Loading...
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 */
56static 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
66static 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 */
96static 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
119void __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 */
129void 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 */
157static 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
173static 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;
308out:
309 kfree(pages);
310 if (mr)
311 rds_mr_put(mr);
312 return ret;
313}
314
315int 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
329int 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 */
356int 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 */
409void 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
444void rds_rdma_free_op(struct rm_rdma_op *ro)
445{
446 unsigned int i;
447
448 for (i = 0; i < ro->op_nents; i++) {
449 struct page *page = sg_page(&ro->op_sg[i]);
450
451 /* Mark page dirty if it was possibly modified, which
452 * is the case for a RDMA_READ which copies from remote
453 * to local memory */
454 if (!ro->op_write) {
455 WARN_ON(!page->mapping && irqs_disabled());
456 set_page_dirty(page);
457 }
458 put_page(page);
459 }
460
461 kfree(ro->op_notifier);
462 ro->op_notifier = NULL;
463 ro->op_active = 0;
464}
465
466void rds_atomic_free_op(struct rm_atomic_op *ao)
467{
468 struct page *page = sg_page(ao->op_sg);
469
470 /* Mark page dirty if it was possibly modified, which
471 * is the case for a RDMA_READ which copies from remote
472 * to local memory */
473 set_page_dirty(page);
474 put_page(page);
475
476 kfree(ao->op_notifier);
477 ao->op_notifier = NULL;
478 ao->op_active = 0;
479}
480
481
482/*
483 * Count the number of pages needed to describe an incoming iovec array.
484 */
485static int rds_rdma_pages(struct rds_iovec iov[], int nr_iovecs)
486{
487 int tot_pages = 0;
488 unsigned int nr_pages;
489 unsigned int i;
490
491 /* figure out the number of pages in the vector */
492 for (i = 0; i < nr_iovecs; i++) {
493 nr_pages = rds_pages_in_vec(&iov[i]);
494 if (nr_pages == 0)
495 return -EINVAL;
496
497 tot_pages += nr_pages;
498
499 /*
500 * nr_pages for one entry is limited to (UINT_MAX>>PAGE_SHIFT)+1,
501 * so tot_pages cannot overflow without first going negative.
502 */
503 if (tot_pages < 0)
504 return -EINVAL;
505 }
506
507 return tot_pages;
508}
509
510int rds_rdma_extra_size(struct rds_rdma_args *args)
511{
512 struct rds_iovec vec;
513 struct rds_iovec __user *local_vec;
514 int tot_pages = 0;
515 unsigned int nr_pages;
516 unsigned int i;
517
518 local_vec = (struct rds_iovec __user *)(unsigned long) args->local_vec_addr;
519
520 /* figure out the number of pages in the vector */
521 for (i = 0; i < args->nr_local; i++) {
522 if (copy_from_user(&vec, &local_vec[i],
523 sizeof(struct rds_iovec)))
524 return -EFAULT;
525
526 nr_pages = rds_pages_in_vec(&vec);
527 if (nr_pages == 0)
528 return -EINVAL;
529
530 tot_pages += nr_pages;
531
532 /*
533 * nr_pages for one entry is limited to (UINT_MAX>>PAGE_SHIFT)+1,
534 * so tot_pages cannot overflow without first going negative.
535 */
536 if (tot_pages < 0)
537 return -EINVAL;
538 }
539
540 return tot_pages * sizeof(struct scatterlist);
541}
542
543/*
544 * The application asks for a RDMA transfer.
545 * Extract all arguments and set up the rdma_op
546 */
547int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
548 struct cmsghdr *cmsg)
549{
550 struct rds_rdma_args *args;
551 struct rm_rdma_op *op = &rm->rdma;
552 int nr_pages;
553 unsigned int nr_bytes;
554 struct page **pages = NULL;
555 struct rds_iovec iovstack[UIO_FASTIOV], *iovs = iovstack;
556 int iov_size;
557 unsigned int i, j;
558 int ret = 0;
559
560 if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_rdma_args))
561 || rm->rdma.op_active)
562 return -EINVAL;
563
564 args = CMSG_DATA(cmsg);
565
566 if (rs->rs_bound_addr == 0) {
567 ret = -ENOTCONN; /* XXX not a great errno */
568 goto out_ret;
569 }
570
571 if (args->nr_local > UIO_MAXIOV) {
572 ret = -EMSGSIZE;
573 goto out_ret;
574 }
575
576 /* Check whether to allocate the iovec area */
577 iov_size = args->nr_local * sizeof(struct rds_iovec);
578 if (args->nr_local > UIO_FASTIOV) {
579 iovs = sock_kmalloc(rds_rs_to_sk(rs), iov_size, GFP_KERNEL);
580 if (!iovs) {
581 ret = -ENOMEM;
582 goto out_ret;
583 }
584 }
585
586 if (copy_from_user(iovs, (struct rds_iovec __user *)(unsigned long) args->local_vec_addr, iov_size)) {
587 ret = -EFAULT;
588 goto out;
589 }
590
591 nr_pages = rds_rdma_pages(iovs, args->nr_local);
592 if (nr_pages < 0) {
593 ret = -EINVAL;
594 goto out;
595 }
596
597 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
598 if (!pages) {
599 ret = -ENOMEM;
600 goto out;
601 }
602
603 op->op_write = !!(args->flags & RDS_RDMA_READWRITE);
604 op->op_fence = !!(args->flags & RDS_RDMA_FENCE);
605 op->op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME);
606 op->op_silent = !!(args->flags & RDS_RDMA_SILENT);
607 op->op_active = 1;
608 op->op_recverr = rs->rs_recverr;
609 WARN_ON(!nr_pages);
610 op->op_sg = rds_message_alloc_sgs(rm, nr_pages);
611 if (!op->op_sg) {
612 ret = -ENOMEM;
613 goto out;
614 }
615
616 if (op->op_notify || op->op_recverr) {
617 /* We allocate an uninitialized notifier here, because
618 * we don't want to do that in the completion handler. We
619 * would have to use GFP_ATOMIC there, and don't want to deal
620 * with failed allocations.
