?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
fuse: llseek optimize SEEK_CUR and SEEK_SET
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / fs / fuse / file.c
blobe0b60acfe7265aa2e2d0655fc8bd37d50cdf279f
1 /*
2 FUSE: Filesystem in Userspace
3 Copyright (C) 2001-2008 Miklos Szeredi <miklos@szeredi.hu>
4
5 This program can be distributed under the terms of the GNU GPL.
6 See the file COPYING.
7 */
8
9 #include "fuse_i.h"
10
11 #include <linux/pagemap.h>
12 #include <linux/slab.h>
13 #include <linux/kernel.h>
14 #include <linux/sched.h>
15 #include <linux/module.h>
16 #include <linux/compat.h>
17 #include <linux/swap.h>
18
19 static const struct file_operations fuse_direct_io_file_operations;
20
21 static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
22 int opcode, struct fuse_open_out *outargp)
23 {
24 struct fuse_open_in inarg;
25 struct fuse_req *req;
26 int err;
27
28 req = fuse_get_req(fc);
29 if (IS_ERR(req))
30 return PTR_ERR(req);
31
32 memset(&inarg, 0, sizeof(inarg));
33 inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
34 if (!fc->atomic_o_trunc)
35 inarg.flags &= ~O_TRUNC;
36 req->in.h.opcode = opcode;
37 req->in.h.nodeid = nodeid;
38 req->in.numargs = 1;
39 req->in.args[0].size = sizeof(inarg);
40 req->in.args[0].value = &inarg;
41 req->out.numargs = 1;
42 req->out.args[0].size = sizeof(*outargp);
43 req->out.args[0].value = outargp;
44 fuse_request_send(fc, req);
45 err = req->out.h.error;
46 fuse_put_request(fc, req);
47
48 return err;
49 }
50
51 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
52 {
53 struct fuse_file *ff;
54
55 ff = kmalloc(sizeof(struct fuse_file), GFP_KERNEL);
56 if (unlikely(!ff))
57 return NULL;
58
59 ff->fc = fc;
60 ff->reserved_req = fuse_request_alloc();
61 if (unlikely(!ff->reserved_req)) {
62 kfree(ff);
63 return NULL;
64 }
65
66 INIT_LIST_HEAD(&ff->write_entry);
67 atomic_set(&ff->count, 0);
68 RB_CLEAR_NODE(&ff->polled_node);
69 init_waitqueue_head(&ff->poll_wait);
70
71 spin_lock(&fc->lock);
72 ff->kh = ++fc->khctr;
73 spin_unlock(&fc->lock);
74
75 return ff;
76 }
77
78 void fuse_file_free(struct fuse_file *ff)
79 {
80 fuse_request_free(ff->reserved_req);
81 kfree(ff);
82 }
83
84 struct fuse_file *fuse_file_get(struct fuse_file *ff)
85 {
86 atomic_inc(&ff->count);
87 return ff;
88 }
89
90 static void fuse_release_async(struct work_struct *work)
91 {
92 struct fuse_req *req;
93 struct fuse_conn *fc;
94 struct path path;
95
96 req = container_of(work, struct fuse_req, misc.release.work);
97 path = req->misc.release.path;
98 fc = get_fuse_conn(path.dentry->d_inode);
99
100 fuse_put_request(fc, req);
101 path_put(&path);
102 }
103
104 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
105 {
106 if (fc->destroy_req) {
107 /*
108 * If this is a fuseblk mount, then it's possible that
109 * releasing the path will result in releasing the
110 * super block and sending the DESTROY request. If
111 * the server is single threaded, this would hang.
112 * For this reason do the path_put() in a separate
113 * thread.
114 */
115 atomic_inc(&req->count);
116 INIT_WORK(&req->misc.release.work, fuse_release_async);
117 schedule_work(&req->misc.release.work);
118 } else {
119 path_put(&req->misc.release.path);
120 }
121 }
122
123 static void fuse_file_put(struct fuse_file *ff, bool sync)
124 {
125 if (atomic_dec_and_test(&ff->count)) {
126 struct fuse_req *req = ff->reserved_req;
127
128 if (sync) {
129 fuse_request_send(ff->fc, req);
130 path_put(&req->misc.release.path);
131 fuse_put_request(ff->fc, req);
132 } else {
133 req->end = fuse_release_end;
134 fuse_request_send_background(ff->fc, req);
135 }
136 kfree(ff);
137 }
138 }
139
140 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
141 bool isdir)
142 {
143 struct fuse_open_out outarg;
144 struct fuse_file *ff;
145 int err;
146 int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
147
148 ff = fuse_file_alloc(fc);
149 if (!ff)
150 return -ENOMEM;
151
152 err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
153 if (err) {
154 fuse_file_free(ff);
155 return err;
156 }
157
158 if (isdir)
159 outarg.open_flags &= ~FOPEN_DIRECT_IO;
160
161 ff->fh = outarg.fh;
162 ff->nodeid = nodeid;
163 ff->open_flags = outarg.open_flags;
164 file->private_data = fuse_file_get(ff);
165
166 return 0;
167 }
168 EXPORT_SYMBOL_GPL(fuse_do_open);
169
170 void fuse_finish_open(struct inode *inode, struct file *file)
171 {
172 struct fuse_file *ff = file->private_data;
173 struct fuse_conn *fc = get_fuse_conn(inode);
174
175 if (ff->open_flags & FOPEN_DIRECT_IO)
176 file->f_op = &fuse_direct_io_file_operations;
177 if (!(ff->open_flags & FOPEN_KEEP_CACHE))
178 invalidate_inode_pages2(inode->i_mapping);
179 if (ff->open_flags & FOPEN_NONSEEKABLE)
180 nonseekable_open(inode, file);
181 if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
182 struct fuse_inode *fi = get_fuse_inode(inode);
183
184 spin_lock(&fc->lock);
185 fi->attr_version = ++fc->attr_version;
186 i_size_write(inode, 0);
187 spin_unlock(&fc->lock);
188 fuse_invalidate_attr(inode);
189 }
190 }
191
192 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
193 {
194 struct fuse_conn *fc = get_fuse_conn(inode);
195 int err;
196
197 /* VFS checks this, but only _after_ ->open() */
198 if (file->f_flags & O_DIRECT)
199 return -EINVAL;
200
201 err = generic_file_open(inode, file);
202 if (err)
203 return err;
204
205 err = fuse_do_open(fc, get_node_id(inode), file, isdir);
206 if (err)
207 return err;
208
209 fuse_finish_open(inode, file);
210
211 return 0;
212 }
213
214 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
215 {
216 struct fuse_conn *fc = ff->fc;
217 struct fuse_req *req = ff->reserved_req;
218 struct fuse_release_in *inarg = &req->misc.release.in;
219
220 spin_lock(&fc->lock);
221 list_del(&ff->write_entry);
222 if (!RB_EMPTY_NODE(&ff->polled_node))
223 rb_erase(&ff->polled_node, &fc->polled_files);
224 spin_unlock(&fc->lock);
225
226 wake_up_interruptible_all(&ff->poll_wait);
227
228 inarg->fh = ff->fh;
229 inarg->flags = flags;
230 req->in.h.opcode = opcode;
231 req->in.h.nodeid = ff->nodeid;
232 req->in.numargs = 1;
233 req->in.args[0].size = sizeof(struct fuse_release_in);
234 req->in.args[0].value = inarg;
235 }
236
237 void fuse_release_common(struct file *file, int opcode)
238 {
239 struct fuse_file *ff;
240 struct fuse_req *req;
241
242 ff = file->private_data;
243 if (unlikely(!ff))
244 return;
245
246 req = ff->reserved_req;
247 fuse_prepare_release(ff, file->f_flags, opcode);
248
249 if (ff->flock) {
250 struct fuse_release_in *inarg = &req->misc.release.in;
251 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
252 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
253 (fl_owner_t) file);
254 }
255 /* Hold vfsmount and dentry until release is finished */
256 path_get(&file->f_path);
257 req->misc.release.path = file->f_path;
258
259 /*
260 * Normally this will send the RELEASE request, however if
261 * some asynchronous READ or WRITE requests are outstanding,
262 * the sending will be delayed.
263 *
264 * Make the release synchronous if this is a fuseblk mount,
265 * synchronous RELEASE is allowed (and desirable) in this case
266 * because the server can be trusted not to screw up.
