]>
Commit | Line | Data |
---|---|---|
eff5f53b AG |
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> | |
5a0e3ad6 | 34 | #include <linux/slab.h> |
eff5f53b AG |
35 | #include <linux/rbtree.h> |
36 | #include <linux/dma-mapping.h> /* for DMA_*_DEVICE */ | |
37 | ||
21f79afa | 38 | #include "rds.h" |
eff5f53b AG |
39 | |
40 | /* | |
41 | * XXX | |
42 | * - build with sparse | |
eff5f53b AG |
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 | */ | |
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 | ||
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); | |
803ea850 | 87 | refcount_inc(&insert->r_refcount); |
eff5f53b AG |
88 | } |
89 | return NULL; | |
90 | } | |
91 | ||
92 | /* | |
93 | * Destroy the transport-specific part of a MR. | |
94 | */ | |
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", | |
803ea850 | 102 | mr->r_key, refcount_read(&mr->r_refcount)); |
eff5f53b AG |
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 | ||
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 | */ | |
128 | void rds_rdma_drop_keys(struct rds_sock *rs) | |
129 | { | |
130 | struct rds_mr *mr; | |
131 | struct rb_node *node; | |
35b52c70 | 132 | unsigned long flags; |
eff5f53b AG |
133 | |
134 | /* Release any MRs associated with this socket */ | |
35b52c70 | 135 | spin_lock_irqsave(&rs->rs_rdma_lock, flags); |
eff5f53b | 136 | while ((node = rb_first(&rs->rs_rdma_keys))) { |
a763f78c | 137 | mr = rb_entry(node, struct rds_mr, r_rb_node); |
eff5f53b AG |
138 | if (mr->r_trans == rs->rs_transport) |
139 | mr->r_invalidate = 0; | |
35b52c70 TY |
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); | |
eff5f53b | 144 | rds_mr_put(mr); |
35b52c70 | 145 | spin_lock_irqsave(&rs->rs_rdma_lock, flags); |
eff5f53b | 146 | } |
35b52c70 | 147 | spin_unlock_irqrestore(&rs->rs_rdma_lock, flags); |
eff5f53b AG |
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 | */ | |
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 | ||
830eb7d5 | 161 | ret = get_user_pages_fast(user_addr, nr_pages, write, pages); |
eff5f53b | 162 | |
7acd4a79 | 163 | if (ret >= 0 && ret < nr_pages) { |
eff5f53b AG |
164 | while (ret--) |
165 | put_page(pages[ret]); | |
166 | ret = -EFAULT; | |
167 | } | |
168 | ||
169 | return ret; | |
170 | } | |
171 | ||
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 | ||
f3069c6d | 186 | if (rs->rs_bound_addr == 0 || !rs->rs_transport) { |
eff5f53b AG |
187 | ret = -ENOTCONN; /* XXX not a great errno */ |
188 | goto out; | |
189 | } | |
190 | ||
8690bfa1 | 191 | if (!rs->rs_transport->get_mr) { |
eff5f53b AG |
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 | ||
f9fb69ad AR |
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 | ||
eff5f53b AG |
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); | |
8690bfa1 | 215 | if (!pages) { |
eff5f53b AG |
216 | ret = -ENOMEM; |
217 | goto out; | |
218 | } | |
219 | ||
220 | mr = kzalloc(sizeof(struct rds_mr), GFP_KERNEL); | |
8690bfa1 | 221 | if (!mr) { |
eff5f53b AG |
222 | ret = -ENOMEM; |
223 | goto out; | |
224 | } | |
225 | ||
803ea850 | 226 | refcount_set(&mr->r_refcount, 1); |