621 */
622 op->op_notifier = kmalloc(sizeof(struct rds_notifier), GFP_KERNEL);
623 if (!op->op_notifier) {
624 ret = -ENOMEM;
625 goto out;
626 }
627 op->op_notifier->n_user_token = args->user_token;
628 op->op_notifier->n_status = RDS_RDMA_SUCCESS;
629 }
630
631 /* The cookie contains the R_Key of the remote memory region, and
632 * optionally an offset into it. This is how we implement RDMA into
633 * unaligned memory.
634 * When setting up the RDMA, we need to add that offset to the
635 * destination address (which is really an offset into the MR)
636 * FIXME: We may want to move this into ib_rdma.c
637 */
638 op->op_rkey = rds_rdma_cookie_key(args->cookie);
639 op->op_remote_addr = args->remote_vec.addr + rds_rdma_cookie_offset(args->cookie);
640
641 nr_bytes = 0;
642
643 rdsdebug("RDS: rdma prepare nr_local %llu rva %llx rkey %x\n",
644 (unsigned long long)args->nr_local,
645 (unsigned long long)args->remote_vec.addr,
646 op->op_rkey);
647
648 for (i = 0; i < args->nr_local; i++) {
649 struct rds_iovec *iov = &iovs[i];
650 /* don't need to check, rds_rdma_pages() verified nr will be +nonzero */
651 unsigned int nr = rds_pages_in_vec(iov);
652
653 rs->rs_user_addr = iov->addr;
654 rs->rs_user_bytes = iov->bytes;
655
656 /* If it's a WRITE operation, we want to pin the pages for reading.
657 * If it's a READ operation, we need to pin the pages for writing.
658 */
659 ret = rds_pin_pages(iov->addr, nr, pages, !op->op_write);
660 if (ret < 0)
661 goto out;
662 else
663 ret = 0;
664
665 rdsdebug("RDS: nr_bytes %u nr %u iov->bytes %llu iov->addr %llx\n",
666 nr_bytes, nr, iov->bytes, iov->addr);
667
668 nr_bytes += iov->bytes;
669
670 for (j = 0; j < nr; j++) {
671 unsigned int offset = iov->addr & ~PAGE_MASK;
672 struct scatterlist *sg;
673
674 sg = &op->op_sg[op->op_nents + j];
675 sg_set_page(sg, pages[j],
676 min_t(unsigned int, iov->bytes, PAGE_SIZE - offset),
677 offset);
678
679 rdsdebug("RDS: sg->offset %x sg->len %x iov->addr %llx iov->bytes %llu\n",
680 sg->offset, sg->length, iov->addr, iov->bytes);
681
682 iov->addr += sg->length;
683 iov->bytes -= sg->length;
684 }
685
686 op->op_nents += nr;
687 }
688
689 if (nr_bytes > args->remote_vec.bytes) {
690 rdsdebug("RDS nr_bytes %u remote_bytes %u do not match\n",
691 nr_bytes,
692 (unsigned int) args->remote_vec.bytes);
693 ret = -EINVAL;
694 goto out;
695 }
696 op->op_bytes = nr_bytes;
697
698out:
699 if (iovs != iovstack)
700 sock_kfree_s(rds_rs_to_sk(rs), iovs, iov_size);
701 kfree(pages);
702out_ret:
703 if (ret)
704 rds_rdma_free_op(op);
705 else
706 rds_stats_inc(s_send_rdma);
707
708 return ret;
709}
710
711/*
712 * The application wants us to pass an RDMA destination (aka MR)
713 * to the remote
714 */
715int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
716 struct cmsghdr *cmsg)
717{
718 unsigned long flags;
719 struct rds_mr *mr;
720 u32 r_key;
721 int err = 0;
722
723 if (cmsg->cmsg_len < CMSG_LEN(sizeof(rds_rdma_cookie_t)) ||
724 rm->m_rdma_cookie != 0)
725 return -EINVAL;
726
727 memcpy(&rm->m_rdma_cookie, CMSG_DATA(cmsg), sizeof(rm->m_rdma_cookie));
728
729 /* We are reusing a previously mapped MR here. Most likely, the
730 * application has written to the buffer, so we need to explicitly
731 * flush those writes to RAM. Otherwise the HCA may not see them
732 * when doing a DMA from that buffer.
733 */
734 r_key = rds_rdma_cookie_key(rm->m_rdma_cookie);
735
736 spin_lock_irqsave(&rs->rs_rdma_lock, flags);
737 mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL);
738 if (!mr)
739 err = -EINVAL; /* invalid r_key */
740 else
741 atomic_inc(&mr->r_refcount);
742 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
743
744 if (mr) {
745 mr->r_trans->sync_mr(mr->r_trans_private, DMA_TO_DEVICE);
746 rm->rdma.op_rdma_mr = mr;
747 }
748 return err;
749}
750
751/*
752 * The application passes us an address range it wants to enable RDMA
753 * to/from. We map the area, and save the <R_Key,offset> pair
754 * in rm->m_rdma_cookie. This causes it to be sent along to the peer
755 * in an extension header.
756 */
757int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
758 struct cmsghdr *cmsg)
759{
760 if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_get_mr_args)) ||
761 rm->m_rdma_cookie != 0)
762 return -EINVAL;
763
764 return __rds_rdma_map(rs, CMSG_DATA(cmsg), &rm->m_rdma_cookie, &rm->rdma.op_rdma_mr);
765}
766
767/*
768 * Fill in rds_message for an atomic request.
