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