]>
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 | |
43 | * - should we limit the size of a mr region? let transport return failure? | |
44 | * - should we detect duplicate keys on a socket? hmm. | |
45 | * - an rdma is an mlock, apply rlimit? | |
46 | */ | |
47 | ||
48 | /* | |
49 | * get the number of pages by looking at the page indices that the start and | |
50 | * end addresses fall in. | |
51 | * | |
52 | * Returns 0 if the vec is invalid. It is invalid if the number of bytes | |
53 | * causes the address to wrap or overflows an unsigned int. This comes | |
54 | * from being stored in the 'length' member of 'struct scatterlist'. | |
55 | */ | |
56 | static unsigned int rds_pages_in_vec(struct rds_iovec *vec) | |
57 | { | |
58 | if ((vec->addr + vec->bytes <= vec->addr) || | |
59 | (vec->bytes > (u64)UINT_MAX)) | |
60 | return 0; | |
61 | ||
62 | return ((vec->addr + vec->bytes + PAGE_SIZE - 1) >> PAGE_SHIFT) - | |
63 | (vec->addr >> PAGE_SHIFT); | |
64 | } | |
65 | ||
66 | static struct rds_mr *rds_mr_tree_walk(struct rb_root *root, u64 key, | |
67 | struct rds_mr *insert) | |
68 | { | |
69 | struct rb_node **p = &root->rb_node; | |
70 | struct rb_node *parent = NULL; | |
71 | struct rds_mr *mr; | |
72 | ||
73 | while (*p) { | |
74 | parent = *p; | |
75 | mr = rb_entry(parent, struct rds_mr, r_rb_node); | |
76 | ||
77 | if (key < mr->r_key) | |
78 | p = &(*p)->rb_left; | |
79 | else if (key > mr->r_key) | |
80 | p = &(*p)->rb_right; | |
81 | else | |
82 | return mr; | |
83 | } | |
84 | ||
85 | if (insert) { | |
86 | rb_link_node(&insert->r_rb_node, parent, p); | |
87 | rb_insert_color(&insert->r_rb_node, root); | |
88 | atomic_inc(&insert->r_refcount); | |
89 | } | |
90 | return NULL; | |
91 | } | |
92 | ||
93 | /* | |
94 | * Destroy the transport-specific part of a MR. | |
95 | */ | |
96 | static void rds_destroy_mr(struct rds_mr *mr) | |
97 | { | |
98 | struct rds_sock *rs = mr->r_sock; | |
99 | void *trans_private = NULL; | |
100 | unsigned long flags; | |
101 | ||
102 | rdsdebug("RDS: destroy mr key is %x refcnt %u\n", | |
103 | mr->r_key, atomic_read(&mr->r_refcount)); | |
104 | ||
105 | if (test_and_set_bit(RDS_MR_DEAD, &mr->r_state)) | |
106 | return; | |
107 | ||
108 | spin_lock_irqsave(&rs->rs_rdma_lock, flags); | |
109 | if (!RB_EMPTY_NODE(&mr->r_rb_node)) | |
110 | rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys); | |
111 | trans_private = mr->r_trans_private; | |
112 | mr->r_trans_private = NULL; | |
113 | spin_unlock_irqrestore(&rs->rs_rdma_lock, flags); | |
114 | ||
115 | if (trans_private) | |
116 | mr->r_trans->free_mr(trans_private, mr->r_invalidate); | |
117 | } | |
118 | ||
119 | void __rds_put_mr_final(struct rds_mr *mr) | |
120 | { | |
121 | rds_destroy_mr(mr); | |
122 | kfree(mr); | |
123 | } | |
124 | ||
125 | /* | |
126 | * By the time this is called we can't have any more ioctls called on | |
127 | * the socket so we don't need to worry about racing with others. | |
128 | */ | |
129 | void rds_rdma_drop_keys(struct rds_sock *rs) | |
130 | { | |
131 | struct rds_mr *mr; | |
132 | struct rb_node *node; | |
35b52c70 | 133 | unsigned long flags; |
eff5f53b AG |
134 | |
135 | /* Release any MRs associated with this socket */ | |