769 */
770int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
771 struct cmsghdr *cmsg)
772{
773 struct page *page = NULL;
774 struct rds_atomic_args *args;
775 int ret = 0;
776
777 if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_atomic_args))
778 || rm->atomic.op_active)
779 return -EINVAL;
780
781 args = CMSG_DATA(cmsg);
782
783 /* Nonmasked & masked cmsg ops converted to masked hw ops */
784 switch (cmsg->cmsg_type) {
785 case RDS_CMSG_ATOMIC_FADD:
786 rm->atomic.op_type = RDS_ATOMIC_TYPE_FADD;
787 rm->atomic.op_m_fadd.add = args->fadd.add;
788 rm->atomic.op_m_fadd.nocarry_mask = 0;
789 break;
790 case RDS_CMSG_MASKED_ATOMIC_FADD:
791 rm->atomic.op_type = RDS_ATOMIC_TYPE_FADD;
792 rm->atomic.op_m_fadd.add = args->m_fadd.add;
793 rm->atomic.op_m_fadd.nocarry_mask = args->m_fadd.nocarry_mask;
794 break;
795 case RDS_CMSG_ATOMIC_CSWP:
796 rm->atomic.op_type = RDS_ATOMIC_TYPE_CSWP;
797 rm->atomic.op_m_cswp.compare = args->cswp.compare;
798 rm->atomic.op_m_cswp.swap = args->cswp.swap;
799 rm->atomic.op_m_cswp.compare_mask = ~0;
800 rm->atomic.op_m_cswp.swap_mask = ~0;
801 break;
802 case RDS_CMSG_MASKED_ATOMIC_CSWP:
803 rm->atomic.op_type = RDS_ATOMIC_TYPE_CSWP;
804 rm->atomic.op_m_cswp.compare = args->m_cswp.compare;
805 rm->atomic.op_m_cswp.swap = args->m_cswp.swap;
806 rm->atomic.op_m_cswp.compare_mask = args->m_cswp.compare_mask;
807 rm->atomic.op_m_cswp.swap_mask = args->m_cswp.swap_mask;
808 break;
809 default:
810 BUG(); /* should never happen */
811 }
812
813 rm->atomic.op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME);
814 rm->atomic.op_silent = !!(args->flags & RDS_RDMA_SILENT);
815 rm->atomic.op_active = 1;
816 rm->atomic.op_recverr = rs->rs_recverr;
817 rm->atomic.op_sg = rds_message_alloc_sgs(rm, 1);
818 if (!rm->atomic.op_sg) {
819 ret = -ENOMEM;
820 goto err;
821 }
822
823 /* verify 8 byte-aligned */
824 if (args->local_addr & 0x7) {
825 ret = -EFAULT;
826 goto err;
827 }
828
829 ret = rds_pin_pages(args->local_addr, 1, &page, 1);
830 if (ret != 1)
831 goto err;
832 ret = 0;
833
834 sg_set_page(rm->atomic.op_sg, page, 8, offset_in_page(args->local_addr));
835
836 if (rm->atomic.op_notify || rm->atomic.op_recverr) {
837 /* We allocate an uninitialized notifier here, because
838 * we don't want to do that in the completion handler. We
839 * would have to use GFP_ATOMIC there, and don't want to deal
840 * with failed allocations.
841 */
842 rm->atomic.op_notifier = kmalloc(sizeof(*rm->atomic.op_notifier), GFP_KERNEL);
843 if (!rm->atomic.op_notifier) {
844 ret = -ENOMEM;
845 goto err;
846 }
847
848 rm->atomic.op_notifier->n_user_token = args->user_token;
849 rm->atomic.op_notifier->n_status = RDS_RDMA_SUCCESS;
850 }
851
852 rm->atomic.op_rkey = rds_rdma_cookie_key(args->cookie);
853 rm->atomic.op_remote_addr = args->remote_addr + rds_rdma_cookie_offset(args->cookie);
854
855 return ret;
856err:
857 if (page)
858 put_page(page);
859 kfree(rm->atomic.op_notifier);
860
861 return ret;
862}
1/*
2 * Copyright (c) 2007, 2020 Oracle and/or its affiliates.
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 detect duplicate keys on a socket? hmm.
44 * - an rdma is an mlock, apply rlimit?
45 */
46
47/*
48 * get the number of pages by looking at the page indices that the start and
49 * end addresses fall in.
50 *
51 * Returns 0 if the vec is invalid. It is invalid if the number of bytes
52 * causes the address to wrap or overflows an unsigned int. This comes
53 * from being stored in the 'length' member of 'struct scatterlist'.
54 */
55static unsigned int rds_pages_in_vec(struct rds_iovec *vec)
56{
57 if ((vec->addr + vec->bytes <= vec->addr) ||
58 (vec->bytes > (u64)UINT_MAX))
59 return 0;
60
61 return ((vec->addr + vec->bytes + PAGE_SIZE - 1) >> PAGE_SHIFT) -
62 (vec->addr >> PAGE_SHIFT);
63}
64
65static struct rds_mr *rds_mr_tree_walk(struct rb_root *root, u64 key,
66 struct rds_mr *insert)
67{
68 struct rb_node **p = &root->rb_node;
69 struct rb_node *parent = NULL;
70 struct rds_mr *mr;
71
72 while (*p) {
73 parent = *p;
74 mr = rb_entry(parent, struct rds_mr, r_rb_node);
75
76 if (key < mr->r_key)
77 p = &(*p)->rb_left;
78 else if (key > mr->r_key)
79 p = &(*p)->rb_right;
80 else
81 return mr;
82 }
83
84 if (insert) {
85 rb_link_node(&insert->r_rb_node, parent, p);
86 rb_insert_color(&insert->r_rb_node, root);
87 kref_get(&insert->r_kref);
88 }
89 return NULL;
90}
91
92/*
93 * Destroy the transport-specific part of a MR.
94 */
95static void rds_destroy_mr(struct rds_mr *mr)
96{
97 struct rds_sock *rs = mr->r_sock;
98 void *trans_private = NULL;
99 unsigned long flags;
100
101 rdsdebug("RDS: destroy mr key is %x refcnt %u\n",
102 mr->r_key, kref_read(&mr->r_kref));
103
104 spin_lock_irqsave(&rs->rs_rdma_lock, flags);
105 if (!RB_EMPTY_NODE(&mr->r_rb_node))
106 rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
107 trans_private = mr->r_trans_private;
108 mr->r_trans_private = NULL;
109 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
110
111 if (trans_private)
112 mr->r_trans->free_mr(trans_private, mr->r_invalidate);
113}
114
115void __rds_put_mr_final(struct kref *kref)
116{
117 struct rds_mr *mr = container_of(kref, struct rds_mr, r_kref);
118
119 rds_destroy_mr(mr);
120 kfree(mr);
121}
122
123/*
124 * By the time this is called we can't have any more ioctls called on
125 * the socket so we don't need to worry about racing with others.