eff5f53b AG |
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 | */ | |
d22faec2 | 248 | ret = rds_pin_pages(args->vec.addr, nr_pages, pages, 1); |
eff5f53b AG |
249 | if (ret < 0) |
250 | goto out; | |
251 | ||
252 | nents = ret; | |
253 | sg = kcalloc(nents, sizeof(*sg), GFP_KERNEL); | |
8690bfa1 | 254 | if (!sg) { |
eff5f53b AG |
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) { | |
803ea850 | 310 | refcount_inc(&mr->r_refcount); |
eff5f53b AG |
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 | ||
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 | ||
244546f0 AG |
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 | ||
eff5f53b AG |
360 | /* |
361 | * Free the MR indicated by the given R_Key | |
362 | */ | |
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 | */ | |
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); | |
3ef13f3c | 424 | if (!mr) { |
c536a068 SS |
425 | pr_debug("rds: trying to unuse MR with unknown r_key %u!\n", |
426 | r_key); | |
3ef13f3c AG |
427 | spin_unlock_irqrestore(&rs->rs_rdma_lock, flags); |
428 | return; | |
429 | } | |
430 | ||
431 | if (mr->r_use_once || force) { | |
eff5f53b AG |
432 | rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys); |
433 | RB_CLEAR_NODE(&mr->r_rb_node); | |
434 | zot_me = 1; | |
3ef13f3c | 435 | } |
eff5f53b AG |
436 | spin_unlock_irqrestore(&rs->rs_rdma_lock, flags); |
437 | ||
438 | /* May have to issue a dma_sync on this memory region. | |
439 | * Note we could avoid this if the operation was a RDMA READ, | |
440 | * but at this point we can't tell. */ | |
3ef13f3c AG |
441 | if (mr->r_trans->sync_mr) |
442 | mr->r_trans->sync_mr(mr->r_trans_private, DMA_FROM_DEVICE); | |
eff5f53b | 443 | |
3ef13f3c AG |
444 | /* If the MR was marked as invalidate, this will |
445 | * trigger an async flush. */ | |
3f6b3143 | 446 | if (zot_me) { |
3ef13f3c | 447 | rds_destroy_mr(mr); |
3f6b3143 | 448 | rds_mr_put(mr); |
449 | } | |
eff5f53b AG |
450 | } |
451 | ||
f8b3aaf2 | 452 | void rds_rdma_free_op(struct rm_rdma_op *ro) |
eff5f53b AG |
453 | { |
454 | unsigned int i; | |
455 | ||
f8b3aaf2 AG |
456 | for (i = 0; i < ro->op_nents; i++) { |
457 | struct page *page = sg_page(&ro->op_sg[i]); | |
eff5f53b AG |
458 | |
459 | /* Mark page dirty if it was possibly modified, which | |
460 | * is the case for a RDMA_READ which copies from remote | |
461 | * to local memory */ | |
f8b3aaf2 | 462 | if (!ro->op_write) { |
5c240fa2 | 463 | WARN_ON(!page->mapping && irqs_disabled()); |
eff5f53b | 464 | set_page_dirty(page); |
561c7df6 | 465 | } |
eff5f53b AG |
466 | put_page(page); |
467 | } | |
468 | ||
f8b3aaf2 AG |
469 | kfree(ro->op_notifier); |
470 | ro->op_notifier = NULL; | |
471 | ro->op_active = 0; | |
ff87e97a AG |
472 | } |
473 | ||
d0ab25a8 AG |
474 | void rds_atomic_free_op(struct rm_atomic_op *ao) |
475 | { | |
476 | struct page *page = sg_page(ao->op_sg); | |
477 | ||
478 | /* Mark page dirty if it was possibly modified, which | |
479 | * is the case for a RDMA_READ which copies from remote | |
480 | * to local memory */ | |
481 | set_page_dirty(page); | |
