]> git.proxmox.com Git - mirror_ubuntu-jammy-kernel.git/blob - fs/fuse/file.c
projects / mirror_ubuntu-jammy-kernel.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Merge tag 'fs.close_range.v5.15' of git://git.kernel.org/pub/scm/linux/kernel/git...
[mirror_ubuntu-jammy-kernel.git] / fs / fuse / file.c
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/sched/signal.h>
16 #include <linux/module.h>
17 #include <linux/swap.h>
18 #include <linux/falloc.h>
19 #include <linux/uio.h>
20 #include <linux/fs.h>
21
22 static int fuse_send_open(struct fuse_mount *fm, u64 nodeid,
23 unsigned int open_flags, int opcode,
24 struct fuse_open_out *outargp)
25 {
26 struct fuse_open_in inarg;
27 FUSE_ARGS(args);
28
29 memset(&inarg, 0, sizeof(inarg));
30 inarg.flags = open_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
31 if (!fm->fc->atomic_o_trunc)
32 inarg.flags &= ~O_TRUNC;
33
34 if (fm->fc->handle_killpriv_v2 &&
35 (inarg.flags & O_TRUNC) && !capable(CAP_FSETID)) {
36 inarg.open_flags |= FUSE_OPEN_KILL_SUIDGID;
37 }
38
39 args.opcode = opcode;
40 args.nodeid = nodeid;
41 args.in_numargs = 1;
42 args.in_args[0].size = sizeof(inarg);
43 args.in_args[0].value = &inarg;
44 args.out_numargs = 1;
45 args.out_args[0].size = sizeof(*outargp);
46 args.out_args[0].value = outargp;
47
48 return fuse_simple_request(fm, &args);
49 }
50
51 struct fuse_release_args {
52 struct fuse_args args;
53 struct fuse_release_in inarg;
54 struct inode *inode;
55 };
56
57 struct fuse_file *fuse_file_alloc(struct fuse_mount *fm)
58 {
59 struct fuse_file *ff;
60
61 ff = kzalloc(sizeof(struct fuse_file), GFP_KERNEL_ACCOUNT);
62 if (unlikely(!ff))
63 return NULL;
64
65 ff->fm = fm;
66 ff->release_args = kzalloc(sizeof(*ff->release_args),
67 GFP_KERNEL_ACCOUNT);
68 if (!ff->release_args) {
69 kfree(ff);
70 return NULL;
71 }
72
73 INIT_LIST_HEAD(&ff->write_entry);
74 mutex_init(&ff->readdir.lock);
75 refcount_set(&ff->count, 1);
76 RB_CLEAR_NODE(&ff->polled_node);
77 init_waitqueue_head(&ff->poll_wait);
78
79 ff->kh = atomic64_inc_return(&fm->fc->khctr);
80
81 return ff;
82 }
83
84 void fuse_file_free(struct fuse_file *ff)
85 {
86 kfree(ff->release_args);
87 mutex_destroy(&ff->readdir.lock);
88 kfree(ff);
89 }
90
91 static struct fuse_file *fuse_file_get(struct fuse_file *ff)
92 {
93 refcount_inc(&ff->count);
94 return ff;
95 }
96
97 static void fuse_release_end(struct fuse_mount *fm, struct fuse_args *args,
98 int error)
99 {
100 struct fuse_release_args *ra = container_of(args, typeof(*ra), args);
101
102 iput(ra->inode);
103 kfree(ra);
104 }
105
106 static void fuse_file_put(struct fuse_file *ff, bool sync, bool isdir)
107 {
108 if (refcount_dec_and_test(&ff->count)) {
109 struct fuse_args *args = &ff->release_args->args;
110
111 if (isdir ? ff->fm->fc->no_opendir : ff->fm->fc->no_open) {
112 /* Do nothing when client does not implement 'open' */
113 fuse_release_end(ff->fm, args, 0);
114 } else if (sync) {
115 fuse_simple_request(ff->fm, args);
116 fuse_release_end(ff->fm, args, 0);
117 } else {
118 args->end = fuse_release_end;
119 if (fuse_simple_background(ff->fm, args,
120 GFP_KERNEL | __GFP_NOFAIL))
121 fuse_release_end(ff->fm, args, -ENOTCONN);
122 }
123 kfree(ff);
124 }
125 }
126
127 struct fuse_file *fuse_file_open(struct fuse_mount *fm, u64 nodeid,
128 unsigned int open_flags, bool isdir)
129 {
130 struct fuse_conn *fc = fm->fc;
131 struct fuse_file *ff;
132 int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
133
134 ff = fuse_file_alloc(fm);
135 if (!ff)
136 return ERR_PTR(-ENOMEM);
137
138 ff->fh = 0;
139 /* Default for no-open */
140 ff->open_flags = FOPEN_KEEP_CACHE | (isdir ? FOPEN_CACHE_DIR : 0);
141 if (isdir ? !fc->no_opendir : !fc->no_open) {
142 struct fuse_open_out outarg;
143 int err;
144
145 err = fuse_send_open(fm, nodeid, open_flags, opcode, &outarg);
146 if (!err) {
147 ff->fh = outarg.fh;
148 ff->open_flags = outarg.open_flags;
149
150 } else if (err != -ENOSYS) {
151 fuse_file_free(ff);
152 return ERR_PTR(err);
153 } else {
154 if (isdir)
155 fc->no_opendir = 1;
156 else
157 fc->no_open = 1;
158 }
159 }
160
161 if (isdir)
162 ff->open_flags &= ~FOPEN_DIRECT_IO;
163
164 ff->nodeid = nodeid;
165
166 return ff;
167 }
168
169 int fuse_do_open(struct fuse_mount *fm, u64 nodeid, struct file *file,
170 bool isdir)
171 {
172 struct fuse_file *ff = fuse_file_open(fm, nodeid, file->f_flags, isdir);
173
174 if (!IS_ERR(ff))
175 file->private_data = ff;
176
177 return PTR_ERR_OR_ZERO(ff);
178 }
179 EXPORT_SYMBOL_GPL(fuse_do_open);
180
181 static void fuse_link_write_file(struct file *file)
182 {
183 struct inode *inode = file_inode(file);
184 struct fuse_inode *fi = get_fuse_inode(inode);
185 struct fuse_file *ff = file->private_data;
186 /*
187 * file may be written through mmap, so chain it onto the
188 * inodes's write_file list
189 */
190 spin_lock(&fi->lock);
191 if (list_empty(&ff->write_entry))
192 list_add(&ff->write_entry, &fi->write_files);
193 spin_unlock(&fi->lock);
194 }
195
196 void fuse_finish_open(struct inode *inode, struct file *file)
197 {
198 struct fuse_file *ff = file->private_data;
199 struct fuse_conn *fc = get_fuse_conn(inode);
200
201 if (!(ff->open_flags & FOPEN_KEEP_CACHE))
202 invalidate_inode_pages2(inode->i_mapping);
203 if (ff->open_flags & FOPEN_STREAM)
204 stream_open(inode, file);
205 else if (ff->open_flags & FOPEN_NONSEEKABLE)
206 nonseekable_open(inode, file);
207 if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
208 struct fuse_inode *fi = get_fuse_inode(inode);
209
210 spin_lock(&fi->lock);
211 fi->attr_version = atomic64_inc_return(&fc->attr_version);
212 i_size_write(inode, 0);
213 spin_unlock(&fi->lock);
214 fuse_invalidate_attr(inode);
215 if (fc->writeback_cache)
216 file_update_time(file);
217 }
218 if ((file->f_mode & FMODE_WRITE) && fc->writeback_cache)
219 fuse_link_write_file(file);
220 }
221
222 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
223 {
224 struct fuse_mount *fm = get_fuse_mount(inode);
225 struct fuse_conn *fc = fm->fc;
226 int err;
227 bool is_wb_truncate = (file->f_flags & O_TRUNC) &&
228 fc->atomic_o_trunc &&
229 fc->writeback_cache;
230 bool dax_truncate = (file->f_flags & O_TRUNC) &&
231 fc->atomic_o_trunc && FUSE_IS_DAX(inode);
232
233 if (fuse_is_bad(inode))
234 return -EIO;
235
236 err = generic_file_open(inode, file);
237 if (err)
238 return err;
239
240 if (is_wb_truncate || dax_truncate) {
241 inode_lock(inode);
242 fuse_set_nowrite(inode);
243 }
244
245 if (dax_truncate) {
246 filemap_invalidate_lock(inode->i_mapping);
247 err = fuse_dax_break_layouts(inode, 0, 0);
248 if (err)
249 goto out;
250 }
251
252 err = fuse_do_open(fm, get_node_id(inode), file, isdir);
253 if (!err)
254 fuse_finish_open(inode, file);
255
256 out:
257 if (dax_truncate)
258 filemap_invalidate_unlock(inode->i_mapping);
259
260 if (is_wb_truncate | dax_truncate) {
261 fuse_release_nowrite(inode);
262 inode_unlock(inode);
263 }
264
265 return err;
266 }
267
268 static void fuse_prepare_release(struct fuse_inode *fi, struct fuse_file *ff,
269 unsigned int flags, int opcode)
270 {
271 struct fuse_conn *fc = ff->fm->fc;
272 struct fuse_release_args *ra = ff->release_args;
273
274 /* Inode is NULL on error path of fuse_create_open() */
275 if (likely(fi)) {
276 spin_lock(&fi->lock);
277 list_del(&ff->write_entry);
278 spin_unlock(&fi->lock);
279 }
280 spin_lock(&fc->lock);
281 if (!RB_EMPTY_NODE(&ff->polled_node))
282 rb_erase(&ff->polled_node, &fc->polled_files);
283 spin_unlock(&fc->lock);
284
285 wake_up_interruptible_all(&ff->poll_wait);
286
287 ra->inarg.fh = ff->fh;
288 ra->inarg.flags = flags;
289 ra->args.in_numargs = 1;
290 ra->args.in_args[0].size = sizeof(struct fuse_release_in);
291 ra->args.in_args[0].value = &ra->inarg;
292 ra->args.opcode = opcode;
293 ra->args.nodeid = ff->nodeid;
294 ra->args.force = true;
295 ra->args.nocreds = true;
296 }
297
298 void fuse_file_release(struct inode *inode, struct fuse_file *ff,
299 unsigned int open_flags, fl_owner_t id, bool isdir)
300 {
301 struct fuse_inode *fi = get_fuse_inode(inode);
302 struct fuse_release_args *ra = ff->release_args;
303 int opcode = isdir ? FUSE_RELEASEDIR : FUSE_RELEASE;
304
305 fuse_prepare_release(fi, ff, open_flags, opcode);
306
307 if (ff->flock) {
308 ra->inarg.release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
309 ra->inarg.lock_owner = fuse_lock_owner_id(ff->fm->fc, id);
310 }
311 /* Hold inode until release is finished */
312 ra->inode = igrab(inode);
313
314 /*
315 * Normally this will send the RELEASE request, however if
316 * some asynchronous READ or WRITE requests are outstanding,
317 * the sending will be delayed.
318 *
319 * Make the release synchronous if this is a fuseblk mount,
320 * synchronous RELEASE is allowed (and desirable) in this case
321 * because the server can be trusted not to screw up.
322 */
323 fuse_file_put(ff, ff->fm->fc->destroy, isdir);
324 }
325
326 void fuse_release_common(struct file *file, bool isdir)
327 {
328 fuse_file_release(file_inode(file), file->private_data, file->f_flags,
329 (fl_owner_t) file, isdir);
330 }
331
332 static int fuse_open(struct inode *inode, struct file *file)
333 {
334 return fuse_open_common(inode, file, false);
335 }
336
337 static int fuse_release(struct inode *inode, struct file *file)
338 {
339 struct fuse_conn *fc = get_fuse_conn(inode);
340
341 /* see fuse_vma_close() for !writeback_cache case */
342 if (fc->writeback_cache)
343 write_inode_now(inode, 1);
344
345 fuse_release_common(file, false);
346
347 /* return value is ignored by VFS */
348 return 0;
349 }
350
351 void fuse_sync_release(struct fuse_inode *fi, struct fuse_file *ff,
352 unsigned int flags)
353 {
354 WARN_ON(refcount_read(&ff->count) > 1);
355 fuse_prepare_release(fi, ff, flags, FUSE_RELEASE);
356 /*
357 * iput(NULL) is a no-op and since the refcount is 1 and everything's
358 * synchronous, we are fine with not doing igrab() here"
359 */
360 fuse_file_put(ff, true, false);
361 }
362 EXPORT_SYMBOL_GPL(fuse_sync_release);
363
364 /*
365 * Scramble the ID space with XTEA, so that the value of the files_struct
366 * pointer is not exposed to userspace.