35b52c70 | 136 | spin_lock_irqsave(&rs->rs_rdma_lock, flags); |
eff5f53b AG |
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; | |
35b52c70 TY |
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); | |
eff5f53b | 145 | rds_mr_put(mr); |
35b52c70 | 146 | spin_lock_irqsave(&rs->rs_rdma_lock, flags); |
eff5f53b | 147 | } |
35b52c70 | 148 | spin_unlock_irqrestore(&rs->rs_rdma_lock, flags); |
eff5f53b AG |
149 | |
150 | if (rs->rs_transport && rs->rs_transport->flush_mrs) | |
151 | rs->rs_transport->flush_mrs(); | |
152 | } | |
153 | ||
154 | /* | |
155 | * Helper function to pin user pages. | |
156 | */ | |
157 | static int rds_pin_pages(unsigned long user_addr, unsigned int nr_pages, | |
158 | struct page **pages, int write) | |
159 | { | |
160 | int ret; | |
161 | ||
830eb7d5 | 162 | ret = get_user_pages_fast(user_addr, nr_pages, write, pages); |
eff5f53b | 163 | |
7acd4a79 | 164 | if (ret >= 0 && ret < nr_pages) { |
eff5f53b AG |
165 | while (ret--) |
166 | put_page(pages[ret]); | |
167 | ret = -EFAULT; | |
168 | } | |
169 | ||
170 | return ret; | |
171 | } | |
172 | ||
173 | static int __rds_rdma_map(struct rds_sock *rs, struct rds_get_mr_args *args, | |
174 | u64 *cookie_ret, struct rds_mr **mr_ret) | |
175 | { | |
176 | struct rds_mr *mr = NULL, *found; | |
177 | unsigned int nr_pages; | |
178 | struct page **pages = NULL; | |
179 | struct scatterlist *sg; | |
180 | void *trans_private; | |
181 | unsigned long flags; | |
182 | rds_rdma_cookie_t cookie; | |
183 | unsigned int nents; | |
184 | long i; | |
185 | int ret; | |
186 | ||
187 | if (rs->rs_bound_addr == 0) { | |
188 | ret = -ENOTCONN; /* XXX not a great errno */ | |
189 | goto out; | |
190 | } | |
191 | ||
8690bfa1 | 192 | if (!rs->rs_transport->get_mr) { |
eff5f53b AG |
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); | |
8690bfa1 | 208 | if (!pages) { |
eff5f53b AG |
209 | ret = -ENOMEM; |
210 | goto out; | |
211 | } | |
212 | ||
213 | mr = kzalloc(sizeof(struct rds_mr), GFP_KERNEL); | |
8690bfa1 | 214 | if (!mr) { |
eff5f53b AG |
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 | */ | |
d22faec2 | 241 | ret = rds_pin_pages(args->vec.addr, nr_pages, pages, 1); |
eff5f53b AG |
242 | if (ret < 0) |
243 | goto out; | |
244 | ||
245 | nents = ret; | |
246 | sg = kcalloc(nents, sizeof(*sg), GFP_KERNEL); | |
8690bfa1 | 247 | if (!sg) { |
eff5f53b AG |
248 | ret = -ENOMEM; |
249 | goto out; | |
250 | } | |
251 | WARN_ON(!nents); | |
252 | sg_init_table(sg, nents); | |
253 | ||
254 | /* Stick all pages into the scatterlist */ | |
255 | for (i = 0 ; i < nents; i++) | |
256 | sg_set_page(&sg[i], pages[i], PAGE_SIZE, 0); | |
257 | ||
258 | rdsdebug("RDS: trans_private nents is %u\n", nents); | |
259 | ||
260 | /* Obtain a transport specific MR. If this succeeds, the | |
261 | * s/g list is now owned by the MR. | |
262 | * Note that dma_map() implies that pending writes are | |
263 | * flushed to RAM, so no dma_sync is needed here. */ | |
264 | trans_private = rs->rs_transport->get_mr(sg, nents, rs, | |
265 | &mr->r_key); | |
266 | ||
267 | if (IS_ERR(trans_private)) { | |
268 | for (i = 0 ; i < nents; i++) | |
269 | put_page(sg_page(&sg[i])); | |
270 | kfree(sg); | |
271 | ret = PTR_ERR(trans_private); | |
272 | goto out; | |
273 | } | |
274 | ||
