xref: /openbmc/linux/net/rds/rdma.c (revision 827634ad)
1 /*
2  * Copyright (c) 2007 Oracle.  All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  *
32  */
33 #include <linux/pagemap.h>
34 #include <linux/slab.h>
35 #include <linux/rbtree.h>
36 #include <linux/dma-mapping.h> /* for DMA_*_DEVICE */
37 
38 #include "rds.h"
39 
40 /*
41  * XXX
42  *  - build with sparse
43  *  - should we limit the size of a mr region?  let transport return failure?
44  *  - should we detect duplicate keys on a socket?  hmm.
45  *  - an rdma is an mlock, apply rlimit?
46  */
47 
48 /*
49  * get the number of pages by looking at the page indices that the start and
50  * end addresses fall in.
51  *
52  * Returns 0 if the vec is invalid.  It is invalid if the number of bytes
53  * causes the address to wrap or overflows an unsigned int.  This comes
54  * from being stored in the 'length' member of 'struct scatterlist'.
55  */
56 static unsigned int rds_pages_in_vec(struct rds_iovec *vec)
57 {
58 	if ((vec->addr + vec->bytes <= vec->addr) ||
59 	    (vec->bytes > (u64)UINT_MAX))
60 		return 0;
61 
62 	return ((vec->addr + vec->bytes + PAGE_SIZE - 1) >> PAGE_SHIFT) -
63 		(vec->addr >> PAGE_SHIFT);
64 }
65 
66 static struct rds_mr *rds_mr_tree_walk(struct rb_root *root, u64 key,
67 				       struct rds_mr *insert)
68 {
69 	struct rb_node **p = &root->rb_node;
70 	struct rb_node *parent = NULL;
71 	struct rds_mr *mr;
72 
73 	while (*p) {
74 		parent = *p;
75 		mr = rb_entry(parent, struct rds_mr, r_rb_node);
76 
77 		if (key < mr->r_key)
78 			p = &(*p)->rb_left;
79 		else if (key > mr->r_key)
80 			p = &(*p)->rb_right;
81 		else
82 			return mr;
83 	}
84 
85 	if (insert) {
86 		rb_link_node(&insert->r_rb_node, parent, p);
87 		rb_insert_color(&insert->r_rb_node, root);
88 		atomic_inc(&insert->r_refcount);
89 	}
90 	return NULL;
91 }
92 
93 /*
94  * Destroy the transport-specific part of a MR.
95  */
96 static void rds_destroy_mr(struct rds_mr *mr)
97 {
98 	struct rds_sock *rs = mr->r_sock;
99 	void *trans_private = NULL;
100 	unsigned long flags;
101 
102 	rdsdebug("RDS: destroy mr key is %x refcnt %u\n",
103 			mr->r_key, atomic_read(&mr->r_refcount));
104 
105 	if (test_and_set_bit(RDS_MR_DEAD, &mr->r_state))
106 		return;
107 
108 	spin_lock_irqsave(&rs->rs_rdma_lock, flags);
109 	if (!RB_EMPTY_NODE(&mr->r_rb_node))
110 		rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
111 	trans_private = mr->r_trans_private;
112 	mr->r_trans_private = NULL;
113 	spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
114 
115 	if (trans_private)
116 		mr->r_trans->free_mr(trans_private, mr->r_invalidate);
117 }
118 
119 void __rds_put_mr_final(struct rds_mr *mr)
120 {
121 	rds_destroy_mr(mr);
122 	kfree(mr);
123 }
124 
125 /*
126  * By the time this is called we can't have any more ioctls called on
127  * the socket so we don't need to worry about racing with others.