126 */
127void rds_rdma_drop_keys(struct rds_sock *rs)
128{
129 struct rds_mr *mr;
130 struct rb_node *node;
131 unsigned long flags;
132
133 /* Release any MRs associated with this socket */
134 spin_lock_irqsave(&rs->rs_rdma_lock, flags);
135 while ((node = rb_first(&rs->rs_rdma_keys))) {
136 mr = rb_entry(node, struct rds_mr, r_rb_node);
137 if (mr->r_trans == rs->rs_transport)
138 mr->r_invalidate = 0;
139 rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
140 RB_CLEAR_NODE(&mr->r_rb_node);
141 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
142 kref_put(&mr->r_kref, __rds_put_mr_final);
143 spin_lock_irqsave(&rs->rs_rdma_lock, flags);
144 }
145 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
146
147 if (rs->rs_transport && rs->rs_transport->flush_mrs)
148 rs->rs_transport->flush_mrs();
149}
150
151/*
152 * Helper function to pin user pages.
153 */
154static int rds_pin_pages(unsigned long user_addr, unsigned int nr_pages,
155 struct page **pages, int write)
156{
157 unsigned int gup_flags = FOLL_LONGTERM;
158 int ret;
159
160 if (write)
161 gup_flags |= FOLL_WRITE;
162
163 ret = pin_user_pages_fast(user_addr, nr_pages, gup_flags, pages);
164 if (ret >= 0 && ret < nr_pages) {
165 unpin_user_pages(pages, ret);
166 ret = -EFAULT;
167 }
168
169 return ret;
170}
171
172static int __rds_rdma_map(struct rds_sock *rs, struct rds_get_mr_args *args,
173 u64 *cookie_ret, struct rds_mr **mr_ret,
174 struct rds_conn_path *cp)
175{
176 struct rds_mr *mr = NULL, *found;
177 struct scatterlist *sg = NULL;
178 unsigned int nr_pages;
179 struct page **pages = NULL;
180 void *trans_private;
181 unsigned long flags;
182 rds_rdma_cookie_t cookie;
183 unsigned int nents = 0;
184 int need_odp = 0;
185 long i;
186 int ret;
187
188 if (ipv6_addr_any(&rs->rs_bound_addr) || !rs->rs_transport) {
189 ret = -ENOTCONN; /* XXX not a great errno */
190 goto out;
191 }
192
193 if (!rs->rs_transport->get_mr) {
194 ret = -EOPNOTSUPP;
195 goto out;
196 }
197
198 /* If the combination of the addr and size requested for this memory
199 * region causes an integer overflow, return error.
200 */
201 if (((args->vec.addr + args->vec.bytes) < args->vec.addr) ||
202 PAGE_ALIGN(args->vec.addr + args->vec.bytes) <
203 (args->vec.addr + args->vec.bytes)) {
204 ret = -EINVAL;
205 goto out;
206 }
207
208 if (!can_do_mlock()) {
209 ret = -EPERM;
210 goto out;
211 }
212
213 nr_pages = rds_pages_in_vec(&args->vec);
214 if (nr_pages == 0) {
215 ret = -EINVAL;
216 goto out;
217 }
218
219 /* Restrict the size of mr irrespective of underlying transport
220 * To account for unaligned mr regions, subtract one from nr_pages
221 */
222 if ((nr_pages - 1) > (RDS_MAX_MSG_SIZE >> PAGE_SHIFT)) {
223 ret = -EMSGSIZE;
224 goto out;
225 }
226
227 rdsdebug("RDS: get_mr addr %llx len %llu nr_pages %u\n",
228 args->vec.addr, args->vec.bytes, nr_pages);
229
230 /* XXX clamp nr_pages to limit the size of this alloc? */
231 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
232 if (!pages) {
233 ret = -ENOMEM;
234 goto out;
235 }
236
237 mr = kzalloc(sizeof(struct rds_mr), GFP_KERNEL);
238 if (!mr) {
239 ret = -ENOMEM;
240 goto out;
241 }
242
243 kref_init(&mr->r_kref);
244 RB_CLEAR_NODE(&mr->r_rb_node);
245 mr->r_trans = rs->rs_transport;
246 mr->r_sock = rs;
247
248 if (args->flags & RDS_RDMA_USE_ONCE)
249 mr->r_use_once = 1;
250 if (args->flags & RDS_RDMA_INVALIDATE)
251 mr->r_invalidate = 1;
252 if (args->flags & RDS_RDMA_READWRITE)
253 mr->r_write = 1;
254
255 /*
256 * Pin the pages that make up the user buffer and transfer the page
257 * pointers to the mr's sg array. We check to see if we've mapped
258 * the whole region after transferring the partial page references
259 * to the sg array so that we can have one page ref cleanup path.
260 *
261 * For now we have no flag that tells us whether the mapping is
262 * r/o or r/w. We need to assume r/w, or we'll do a lot of RDMA to
263 * the zero page.
264 */
265 ret = rds_pin_pages(args->vec.addr, nr_pages, pages, 1);
266 if (ret == -EOPNOTSUPP) {
267 need_odp = 1;
268 } else if (ret <= 0) {
269 goto out;
270 } else {
271 nents = ret;
272 sg = kmalloc_array(nents, sizeof(*sg), GFP_KERNEL);
273 if (!sg) {
274 ret = -ENOMEM;
275 goto out;
276 }
277 WARN_ON(!nents);
278 sg_init_table(sg, nents);
279
280 /* Stick all pages into the scatterlist */
281 for (i = 0 ; i < nents; i++)
282 sg_set_page(&sg[i], pages[i], PAGE_SIZE, 0);
283
284 rdsdebug("RDS: trans_private nents is %u\n", nents);
285 }
286 /* Obtain a transport specific MR. If this succeeds, the
287 * s/g list is now owned by the MR.