275 | mr->r_trans_private = trans_private; | |
276 | ||
277 | rdsdebug("RDS: get_mr put_user key is %x cookie_addr %p\n", | |
278 | mr->r_key, (void *)(unsigned long) args->cookie_addr); | |
279 | ||
280 | /* The user may pass us an unaligned address, but we can only | |
281 | * map page aligned regions. So we keep the offset, and build | |
282 | * a 64bit cookie containing <R_Key, offset> and pass that | |
283 | * around. */ | |
284 | cookie = rds_rdma_make_cookie(mr->r_key, args->vec.addr & ~PAGE_MASK); | |
285 | if (cookie_ret) | |
286 | *cookie_ret = cookie; | |
287 | ||
288 | if (args->cookie_addr && put_user(cookie, (u64 __user *)(unsigned long) args->cookie_addr)) { | |
289 | ret = -EFAULT; | |
290 | goto out; | |
291 | } | |
292 | ||
293 | /* Inserting the new MR into the rbtree bumps its | |
294 | * reference count. */ | |
295 | spin_lock_irqsave(&rs->rs_rdma_lock, flags); | |
296 | found = rds_mr_tree_walk(&rs->rs_rdma_keys, mr->r_key, mr); | |
297 | spin_unlock_irqrestore(&rs->rs_rdma_lock, flags); | |
298 | ||
299 | BUG_ON(found && found != mr); | |
300 | ||
301 | rdsdebug("RDS: get_mr key is %x\n", mr->r_key); | |
302 | if (mr_ret) { | |
303 | atomic_inc(&mr->r_refcount); | |
304 | *mr_ret = mr; | |
305 | } | |
306 | ||
307 | ret = 0; | |
308 | out: | |
309 | kfree(pages); | |
310 | if (mr) | |
311 | rds_mr_put(mr); | |
312 | return ret; | |
313 | } | |
314 | ||
315 | int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen) | |
316 | { | |
317 | struct rds_get_mr_args args; | |
318 | ||
319 | if (optlen != sizeof(struct rds_get_mr_args)) | |
320 | return -EINVAL; | |
321 | ||
322 | if (copy_from_user(&args, (struct rds_get_mr_args __user *)optval, | |
323 | sizeof(struct rds_get_mr_args))) | |
324 | return -EFAULT; | |
325 | ||
326 | return __rds_rdma_map(rs, &args, NULL, NULL); | |
327 | } | |
328 | ||
244546f0 AG |
329 | int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen) |
330 | { | |
331 | struct rds_get_mr_for_dest_args args; | |
332 | struct rds_get_mr_args new_args; | |
333 | ||
334 | if (optlen != sizeof(struct rds_get_mr_for_dest_args)) | |
335 | return -EINVAL; | |
336 | ||
337 | if (copy_from_user(&args, (struct rds_get_mr_for_dest_args __user *)optval, | |
338 | sizeof(struct rds_get_mr_for_dest_args))) | |
339 | return -EFAULT; | |
340 | ||
341 | /* | |
342 | * Initially, just behave like get_mr(). | |
343 | * TODO: Implement get_mr as wrapper around this | |
344 | * and deprecate it. | |
345 | */ | |
346 | new_args.vec = args.vec; | |
347 | new_args.cookie_addr = args.cookie_addr; | |
348 | new_args.flags = args.flags; | |
349 | ||
350 | return __rds_rdma_map(rs, &new_args, NULL, NULL); | |
351 | } | |
352 | ||
eff5f53b AG |
353 | /* |
354 | * Free the MR indicated by the given R_Key | |
355 | */ | |
356 | int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen) | |
357 | { | |
358 | struct rds_free_mr_args args; | |
359 | struct rds_mr *mr; | |
360 | unsigned long flags; | |
361 | ||
362 | if (optlen != sizeof(struct rds_free_mr_args)) | |
363 | return -EINVAL; | |
364 | ||
365 | if (copy_from_user(&args, (struct rds_free_mr_args __user *)optval, | |
366 | sizeof(struct rds_free_mr_args))) | |
367 | return -EFAULT; | |
368 | ||
