blob: eaeeb91e11196a07405ee9cbfc417c7da3793ea3 [file] [log] [blame]
Andy Grovereff5f532009-02-24 15:30:29 +00001/*
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/rbtree.h>
35#include <linux/dma-mapping.h> /* for DMA_*_DEVICE */
36
37#include "rdma.h"
38
39/*
40 * XXX
41 * - build with sparse
42 * - should we limit the size of a mr region? let transport return failure?
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 atomic_inc(&insert->r_refcount);
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, 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
118void __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 */
128void rds_rdma_drop_keys(struct rds_sock *rs)
129{
130 struct rds_mr *mr;
131 struct rb_node *node;
132
133 /* Release any MRs associated with this socket */
134 while ((node = rb_first(&rs->rs_rdma_keys))) {
135 mr = container_of(node, struct rds_mr, r_rb_node);
136 if (mr->r_trans == rs->rs_transport)
137 mr->r_invalidate = 0;
138 rds_mr_put(mr);
139 }
140
141 if (rs->rs_transport && rs->rs_transport->flush_mrs)
142 rs->rs_transport->flush_mrs();
143}
144
145/*
146 * Helper function to pin user pages.
147 */
148static int rds_pin_pages(unsigned long user_addr, unsigned int nr_pages,
149 struct page **pages, int write)
150{
151 int ret;
152
153 down_read(&current->mm->mmap_sem);
154 ret = get_user_pages(current, current->mm, user_addr,
155 nr_pages, write, 0, pages, NULL);
156 up_read(&current->mm->mmap_sem);
157
158 if (0 <= ret && (unsigned) ret < nr_pages) {
159 while (ret--)
160 put_page(pages[ret]);
161 ret = -EFAULT;
162 }
163
164 return ret;
165}
166
167static int __rds_rdma_map(struct rds_sock *rs, struct rds_get_mr_args *args,
168 u64 *cookie_ret, struct rds_mr **mr_ret)
169{
170 struct rds_mr *mr = NULL, *found;
171 unsigned int nr_pages;
172 struct page **pages = NULL;
173 struct scatterlist *sg;
174 void *trans_private;
175 unsigned long flags;
176 rds_rdma_cookie_t cookie;
177 unsigned int nents;
178 long i;
179 int ret;
180
181 if (rs->rs_bound_addr == 0) {
182 ret = -ENOTCONN; /* XXX not a great errno */
183 goto out;
184 }
185
186 if (rs->rs_transport->get_mr == NULL) {
187 ret = -EOPNOTSUPP;
188 goto out;
189 }
190
191 nr_pages = rds_pages_in_vec(&args->vec);
192 if (nr_pages == 0) {
193 ret = -EINVAL;
194 goto out;
195 }
196
197 rdsdebug("RDS: get_mr addr %llx len %llu nr_pages %u\n",
198 args->vec.addr, args->vec.bytes, nr_pages);
199
200 /* XXX clamp nr_pages to limit the size of this alloc? */
201 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
202 if (pages == NULL) {
203 ret = -ENOMEM;
204 goto out;
205 }
206
207 mr = kzalloc(sizeof(struct rds_mr), GFP_KERNEL);
208 if (mr == NULL) {
209 ret = -ENOMEM;
210 goto out;
211 }
212
213 atomic_set(&mr->r_refcount, 1);
214 RB_CLEAR_NODE(&mr->r_rb_node);
215 mr->r_trans = rs->rs_transport;
216 mr->r_sock = rs;
217
218 if (args->flags & RDS_RDMA_USE_ONCE)
219 mr->r_use_once = 1;
220 if (args->flags & RDS_RDMA_INVALIDATE)
221 mr->r_invalidate = 1;
222 if (args->flags & RDS_RDMA_READWRITE)
223 mr->r_write = 1;
224
225 /*
226 * Pin the pages that make up the user buffer and transfer the page
227 * pointers to the mr's sg array. We check to see if we've mapped
228 * the whole region after transferring the partial page references
229 * to the sg array so that we can have one page ref cleanup path.