367 */
368 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
369 {
370 u32 *k = fc->scramble_key;
371 u64 v = (unsigned long) id;
372 u32 v0 = v;
373 u32 v1 = v >> 32;
374 u32 sum = 0;
375 int i;
376
377 for (i = 0; i < 32; i++) {
378 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
379 sum += 0x9E3779B9;
380 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
381 }
382
383 return (u64) v0 + ((u64) v1 << 32);
384 }
385
386 struct fuse_writepage_args {
387 struct fuse_io_args ia;
388 struct rb_node writepages_entry;
389 struct list_head queue_entry;
390 struct fuse_writepage_args *next;
391 struct inode *inode;
392 };
393
394 static struct fuse_writepage_args *fuse_find_writeback(struct fuse_inode *fi,
395 pgoff_t idx_from, pgoff_t idx_to)
396 {
397 struct rb_node *n;
398
399 n = fi->writepages.rb_node;
400
401 while (n) {
402 struct fuse_writepage_args *wpa;
403 pgoff_t curr_index;
404
405 wpa = rb_entry(n, struct fuse_writepage_args, writepages_entry);
406 WARN_ON(get_fuse_inode(wpa->inode) != fi);
407 curr_index = wpa->ia.write.in.offset >> PAGE_SHIFT;
408 if (idx_from >= curr_index + wpa->ia.ap.num_pages)
409 n = n->rb_right;
410 else if (idx_to < curr_index)
411 n = n->rb_left;
412 else
413 return wpa;
414 }
415 return NULL;
416 }
417
418 /*
419 * Check if any page in a range is under writeback
420 *
421 * This is currently done by walking the list of writepage requests
422 * for the inode, which can be pretty inefficient.
423 */
424 static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
425 pgoff_t idx_to)
426 {
427 struct fuse_inode *fi = get_fuse_inode(inode);
428 bool found;
429
430 spin_lock(&fi->lock);
431 found = fuse_find_writeback(fi, idx_from, idx_to);
432 spin_unlock(&fi->lock);
433
434 return found;
435 }
436
437 static inline bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
438 {
439 return fuse_range_is_writeback(inode, index, index);
440 }
441
442 /*
443 * Wait for page writeback to be completed.
444 *
445 * Since fuse doesn't rely on the VM writeback tracking, this has to
446 * use some other means.
447 */
448 static void fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
449 {
450 struct fuse_inode *fi = get_fuse_inode(inode);
451
452 wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
453 }
454
455 /*
456 * Wait for all pending writepages on the inode to finish.
457 *
458 * This is currently done by blocking further writes with FUSE_NOWRITE
459 * and waiting for all sent writes to complete.
460 *
461 * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
462 * could conflict with truncation.
463 */
464 static void fuse_sync_writes(struct inode *inode)
465 {
466 fuse_set_nowrite(inode);
467 fuse_release_nowrite(inode);
468 }
469
470 static int fuse_flush(struct file *file, fl_owner_t id)
471 {
472 struct inode *inode = file_inode(file);
473 struct fuse_mount *fm = get_fuse_mount(inode);
474 struct fuse_file *ff = file->private_data;
475 struct fuse_flush_in inarg;
476 FUSE_ARGS(args);
477 int err;
478
479 if (fuse_is_bad(inode))
480 return -EIO;
481
482 err = write_inode_now(inode, 1);
483 if (err)
484 return err;
485
486 inode_lock(inode);
487 fuse_sync_writes(inode);
488 inode_unlock(inode);
489
490 err = filemap_check_errors(file->f_mapping);
491 if (err)
492 return err;
493
494 err = 0;
495 if (fm->fc->no_flush)
496 goto inval_attr_out;
497
498 memset(&inarg, 0, sizeof(inarg));
499 inarg.fh = ff->fh;
500 inarg.lock_owner = fuse_lock_owner_id(fm->fc, id);
501 args.opcode = FUSE_FLUSH;
502 args.nodeid = get_node_id(inode);
503 args.in_numargs = 1;
504 args.in_args[0].size = sizeof(inarg);
505 args.in_args[0].value = &inarg;
506 args.force = true;
507
508 err = fuse_simple_request(fm, &args);
509 if (err == -ENOSYS) {
510 fm->fc->no_flush = 1;
511 err = 0;
512 }
513
514 inval_attr_out:
515 /*
516 * In memory i_blocks is not maintained by fuse, if writeback cache is
517 * enabled, i_blocks from cached attr may not be accurate.
518 */
519 if (!err && fm->fc->writeback_cache)
520 fuse_invalidate_attr(inode);
521 return err;
522 }
523
524 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
525 int datasync, int opcode)
526 {
527 struct inode *inode = file->f_mapping->host;
528 struct fuse_mount *fm = get_fuse_mount(inode);
529 struct fuse_file *ff = file->private_data;
530 FUSE_ARGS(args);
531 struct fuse_fsync_in inarg;
532
533 memset(&inarg, 0, sizeof(inarg));
534 inarg.fh = ff->fh;
535 inarg.fsync_flags = datasync ? FUSE_FSYNC_FDATASYNC : 0;
536 args.opcode = opcode;
537 args.nodeid = get_node_id(inode);
538 args.in_numargs = 1;
539 args.in_args[0].size = sizeof(inarg);
540 args.in_args[0].value = &inarg;
541 return fuse_simple_request(fm, &args);
542 }
543
544 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
545 int datasync)
546 {
547 struct inode *inode = file->f_mapping->host;
548 struct fuse_conn *fc = get_fuse_conn(inode);
549 int err;
550
551 if (fuse_is_bad(inode))
552 return -EIO;
553
554 inode_lock(inode);
555
556 /*
557 * Start writeback against all dirty pages of the inode, then
558 * wait for all outstanding writes, before sending the FSYNC
559 * request.
560 */
561 err = file_write_and_wait_range(file, start, end);
562 if (err)
563 goto out;
564
565 fuse_sync_writes(inode);
566
567 /*
568 * Due to implementation of fuse writeback
569 * file_write_and_wait_range() does not catch errors.
570 * We have to do this directly after fuse_sync_writes()
571 */
572 err = file_check_and_advance_wb_err(file);
573 if (err)
574 goto out;
575
576 err = sync_inode_metadata(inode, 1);
577 if (err)
578 goto out;
579
580 if (fc->no_fsync)
581 goto out;
582
583 err = fuse_fsync_common(file, start, end, datasync, FUSE_FSYNC);
584 if (err == -ENOSYS) {
585 fc->no_fsync = 1;
586 err = 0;
587 }
588 out:
589 inode_unlock(inode);
590
591 return err;
592 }
593
594 void fuse_read_args_fill(struct fuse_io_args *ia, struct file *file, loff_t pos,
595 size_t count, int opcode)
596 {
597 struct fuse_file *ff = file->private_data;
598 struct fuse_args *args = &ia->ap.args;
599
600 ia->read.in.fh = ff->fh;
601 ia->read.in.offset = pos;
602 ia->read.in.size = count;
603 ia->read.in.flags = file->f_flags;
604 args->opcode = opcode;
605 args->nodeid = ff->nodeid;
606 args->in_numargs = 1;
607 args->in_args[0].size = sizeof(ia->read.in);
608 args->in_args[0].value = &ia->read.in;
609 args->out_argvar = true;
610 args->out_numargs = 1;
611 args->out_args[0].size = count;
612 }
613
614 static void fuse_release_user_pages(struct fuse_args_pages *ap,
615 bool should_dirty)
616 {
617 unsigned int i;
618
619 for (i = 0; i < ap->num_pages; i++) {
620 if (should_dirty)
621 set_page_dirty_lock(ap->pages[i]);
622 put_page(ap->pages[i]);
623 }
624 }
625
626 static void fuse_io_release(struct kref *kref)
627 {
628 kfree(container_of(kref, struct fuse_io_priv, refcnt));
629 }
630
631 static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
632 {
633 if (io->err)
634 return io->err;
635
636 if (io->bytes >= 0 && io->write)
637 return -EIO;
638
639 return io->bytes < 0 ? io->size : io->bytes;
640 }
641
642 /**
643 * In case of short read, the caller sets 'pos' to the position of
644 * actual end of fuse request in IO request. Otherwise, if bytes_requested
645 * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
646 *
647 * An example:
648 * User requested DIO read of 64K. It was split into two 32K fuse requests,
649 * both submitted asynchronously. The first of them was ACKed by userspace as
650 * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
651 * second request was ACKed as short, e.g. only 1K was read, resulting in
652 * pos == 33K.
653 *
654 * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
655 * will be equal to the length of the longest contiguous fragment of
656 * transferred data starting from the beginning of IO request.
657 */
658 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
659 {
660 int left;
661
662 spin_lock(&io->lock);
663 if (err)
664 io->err = io->err ? : err;
665 else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
666 io->bytes = pos;
667
668 left = --io->reqs;
669 if (!left && io->blocking)
670 complete(io->done);
671 spin_unlock(&io->lock);
672
673 if (!left && !io->blocking) {
674 ssize_t res = fuse_get_res_by_io(io);
675
676 if (res >= 0) {
677 struct inode *inode = file_inode(io->iocb->ki_filp);
678 struct fuse_conn *fc = get_fuse_conn(inode);
679 struct fuse_inode *fi = get_fuse_inode(inode);
680
681 spin_lock(&fi->lock);
682 fi->attr_version = atomic64_inc_return(&fc->attr_version);
683 spin_unlock(&fi->lock);
684 }
685
686 io->iocb->ki_complete(io->iocb, res, 0);
687 }
688
689 kref_put(&io->refcnt, fuse_io_release);
690 }
691
692 static struct fuse_io_args *fuse_io_alloc(struct fuse_io_priv *io,
693 unsigned int npages)
694 {
695 struct fuse_io_args *ia;
696
697 ia = kzalloc(sizeof(*ia), GFP_KERNEL);
698 if (ia) {
699 ia->io = io;
700 ia->ap.pages = fuse_pages_alloc(npages, GFP_KERNEL,
701 &ia->ap.descs);
702 if (!ia->ap.pages) {
703 kfree(ia);
704 ia = NULL;
705 }
706 }
707 return ia;
708 }
709
710 static void fuse_io_free(struct fuse_io_args *ia)
711 {
712 kfree(ia->ap.pages);
713 kfree(ia);
714 }
715
716 static void fuse_aio_complete_req(struct fuse_mount *fm, struct fuse_args *args,
717 int err)
718 {
719 struct fuse_io_args *ia = container_of(args, typeof(*ia), ap.args);
720 struct fuse_io_priv *io = ia->io;
721 ssize_t pos = -1;
722
723 fuse_release_user_pages(&ia->ap, io->should_dirty);
724
725 if (err) {
726 /* Nothing */
727 } else if (io->write) {
728 if (ia->write.out.size > ia->write.in.size) {
729 err = -EIO;
730 } else if (ia->write.in.size != ia->write.out.size) {
731 pos = ia->write.in.offset - io->offset +
732 ia->write.out.size;
733 }
734 } else {
735 u32 outsize = args->out_args[0].size;
736
737 if (ia->read.in.size != outsize)
738 pos = ia->read.in.offset - io->offset + outsize;
739 }
740
741 fuse_aio_complete(io, err, pos);
742 fuse_io_free(ia);
743 }
744
745 static ssize_t fuse_async_req_send(struct fuse_mount *fm,
746 struct fuse_io_args *ia, size_t num_bytes)
747 {
748 ssize_t err;
749 struct fuse_io_priv *io = ia->io;
750
751 spin_lock(&io->lock);
752 kref_get(&io->refcnt);
753 io->size += num_bytes;
754 io->reqs++;
755 spin_unlock(&io->lock);
756
757 ia->ap.args.end = fuse_aio_complete_req;
758 ia->ap.args.may_block = io->should_dirty;
759 err = fuse_simple_background(fm, &ia->ap.args, GFP_KERNEL);
760 if (err)
761 fuse_aio_complete_req(fm, &ia->ap.args, err);
762
763 return num_bytes;
764 }
765
766 static ssize_t fuse_send_read(struct fuse_io_args *ia, loff_t pos, size_t count,
767 fl_owner_t owner)
768 {
769 struct file *file = ia->io->iocb->ki_filp;
770 struct fuse_file *ff = file->private_data;
771 struct fuse_mount *fm = ff->fm;
772
773 fuse_read_args_fill(ia, file, pos, count, FUSE_READ);
774 if (owner != NULL) {
775 ia->read.in.read_flags |= FUSE_READ_LOCKOWNER;
776 ia->read.in.lock_owner = fuse_lock_owner_id(fm->fc, owner);
777 }
778
779 if (ia->io->async)
780 return fuse_async_req_send(fm, ia, count);
781
782 return fuse_simple_request(fm, &ia->ap.args);
783 }
784
785 static void fuse_read_update_size(struct inode *inode, loff_t size,
786 u64 attr_ver)
787 {
788 struct fuse_conn *fc = get_fuse_conn(inode);
789 struct fuse_inode *fi = get_fuse_inode(inode);
790
791 spin_lock(&fi->lock);
792 if (attr_ver == fi->attr_version && size < inode->i_size &&
793 !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
794 fi->attr_version = atomic64_inc_return(&fc->attr_version);
795 i_size_write(inode, size);
796 }
797 spin_unlock(&fi->lock);
798 }
799
800 static void fuse_short_read(struct inode *inode, u64 attr_ver, size_t num_read,
801 struct fuse_args_pages *ap)
802 {
803 struct fuse_conn *fc = get_fuse_conn(inode);
804
805 /*
806 * If writeback_cache is enabled, a short read means there's a hole in
807 * the file. Some data after the hole is in page cache, but has not
808 * reached the client fs yet. So the hole is not present there.
809 */
810 if (!fc->writeback_cache) {
811 loff_t pos = page_offset(ap->pages[0]) + num_read;
812 fuse_read_update_size(inode, pos, attr_ver);
813 }
814 }
815
816 static int fuse_do_readpage(struct file *file, struct page *page)
817 {
818 struct inode *inode = page->mapping->host;
819 struct fuse_mount *fm = get_fuse_mount(inode);
820 loff_t pos = page_offset(page);
821 struct fuse_page_desc desc = { .length = PAGE_SIZE };
822 struct fuse_io_args ia = {
823 .ap.args.page_zeroing = true,
824 .ap.args.out_pages = true,
825 .ap.num_pages = 1,
826 .ap.pages = &page,
827 .ap.descs = &desc,
828 };
829 ssize_t res;
830 u64 attr_ver;
831
832 /*
833 * Page writeback can extend beyond the lifetime of the
834 * page-cache page, so make sure we read a properly synced
835 * page.