267 */
268 fuse_file_put(ff, ff->fc->destroy_req != NULL);
269 }
270
271 static int fuse_open(struct inode *inode, struct file *file)
272 {
273 return fuse_open_common(inode, file, false);
274 }
275
276 static int fuse_release(struct inode *inode, struct file *file)
277 {
278 fuse_release_common(file, FUSE_RELEASE);
279
280 /* return value is ignored by VFS */
281 return 0;
282 }
283
284 void fuse_sync_release(struct fuse_file *ff, int flags)
285 {
286 WARN_ON(atomic_read(&ff->count) > 1);
287 fuse_prepare_release(ff, flags, FUSE_RELEASE);
288 ff->reserved_req->force = 1;
289 fuse_request_send(ff->fc, ff->reserved_req);
290 fuse_put_request(ff->fc, ff->reserved_req);
291 kfree(ff);
292 }
293 EXPORT_SYMBOL_GPL(fuse_sync_release);
294
295 /*
296 * Scramble the ID space with XTEA, so that the value of the files_struct
297 * pointer is not exposed to userspace.
298 */
299 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
300 {
301 u32 *k = fc->scramble_key;
302 u64 v = (unsigned long) id;
303 u32 v0 = v;
304 u32 v1 = v >> 32;
305 u32 sum = 0;
306 int i;
307
308 for (i = 0; i < 32; i++) {
309 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
310 sum += 0x9E3779B9;
311 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
312 }
313
314 return (u64) v0 + ((u64) v1 << 32);
315 }
316
317 /*
318 * Check if page is under writeback
319 *
320 * This is currently done by walking the list of writepage requests
321 * for the inode, which can be pretty inefficient.
322 */
323 static bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
324 {
325 struct fuse_conn *fc = get_fuse_conn(inode);
326 struct fuse_inode *fi = get_fuse_inode(inode);
327 struct fuse_req *req;
328 bool found = false;
329
330 spin_lock(&fc->lock);
331 list_for_each_entry(req, &fi->writepages, writepages_entry) {
332 pgoff_t curr_index;
333
334 BUG_ON(req->inode != inode);
335 curr_index = req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
336 if (curr_index == index) {
337 found = true;
338 break;
339 }
340 }
341 spin_unlock(&fc->lock);
342
343 return found;
344 }
345
346 /*
347 * Wait for page writeback to be completed.
348 *
349 * Since fuse doesn't rely on the VM writeback tracking, this has to
350 * use some other means.
351 */
352 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
353 {
354 struct fuse_inode *fi = get_fuse_inode(inode);
355
356 wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
357 return 0;
358 }
359
360 static int fuse_flush(struct file *file, fl_owner_t id)
361 {
362 struct inode *inode = file->f_path.dentry->d_inode;
363 struct fuse_conn *fc = get_fuse_conn(inode);
364 struct fuse_file *ff = file->private_data;
365 struct fuse_req *req;
366 struct fuse_flush_in inarg;
367 int err;
368
369 if (is_bad_inode(inode))
370 return -EIO;
371
372 if (fc->no_flush)
373 return 0;
374
375 req = fuse_get_req_nofail(fc, file);
376 memset(&inarg, 0, sizeof(inarg));
377 inarg.fh = ff->fh;
378 inarg.lock_owner = fuse_lock_owner_id(fc, id);
379 req->in.h.opcode = FUSE_FLUSH;
380 req->in.h.nodeid = get_node_id(inode);
381 req->in.numargs = 1;
382 req->in.args[0].size = sizeof(inarg);
383 req->in.args[0].value = &inarg;
384 req->force = 1;
385 fuse_request_send(fc, req);
386 err = req->out.h.error;
387 fuse_put_request(fc, req);
388 if (err == -ENOSYS) {
389 fc->no_flush = 1;
390 err = 0;
391 }
392 return err;
393 }
394
395 /*
396 * Wait for all pending writepages on the inode to finish.
397 *
398 * This is currently done by blocking further writes with FUSE_NOWRITE
399 * and waiting for all sent writes to complete.
400 *
401 * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
402 * could conflict with truncation.
403 */
404 static void fuse_sync_writes(struct inode *inode)
405 {
406 fuse_set_nowrite(inode);
407 fuse_release_nowrite(inode);
408 }
409
410 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
411 int datasync, int isdir)
412 {
413 struct inode *inode = file->f_mapping->host;
414 struct fuse_conn *fc = get_fuse_conn(inode);
415 struct fuse_file *ff = file->private_data;
416 struct fuse_req *req;
417 struct fuse_fsync_in inarg;
418 int err;
419
420 if (is_bad_inode(inode))
421 return -EIO;
422
423 err = filemap_write_and_wait_range(inode->i_mapping, start, end);
424 if (err)
425 return err;
426
427 if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
428 return 0;
429
430 mutex_lock(&inode->i_mutex);
431
432 /*
433 * Start writeback against all dirty pages of the inode, then
434 * wait for all outstanding writes, before sending the FSYNC
435 * request.
436 */
437 err = write_inode_now(inode, 0);
438 if (err)
439 goto out;
440
441 fuse_sync_writes(inode);
442
443 req = fuse_get_req(fc);
444 if (IS_ERR(req)) {
445 err = PTR_ERR(req);
446 goto out;
447 }
448
449 memset(&inarg, 0, sizeof(inarg));
450 inarg.fh = ff->fh;
451 inarg.fsync_flags = datasync ? 1 : 0;
452 req->in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
453 req->in.h.nodeid = get_node_id(inode);
454 req->in.numargs = 1;
455 req->in.args[0].size = sizeof(inarg);
456 req->in.args[0].value = &inarg;
457 fuse_request_send(fc, req);
458 err = req->out.h.error;
459 fuse_put_request(fc, req);
460 if (err == -ENOSYS) {
461 if (isdir)
462 fc->no_fsyncdir = 1;
463 else
464 fc->no_fsync = 1;
465 err = 0;
466 }
467 out:
468 mutex_unlock(&inode->i_mutex);
469 return err;
470 }
471
472 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
473 int datasync)
474 {
475 return fuse_fsync_common(file, start, end, datasync, 0);
476 }
477
478 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
479 size_t count, int opcode)
480 {
481 struct fuse_read_in *inarg = &req->misc.read.in;
482 struct fuse_file *ff = file->private_data;
483
484 inarg->fh = ff->fh;
485 inarg->offset = pos;
486 inarg->size = count;
487 inarg->flags = file->f_flags;
488 req->in.h.opcode = opcode;
489 req->in.h.nodeid = ff->nodeid;
490 req->in.numargs = 1;
491 req->in.args[0].size = sizeof(struct fuse_read_in);
492 req->in.args[0].value = inarg;
493 req->out.argvar = 1;
494 req->out.numargs = 1;
495 req->out.args[0].size = count;
496 }
497
498 static size_t fuse_send_read(struct fuse_req *req, struct file *file,
499 loff_t pos, size_t count, fl_owner_t owner)
500 {
501 struct fuse_file *ff = file->private_data;
502 struct fuse_conn *fc = ff->fc;
503
504 fuse_read_fill(req, file, pos, count, FUSE_READ);
505 if (owner != NULL) {
506 struct fuse_read_in *inarg = &req->misc.read.in;
507
508 inarg->read_flags |= FUSE_READ_LOCKOWNER;
509 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
510 }
511 fuse_request_send(fc, req);
512 return req->out.args[0].size;
513 }
514
515 static void fuse_read_update_size(struct inode *inode, loff_t size,
516 u64 attr_ver)
517 {
518 struct fuse_conn *fc = get_fuse_conn(inode);
519 struct fuse_inode *fi = get_fuse_inode(inode);
520
521 spin_lock(&fc->lock);
522 if (attr_ver == fi->attr_version && size < inode->i_size) {
523 fi->attr_version = ++fc->attr_version;
524 i_size_write(inode, size);
525 }
526 spin_unlock(&fc->lock);
527 }
528
529 static int fuse_readpage(struct file *file, struct page *page)
530 {
531 struct inode *inode = page->mapping->host;
532 struct fuse_conn *fc = get_fuse_conn(inode);
533 struct fuse_req *req;
534 size_t num_read;
535 loff_t pos = page_offset(page);
536 size_t count = PAGE_CACHE_SIZE;
537 u64 attr_ver;
538 int err;
539
540 err = -EIO;
541 if (is_bad_inode(inode))
542 goto out;
543
544 /*
545 * Page writeback can extend beyond the lifetime of the
546 * page-cache page, so make sure we read a properly synced
547 * page.