288 * Note that dma_map() implies that pending writes are
289 * flushed to RAM, so no dma_sync is needed here. */
290 trans_private = rs->rs_transport->get_mr(
291 sg, nents, rs, &mr->r_key, cp ? cp->cp_conn : NULL,
292 args->vec.addr, args->vec.bytes,
293 need_odp ? ODP_ZEROBASED : ODP_NOT_NEEDED);
294
295 if (IS_ERR(trans_private)) {
296 /* In ODP case, we don't GUP pages, so don't need
297 * to release anything.
298 */
299 if (!need_odp) {
300 unpin_user_pages(pages, nr_pages);
301 kfree(sg);
302 }
303 ret = PTR_ERR(trans_private);
304 goto out;
305 }
306
307 mr->r_trans_private = trans_private;
308
309 rdsdebug("RDS: get_mr put_user key is %x cookie_addr %p\n",
310 mr->r_key, (void *)(unsigned long) args->cookie_addr);
311
312 /* The user may pass us an unaligned address, but we can only
313 * map page aligned regions. So we keep the offset, and build
314 * a 64bit cookie containing <R_Key, offset> and pass that
315 * around. */
316 if (need_odp)
317 cookie = rds_rdma_make_cookie(mr->r_key, 0);
318 else
319 cookie = rds_rdma_make_cookie(mr->r_key,
320 args->vec.addr & ~PAGE_MASK);
321 if (cookie_ret)
322 *cookie_ret = cookie;
323
324 if (args->cookie_addr &&
325 put_user(cookie, (u64 __user *)(unsigned long)args->cookie_addr)) {
326 if (!need_odp) {
327 unpin_user_pages(pages, nr_pages);
328 kfree(sg);
329 }
330 ret = -EFAULT;
331 goto out;
332 }
333
334 /* Inserting the new MR into the rbtree bumps its
335 * reference count. */
336 spin_lock_irqsave(&rs->rs_rdma_lock, flags);
337 found = rds_mr_tree_walk(&rs->rs_rdma_keys, mr->r_key, mr);
338 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
339
340 BUG_ON(found && found != mr);
341
342 rdsdebug("RDS: get_mr key is %x\n", mr->r_key);
343 if (mr_ret) {
344 kref_get(&mr->r_kref);
345 *mr_ret = mr;
346 }
347
348 ret = 0;
349out:
350 kfree(pages);
351 if (mr)
352 kref_put(&mr->r_kref, __rds_put_mr_final);
353 return ret;
354}
355
356int rds_get_mr(struct rds_sock *rs, sockptr_t optval, int optlen)
357{
358 struct rds_get_mr_args args;
359
360 if (optlen != sizeof(struct rds_get_mr_args))
361 return -EINVAL;
362
363 if (copy_from_sockptr(&args, optval, sizeof(struct rds_get_mr_args)))
364 return -EFAULT;
365
366 return __rds_rdma_map(rs, &args, NULL, NULL, NULL);
367}
368
369int rds_get_mr_for_dest(struct rds_sock *rs, sockptr_t optval, int optlen)
370{
371 struct rds_get_mr_for_dest_args args;
372 struct rds_get_mr_args new_args;
373
374 if (optlen != sizeof(struct rds_get_mr_for_dest_args))
375 return -EINVAL;
376
377 if (copy_from_sockptr(&args, optval,
378 sizeof(struct rds_get_mr_for_dest_args)))
379 return -EFAULT;
380
381 /*
382 * Initially, just behave like get_mr().
383 * TODO: Implement get_mr as wrapper around this
384 * and deprecate it.
385 */
386 new_args.vec = args.vec;
387 new_args.cookie_addr = args.cookie_addr;
388 new_args.flags = args.flags;
389
390 return __rds_rdma_map(rs, &new_args, NULL, NULL, NULL);
391}
392
393/*
394 * Free the MR indicated by the given R_Key
395 */
396int rds_free_mr(struct rds_sock *rs, sockptr_t optval, int optlen)
397{
398 struct rds_free_mr_args args;
399 struct rds_mr *mr;
400 unsigned long flags;
401
402 if (optlen != sizeof(struct rds_free_mr_args))
403 return -EINVAL;
404
405 if (copy_from_sockptr(&args, optval, sizeof(struct rds_free_mr_args)))
406 return -EFAULT;
407
408 /* Special case - a null cookie means flush all unused MRs */
409 if (args.cookie == 0) {
410 if (!rs->rs_transport || !rs->rs_transport->flush_mrs)
411 return -EINVAL;
412 rs->rs_transport->flush_mrs();
413 return 0;
414 }
415
416 /* Look up the MR given its R_key and remove it from the rbtree
417 * so nobody else finds it.
418 * This should also prevent races with rds_rdma_unuse.
419 */
420 spin_lock_irqsave(&rs->rs_rdma_lock, flags);
421 mr = rds_mr_tree_walk(&rs->rs_rdma_keys, rds_rdma_cookie_key(args.cookie), NULL);
422 if (mr) {
423 rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
424 RB_CLEAR_NODE(&mr->r_rb_node);
425 if (args.flags & RDS_RDMA_INVALIDATE)
426 mr->r_invalidate = 1;
427 }
428 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
429
430 if (!mr)
431 return -EINVAL;
432
433 kref_put(&mr->r_kref, __rds_put_mr_final);
434 return 0;
435}
436
437/*
438 * This is called when we receive an extension header that
439 * tells us this MR was used. It allows us to implement
440 * use_once semantics
441 */
442void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force)
443{
444 struct rds_mr *mr;
445 unsigned long flags;
446 int zot_me = 0;
447
448 spin_lock_irqsave(&rs->rs_rdma_lock, flags);
449 mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL);
450 if (!mr) {
451 pr_debug("rds: trying to unuse MR with unknown r_key %u!\n",
452 r_key);
453 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
454 return;
455 }
456
457 /* Get a reference so that the MR won't go away before calling
458 * sync_mr() below.
459 */
460 kref_get(&mr->r_kref);
461
462 /* If it is going to be freed, remove it from the tree now so
463 * that no other thread can find it and free it.
