1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
23  */
24 
25 /*
26  * Copyright (c) 2007 Oracle.  All rights reserved.
27  *
28  * This software is available to you under a choice of one of two
29  * licenses.  You may choose to be licensed under the terms of the GNU
30  * General Public License (GPL) Version 2, available from the file
31  * COPYING in the main directory of this source tree, or the
32  * OpenIB.org BSD license below:
33  *
34  *     Redistribution and use in source and binary forms, with or
35  *     without modification, are permitted provided that the following
36  *     conditions are met:
37  *
38  *      - Redistributions of source code must retain the above
39  *        copyright notice, this list of conditions and the following
40  *        disclaimer.
41  *
42  *      - Redistributions in binary form must reproduce the above
43  *        copyright notice, this list of conditions and the following
44  *        disclaimer in the documentation and/or other materials
45  *        provided with the distribution.
46  *
47  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
48  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
49  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
50  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
51  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
52  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
53  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
54  * SOFTWARE.
55  *
56  */
57 #include <sys/ib/clients/of/rdma/ib_verbs.h>
58 #include <sys/ib/clients/of/rdma/ib_addr.h>
59 #include <sys/ib/clients/of/rdma/rdma_cm.h>
60 
61 #include <sys/ib/clients/rdsv3/ib.h>
62 #include <sys/ib/clients/rdsv3/rdma.h>
63 #include <sys/ib/clients/rdsv3/rdsv3_debug.h>
64 
65 #define	DMA_TO_DEVICE 0
66 #define	DMA_FROM_DEVICE 1
67 #define	RB_CLEAR_NODE(nodep) AVL_SETPARENT(nodep, nodep);
68 
69 /*
70  * XXX
71  *  - build with sparse
72  *  - should we limit the size of a mr region?  let transport return failure?
73  *  - should we detect duplicate keys on a socket?  hmm.
74  *  - an rdma is an mlock, apply rlimit?
75  */
76 
77 /*
78  * get the number of pages by looking at the page indices that the start and
79  * end addresses fall in.
80  *
81  * Returns 0 if the vec is invalid.  It is invalid if the number of bytes
82  * causes the address to wrap or overflows an unsigned int.  This comes
83  * from being stored in the 'length' member of 'struct rdsv3_scatterlist'.
84  */
85 static unsigned int
86 rdsv3_pages_in_vec(struct rdsv3_iovec *vec)
87 {
88 	if ((vec->addr + vec->bytes <= vec->addr) ||
89 	    (vec->bytes > (uint64_t)UINT_MAX)) {
90 		return (0);
91 	}
92 
93 	return (((vec->addr + vec->bytes + PAGESIZE - 1) >>
94 	    PAGESHIFT) - (vec->addr >> PAGESHIFT));
95 }
96 
97 static struct rdsv3_mr *
98 rdsv3_mr_tree_walk(struct avl_tree *root, uint32_t key,
99 	struct rdsv3_mr *insert)
100 {
101 	struct rdsv3_mr *mr;
102 	avl_index_t where;
103 
104 	mr = avl_find(root, &key, &where);
105 	if ((mr == NULL) && (insert != NULL)) {
106 		avl_insert(root, (void *)insert, where);
107 		atomic_add_32(&insert->r_refcount, 1);
108 		return (NULL);
109 	}
110 
111 	return (mr);
112 }
113 
114 /*
115  * Destroy the transport-specific part of a MR.
116  */
117 static void
118 rdsv3_destroy_mr(struct rdsv3_mr *mr)
119 {
120 	struct rdsv3_sock *rs = mr->r_sock;
121 	void *trans_private = NULL;
122 	avl_node_t *np;
123 
124 	RDSV3_DPRINTF5("rdsv3_destroy_mr",
125 	    "RDS: destroy mr key is %x refcnt %u",
126 	    mr->r_key, atomic_get(&mr->r_refcount));
127 
128 	if (test_and_set_bit(RDSV3_MR_DEAD, &mr->r_state))
129 		return;
130 
131 	mutex_enter(&rs->rs_rdma_lock);
132 	np = &mr->r_rb_node;
133 	if (AVL_XPARENT(np) != np)
134 		avl_remove(&rs->rs_rdma_keys, mr);
135 	trans_private = mr->r_trans_private;
136 	mr->r_trans_private = NULL;
137 	mutex_exit(&rs->rs_rdma_lock);
138 
139 	if (trans_private)
140 		mr->r_trans->free_mr(trans_private, mr->r_invalidate);
141 }
142 
143 void
144 __rdsv3_put_mr_final(struct rdsv3_mr *mr)
145 {
146 	rdsv3_destroy_mr(mr);
147 	kmem_free(mr, sizeof (*mr));
148 }
149 
150 /*
151  * By the time this is called we can't have any more ioctls called on
152  * the socket so we don't need to worry about racing with others.