482 | put_page(page); | |
483 | ||
484 | kfree(ao->op_notifier); | |
485 | ao->op_notifier = NULL; | |
486 | ao->op_active = 0; | |
487 | } | |
488 | ||
489 | ||
ff87e97a | 490 | /* |
fc8162e3 | 491 | * Count the number of pages needed to describe an incoming iovec array. |
ff87e97a | 492 | */ |
fc8162e3 AG |
493 | static int rds_rdma_pages(struct rds_iovec iov[], int nr_iovecs) |
494 | { | |
495 | int tot_pages = 0; | |
496 | unsigned int nr_pages; | |
497 | unsigned int i; | |
498 | ||
499 | /* figure out the number of pages in the vector */ | |
500 | for (i = 0; i < nr_iovecs; i++) { | |
501 | nr_pages = rds_pages_in_vec(&iov[i]); | |
502 | if (nr_pages == 0) | |
503 | return -EINVAL; | |
504 | ||
505 | tot_pages += nr_pages; | |
506 | ||
507 | /* | |
508 | * nr_pages for one entry is limited to (UINT_MAX>>PAGE_SHIFT)+1, | |
509 | * so tot_pages cannot overflow without first going negative. | |
510 | */ | |
511 | if (tot_pages < 0) | |
512 | return -EINVAL; | |
513 | } | |
514 | ||
515 | return tot_pages; | |
516 | } | |
517 | ||
74ef8bc0 | 518 | int rds_rdma_extra_size(struct rds_rdma_args *args, |
519 | struct rds_iov_vector *iov) | |
ff87e97a | 520 | { |
74ef8bc0 | 521 | struct rds_iovec *vec; |
ff87e97a | 522 | struct rds_iovec __user *local_vec; |
fc8162e3 | 523 | int tot_pages = 0; |
ff87e97a AG |
524 | unsigned int nr_pages; |
525 | unsigned int i; | |
526 | ||
527 | local_vec = (struct rds_iovec __user *)(unsigned long) args->local_vec_addr; | |
528 | ||
c0955087 MG |
529 | if (args->nr_local == 0) |
530 | return -EINVAL; | |
531 | ||
74ef8bc0 | 532 | iov->iov = kcalloc(args->nr_local, |
533 | sizeof(struct rds_iovec), | |
534 | GFP_KERNEL); | |
535 | if (!iov->iov) | |
536 | return -ENOMEM; | |
537 | ||
538 | vec = &iov->iov[0]; | |
539 | ||
540 | if (copy_from_user(vec, local_vec, args->nr_local * | |
541 | sizeof(struct rds_iovec))) | |
542 | return -EFAULT; | |
543 | iov->len = args->nr_local; | |
544 | ||
ff87e97a | 545 | /* figure out the number of pages in the vector */ |
74ef8bc0 | 546 | for (i = 0; i < args->nr_local; i++, vec++) { |
ff87e97a | 547 | |
74ef8bc0 | 548 | nr_pages = rds_pages_in_vec(vec); |
ff87e97a AG |
549 | if (nr_pages == 0) |
550 | return -EINVAL; | |
551 | ||
552 | tot_pages += nr_pages; | |
1b1f693d LT |
553 | |
554 | /* | |
555 | * nr_pages for one entry is limited to (UINT_MAX>>PAGE_SHIFT)+1, | |
556 | * so tot_pages cannot overflow without first going negative. | |
557 | */ | |
fc8162e3 | 558 | if (tot_pages < 0) |
1b1f693d | 559 | return -EINVAL; |
ff87e97a AG |
560 | } |
561 | ||
fc8162e3 | 562 | return tot_pages * sizeof(struct scatterlist); |
eff5f53b AG |
563 | } |
564 | ||
565 | /* | |
4324879d AG |
566 | * The application asks for a RDMA transfer. |
567 | * Extract all arguments and set up the rdma_op | |
eff5f53b | 568 | */ |
4324879d | 569 | int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm, |
74ef8bc0 | 570 | struct cmsghdr *cmsg, |
571 | struct rds_iov_vector *vec) | |
eff5f53b | 572 | { |
4324879d | 573 | struct rds_rdma_args *args; |
f8b3aaf2 | 574 | struct rm_rdma_op *op = &rm->rdma; |