369 | /* Special case - a null cookie means flush all unused MRs */ | |
370 | if (args.cookie == 0) { | |
371 | if (!rs->rs_transport || !rs->rs_transport->flush_mrs) | |
372 | return -EINVAL; | |
373 | rs->rs_transport->flush_mrs(); | |
374 | return 0; | |
375 | } | |
376 | ||
377 | /* Look up the MR given its R_key and remove it from the rbtree | |
378 | * so nobody else finds it. | |
379 | * This should also prevent races with rds_rdma_unuse. | |
380 | */ | |
381 | spin_lock_irqsave(&rs->rs_rdma_lock, flags); | |
382 | mr = rds_mr_tree_walk(&rs->rs_rdma_keys, rds_rdma_cookie_key(args.cookie), NULL); | |
383 | if (mr) { | |
384 | rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys); | |
385 | RB_CLEAR_NODE(&mr->r_rb_node); | |
386 | if (args.flags & RDS_RDMA_INVALIDATE) | |
387 | mr->r_invalidate = 1; | |
388 | } | |
389 | spin_unlock_irqrestore(&rs->rs_rdma_lock, flags); | |
390 | ||
391 | if (!mr) | |
392 | return -EINVAL; | |
393 | ||
394 | /* | |
395 | * call rds_destroy_mr() ourselves so that we're sure it's done by the time | |
396 | * we return. If we let rds_mr_put() do it it might not happen until | |
397 | * someone else drops their ref. | |
398 | */ | |
399 | rds_destroy_mr(mr); | |
400 | rds_mr_put(mr); | |
401 | return 0; | |
402 | } | |
403 | ||
404 | /* | |
405 | * This is called when we receive an extension header that | |
406 | * tells us this MR was used. It allows us to implement | |
407 | * use_once semantics | |
408 | */ | |
409 | void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force) | |
410 | { | |
411 | struct rds_mr *mr; | |
412 | unsigned long flags; | |
413 | int zot_me = 0; | |
414 | ||
415 | spin_lock_irqsave(&rs->rs_rdma_lock, flags); | |
416 | mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL); | |
3ef13f3c AG |
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) { | |
eff5f53b AG |
424 | rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys); |
425 | RB_CLEAR_NODE(&mr->r_rb_node); | |
426 | zot_me = 1; | |
3ef13f3c | 427 | } |
eff5f53b AG |
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. */ | |
3ef13f3c AG |
433 | if (mr->r_trans->sync_mr) |
434 | mr->r_trans->sync_mr(mr->r_trans_private, DMA_FROM_DEVICE); | |
eff5f53b | 435 | |
3ef13f3c AG |
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); | |
eff5f53b AG |
441 | } |
442 | ||
f8b3aaf2 | 443 | void rds_rdma_free_op(struct rm_rdma_op *ro) |
eff5f53b AG |
444 | { |
445 | unsigned int i; | |
446 | ||
f8b3aaf2 AG |
447 | for (i = 0; i < ro->op_nents; i++) { |
448 | struct page *page = sg_page(&ro->op_sg[i]); | |
eff5f53b AG |
449 | |
450 | /* Mark page dirty if it was possibly modified, which | |
451 | * is the case for a RDMA_READ which copies from remote | |
452 | * to local memory */ | |
f8b3aaf2 | 453 | if (!ro->op_write) { |
9e2effba | 454 | BUG_ON(irqs_disabled()); |
eff5f53b | 455 | set_page_dirty(page); |
561c7df6 | 456 | } |
eff5f53b AG |
457 | put_page(page); |
458 | } | |
459 | ||
f8b3aaf2 AG |
460 | kfree(ro->op_notifier); |
461 | ro->op_notifier = NULL; | |
462 | ro->op_active = 0; | |
ff87e97a AG |
463 | } |
464 | ||
d0ab25a8 AG |
465 | void rds_atomic_free_op(struct rm_atomic_op *ao) |
466 | { | |
467 | struct page *page = sg_page(ao->op_sg); | |
468 | ||
469 | /* Mark page dirty if it was possibly modified, which | |