153  */
154 void
155 rdsv3_rdma_drop_keys(struct rdsv3_sock *rs)
156 {
157 	struct rdsv3_mr *mr;
158 	struct avl_node *node;
159 
160 	/* Release any MRs associated with this socket */
161 	mutex_enter(&rs->rs_rdma_lock);
162 	while ((node = avl_first(&rs->rs_rdma_keys))) {
163 		mr = container_of(node, struct rdsv3_mr, r_rb_node);
164 		if (mr->r_trans == rs->rs_transport)
165 			mr->r_invalidate = 0;
166 		avl_remove(&rs->rs_rdma_keys, &mr->r_rb_node);
167 		RB_CLEAR_NODE(&mr->r_rb_node)
168 		mutex_exit(&rs->rs_rdma_lock);
169 		rdsv3_destroy_mr(mr);
170 		rdsv3_mr_put(mr);
171 		mutex_enter(&rs->rs_rdma_lock);
172 	}
173 	mutex_exit(&rs->rs_rdma_lock);
174 
175 	if (rs->rs_transport && rs->rs_transport->flush_mrs)
176 		rs->rs_transport->flush_mrs();
177 }
178 
179 static int
180 __rdsv3_rdma_map(struct rdsv3_sock *rs, struct rdsv3_get_mr_args *args,
181 	uint64_t *cookie_ret, struct rdsv3_mr **mr_ret)
182 {
183 	struct rdsv3_mr *mr = NULL, *found;
184 	void *trans_private;
185 	rdsv3_rdma_cookie_t cookie;
186 	unsigned int nents = 0;
187 	int ret;
188 
189 	if (rs->rs_bound_addr == 0) {
190 		ret = -ENOTCONN; /* XXX not a great errno */
191 		goto out;
192 	}
193 
194 	if (!rs->rs_transport->get_mr) {
195 		ret = -EOPNOTSUPP;
196 		goto out;
197 	}
198 
199 	mr = kmem_zalloc(sizeof (struct rdsv3_mr), KM_NOSLEEP);
200 	if (!mr) {
201 		ret = -ENOMEM;
202 		goto out;
203 	}
204 
205 	mr->r_refcount = 1;
206 	RB_CLEAR_NODE(&mr->r_rb_node);
207 	mr->r_trans = rs->rs_transport;
208 	mr->r_sock = rs;
209 
210 	if (args->flags & RDSV3_RDMA_USE_ONCE)
211 		mr->r_use_once = 1;
212 	if (args->flags & RDSV3_RDMA_INVALIDATE)
213 		mr->r_invalidate = 1;
214 	if (args->flags & RDSV3_RDMA_READWRITE)
215 		mr->r_write = 1;
216 
217 	/*
218 	 * Obtain a transport specific MR. If this succeeds, the
219 	 * s/g list is now owned by the MR.
220 	 * Note that dma_map() implies that pending writes are
221 	 * flushed to RAM, so no dma_sync is needed here.
222 	 */
223 	trans_private = rs->rs_transport->get_mr(&args->vec, nents, rs,
224 	    &mr->r_key);
225 
226 	if (IS_ERR(trans_private)) {
227 		ret = PTR_ERR(trans_private);
228 		goto out;
229 	}
230 
231 	mr->r_trans_private = trans_private;
232 
233 	/*
234 	 * The user may pass us an unaligned address, but we can only
235 	 * map page aligned regions. So we keep the offset, and build
236 	 * a 64bit cookie containing <R_Key, offset> and pass that
237 	 * around.
238 	 */
239 	cookie = rdsv3_rdma_make_cookie(mr->r_key, args->vec.addr & ~PAGEMASK);
240 	if (cookie_ret)
241 		*cookie_ret = cookie;
242 
243 	/*
244 	 * copy value of cookie to user address at args->cookie_addr
245 	 */
246 	if (args->cookie_addr) {
247 		ret = ddi_copyout((void *)&cookie,
248 		    (void *)((intptr_t)args->cookie_addr),
249 		    sizeof (rdsv3_rdma_cookie_t), 0);
250 		if (ret != 0) {
251 			ret = -EFAULT;
252 			goto out;
253 		}
254 	}
255 
256 	RDSV3_DPRINTF5("__rdsv3_rdma_map",
257 	    "RDS: get_mr mr 0x%p addr 0x%llx key 0x%x",
258 	    mr, args->vec.addr, mr->r_key);
259 	/*
260 	 * Inserting the new MR into the rbtree bumps its
261 	 * reference count.