128  */
129 void rds_rdma_drop_keys(struct rds_sock *rs)
130 {
131 	struct rds_mr *mr;
132 	struct rb_node *node;
133 	unsigned long flags;
134 
135 	/* Release any MRs associated with this socket */
136 	spin_lock_irqsave(&rs->rs_rdma_lock, flags);
137 	while ((node = rb_first(&rs->rs_rdma_keys))) {
138 		mr = container_of(node, struct rds_mr, r_rb_node);
139 		if (mr->r_trans == rs->rs_transport)
140 			mr->r_invalidate = 0;
141 		rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
142 		RB_CLEAR_NODE(&mr->r_rb_node);
143 		spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
144 		rds_destroy_mr(mr);
145 		rds_mr_put(mr);
146 		spin_lock_irqsave(&rs->rs_rdma_lock, flags);
147 	}
148 	spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
149 
150 	if (rs->rs_transport && rs->rs_transport->flush_mrs)
151 		rs->rs_transport->flush_mrs();
152 }
153 
154 /*
155  * Helper function to pin user pages.
156  */
157 static int rds_pin_pages(unsigned long user_addr, unsigned int nr_pages,
158 			struct page **pages, int write)
159 {
160 	int ret;
161 
162 	ret = get_user_pages_fast(user_addr, nr_pages, write, pages);
163 
164 	if (ret >= 0 && ret < nr_pages) {
165 		while (ret--)
166 			put_page(pages[ret]);
167 		ret = -EFAULT;
168 	}
169 
170 	return ret;
171 }
172 
173 static int __rds_rdma_map(struct rds_sock *rs, struct rds_get_mr_args *args,
174 				u64 *cookie_ret, struct rds_mr **mr_ret)
175 {
176 	struct rds_mr *mr = NULL, *found;
177 	unsigned int nr_pages;
178 	struct page **pages = NULL;
179 	struct scatterlist *sg;
180 	void *trans_private;
181 	unsigned long flags;
182 	rds_rdma_cookie_t cookie;
183 	unsigned int nents;
184 	long i;
185 	int ret;
186 
187 	if (rs->rs_bound_addr == 0) {
188 		ret = -ENOTCONN; /* XXX not a great errno */
189 		goto out;
190 	}
191 
192 	if (!rs->rs_transport->get_mr) {
193 		ret = -EOPNOTSUPP;
194 		goto out;
195 	}
196 
197 	nr_pages = rds_pages_in_vec(&args->vec);
198 	if (nr_pages == 0) {
199 		ret = -EINVAL;
200 		goto out;
201 	}
202 
203 	rdsdebug("RDS: get_mr addr %llx len %llu nr_pages %u\n",
204 		args->vec.addr, args->vec.bytes, nr_pages);
205 
206 	/* XXX clamp nr_pages to limit the size of this alloc? */
207 	pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
208 	if (!pages) {
209 		ret = -ENOMEM;
210 		goto out;
211 	}
212 
213 	mr = kzalloc(sizeof(struct rds_mr), GFP_KERNEL);
214 	if (!mr) {
215 		ret = -ENOMEM;
216 		goto out;
217 	}
218 
219 	atomic_set(&mr->r_refcount, 1);
220 	RB_CLEAR_NODE(&mr->r_rb_node);
221 	mr->r_trans = rs->rs_transport;
222 	mr->r_sock = rs;
223 
224 	if (args->flags & RDS_RDMA_USE_ONCE)
225 		mr->r_use_once = 1;
226 	if (args->flags & RDS_RDMA_INVALIDATE)
227 		mr->r_invalidate = 1;
228 	if (args->flags & RDS_RDMA_READWRITE)
229 		mr->r_write = 1;
230 
231 	/*
232 	 * Pin the pages that make up the user buffer and transfer the page
233 	 * pointers to the mr's sg array.  We check to see if we've mapped
234 	 * the whole region after transferring the partial page references
235 	 * to the sg array so that we can have one page ref cleanup path.
236 	 *
237 	 * For now we have no flag that tells us whether the mapping is
238 	 * r/o or r/w. We need to assume r/w, or we'll do a lot of RDMA to
239 	 * the zero page.
