Loading...
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
22static 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
51struct fuse_release_args {
52 struct fuse_args args;
53 struct fuse_release_in inarg;
54 struct inode *inode;
55};
56
57struct 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
84void fuse_file_free(struct fuse_file *ff)
85{
86 kfree(ff->release_args);
87 mutex_destroy(&ff->readdir.lock);
88 kfree(ff);
89}
90
91static struct fuse_file *fuse_file_get(struct fuse_file *ff)
92{
93 refcount_inc(&ff->count);
94 return ff;
95}
96
97static 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
106static 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
127struct 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
169int 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}
179EXPORT_SYMBOL_GPL(fuse_do_open);
180
181static 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
196void 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_STREAM)
202 stream_open(inode, file);
203 else if (ff->open_flags & FOPEN_NONSEEKABLE)
204 nonseekable_open(inode, file);
205
206 if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
207 struct fuse_inode *fi = get_fuse_inode(inode);
208
209 spin_lock(&fi->lock);
210 fi->attr_version = atomic64_inc_return(&fc->attr_version);
211 i_size_write(inode, 0);
212 spin_unlock(&fi->lock);
213 file_update_time(file);
214 fuse_invalidate_attr_mask(inode, FUSE_STATX_MODSIZE);
215 }
216 if ((file->f_mode & FMODE_WRITE) && fc->writeback_cache)
217 fuse_link_write_file(file);
218}
219
220int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
221{
222 struct fuse_mount *fm = get_fuse_mount(inode);
223 struct fuse_conn *fc = fm->fc;
224 int err;
225 bool is_wb_truncate = (file->f_flags & O_TRUNC) &&
226 fc->atomic_o_trunc &&
227 fc->writeback_cache;
228 bool dax_truncate = (file->f_flags & O_TRUNC) &&
229 fc->atomic_o_trunc && FUSE_IS_DAX(inode);
230
231 if (fuse_is_bad(inode))
232 return -EIO;
233
234 err = generic_file_open(inode, file);
235 if (err)
236 return err;
237
238 if (is_wb_truncate || dax_truncate)
239 inode_lock(inode);
240
241 if (dax_truncate) {
242 filemap_invalidate_lock(inode->i_mapping);
243 err = fuse_dax_break_layouts(inode, 0, 0);
244 if (err)
245 goto out_inode_unlock;
246 }
247
248 if (is_wb_truncate || dax_truncate)
249 fuse_set_nowrite(inode);
250
251 err = fuse_do_open(fm, get_node_id(inode), file, isdir);
252 if (!err)
253 fuse_finish_open(inode, file);
254
255 if (is_wb_truncate || dax_truncate)
256 fuse_release_nowrite(inode);
257 if (!err) {
258 struct fuse_file *ff = file->private_data;
259
260 if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC))
261 truncate_pagecache(inode, 0);
262 else if (!(ff->open_flags & FOPEN_KEEP_CACHE))
263 invalidate_inode_pages2(inode->i_mapping);
264 }
265 if (dax_truncate)
266 filemap_invalidate_unlock(inode->i_mapping);
267out_inode_unlock:
268 if (is_wb_truncate || dax_truncate)
269 inode_unlock(inode);
270
271 return err;
272}
273
274static void fuse_prepare_release(struct fuse_inode *fi, struct fuse_file *ff,
275 unsigned int flags, int opcode)
276{
277 struct fuse_conn *fc = ff->fm->fc;
278 struct fuse_release_args *ra = ff->release_args;
279
280 /* Inode is NULL on error path of fuse_create_open() */
281 if (likely(fi)) {
282 spin_lock(&fi->lock);
283 list_del(&ff->write_entry);
284 spin_unlock(&fi->lock);
285 }
286 spin_lock(&fc->lock);
287 if (!RB_EMPTY_NODE(&ff->polled_node))
288 rb_erase(&ff->polled_node, &fc->polled_files);
289 spin_unlock(&fc->lock);
290
291 wake_up_interruptible_all(&ff->poll_wait);
292
293 ra->inarg.fh = ff->fh;
294 ra->inarg.flags = flags;
295 ra->args.in_numargs = 1;
296 ra->args.in_args[0].size = sizeof(struct fuse_release_in);
297 ra->args.in_args[0].value = &ra->inarg;
298 ra->args.opcode = opcode;
299 ra->args.nodeid = ff->nodeid;
300 ra->args.force = true;
301 ra->args.nocreds = true;
302}
303
304void fuse_file_release(struct inode *inode, struct fuse_file *ff,
305 unsigned int open_flags, fl_owner_t id, bool isdir)
306{
307 struct fuse_inode *fi = get_fuse_inode(inode);
308 struct fuse_release_args *ra = ff->release_args;
309 int opcode = isdir ? FUSE_RELEASEDIR : FUSE_RELEASE;
310
311 fuse_prepare_release(fi, ff, open_flags, opcode);
312
313 if (ff->flock) {
314 ra->inarg.release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
315 ra->inarg.lock_owner = fuse_lock_owner_id(ff->fm->fc, id);
316 }
317 /* Hold inode until release is finished */
318 ra->inode = igrab(inode);
319
320 /*
321 * Normally this will send the RELEASE request, however if
322 * some asynchronous READ or WRITE requests are outstanding,
323 * the sending will be delayed.