836 */
837 fuse_wait_on_page_writeback(inode, page->index);
838
839 attr_ver = fuse_get_attr_version(fm->fc);
840
841 /* Don't overflow end offset */
842 if (pos + (desc.length - 1) == LLONG_MAX)
843 desc.length--;
844
845 fuse_read_args_fill(&ia, file, pos, desc.length, FUSE_READ);
846 res = fuse_simple_request(fm, &ia.ap.args);
847 if (res < 0)
848 return res;
849 /*
850 * Short read means EOF. If file size is larger, truncate it
851 */
852 if (res < desc.length)
853 fuse_short_read(inode, attr_ver, res, &ia.ap);
854
855 SetPageUptodate(page);
856
857 return 0;
858 }
859
860 static int fuse_readpage(struct file *file, struct page *page)
861 {
862 struct inode *inode = page->mapping->host;
863 int err;
864
865 err = -EIO;
866 if (fuse_is_bad(inode))
867 goto out;
868
869 err = fuse_do_readpage(file, page);
870 fuse_invalidate_atime(inode);
871 out:
872 unlock_page(page);
873 return err;
874 }
875
876 static void fuse_readpages_end(struct fuse_mount *fm, struct fuse_args *args,
877 int err)
878 {
879 int i;
880 struct fuse_io_args *ia = container_of(args, typeof(*ia), ap.args);
881 struct fuse_args_pages *ap = &ia->ap;
882 size_t count = ia->read.in.size;
883 size_t num_read = args->out_args[0].size;
884 struct address_space *mapping = NULL;
885
886 for (i = 0; mapping == NULL && i < ap->num_pages; i++)
887 mapping = ap->pages[i]->mapping;
888
889 if (mapping) {
890 struct inode *inode = mapping->host;
891
892 /*
893 * Short read means EOF. If file size is larger, truncate it
894 */
895 if (!err && num_read < count)
896 fuse_short_read(inode, ia->read.attr_ver, num_read, ap);
897
898 fuse_invalidate_atime(inode);
899 }
900
901 for (i = 0; i < ap->num_pages; i++) {
902 struct page *page = ap->pages[i];
903
904 if (!err)
905 SetPageUptodate(page);
906 else
907 SetPageError(page);
908 unlock_page(page);
909 put_page(page);
910 }
911 if (ia->ff)
912 fuse_file_put(ia->ff, false, false);
913
914 fuse_io_free(ia);
915 }
916
917 static void fuse_send_readpages(struct fuse_io_args *ia, struct file *file)
918 {
919 struct fuse_file *ff = file->private_data;
920 struct fuse_mount *fm = ff->fm;
921 struct fuse_args_pages *ap = &ia->ap;
922 loff_t pos = page_offset(ap->pages[0]);
923 size_t count = ap->num_pages << PAGE_SHIFT;
924 ssize_t res;
925 int err;
926
927 ap->args.out_pages = true;
928 ap->args.page_zeroing = true;
929 ap->args.page_replace = true;
930
931 /* Don't overflow end offset */
932 if (pos + (count - 1) == LLONG_MAX) {
933 count--;
934 ap->descs[ap->num_pages - 1].length--;
935 }
936 WARN_ON((loff_t) (pos + count) < 0);
937
938 fuse_read_args_fill(ia, file, pos, count, FUSE_READ);
939 ia->read.attr_ver = fuse_get_attr_version(fm->fc);
940 if (fm->fc->async_read) {
941 ia->ff = fuse_file_get(ff);
942 ap->args.end = fuse_readpages_end;
943 err = fuse_simple_background(fm, &ap->args, GFP_KERNEL);
944 if (!err)
945 return;
946 } else {
947 res = fuse_simple_request(fm, &ap->args);
948 err = res < 0 ? res : 0;
949 }
950 fuse_readpages_end(fm, &ap->args, err);
951 }
952
953 static void fuse_readahead(struct readahead_control *rac)
954 {
955 struct inode *inode = rac->mapping->host;
956 struct fuse_conn *fc = get_fuse_conn(inode);
957 unsigned int i, max_pages, nr_pages = 0;
958
959 if (fuse_is_bad(inode))
960 return;
961
962 max_pages = min_t(unsigned int, fc->max_pages,
963 fc->max_read / PAGE_SIZE);
964
965 for (;;) {
966 struct fuse_io_args *ia;
967 struct fuse_args_pages *ap;
968
969 nr_pages = readahead_count(rac) - nr_pages;
970 if (nr_pages > max_pages)
971 nr_pages = max_pages;
972 if (nr_pages == 0)
973 break;
974 ia = fuse_io_alloc(NULL, nr_pages);
975 if (!ia)
976 return;
977 ap = &ia->ap;
978 nr_pages = __readahead_batch(rac, ap->pages, nr_pages);
979 for (i = 0; i < nr_pages; i++) {
980 fuse_wait_on_page_writeback(inode,
981 readahead_index(rac) + i);
982 ap->descs[i].length = PAGE_SIZE;
983 }
984 ap->num_pages = nr_pages;
985 fuse_send_readpages(ia, rac->file);
986 }
987 }
988
989 static ssize_t fuse_cache_read_iter(struct kiocb *iocb, struct iov_iter *to)
990 {
991 struct inode *inode = iocb->ki_filp->f_mapping->host;
992 struct fuse_conn *fc = get_fuse_conn(inode);
993
994 /*
995 * In auto invalidate mode, always update attributes on read.
996 * Otherwise, only update if we attempt to read past EOF (to ensure
997 * i_size is up to date).
998 */
999 if (fc->auto_inval_data ||
1000 (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
1001 int err;
1002 err = fuse_update_attributes(inode, iocb->ki_filp);
1003 if (err)
1004 return err;
1005 }
1006
1007 return generic_file_read_iter(iocb, to);
1008 }
1009
1010 static void fuse_write_args_fill(struct fuse_io_args *ia, struct fuse_file *ff,
1011 loff_t pos, size_t count)
1012 {
1013 struct fuse_args *args = &ia->ap.args;
1014
1015 ia->write.in.fh = ff->fh;
1016 ia->write.in.offset = pos;
1017 ia->write.in.size = count;
1018 args->opcode = FUSE_WRITE;
1019 args->nodeid = ff->nodeid;
1020 args->in_numargs = 2;
1021 if (ff->fm->fc->minor < 9)
1022 args->in_args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
1023 else
1024 args->in_args[0].size = sizeof(ia->write.in);
1025 args->in_args[0].value = &ia->write.in;
1026 args->in_args[1].size = count;
1027 args->out_numargs = 1;
1028 args->out_args[0].size = sizeof(ia->write.out);
1029 args->out_args[0].value = &ia->write.out;
1030 }
1031
1032 static unsigned int fuse_write_flags(struct kiocb *iocb)
1033 {
1034 unsigned int flags = iocb->ki_filp->f_flags;
1035
1036 if (iocb->ki_flags & IOCB_DSYNC)
1037 flags |= O_DSYNC;
1038 if (iocb->ki_flags & IOCB_SYNC)
1039 flags |= O_SYNC;
1040
1041 return flags;
1042 }
1043
1044 static ssize_t fuse_send_write(struct fuse_io_args *ia, loff_t pos,
1045 size_t count, fl_owner_t owner)
1046 {
1047 struct kiocb *iocb = ia->io->iocb;
1048 struct file *file = iocb->ki_filp;
1049 struct fuse_file *ff = file->private_data;
1050 struct fuse_mount *fm = ff->fm;
1051 struct fuse_write_in *inarg = &ia->write.in;
1052 ssize_t err;
1053
1054 fuse_write_args_fill(ia, ff, pos, count);
1055 inarg->flags = fuse_write_flags(iocb);
1056 if (owner != NULL) {
1057 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
1058 inarg->lock_owner = fuse_lock_owner_id(fm->fc, owner);
1059 }
1060
1061 if (ia->io->async)
1062 return fuse_async_req_send(fm, ia, count);
1063
1064 err = fuse_simple_request(fm, &ia->ap.args);
1065 if (!err && ia->write.out.size > count)
1066 err = -EIO;
1067
1068 return err ?: ia->write.out.size;
1069 }
1070
1071 bool fuse_write_update_size(struct inode *inode, loff_t pos)
1072 {
1073 struct fuse_conn *fc = get_fuse_conn(inode);
1074 struct fuse_inode *fi = get_fuse_inode(inode);
1075 bool ret = false;
1076
1077 spin_lock(&fi->lock);
1078 fi->attr_version = atomic64_inc_return(&fc->attr_version);
1079 if (pos > inode->i_size) {
1080 i_size_write(inode, pos);
1081 ret = true;
1082 }
1083 spin_unlock(&fi->lock);
1084
1085 return ret;
1086 }
1087
1088 static ssize_t fuse_send_write_pages(struct fuse_io_args *ia,
1089 struct kiocb *iocb, struct inode *inode,
1090 loff_t pos, size_t count)
1091 {
1092 struct fuse_args_pages *ap = &ia->ap;
1093 struct file *file = iocb->ki_filp;
1094 struct fuse_file *ff = file->private_data;
1095 struct fuse_mount *fm = ff->fm;
1096 unsigned int offset, i;
1097 bool short_write;
1098 int err;
1099
1100 for (i = 0; i < ap->num_pages; i++)
1101 fuse_wait_on_page_writeback(inode, ap->pages[i]->index);
1102
1103 fuse_write_args_fill(ia, ff, pos, count);
1104 ia->write.in.flags = fuse_write_flags(iocb);
1105 if (fm->fc->handle_killpriv_v2 && !capable(CAP_FSETID))
1106 ia->write.in.write_flags |= FUSE_WRITE_KILL_SUIDGID;
1107
1108 err = fuse_simple_request(fm, &ap->args);
1109 if (!err && ia->write.out.size > count)
1110 err = -EIO;
1111
1112 short_write = ia->write.out.size < count;
1113 offset = ap->descs[0].offset;
1114 count = ia->write.out.size;
1115 for (i = 0; i < ap->num_pages; i++) {
1116 struct page *page = ap->pages[i];
1117
1118 if (err) {
1119 ClearPageUptodate(page);
1120 } else {
1121 if (count >= PAGE_SIZE - offset)
1122 count -= PAGE_SIZE - offset;
1123 else {
1124 if (short_write)
1125 ClearPageUptodate(page);
1126 count = 0;
1127 }
1128 offset = 0;
1129 }
1130 if (ia->write.page_locked && (i == ap->num_pages - 1))
1131 unlock_page(page);
1132 put_page(page);
1133 }
1134
1135 return err;
1136 }
1137
1138 static ssize_t fuse_fill_write_pages(struct fuse_io_args *ia,
1139 struct address_space *mapping,
1140 struct iov_iter *ii, loff_t pos,
1141 unsigned int max_pages)
1142 {
1143 struct fuse_args_pages *ap = &ia->ap;
1144 struct fuse_conn *fc = get_fuse_conn(mapping->host);
1145 unsigned offset = pos & (PAGE_SIZE - 1);
1146 size_t count = 0;
1147 int err;
1148
1149 ap->args.in_pages = true;
1150 ap->descs[0].offset = offset;
1151
1152 do {
1153 size_t tmp;
1154 struct page *page;
1155 pgoff_t index = pos >> PAGE_SHIFT;
1156 size_t bytes = min_t(size_t, PAGE_SIZE - offset,
1157 iov_iter_count(ii));
1158
1159 bytes = min_t(size_t, bytes, fc->max_write - count);
1160
1161 again:
1162 err = -EFAULT;
1163 if (iov_iter_fault_in_readable(ii, bytes))
1164 break;
1165
1166 err = -ENOMEM;
1167 page = grab_cache_page_write_begin(mapping, index, 0);
1168 if (!page)
1169 break;
1170
1171 if (mapping_writably_mapped(mapping))
1172 flush_dcache_page(page);
1173
1174 tmp = copy_page_from_iter_atomic(page, offset, bytes, ii);
1175 flush_dcache_page(page);
1176
1177 if (!tmp) {
1178 unlock_page(page);
1179 put_page(page);
1180 goto again;
1181 }
1182
1183 err = 0;
1184 ap->pages[ap->num_pages] = page;
1185 ap->descs[ap->num_pages].length = tmp;
1186 ap->num_pages++;
1187
1188 count += tmp;
1189 pos += tmp;
1190 offset += tmp;
1191 if (offset == PAGE_SIZE)
1192 offset = 0;
1193
1194 /* If we copied full page, mark it uptodate */
1195 if (tmp == PAGE_SIZE)
1196 SetPageUptodate(page);
1197
1198 if (PageUptodate(page)) {
1199 unlock_page(page);
1200 } else {
1201 ia->write.page_locked = true;
1202 break;
1203 }
1204 if (!fc->big_writes)
1205 break;
1206 } while (iov_iter_count(ii) && count < fc->max_write &&
1207 ap->num_pages < max_pages && offset == 0);
1208
1209 return count > 0 ? count : err;
1210 }
1211
1212 static inline unsigned int fuse_wr_pages(loff_t pos, size_t len,
1213 unsigned int max_pages)
1214 {
1215 return min_t(unsigned int,
1216 ((pos + len - 1) >> PAGE_SHIFT) -
1217 (pos >> PAGE_SHIFT) + 1,
1218 max_pages);
1219 }
1220
1221 static ssize_t fuse_perform_write(struct kiocb *iocb,
1222 struct address_space *mapping,
1223 struct iov_iter *ii, loff_t pos)
1224 {
1225 struct inode *inode = mapping->host;
1226 struct fuse_conn *fc = get_fuse_conn(inode);
1227 struct fuse_inode *fi = get_fuse_inode(inode);
1228 int err = 0;
1229 ssize_t res = 0;
1230
1231 if (inode->i_size < pos + iov_iter_count(ii))
1232 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1233
1234 do {
1235 ssize_t count;
1236 struct fuse_io_args ia = {};
1237 struct fuse_args_pages *ap = &ia.ap;
1238 unsigned int nr_pages = fuse_wr_pages(pos, iov_iter_count(ii),
1239 fc->max_pages);
1240
1241 ap->pages = fuse_pages_alloc(nr_pages, GFP_KERNEL, &ap->descs);
1242 if (!ap->pages) {
1243 err = -ENOMEM;
1244 break;
1245 }
1246
1247 count = fuse_fill_write_pages(&ia, mapping, ii, pos, nr_pages);
1248 if (count <= 0) {
1249 err = count;
1250 } else {
1251 err = fuse_send_write_pages(&ia, iocb, inode,
1252 pos, count);
1253 if (!err) {
1254 size_t num_written = ia.write.out.size;
1255
1256 res += num_written;
1257 pos += num_written;
1258
1259 /* break out of the loop on short write */
1260 if (num_written != count)
1261 err = -EIO;
1262 }
1263 }
1264 kfree(ap->pages);
1265 } while (!err && iov_iter_count(ii));