548 */
549 fuse_wait_on_page_writeback(inode, page->index);
550
551 req = fuse_get_req(fc);
552 err = PTR_ERR(req);
553 if (IS_ERR(req))
554 goto out;
555
556 attr_ver = fuse_get_attr_version(fc);
557
558 req->out.page_zeroing = 1;
559 req->out.argpages = 1;
560 req->num_pages = 1;
561 req->pages[0] = page;
562 num_read = fuse_send_read(req, file, pos, count, NULL);
563 err = req->out.h.error;
564 fuse_put_request(fc, req);
565
566 if (!err) {
567 /*
568 * Short read means EOF. If file size is larger, truncate it
569 */
570 if (num_read < count)
571 fuse_read_update_size(inode, pos + num_read, attr_ver);
572
573 SetPageUptodate(page);
574 }
575
576 fuse_invalidate_attr(inode); /* atime changed */
577 out:
578 unlock_page(page);
579 return err;
580 }
581
582 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
583 {
584 int i;
585 size_t count = req->misc.read.in.size;
586 size_t num_read = req->out.args[0].size;
587 struct address_space *mapping = NULL;
588
589 for (i = 0; mapping == NULL && i < req->num_pages; i++)
590 mapping = req->pages[i]->mapping;
591
592 if (mapping) {
593 struct inode *inode = mapping->host;
594
595 /*
596 * Short read means EOF. If file size is larger, truncate it
597 */
598 if (!req->out.h.error && num_read < count) {
599 loff_t pos;
600
601 pos = page_offset(req->pages[0]) + num_read;
602 fuse_read_update_size(inode, pos,
603 req->misc.read.attr_ver);
604 }
605 fuse_invalidate_attr(inode); /* atime changed */
606 }
607
608 for (i = 0; i < req->num_pages; i++) {
609 struct page *page = req->pages[i];
610 if (!req->out.h.error)
611 SetPageUptodate(page);
612 else
613 SetPageError(page);
614 unlock_page(page);
615 page_cache_release(page);
616 }
617 if (req->ff)
618 fuse_file_put(req->ff, false);
619 }
620
621 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
622 {
623 struct fuse_file *ff = file->private_data;
624 struct fuse_conn *fc = ff->fc;
625 loff_t pos = page_offset(req->pages[0]);
626 size_t count = req->num_pages << PAGE_CACHE_SHIFT;
627
628 req->out.argpages = 1;
629 req->out.page_zeroing = 1;
630 req->out.page_replace = 1;
631 fuse_read_fill(req, file, pos, count, FUSE_READ);
632 req->misc.read.attr_ver = fuse_get_attr_version(fc);
633 if (fc->async_read) {
634 req->ff = fuse_file_get(ff);
635 req->end = fuse_readpages_end;
636 fuse_request_send_background(fc, req);
637 } else {
638 fuse_request_send(fc, req);
639 fuse_readpages_end(fc, req);
640 fuse_put_request(fc, req);
641 }
642 }
643
644 struct fuse_fill_data {
645 struct fuse_req *req;
646 struct file *file;
647 struct inode *inode;
648 };
649
650 static int fuse_readpages_fill(void *_data, struct page *page)
651 {
652 struct fuse_fill_data *data = _data;
653 struct fuse_req *req = data->req;
654 struct inode *inode = data->inode;
655 struct fuse_conn *fc = get_fuse_conn(inode);
656
657 fuse_wait_on_page_writeback(inode, page->index);
658
659 if (req->num_pages &&
660 (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
661 (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
662 req->pages[req->num_pages - 1]->index + 1 != page->index)) {
663 fuse_send_readpages(req, data->file);
664 data->req = req = fuse_get_req(fc);
665 if (IS_ERR(req)) {
666 unlock_page(page);
667 return PTR_ERR(req);
668 }
669 }
670 page_cache_get(page);
671 req->pages[req->num_pages] = page;
672 req->num_pages++;
673 return 0;
674 }
675
676 static int fuse_readpages(struct file *file, struct address_space *mapping,
677 struct list_head *pages, unsigned nr_pages)
678 {
679 struct inode *inode = mapping->host;
680 struct fuse_conn *fc = get_fuse_conn(inode);
681 struct fuse_fill_data data;
682 int err;
683
684 err = -EIO;
685 if (is_bad_inode(inode))
686 goto out;
687
688 data.file = file;
689 data.inode = inode;
690 data.req = fuse_get_req(fc);
691 err = PTR_ERR(data.req);
692 if (IS_ERR(data.req))
693 goto out;
694
695 err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
696 if (!err) {
697 if (data.req->num_pages)
698 fuse_send_readpages(data.req, file);
699 else
700 fuse_put_request(fc, data.req);
701 }
702 out:
703 return err;
704 }
705
706 static ssize_t fuse_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
707 unsigned long nr_segs, loff_t pos)
708 {
709 struct inode *inode = iocb->ki_filp->f_mapping->host;
710
711 if (pos + iov_length(iov, nr_segs) > i_size_read(inode)) {
712 int err;
713 /*
714 * If trying to read past EOF, make sure the i_size
715 * attribute is up-to-date.
716 */
717 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
718 if (err)
719 return err;
720 }
721
722 return generic_file_aio_read(iocb, iov, nr_segs, pos);
723 }
724
725 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
726 loff_t pos, size_t count)
727 {
728 struct fuse_write_in *inarg = &req->misc.write.in;
729 struct fuse_write_out *outarg = &req->misc.write.out;
730
731 inarg->fh = ff->fh;
732 inarg->offset = pos;
733 inarg->size = count;
734 req->in.h.opcode = FUSE_WRITE;
735 req->in.h.nodeid = ff->nodeid;
736 req->in.numargs = 2;
737 if (ff->fc->minor < 9)
738 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
739 else
740 req->in.args[0].size = sizeof(struct fuse_write_in);
741 req->in.args[0].value = inarg;
742 req->in.args[1].size = count;
743 req->out.numargs = 1;
744 req->out.args[0].size = sizeof(struct fuse_write_out);
745 req->out.args[0].value = outarg;
746 }
747
748 static size_t fuse_send_write(struct fuse_req *req, struct file *file,
749 loff_t pos, size_t count, fl_owner_t owner)
750 {
751 struct fuse_file *ff = file->private_data;
752 struct fuse_conn *fc = ff->fc;
753 struct fuse_write_in *inarg = &req->misc.write.in;
754
755 fuse_write_fill(req, ff, pos, count);
756 inarg->flags = file->f_flags;
757 if (owner != NULL) {
758 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
759 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
760 }
761 fuse_request_send(fc, req);
762 return req->misc.write.out.size;
763 }
764
765 void fuse_write_update_size(struct inode *inode, loff_t pos)
766 {
767 struct fuse_conn *fc = get_fuse_conn(inode);
768 struct fuse_inode *fi = get_fuse_inode(inode);
769
770 spin_lock(&fc->lock);
771 fi->attr_version = ++fc->attr_version;
772 if (pos > inode->i_size)
773 i_size_write(inode, pos);
774 spin_unlock(&fc->lock);
775 }
776
777 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
778 struct inode *inode, loff_t pos,
779 size_t count)
780 {
781 size_t res;
782 unsigned offset;
783 unsigned i;
784
785 for (i = 0; i < req->num_pages; i++)
786 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
787
788 res = fuse_send_write(req, file, pos, count, NULL);
789
790 offset = req->page_offset;
791 count = res;
792 for (i = 0; i < req->num_pages; i++) {
793 struct page *page = req->pages[i];
794
795 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
796 SetPageUptodate(page);
797