9b9d2e00 | 575 | int nr_pages; |
eff5f53b AG |
576 | unsigned int nr_bytes; |
577 | struct page **pages = NULL; | |
74ef8bc0 | 578 | struct rds_iovec *iovs; |
eff5f53b | 579 | unsigned int i, j; |
ff87e97a | 580 | int ret = 0; |
eff5f53b | 581 | |
4324879d | 582 | if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_rdma_args)) |
f8b3aaf2 | 583 | || rm->rdma.op_active) |
4324879d AG |
584 | return -EINVAL; |
585 | ||
586 | args = CMSG_DATA(cmsg); | |
eff5f53b AG |
587 | |
588 | if (rs->rs_bound_addr == 0) { | |
589 | ret = -ENOTCONN; /* XXX not a great errno */ | |
dee49f20 | 590 | goto out_ret; |
eff5f53b AG |
591 | } |
592 | ||
218854af | 593 | if (args->nr_local > UIO_MAXIOV) { |
eff5f53b | 594 | ret = -EMSGSIZE; |
dee49f20 | 595 | goto out_ret; |
eff5f53b AG |
596 | } |
597 | ||
74ef8bc0 | 598 | if (vec->len != args->nr_local) { |
599 | ret = -EINVAL; | |
600 | goto out_ret; | |
fc8162e3 AG |
601 | } |
602 | ||
74ef8bc0 | 603 | iovs = vec->iov; |
fc8162e3 AG |
604 | |
605 | nr_pages = rds_rdma_pages(iovs, args->nr_local); | |
a09f69c4 AG |
606 | if (nr_pages < 0) { |
607 | ret = -EINVAL; | |
74ef8bc0 | 608 | goto out_ret; |
a09f69c4 | 609 | } |
eff5f53b | 610 | |
ff87e97a AG |
611 | pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL); |
612 | if (!pages) { | |
eff5f53b | 613 | ret = -ENOMEM; |
74ef8bc0 | 614 | goto out_ret; |
eff5f53b AG |
615 | } |
616 | ||
f8b3aaf2 AG |
617 | op->op_write = !!(args->flags & RDS_RDMA_READWRITE); |
618 | op->op_fence = !!(args->flags & RDS_RDMA_FENCE); | |
619 | op->op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME); | |
2c3a5f9a | 620 | op->op_silent = !!(args->flags & RDS_RDMA_SILENT); |
f8b3aaf2 AG |
621 | op->op_active = 1; |
622 | op->op_recverr = rs->rs_recverr; | |
eff5f53b | 623 | WARN_ON(!nr_pages); |
f8b3aaf2 | 624 | op->op_sg = rds_message_alloc_sgs(rm, nr_pages); |
d139ff09 AG |
625 | if (!op->op_sg) { |
626 | ret = -ENOMEM; | |
74ef8bc0 | 627 | goto out_pages; |
d139ff09 | 628 | } |
eff5f53b | 629 | |
f8b3aaf2 | 630 | if (op->op_notify || op->op_recverr) { |
eff5f53b AG |
631 | /* We allocate an uninitialized notifier here, because |
632 | * we don't want to do that in the completion handler. We | |
633 | * would have to use GFP_ATOMIC there, and don't want to deal | |
634 | * with failed allocations. | |
635 | */ | |
f8b3aaf2 AG |
636 | op->op_notifier = kmalloc(sizeof(struct rds_notifier), GFP_KERNEL); |
637 | if (!op->op_notifier) { | |
eff5f53b | 638 | ret = -ENOMEM; |
74ef8bc0 | 639 | goto out_pages; |
eff5f53b | 640 | } |
f8b3aaf2 AG |
641 | op->op_notifier->n_user_token = args->user_token; |
642 | op->op_notifier->n_status = RDS_RDMA_SUCCESS; | |
941f8d55 SS |
643 | |
644 | /* Enable rmda notification on data operation for composite | |
645 | * rds messages and make sure notification is enabled only | |
646 | * for the data operation which follows it so that application | |
647 | * gets notified only after full message gets delivered. | |
648 | */ | |
649 | if (rm->data.op_sg) { | |
650 | rm->rdma.op_notify = 0; | |
651 | rm->data.op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME); | |