1266
1267 if (res > 0)
1268 fuse_write_update_size(inode, pos);
1269
1270 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1271 fuse_invalidate_attr(inode);
1272
1273 return res > 0 ? res : err;
1274 }
1275
1276 static ssize_t fuse_cache_write_iter(struct kiocb *iocb, struct iov_iter *from)
1277 {
1278 struct file *file = iocb->ki_filp;
1279 struct address_space *mapping = file->f_mapping;
1280 ssize_t written = 0;
1281 ssize_t written_buffered = 0;
1282 struct inode *inode = mapping->host;
1283 ssize_t err;
1284 struct fuse_conn *fc = get_fuse_conn(inode);
1285 loff_t endbyte = 0;
1286
1287 if (fc->writeback_cache) {
1288 /* Update size (EOF optimization) and mode (SUID clearing) */
1289 err = fuse_update_attributes(mapping->host, file);
1290 if (err)
1291 return err;
1292
1293 if (fc->handle_killpriv_v2 &&
1294 should_remove_suid(file_dentry(file))) {
1295 goto writethrough;
1296 }
1297
1298 return generic_file_write_iter(iocb, from);
1299 }
1300
1301 writethrough:
1302 inode_lock(inode);
1303
1304 /* We can write back this queue in page reclaim */
1305 current->backing_dev_info = inode_to_bdi(inode);
1306
1307 err = generic_write_checks(iocb, from);
1308 if (err <= 0)
1309 goto out;
1310
1311 err = file_remove_privs(file);
1312 if (err)
1313 goto out;
1314
1315 err = file_update_time(file);
1316 if (err)
1317 goto out;
1318
1319 if (iocb->ki_flags & IOCB_DIRECT) {
1320 loff_t pos = iocb->ki_pos;
1321 written = generic_file_direct_write(iocb, from);
1322 if (written < 0 || !iov_iter_count(from))
1323 goto out;
1324
1325 pos += written;
1326
1327 written_buffered = fuse_perform_write(iocb, mapping, from, pos);
1328 if (written_buffered < 0) {
1329 err = written_buffered;
1330 goto out;
1331 }
1332 endbyte = pos + written_buffered - 1;
1333
1334 err = filemap_write_and_wait_range(file->f_mapping, pos,
1335 endbyte);
1336 if (err)
1337 goto out;
1338
1339 invalidate_mapping_pages(file->f_mapping,
1340 pos >> PAGE_SHIFT,
1341 endbyte >> PAGE_SHIFT);
1342
1343 written += written_buffered;
1344 iocb->ki_pos = pos + written_buffered;
1345 } else {
1346 written = fuse_perform_write(iocb, mapping, from, iocb->ki_pos);
1347 if (written >= 0)
1348 iocb->ki_pos += written;
1349 }
1350 out:
1351 current->backing_dev_info = NULL;
1352 inode_unlock(inode);
1353 if (written > 0)
1354 written = generic_write_sync(iocb, written);
1355
1356 return written ? written : err;
1357 }
1358
1359 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1360 {
1361 return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1362 }
1363
1364 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1365 size_t max_size)
1366 {
1367 return min(iov_iter_single_seg_count(ii), max_size);
1368 }
1369
1370 static int fuse_get_user_pages(struct fuse_args_pages *ap, struct iov_iter *ii,
1371 size_t *nbytesp, int write,
1372 unsigned int max_pages)
1373 {
1374 size_t nbytes = 0; /* # bytes already packed in req */
1375 ssize_t ret = 0;
1376
1377 /* Special case for kernel I/O: can copy directly into the buffer */
1378 if (iov_iter_is_kvec(ii)) {
1379 unsigned long user_addr = fuse_get_user_addr(ii);
1380 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1381
1382 if (write)
1383 ap->args.in_args[1].value = (void *) user_addr;
1384 else
1385 ap->args.out_args[0].value = (void *) user_addr;
1386
1387 iov_iter_advance(ii, frag_size);
1388 *nbytesp = frag_size;
1389 return 0;
1390 }
1391
1392 while (nbytes < *nbytesp && ap->num_pages < max_pages) {
1393 unsigned npages;
1394 size_t start;
1395 ret = iov_iter_get_pages(ii, &ap->pages[ap->num_pages],
1396 *nbytesp - nbytes,
1397 max_pages - ap->num_pages,
1398 &start);
1399 if (ret < 0)
1400 break;
1401
1402 iov_iter_advance(ii, ret);
1403 nbytes += ret;
1404
1405 ret += start;
1406 npages = DIV_ROUND_UP(ret, PAGE_SIZE);
1407
1408 ap->descs[ap->num_pages].offset = start;
1409 fuse_page_descs_length_init(ap->descs, ap->num_pages, npages);
1410
1411 ap->num_pages += npages;
1412 ap->descs[ap->num_pages - 1].length -=
1413 (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1414 }
1415
1416 if (write)
1417 ap->args.in_pages = true;
1418 else
1419 ap->args.out_pages = true;
1420
1421 *nbytesp = nbytes;
1422
1423 return ret < 0 ? ret : 0;
1424 }
1425
1426 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1427 loff_t *ppos, int flags)
1428 {
1429 int write = flags & FUSE_DIO_WRITE;
1430 int cuse = flags & FUSE_DIO_CUSE;
1431 struct file *file = io->iocb->ki_filp;
1432 struct inode *inode = file->f_mapping->host;
1433 struct fuse_file *ff = file->private_data;
1434 struct fuse_conn *fc = ff->fm->fc;
1435 size_t nmax = write ? fc->max_write : fc->max_read;
1436 loff_t pos = *ppos;
1437 size_t count = iov_iter_count(iter);
1438 pgoff_t idx_from = pos >> PAGE_SHIFT;
1439 pgoff_t idx_to = (pos + count - 1) >> PAGE_SHIFT;
1440 ssize_t res = 0;
1441 int err = 0;
1442 struct fuse_io_args *ia;
1443 unsigned int max_pages;
1444
1445 max_pages = iov_iter_npages(iter, fc->max_pages);
1446 ia = fuse_io_alloc(io, max_pages);
1447 if (!ia)
1448 return -ENOMEM;
1449
1450 ia->io = io;
1451 if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1452 if (!write)
1453 inode_lock(inode);
1454 fuse_sync_writes(inode);
1455 if (!write)
1456 inode_unlock(inode);
1457 }
1458
1459 io->should_dirty = !write && iter_is_iovec(iter);
1460 while (count) {
1461 ssize_t nres;
1462 fl_owner_t owner = current->files;
1463 size_t nbytes = min(count, nmax);
1464
1465 err = fuse_get_user_pages(&ia->ap, iter, &nbytes, write,
1466 max_pages);
1467 if (err && !nbytes)
1468 break;
1469
1470 if (write) {
1471 if (!capable(CAP_FSETID))
1472 ia->write.in.write_flags |= FUSE_WRITE_KILL_SUIDGID;
1473
1474 nres = fuse_send_write(ia, pos, nbytes, owner);
1475 } else {
1476 nres = fuse_send_read(ia, pos, nbytes, owner);
1477 }
1478
1479 if (!io->async || nres < 0) {
1480 fuse_release_user_pages(&ia->ap, io->should_dirty);
1481 fuse_io_free(ia);
1482 }
1483 ia = NULL;
1484 if (nres < 0) {
1485 iov_iter_revert(iter, nbytes);
1486 err = nres;
1487 break;
1488 }
1489 WARN_ON(nres > nbytes);
1490
1491 count -= nres;
1492 res += nres;
1493 pos += nres;
1494 if (nres != nbytes) {
1495 iov_iter_revert(iter, nbytes - nres);
1496 break;
1497 }
1498 if (count) {
1499 max_pages = iov_iter_npages(iter, fc->max_pages);
1500 ia = fuse_io_alloc(io, max_pages);
1501 if (!ia)
1502 break;
1503 }
1504 }
1505 if (ia)
1506 fuse_io_free(ia);
1507 if (res > 0)
1508 *ppos = pos;
1509
1510 return res > 0 ? res : err;
1511 }
1512 EXPORT_SYMBOL_GPL(fuse_direct_io);
1513
1514 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1515 struct iov_iter *iter,
1516 loff_t *ppos)
1517 {
1518 ssize_t res;
1519 struct inode *inode = file_inode(io->iocb->ki_filp);
1520
1521 res = fuse_direct_io(io, iter, ppos, 0);
1522
1523 fuse_invalidate_atime(inode);
1524
1525 return res;
1526 }
1527
1528 static ssize_t fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter);
1529
1530 static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1531 {
1532 ssize_t res;
1533
1534 if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1535 res = fuse_direct_IO(iocb, to);
1536 } else {
1537 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1538
1539 res = __fuse_direct_read(&io, to, &iocb->ki_pos);
1540 }
1541
1542 return res;
1543 }
1544
1545 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1546 {
1547 struct inode *inode = file_inode(iocb->ki_filp);
1548 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1549 ssize_t res;
1550
1551 /* Don't allow parallel writes to the same file */
1552 inode_lock(inode);
1553 res = generic_write_checks(iocb, from);
1554 if (res > 0) {
1555 if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1556 res = fuse_direct_IO(iocb, from);
1557 } else {
1558 res = fuse_direct_io(&io, from, &iocb->ki_pos,
1559 FUSE_DIO_WRITE);
1560 }
1561 }
1562 fuse_invalidate_attr(inode);
1563 if (res > 0)
1564 fuse_write_update_size(inode, iocb->ki_pos);
1565 inode_unlock(inode);
1566
1567 return res;
1568 }
1569
1570 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
1571 {
1572 struct file *file = iocb->ki_filp;
1573 struct fuse_file *ff = file->private_data;
1574 struct inode *inode = file_inode(file);
1575
1576 if (fuse_is_bad(inode))
1577 return -EIO;
1578
1579 if (FUSE_IS_DAX(inode))
1580 return fuse_dax_read_iter(iocb, to);
1581
1582 if (!(ff->open_flags & FOPEN_DIRECT_IO))
1583 return fuse_cache_read_iter(iocb, to);
1584 else
1585 return fuse_direct_read_iter(iocb, to);
1586 }
1587
1588 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1589 {
1590 struct file *file = iocb->ki_filp;
1591 struct fuse_file *ff = file->private_data;
1592 struct inode *inode = file_inode(file);
1593
1594 if (fuse_is_bad(inode))
1595 return -EIO;
1596
1597 if (FUSE_IS_DAX(inode))
1598 return fuse_dax_write_iter(iocb, from);
1599
1600 if (!(ff->open_flags & FOPEN_DIRECT_IO))
1601 return fuse_cache_write_iter(iocb, from);
1602 else
1603 return fuse_direct_write_iter(iocb, from);
1604 }
1605
1606 static void fuse_writepage_free(struct fuse_writepage_args *wpa)
1607 {
1608 struct fuse_args_pages *ap = &wpa->ia.ap;
1609 int i;
1610
1611 for (i = 0; i < ap->num_pages; i++)
1612 __free_page(ap->pages[i]);
1613
1614 if (wpa->ia.ff)
1615 fuse_file_put(wpa->ia.ff, false, false);
1616
1617 kfree(ap->pages);
1618 kfree(wpa);
1619 }
1620
1621 static void fuse_writepage_finish(struct fuse_mount *fm,
1622 struct fuse_writepage_args *wpa)
1623 {
1624 struct fuse_args_pages *ap = &wpa->ia.ap;
1625 struct inode *inode = wpa->inode;
1626 struct fuse_inode *fi = get_fuse_inode(inode);
1627 struct backing_dev_info *bdi = inode_to_bdi(inode);
1628 int i;
1629
1630 for (i = 0; i < ap->num_pages; i++) {
1631 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1632 dec_node_page_state(ap->pages[i], NR_WRITEBACK_TEMP);
1633 wb_writeout_inc(&bdi->wb);
1634 }
1635 wake_up(&fi->page_waitq);
1636 }
1637
1638 /* Called under fi->lock, may release and reacquire it */
1639 static void fuse_send_writepage(struct fuse_mount *fm,
1640 struct fuse_writepage_args *wpa, loff_t size)
1641 __releases(fi->lock)
1642 __acquires(fi->lock)
1643 {
1644 struct fuse_writepage_args *aux, *next;
1645 struct fuse_inode *fi = get_fuse_inode(wpa->inode);
1646 struct fuse_write_in *inarg = &wpa->ia.write.in;
1647 struct fuse_args *args = &wpa->ia.ap.args;
1648 __u64 data_size = wpa->ia.ap.num_pages * PAGE_SIZE;
1649 int err;
1650
1651 fi->writectr++;
1652 if (inarg->offset + data_size <= size) {
1653 inarg->size = data_size;
1654 } else if (inarg->offset < size) {
1655 inarg->size = size - inarg->offset;
1656 } else {
1657 /* Got truncated off completely */
1658 goto out_free;
1659 }
1660
1661 args->in_args[1].size = inarg->size;
1662 args->force = true;
1663 args->nocreds = true;
1664
1665 err = fuse_simple_background(fm, args, GFP_ATOMIC);
1666 if (err == -ENOMEM) {
1667 spin_unlock(&fi->lock);
1668 err = fuse_simple_background(fm, args, GFP_NOFS | __GFP_NOFAIL);
1669 spin_lock(&fi->lock);
1670 }
1671
1672 /* Fails on broken connection only */
1673 if (unlikely(err))
1674 goto out_free;
1675
1676 return;
1677
1678 out_free:
1679 fi->writectr--;
1680 rb_erase(&wpa->writepages_entry, &fi->writepages);
1681 fuse_writepage_finish(fm, wpa);
1682 spin_unlock(&fi->lock);
1683
1684 /* After fuse_writepage_finish() aux request list is private */
1685 for (aux = wpa->next; aux; aux = next) {
1686 next = aux->next;
1687 aux->next = NULL;
1688 fuse_writepage_free(aux);
1689 }
1690
1691 fuse_writepage_free(wpa);
1692 spin_lock(&fi->lock);
1693 }
1694
1695 /*
1696 * If fi->writectr is positive (no truncate or fsync going on) send
1697 * all queued writepage requests.