230 *
231 * For now we have no flag that tells us whether the mapping is
232 * r/o or r/w. We need to assume r/w, or we'll do a lot of RDMA to
233 * the zero page.
234 */
235 ret = rds_pin_pages(args->vec.addr & PAGE_MASK, nr_pages, pages, 1);
236 if (ret < 0)
237 goto out;
238
239 nents = ret;
240 sg = kcalloc(nents, sizeof(*sg), GFP_KERNEL);
241 if (sg == NULL) {
242 ret = -ENOMEM;
243 goto out;
244 }
245 WARN_ON(!nents);
246 sg_init_table(sg, nents);
247
248 /* Stick all pages into the scatterlist */
249 for (i = 0 ; i < nents; i++)
250 sg_set_page(&sg[i], pages[i], PAGE_SIZE, 0);
251
252 rdsdebug("RDS: trans_private nents is %u\n", nents);
253
254 /* Obtain a transport specific MR. If this succeeds, the
255 * s/g list is now owned by the MR.
256 * Note that dma_map() implies that pending writes are
257 * flushed to RAM, so no dma_sync is needed here. */
258 trans_private = rs->rs_transport->get_mr(sg, nents, rs,
259 &mr->r_key);
260
261 if (IS_ERR(trans_private)) {
262 for (i = 0 ; i < nents; i++)
263 put_page(sg_page(&sg[i]));
264 kfree(sg);
265 ret = PTR_ERR(trans_private);
266 goto out;
267 }
268
269 mr->r_trans_private = trans_private;
270
271 rdsdebug("RDS: get_mr put_user key is %x cookie_addr %p\n",
272 mr->r_key, (void *)(unsigned long) args->cookie_addr);
273
274 /* The user may pass us an unaligned address, but we can only
275 * map page aligned regions. So we keep the offset, and build
276 * a 64bit cookie containing <R_Key, offset> and pass that
277 * around. */
278 cookie = rds_rdma_make_cookie(mr->r_key, args->vec.addr & ~PAGE_MASK);
279 if (cookie_ret)
280 *cookie_ret = cookie;
281
282 if (args->cookie_addr && put_user(cookie, (u64 __user *)(unsigned long) args->cookie_addr)) {
283 ret = -EFAULT;
284 goto out;
285 }
286
287 /* Inserting the new MR into the rbtree bumps its
288 * reference count. */
289 spin_lock_irqsave(&rs->rs_rdma_lock, flags);
290 found = rds_mr_tree_walk(&rs->rs_rdma_keys, mr->r_key, mr);
291 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
292
293 BUG_ON(found && found != mr);
294
295 rdsdebug("RDS: get_mr key is %x\n", mr->r_key);
296 if (mr_ret) {
297 atomic_inc(&mr->r_refcount);
298 *mr_ret = mr;
299 }
300
301 ret = 0;
302out:
303 kfree(pages);
304 if (mr)
305 rds_mr_put(mr);
306 return ret;
307}
308
309int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen)
310{
311 struct rds_get_mr_args args;
312
313 if (optlen != sizeof(struct rds_get_mr_args))
314 return -EINVAL;
315
316 if (copy_from_user(&args, (struct rds_get_mr_args __user *)optval,
317 sizeof(struct rds_get_mr_args)))
318 return -EFAULT;
319
320 return __rds_rdma_map(rs, &args, NULL, NULL);
321}
322
323/*
324 * Free the MR indicated by the given R_Key
325 */
326int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen)
327{
328 struct rds_free_mr_args args;
329 struct rds_mr *mr;
330 unsigned long flags;
331
332 if (optlen != sizeof(struct rds_free_mr_args))
333 return -EINVAL;
334
335 if (copy_from_user(&args, (struct rds_free_mr_args __user *)optval,
336 sizeof(struct rds_free_mr_args)))
337 return -EFAULT;
338
339 /* Special case - a null cookie means flush all unused MRs */
340 if (args.cookie == 0) {
341 if (!rs->rs_transport || !rs->rs_transport->flush_mrs)
342 return -EINVAL;
343 rs->rs_transport->flush_mrs();
344 return 0;
345 }
346
347 /* Look up the MR given its R_key and remove it from the rbtree
348 * so nobody else finds it.