240 	 */
241 	ret = rds_pin_pages(args->vec.addr, nr_pages, pages, 1);
242 	if (ret < 0)
243 		goto out;
244 
245 	nents = ret;
246 	sg = kcalloc(nents, sizeof(*sg), GFP_KERNEL);
247 	if (!sg) {
248 		ret = -ENOMEM;
249 		goto out;
250 	}
251 	WARN_ON(!nents);
252 	sg_init_table(sg, nents);
253 
254 	/* Stick all pages into the scatterlist */
255 	for (i = 0 ; i < nents; i++)
256 		sg_set_page(&sg[i], pages[i], PAGE_SIZE, 0);
257 
258 	rdsdebug("RDS: trans_private nents is %u\n", nents);
259 
260 	/* Obtain a transport specific MR. If this succeeds, the
261 	 * s/g list is now owned by the MR.
262 	 * Note that dma_map() implies that pending writes are
263 	 * flushed to RAM, so no dma_sync is needed here. */
264 	trans_private = rs->rs_transport->get_mr(sg, nents, rs,
265 						 &mr->r_key);
266 
267 	if (IS_ERR(trans_private)) {
268 		for (i = 0 ; i < nents; i++)
269 			put_page(sg_page(&sg[i]));
270 		kfree(sg);
271 		ret = PTR_ERR(trans_private);
272 		goto out;
273 	}
274 
275 	mr->r_trans_private = trans_private;
276 
277 	rdsdebug("RDS: get_mr put_user key is %x cookie_addr %p\n",
278 	       mr->r_key, (void *)(unsigned long) args->cookie_addr);
279 
280 	/* The user may pass us an unaligned address, but we can only
281 	 * map page aligned regions. So we keep the offset, and build
282 	 * a 64bit cookie containing <R_Key, offset> and pass that
283 	 * around. */
284 	cookie = rds_rdma_make_cookie(mr->r_key, args->vec.addr & ~PAGE_MASK);
285 	if (cookie_ret)
286 		*cookie_ret = cookie;
287 
288 	if (args->cookie_addr && put_user(cookie, (u64 __user *)(unsigned long) args->cookie_addr)) {
289 		ret = -EFAULT;
290 		goto out;
291 	}
292 
293 	/* Inserting the new MR into the rbtree bumps its
294 	 * reference count. */
295 	spin_lock_irqsave(&rs->rs_rdma_lock, flags);
296 	found = rds_mr_tree_walk(&rs->rs_rdma_keys, mr->r_key, mr);
297 	spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
298 
299 	BUG_ON(found && found != mr);
300 
301 	rdsdebug("RDS: get_mr key is %x\n", mr->r_key);
302 	if (mr_ret) {
303 		atomic_inc(&mr->r_refcount);
304 		*mr_ret = mr;
305 	}
306 
307 	ret = 0;
308 out:
309 	kfree(pages);
310 	if (mr)
311 		rds_mr_put(mr);
312 	return ret;
313 }
314 
315 int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen)
316 {
317 	struct rds_get_mr_args args;
318 
319 	if (optlen != sizeof(struct rds_get_mr_args))
320 		return -EINVAL;
321 
322 	if (copy_from_user(&args, (struct rds_get_mr_args __user *)optval,
323 			   sizeof(struct rds_get_mr_args)))
324 		return -EFAULT;
325 
326 	return __rds_rdma_map(rs, &args, NULL, NULL);
327 }
328 
329 int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen)
330 {
331 	struct rds_get_mr_for_dest_args args;
332 	struct rds_get_mr_args new_args;
333 
334 	if (optlen != sizeof(struct rds_get_mr_for_dest_args))
335 		return -EINVAL;
336 
337 	if (copy_from_user(&args, (struct rds_get_mr_for_dest_args __user *)optval,
338 			   sizeof(struct rds_get_mr_for_dest_args)))
339 		return -EFAULT;
340 
341 	/*
342 	 * Initially, just behave like get_mr().
343 	 * TODO: Implement get_mr as wrapper around this
344 	 *	 and deprecate it.