1698 *
1699 * Called with fi->lock
1700 */
1701 void fuse_flush_writepages(struct inode *inode)
1702 __releases(fi->lock)
1703 __acquires(fi->lock)
1704 {
1705 struct fuse_mount *fm = get_fuse_mount(inode);
1706 struct fuse_inode *fi = get_fuse_inode(inode);
1707 loff_t crop = i_size_read(inode);
1708 struct fuse_writepage_args *wpa;
1709
1710 while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1711 wpa = list_entry(fi->queued_writes.next,
1712 struct fuse_writepage_args, queue_entry);
1713 list_del_init(&wpa->queue_entry);
1714 fuse_send_writepage(fm, wpa, crop);
1715 }
1716 }
1717
1718 static struct fuse_writepage_args *fuse_insert_writeback(struct rb_root *root,
1719 struct fuse_writepage_args *wpa)
1720 {
1721 pgoff_t idx_from = wpa->ia.write.in.offset >> PAGE_SHIFT;
1722 pgoff_t idx_to = idx_from + wpa->ia.ap.num_pages - 1;
1723 struct rb_node **p = &root->rb_node;
1724 struct rb_node *parent = NULL;
1725
1726 WARN_ON(!wpa->ia.ap.num_pages);
1727 while (*p) {
1728 struct fuse_writepage_args *curr;
1729 pgoff_t curr_index;
1730
1731 parent = *p;
1732 curr = rb_entry(parent, struct fuse_writepage_args,
1733 writepages_entry);
1734 WARN_ON(curr->inode != wpa->inode);
1735 curr_index = curr->ia.write.in.offset >> PAGE_SHIFT;
1736
1737 if (idx_from >= curr_index + curr->ia.ap.num_pages)
1738 p = &(*p)->rb_right;
1739 else if (idx_to < curr_index)
1740 p = &(*p)->rb_left;
1741 else
1742 return curr;
1743 }
1744
1745 rb_link_node(&wpa->writepages_entry, parent, p);
1746 rb_insert_color(&wpa->writepages_entry, root);
1747 return NULL;
1748 }
1749
1750 static void tree_insert(struct rb_root *root, struct fuse_writepage_args *wpa)
1751 {
1752 WARN_ON(fuse_insert_writeback(root, wpa));
1753 }
1754
1755 static void fuse_writepage_end(struct fuse_mount *fm, struct fuse_args *args,
1756 int error)
1757 {
1758 struct fuse_writepage_args *wpa =
1759 container_of(args, typeof(*wpa), ia.ap.args);
1760 struct inode *inode = wpa->inode;
1761 struct fuse_inode *fi = get_fuse_inode(inode);
1762 struct fuse_conn *fc = get_fuse_conn(inode);
1763
1764 mapping_set_error(inode->i_mapping, error);
1765 /*
1766 * A writeback finished and this might have updated mtime/ctime on
1767 * server making local mtime/ctime stale. Hence invalidate attrs.
1768 * Do this only if writeback_cache is not enabled. If writeback_cache
1769 * is enabled, we trust local ctime/mtime.
1770 */
1771 if (!fc->writeback_cache)
1772 fuse_invalidate_attr(inode);
1773 spin_lock(&fi->lock);
1774 rb_erase(&wpa->writepages_entry, &fi->writepages);
1775 while (wpa->next) {
1776 struct fuse_mount *fm = get_fuse_mount(inode);
1777 struct fuse_write_in *inarg = &wpa->ia.write.in;
1778 struct fuse_writepage_args *next = wpa->next;
1779
1780 wpa->next = next->next;
1781 next->next = NULL;
1782 next->ia.ff = fuse_file_get(wpa->ia.ff);
1783 tree_insert(&fi->writepages, next);
1784
1785 /*
1786 * Skip fuse_flush_writepages() to make it easy to crop requests
1787 * based on primary request size.
1788 *
1789 * 1st case (trivial): there are no concurrent activities using
1790 * fuse_set/release_nowrite. Then we're on safe side because
1791 * fuse_flush_writepages() would call fuse_send_writepage()
1792 * anyway.
1793 *
1794 * 2nd case: someone called fuse_set_nowrite and it is waiting
1795 * now for completion of all in-flight requests. This happens
1796 * rarely and no more than once per page, so this should be
1797 * okay.
1798 *
1799 * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1800 * of fuse_set_nowrite..fuse_release_nowrite section. The fact
1801 * that fuse_set_nowrite returned implies that all in-flight
1802 * requests were completed along with all of their secondary
1803 * requests. Further primary requests are blocked by negative
1804 * writectr. Hence there cannot be any in-flight requests and
1805 * no invocations of fuse_writepage_end() while we're in
1806 * fuse_set_nowrite..fuse_release_nowrite section.
1807 */
1808 fuse_send_writepage(fm, next, inarg->offset + inarg->size);
1809 }
1810 fi->writectr--;
1811 fuse_writepage_finish(fm, wpa);
1812 spin_unlock(&fi->lock);
1813 fuse_writepage_free(wpa);
1814 }
1815
1816 static struct fuse_file *__fuse_write_file_get(struct fuse_conn *fc,
1817 struct fuse_inode *fi)
1818 {
1819 struct fuse_file *ff = NULL;
1820
1821 spin_lock(&fi->lock);
1822 if (!list_empty(&fi->write_files)) {
1823 ff = list_entry(fi->write_files.next, struct fuse_file,
1824 write_entry);
1825 fuse_file_get(ff);
1826 }
1827 spin_unlock(&fi->lock);
1828
1829 return ff;
1830 }
1831
1832 static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1833 struct fuse_inode *fi)
1834 {
1835 struct fuse_file *ff = __fuse_write_file_get(fc, fi);
1836 WARN_ON(!ff);
1837 return ff;
1838 }
1839
1840 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1841 {
1842 struct fuse_conn *fc = get_fuse_conn(inode);
1843 struct fuse_inode *fi = get_fuse_inode(inode);
1844 struct fuse_file *ff;
1845 int err;
1846
1847 ff = __fuse_write_file_get(fc, fi);
1848 err = fuse_flush_times(inode, ff);
1849 if (ff)
1850 fuse_file_put(ff, false, false);
1851
1852 return err;
1853 }
1854
1855 static struct fuse_writepage_args *fuse_writepage_args_alloc(void)
1856 {
1857 struct fuse_writepage_args *wpa;
1858 struct fuse_args_pages *ap;
1859
1860 wpa = kzalloc(sizeof(*wpa), GFP_NOFS);
1861 if (wpa) {
1862 ap = &wpa->ia.ap;
1863 ap->num_pages = 0;
1864 ap->pages = fuse_pages_alloc(1, GFP_NOFS, &ap->descs);
1865 if (!ap->pages) {
1866 kfree(wpa);
1867 wpa = NULL;
1868 }
1869 }
1870 return wpa;
1871
1872 }
1873
1874 static int fuse_writepage_locked(struct page *page)
1875 {
1876 struct address_space *mapping = page->mapping;
1877 struct inode *inode = mapping->host;
1878 struct fuse_conn *fc = get_fuse_conn(inode);
1879 struct fuse_inode *fi = get_fuse_inode(inode);
1880 struct fuse_writepage_args *wpa;
1881 struct fuse_args_pages *ap;
1882 struct page *tmp_page;
1883 int error = -ENOMEM;
1884
1885 set_page_writeback(page);
1886
1887 wpa = fuse_writepage_args_alloc();
1888 if (!wpa)
1889 goto err;
1890 ap = &wpa->ia.ap;
1891
1892 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1893 if (!tmp_page)
1894 goto err_free;
1895
1896 error = -EIO;
1897 wpa->ia.ff = fuse_write_file_get(fc, fi);
1898 if (!wpa->ia.ff)
1899 goto err_nofile;
1900
1901 fuse_write_args_fill(&wpa->ia, wpa->ia.ff, page_offset(page), 0);
1902
1903 copy_highpage(tmp_page, page);
1904 wpa->ia.write.in.write_flags |= FUSE_WRITE_CACHE;
1905 wpa->next = NULL;
1906 ap->args.in_pages = true;
1907 ap->num_pages = 1;
1908 ap->pages[0] = tmp_page;
1909 ap->descs[0].offset = 0;
1910 ap->descs[0].length = PAGE_SIZE;
1911 ap->args.end = fuse_writepage_end;
1912 wpa->inode = inode;
1913
1914 inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1915 inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1916
1917 spin_lock(&fi->lock);
1918 tree_insert(&fi->writepages, wpa);
1919 list_add_tail(&wpa->queue_entry, &fi->queued_writes);
1920 fuse_flush_writepages(inode);
1921 spin_unlock(&fi->lock);
1922
1923 end_page_writeback(page);
1924
1925 return 0;
1926
1927 err_nofile:
1928 __free_page(tmp_page);
1929 err_free:
1930 kfree(wpa);
1931 err:
1932 mapping_set_error(page->mapping, error);
1933 end_page_writeback(page);
1934 return error;
1935 }
1936
1937 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1938 {
1939 int err;
1940
1941 if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1942 /*
1943 * ->writepages() should be called for sync() and friends. We
1944 * should only get here on direct reclaim and then we are
1945 * allowed to skip a page which is already in flight
1946 */
1947 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1948
1949 redirty_page_for_writepage(wbc, page);
1950 unlock_page(page);
1951 return 0;
1952 }
1953
1954 err = fuse_writepage_locked(page);
1955 unlock_page(page);
1956
1957 return err;
1958 }
1959
1960 struct fuse_fill_wb_data {
1961 struct fuse_writepage_args *wpa;
1962 struct fuse_file *ff;
1963 struct inode *inode;
1964 struct page **orig_pages;
1965 unsigned int max_pages;
1966 };
1967
1968 static bool fuse_pages_realloc(struct fuse_fill_wb_data *data)
1969 {
1970 struct fuse_args_pages *ap = &data->wpa->ia.ap;
1971 struct fuse_conn *fc = get_fuse_conn(data->inode);
1972 struct page **pages;
1973 struct fuse_page_desc *descs;
1974 unsigned int npages = min_t(unsigned int,
1975 max_t(unsigned int, data->max_pages * 2,
1976 FUSE_DEFAULT_MAX_PAGES_PER_REQ),
1977 fc->max_pages);
1978 WARN_ON(npages <= data->max_pages);
1979
1980 pages = fuse_pages_alloc(npages, GFP_NOFS, &descs);
1981 if (!pages)
1982 return false;
1983
1984 memcpy(pages, ap->pages, sizeof(struct page *) * ap->num_pages);
1985 memcpy(descs, ap->descs, sizeof(struct fuse_page_desc) * ap->num_pages);
1986 kfree(ap->pages);
1987 ap->pages = pages;
1988 ap->descs = descs;
1989 data->max_pages = npages;
1990
1991 return true;
1992 }
1993
1994 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1995 {
1996 struct fuse_writepage_args *wpa = data->wpa;
1997 struct inode *inode = data->inode;
1998 struct fuse_inode *fi = get_fuse_inode(inode);
1999 int num_pages = wpa->ia.ap.num_pages;
2000 int i;
2001
2002 wpa->ia.ff = fuse_file_get(data->ff);
2003 spin_lock(&fi->lock);
2004 list_add_tail(&wpa->queue_entry, &fi->queued_writes);
2005 fuse_flush_writepages(inode);
2006 spin_unlock(&fi->lock);
2007
2008 for (i = 0; i < num_pages; i++)
2009 end_page_writeback(data->orig_pages[i]);
2010 }
2011
2012 /*
2013 * Check under fi->lock if the page is under writeback, and insert it onto the
2014 * rb_tree if not. Otherwise iterate auxiliary write requests, to see if there's
2015 * one already added for a page at this offset. If there's none, then insert
2016 * this new request onto the auxiliary list, otherwise reuse the existing one by
2017 * swapping the new temp page with the old one.
