/*
630 	 * We are reusing a previously mapped MR here. Most likely, the
631 	 * application has written to the buffer, so we need to explicitly
632 	 * flush those writes to RAM. Otherwise the HCA may not see them
633 	 * when doing a DMA from that buffer.
634 	 */
635 	r_key = rdsv3_rdma_cookie_key(rm->m_rdma_cookie);
636 
637 	mutex_enter(&rs->rs_rdma_lock);
638 	mr = rdsv3_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL);
639 	if (!mr)
640 		err = -EINVAL;	/* invalid r_key */
641 	else
642 		atomic_inc_32(&mr->r_refcount);
643 	mutex_exit(&rs->rs_rdma_lock);
644 
645 	if (mr) {
646 		mr->r_trans->sync_mr(mr->r_trans_private, DMA_TO_DEVICE);
647 		rm->m_rdma_mr = mr;
648 	}
649 	return (err);
650 }
651 
652 /*
653  * The application passes us an address range it wants to enable RDMA
654  * to/from. We map the area, and save the <R_Key,offset> pair
655  * in rm->m_rdma_cookie. This causes it to be sent along to the peer
656  * in an extension header.
657  */
658 int
659 rdsv3_cmsg_rdma_map(struct rdsv3_sock *rs, struct rdsv3_message *rm,
660     struct cmsghdr *cmsg)
661 {
662 	/* uint64_t alignment on the buffer */
663 	uint64_t buf[CEIL(CMSG_LEN(sizeof (struct rds_get_mr_args)),
664 	    sizeof (uint64_t))];
665 	int status;
666 
667 	if (cmsg->cmsg_len != CMSG_LEN(sizeof (struct rds_get_mr_args)) ||
668 	    rm->m_rdma_cookie != 0)
669 		return (-EINVAL);
670 
671 	ASSERT(sizeof (buf) >= cmsg->cmsg_len && ((uintptr_t)buf & 0x7) == 0);
672 
673 	bcopy(CMSG_DATA(cmsg), (char *)buf, cmsg->cmsg_len);
674 	status = __rdsv3_rdma_map(rs, (struct rds_get_mr_args *)buf,
675 	    &rm->m_rdma_cookie, &rm->m_rdma_mr);
676 
677 	return (status);
678 }
679