470 | * is the case for a RDMA_READ which copies from remote | |
471 | * to local memory */ | |
472 | set_page_dirty(page); | |
473 | put_page(page); | |
474 | ||
475 | kfree(ao->op_notifier); | |
476 | ao->op_notifier = NULL; | |
477 | ao->op_active = 0; | |
478 | } | |
479 | ||
480 | ||
ff87e97a AG |
481 | /* |
482 | * Count the number of pages needed to describe an incoming iovec. | |
483 | */ | |
484 | static int rds_rdma_pages(struct rds_rdma_args *args) | |
485 | { | |
486 | struct rds_iovec vec; | |
487 | struct rds_iovec __user *local_vec; | |
488 | unsigned int tot_pages = 0; | |
489 | unsigned int nr_pages; | |
490 | unsigned int i; | |
491 | ||
492 | local_vec = (struct rds_iovec __user *)(unsigned long) args->local_vec_addr; | |
493 | ||
494 | /* figure out the number of pages in the vector */ | |
495 | for (i = 0; i < args->nr_local; i++) { | |
496 | if (copy_from_user(&vec, &local_vec[i], | |
497 | sizeof(struct rds_iovec))) | |
498 | return -EFAULT; | |
499 | ||
500 | nr_pages = rds_pages_in_vec(&vec); | |
501 | if (nr_pages == 0) | |
502 | return -EINVAL; | |
503 | ||
504 | tot_pages += nr_pages; | |
505 | } | |
506 | ||
507 | return tot_pages; | |
508 | } | |
509 | ||
510 | int rds_rdma_extra_size(struct rds_rdma_args *args) | |
511 | { | |
512 | return rds_rdma_pages(args) * sizeof(struct scatterlist); | |
eff5f53b AG |
513 | } |
514 | ||
515 | /* | |
4324879d AG |
516 | * The application asks for a RDMA transfer. |
517 | * Extract all arguments and set up the rdma_op | |
eff5f53b | 518 | */ |
4324879d AG |
519 | int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm, |
520 | struct cmsghdr *cmsg) | |
eff5f53b | 521 | { |
4324879d | 522 | struct rds_rdma_args *args; |
eff5f53b | 523 | struct rds_iovec vec; |
f8b3aaf2 | 524 | struct rm_rdma_op *op = &rm->rdma; |
9b9d2e00 | 525 | int nr_pages; |
eff5f53b AG |
526 | unsigned int nr_bytes; |
527 | struct page **pages = NULL; | |
528 | struct rds_iovec __user *local_vec; | |
eff5f53b AG |
529 | unsigned int nr; |
530 | unsigned int i, j; | |
ff87e97a | 531 | int ret = 0; |
eff5f53b | 532 | |
4324879d | 533 | if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_rdma_args)) |
f8b3aaf2 | 534 | || rm->rdma.op_active) |
4324879d AG |
535 | return -EINVAL; |
536 | ||
537 | args = CMSG_DATA(cmsg); | |
eff5f53b AG |
538 | |
539 | if (rs->rs_bound_addr == 0) { | |
540 | ret = -ENOTCONN; /* XXX not a great errno */ | |
541 | goto out; | |
542 | } | |
543 | ||
544 | if (args->nr_local > (u64)UINT_MAX) { | |
545 | ret = -EMSGSIZE; | |
546 | goto out; | |
547 | } | |
548 | ||
ff87e97a AG |
549 | nr_pages = rds_rdma_pages(args); |
550 | if (nr_pages < 0) | |
eff5f53b | 551 | goto out; |
eff5f53b | 552 | |
ff87e97a AG |
553 | pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL); |
554 | if (!pages) { | |
eff5f53b AG |
555 | ret = -ENOMEM; |
556 | goto out; | |
557 | } | |
558 | ||
f8b3aaf2 AG |
559 | op->op_write = !!(args->flags & RDS_RDMA_READWRITE); |
560 | op->op_fence = !!(args->flags & RDS_RDMA_FENCE); | |
561 | op->op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME); | |
2c3a5f9a | 562 | op->op_silent = !!(args->flags & RDS_RDMA_SILENT); |
f8b3aaf2 AG |
563 | op->op_active = 1; |
564 | op->op_recverr = rs->rs_recverr; | |