345 	 */
346 	new_args.vec = args.vec;
347 	new_args.cookie_addr = args.cookie_addr;
348 	new_args.flags = args.flags;
349 
350 	return __rds_rdma_map(rs, &new_args, NULL, NULL);
351 }
352 
353 /*
354  * Free the MR indicated by the given R_Key
355  */
356 int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen)
357 {
358 	struct rds_free_mr_args args;
359 	struct rds_mr *mr;
360 	unsigned long flags;
361 
362 	if (optlen != sizeof(struct rds_free_mr_args))
363 		return -EINVAL;
364 
365 	if (copy_from_user(&args, (struct rds_free_mr_args __user *)optval,
366 			   sizeof(struct rds_free_mr_args)))
367 		return -EFAULT;
368 
369 	/* Special case - a null cookie means flush all unused MRs */
370 	if (args.cookie == 0) {
371 		if (!rs->rs_transport || !rs->rs_transport->flush_mrs)
372 			return -EINVAL;
373 		rs->rs_transport->flush_mrs();
374 		return 0;
375 	}
376 
377 	/* Look up the MR given its R_key and remove it from the rbtree
378 	 * so nobody else finds it.
379 	 * This should also prevent races with rds_rdma_unuse.
380 	 */
381 	spin_lock_irqsave(&rs->rs_rdma_lock, flags);
382 	mr = rds_mr_tree_walk(&rs->rs_rdma_keys, rds_rdma_cookie_key(args.cookie), NULL);
383 	if (mr) {
384 		rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
385 		RB_CLEAR_NODE(&mr->r_rb_node);
386 		if (args.flags & RDS_RDMA_INVALIDATE)
387 			mr->r_invalidate = 1;
388 	}
389 	spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
390 
391 	if (!mr)
392 		return -EINVAL;
393 
394 	/*
395 	 * call rds_destroy_mr() ourselves so that we're sure it's done by the time
396 	 * we return.  If we let rds_mr_put() do it it might not happen until
397 	 * someone else drops their ref.
398 	 */
399 	rds_destroy_mr(mr);
400 	rds_mr_put(mr);
401 	return 0;
402 }
403 
404 /*
405  * This is called when we receive an extension header that
406  * tells us this MR was used. It allows us to implement
407  * use_once semantics
408  */
409 void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force)
410 {
411 	struct rds_mr *mr;
412 	unsigned long flags;
413 	int zot_me = 0;
414 
415 	spin_lock_irqsave(&rs->rs_rdma_lock, flags);
416 	mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL);
417 	if (!mr) {
418 		printk(KERN_ERR "rds: trying to unuse MR with unknown r_key %u!\n", r_key);
419 		spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
420 		return;
421 	}
422 
423 	if (mr->r_use_once || force) {
424 		rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
425 		RB_CLEAR_NODE(&mr->r_rb_node);
426 		zot_me = 1;
427 	}
428 	spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
429 
430 	/* May have to issue a dma_sync on this memory region.
431 	 * Note we could avoid this if the operation was a RDMA READ,
432 	 * but at this point we can't tell. */
433 	if (mr->r_trans->sync_mr)
434 		mr->r_trans->sync_mr(mr->r_trans_private, DMA_FROM_DEVICE);
435 
436 	/* If the MR was marked as invalidate, this will
437 	 * trigger an async flush. */
438 	if (zot_me)
439 		rds_destroy_mr(mr);
440 	rds_mr_put(mr);
441 }
442 
443 void rds_rdma_free_op(struct rm_rdma_op *ro)
444 {
445 	unsigned int i;
446 
447 	for (i = 0; i < ro->op_nents; i++) {
448 		struct page *page = sg_page(&ro->op_sg[i]);
449 
450 		/* Mark page dirty if it was possibly modified, which
451 		 * is the case for a RDMA_READ which copies from remote
452 		 * to local memory */
453 		if (!ro->op_write) {
454 			BUG_ON(irqs_disabled());
455 			set_page_dirty(page);
456 		}
457 		put_page(page);
458 	}
459 
460 	kfree(ro->op_notifier);
461 	ro->op_notifier = NULL;
462 	ro->op_active = 0;
463 }
464 
465 void rds_atomic_free_op(struct rm_atomic_op *ao)
466 {
467 	struct page *page = sg_page(ao->op_sg);
468 
469 	/* Mark page dirty if it was possibly modified, which
470 	 * is the case for a RDMA_READ which copies from remote
471 	 * to local memory */
472 	set_page_dirty(page);
473 	put_page(page);
474 
475 	kfree(ao->op_notifier);
476 	ao->op_notifier = NULL;
477 	ao->op_active = 0;
478 }
479 
480 
481 /*
482  * Count the number of pages needed to describe an incoming iovec array.