349 * This should also prevent races with rds_rdma_unuse.
350 */
351 spin_lock_irqsave(&rs->rs_rdma_lock, flags);
352 mr = rds_mr_tree_walk(&rs->rs_rdma_keys, rds_rdma_cookie_key(args.cookie), NULL);
353 if (mr) {
354 rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
355 RB_CLEAR_NODE(&mr->r_rb_node);
356 if (args.flags & RDS_RDMA_INVALIDATE)
357 mr->r_invalidate = 1;
358 }
359 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
360
361 if (!mr)
362 return -EINVAL;
363
364 /*
365 * call rds_destroy_mr() ourselves so that we're sure it's done by the time
366 * we return. If we let rds_mr_put() do it it might not happen until
367 * someone else drops their ref.
368 */
369 rds_destroy_mr(mr);
370 rds_mr_put(mr);
371 return 0;
372}
373
374/*
375 * This is called when we receive an extension header that
376 * tells us this MR was used. It allows us to implement
377 * use_once semantics
378 */
379void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force)
380{
381 struct rds_mr *mr;
382 unsigned long flags;
383 int zot_me = 0;
384
385 spin_lock_irqsave(&rs->rs_rdma_lock, flags);
386 mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL);
387 if (mr && (mr->r_use_once || force)) {
388 rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
389 RB_CLEAR_NODE(&mr->r_rb_node);
390 zot_me = 1;
391 } else if (mr)
392 atomic_inc(&mr->r_refcount);
393 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
394
395 /* May have to issue a dma_sync on this memory region.
396 * Note we could avoid this if the operation was a RDMA READ,
397 * but at this point we can't tell. */
398 if (mr != NULL) {
399 if (mr->r_trans->sync_mr)
400 mr->r_trans->sync_mr(mr->r_trans_private, DMA_FROM_DEVICE);
401
402 /* If the MR was marked as invalidate, this will
403 * trigger an async flush. */
404 if (zot_me)
405 rds_destroy_mr(mr);
406 rds_mr_put(mr);
407 }
408}
409
410void rds_rdma_free_op(struct rds_rdma_op *ro)
411{
412 unsigned int i;
413
414 for (i = 0; i < ro->r_nents; i++) {
415 struct page *page = sg_page(&ro->r_sg[i]);
416
417 /* Mark page dirty if it was possibly modified, which
418 * is the case for a RDMA_READ which copies from remote
419 * to local memory */
420 if (!ro->r_write)
421 set_page_dirty(page);
422 put_page(page);
423 }
424
425 kfree(ro->r_notifier);
426 kfree(ro);
427}
428
429/*
430 * args is a pointer to an in-kernel copy in the sendmsg cmsg.