eff5f53b | 565 | WARN_ON(!nr_pages); |
f8b3aaf2 | 566 | op->op_sg = rds_message_alloc_sgs(rm, nr_pages); |
eff5f53b | 567 | |
f8b3aaf2 | 568 | if (op->op_notify || op->op_recverr) { |
eff5f53b AG |
569 | /* We allocate an uninitialized notifier here, because |
570 | * we don't want to do that in the completion handler. We | |
571 | * would have to use GFP_ATOMIC there, and don't want to deal | |
572 | * with failed allocations. | |
573 | */ | |
f8b3aaf2 AG |
574 | op->op_notifier = kmalloc(sizeof(struct rds_notifier), GFP_KERNEL); |
575 | if (!op->op_notifier) { | |
eff5f53b AG |
576 | ret = -ENOMEM; |
577 | goto out; | |
578 | } | |
f8b3aaf2 AG |
579 | op->op_notifier->n_user_token = args->user_token; |
580 | op->op_notifier->n_status = RDS_RDMA_SUCCESS; | |
eff5f53b AG |
581 | } |
582 | ||
583 | /* The cookie contains the R_Key of the remote memory region, and | |
584 | * optionally an offset into it. This is how we implement RDMA into | |
585 | * unaligned memory. | |
586 | * When setting up the RDMA, we need to add that offset to the | |
587 | * destination address (which is really an offset into the MR) | |
588 | * FIXME: We may want to move this into ib_rdma.c | |
589 | */ | |
f8b3aaf2 AG |
590 | op->op_rkey = rds_rdma_cookie_key(args->cookie); |
591 | op->op_remote_addr = args->remote_vec.addr + rds_rdma_cookie_offset(args->cookie); | |
eff5f53b AG |
592 | |
593 | nr_bytes = 0; | |
594 | ||
595 | rdsdebug("RDS: rdma prepare nr_local %llu rva %llx rkey %x\n", | |
596 | (unsigned long long)args->nr_local, | |
597 | (unsigned long long)args->remote_vec.addr, | |
f8b3aaf2 | 598 | op->op_rkey); |
eff5f53b | 599 | |
ff87e97a AG |
600 | local_vec = (struct rds_iovec __user *)(unsigned long) args->local_vec_addr; |
601 | ||
eff5f53b AG |
602 | for (i = 0; i < args->nr_local; i++) { |
603 | if (copy_from_user(&vec, &local_vec[i], | |
604 | sizeof(struct rds_iovec))) { | |
605 | ret = -EFAULT; | |
606 | goto out; | |
607 | } | |
608 | ||
609 | nr = rds_pages_in_vec(&vec); | |
610 | if (nr == 0) { | |
611 | ret = -EINVAL; | |
612 | goto out; | |
613 | } | |
614 | ||
615 | rs->rs_user_addr = vec.addr; | |
616 | rs->rs_user_bytes = vec.bytes; | |
617 | ||
eff5f53b AG |
618 | /* If it's a WRITE operation, we want to pin the pages for reading. |
619 | * If it's a READ operation, we need to pin the pages for writing. | |
620 | */ | |
f8b3aaf2 | 621 | ret = rds_pin_pages(vec.addr, nr, pages, !op->op_write); |
eff5f53b AG |
622 | if (ret < 0) |
623 | goto out; | |
624 | ||
625 | rdsdebug("RDS: nr_bytes %u nr %u vec.bytes %llu vec.addr %llx\n", | |
626 | nr_bytes, nr, vec.bytes, vec.addr); | |
627 | ||
628 | nr_bytes += vec.bytes; | |
629 | ||
630 | for (j = 0; j < nr; j++) { | |
631 | unsigned int offset = vec.addr & ~PAGE_MASK; | |
ff87e97a | 632 | struct scatterlist *sg; |
eff5f53b | 633 | |
f8b3aaf2 | 634 | sg = &op->op_sg[op->op_nents + j]; |
eff5f53b AG |
635 | sg_set_page(sg, pages[j], |
636 | min_t(unsigned int, vec.bytes, PAGE_SIZE - offset), | |
637 | offset); | |
638 | ||
639 | rdsdebug("RDS: sg->offset %x sg->len %x vec.addr %llx vec.bytes %llu\n", | |
640 | sg->offset, sg->length, vec.addr, vec.bytes); | |
641 | ||
642 | vec.addr += sg->length; | |
643 | vec.bytes -= sg->length; | |