798 if (count > PAGE_CACHE_SIZE - offset)
799 count -= PAGE_CACHE_SIZE - offset;
800 else
801 count = 0;
802 offset = 0;
803
804 unlock_page(page);
805 page_cache_release(page);
806 }
807
808 return res;
809 }
810
811 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
812 struct address_space *mapping,
813 struct iov_iter *ii, loff_t pos)
814 {
815 struct fuse_conn *fc = get_fuse_conn(mapping->host);
816 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
817 size_t count = 0;
818 int err;
819
820 req->in.argpages = 1;
821 req->page_offset = offset;
822
823 do {
824 size_t tmp;
825 struct page *page;
826 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
827 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
828 iov_iter_count(ii));
829
830 bytes = min_t(size_t, bytes, fc->max_write - count);
831
832 again:
833 err = -EFAULT;
834 if (iov_iter_fault_in_readable(ii, bytes))
835 break;
836
837 err = -ENOMEM;
838 page = grab_cache_page_write_begin(mapping, index, 0);
839 if (!page)
840 break;
841
842 if (mapping_writably_mapped(mapping))
843 flush_dcache_page(page);
844
845 pagefault_disable();
846 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
847 pagefault_enable();
848 flush_dcache_page(page);
849
850 mark_page_accessed(page);
851
852 if (!tmp) {
853 unlock_page(page);
854 page_cache_release(page);
855 bytes = min(bytes, iov_iter_single_seg_count(ii));
856 goto again;
857 }
858
859 err = 0;
860 req->pages[req->num_pages] = page;
861 req->num_pages++;
862
863 iov_iter_advance(ii, tmp);
864 count += tmp;
865 pos += tmp;
866 offset += tmp;
867 if (offset == PAGE_CACHE_SIZE)
868 offset = 0;
869
870 if (!fc->big_writes)
871 break;
872 } while (iov_iter_count(ii) && count < fc->max_write &&
873 req->num_pages < FUSE_MAX_PAGES_PER_REQ && offset == 0);
874
875 return count > 0 ? count : err;
876 }
877
878 static ssize_t fuse_perform_write(struct file *file,
879 struct address_space *mapping,
880 struct iov_iter *ii, loff_t pos)
881 {
882 struct inode *inode = mapping->host;
883 struct fuse_conn *fc = get_fuse_conn(inode);
884 int err = 0;
885 ssize_t res = 0;
886
887 if (is_bad_inode(inode))
888 return -EIO;
889
890 do {
891 struct fuse_req *req;
892 ssize_t count;
893
894 req = fuse_get_req(fc);
895 if (IS_ERR(req)) {
896 err = PTR_ERR(req);
897 break;
898 }
899
900 count = fuse_fill_write_pages(req, mapping, ii, pos);
901 if (count <= 0) {
902 err = count;
903 } else {
904 size_t num_written;
905
906 num_written = fuse_send_write_pages(req, file, inode,
907 pos, count);
908 err = req->out.h.error;
909 if (!err) {
910 res += num_written;
911 pos += num_written;
912
913 /* break out of the loop on short write */
914 if (num_written != count)
915 err = -EIO;
916 }
917 }
918 fuse_put_request(fc, req);
919 } while (!err && iov_iter_count(ii));
920
921 if (res > 0)
922 fuse_write_update_size(inode, pos);
923
924 fuse_invalidate_attr(inode);
925
926 return res > 0 ? res : err;
927 }
928
929 static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
930 unsigned long nr_segs, loff_t pos)
931 {
932 struct file *file = iocb->ki_filp;
933 struct address_space *mapping = file->f_mapping;
934 size_t count = 0;
935 ssize_t written = 0;
936 struct inode *inode = mapping->host;
937 ssize_t err;
938 struct iov_iter i;
939
940 WARN_ON(iocb->ki_pos != pos);
941
942 err = generic_segment_checks(iov, &nr_segs, &count, VERIFY_READ);
943 if (err)
944 return err;
945
946 mutex_lock(&inode->i_mutex);
947 vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
948
949 /* We can write back this queue in page reclaim */
950 current->backing_dev_info = mapping->backing_dev_info;
951
952 err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
953 if (err)
954 goto out;
955
956 if (count == 0)
957 goto out;
958
959 err = file_remove_suid(file);
960 if (err)
961 goto out;
962
963 file_update_time(file);
964
965 iov_iter_init(&i, iov, nr_segs, count, 0);
966 written = fuse_perform_write(file, mapping, &i, pos);
967 if (written >= 0)
968 iocb->ki_pos = pos + written;
969
970 out:
971 current->backing_dev_info = NULL;
972 mutex_unlock(&inode->i_mutex);
973
974 return written ? written : err;
975 }
976
977 static void fuse_release_user_pages(struct fuse_req *req, int write)
978 {
979 unsigned i;
980
981 for (i = 0; i < req->num_pages; i++) {
982 struct page *page = req->pages[i];
983 if (write)
984 set_page_dirty_lock(page);
985 put_page(page);
986 }
987 }
988
989 static int fuse_get_user_pages(struct fuse_req *req, const char __user *buf,
990 size_t *nbytesp, int write)
991 {
992 size_t nbytes = *nbytesp;
993 unsigned long user_addr = (unsigned long) buf;
994 unsigned offset = user_addr & ~PAGE_MASK;
995 int npages;
996
997 /* Special case for kernel I/O: can copy directly into the buffer */
998 if (segment_eq(get_fs(), KERNEL_DS)) {
999 if (write)
1000 req->in.args[1].value = (void *) user_addr;
1001 else
1002 req->out.args[0].value = (void *) user_addr;
1003
1004 return 0;
1005 }
1006
1007 nbytes = min_t(size_t, nbytes, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
1008 npages = (nbytes + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1009 npages = clamp(npages, 1, FUSE_MAX_PAGES_PER_REQ);
1010 npages = get_user_pages_fast(user_addr, npages, !write, req->pages);
1011 if (npages < 0)
1012 return npages;
1013
1014 req->num_pages = npages;
1015 req->page_offset = offset;
1016
1017 if (write)
1018 req->in.argpages = 1;
1019 else
1020 req->out.argpages = 1;
1021
1022 nbytes = (req->num_pages << PAGE_SHIFT) - req->page_offset;
1023 *nbytesp = min(*nbytesp, nbytes);
1024
1025 return 0;
1026 }
1027
1028 ssize_t fuse_direct_io(struct file *file, const char __user *buf,
1029 size_t count, loff_t *ppos, int write)
1030 {
1031 struct fuse_file *ff = file->private_data;
1032 struct fuse_conn *fc = ff->fc;
1033 size_t nmax = write ? fc->max_write : fc->max_read;
1034 loff_t pos = *ppos;
1035 ssize_t res = 0;
1036 struct fuse_req *req;
1037
1038 req = fuse_get_req(fc);
1039 if (IS_ERR(req))
1040 return PTR_ERR(req);
1041
1042 while (count) {
1043 size_t nres;
1044 fl_owner_t owner = current->files;
1045 size_t nbytes = min(count, nmax);
1046 int err = fuse_get_user_pages(req, buf, &nbytes, write);
1047 if (err) {
1048 res = err;
1049 break;
1050 }
1051
1052 if (write)
1053 nres = fuse_send_write(req, file, pos, nbytes, owner);
1054 else
1055 nres = fuse_send_read(req, file, pos, nbytes, owner);
1056
1057 fuse_release_user_pages(req, !write);
1058 if (req->out.h.error) {
1059 if (!res)
1060 res = req->out.h.error;
1061 break;
1062 } else if (nres > nbytes) {
1063 res = -EIO;
1064 break;
1065 }
1066 count -= nres;
1067 res += nres;
1068 pos += nres;
1069 buf += nres;
1070 if (nres != nbytes)
1071 break;
1072 if (count) {
1073 fuse_put_request(fc, req);
1074 req = fuse_get_req(fc);
1075 if (IS_ERR(req))
1076 break;
1077 }