464 */
465 if (mr->r_use_once || force) {
466 rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
467 RB_CLEAR_NODE(&mr->r_rb_node);
468 zot_me = 1;
469 }
470 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
471
472 /* May have to issue a dma_sync on this memory region.
473 * Note we could avoid this if the operation was a RDMA READ,
474 * but at this point we can't tell. */
475 if (mr->r_trans->sync_mr)
476 mr->r_trans->sync_mr(mr->r_trans_private, DMA_FROM_DEVICE);
477
478 /* Release the reference held above. */
479 kref_put(&mr->r_kref, __rds_put_mr_final);
480
481 /* If the MR was marked as invalidate, this will
482 * trigger an async flush. */
483 if (zot_me)
484 kref_put(&mr->r_kref, __rds_put_mr_final);
485}
486
487void rds_rdma_free_op(struct rm_rdma_op *ro)
488{
489 unsigned int i;
490
491 if (ro->op_odp_mr) {
492 kref_put(&ro->op_odp_mr->r_kref, __rds_put_mr_final);
493 } else {
494 for (i = 0; i < ro->op_nents; i++) {
495 struct page *page = sg_page(&ro->op_sg[i]);
496
497 /* Mark page dirty if it was possibly modified, which
498 * is the case for a RDMA_READ which copies from remote
499 * to local memory
500 */
501 unpin_user_pages_dirty_lock(&page, 1, !ro->op_write);
502 }
503 }
504
505 kfree(ro->op_notifier);
506 ro->op_notifier = NULL;
507 ro->op_active = 0;
508 ro->op_odp_mr = NULL;
509}
510
511void rds_atomic_free_op(struct rm_atomic_op *ao)
512{
513 struct page *page = sg_page(ao->op_sg);
514
515 /* Mark page dirty if it was possibly modified, which
516 * is the case for a RDMA_READ which copies from remote
517 * to local memory */
518 unpin_user_pages_dirty_lock(&page, 1, true);
519
520 kfree(ao->op_notifier);
521 ao->op_notifier = NULL;
522 ao->op_active = 0;
523}
524
525
526/*
527 * Count the number of pages needed to describe an incoming iovec array.
528 */
529static int rds_rdma_pages(struct rds_iovec iov[], int nr_iovecs)
530{
531 int tot_pages = 0;
532 unsigned int nr_pages;
533 unsigned int i;
534
535 /* figure out the number of pages in the vector */
536 for (i = 0; i < nr_iovecs; i++) {
537 nr_pages = rds_pages_in_vec(&iov[i]);
538 if (nr_pages == 0)
539 return -EINVAL;
540
541 tot_pages += nr_pages;
542
543 /*
544 * nr_pages for one entry is limited to (UINT_MAX>>PAGE_SHIFT)+1,
545 * so tot_pages cannot overflow without first going negative.
546 */
547 if (tot_pages < 0)
548 return -EINVAL;
549 }
550
551 return tot_pages;
552}
553
554int rds_rdma_extra_size(struct rds_rdma_args *args,
555 struct rds_iov_vector *iov)
556{
557 struct rds_iovec *vec;
558 struct rds_iovec __user *local_vec;
559 int tot_pages = 0;
560 unsigned int nr_pages;
561 unsigned int i;
562
563 local_vec = (struct rds_iovec __user *)(unsigned long) args->local_vec_addr;
564
565 if (args->nr_local == 0)
566 return -EINVAL;
567
568 if (args->nr_local > UIO_MAXIOV)
569 return -EMSGSIZE;
570
571 iov->iov = kcalloc(args->nr_local,
572 sizeof(struct rds_iovec),
573 GFP_KERNEL);
574 if (!iov->iov)
575 return -ENOMEM;
576
577 vec = &iov->iov[0];
578
579 if (copy_from_user(vec, local_vec, args->nr_local *
580 sizeof(struct rds_iovec)))
581 return -EFAULT;
582 iov->len = args->nr_local;
583
584 /* figure out the number of pages in the vector */
585 for (i = 0; i < args->nr_local; i++, vec++) {
586
587 nr_pages = rds_pages_in_vec(vec);
588 if (nr_pages == 0)
589 return -EINVAL;
590
591 tot_pages += nr_pages;
592
593 /*
594 * nr_pages for one entry is limited to (UINT_MAX>>PAGE_SHIFT)+1,
595 * so tot_pages cannot overflow without first going negative.
596 */
597 if (tot_pages < 0)
598 return -EINVAL;
599 }
600
601 return tot_pages * sizeof(struct scatterlist);
602}
603
604/*
605 * The application asks for a RDMA transfer.
606 * Extract all arguments and set up the rdma_op
607 */
608int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
609 struct cmsghdr *cmsg,
610 struct rds_iov_vector *vec)
611{
612 struct rds_rdma_args *args;
613 struct rm_rdma_op *op = &rm->rdma;
614 int nr_pages;
615 unsigned int nr_bytes;
616 struct page **pages = NULL;
617 struct rds_iovec *iovs;
618 unsigned int i, j;
619 int ret = 0;
620 bool odp_supported = true;
621
622 if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_rdma_args))
623 || rm->rdma.op_active)
624 return -EINVAL;
625
626 args = CMSG_DATA(cmsg);
627
628 if (ipv6_addr_any(&rs->rs_bound_addr)) {
629 ret = -ENOTCONN; /* XXX not a great errno */
630 goto out_ret;
631 }
632
633 if (args->nr_local > UIO_MAXIOV) {
634 ret = -EMSGSIZE;
635 goto out_ret;
636 }
637
638 if (vec->len != args->nr_local) {
639 ret = -EINVAL;
640 goto out_ret;
641 }
642 /* odp-mr is not supported for multiple requests within one message */
643 if (args->nr_local != 1)
644 odp_supported = false;
645
646 iovs = vec->iov;
647
648 nr_pages = rds_rdma_pages(iovs, args->nr_local);
649 if (nr_pages < 0) {
650 ret = -EINVAL;
651 goto out_ret;
652 }
653
654 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
655 if (!pages) {
656 ret = -ENOMEM;
657 goto out_ret;
658 }
659
660 op->op_write = !!(args->flags & RDS_RDMA_READWRITE);
661 op->op_fence = !!(args->flags & RDS_RDMA_FENCE);
662 op->op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME);
663 op->op_silent = !!(args->flags & RDS_RDMA_SILENT);
664 op->op_active = 1;
665 op->op_recverr = rs->rs_recverr;
666 op->op_odp_mr = NULL;
667
668 WARN_ON(!nr_pages);
669 op->op_sg = rds_message_alloc_sgs(rm, nr_pages);
670 if (IS_ERR(op->op_sg)) {
671 ret = PTR_ERR(op->op_sg);
672 goto out_pages;
673 }
674
675 if (op->op_notify || op->op_recverr) {
676 /* We allocate an uninitialized notifier here, because
677 * we don't want to do that in the completion handler. We
678 * would have to use GFP_ATOMIC there, and don't want to deal
679 * with failed allocations.