644 | } | |
645 | ||
f8b3aaf2 | 646 | op->op_nents += nr; |
eff5f53b AG |
647 | } |
648 | ||
eff5f53b AG |
649 | if (nr_bytes > args->remote_vec.bytes) { |
650 | rdsdebug("RDS nr_bytes %u remote_bytes %u do not match\n", | |
651 | nr_bytes, | |
652 | (unsigned int) args->remote_vec.bytes); | |
653 | ret = -EINVAL; | |
654 | goto out; | |
655 | } | |
f8b3aaf2 | 656 | op->op_bytes = nr_bytes; |
eff5f53b AG |
657 | |
658 | ret = 0; | |
659 | out: | |
660 | kfree(pages); | |
ff87e97a AG |
661 | if (ret) |
662 | rds_rdma_free_op(op); | |
663 | ||
eff5f53b | 664 | rds_stats_inc(s_send_rdma); |
4324879d AG |
665 | |
666 | return ret; | |
eff5f53b AG |
667 | } |
668 | ||
669 | /* | |
670 | * The application wants us to pass an RDMA destination (aka MR) | |
671 | * to the remote | |
672 | */ | |
673 | int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm, | |
674 | struct cmsghdr *cmsg) | |
675 | { | |
676 | unsigned long flags; | |
677 | struct rds_mr *mr; | |
678 | u32 r_key; | |
679 | int err = 0; | |
680 | ||
f64f9e71 JP |
681 | if (cmsg->cmsg_len < CMSG_LEN(sizeof(rds_rdma_cookie_t)) || |
682 | rm->m_rdma_cookie != 0) | |
eff5f53b AG |
683 | return -EINVAL; |
684 | ||
685 | memcpy(&rm->m_rdma_cookie, CMSG_DATA(cmsg), sizeof(rm->m_rdma_cookie)); | |
686 | ||
687 | /* We are reusing a previously mapped MR here. Most likely, the | |
688 | * application has written to the buffer, so we need to explicitly | |
689 | * flush those writes to RAM. Otherwise the HCA may not see them | |
690 | * when doing a DMA from that buffer. | |
691 | */ | |
692 | r_key = rds_rdma_cookie_key(rm->m_rdma_cookie); | |
693 | ||
694 | spin_lock_irqsave(&rs->rs_rdma_lock, flags); | |
695 | mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL); | |
8690bfa1 | 696 | if (!mr) |
eff5f53b AG |
697 | err = -EINVAL; /* invalid r_key */ |
698 | else | |
699 | atomic_inc(&mr->r_refcount); | |
700 | spin_unlock_irqrestore(&rs->rs_rdma_lock, flags); | |
701 | ||
702 | if (mr) { | |
703 | mr->r_trans->sync_mr(mr->r_trans_private, DMA_TO_DEVICE); | |
f8b3aaf2 | 704 | rm->rdma.op_rdma_mr = mr; |
eff5f53b AG |
705 | } |
706 | return err; | |
707 | } | |
708 | ||
709 | /* | |
710 | * The application passes us an address range it wants to enable RDMA | |
711 | * to/from. We map the area, and save the <R_Key,offset> pair | |
712 | * in rm->m_rdma_cookie. This causes it to be sent along to the peer | |
713 | * in an extension header. | |
714 | */ | |
715 | int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm, | |
716 | struct cmsghdr *cmsg) | |
717 | { | |
f64f9e71 JP |
718 | if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_get_mr_args)) || |
719 | rm->m_rdma_cookie != 0) | |
eff5f53b AG |
720 | return -EINVAL; |
721 | ||
f8b3aaf2 | 722 | return __rds_rdma_map(rs, CMSG_DATA(cmsg), &rm->m_rdma_cookie, &rm->rdma.op_rdma_mr); |
eff5f53b | 723 | } |
15133f6e AG |
724 | |
725 | /* | |
726 | * Fill in rds_message for an atomic request. | |
727 | */ | |
728 | int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm, | |
729 | struct cmsghdr *cmsg) | |
730 | { | |
731 | struct page *page = NULL; | |
732 | struct rds_atomic_args *args; | |
733 | int ret = 0; | |
734 | ||
735 | if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_atomic_args)) | |