2018 */
2019 static bool fuse_writepage_add(struct fuse_writepage_args *new_wpa,
2020 struct page *page)
2021 {
2022 struct fuse_inode *fi = get_fuse_inode(new_wpa->inode);
2023 struct fuse_writepage_args *tmp;
2024 struct fuse_writepage_args *old_wpa;
2025 struct fuse_args_pages *new_ap = &new_wpa->ia.ap;
2026
2027 WARN_ON(new_ap->num_pages != 0);
2028 new_ap->num_pages = 1;
2029
2030 spin_lock(&fi->lock);
2031 old_wpa = fuse_insert_writeback(&fi->writepages, new_wpa);
2032 if (!old_wpa) {
2033 spin_unlock(&fi->lock);
2034 return true;
2035 }
2036
2037 for (tmp = old_wpa->next; tmp; tmp = tmp->next) {
2038 pgoff_t curr_index;
2039
2040 WARN_ON(tmp->inode != new_wpa->inode);
2041 curr_index = tmp->ia.write.in.offset >> PAGE_SHIFT;
2042 if (curr_index == page->index) {
2043 WARN_ON(tmp->ia.ap.num_pages != 1);
2044 swap(tmp->ia.ap.pages[0], new_ap->pages[0]);
2045 break;
2046 }
2047 }
2048
2049 if (!tmp) {
2050 new_wpa->next = old_wpa->next;
2051 old_wpa->next = new_wpa;
2052 }
2053
2054 spin_unlock(&fi->lock);
2055
2056 if (tmp) {
2057 struct backing_dev_info *bdi = inode_to_bdi(new_wpa->inode);
2058
2059 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
2060 dec_node_page_state(new_ap->pages[0], NR_WRITEBACK_TEMP);
2061 wb_writeout_inc(&bdi->wb);
2062 fuse_writepage_free(new_wpa);
2063 }
2064
2065 return false;
2066 }
2067
2068 static bool fuse_writepage_need_send(struct fuse_conn *fc, struct page *page,
2069 struct fuse_args_pages *ap,
2070 struct fuse_fill_wb_data *data)
2071 {
2072 WARN_ON(!ap->num_pages);
2073
2074 /*
2075 * Being under writeback is unlikely but possible. For example direct
2076 * read to an mmaped fuse file will set the page dirty twice; once when
2077 * the pages are faulted with get_user_pages(), and then after the read
2078 * completed.
2079 */
2080 if (fuse_page_is_writeback(data->inode, page->index))
2081 return true;
2082
2083 /* Reached max pages */
2084 if (ap->num_pages == fc->max_pages)
2085 return true;
2086
2087 /* Reached max write bytes */
2088 if ((ap->num_pages + 1) * PAGE_SIZE > fc->max_write)
2089 return true;
2090
2091 /* Discontinuity */
2092 if (data->orig_pages[ap->num_pages - 1]->index + 1 != page->index)
2093 return true;
2094
2095 /* Need to grow the pages array? If so, did the expansion fail? */
2096 if (ap->num_pages == data->max_pages && !fuse_pages_realloc(data))
2097 return true;
2098
2099 return false;
2100 }
2101
2102 static int fuse_writepages_fill(struct page *page,
2103 struct writeback_control *wbc, void *_data)
2104 {
2105 struct fuse_fill_wb_data *data = _data;
2106 struct fuse_writepage_args *wpa = data->wpa;
2107 struct fuse_args_pages *ap = &wpa->ia.ap;
2108 struct inode *inode = data->inode;
2109 struct fuse_inode *fi = get_fuse_inode(inode);
2110 struct fuse_conn *fc = get_fuse_conn(inode);
2111 struct page *tmp_page;
2112 int err;
2113
2114 if (!data->ff) {
2115 err = -EIO;
2116 data->ff = fuse_write_file_get(fc, fi);
2117 if (!data->ff)
2118 goto out_unlock;
2119 }
2120
2121 if (wpa && fuse_writepage_need_send(fc, page, ap, data)) {
2122 fuse_writepages_send(data);
2123 data->wpa = NULL;
2124 }
2125
2126 err = -ENOMEM;
2127 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2128 if (!tmp_page)
2129 goto out_unlock;
2130
2131 /*
2132 * The page must not be redirtied until the writeout is completed
2133 * (i.e. userspace has sent a reply to the write request). Otherwise
2134 * there could be more than one temporary page instance for each real
2135 * page.
2136 *
2137 * This is ensured by holding the page lock in page_mkwrite() while
2138 * checking fuse_page_is_writeback(). We already hold the page lock
2139 * since clear_page_dirty_for_io() and keep it held until we add the
2140 * request to the fi->writepages list and increment ap->num_pages.
2141 * After this fuse_page_is_writeback() will indicate that the page is
2142 * under writeback, so we can release the page lock.
2143 */
2144 if (data->wpa == NULL) {
2145 err = -ENOMEM;
2146 wpa = fuse_writepage_args_alloc();
2147 if (!wpa) {
2148 __free_page(tmp_page);
2149 goto out_unlock;
2150 }
2151 data->max_pages = 1;
2152
2153 ap = &wpa->ia.ap;
2154 fuse_write_args_fill(&wpa->ia, data->ff, page_offset(page), 0);
2155 wpa->ia.write.in.write_flags |= FUSE_WRITE_CACHE;
2156 wpa->next = NULL;
2157 ap->args.in_pages = true;
2158 ap->args.end = fuse_writepage_end;
2159 ap->num_pages = 0;
2160 wpa->inode = inode;
2161 }
2162 set_page_writeback(page);
2163
2164 copy_highpage(tmp_page, page);
2165 ap->pages[ap->num_pages] = tmp_page;
2166 ap->descs[ap->num_pages].offset = 0;
2167 ap->descs[ap->num_pages].length = PAGE_SIZE;
2168 data->orig_pages[ap->num_pages] = page;
2169
2170 inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
2171 inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
2172
2173 err = 0;
2174 if (data->wpa) {
2175 /*
2176 * Protected by fi->lock against concurrent access by
2177 * fuse_page_is_writeback().
2178 */
2179 spin_lock(&fi->lock);
2180 ap->num_pages++;
2181 spin_unlock(&fi->lock);
2182 } else if (fuse_writepage_add(wpa, page)) {
2183 data->wpa = wpa;
2184 } else {
2185 end_page_writeback(page);
2186 }
2187 out_unlock:
2188 unlock_page(page);
2189
2190 return err;
2191 }
2192
2193 static int fuse_writepages(struct address_space *mapping,
2194 struct writeback_control *wbc)
2195 {
2196 struct inode *inode = mapping->host;
2197 struct fuse_conn *fc = get_fuse_conn(inode);
2198 struct fuse_fill_wb_data data;
2199 int err;
2200
2201 err = -EIO;
2202 if (fuse_is_bad(inode))
2203 goto out;
2204
2205 data.inode = inode;
2206 data.wpa = NULL;
2207 data.ff = NULL;
2208
2209 err = -ENOMEM;
2210 data.orig_pages = kcalloc(fc->max_pages,
2211 sizeof(struct page *),
2212 GFP_NOFS);
2213 if (!data.orig_pages)
2214 goto out;
2215
2216 err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
2217 if (data.wpa) {
2218 WARN_ON(!data.wpa->ia.ap.num_pages);
2219 fuse_writepages_send(&data);
2220 }
2221 if (data.ff)
2222 fuse_file_put(data.ff, false, false);
2223
2224 kfree(data.orig_pages);
2225 out:
2226 return err;
2227 }
2228
2229 /*
2230 * It's worthy to make sure that space is reserved on disk for the write,
2231 * but how to implement it without killing performance need more thinking.
2232 */
2233 static int fuse_write_begin(struct file *file, struct address_space *mapping,
2234 loff_t pos, unsigned len, unsigned flags,
2235 struct page **pagep, void **fsdata)
2236 {
2237 pgoff_t index = pos >> PAGE_SHIFT;
2238 struct fuse_conn *fc = get_fuse_conn(file_inode(file));
2239 struct page *page;
2240 loff_t fsize;
2241 int err = -ENOMEM;
2242
2243 WARN_ON(!fc->writeback_cache);
2244
2245 page = grab_cache_page_write_begin(mapping, index, flags);
2246 if (!page)
2247 goto error;
2248
2249 fuse_wait_on_page_writeback(mapping->host, page->index);
2250
2251 if (PageUptodate(page) || len == PAGE_SIZE)
2252 goto success;
2253 /*
2254 * Check if the start this page comes after the end of file, in which
2255 * case the readpage can be optimized away.
2256 */
2257 fsize = i_size_read(mapping->host);
2258 if (fsize <= (pos & PAGE_MASK)) {
2259 size_t off = pos & ~PAGE_MASK;
2260 if (off)
2261 zero_user_segment(page, 0, off);
2262 goto success;
2263 }
2264 err = fuse_do_readpage(file, page);
2265 if (err)
2266 goto cleanup;
2267 success:
2268 *pagep = page;
2269 return 0;
2270
2271 cleanup:
2272 unlock_page(page);
2273 put_page(page);
2274 error:
2275 return err;
2276 }
2277
2278 static int fuse_write_end(struct file *file, struct address_space *mapping,
2279 loff_t pos, unsigned len, unsigned copied,
2280 struct page *page, void *fsdata)
2281 {
2282 struct inode *inode = page->mapping->host;
2283
2284 /* Haven't copied anything? Skip zeroing, size extending, dirtying. */
2285 if (!copied)
2286 goto unlock;
2287
2288 if (!PageUptodate(page)) {
2289 /* Zero any unwritten bytes at the end of the page */
2290 size_t endoff = (pos + copied) & ~PAGE_MASK;
2291 if (endoff)
2292 zero_user_segment(page, endoff, PAGE_SIZE);
2293 SetPageUptodate(page);
2294 }
2295
2296 fuse_write_update_size(inode, pos + copied);
2297 set_page_dirty(page);
2298
2299 unlock:
2300 unlock_page(page);
2301 put_page(page);
2302
2303 return copied;
2304 }
2305
2306 static int fuse_launder_page(struct page *page)
2307 {
2308 int err = 0;
2309 if (clear_page_dirty_for_io(page)) {
2310 struct inode *inode = page->mapping->host;
2311
2312 /* Serialize with pending writeback for the same page */
2313 fuse_wait_on_page_writeback(inode, page->index);
2314 err = fuse_writepage_locked(page);
2315 if (!err)
2316 fuse_wait_on_page_writeback(inode, page->index);
2317 }
2318 return err;
2319 }
2320
2321 /*
2322 * Write back dirty pages now, because there may not be any suitable
2323 * open files later
2324 */
2325 static void fuse_vma_close(struct vm_area_struct *vma)
2326 {
2327 filemap_write_and_wait(vma->vm_file->f_mapping);
2328 }
2329
2330 /*
2331 * Wait for writeback against this page to complete before allowing it
2332 * to be marked dirty again, and hence written back again, possibly
2333 * before the previous writepage completed.
2334 *
2335 * Block here, instead of in ->writepage(), so that the userspace fs
2336 * can only block processes actually operating on the filesystem.
2337 *
2338 * Otherwise unprivileged userspace fs would be able to block
2339 * unrelated:
2340 *
2341 * - page migration
2342 * - sync(2)
2343 * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2344 */
2345 static vm_fault_t fuse_page_mkwrite(struct vm_fault *vmf)
2346 {
2347 struct page *page = vmf->page;
2348 struct inode *inode = file_inode(vmf->vma->vm_file);
2349
2350 file_update_time(vmf->vma->vm_file);
2351 lock_page(page);
2352 if (page->mapping != inode->i_mapping) {
2353 unlock_page(page);
2354 return VM_FAULT_NOPAGE;
2355 }
2356
2357 fuse_wait_on_page_writeback(inode, page->index);
2358 return VM_FAULT_LOCKED;
2359 }
2360
2361 static const struct vm_operations_struct fuse_file_vm_ops = {
2362 .close = fuse_vma_close,
2363 .fault = filemap_fault,
2364 .map_pages = filemap_map_pages,
2365 .page_mkwrite = fuse_page_mkwrite,
2366 };
2367
2368 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2369 {
2370 struct fuse_file *ff = file->private_data;
2371
2372 /* DAX mmap is superior to direct_io mmap */
2373 if (FUSE_IS_DAX(file_inode(file)))
2374 return fuse_dax_mmap(file, vma);
2375
2376 if (ff->open_flags & FOPEN_DIRECT_IO) {
2377 /* Can't provide the coherency needed for MAP_SHARED */
2378 if (vma->vm_flags & VM_MAYSHARE)
2379 return -ENODEV;
2380
2381 invalidate_inode_pages2(file->f_mapping);
2382
2383 return generic_file_mmap(file, vma);
2384 }
2385
2386 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2387 fuse_link_write_file(file);
2388
2389 file_accessed(file);
2390 vma->vm_ops = &fuse_file_vm_ops;
2391 return 0;
2392 }
2393
2394 static int convert_fuse_file_lock(struct fuse_conn *fc,
2395 const struct fuse_file_lock *ffl,
2396 struct file_lock *fl)
2397 {
2398 switch (ffl->type) {
2399 case F_UNLCK:
2400 break;
2401
2402 case F_RDLCK:
2403 case F_WRLCK:
2404 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2405 ffl->end < ffl->start)
2406 return -EIO;
2407
2408 fl->fl_start = ffl->start;
2409 fl->fl_end = ffl->end;
2410
2411 /*
2412 * Convert pid into init's pid namespace. The locks API will
2413 * translate it into the caller's pid namespace.