652 | } | |
eff5f53b AG |
653 | } |
654 | ||
655 | /* The cookie contains the R_Key of the remote memory region, and | |
656 | * optionally an offset into it. This is how we implement RDMA into | |
657 | * unaligned memory. | |
658 | * When setting up the RDMA, we need to add that offset to the | |
659 | * destination address (which is really an offset into the MR) | |
660 | * FIXME: We may want to move this into ib_rdma.c | |
661 | */ | |
f8b3aaf2 AG |
662 | op->op_rkey = rds_rdma_cookie_key(args->cookie); |
663 | op->op_remote_addr = args->remote_vec.addr + rds_rdma_cookie_offset(args->cookie); | |
eff5f53b AG |
664 | |
665 | nr_bytes = 0; | |
666 | ||
667 | rdsdebug("RDS: rdma prepare nr_local %llu rva %llx rkey %x\n", | |
668 | (unsigned long long)args->nr_local, | |
669 | (unsigned long long)args->remote_vec.addr, | |
f8b3aaf2 | 670 | op->op_rkey); |
eff5f53b AG |
671 | |
672 | for (i = 0; i < args->nr_local; i++) { | |
fc8162e3 AG |
673 | struct rds_iovec *iov = &iovs[i]; |
674 | /* don't need to check, rds_rdma_pages() verified nr will be +nonzero */ | |
675 | unsigned int nr = rds_pages_in_vec(iov); | |
eff5f53b | 676 | |
fc8162e3 AG |
677 | rs->rs_user_addr = iov->addr; |
678 | rs->rs_user_bytes = iov->bytes; | |
eff5f53b | 679 | |
eff5f53b AG |
680 | /* If it's a WRITE operation, we want to pin the pages for reading. |
681 | * If it's a READ operation, we need to pin the pages for writing. | |
682 | */ | |
fc8162e3 | 683 | ret = rds_pin_pages(iov->addr, nr, pages, !op->op_write); |
eff5f53b | 684 | if (ret < 0) |
74ef8bc0 | 685 | goto out_pages; |
1d2e3f39 | 686 | else |
687 | ret = 0; | |
eff5f53b | 688 | |
fc8162e3 AG |
689 | rdsdebug("RDS: nr_bytes %u nr %u iov->bytes %llu iov->addr %llx\n", |
690 | nr_bytes, nr, iov->bytes, iov->addr); | |
eff5f53b | 691 | |
fc8162e3 | 692 | nr_bytes += iov->bytes; |
eff5f53b AG |
693 | |
694 | for (j = 0; j < nr; j++) { | |
fc8162e3 | 695 | unsigned int offset = iov->addr & ~PAGE_MASK; |
ff87e97a | 696 | struct scatterlist *sg; |
eff5f53b | 697 | |
f8b3aaf2 | 698 | sg = &op->op_sg[op->op_nents + j]; |
eff5f53b | 699 | sg_set_page(sg, pages[j], |
fc8162e3 | 700 | min_t(unsigned int, iov->bytes, PAGE_SIZE - offset), |
eff5f53b AG |
701 | offset); |
702 | ||
fc8162e3 AG |
703 | rdsdebug("RDS: sg->offset %x sg->len %x iov->addr %llx iov->bytes %llu\n", |
704 | sg->offset, sg->length, iov->addr, iov->bytes); | |
eff5f53b | 705 | |
fc8162e3 AG |
706 | iov->addr += sg->length; |
707 | iov->bytes -= sg->length; | |
eff5f53b AG |
708 | } |
709 | ||
f8b3aaf2 | 710 | op->op_nents += nr; |
eff5f53b AG |
711 | } |
712 | ||
eff5f53b AG |
713 | if (nr_bytes > args->remote_vec.bytes) { |
714 | rdsdebug("RDS nr_bytes %u remote_bytes %u do not match\n", | |
715 | nr_bytes, | |
716 | (unsigned int) args->remote_vec.bytes); | |
717 | ret = -EINVAL; | |
74ef8bc0 | 718 | goto out_pages; |
eff5f53b | 719 | } |
f8b3aaf2 | 720 | op->op_bytes = nr_bytes; |
eff5f53b | 721 | |
74ef8bc0 | 722 | out_pages: |
eff5f53b | 723 | kfree(pages); |
dee49f20 | 724 | out_ret: |
ff87e97a AG |
725 | if (ret) |