483  */
484 static int rds_rdma_pages(struct rds_iovec iov[], int nr_iovecs)
485 {
486 	int tot_pages = 0;
487 	unsigned int nr_pages;
488 	unsigned int i;
489 
490 	/* figure out the number of pages in the vector */
491 	for (i = 0; i < nr_iovecs; i++) {
492 		nr_pages = rds_pages_in_vec(&iov[i]);
493 		if (nr_pages == 0)
494 			return -EINVAL;
495 
496 		tot_pages += nr_pages;
497 
498 		/*
499 		 * nr_pages for one entry is limited to (UINT_MAX>>PAGE_SHIFT)+1,
500 		 * so tot_pages cannot overflow without first going negative.
501 		 */
502 		if (tot_pages < 0)
503 			return -EINVAL;
504 	}
505 
506 	return tot_pages;
507 }
508 
509 int rds_rdma_extra_size(struct rds_rdma_args *args)
510 {
511 	struct rds_iovec vec;
512 	struct rds_iovec __user *local_vec;
513 	int tot_pages = 0;
514 	unsigned int nr_pages;
515 	unsigned int i;
516 
517 	local_vec = (struct rds_iovec __user *)(unsigned long) args->local_vec_addr;
518 
519 	/* figure out the number of pages in the vector */
520 	for (i = 0; i < args->nr_local; i++) {
521 		if (copy_from_user(&vec, &local_vec[i],
522 				   sizeof(struct rds_iovec)))
523 			return -EFAULT;
524 
525 		nr_pages = rds_pages_in_vec(&vec);
526 		if (nr_pages == 0)
527 			return -EINVAL;
528 
529 		tot_pages += nr_pages;
530 
531 		/*
532 		 * nr_pages for one entry is limited to (UINT_MAX>>PAGE_SHIFT)+1,
533 		 * so tot_pages cannot overflow without first going negative.
534 		 */
535 		if (tot_pages < 0)
536 			return -EINVAL;
537 	}
538 
539 	return tot_pages * sizeof(struct scatterlist);
540 }
541 
542 /*
543  * The application asks for a RDMA transfer.