1078 }
1079 if (!IS_ERR(req))
1080 fuse_put_request(fc, req);
1081 if (res > 0)
1082 *ppos = pos;
1083
1084 return res;
1085 }
1086 EXPORT_SYMBOL_GPL(fuse_direct_io);
1087
1088 static ssize_t fuse_direct_read(struct file *file, char __user *buf,
1089 size_t count, loff_t *ppos)
1090 {
1091 ssize_t res;
1092 struct inode *inode = file->f_path.dentry->d_inode;
1093
1094 if (is_bad_inode(inode))
1095 return -EIO;
1096
1097 res = fuse_direct_io(file, buf, count, ppos, 0);
1098
1099 fuse_invalidate_attr(inode);
1100
1101 return res;
1102 }
1103
1104 static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
1105 size_t count, loff_t *ppos)
1106 {
1107 struct inode *inode = file->f_path.dentry->d_inode;
1108 ssize_t res;
1109
1110 if (is_bad_inode(inode))
1111 return -EIO;
1112
1113 /* Don't allow parallel writes to the same file */
1114 mutex_lock(&inode->i_mutex);
1115 res = generic_write_checks(file, ppos, &count, 0);
1116 if (!res) {
1117 res = fuse_direct_io(file, buf, count, ppos, 1);
1118 if (res > 0)
1119 fuse_write_update_size(inode, *ppos);
1120 }
1121 mutex_unlock(&inode->i_mutex);
1122
1123 fuse_invalidate_attr(inode);
1124
1125 return res;
1126 }
1127
1128 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1129 {
1130 __free_page(req->pages[0]);
1131 fuse_file_put(req->ff, false);
1132 }
1133
1134 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1135 {
1136 struct inode *inode = req->inode;
1137 struct fuse_inode *fi = get_fuse_inode(inode);
1138 struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
1139
1140 list_del(&req->writepages_entry);
1141 dec_bdi_stat(bdi, BDI_WRITEBACK);
1142 dec_zone_page_state(req->pages[0], NR_WRITEBACK_TEMP);
1143 bdi_writeout_inc(bdi);
1144 wake_up(&fi->page_waitq);
1145 }
1146
1147 /* Called under fc->lock, may release and reacquire it */
1148 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req)
1149 __releases(fc->lock)
1150 __acquires(fc->lock)
1151 {
1152 struct fuse_inode *fi = get_fuse_inode(req->inode);
1153 loff_t size = i_size_read(req->inode);
1154 struct fuse_write_in *inarg = &req->misc.write.in;
1155
1156 if (!fc->connected)
1157 goto out_free;
1158
1159 if (inarg->offset + PAGE_CACHE_SIZE <= size) {
1160 inarg->size = PAGE_CACHE_SIZE;
1161 } else if (inarg->offset < size) {
1162 inarg->size = size & (PAGE_CACHE_SIZE - 1);
1163 } else {
1164 /* Got truncated off completely */
1165 goto out_free;
1166 }
1167
1168 req->in.args[1].size = inarg->size;
1169 fi->writectr++;
1170 fuse_request_send_background_locked(fc, req);
1171 return;
1172
1173 out_free:
1174 fuse_writepage_finish(fc, req);
1175 spin_unlock(&fc->lock);
1176 fuse_writepage_free(fc, req);
1177 fuse_put_request(fc, req);
1178 spin_lock(&fc->lock);
1179 }
1180
1181 /*
1182 * If fi->writectr is positive (no truncate or fsync going on) send
1183 * all queued writepage requests.
1184 *
1185 * Called with fc->lock
1186 */
1187 void fuse_flush_writepages(struct inode *inode)
1188 __releases(fc->lock)
1189 __acquires(fc->lock)
1190 {
1191 struct fuse_conn *fc = get_fuse_conn(inode);
1192 struct fuse_inode *fi = get_fuse_inode(inode);
1193 struct fuse_req *req;
1194
1195 while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1196 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1197 list_del_init(&req->list);
1198 fuse_send_writepage(fc, req);
1199 }
1200 }
1201
1202 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1203 {
1204 struct inode *inode = req->inode;
1205 struct fuse_inode *fi = get_fuse_inode(inode);
1206
1207 mapping_set_error(inode->i_mapping, req->out.h.error);
1208 spin_lock(&fc->lock);
1209 fi->writectr--;
1210 fuse_writepage_finish(fc, req);
1211 spin_unlock(&fc->lock);
1212 fuse_writepage_free(fc, req);
1213 }
1214
1215 static int fuse_writepage_locked(struct page *page)
1216 {
1217 struct address_space *mapping = page->mapping;
1218 struct inode *inode = mapping->host;
1219 struct fuse_conn *fc = get_fuse_conn(inode);
1220 struct fuse_inode *fi = get_fuse_inode(inode);
1221 struct fuse_req *req;
1222 struct fuse_file *ff;
1223 struct page *tmp_page;
1224
1225 set_page_writeback(page);
1226
1227 req = fuse_request_alloc_nofs();
1228 if (!req)
1229 goto err;
1230
1231 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1232 if (!tmp_page)
1233 goto err_free;
1234
1235 spin_lock(&fc->lock);
1236 BUG_ON(list_empty(&fi->write_files));
1237 ff = list_entry(fi->write_files.next, struct fuse_file, write_entry);
1238 req->ff = fuse_file_get(ff);
1239 spin_unlock(&fc->lock);
1240
1241 fuse_write_fill(req, ff, page_offset(page), 0);
1242
1243 copy_highpage(tmp_page, page);
1244 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1245 req->in.argpages = 1;
1246 req->num_pages = 1;
1247 req->pages[0] = tmp_page;
1248 req->page_offset = 0;
1249 req->end = fuse_writepage_end;
1250 req->inode = inode;
1251
1252 inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
1253 inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1254 end_page_writeback(page);
1255
1256 spin_lock(&fc->lock);
1257 list_add(&req->writepages_entry, &fi->writepages);
1258 list_add_tail(&req->list, &fi->queued_writes);
1259 fuse_flush_writepages(inode);
1260 spin_unlock(&fc->lock);
1261
1262 return 0;
1263
1264 err_free:
1265 fuse_request_free(req);
1266 err:
1267 end_page_writeback(page);
1268 return -ENOMEM;
1269 }
1270
1271 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1272 {
1273 int err;
1274
1275 err = fuse_writepage_locked(page);
1276 unlock_page(page);
1277
1278 return err;
1279 }
1280
1281 static int fuse_launder_page(struct page *page)
1282 {
1283 int err = 0;
1284 if (clear_page_dirty_for_io(page)) {
1285 struct inode *inode = page->mapping->host;
1286 err = fuse_writepage_locked(page);
1287 if (!err)
1288 fuse_wait_on_page_writeback(inode, page->index);
1289 }
1290 return err;
1291 }
1292
1293 /*
1294 * Write back dirty pages now, because there may not be any suitable
1295 * open files later
1296 */
1297 static void fuse_vma_close(struct vm_area_struct *vma)
1298 {
1299 filemap_write_and_wait(vma->vm_file->f_mapping);
1300 }
1301
1302 /*
1303 * Wait for writeback against this page to complete before allowing it
1304 * to be marked dirty again, and hence written back again, possibly
1305 * before the previous writepage completed.
1306 *
1307 * Block here, instead of in ->writepage(), so that the userspace fs
1308 * can only block processes actually operating on the filesystem.
1309 *
1310 * Otherwise unprivileged userspace fs would be able to block
1311 * unrelated:
1312 *
1313 * - page migration
1314 * - sync(2)
1315 * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
1316 */
1317 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1318 {
1319 struct page *page = vmf->page;
1320 /*
1321 * Don't use page->mapping as it may become NULL from a
1322 * concurrent truncate.