2414 */
2415 rcu_read_lock();
2416 fl->fl_pid = pid_nr_ns(find_pid_ns(ffl->pid, fc->pid_ns), &init_pid_ns);
2417 rcu_read_unlock();
2418 break;
2419
2420 default:
2421 return -EIO;
2422 }
2423 fl->fl_type = ffl->type;
2424 return 0;
2425 }
2426
2427 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2428 const struct file_lock *fl, int opcode, pid_t pid,
2429 int flock, struct fuse_lk_in *inarg)
2430 {
2431 struct inode *inode = file_inode(file);
2432 struct fuse_conn *fc = get_fuse_conn(inode);
2433 struct fuse_file *ff = file->private_data;
2434
2435 memset(inarg, 0, sizeof(*inarg));
2436 inarg->fh = ff->fh;
2437 inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2438 inarg->lk.start = fl->fl_start;
2439 inarg->lk.end = fl->fl_end;
2440 inarg->lk.type = fl->fl_type;
2441 inarg->lk.pid = pid;
2442 if (flock)
2443 inarg->lk_flags |= FUSE_LK_FLOCK;
2444 args->opcode = opcode;
2445 args->nodeid = get_node_id(inode);
2446 args->in_numargs = 1;
2447 args->in_args[0].size = sizeof(*inarg);
2448 args->in_args[0].value = inarg;
2449 }
2450
2451 static int fuse_getlk(struct file *file, struct file_lock *fl)
2452 {
2453 struct inode *inode = file_inode(file);
2454 struct fuse_mount *fm = get_fuse_mount(inode);
2455 FUSE_ARGS(args);
2456 struct fuse_lk_in inarg;
2457 struct fuse_lk_out outarg;
2458 int err;
2459
2460 fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2461 args.out_numargs = 1;
2462 args.out_args[0].size = sizeof(outarg);
2463 args.out_args[0].value = &outarg;
2464 err = fuse_simple_request(fm, &args);
2465 if (!err)
2466 err = convert_fuse_file_lock(fm->fc, &outarg.lk, fl);
2467
2468 return err;
2469 }
2470
2471 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2472 {
2473 struct inode *inode = file_inode(file);
2474 struct fuse_mount *fm = get_fuse_mount(inode);
2475 FUSE_ARGS(args);
2476 struct fuse_lk_in inarg;
2477 int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2478 struct pid *pid = fl->fl_type != F_UNLCK ? task_tgid(current) : NULL;
2479 pid_t pid_nr = pid_nr_ns(pid, fm->fc->pid_ns);
2480 int err;
2481
2482 if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2483 /* NLM needs asynchronous locks, which we don't support yet */
2484 return -ENOLCK;
2485 }
2486
2487 /* Unlock on close is handled by the flush method */
2488 if ((fl->fl_flags & FL_CLOSE_POSIX) == FL_CLOSE_POSIX)
2489 return 0;
2490
2491 fuse_lk_fill(&args, file, fl, opcode, pid_nr, flock, &inarg);
2492 err = fuse_simple_request(fm, &args);
2493
2494 /* locking is restartable */
2495 if (err == -EINTR)
2496 err = -ERESTARTSYS;
2497
2498 return err;
2499 }
2500
2501 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2502 {
2503 struct inode *inode = file_inode(file);
2504 struct fuse_conn *fc = get_fuse_conn(inode);
2505 int err;
2506
2507 if (cmd == F_CANCELLK) {
2508 err = 0;
2509 } else if (cmd == F_GETLK) {
2510 if (fc->no_lock) {
2511 posix_test_lock(file, fl);
2512 err = 0;
2513 } else
2514 err = fuse_getlk(file, fl);
2515 } else {
2516 if (fc->no_lock)
2517 err = posix_lock_file(file, fl, NULL);
2518 else
2519 err = fuse_setlk(file, fl, 0);
2520 }
2521 return err;
2522 }
2523
2524 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2525 {
2526 struct inode *inode = file_inode(file);
2527 struct fuse_conn *fc = get_fuse_conn(inode);
2528 int err;
2529
2530 if (fc->no_flock) {
2531 err = locks_lock_file_wait(file, fl);
2532 } else {
2533 struct fuse_file *ff = file->private_data;
2534
2535 /* emulate flock with POSIX locks */
2536 ff->flock = true;
2537 err = fuse_setlk(file, fl, 1);
2538 }
2539
2540 return err;
2541 }
2542
2543 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2544 {
2545 struct inode *inode = mapping->host;
2546 struct fuse_mount *fm = get_fuse_mount(inode);
2547 FUSE_ARGS(args);
2548 struct fuse_bmap_in inarg;
2549 struct fuse_bmap_out outarg;
2550 int err;
2551
2552 if (!inode->i_sb->s_bdev || fm->fc->no_bmap)
2553 return 0;
2554
2555 memset(&inarg, 0, sizeof(inarg));
2556 inarg.block = block;
2557 inarg.blocksize = inode->i_sb->s_blocksize;
2558 args.opcode = FUSE_BMAP;
2559 args.nodeid = get_node_id(inode);
2560 args.in_numargs = 1;
2561 args.in_args[0].size = sizeof(inarg);
2562 args.in_args[0].value = &inarg;
2563 args.out_numargs = 1;
2564 args.out_args[0].size = sizeof(outarg);
2565 args.out_args[0].value = &outarg;
2566 err = fuse_simple_request(fm, &args);
2567 if (err == -ENOSYS)
2568 fm->fc->no_bmap = 1;
2569
2570 return err ? 0 : outarg.block;
2571 }
2572
2573 static loff_t fuse_lseek(struct file *file, loff_t offset, int whence)
2574 {
2575 struct inode *inode = file->f_mapping->host;
2576 struct fuse_mount *fm = get_fuse_mount(inode);
2577 struct fuse_file *ff = file->private_data;
2578 FUSE_ARGS(args);
2579 struct fuse_lseek_in inarg = {
2580 .fh = ff->fh,
2581 .offset = offset,
2582 .whence = whence
2583 };
2584 struct fuse_lseek_out outarg;
2585 int err;
2586
2587 if (fm->fc->no_lseek)
2588 goto fallback;
2589
2590 args.opcode = FUSE_LSEEK;
2591 args.nodeid = ff->nodeid;
2592 args.in_numargs = 1;
2593 args.in_args[0].size = sizeof(inarg);
2594 args.in_args[0].value = &inarg;
2595 args.out_numargs = 1;
2596 args.out_args[0].size = sizeof(outarg);
2597 args.out_args[0].value = &outarg;
2598 err = fuse_simple_request(fm, &args);
2599 if (err) {
2600 if (err == -ENOSYS) {
2601 fm->fc->no_lseek = 1;
2602 goto fallback;
2603 }
2604 return err;
2605 }
2606
2607 return vfs_setpos(file, outarg.offset, inode->i_sb->s_maxbytes);
2608
2609 fallback:
2610 err = fuse_update_attributes(inode, file);
2611 if (!err)
2612 return generic_file_llseek(file, offset, whence);
2613 else
2614 return err;
2615 }
2616
2617 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2618 {
2619 loff_t retval;
2620 struct inode *inode = file_inode(file);
2621
2622 switch (whence) {
2623 case SEEK_SET:
2624 case SEEK_CUR:
2625 /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2626 retval = generic_file_llseek(file, offset, whence);
2627 break;
2628 case SEEK_END:
2629 inode_lock(inode);
2630 retval = fuse_update_attributes(inode, file);
2631 if (!retval)
2632 retval = generic_file_llseek(file, offset, whence);
2633 inode_unlock(inode);
2634 break;
2635 case SEEK_HOLE:
2636 case SEEK_DATA:
2637 inode_lock(inode);
2638 retval = fuse_lseek(file, offset, whence);
2639 inode_unlock(inode);
2640 break;
2641 default:
2642 retval = -EINVAL;
2643 }
2644
2645 return retval;
2646 }
2647
2648 /*
2649 * All files which have been polled are linked to RB tree
2650 * fuse_conn->polled_files which is indexed by kh. Walk the tree and
2651 * find the matching one.
2652 */
2653 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2654 struct rb_node **parent_out)
2655 {
2656 struct rb_node **link = &fc->polled_files.rb_node;
2657 struct rb_node *last = NULL;
2658
2659 while (*link) {
2660 struct fuse_file *ff;
2661
2662 last = *link;
2663 ff = rb_entry(last, struct fuse_file, polled_node);
2664
2665 if (kh < ff->kh)
2666 link = &last->rb_left;
2667 else if (kh > ff->kh)
2668 link = &last->rb_right;
2669 else
2670 return link;
2671 }
2672
2673 if (parent_out)
2674 *parent_out = last;
2675 return link;
2676 }
2677
2678 /*
2679 * The file is about to be polled. Make sure it's on the polled_files
2680 * RB tree. Note that files once added to the polled_files tree are
2681 * not removed before the file is released. This is because a file
2682 * polled once is likely to be polled again.
2683 */
2684 static void fuse_register_polled_file(struct fuse_conn *fc,
2685 struct fuse_file *ff)
2686 {
2687 spin_lock(&fc->lock);
2688 if (RB_EMPTY_NODE(&ff->polled_node)) {
2689 struct rb_node **link, *parent;
2690
2691 link = fuse_find_polled_node(fc, ff->kh, &parent);
2692 BUG_ON(*link);
2693 rb_link_node(&ff->polled_node, parent, link);
2694 rb_insert_color(&ff->polled_node, &fc->polled_files);
2695 }
2696 spin_unlock(&fc->lock);
2697 }
2698
2699 __poll_t fuse_file_poll(struct file *file, poll_table *wait)
2700 {
2701 struct fuse_file *ff = file->private_data;
2702 struct fuse_mount *fm = ff->fm;
2703 struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2704 struct fuse_poll_out outarg;
2705 FUSE_ARGS(args);
2706 int err;
2707
2708 if (fm->fc->no_poll)
2709 return DEFAULT_POLLMASK;
2710
2711 poll_wait(file, &ff->poll_wait, wait);
2712 inarg.events = mangle_poll(poll_requested_events(wait));
2713
2714 /*
2715 * Ask for notification iff there's someone waiting for it.
2716 * The client may ignore the flag and always notify.
2717 */
2718 if (waitqueue_active(&ff->poll_wait)) {
2719 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2720 fuse_register_polled_file(fm->fc, ff);
2721 }
2722
2723 args.opcode = FUSE_POLL;
2724 args.nodeid = ff->nodeid;
2725 args.in_numargs = 1;
2726 args.in_args[0].size = sizeof(inarg);
2727 args.in_args[0].value = &inarg;
2728 args.out_numargs = 1;
2729 args.out_args[0].size = sizeof(outarg);
2730 args.out_args[0].value = &outarg;
2731 err = fuse_simple_request(fm, &args);
2732
2733 if (!err)
2734 return demangle_poll(outarg.revents);
2735 if (err == -ENOSYS) {
2736 fm->fc->no_poll = 1;
2737 return DEFAULT_POLLMASK;
2738 }
2739 return EPOLLERR;
2740 }
2741 EXPORT_SYMBOL_GPL(fuse_file_poll);
2742
2743 /*
2744 * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2745 * wakes up the poll waiters.