262 	 */
263 	mutex_enter(&rs->rs_rdma_lock);
264 	found = rdsv3_mr_tree_walk(&rs->rs_rdma_keys, mr->r_key, mr);
265 	mutex_exit(&rs->rs_rdma_lock);
266 
267 	ASSERT(!(found && found != mr));
268 
269 	if (mr_ret) {
270 		atomic_add_32(&mr->r_refcount, 1);
271 		*mr_ret = mr;
272 	}
273 
274 	ret = 0;
275 out:
276 	if (mr)
277 		rdsv3_mr_put(mr);
278 	return (ret);
279 }
280 
281 int
282 rdsv3_get_mr(struct rdsv3_sock *rs, const void *optval, int optlen)
283 {
284 	struct rdsv3_get_mr_args args;
285 
286 	if (optlen != sizeof (struct rdsv3_get_mr_args))
287 		return (-EINVAL);
288 
289 #if 1
290 	bcopy((struct rdsv3_get_mr_args *)optval, &args,
291 	    sizeof (struct rdsv3_get_mr_args));
292 #else
293 	if (ddi_copyin(optval, &args, optlen, 0))
294 		return (-EFAULT);
295 #endif
296 
297 	return (__rdsv3_rdma_map(rs, &args, NULL, NULL));
298 }
299 
300 int
301 rdsv3_get_mr_for_dest(struct rdsv3_sock *rs, const void *optval,
302     int optlen)
303 {
304 	struct rdsv3_get_mr_for_dest_args args;
305 	struct rdsv3_get_mr_args new_args;
306 
307 	if (optlen != sizeof (struct rdsv3_get_mr_for_dest_args))
308 		return (-EINVAL);
309 
310 #if 1
311 	bcopy((struct rdsv3_get_mr_for_dest_args *)optval, &args,
312 	    sizeof (struct rdsv3_get_mr_for_dest_args));
313 #else
314 	if (ddi_copyin(optval, &args, optlen, 0))
315 		return (-EFAULT);
316 #endif
317 
318 	/*
319 	 * Initially, just behave like get_mr().
320 	 * TODO: Implement get_mr as wrapper around this
321 	 *	 and deprecate it.
322 	 */
323 	new_args.vec = args.vec;
324 	new_args.cookie_addr = args.cookie_addr;
325 	new_args.flags = args.flags;
326 
327 	return (__rdsv3_rdma_map(rs, &new_args, NULL, NULL));
328 }
329 
330 /*
331  * Free the MR indicated by the given R_Key
332  */
333 int
334 rdsv3_free_mr(struct rdsv3_sock *rs, const void *optval, int optlen)
335 {
336 	struct rdsv3_free_mr_args args;
337 	struct rdsv3_mr *mr;
338 
339 	if (optlen != sizeof (struct rdsv3_free_mr_args))
340 		return (-EINVAL);
341 
342 #if 1
343 	bcopy((struct rdsv3_free_mr_args *)optval, &args,
344 	    sizeof (struct rdsv3_free_mr_args));
345 #else
346 	if (ddi_copyin((struct rdsv3_free_mr_args *)optval, &args,
347 	    sizeof (struct rdsv3_free_mr_args), 0))
348 		return (-EFAULT);
349 #endif
350 
351 	/* Special case - a null cookie means flush all unused MRs */
352 	if (args.cookie == 0) {
353 		if (!rs->rs_transport || !rs->rs_transport->flush_mrs)
354 			return (-EINVAL);
355 		rs->rs_transport->flush_mrs();
356 		return (0);
357 	}
358 
359 	/*
360 	 * Look up the MR given its R_key and remove it from the rbtree
361 	 * so nobody else finds it.
362 	 * This should also prevent races with rdsv3_rdma_unuse.