324 *
325 * Make the release synchronous if this is a fuseblk mount,
326 * synchronous RELEASE is allowed (and desirable) in this case
327 * because the server can be trusted not to screw up.
328 */
329 fuse_file_put(ff, ff->fm->fc->destroy, isdir);
330}
331
332void fuse_release_common(struct file *file, bool isdir)
333{
334 fuse_file_release(file_inode(file), file->private_data, file->f_flags,
335 (fl_owner_t) file, isdir);
336}
337
338static int fuse_open(struct inode *inode, struct file *file)
339{
340 return fuse_open_common(inode, file, false);
341}
342
343static int fuse_release(struct inode *inode, struct file *file)
344{
345 struct fuse_conn *fc = get_fuse_conn(inode);
346
347 /*
348 * Dirty pages might remain despite write_inode_now() call from
349 * fuse_flush() due to writes racing with the close.
350 */
351 if (fc->writeback_cache)
352 write_inode_now(inode, 1);
353
354 fuse_release_common(file, false);
355
356 /* return value is ignored by VFS */
357 return 0;
358}
359
360void fuse_sync_release(struct fuse_inode *fi, struct fuse_file *ff,
361 unsigned int flags)
362{
363 WARN_ON(refcount_read(&ff->count) > 1);
364 fuse_prepare_release(fi, ff, flags, FUSE_RELEASE);
365 /*
366 * iput(NULL) is a no-op and since the refcount is 1 and everything's
367 * synchronous, we are fine with not doing igrab() here"
368 */
369 fuse_file_put(ff, true, false);
370}
371EXPORT_SYMBOL_GPL(fuse_sync_release);
372
373/*
374 * Scramble the ID space with XTEA, so that the value of the files_struct
375 * pointer is not exposed to userspace.
376 */
377u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
378{
379 u32 *k = fc->scramble_key;
380 u64 v = (unsigned long) id;
381 u32 v0 = v;
382 u32 v1 = v >> 32;
383 u32 sum = 0;
384 int i;
385
386 for (i = 0; i < 32; i++) {
387 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
388 sum += 0x9E3779B9;
389 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
390 }
391
392 return (u64) v0 + ((u64) v1 << 32);
393}
394
395struct fuse_writepage_args {
396 struct fuse_io_args ia;
397 struct rb_node writepages_entry;
398 struct list_head queue_entry;
399 struct fuse_writepage_args *next;
400 struct inode *inode;
401 struct fuse_sync_bucket *bucket;
402};
403
404static struct fuse_writepage_args *fuse_find_writeback(struct fuse_inode *fi,
405 pgoff_t idx_from, pgoff_t idx_to)
406{
407 struct rb_node *n;
408
409 n = fi->writepages.rb_node;
410
411 while (n) {
412 struct fuse_writepage_args *wpa;
413 pgoff_t curr_index;
414
415 wpa = rb_entry(n, struct fuse_writepage_args, writepages_entry);
416 WARN_ON(get_fuse_inode(wpa->inode) != fi);
417 curr_index = wpa->ia.write.in.offset >> PAGE_SHIFT;
418 if (idx_from >= curr_index + wpa->ia.ap.num_pages)
419 n = n->rb_right;
420 else if (idx_to < curr_index)
421 n = n->rb_left;
422 else
423 return wpa;
424 }
425 return NULL;
426}
427
428/*
429 * Check if any page in a range is under writeback
430 *
431 * This is currently done by walking the list of writepage requests
432 * for the inode, which can be pretty inefficient.
433 */
434static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
435 pgoff_t idx_to)
436{
437 struct fuse_inode *fi = get_fuse_inode(inode);
438 bool found;
439
440 spin_lock(&fi->lock);
441 found = fuse_find_writeback(fi, idx_from, idx_to);
442 spin_unlock(&fi->lock);
443
444 return found;
445}
446
447static inline bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
448{
449 return fuse_range_is_writeback(inode, index, index);
450}
451
452/*
453 * Wait for page writeback to be completed.
454 *
455 * Since fuse doesn't rely on the VM writeback tracking, this has to
456 * use some other means.