2746 */
2747 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2748 struct fuse_notify_poll_wakeup_out *outarg)
2749 {
2750 u64 kh = outarg->kh;
2751 struct rb_node **link;
2752
2753 spin_lock(&fc->lock);
2754
2755 link = fuse_find_polled_node(fc, kh, NULL);
2756 if (*link) {
2757 struct fuse_file *ff;
2758
2759 ff = rb_entry(*link, struct fuse_file, polled_node);
2760 wake_up_interruptible_sync(&ff->poll_wait);
2761 }
2762
2763 spin_unlock(&fc->lock);
2764 return 0;
2765 }
2766
2767 static void fuse_do_truncate(struct file *file)
2768 {
2769 struct inode *inode = file->f_mapping->host;
2770 struct iattr attr;
2771
2772 attr.ia_valid = ATTR_SIZE;
2773 attr.ia_size = i_size_read(inode);
2774
2775 attr.ia_file = file;
2776 attr.ia_valid |= ATTR_FILE;
2777
2778 fuse_do_setattr(file_dentry(file), &attr, file);
2779 }
2780
2781 static inline loff_t fuse_round_up(struct fuse_conn *fc, loff_t off)
2782 {
2783 return round_up(off, fc->max_pages << PAGE_SHIFT);
2784 }
2785
2786 static ssize_t
2787 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
2788 {
2789 DECLARE_COMPLETION_ONSTACK(wait);
2790 ssize_t ret = 0;
2791 struct file *file = iocb->ki_filp;
2792 struct fuse_file *ff = file->private_data;
2793 loff_t pos = 0;
2794 struct inode *inode;
2795 loff_t i_size;
2796 size_t count = iov_iter_count(iter), shortened = 0;
2797 loff_t offset = iocb->ki_pos;
2798 struct fuse_io_priv *io;
2799
2800 pos = offset;
2801 inode = file->f_mapping->host;
2802 i_size = i_size_read(inode);
2803
2804 if ((iov_iter_rw(iter) == READ) && (offset >= i_size))
2805 return 0;
2806
2807 io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2808 if (!io)
2809 return -ENOMEM;
2810 spin_lock_init(&io->lock);
2811 kref_init(&io->refcnt);
2812 io->reqs = 1;
2813 io->bytes = -1;
2814 io->size = 0;
2815 io->offset = offset;
2816 io->write = (iov_iter_rw(iter) == WRITE);
2817 io->err = 0;
2818 /*
2819 * By default, we want to optimize all I/Os with async request
2820 * submission to the client filesystem if supported.
2821 */
2822 io->async = ff->fm->fc->async_dio;
2823 io->iocb = iocb;
2824 io->blocking = is_sync_kiocb(iocb);
2825
2826 /* optimization for short read */
2827 if (io->async && !io->write && offset + count > i_size) {
2828 iov_iter_truncate(iter, fuse_round_up(ff->fm->fc, i_size - offset));
2829 shortened = count - iov_iter_count(iter);
2830 count -= shortened;
2831 }
2832
2833 /*
2834 * We cannot asynchronously extend the size of a file.
2835 * In such case the aio will behave exactly like sync io.
2836 */
2837 if ((offset + count > i_size) && io->write)
2838 io->blocking = true;
2839
2840 if (io->async && io->blocking) {
2841 /*
2842 * Additional reference to keep io around after
2843 * calling fuse_aio_complete()
2844 */
2845 kref_get(&io->refcnt);
2846 io->done = &wait;
2847 }
2848
2849 if (iov_iter_rw(iter) == WRITE) {
2850 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
2851 fuse_invalidate_attr(inode);
2852 } else {
2853 ret = __fuse_direct_read(io, iter, &pos);
2854 }
2855 iov_iter_reexpand(iter, iov_iter_count(iter) + shortened);
2856
2857 if (io->async) {
2858 bool blocking = io->blocking;
2859
2860 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2861
2862 /* we have a non-extending, async request, so return */
2863 if (!blocking)
2864 return -EIOCBQUEUED;
2865
2866 wait_for_completion(&wait);
2867 ret = fuse_get_res_by_io(io);
2868 }
2869
2870 kref_put(&io->refcnt, fuse_io_release);
2871
2872 if (iov_iter_rw(iter) == WRITE) {
2873 if (ret > 0)
2874 fuse_write_update_size(inode, pos);
2875 else if (ret < 0 && offset + count > i_size)
2876 fuse_do_truncate(file);
2877 }
2878
2879 return ret;
2880 }
2881
2882 static int fuse_writeback_range(struct inode *inode, loff_t start, loff_t end)
2883 {
2884 int err = filemap_write_and_wait_range(inode->i_mapping, start, end);
2885
2886 if (!err)
2887 fuse_sync_writes(inode);
2888
2889 return err;
2890 }
2891
2892 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2893 loff_t length)
2894 {
2895 struct fuse_file *ff = file->private_data;
2896 struct inode *inode = file_inode(file);
2897 struct fuse_inode *fi = get_fuse_inode(inode);
2898 struct fuse_mount *fm = ff->fm;
2899 FUSE_ARGS(args);
2900 struct fuse_fallocate_in inarg = {
2901 .fh = ff->fh,
2902 .offset = offset,
2903 .length = length,
2904 .mode = mode
2905 };
2906 int err;
2907 bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
2908 (mode & (FALLOC_FL_PUNCH_HOLE |
2909 FALLOC_FL_ZERO_RANGE));
2910
2911 bool block_faults = FUSE_IS_DAX(inode) && lock_inode;
2912
2913 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
2914 FALLOC_FL_ZERO_RANGE))
2915 return -EOPNOTSUPP;
2916
2917 if (fm->fc->no_fallocate)
2918 return -EOPNOTSUPP;
2919
2920 if (lock_inode) {
2921 inode_lock(inode);
2922 if (block_faults) {
2923 filemap_invalidate_lock(inode->i_mapping);
2924 err = fuse_dax_break_layouts(inode, 0, 0);
2925 if (err)
2926 goto out;
2927 }
2928
2929 if (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE)) {
2930 loff_t endbyte = offset + length - 1;
2931
2932 err = fuse_writeback_range(inode, offset, endbyte);
2933 if (err)
2934 goto out;
2935 }
2936 }
2937
2938 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
2939 offset + length > i_size_read(inode)) {
2940 err = inode_newsize_ok(inode, offset + length);
2941 if (err)
2942 goto out;
2943 }
2944
2945 if (!(mode & FALLOC_FL_KEEP_SIZE))
2946 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2947
2948 args.opcode = FUSE_FALLOCATE;
2949 args.nodeid = ff->nodeid;
2950 args.in_numargs = 1;
2951 args.in_args[0].size = sizeof(inarg);
2952 args.in_args[0].value = &inarg;
2953 err = fuse_simple_request(fm, &args);
2954 if (err == -ENOSYS) {
2955 fm->fc->no_fallocate = 1;
2956 err = -EOPNOTSUPP;
2957 }
2958 if (err)
2959 goto out;
2960
2961 /* we could have extended the file */
2962 if (!(mode & FALLOC_FL_KEEP_SIZE)) {
2963 bool changed = fuse_write_update_size(inode, offset + length);
2964
2965 if (changed && fm->fc->writeback_cache)
2966 file_update_time(file);
2967 }
2968
2969 if (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE))
2970 truncate_pagecache_range(inode, offset, offset + length - 1);
2971
2972 fuse_invalidate_attr(inode);
2973
2974 out:
2975 if (!(mode & FALLOC_FL_KEEP_SIZE))
2976 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2977
2978 if (block_faults)
2979 filemap_invalidate_unlock(inode->i_mapping);
2980
2981 if (lock_inode)
2982 inode_unlock(inode);
2983
2984 return err;
2985 }
2986
2987 static ssize_t __fuse_copy_file_range(struct file *file_in, loff_t pos_in,
2988 struct file *file_out, loff_t pos_out,
2989 size_t len, unsigned int flags)
2990 {
2991 struct fuse_file *ff_in = file_in->private_data;
2992 struct fuse_file *ff_out = file_out->private_data;
2993 struct inode *inode_in = file_inode(file_in);
2994 struct inode *inode_out = file_inode(file_out);
2995 struct fuse_inode *fi_out = get_fuse_inode(inode_out);
2996 struct fuse_mount *fm = ff_in->fm;
2997 struct fuse_conn *fc = fm->fc;
2998 FUSE_ARGS(args);
2999 struct fuse_copy_file_range_in inarg = {
3000 .fh_in = ff_in->fh,
3001 .off_in = pos_in,
3002 .nodeid_out = ff_out->nodeid,
3003 .fh_out = ff_out->fh,
3004 .off_out = pos_out,
3005 .len = len,
3006 .flags = flags
3007 };
3008 struct fuse_write_out outarg;
3009 ssize_t err;
3010 /* mark unstable when write-back is not used, and file_out gets
3011 * extended */
3012 bool is_unstable = (!fc->writeback_cache) &&
3013 ((pos_out + len) > inode_out->i_size);
3014
3015 if (fc->no_copy_file_range)
3016 return -EOPNOTSUPP;
3017
3018 if (file_inode(file_in)->i_sb != file_inode(file_out)->i_sb)
3019 return -EXDEV;
3020
3021 inode_lock(inode_in);
3022 err = fuse_writeback_range(inode_in, pos_in, pos_in + len - 1);
3023 inode_unlock(inode_in);
3024 if (err)
3025 return err;
3026
3027 inode_lock(inode_out);
3028
3029 err = file_modified(file_out);
3030 if (err)
3031 goto out;
3032
3033 /*
3034 * Write out dirty pages in the destination file before sending the COPY
3035 * request to userspace. After the request is completed, truncate off
3036 * pages (including partial ones) from the cache that have been copied,
3037 * since these contain stale data at that point.
3038 *
3039 * This should be mostly correct, but if the COPY writes to partial
3040 * pages (at the start or end) and the parts not covered by the COPY are
3041 * written through a memory map after calling fuse_writeback_range(),
3042 * then these partial page modifications will be lost on truncation.
3043 *
3044 * It is unlikely that someone would rely on such mixed style
3045 * modifications. Yet this does give less guarantees than if the
3046 * copying was performed with write(2).
3047 *
3048 * To fix this a mapping->invalidate_lock could be used to prevent new
3049 * faults while the copy is ongoing.
3050 */
3051 err = fuse_writeback_range(inode_out, pos_out, pos_out + len - 1);
3052 if (err)
3053 goto out;
3054
3055 if (is_unstable)
3056 set_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3057
3058 args.opcode = FUSE_COPY_FILE_RANGE;
3059 args.nodeid = ff_in->nodeid;
3060 args.in_numargs = 1;
3061 args.in_args[0].size = sizeof(inarg);
3062 args.in_args[0].value = &inarg;
3063 args.out_numargs = 1;
3064 args.out_args[0].size = sizeof(outarg);
3065 args.out_args[0].value = &outarg;
3066 err = fuse_simple_request(fm, &args);
3067 if (err == -ENOSYS) {
3068 fc->no_copy_file_range = 1;
3069 err = -EOPNOTSUPP;
3070 }
3071 if (err)
3072 goto out;
3073
3074 truncate_inode_pages_range(inode_out->i_mapping,
3075 ALIGN_DOWN(pos_out, PAGE_SIZE),
3076 ALIGN(pos_out + outarg.size, PAGE_SIZE) - 1);
3077
3078 if (fc->writeback_cache) {
3079 fuse_write_update_size(inode_out, pos_out + outarg.size);
3080 file_update_time(file_out);
3081 }
3082
3083 fuse_invalidate_attr(inode_out);
3084
3085 err = outarg.size;
3086 out:
3087 if (is_unstable)
3088 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3089
3090 inode_unlock(inode_out);
3091 file_accessed(file_in);
3092
3093 return err;
3094 }
3095
3096 static ssize_t fuse_copy_file_range(struct file *src_file, loff_t src_off,
3097 struct file *dst_file, loff_t dst_off,
3098 size_t len, unsigned int flags)
3099 {
3100 ssize_t ret;
3101
3102 ret = __fuse_copy_file_range(src_file, src_off, dst_file, dst_off,
3103 len, flags);
3104
3105 if (ret == -EOPNOTSUPP || ret == -EXDEV)
3106 ret = generic_copy_file_range(src_file, src_off, dst_file,
3107 dst_off, len, flags);
3108 return ret;
3109 }
3110
3111 static const struct file_operations fuse_file_operations = {
3112 .llseek = fuse_file_llseek,
3113 .read_iter = fuse_file_read_iter,
3114 .write_iter = fuse_file_write_iter,
3115 .mmap = fuse_file_mmap,
3116 .open = fuse_open,
3117 .flush = fuse_flush,
3118 .release = fuse_release,
3119 .fsync = fuse_fsync,
3120 .lock = fuse_file_lock,
3121 .get_unmapped_area = thp_get_unmapped_area,
3122 .flock = fuse_file_flock,
3123 .splice_read = generic_file_splice_read,
3124 .splice_write = iter_file_splice_write,
3125 .unlocked_ioctl = fuse_file_ioctl,
3126 .compat_ioctl = fuse_file_compat_ioctl,
3127 .poll = fuse_file_poll,
3128 .fallocate = fuse_file_fallocate,
3129 .copy_file_range = fuse_copy_file_range,
3130 };
3131
3132 static const struct address_space_operations fuse_file_aops = {
3133 .readpage = fuse_readpage,
3134 .readahead = fuse_readahead,
3135 .writepage = fuse_writepage,
3136 .writepages = fuse_writepages,
3137 .launder_page = fuse_launder_page,
3138 .set_page_dirty = __set_page_dirty_nobuffers,
3139 .bmap = fuse_bmap,
3140 .direct_IO = fuse_direct_IO,
3141 .write_begin = fuse_write_begin,
3142 .write_end = fuse_write_end,
3143 };
3144
3145 void fuse_init_file_inode(struct inode *inode)
3146 {
3147 struct fuse_inode *fi = get_fuse_inode(inode);
3148
3149 inode->i_fop = &fuse_file_operations;
3150 inode->i_data.a_ops = &fuse_file_aops;
3151
3152 INIT_LIST_HEAD(&fi->write_files);
3153 INIT_LIST_HEAD(&fi->queued_writes);
3154 fi->writectr = 0;
3155 init_waitqueue_head(&fi->page_waitq);
3156 fi->writepages = RB_ROOT;
3157
3158 if (IS_ENABLED(CONFIG_FUSE_DAX))
3159 fuse_dax_inode_init(inode);
3160 }
Ubuntu Jammy Linux kernel mirror