363 	 */
364 	mutex_enter(&rs->rs_rdma_lock);
365 	mr = rdsv3_mr_tree_walk(&rs->rs_rdma_keys,
366 	    rdsv3_rdma_cookie_key(args.cookie), NULL);
367 	if (mr) {
368 		avl_remove(&rs->rs_rdma_keys, &mr->r_rb_node);
369 		RB_CLEAR_NODE(&mr->r_rb_node);
370 		if (args.flags & RDSV3_RDMA_INVALIDATE)
371 			mr->r_invalidate = 1;
372 	}
373 	mutex_exit(&rs->rs_rdma_lock);
374 
375 	if (!mr)
376 		return (-EINVAL);
377 
378 	/*
379 	 * call rdsv3_destroy_mr() ourselves so that we're sure it's done
380 	 * by time we return.  If we let rdsv3_mr_put() do it it might not
381 	 * happen until someone else drops their ref.
382 	 */
383 	rdsv3_destroy_mr(mr);
384 	rdsv3_mr_put(mr);
385 	return (0);
386 }
387 
388 /*
389  * This is called when we receive an extension header that
390  * tells us this MR was used. It allows us to implement
391  * use_once semantics
392  */
393 void
394 rdsv3_rdma_unuse(struct rdsv3_sock *rs, uint32_t r_key, int force)
395 {
396 	struct rdsv3_mr *mr;
397 	int zot_me = 0;
398 
399 	RDSV3_DPRINTF4("rdsv3_rdma_unuse", "Enter rkey: 0x%x", r_key);
400 
401 	mutex_enter(&rs->rs_rdma_lock);
402 	mr = rdsv3_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL);
403 	if (!mr) {
404 		RDSV3_DPRINTF4("rdsv3_rdma_unuse",
405 		    "rdsv3: trying to unuse MR with unknown r_key %u!", r_key);
406 		mutex_exit(&rs->rs_rdma_lock);
407 		return;
408 	}
409 
410 	if (mr->r_use_once || force) {
411 		avl_remove(&rs->rs_rdma_keys, &mr->r_rb_node);
412 		RB_CLEAR_NODE(&mr->r_rb_node);
413 		zot_me = 1;
414 	}
415 	mutex_exit(&rs->rs_rdma_lock);
416 
417 	/*
418 	 * May have to issue a dma_sync on this memory region.
419 	 * Note we could avoid this if the operation was a RDMA READ,
420 	 * but at this point we can't tell.
421 	 */
422 	if (mr->r_trans->sync_mr)
423 		mr->r_trans->sync_mr(mr->r_trans_private, DMA_FROM_DEVICE);
424 
425 	/*
426 	 * If the MR was marked as invalidate, this will
427 	 * trigger an async flush.
428 	 */
429 	if (zot_me)
430 		rdsv3_destroy_mr(mr);
431 	rdsv3_mr_put(mr);
432 	RDSV3_DPRINTF4("rdsv3_rdma_unuse", "Return");
433 }
434 
435 void
436 rdsv3_rdma_free_op(struct rdsv3_rdma_op *ro)
437 {
438 	unsigned int i;
439 
440 	/* deallocate RDMA resources on rdsv3_message */
441 	for (i = 0; i < ro->r_nents; i++) {
442 		ddi_umem_unlock(ro->r_rdma_sg[i].umem_cookie);
443 	}
444 
445 	if (ro->r_notifier)
446 		kmem_free(ro->r_notifier, sizeof (*ro->r_notifier));
447 	kmem_free(ro, sizeof (*ro));
448 }
449 
450 /*
451  * args is a pointer to an in-kernel copy in the sendmsg cmsg.
452  */
453 static struct rdsv3_rdma_op *
454 rdsv3_rdma_prepare(struct rdsv3_sock *rs, struct rdsv3_rdma_args *args)
455 {
456 	struct rdsv3_iovec vec;
457 	struct rdsv3_rdma_op *op = NULL;
458 	unsigned int nr_bytes;
459 	struct rdsv3_iovec *local_vec;
460 	unsigned int nr;
461 	unsigned int i;
462 	ddi_umem_cookie_t umem_cookie;
463 	size_t umem_len;
464 	caddr_t umem_addr;
465 	int ret;
466 
467 	if (rs->rs_bound_addr == 0) {
468 		ret = -ENOTCONN; /* XXX not a great errno */
469 		goto out;
470 	}
471 
472 	if (args->nr_local > (uint64_t)UINT_MAX) {
473 		ret = -EMSGSIZE;
474 		goto out;
475 	}
476 
477 	op = kmem_zalloc(offsetof(struct rdsv3_rdma_op,
478 	    r_rdma_sg[args->nr_local]), KM_NOSLEEP);
479 	if (op == NULL) {
480 		ret = -ENOMEM;
481 		goto out;
482 	}
483 
484 	op->r_write = !!(args->flags & RDSV3_RDMA_READWRITE);
485 	op->r_fence = !!(args->flags & RDSV3_RDMA_FENCE);
486 	op->r_notify = !!(args->flags & RDSV3_RDMA_NOTIFY_ME);
487 	op->r_recverr = rs->rs_recverr;
488 
489 	if (op->r_notify || op->r_recverr) {
490 		/*
491 		 * We allocate an uninitialized notifier here, because
492 		 * we don't want to do that in the completion handler. We
493 		 * would have to use GFP_ATOMIC there, and don't want to deal
494 		 * with failed allocations.