680 */
681 op->op_notifier = kmalloc(sizeof(struct rds_notifier), GFP_KERNEL);
682 if (!op->op_notifier) {
683 ret = -ENOMEM;
684 goto out_pages;
685 }
686 op->op_notifier->n_user_token = args->user_token;
687 op->op_notifier->n_status = RDS_RDMA_SUCCESS;
688 }
689
690 /* The cookie contains the R_Key of the remote memory region, and
691 * optionally an offset into it. This is how we implement RDMA into
692 * unaligned memory.
693 * When setting up the RDMA, we need to add that offset to the
694 * destination address (which is really an offset into the MR)
695 * FIXME: We may want to move this into ib_rdma.c
696 */
697 op->op_rkey = rds_rdma_cookie_key(args->cookie);
698 op->op_remote_addr = args->remote_vec.addr + rds_rdma_cookie_offset(args->cookie);
699
700 nr_bytes = 0;
701
702 rdsdebug("RDS: rdma prepare nr_local %llu rva %llx rkey %x\n",
703 (unsigned long long)args->nr_local,
704 (unsigned long long)args->remote_vec.addr,
705 op->op_rkey);
706
707 for (i = 0; i < args->nr_local; i++) {
708 struct rds_iovec *iov = &iovs[i];
709 /* don't need to check, rds_rdma_pages() verified nr will be +nonzero */
710 unsigned int nr = rds_pages_in_vec(iov);
711
712 rs->rs_user_addr = iov->addr;
713 rs->rs_user_bytes = iov->bytes;
714
715 /* If it's a WRITE operation, we want to pin the pages for reading.
716 * If it's a READ operation, we need to pin the pages for writing.
717 */
718 ret = rds_pin_pages(iov->addr, nr, pages, !op->op_write);
719 if ((!odp_supported && ret <= 0) ||
720 (odp_supported && ret <= 0 && ret != -EOPNOTSUPP))
721 goto out_pages;
722
723 if (ret == -EOPNOTSUPP) {
724 struct rds_mr *local_odp_mr;
725
726 if (!rs->rs_transport->get_mr) {
727 ret = -EOPNOTSUPP;
728 goto out_pages;
729 }
730 local_odp_mr =
731 kzalloc(sizeof(*local_odp_mr), GFP_KERNEL);
732 if (!local_odp_mr) {
733 ret = -ENOMEM;
734 goto out_pages;
735 }
736 RB_CLEAR_NODE(&local_odp_mr->r_rb_node);
737 kref_init(&local_odp_mr->r_kref);
738 local_odp_mr->r_trans = rs->rs_transport;
739 local_odp_mr->r_sock = rs;
740 local_odp_mr->r_trans_private =
741 rs->rs_transport->get_mr(
742 NULL, 0, rs, &local_odp_mr->r_key, NULL,
743 iov->addr, iov->bytes, ODP_VIRTUAL);
744 if (IS_ERR(local_odp_mr->r_trans_private)) {
745 ret = PTR_ERR(local_odp_mr->r_trans_private);
746 rdsdebug("get_mr ret %d %p\"", ret,
747 local_odp_mr->r_trans_private);
748 kfree(local_odp_mr);
749 ret = -EOPNOTSUPP;
750 goto out_pages;
751 }
752 rdsdebug("Need odp; local_odp_mr %p trans_private %p\n",
753 local_odp_mr, local_odp_mr->r_trans_private);
754 op->op_odp_mr = local_odp_mr;
755 op->op_odp_addr = iov->addr;
756 }
757
758 rdsdebug("RDS: nr_bytes %u nr %u iov->bytes %llu iov->addr %llx\n",
759 nr_bytes, nr, iov->bytes, iov->addr);
760
761 nr_bytes += iov->bytes;
762
763 for (j = 0; j < nr; j++) {
764 unsigned int offset = iov->addr & ~PAGE_MASK;
765 struct scatterlist *sg;
766
767 sg = &op->op_sg[op->op_nents + j];
768 sg_set_page(sg, pages[j],
769 min_t(unsigned int, iov->bytes, PAGE_SIZE - offset),
770 offset);
771
772 sg_dma_len(sg) = sg->length;
773 rdsdebug("RDS: sg->offset %x sg->len %x iov->addr %llx iov->bytes %llu\n",
774 sg->offset, sg->length, iov->addr, iov->bytes);
775
776 iov->addr += sg->length;
777 iov->bytes -= sg->length;
778 }
779
780 op->op_nents += nr;
781 }
782
783 if (nr_bytes > args->remote_vec.bytes) {
784 rdsdebug("RDS nr_bytes %u remote_bytes %u do not match\n",
785 nr_bytes,
786 (unsigned int) args->remote_vec.bytes);
787 ret = -EINVAL;
788 goto out_pages;
789 }
790 op->op_bytes = nr_bytes;
791 ret = 0;
792
793out_pages:
794 kfree(pages);
795out_ret:
796 if (ret)
797 rds_rdma_free_op(op);
798 else
799 rds_stats_inc(s_send_rdma);
800
801 return ret;
802}
803
804/*
805 * The application wants us to pass an RDMA destination (aka MR)
806 * to the remote
807 */
808int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
809 struct cmsghdr *cmsg)
810{
811 unsigned long flags;
812 struct rds_mr *mr;
813 u32 r_key;
814 int err = 0;
815
816 if (cmsg->cmsg_len < CMSG_LEN(sizeof(rds_rdma_cookie_t)) ||
817 rm->m_rdma_cookie != 0)
818 return -EINVAL;
819
820 memcpy(&rm->m_rdma_cookie, CMSG_DATA(cmsg), sizeof(rm->m_rdma_cookie));
821
822 /* We are reusing a previously mapped MR here. Most likely, the
823 * application has written to the buffer, so we need to explicitly
824 * flush those writes to RAM. Otherwise the HCA may not see them
825 * when doing a DMA from that buffer.