1323 */
1324 struct inode *inode = vma->vm_file->f_mapping->host;
1325
1326 fuse_wait_on_page_writeback(inode, page->index);
1327 return 0;
1328 }
1329
1330 static const struct vm_operations_struct fuse_file_vm_ops = {
1331 .close = fuse_vma_close,
1332 .fault = filemap_fault,
1333 .page_mkwrite = fuse_page_mkwrite,
1334 };
1335
1336 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
1337 {
1338 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE)) {
1339 struct inode *inode = file->f_dentry->d_inode;
1340 struct fuse_conn *fc = get_fuse_conn(inode);
1341 struct fuse_inode *fi = get_fuse_inode(inode);
1342 struct fuse_file *ff = file->private_data;
1343 /*
1344 * file may be written through mmap, so chain it onto the
1345 * inodes's write_file list
1346 */
1347 spin_lock(&fc->lock);
1348 if (list_empty(&ff->write_entry))
1349 list_add(&ff->write_entry, &fi->write_files);
1350 spin_unlock(&fc->lock);
1351 }
1352 file_accessed(file);
1353 vma->vm_ops = &fuse_file_vm_ops;
1354 return 0;
1355 }
1356
1357 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
1358 {
1359 /* Can't provide the coherency needed for MAP_SHARED */
1360 if (vma->vm_flags & VM_MAYSHARE)
1361 return -ENODEV;
1362
1363 invalidate_inode_pages2(file->f_mapping);
1364
1365 return generic_file_mmap(file, vma);
1366 }
1367
1368 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
1369 struct file_lock *fl)
1370 {
1371 switch (ffl->type) {
1372 case F_UNLCK:
1373 break;
1374
1375 case F_RDLCK:
1376 case F_WRLCK:
1377 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
1378 ffl->end < ffl->start)
1379 return -EIO;
1380
1381 fl->fl_start = ffl->start;
1382 fl->fl_end = ffl->end;
1383 fl->fl_pid = ffl->pid;
1384 break;
1385
1386 default:
1387 return -EIO;
1388 }
1389 fl->fl_type = ffl->type;
1390 return 0;
1391 }
1392
1393 static void fuse_lk_fill(struct fuse_req *req, struct file *file,
1394 const struct file_lock *fl, int opcode, pid_t pid,
1395 int flock)
1396 {
1397 struct inode *inode = file->f_path.dentry->d_inode;
1398 struct fuse_conn *fc = get_fuse_conn(inode);
1399 struct fuse_file *ff = file->private_data;
1400 struct fuse_lk_in *arg = &req->misc.lk_in;
1401
1402 arg->fh = ff->fh;
1403 arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
1404 arg->lk.start = fl->fl_start;
1405 arg->lk.end = fl->fl_end;
1406 arg->lk.type = fl->fl_type;
1407 arg->lk.pid = pid;
1408 if (flock)
1409 arg->lk_flags |= FUSE_LK_FLOCK;
1410 req->in.h.opcode = opcode;
1411 req->in.h.nodeid = get_node_id(inode);
1412 req->in.numargs = 1;
1413 req->in.args[0].size = sizeof(*arg);
1414 req->in.args[0].value = arg;
1415 }
1416
1417 static int fuse_getlk(struct file *file, struct file_lock *fl)
1418 {
1419 struct inode *inode = file->f_path.dentry->d_inode;
1420 struct fuse_conn *fc = get_fuse_conn(inode);
1421 struct fuse_req *req;
1422 struct fuse_lk_out outarg;
1423 int err;
1424
1425 req = fuse_get_req(fc);
1426 if (IS_ERR(req))
1427 return PTR_ERR(req);
1428
1429 fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
1430 req->out.numargs = 1;
1431 req->out.args[0].size = sizeof(outarg);
1432 req->out.args[0].value = &outarg;
1433 fuse_request_send(fc, req);
1434 err = req->out.h.error;
1435 fuse_put_request(fc, req);
1436 if (!err)
1437 err = convert_fuse_file_lock(&outarg.lk, fl);
1438
1439 return err;
1440 }
1441
1442 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
1443 {
1444 struct inode *inode = file->f_path.dentry->d_inode;
1445 struct fuse_conn *fc = get_fuse_conn(inode);
1446 struct fuse_req *req;
1447 int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
1448 pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
1449 int err;
1450
1451 if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
1452 /* NLM needs asynchronous locks, which we don't support yet */
1453 return -ENOLCK;
1454 }
1455
1456 /* Unlock on close is handled by the flush method */
1457 if (fl->fl_flags & FL_CLOSE)
1458 return 0;
1459
1460 req = fuse_get_req(fc);
1461 if (IS_ERR(req))
1462 return PTR_ERR(req);
1463
1464 fuse_lk_fill(req, file, fl, opcode, pid, flock);
1465 fuse_request_send(fc, req);
1466 err = req->out.h.error;
1467 /* locking is restartable */
1468 if (err == -EINTR)
1469 err = -ERESTARTSYS;
1470 fuse_put_request(fc, req);
1471 return err;
1472 }
1473
1474 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
1475 {
1476 struct inode *inode = file->f_path.dentry->d_inode;
1477 struct fuse_conn *fc = get_fuse_conn(inode);
1478 int err;
1479
1480 if (cmd == F_CANCELLK) {
1481 err = 0;
1482 } else if (cmd == F_GETLK) {
1483 if (fc->no_lock) {
1484 posix_test_lock(file, fl);
1485 err = 0;
1486 } else
1487 err = fuse_getlk(file, fl);
1488 } else {
1489 if (fc->no_lock)
1490 err = posix_lock_file(file, fl, NULL);
1491 else
1492 err = fuse_setlk(file, fl, 0);
1493 }
1494 return err;
1495 }
1496
1497 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
1498 {
1499 struct inode *inode = file->f_path.dentry->d_inode;
1500 struct fuse_conn *fc = get_fuse_conn(inode);
1501 int err;
1502
1503 if (fc->no_flock) {
1504 err = flock_lock_file_wait(file, fl);
1505 } else {
1506 struct fuse_file *ff = file->private_data;
1507
1508 /* emulate flock with POSIX locks */
1509 fl->fl_owner = (fl_owner_t) file;
1510 ff->flock = true;
1511 err = fuse_setlk(file, fl, 1);
1512 }
1513
1514 return err;
1515 }
1516
1517 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
1518 {
1519 struct inode *inode = mapping->host;
1520 struct fuse_conn *fc = get_fuse_conn(inode);
1521 struct fuse_req *req;
1522 struct fuse_bmap_in inarg;
1523 struct fuse_bmap_out outarg;
1524 int err;
1525
1526 if (!inode->i_sb->s_bdev || fc->no_bmap)
1527 return 0;
1528
1529 req = fuse_get_req(fc);
1530 if (IS_ERR(req))
1531 return 0;
1532
1533 memset(&inarg, 0, sizeof(inarg));
1534 inarg.block = block;
1535 inarg.blocksize = inode->i_sb->s_blocksize;
1536 req->in.h.opcode = FUSE_BMAP;
1537 req->in.h.nodeid = get_node_id(inode);
1538 req->in.numargs = 1;
1539 req->in.args[0].size = sizeof(inarg);
1540 req->in.args[0].value = &inarg;
1541 req->out.numargs = 1;
1542 req->out.args[0].size = sizeof(outarg);
1543 req->out.args[0].value = &outarg;
1544 fuse_request_send(fc, req);
1545 err = req->out.h.error;
1546 fuse_put_request(fc, req);
1547 if (err == -ENOSYS)
1548 fc->no_bmap = 1;
1549
1550 return err ? 0 : outarg.block;
1551 }
1552
1553 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int origin)
1554 {
1555 loff_t retval;
1556 struct inode *inode = file->f_path.dentry->d_inode;
1557
1558 /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
1559 if (origin == SEEK_CUR || origin == SEEK_SET)
1560 return generic_file_llseek(file, offset, origin);
1561
1562 mutex_lock(&inode->i_mutex);
1563 retval = fuse_update_attributes(inode, NULL, file, NULL);
1564 if (!retval)
1565 retval = generic_file_llseek(file, offset, origin);
1566 mutex_unlock(&inode->i_mutex);
1567
1568 return retval;
1569 }
1570
1571 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
1572 unsigned int nr_segs, size_t bytes, bool to_user)
1573 {
1574 struct iov_iter ii;
1575 int page_idx = 0;
1576
1577 if (!bytes)
1578 return 0;
1579
1580 iov_iter_init(&ii, iov, nr_segs, bytes, 0);
1581
1582 while (iov_iter_count(&ii)) {
1583 struct page *page = pages[page_idx++];
1584 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
1585 void *kaddr;
1586
1587 kaddr = kmap(page);
1588
1589 while (todo) {
1590 char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
1591 size_t iov_len = ii.iov->iov_len - ii.iov_offset;
1592 size_t copy = min(todo, iov_len);
1593 size_t left;
1594
1595 if (!to_user)
1596 left = copy_from_user(kaddr, uaddr, copy);
1597 else
1598 left = copy_to_user(uaddr, kaddr, copy);
1599
1600 if (unlikely(left))
1601 return -EFAULT;
1602
1603 iov_iter_advance(&ii, copy);
1604 todo -= copy;
1605 kaddr += copy;
1606 }
1607
1608 kunmap(page);
1609 }
1610
1611 return 0;
1612 }
1613
1614 /*
1615 * CUSE servers compiled on 32bit broke on 64bit kernels because the
1616 * ABI was defined to be 'struct iovec' which is different on 32bit
1617 * and 64bit. Fortunately we can determine which structure the server
1618 * used from the size of the reply.