736 | || rm->atomic.op_active) | |
737 | return -EINVAL; | |
738 | ||
739 | args = CMSG_DATA(cmsg); | |
740 | ||
20c72bd5 AG |
741 | /* Nonmasked & masked cmsg ops converted to masked hw ops */ |
742 | switch (cmsg->cmsg_type) { | |
743 | case RDS_CMSG_ATOMIC_FADD: | |
744 | rm->atomic.op_type = RDS_ATOMIC_TYPE_FADD; | |
745 | rm->atomic.op_m_fadd.add = args->fadd.add; | |
746 | rm->atomic.op_m_fadd.nocarry_mask = 0; | |
747 | break; | |
748 | case RDS_CMSG_MASKED_ATOMIC_FADD: | |
15133f6e | 749 | rm->atomic.op_type = RDS_ATOMIC_TYPE_FADD; |
20c72bd5 AG |
750 | rm->atomic.op_m_fadd.add = args->m_fadd.add; |
751 | rm->atomic.op_m_fadd.nocarry_mask = args->m_fadd.nocarry_mask; | |
752 | break; | |
753 | case RDS_CMSG_ATOMIC_CSWP: | |
754 | rm->atomic.op_type = RDS_ATOMIC_TYPE_CSWP; | |
755 | rm->atomic.op_m_cswp.compare = args->cswp.compare; | |
756 | rm->atomic.op_m_cswp.swap = args->cswp.swap; | |
757 | rm->atomic.op_m_cswp.compare_mask = ~0; | |
758 | rm->atomic.op_m_cswp.swap_mask = ~0; | |
759 | break; | |
760 | case RDS_CMSG_MASKED_ATOMIC_CSWP: | |
761 | rm->atomic.op_type = RDS_ATOMIC_TYPE_CSWP; | |
762 | rm->atomic.op_m_cswp.compare = args->m_cswp.compare; | |
763 | rm->atomic.op_m_cswp.swap = args->m_cswp.swap; | |
764 | rm->atomic.op_m_cswp.compare_mask = args->m_cswp.compare_mask; | |
765 | rm->atomic.op_m_cswp.swap_mask = args->m_cswp.swap_mask; | |
766 | break; | |
767 | default: | |
768 | BUG(); /* should never happen */ | |
15133f6e AG |
769 | } |
770 | ||
15133f6e | 771 | rm->atomic.op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME); |
2c3a5f9a | 772 | rm->atomic.op_silent = !!(args->flags & RDS_RDMA_SILENT); |
7e3bd65e | 773 | rm->atomic.op_active = 1; |
15133f6e AG |
774 | rm->atomic.op_recverr = rs->rs_recverr; |
775 | rm->atomic.op_sg = rds_message_alloc_sgs(rm, 1); | |
776 | ||
777 | /* verify 8 byte-aligned */ | |
778 | if (args->local_addr & 0x7) { | |
779 | ret = -EFAULT; | |
780 | goto err; | |
781 | } | |
782 | ||
783 | ret = rds_pin_pages(args->local_addr, 1, &page, 1); | |
784 | if (ret != 1) | |
785 | goto err; | |
786 | ret = 0; | |
787 | ||
788 | sg_set_page(rm->atomic.op_sg, page, 8, offset_in_page(args->local_addr)); | |
789 | ||
790 | if (rm->atomic.op_notify || rm->atomic.op_recverr) { | |
791 | /* We allocate an uninitialized notifier here, because | |
792 | * we don't want to do that in the completion handler. We | |
793 | * would have to use GFP_ATOMIC there, and don't want to deal | |
794 | * with failed allocations. | |
795 | */ | |
796 | rm->atomic.op_notifier = kmalloc(sizeof(*rm->atomic.op_notifier), GFP_KERNEL); | |
797 | if (!rm->atomic.op_notifier) { | |
798 | ret = -ENOMEM; | |
799 | goto err; | |
800 | } | |
801 | ||
802 | rm->atomic.op_notifier->n_user_token = args->user_token; | |
803 | rm->atomic.op_notifier->n_status = RDS_RDMA_SUCCESS; | |
804 | } | |
805 | ||
40589e74 | 806 | rm->atomic.op_rkey = rds_rdma_cookie_key(args->cookie); |
15133f6e AG |
807 | rm->atomic.op_remote_addr = args->remote_addr + rds_rdma_cookie_offset(args->cookie); |
808 | ||
15133f6e AG |
809 | return ret; |
810 | err: | |
811 | if (page) | |
812 | put_page(page); | |
813 | kfree(rm->atomic.op_notifier); | |
814 | ||
815 | return ret; | |
816 | } |