826 */
827 r_key = rds_rdma_cookie_key(rm->m_rdma_cookie);
828
829 spin_lock_irqsave(&rs->rs_rdma_lock, flags);
830 mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL);
831 if (!mr)
832 err = -EINVAL; /* invalid r_key */
833 else
834 kref_get(&mr->r_kref);
835 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
836
837 if (mr) {
838 mr->r_trans->sync_mr(mr->r_trans_private,
839 DMA_TO_DEVICE);
840 rm->rdma.op_rdma_mr = mr;
841 }
842 return err;
843}
844
845/*
846 * The application passes us an address range it wants to enable RDMA
847 * to/from. We map the area, and save the <R_Key,offset> pair
848 * in rm->m_rdma_cookie. This causes it to be sent along to the peer
849 * in an extension header.
850 */
851int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
852 struct cmsghdr *cmsg)
853{
854 if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_get_mr_args)) ||
855 rm->m_rdma_cookie != 0)
856 return -EINVAL;
857
858 return __rds_rdma_map(rs, CMSG_DATA(cmsg), &rm->m_rdma_cookie,
859 &rm->rdma.op_rdma_mr, rm->m_conn_path);
860}
861
862/*
863 * Fill in rds_message for an atomic request.
864 */
865int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
866 struct cmsghdr *cmsg)
867{
868 struct page *page = NULL;
869 struct rds_atomic_args *args;
870 int ret = 0;
871
872 if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_atomic_args))
873 || rm->atomic.op_active)
874 return -EINVAL;
875
876 args = CMSG_DATA(cmsg);
877
878 /* Nonmasked & masked cmsg ops converted to masked hw ops */
879 switch (cmsg->cmsg_type) {
880 case RDS_CMSG_ATOMIC_FADD:
881 rm->atomic.op_type = RDS_ATOMIC_TYPE_FADD;
882 rm->atomic.op_m_fadd.add = args->fadd.add;
883 rm->atomic.op_m_fadd.nocarry_mask = 0;
884 break;
885 case RDS_CMSG_MASKED_ATOMIC_FADD:
886 rm->atomic.op_type = RDS_ATOMIC_TYPE_FADD;
887 rm->atomic.op_m_fadd.add = args->m_fadd.add;
888 rm->atomic.op_m_fadd.nocarry_mask = args->m_fadd.nocarry_mask;
889 break;
890 case RDS_CMSG_ATOMIC_CSWP:
891 rm->atomic.op_type = RDS_ATOMIC_TYPE_CSWP;
892 rm->atomic.op_m_cswp.compare = args->cswp.compare;
893 rm->atomic.op_m_cswp.swap = args->cswp.swap;
894 rm->atomic.op_m_cswp.compare_mask = ~0;
895 rm->atomic.op_m_cswp.swap_mask = ~0;
896 break;
897 case RDS_CMSG_MASKED_ATOMIC_CSWP:
898 rm->atomic.op_type = RDS_ATOMIC_TYPE_CSWP;
899 rm->atomic.op_m_cswp.compare = args->m_cswp.compare;
900 rm->atomic.op_m_cswp.swap = args->m_cswp.swap;
901 rm->atomic.op_m_cswp.compare_mask = args->m_cswp.compare_mask;
902 rm->atomic.op_m_cswp.swap_mask = args->m_cswp.swap_mask;
903 break;
904 default:
905 BUG(); /* should never happen */
906 }
907
908 rm->atomic.op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME);
909 rm->atomic.op_silent = !!(args->flags & RDS_RDMA_SILENT);
910 rm->atomic.op_active = 1;
911 rm->atomic.op_recverr = rs->rs_recverr;
912 rm->atomic.op_sg = rds_message_alloc_sgs(rm, 1);
913 if (IS_ERR(rm->atomic.op_sg)) {
914 ret = PTR_ERR(rm->atomic.op_sg);
915 goto err;
916 }
917
918 /* verify 8 byte-aligned */
919 if (args->local_addr & 0x7) {
920 ret = -EFAULT;
921 goto err;
922 }
923
924 ret = rds_pin_pages(args->local_addr, 1, &page, 1);
925 if (ret != 1)
926 goto err;
927 ret = 0;
928
929 sg_set_page(rm->atomic.op_sg, page, 8, offset_in_page(args->local_addr));
930
931 if (rm->atomic.op_notify || rm->atomic.op_recverr) {
932 /* We allocate an uninitialized notifier here, because
933 * we don't want to do that in the completion handler. We
934 * would have to use GFP_ATOMIC there, and don't want to deal
935 * with failed allocations.
936 */
937 rm->atomic.op_notifier = kmalloc(sizeof(*rm->atomic.op_notifier), GFP_KERNEL);
938 if (!rm->atomic.op_notifier) {
939 ret = -ENOMEM;
940 goto err;
941 }
942
943 rm->atomic.op_notifier->n_user_token = args->user_token;
944 rm->atomic.op_notifier->n_status = RDS_RDMA_SUCCESS;
945 }
946
947 rm->atomic.op_rkey = rds_rdma_cookie_key(args->cookie);
948 rm->atomic.op_remote_addr = args->remote_addr + rds_rdma_cookie_offset(args->cookie);
949
950 return ret;
951err:
952 if (page)
953 unpin_user_page(page);
954 rm->atomic.op_active = 0;
955 kfree(rm->atomic.op_notifier);
956
957 return ret;
958}