1619 */
1620 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
1621 size_t transferred, unsigned count,
1622 bool is_compat)
1623 {
1624 #ifdef CONFIG_COMPAT
1625 if (count * sizeof(struct compat_iovec) == transferred) {
1626 struct compat_iovec *ciov = src;
1627 unsigned i;
1628
1629 /*
1630 * With this interface a 32bit server cannot support
1631 * non-compat (i.e. ones coming from 64bit apps) ioctl
1632 * requests
1633 */
1634 if (!is_compat)
1635 return -EINVAL;
1636
1637 for (i = 0; i < count; i++) {
1638 dst[i].iov_base = compat_ptr(ciov[i].iov_base);
1639 dst[i].iov_len = ciov[i].iov_len;
1640 }
1641 return 0;
1642 }
1643 #endif
1644
1645 if (count * sizeof(struct iovec) != transferred)
1646 return -EIO;
1647
1648 memcpy(dst, src, transferred);
1649 return 0;
1650 }
1651
1652 /* Make sure iov_length() won't overflow */
1653 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
1654 {
1655 size_t n;
1656 u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
1657
1658 for (n = 0; n < count; n++) {
1659 if (iov->iov_len > (size_t) max)
1660 return -ENOMEM;
1661 max -= iov->iov_len;
1662 }
1663 return 0;
1664 }
1665
1666 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
1667 void *src, size_t transferred, unsigned count,
1668 bool is_compat)
1669 {
1670 unsigned i;
1671 struct fuse_ioctl_iovec *fiov = src;
1672
1673 if (fc->minor < 16) {
1674 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
1675 count, is_compat);
1676 }
1677
1678 if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
1679 return -EIO;
1680
1681 for (i = 0; i < count; i++) {
1682 /* Did the server supply an inappropriate value? */
1683 if (fiov[i].base != (unsigned long) fiov[i].base ||
1684 fiov[i].len != (unsigned long) fiov[i].len)
1685 return -EIO;
1686
1687 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
1688 dst[i].iov_len = (size_t) fiov[i].len;
1689
1690 #ifdef CONFIG_COMPAT
1691 if (is_compat &&
1692 (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
1693 (compat_size_t) dst[i].iov_len != fiov[i].len))
1694 return -EIO;
1695 #endif
1696 }
1697
1698 return 0;
1699 }
1700
1701
1702 /*
1703 * For ioctls, there is no generic way to determine how much memory
1704 * needs to be read and/or written. Furthermore, ioctls are allowed
1705 * to dereference the passed pointer, so the parameter requires deep
1706 * copying but FUSE has no idea whatsoever about what to copy in or
1707 * out.
1708 *
1709 * This is solved by allowing FUSE server to retry ioctl with
1710 * necessary in/out iovecs. Let's assume the ioctl implementation
1711 * needs to read in the following structure.
1712 *
1713 * struct a {
1714 * char *buf;
1715 * size_t buflen;
1716 * }
1717 *
1718 * On the first callout to FUSE server, inarg->in_size and
1719 * inarg->out_size will be NULL; then, the server completes the ioctl
1720 * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
1721 * the actual iov array to
1722 *
1723 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) } }
1724 *
1725 * which tells FUSE to copy in the requested area and retry the ioctl.
1726 * On the second round, the server has access to the structure and
1727 * from that it can tell what to look for next, so on the invocation,
1728 * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
1729 *
1730 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) },
1731 * { .iov_base = a.buf, .iov_len = a.buflen } }
1732 *
1733 * FUSE will copy both struct a and the pointed buffer from the
1734 * process doing the ioctl and retry ioctl with both struct a and the
1735 * buffer.
1736 *
1737 * This time, FUSE server has everything it needs and completes ioctl
1738 * without FUSE_IOCTL_RETRY which finishes the ioctl call.
1739 *
1740 * Copying data out works the same way.
1741 *
1742 * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
1743 * automatically initializes in and out iovs by decoding @cmd with
1744 * _IOC_* macros and the server is not allowed to request RETRY. This
1745 * limits ioctl data transfers to well-formed ioctls and is the forced
1746 * behavior for all FUSE servers.
1747 */
1748 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
1749 unsigned int flags)
1750 {
1751 struct fuse_file *ff = file->private_data;
1752 struct fuse_conn *fc = ff->fc;
1753 struct fuse_ioctl_in inarg = {
1754 .fh = ff->fh,
1755 .cmd = cmd,
1756 .arg = arg,
1757 .flags = flags
1758 };
1759 struct fuse_ioctl_out outarg;
1760 struct fuse_req *req = NULL;
1761 struct page **pages = NULL;
1762 struct iovec *iov_page = NULL;
1763 struct iovec *in_iov = NULL, *out_iov = NULL;
1764 unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
1765 size_t in_size, out_size, transferred;
1766 int err;
1767
1768 #if BITS_PER_LONG == 32
1769 inarg.flags |= FUSE_IOCTL_32BIT;
1770 #else
1771 if (flags & FUSE_IOCTL_COMPAT)
1772 inarg.flags |= FUSE_IOCTL_32BIT;
1773 #endif
1774
1775 /* assume all the iovs returned by client always fits in a page */
1776 BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
1777
1778 err = -ENOMEM;
1779 pages = kzalloc(sizeof(pages[0]) * FUSE_MAX_PAGES_PER_REQ, GFP_KERNEL);
1780 iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
1781 if (!pages || !iov_page)
1782 goto out;
1783
1784 /*
1785 * If restricted, initialize IO parameters as encoded in @cmd.
1786 * RETRY from server is not allowed.
1787 */
1788 if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
1789 struct iovec *iov = iov_page;
1790
1791 iov->iov_base = (void __user *)arg;
1792 iov->iov_len = _IOC_SIZE(cmd);
1793
1794 if (_IOC_DIR(cmd) & _IOC_WRITE) {
1795 in_iov = iov;
1796 in_iovs = 1;
1797 }
1798
1799 if (_IOC_DIR(cmd) & _IOC_READ) {
1800 out_iov = iov;
1801 out_iovs = 1;
1802 }
1803 }
1804
1805 retry:
1806 inarg.in_size = in_size = iov_length(in_iov, in_iovs);
1807 inarg.out_size = out_size = iov_length(out_iov, out_iovs);
1808
1809 /*
1810 * Out data can be used either for actual out data or iovs,
1811 * make sure there always is at least one page.
1812 */
1813 out_size = max_t(size_t, out_size, PAGE_SIZE);
1814 max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
1815
1816 /* make sure there are enough buffer pages and init request with them */
1817 err = -ENOMEM;
1818 if (max_pages > FUSE_MAX_PAGES_PER_REQ)
1819 goto out;
1820 while (num_pages < max_pages) {
1821 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
1822 if (!pages[num_pages])
1823 goto out;
1824 num_pages++;
1825 }
1826
1827 req = fuse_get_req(fc);
1828 if (IS_ERR(req)) {
1829 err = PTR_ERR(req);
1830 req = NULL;
1831 goto out;
1832 }
1833 memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
1834 req->num_pages = num_pages;
1835
1836 /* okay, let's send it to the client */
1837 req->in.h.opcode = FUSE_IOCTL;
1838 req->in.h.nodeid = ff->nodeid;
1839 req->in.numargs = 1;
1840 req->in.args[0].size = sizeof(inarg);
1841 req->in.args[0].value = &inarg;
1842 if (in_size) {
1843 req->in.numargs++;
1844 req->in.args[1].size = in_size;
1845 req->in.argpages = 1;
1846
1847 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
1848 false);
1849 if (err)
1850 goto out;
1851 }
1852
1853 req->out.numargs = 2;
1854 req->out.args[0].size = sizeof(outarg);
1855 req->out.args[0].value = &outarg;
1856 req->out.args[1].size = out_size;
1857 req->out.argpages = 1;
1858 req->out.argvar = 1;
1859
1860 fuse_request_send(fc, req);
1861 err = req->out.h.error;
1862 transferred = req->out.args[1].size;
1863 fuse_put_request(fc, req);
1864 req = NULL;
1865 if (err)
1866 goto out;
1867
1868 /* did it ask for retry? */
1869 if (outarg.flags & FUSE_IOCTL_RETRY) {
1870 void *vaddr;
1871
1872 /* no retry if in restricted mode */
1873 err = -EIO;
1874 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
1875 goto out;
1876
1877 in_iovs = outarg.in_iovs;
1878 out_iovs = outarg.out_iovs;
1879
1880 /*
1881 * Make sure things are in boundary, separate checks
1882 * are to protect against overflow.