544  * Extract all arguments and set up the rdma_op
545  */
546 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
547 			  struct cmsghdr *cmsg)
548 {
549 	struct rds_rdma_args *args;
550 	struct rm_rdma_op *op = &rm->rdma;
551 	int nr_pages;
552 	unsigned int nr_bytes;
553 	struct page **pages = NULL;
554 	struct rds_iovec iovstack[UIO_FASTIOV], *iovs = iovstack;
555 	int iov_size;
556 	unsigned int i, j;
557 	int ret = 0;
558 
559 	if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_rdma_args))
560 	    || rm->rdma.op_active)
561 		return -EINVAL;
562 
563 	args = CMSG_DATA(cmsg);
564 
565 	if (rs->rs_bound_addr == 0) {
566 		ret = -ENOTCONN; /* XXX not a great errno */
567 		goto out_ret;
568 	}
569 
570 	if (args->nr_local > UIO_MAXIOV) {
571 		ret = -EMSGSIZE;
572 		goto out_ret;
573 	}
574 
575 	/* Check whether to allocate the iovec area */
576 	iov_size = args->nr_local * sizeof(struct rds_iovec);
577 	if (args->nr_local > UIO_FASTIOV) {
578 		iovs = sock_kmalloc(rds_rs_to_sk(rs), iov_size, GFP_KERNEL);
579 		if (!iovs) {
580 			ret = -ENOMEM;
581 			goto out_ret;
582 		}
583 	}
584 
585 	if (copy_from_user(iovs, (struct rds_iovec __user *)(unsigned long) args->local_vec_addr, iov_size)) {
586 		ret = -EFAULT;
587 		goto out;
588 	}
589 
590 	nr_pages = rds_rdma_pages(iovs, args->nr_local);
591 	if (nr_pages < 0) {
592 		ret = -EINVAL;
593 		goto out;
594 	}
595 
596 	pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
597 	if (!pages) {
598 		ret = -ENOMEM;
599 		goto out;
600 	}
601 
602 	op->op_write = !!(args->flags & RDS_RDMA_READWRITE);
603 	op->op_fence = !!(args->flags & RDS_RDMA_FENCE);
604 	op->op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME);
605 	op->op_silent = !!(args->flags & RDS_RDMA_SILENT);
606 	op->op_active = 1;
607 	op->op_recverr = rs->rs_recverr;
608 	WARN_ON(!nr_pages);
609 	op->op_sg = rds_message_alloc_sgs(rm, nr_pages);
610 	if (!op->op_sg) {
611 		ret = -ENOMEM;
612 		goto out;
613 	}
614 
615 	if (op->op_notify || op->op_recverr) {
616 		/* We allocate an uninitialized notifier here, because
617 		 * we don't want to do that in the completion handler. We
618 		 * would have to use GFP_ATOMIC there, and don't want to deal
619 		 * with failed allocations.
620 		 */
621 		op->op_notifier = kmalloc(sizeof(struct rds_notifier), GFP_KERNEL);
622 		if (!op->op_notifier) {
623 			ret = -ENOMEM;
624 			goto out;
625 		}
626 		op->op_notifier->n_user_token = args->user_token;
627 		op->op_notifier->n_status = RDS_RDMA_SUCCESS;
628 	}
629 
630 	/* The cookie contains the R_Key of the remote memory region, and
631 	 * optionally an offset into it. This is how we implement RDMA into
632 	 * unaligned memory.
633 	 * When setting up the RDMA, we need to add that offset to the
634 	 * destination address (which is really an offset into the MR)
635 	 * FIXME: We may want to move this into ib_rdma.c
636 	 */
637 	op->op_rkey = rds_rdma_cookie_key(args->cookie);
638 	op->op_remote_addr = args->remote_vec.addr + rds_rdma_cookie_offset(args->cookie);
639 
640 	nr_bytes = 0;
641 
642 	rdsdebug("RDS: rdma prepare nr_local %llu rva %llx rkey %x\n",
643 	       (unsigned long long)args->nr_local,
644 	       (unsigned long long)args->remote_vec.addr,
645 	       op->op_rkey);
646 
647 	for (i = 0; i < args->nr_local; i++) {
648 		struct rds_iovec *iov = &iovs[i];
649 		/* don't need to check, rds_rdma_pages() verified nr will be +nonzero */
650 		unsigned int nr = rds_pages_in_vec(iov);
651 
652 		rs->rs_user_addr = iov->addr;
653 		rs->rs_user_bytes = iov->bytes;
654 
655 		/* If it's a WRITE operation, we want to pin the pages for reading.
656 		 * If it's a READ operation, we need to pin the pages for writing.