431 */
432static struct rds_rdma_op *rds_rdma_prepare(struct rds_sock *rs,
433 struct rds_rdma_args *args)
434{
435 struct rds_iovec vec;
436 struct rds_rdma_op *op = NULL;
437 unsigned int nr_pages;
438 unsigned int max_pages;
439 unsigned int nr_bytes;
440 struct page **pages = NULL;
441 struct rds_iovec __user *local_vec;
442 struct scatterlist *sg;
443 unsigned int nr;
444 unsigned int i, j;
445 int ret;
446
447
448 if (rs->rs_bound_addr == 0) {
449 ret = -ENOTCONN; /* XXX not a great errno */
450 goto out;
451 }
452
453 if (args->nr_local > (u64)UINT_MAX) {
454 ret = -EMSGSIZE;
455 goto out;
456 }
457
458 nr_pages = 0;
459 max_pages = 0;
460
461 local_vec = (struct rds_iovec __user *)(unsigned long) args->local_vec_addr;
462
463 /* figure out the number of pages in the vector */
464 for (i = 0; i < args->nr_local; i++) {
465 if (copy_from_user(&vec, &local_vec[i],
466 sizeof(struct rds_iovec))) {
467 ret = -EFAULT;
468 goto out;
469 }
470
471 nr = rds_pages_in_vec(&vec);
472 if (nr == 0) {
473 ret = -EINVAL;
474 goto out;
475 }
476
477 max_pages = max(nr, max_pages);
478 nr_pages += nr;
479 }
480
481 pages = kcalloc(max_pages, sizeof(struct page *), GFP_KERNEL);
482 if (pages == NULL) {
483 ret = -ENOMEM;
484 goto out;
485 }
486
487 op = kzalloc(offsetof(struct rds_rdma_op, r_sg[nr_pages]), GFP_KERNEL);
488 if (op == NULL) {
489 ret = -ENOMEM;
490 goto out;
491 }
492
493 op->r_write = !!(args->flags & RDS_RDMA_READWRITE);
494 op->r_fence = !!(args->flags & RDS_RDMA_FENCE);
495 op->r_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME);
496 op->r_recverr = rs->rs_recverr;
497 WARN_ON(!nr_pages);
498 sg_init_table(op->r_sg, nr_pages);
499
500 if (op->r_notify || op->r_recverr) {
501 /* We allocate an uninitialized notifier here, because
502 * we don't want to do that in the completion handler. We
503 * would have to use GFP_ATOMIC there, and don't want to deal
504 * with failed allocations.
505 */
506 op->r_notifier = kmalloc(sizeof(struct rds_notifier), GFP_KERNEL);
507 if (!op->r_notifier) {
508 ret = -ENOMEM;
509 goto out;
510 }
511 op->r_notifier->n_user_token = args->user_token;
512 op->r_notifier->n_status = RDS_RDMA_SUCCESS;
513 }
514
515 /* The cookie contains the R_Key of the remote memory region, and
516 * optionally an offset into it. This is how we implement RDMA into
517 * unaligned memory.
518 * When setting up the RDMA, we need to add that offset to the
519 * destination address (which is really an offset into the MR)
520 * FIXME: We may want to move this into ib_rdma.c
521 */
522 op->r_key = rds_rdma_cookie_key(args->cookie);
523 op->r_remote_addr = args->remote_vec.addr + rds_rdma_cookie_offset(args->cookie);
524
525 nr_bytes = 0;
526
527 rdsdebug("RDS: rdma prepare nr_local %llu rva %llx rkey %x\n",
528 (unsigned long long)args->nr_local,
529 (unsigned long long)args->remote_vec.addr,
530 op->r_key);
531
532 for (i = 0; i < args->nr_local; i++) {
533 if (copy_from_user(&vec, &local_vec[i],
534 sizeof(struct rds_iovec))) {
535 ret = -EFAULT;
536 goto out;
537 }
538
539 nr = rds_pages_in_vec(&vec);
540 if (nr == 0) {
541 ret = -EINVAL;
542 goto out;
543 }
544
545 rs->rs_user_addr = vec.addr;
546 rs->rs_user_bytes = vec.bytes;
547
548 /* did the user change the vec under us? */
549 if (nr > max_pages || op->r_nents + nr > nr_pages) {
550 ret = -EINVAL;
551 goto out;
552 }
553 /* If it's a WRITE operation, we want to pin the pages for reading.
554 * If it's a READ operation, we need to pin the pages for writing.