726 | rds_rdma_free_op(op); | |
f4a3fc03 AG |
727 | else |
728 | rds_stats_inc(s_send_rdma); | |
4324879d AG |
729 | |
730 | return ret; | |
eff5f53b AG |
731 | } |
732 | ||
733 | /* | |
734 | * The application wants us to pass an RDMA destination (aka MR) | |
735 | * to the remote | |
736 | */ | |
737 | int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm, | |
738 | struct cmsghdr *cmsg) | |
739 | { | |
740 | unsigned long flags; | |
741 | struct rds_mr *mr; | |
742 | u32 r_key; | |
743 | int err = 0; | |
744 | ||
f64f9e71 JP |
745 | if (cmsg->cmsg_len < CMSG_LEN(sizeof(rds_rdma_cookie_t)) || |
746 | rm->m_rdma_cookie != 0) | |
eff5f53b AG |
747 | return -EINVAL; |
748 | ||
749 | memcpy(&rm->m_rdma_cookie, CMSG_DATA(cmsg), sizeof(rm->m_rdma_cookie)); | |
750 | ||
751 | /* We are reusing a previously mapped MR here. Most likely, the | |
752 | * application has written to the buffer, so we need to explicitly | |
753 | * flush those writes to RAM. Otherwise the HCA may not see them | |
754 | * when doing a DMA from that buffer. | |
755 | */ | |
756 | r_key = rds_rdma_cookie_key(rm->m_rdma_cookie); | |
757 | ||
758 | spin_lock_irqsave(&rs->rs_rdma_lock, flags); | |
759 | mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL); | |
8690bfa1 | 760 | if (!mr) |
eff5f53b AG |
761 | err = -EINVAL; /* invalid r_key */ |
762 | else | |
803ea850 | 763 | refcount_inc(&mr->r_refcount); |
eff5f53b AG |
764 | spin_unlock_irqrestore(&rs->rs_rdma_lock, flags); |
765 | ||
766 | if (mr) { | |
767 | mr->r_trans->sync_mr(mr->r_trans_private, DMA_TO_DEVICE); | |
f8b3aaf2 | 768 | rm->rdma.op_rdma_mr = mr; |
eff5f53b AG |
769 | } |
770 | return err; | |
771 | } | |
772 | ||
773 | /* | |
774 | * The application passes us an address range it wants to enable RDMA | |
775 | * to/from. We map the area, and save the <R_Key,offset> pair | |
776 | * in rm->m_rdma_cookie. This causes it to be sent along to the peer | |
777 | * in an extension header. | |
778 | */ | |
779 | int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm, | |
780 | struct cmsghdr *cmsg) | |
781 | { | |
f64f9e71 JP |
782 | if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_get_mr_args)) || |
783 | rm->m_rdma_cookie != 0) | |
eff5f53b AG |
784 | return -EINVAL; |
785 | ||
f8b3aaf2 | 786 | return __rds_rdma_map(rs, CMSG_DATA(cmsg), &rm->m_rdma_cookie, &rm->rdma.op_rdma_mr); |
eff5f53b | 787 | } |
15133f6e AG |
788 | |
789 | /* | |
790 | * Fill in rds_message for an atomic request. | |
791 | */ | |
792 | int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm, | |
793 | struct cmsghdr *cmsg) | |
794 | { | |
795 | struct page *page = NULL; | |
796 | struct rds_atomic_args *args; | |
797 | int ret = 0; | |
798 | ||
799 | if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_atomic_args)) | |
800 | || rm->atomic.op_active) | |
801 | return -EINVAL; | |
802 | ||
803 | args = CMSG_DATA(cmsg); | |
804 | ||
20c72bd5 AG |
805 | /* Nonmasked & masked cmsg ops converted to masked hw ops */ |
806 | switch (cmsg->cmsg_type) { | |
807 | case RDS_CMSG_ATOMIC_FADD: | |
808 | rm->atomic.op_type = RDS_ATOMIC_TYPE_FADD; | |