495 		 */
496 		op->r_notifier = kmem_alloc(sizeof (struct rdsv3_notifier),
497 		    KM_NOSLEEP);
498 		if (!op->r_notifier) {
499 			ret = -ENOMEM;
500 			goto out;
501 		}
502 		op->r_notifier->n_user_token = args->user_token;
503 		op->r_notifier->n_status = RDSV3_RDMA_SUCCESS;
504 	}
505 
506 	/*
507 	 * The cookie contains the R_Key of the remote memory region, and
508 	 * optionally an offset into it. This is how we implement RDMA into
509 	 * unaligned memory.
510 	 * When setting up the RDMA, we need to add that offset to the
511 	 * destination address (which is really an offset into the MR)
512 	 * FIXME: We may want to move this into ib_rdma.c
513 	 */
514 	op->r_key = rdsv3_rdma_cookie_key(args->cookie);
515 	op->r_remote_addr = args->remote_vec.addr +
516 	    rdsv3_rdma_cookie_offset(args->cookie);
517 
518 	nr_bytes = 0;
519 
520 	RDSV3_DPRINTF5("rdsv3_rdma_prepare",
521 	    "RDS: rdma prepare nr_local %llu rva %llx rkey %x",
522 	    (unsigned long long)args->nr_local,
523 	    (unsigned long long)args->remote_vec.addr,
524 	    op->r_key);
525 
526 	local_vec = (struct rdsv3_iovec *)(unsigned long) args->local_vec_addr;
527 
528 	/* pin the scatter list of user buffers */
529 	for (i = 0; i < args->nr_local; i++) {
530 		if (ddi_copyin(&local_vec[i], &vec,
531 		    sizeof (struct rdsv3_iovec), 0)) {
532 			ret = -EFAULT;
533 			goto out;
534 		}
535 
536 		nr = rdsv3_pages_in_vec(&vec);
537 		if (nr == 0) {
538 			RDSV3_DPRINTF2("rdsv3_rdma_prepare",
539 			    "rdsv3_pages_in_vec returned 0");
540 			ret = -EINVAL;
541 			goto out;
542 		}
543 
544 		rs->rs_user_addr = vec.addr;
545 		rs->rs_user_bytes = vec.bytes;
546 
547 		/* pin user memory pages */
548 		umem_len = ptob(btopr(vec.bytes +
549 		    ((uintptr_t)vec.addr & PAGEOFFSET)));
550 		umem_addr = (caddr_t)((uintptr_t)vec.addr & ~PAGEOFFSET);
551 		ret = umem_lockmemory(umem_addr, umem_len,
552 		    DDI_UMEMLOCK_WRITE | DDI_UMEMLOCK_READ,
553 		    &umem_cookie, NULL, NULL);
554 		if (ret != 0) {
555 			RDSV3_DPRINTF2("rdsv3_rdma_prepare",
556 			    "umem_lockmemory() returned %d", ret);
557 			ret = -EFAULT;
558 			goto out;
559 		}
560 		op->r_rdma_sg[i].umem_cookie = umem_cookie;
561 		op->r_rdma_sg[i].iovec = vec;
562 		nr_bytes += vec.bytes;
563 
564 		RDSV3_DPRINTF5("rdsv3_rdma_prepare",
565 		    "RDS: nr_bytes %u nr %u vec.bytes %llu vec.addr %llx",
566 		    nr_bytes, nr, vec.bytes, vec.addr);
567 	}
568 	op->r_nents = i;
569 
570 	if (nr_bytes > args->remote_vec.bytes) {
571 		RDSV3_DPRINTF2("rdsv3_rdma_prepare",
572 		    "RDS nr_bytes %u remote_bytes %u do not match",
573 		    nr_bytes, (unsigned int) args->remote_vec.bytes);
574 		ret = -EINVAL;
575 		goto out;
576 	}
577 	op->r_bytes = nr_bytes;
578 
579 	ret = 0;
580 out:
581 	if (ret) {
582 		if (op)
583 			rdsv3_rdma_free_op(op);
584 		op = ERR_PTR(ret);
585 	}
586 	return (op);
587 }
588 
589 /*
590  * The application asks for a RDMA transfer.