657 		 */
658 		ret = rds_pin_pages(iov->addr, nr, pages, !op->op_write);
659 		if (ret < 0)
660 			goto out;
661 
662 		rdsdebug("RDS: nr_bytes %u nr %u iov->bytes %llu iov->addr %llx\n",
663 			 nr_bytes, nr, iov->bytes, iov->addr);
664 
665 		nr_bytes += iov->bytes;
666 
667 		for (j = 0; j < nr; j++) {
668 			unsigned int offset = iov->addr & ~PAGE_MASK;
669 			struct scatterlist *sg;
670 
671 			sg = &op->op_sg[op->op_nents + j];
672 			sg_set_page(sg, pages[j],
673 					min_t(unsigned int, iov->bytes, PAGE_SIZE - offset),
674 					offset);
675 
676 			rdsdebug("RDS: sg->offset %x sg->len %x iov->addr %llx iov->bytes %llu\n",
677 			       sg->offset, sg->length, iov->addr, iov->bytes);
678 
679 			iov->addr += sg->length;
680 			iov->bytes -= sg->length;
681 		}
682 
683 		op->op_nents += nr;
684 	}
685 
686 	if (nr_bytes > args->remote_vec.bytes) {
687 		rdsdebug("RDS nr_bytes %u remote_bytes %u do not match\n",
688 				nr_bytes,
689 				(unsigned int) args->remote_vec.bytes);
690 		ret = -EINVAL;
691 		goto out;
692 	}
693 	op->op_bytes = nr_bytes;
694 
695 out:
696 	if (iovs != iovstack)
697 		sock_kfree_s(rds_rs_to_sk(rs), iovs, iov_size);
698 	kfree(pages);
699 out_ret:
700 	if (ret)
701 		rds_rdma_free_op(op);
702 	else
703 		rds_stats_inc(s_send_rdma);
704 
705 	return ret;
706 }
707 
708 /*
709  * The application wants us to pass an RDMA destination (aka MR)
710  * to the remote
711  */
712 int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
713 			  struct cmsghdr *cmsg)
714 {
715 	unsigned long flags;
716 	struct rds_mr *mr;
717 	u32 r_key;
718 	int err = 0;
719 
720 	if (cmsg->cmsg_len < CMSG_LEN(sizeof(rds_rdma_cookie_t)) ||
721 	    rm->m_rdma_cookie != 0)
722 		return -EINVAL;
723 
724 	memcpy(&rm->m_rdma_cookie, CMSG_DATA(cmsg), sizeof(rm->m_rdma_cookie));
725 
726 	/* We are reusing a previously mapped MR here. Most likely, the
727 	 * application has written to the buffer, so we need to explicitly
728 	 * flush those writes to RAM. Otherwise the HCA may not see them
729 	 * when doing a DMA from that buffer.
730 	 */
731 	r_key = rds_rdma_cookie_key(rm->m_rdma_cookie);
732 
733 	spin_lock_irqsave(&rs->rs_rdma_lock, flags);
734 	mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL);
735 	if (!mr)
736 		err = -EINVAL;	/* invalid r_key */
737 	else
738 		atomic_inc(&mr->r_refcount);
739 	spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
740 
741 	if (mr) {
742 		mr->r_trans->sync_mr(mr->r_trans_private, DMA_TO_DEVICE);
743 		rm->rdma.op_rdma_mr = mr;
744 	}
745 	return err;
746 }
747 
748 /*
749  * The application passes us an address range it wants to enable RDMA
750  * to/from. We map the area, and save the <R_Key,offset> pair
751  * in rm->m_rdma_cookie. This causes it to be sent along to the peer
752  * in an extension header.
753  */
754 int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
755 			  struct cmsghdr *cmsg)
756 {
757 	if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_get_mr_args)) ||
758 	    rm->m_rdma_cookie != 0)
759 		return -EINVAL;
760 
761 	return __rds_rdma_map(rs, CMSG_DATA(cmsg), &rm->m_rdma_cookie, &rm->rdma.op_rdma_mr);
762 }
763 
764 /*
765  * Fill in rds_message for an atomic request.