1883 */
1884 err = -ENOMEM;
1885 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
1886 out_iovs > FUSE_IOCTL_MAX_IOV ||
1887 in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
1888 goto out;
1889
1890 vaddr = kmap_atomic(pages[0], KM_USER0);
1891 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
1892 transferred, in_iovs + out_iovs,
1893 (flags & FUSE_IOCTL_COMPAT) != 0);
1894 kunmap_atomic(vaddr, KM_USER0);
1895 if (err)
1896 goto out;
1897
1898 in_iov = iov_page;
1899 out_iov = in_iov + in_iovs;
1900
1901 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
1902 if (err)
1903 goto out;
1904
1905 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
1906 if (err)
1907 goto out;
1908
1909 goto retry;
1910 }
1911
1912 err = -EIO;
1913 if (transferred > inarg.out_size)
1914 goto out;
1915
1916 err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
1917 out:
1918 if (req)
1919 fuse_put_request(fc, req);
1920 free_page((unsigned long) iov_page);
1921 while (num_pages)
1922 __free_page(pages[--num_pages]);
1923 kfree(pages);
1924
1925 return err ? err : outarg.result;
1926 }
1927 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
1928
1929 static long fuse_file_ioctl_common(struct file *file, unsigned int cmd,
1930 unsigned long arg, unsigned int flags)
1931 {
1932 struct inode *inode = file->f_dentry->d_inode;
1933 struct fuse_conn *fc = get_fuse_conn(inode);
1934
1935 if (!fuse_allow_task(fc, current))
1936 return -EACCES;
1937
1938 if (is_bad_inode(inode))
1939 return -EIO;
1940
1941 return fuse_do_ioctl(file, cmd, arg, flags);
1942 }
1943
1944 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
1945 unsigned long arg)
1946 {
1947 return fuse_file_ioctl_common(file, cmd, arg, 0);
1948 }
1949
1950 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
1951 unsigned long arg)
1952 {
1953 return fuse_file_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
1954 }
1955
1956 /*
1957 * All files which have been polled are linked to RB tree
1958 * fuse_conn->polled_files which is indexed by kh. Walk the tree and
1959 * find the matching one.
1960 */
1961 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
1962 struct rb_node **parent_out)
1963 {
1964 struct rb_node **link = &fc->polled_files.rb_node;
1965 struct rb_node *last = NULL;
1966
1967 while (*link) {
1968 struct fuse_file *ff;
1969
1970 last = *link;
1971 ff = rb_entry(last, struct fuse_file, polled_node);
1972
1973 if (kh < ff->kh)
1974 link = &last->rb_left;
1975 else if (kh > ff->kh)
1976 link = &last->rb_right;
1977 else
1978 return link;
1979 }
1980
1981 if (parent_out)
1982 *parent_out = last;
1983 return link;
1984 }
1985
1986 /*
1987 * The file is about to be polled. Make sure it's on the polled_files
1988 * RB tree. Note that files once added to the polled_files tree are
1989 * not removed before the file is released. This is because a file
1990 * polled once is likely to be polled again.
1991 */
1992 static void fuse_register_polled_file(struct fuse_conn *fc,
1993 struct fuse_file *ff)
1994 {
1995 spin_lock(&fc->lock);
1996 if (RB_EMPTY_NODE(&ff->polled_node)) {
1997 struct rb_node **link, *parent;
1998
1999 link = fuse_find_polled_node(fc, ff->kh, &parent);
2000 BUG_ON(*link);
2001 rb_link_node(&ff->polled_node, parent, link);
2002 rb_insert_color(&ff->polled_node, &fc->polled_files);
2003 }
2004 spin_unlock(&fc->lock);
2005 }
2006
2007 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2008 {
2009 struct fuse_file *ff = file->private_data;
2010 struct fuse_conn *fc = ff->fc;
2011 struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2012 struct fuse_poll_out outarg;
2013 struct fuse_req *req;
2014 int err;
2015
2016 if (fc->no_poll)
2017 return DEFAULT_POLLMASK;
2018
2019 poll_wait(file, &ff->poll_wait, wait);
2020
2021 /*
2022 * Ask for notification iff there's someone waiting for it.
2023 * The client may ignore the flag and always notify.
2024 */
2025 if (waitqueue_active(&ff->poll_wait)) {
2026 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2027 fuse_register_polled_file(fc, ff);
2028 }
2029
2030 req = fuse_get_req(fc);
2031 if (IS_ERR(req))
2032 return POLLERR;
2033
2034 req->in.h.opcode = FUSE_POLL;
2035 req->in.h.nodeid = ff->nodeid;
2036 req->in.numargs = 1;
2037 req->in.args[0].size = sizeof(inarg);
2038 req->in.args[0].value = &inarg;
2039 req->out.numargs = 1;
2040 req->out.args[0].size = sizeof(outarg);
2041 req->out.args[0].value = &outarg;
2042 fuse_request_send(fc, req);
2043 err = req->out.h.error;
2044 fuse_put_request(fc, req);
2045
2046 if (!err)
2047 return outarg.revents;
2048 if (err == -ENOSYS) {
2049 fc->no_poll = 1;
2050 return DEFAULT_POLLMASK;
2051 }
2052 return POLLERR;
2053 }
2054 EXPORT_SYMBOL_GPL(fuse_file_poll);
2055
2056 /*
2057 * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2058 * wakes up the poll waiters.
2059 */
2060 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2061 struct fuse_notify_poll_wakeup_out *outarg)
2062 {
2063 u64 kh = outarg->kh;
2064 struct rb_node **link;
2065
2066 spin_lock(&fc->lock);
2067
2068 link = fuse_find_polled_node(fc, kh, NULL);
2069 if (*link) {
2070 struct fuse_file *ff;
2071
2072 ff = rb_entry(*link, struct fuse_file, polled_node);
2073 wake_up_interruptible_sync(&ff->poll_wait);
2074 }
2075
2076 spin_unlock(&fc->lock);
2077 return 0;
2078 }
2079
2080 static const struct file_operations fuse_file_operations = {
2081 .llseek = fuse_file_llseek,
2082 .read = do_sync_read,
2083 .aio_read = fuse_file_aio_read,
2084 .write = do_sync_write,
2085 .aio_write = fuse_file_aio_write,
2086 .mmap = fuse_file_mmap,
2087 .open = fuse_open,
2088 .flush = fuse_flush,
2089 .release = fuse_release,
2090 .fsync = fuse_fsync,
2091 .lock = fuse_file_lock,
2092 .flock = fuse_file_flock,
2093 .splice_read = generic_file_splice_read,
2094 .unlocked_ioctl = fuse_file_ioctl,
2095 .compat_ioctl = fuse_file_compat_ioctl,
2096 .poll = fuse_file_poll,
2097 };
2098
2099 static const struct file_operations fuse_direct_io_file_operations = {
2100 .llseek = fuse_file_llseek,
2101 .read = fuse_direct_read,
2102 .write = fuse_direct_write,
2103 .mmap = fuse_direct_mmap,
2104 .open = fuse_open,
2105 .flush = fuse_flush,
2106 .release = fuse_release,
2107 .fsync = fuse_fsync,
2108 .lock = fuse_file_lock,
2109 .flock = fuse_file_flock,
2110 .unlocked_ioctl = fuse_file_ioctl,
2111 .compat_ioctl = fuse_file_compat_ioctl,
2112 .poll = fuse_file_poll,
2113 /* no splice_read */
2114 };
2115
2116 static const struct address_space_operations fuse_file_aops = {
2117 .readpage = fuse_readpage,
2118 .writepage = fuse_writepage,
2119 .launder_page = fuse_launder_page,
2120 .readpages = fuse_readpages,
2121 .set_page_dirty = __set_page_dirty_nobuffers,
2122 .bmap = fuse_bmap,
2123 };
2124
2125 void fuse_init_file_inode(struct inode *inode)
2126 {
2127 inode->i_fop = &fuse_file_operations;
2128 inode->i_data.a_ops = &fuse_file_aops;
2129 }
Linux 2.6 ibm-acpi/thinkpad-acpi driver git tree