591  * Extract all arguments and set up the rdma_op
592  */
593 int
594 rdsv3_cmsg_rdma_args(struct rdsv3_sock *rs, struct rdsv3_message *rm,
595 	struct cmsghdr *cmsg)
596 {
597 	struct rdsv3_rdma_op *op;
598 	struct rdsv3_rdma_args *ap;
599 
600 	if (cmsg->cmsg_len < CMSG_LEN(sizeof (struct rdsv3_rdma_args)) ||
601 	    rm->m_rdma_op != NULL)
602 		return (-EINVAL);
603 
604 	/* uint64_t alignment on struct rdsv3_get_mr_args */
605 	ap = (struct rdsv3_rdma_args *)kmem_alloc(cmsg->cmsg_len, KM_SLEEP);
606 	bcopy(CMSG_DATA(cmsg), ap, cmsg->cmsg_len);
607 	op = rdsv3_rdma_prepare(rs, ap);
608 	kmem_free(ap, cmsg->cmsg_len);
609 	if (IS_ERR(op))
610 		return (PTR_ERR(op));
611 	rdsv3_stats_inc(s_send_rdma);
612 	rm->m_rdma_op = op;
613 	return (0);
614 }
615 
616 /*
617  * The application wants us to pass an RDMA destination (aka MR)
618  * to the remote
619  */
620 int
621 rdsv3_cmsg_rdma_dest(struct rdsv3_sock *rs, struct rdsv3_message *rm,
622 	struct cmsghdr *cmsg)
623 {
624 	struct rdsv3_mr *mr;
625 	uint32_t r_key;
626 	int err = 0;
627 
628 	if (cmsg->cmsg_len < CMSG_LEN(sizeof (rdsv3_rdma_cookie_t)) ||
629 	    rm->m_rdma_cookie != 0)
630 		return (-EINVAL);
631 
632 	(void) memcpy(&rm->m_rdma_cookie, CMSG_DATA(cmsg),
633 	    sizeof (rm->m_rdma_cookie));
634 
635 	/*
636 	 * We are reusing a previously mapped MR here. Most likely, the
637 	 * application has written to the buffer, so we need to explicitly
638 	 * flush those writes to RAM. Otherwise the HCA may not see them
639 	 * when doing a DMA from that buffer.
640 	 */
641 	r_key = rdsv3_rdma_cookie_key(rm->m_rdma_cookie);
642 
643 	mutex_enter(&rs->rs_rdma_lock);
644 	mr = rdsv3_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL);
645 	if (!mr)
646 		err = -EINVAL;	/* invalid r_key */
647 	else
648 		atomic_add_32(&mr->r_refcount, 1);
649 	mutex_exit(&rs->rs_rdma_lock);
650 
651 	if (mr) {
652 		mr->r_trans->sync_mr(mr->r_trans_private, DMA_TO_DEVICE);
653 		rm->m_rdma_mr = mr;
654 	}
655 	return (err);
656 }
657 
658 /*
659  * The application passes us an address range it wants to enable RDMA
660  * to/from. We map the area, and save the <R_Key,offset> pair
661  * in rm->m_rdma_cookie. This causes it to be sent along to the peer
662  * in an extension header.
663  */
664 int
665 rdsv3_cmsg_rdma_map(struct rdsv3_sock *rs, struct rdsv3_message *rm,
666 	struct cmsghdr *cmsg)
667 {
668 	struct rdsv3_get_mr_args *mrp;
669 	int status;
670 
671 	if (cmsg->cmsg_len < CMSG_LEN(sizeof (struct rdsv3_get_mr_args)) ||
672 	    rm->m_rdma_cookie != 0)
673 		return (-EINVAL);
674 
675 	/* uint64_t alignment on struct rdsv3_get_mr_args */
676 	mrp = (struct rdsv3_get_mr_args *)kmem_alloc(cmsg->cmsg_len, KM_SLEEP);
677 	bcopy(CMSG_DATA(cmsg), mrp, cmsg->cmsg_len);
678 	status = __rdsv3_rdma_map(rs, mrp, &rm->m_rdma_cookie, &rm->m_rdma_mr);
679 	kmem_free(mrp, cmsg->cmsg_len);
680 	return (status);
681 }
682