766  */
767 int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
768 		    struct cmsghdr *cmsg)
769 {
770 	struct page *page = NULL;
771 	struct rds_atomic_args *args;
772 	int ret = 0;
773 
774 	if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_atomic_args))
775 	 || rm->atomic.op_active)
776 		return -EINVAL;
777 
778 	args = CMSG_DATA(cmsg);
779 
780 	/* Nonmasked & masked cmsg ops converted to masked hw ops */
781 	switch (cmsg->cmsg_type) {
782 	case RDS_CMSG_ATOMIC_FADD:
783 		rm->atomic.op_type = RDS_ATOMIC_TYPE_FADD;
784 		rm->atomic.op_m_fadd.add = args->fadd.add;
785 		rm->atomic.op_m_fadd.nocarry_mask = 0;
786 		break;
787 	case RDS_CMSG_MASKED_ATOMIC_FADD:
788 		rm->atomic.op_type = RDS_ATOMIC_TYPE_FADD;
789 		rm->atomic.op_m_fadd.add = args->m_fadd.add;
790 		rm->atomic.op_m_fadd.nocarry_mask = args->m_fadd.nocarry_mask;
791 		break;
792 	case RDS_CMSG_ATOMIC_CSWP:
793 		rm->atomic.op_type = RDS_ATOMIC_TYPE_CSWP;
794 		rm->atomic.op_m_cswp.compare = args->cswp.compare;
795 		rm->atomic.op_m_cswp.swap = args->cswp.swap;
796 		rm->atomic.op_m_cswp.compare_mask = ~0;
797 		rm->atomic.op_m_cswp.swap_mask = ~0;
798 		break;
799 	case RDS_CMSG_MASKED_ATOMIC_CSWP:
800 		rm->atomic.op_type = RDS_ATOMIC_TYPE_CSWP;
801 		rm->atomic.op_m_cswp.compare = args->m_cswp.compare;
802 		rm->atomic.op_m_cswp.swap = args->m_cswp.swap;
803 		rm->atomic.op_m_cswp.compare_mask = args->m_cswp.compare_mask;
804 		rm->atomic.op_m_cswp.swap_mask = args->m_cswp.swap_mask;
805 		break;
806 	default:
807 		BUG(); /* should never happen */
808 	}
809 
810 	rm->atomic.op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME);
811 	rm->atomic.op_silent = !!(args->flags & RDS_RDMA_SILENT);
812 	rm->atomic.op_active = 1;
813 	rm->atomic.op_recverr = rs->rs_recverr;
814 	rm->atomic.op_sg = rds_message_alloc_sgs(rm, 1);
815 	if (!rm->atomic.op_sg) {
816 		ret = -ENOMEM;
817 		goto err;
818 	}
819 
820 	/* verify 8 byte-aligned */
821 	if (args->local_addr & 0x7) {
822 		ret = -EFAULT;
823 		goto err;
824 	}
825 
826 	ret = rds_pin_pages(args->local_addr, 1, &page, 1);
827 	if (ret != 1)
828 		goto err;
829 	ret = 0;
830 
831 	sg_set_page(rm->atomic.op_sg, page, 8, offset_in_page(args->local_addr));
832 
833 	if (rm->atomic.op_notify || rm->atomic.op_recverr) {
834 		/* We allocate an uninitialized notifier here, because
835 		 * we don't want to do that in the completion handler. We
836 		 * would have to use GFP_ATOMIC there, and don't want to deal
837 		 * with failed allocations.
838 		 */
839 		rm->atomic.op_notifier = kmalloc(sizeof(*rm->atomic.op_notifier), GFP_KERNEL);
840 		if (!rm->atomic.op_notifier) {
841 			ret = -ENOMEM;
842 			goto err;
843 		}
844 
845 		rm->atomic.op_notifier->n_user_token = args->user_token;
846 		rm->atomic.op_notifier->n_status = RDS_RDMA_SUCCESS;
847 	}
848 
849 	rm->atomic.op_rkey = rds_rdma_cookie_key(args->cookie);
850 	rm->atomic.op_remote_addr = args->remote_addr + rds_rdma_cookie_offset(args->cookie);
851 
852 	return ret;
853 err:
854 	if (page)
855 		put_page(page);
856 	kfree(rm->atomic.op_notifier);
857 
858 	return ret;
859 }
860