809 | rm->atomic.op_m_fadd.add = args->fadd.add; | |
810 | rm->atomic.op_m_fadd.nocarry_mask = 0; | |
811 | break; | |
812 | case RDS_CMSG_MASKED_ATOMIC_FADD: | |
15133f6e | 813 | rm->atomic.op_type = RDS_ATOMIC_TYPE_FADD; |
20c72bd5 AG |
814 | rm->atomic.op_m_fadd.add = args->m_fadd.add; |
815 | rm->atomic.op_m_fadd.nocarry_mask = args->m_fadd.nocarry_mask; | |
816 | break; | |
817 | case RDS_CMSG_ATOMIC_CSWP: | |
818 | rm->atomic.op_type = RDS_ATOMIC_TYPE_CSWP; | |
819 | rm->atomic.op_m_cswp.compare = args->cswp.compare; | |
820 | rm->atomic.op_m_cswp.swap = args->cswp.swap; | |
821 | rm->atomic.op_m_cswp.compare_mask = ~0; | |
822 | rm->atomic.op_m_cswp.swap_mask = ~0; | |
823 | break; | |
824 | case RDS_CMSG_MASKED_ATOMIC_CSWP: | |
825 | rm->atomic.op_type = RDS_ATOMIC_TYPE_CSWP; | |
826 | rm->atomic.op_m_cswp.compare = args->m_cswp.compare; | |
827 | rm->atomic.op_m_cswp.swap = args->m_cswp.swap; | |
828 | rm->atomic.op_m_cswp.compare_mask = args->m_cswp.compare_mask; | |
829 | rm->atomic.op_m_cswp.swap_mask = args->m_cswp.swap_mask; | |
830 | break; | |
831 | default: | |
832 | BUG(); /* should never happen */ | |
15133f6e AG |
833 | } |
834 | ||
15133f6e | 835 | rm->atomic.op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME); |
2c3a5f9a | 836 | rm->atomic.op_silent = !!(args->flags & RDS_RDMA_SILENT); |
7e3bd65e | 837 | rm->atomic.op_active = 1; |
15133f6e AG |
838 | rm->atomic.op_recverr = rs->rs_recverr; |
839 | rm->atomic.op_sg = rds_message_alloc_sgs(rm, 1); | |
d139ff09 AG |
840 | if (!rm->atomic.op_sg) { |
841 | ret = -ENOMEM; | |
842 | goto err; | |
843 | } | |
15133f6e AG |
844 | |
845 | /* verify 8 byte-aligned */ | |
846 | if (args->local_addr & 0x7) { | |
847 | ret = -EFAULT; | |
848 | goto err; | |
849 | } | |
850 | ||
851 | ret = rds_pin_pages(args->local_addr, 1, &page, 1); | |
852 | if (ret != 1) | |
853 | goto err; | |
854 | ret = 0; | |
855 | ||
856 | sg_set_page(rm->atomic.op_sg, page, 8, offset_in_page(args->local_addr)); | |
857 | ||
858 | if (rm->atomic.op_notify || rm->atomic.op_recverr) { | |
859 | /* We allocate an uninitialized notifier here, because | |
860 | * we don't want to do that in the completion handler. We | |
861 | * would have to use GFP_ATOMIC there, and don't want to deal | |
862 | * with failed allocations. | |
863 | */ | |
864 | rm->atomic.op_notifier = kmalloc(sizeof(*rm->atomic.op_notifier), GFP_KERNEL); | |
865 | if (!rm->atomic.op_notifier) { | |
866 | ret = -ENOMEM; | |
867 | goto err; | |
868 | } | |
869 | ||
870 | rm->atomic.op_notifier->n_user_token = args->user_token; | |
871 | rm->atomic.op_notifier->n_status = RDS_RDMA_SUCCESS; | |
872 | } | |
873 | ||
40589e74 | 874 | rm->atomic.op_rkey = rds_rdma_cookie_key(args->cookie); |
15133f6e AG |
875 | rm->atomic.op_remote_addr = args->remote_addr + rds_rdma_cookie_offset(args->cookie); |
876 | ||
15133f6e AG |
877 | return ret; |
878 | err: | |
879 | if (page) | |
880 | put_page(page); | |
7d11f77f | 881 | rm->atomic.op_active = 0; |
15133f6e AG |
882 | kfree(rm->atomic.op_notifier); |
883 | ||
884 | return ret; | |
885 | } |