555 */
556 ret = rds_pin_pages(vec.addr & PAGE_MASK, nr, pages, !op->r_write);
557 if (ret < 0)
558 goto out;
559
560 rdsdebug("RDS: nr_bytes %u nr %u vec.bytes %llu vec.addr %llx\n",
561 nr_bytes, nr, vec.bytes, vec.addr);
562
563 nr_bytes += vec.bytes;
564
565 for (j = 0; j < nr; j++) {
566 unsigned int offset = vec.addr & ~PAGE_MASK;
567
568 sg = &op->r_sg[op->r_nents + j];
569 sg_set_page(sg, pages[j],
570 min_t(unsigned int, vec.bytes, PAGE_SIZE - offset),
571 offset);
572
573 rdsdebug("RDS: sg->offset %x sg->len %x vec.addr %llx vec.bytes %llu\n",
574 sg->offset, sg->length, vec.addr, vec.bytes);
575
576 vec.addr += sg->length;
577 vec.bytes -= sg->length;
578 }
579
580 op->r_nents += nr;
581 }
582
583
584 if (nr_bytes > args->remote_vec.bytes) {
585 rdsdebug("RDS nr_bytes %u remote_bytes %u do not match\n",
586 nr_bytes,
587 (unsigned int) args->remote_vec.bytes);
588 ret = -EINVAL;
589 goto out;
590 }
591 op->r_bytes = nr_bytes;
592
593 ret = 0;
594out:
595 kfree(pages);
596 if (ret) {
597 if (op)
598 rds_rdma_free_op(op);
599 op = ERR_PTR(ret);
600 }
601 return op;
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{
611 struct rds_rdma_op *op;
612
613 if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_rdma_args))
614 || rm->m_rdma_op != NULL)
615 return -EINVAL;
616
617 op = rds_rdma_prepare(rs, CMSG_DATA(cmsg));
618 if (IS_ERR(op))
619 return PTR_ERR(op);
620 rds_stats_inc(s_send_rdma);
621 rm->m_rdma_op = op;
622 return 0;
623}
624
625/*
626 * The application wants us to pass an RDMA destination (aka MR)
627 * to the remote
628 */
629int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
630 struct cmsghdr *cmsg)
631{
632 unsigned long flags;
633 struct rds_mr *mr;
634 u32 r_key;
635 int err = 0;
636
637 if (cmsg->cmsg_len < CMSG_LEN(sizeof(rds_rdma_cookie_t))
638 || rm->m_rdma_cookie != 0)
639 return -EINVAL;
640
641 memcpy(&rm->m_rdma_cookie, CMSG_DATA(cmsg), sizeof(rm->m_rdma_cookie));
642
643 /* We are reusing a previously mapped MR here. Most likely, the
644 * application has written to the buffer, so we need to explicitly
645 * flush those writes to RAM. Otherwise the HCA may not see them
646 * when doing a DMA from that buffer.
647 */
648 r_key = rds_rdma_cookie_key(rm->m_rdma_cookie);
649
650 spin_lock_irqsave(&rs->rs_rdma_lock, flags);
651 mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL);
652 if (mr == NULL)
653 err = -EINVAL; /* invalid r_key */
654 else
655 atomic_inc(&mr->r_refcount);
656 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
657
658 if (mr) {
659 mr->r_trans->sync_mr(mr->r_trans_private, DMA_TO_DEVICE);
660 rm->m_rdma_mr = mr;
661 }
662 return err;
663}
664
665/*
666 * The application passes us an address range it wants to enable RDMA
667 * to/from. We map the area, and save the <R_Key,offset> pair
668 * in rm->m_rdma_cookie. This causes it to be sent along to the peer
669 * in an extension header.
670 */
671int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
672 struct cmsghdr *cmsg)
673{
674 if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_get_mr_args))
675 || rm->m_rdma_cookie != 0)
676 return -EINVAL;
677
678 return __rds_rdma_map(rs, CMSG_DATA(cmsg), &rm->m_rdma_cookie, &rm->m_rdma_mr);
679}