457 */
458static void fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
459{
460 struct fuse_inode *fi = get_fuse_inode(inode);
461
462 wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
463}
464
465/*
466 * Wait for all pending writepages on the inode to finish.
467 *
468 * This is currently done by blocking further writes with FUSE_NOWRITE
469 * and waiting for all sent writes to complete.
470 *
471 * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
472 * could conflict with truncation.
473 */
474static void fuse_sync_writes(struct inode *inode)
475{
476 fuse_set_nowrite(inode);
477 fuse_release_nowrite(inode);
478}
479
480static int fuse_flush(struct file *file, fl_owner_t id)
481{
482 struct inode *inode = file_inode(file);
483 struct fuse_mount *fm = get_fuse_mount(inode);
484 struct fuse_file *ff = file->private_data;
485 struct fuse_flush_in inarg;
486 FUSE_ARGS(args);
487 int err;
488
489 if (fuse_is_bad(inode))
490 return -EIO;
491
492 if (ff->open_flags & FOPEN_NOFLUSH && !fm->fc->writeback_cache)
493 return 0;
494
495 err = write_inode_now(inode, 1);
496 if (err)
497 return err;
498
499 inode_lock(inode);
500 fuse_sync_writes(inode);
501 inode_unlock(inode);
502
503 err = filemap_check_errors(file->f_mapping);
504 if (err)
505 return err;
506
507 err = 0;
508 if (fm->fc->no_flush)
509 goto inval_attr_out;
510
511 memset(&inarg, 0, sizeof(inarg));
512 inarg.fh = ff->fh;
513 inarg.lock_owner = fuse_lock_owner_id(fm->fc, id);
514 args.opcode = FUSE_FLUSH;
515 args.nodeid = get_node_id(inode);
516 args.in_numargs = 1;
517 args.in_args[0].size = sizeof(inarg);
518 args.in_args[0].value = &inarg;
519 args.force = true;
520
521 err = fuse_simple_request(fm, &args);
522 if (err == -ENOSYS) {
523 fm->fc->no_flush = 1;
524 err = 0;
525 }
526
527inval_attr_out:
528 /*
529 * In memory i_blocks is not maintained by fuse, if writeback cache is
530 * enabled, i_blocks from cached attr may not be accurate.
531 */
532 if (!err && fm->fc->writeback_cache)
533 fuse_invalidate_attr_mask(inode, STATX_BLOCKS);
534 return err;
535}
536
537int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
538 int datasync, int opcode)
539{
540 struct inode *inode = file->f_mapping->host;
541 struct fuse_mount *fm = get_fuse_mount(inode);
542 struct fuse_file *ff = file->private_data;
543 FUSE_ARGS(args);
544 struct fuse_fsync_in inarg;
545
546 memset(&inarg, 0, sizeof(inarg));
547 inarg.fh = ff->fh;
548 inarg.fsync_flags = datasync ? FUSE_FSYNC_FDATASYNC : 0;
549 args.opcode = opcode;
550 args.nodeid = get_node_id(inode);
551 args.in_numargs = 1;
552 args.in_args[0].size = sizeof(inarg);
553 args.in_args[0].value = &inarg;
554 return fuse_simple_request(fm, &args);
555}
556
557static int fuse_fsync(struct file *file, loff_t start, loff_t end,
558 int datasync)
559{
560 struct inode *inode = file->f_mapping->host;
561 struct fuse_conn *fc = get_fuse_conn(inode);
562 int err;
563
564 if (fuse_is_bad(inode))
565 return -EIO;
566
567 inode_lock(inode);
568
569 /*
570 * Start writeback against all dirty pages of the inode, then
571 * wait for all outstanding writes, before sending the FSYNC
572 * request.
573 */
574 err = file_write_and_wait_range(file, start, end);
575 if (err)
576 goto out;
577
578 fuse_sync_writes(inode);
579
580 /*
581 * Due to implementation of fuse writeback
582 * file_write_and_wait_range() does not catch errors.
583 * We have to do this directly after fuse_sync_writes()
584 */
585 err = file_check_and_advance_wb_err(file);
586 if (err)
587 goto out;
588
589 err = sync_inode_metadata(inode, 1);
590 if (err)
591 goto out;
592
593 if (fc->no_fsync)
594 goto out;
595
596 err = fuse_fsync_common(file, start, end, datasync, FUSE_FSYNC);
597 if (err == -ENOSYS) {
598 fc->no_fsync = 1;
599 err = 0;
600 }
601out:
602 inode_unlock(inode);
603
604 return err;
605}
606
607void fuse_read_args_fill(struct fuse_io_args *ia, struct file *file, loff_t pos,
608 size_t count, int opcode)
609{
610 struct fuse_file *ff = file->private_data;
611 struct fuse_args *args = &ia->ap.args;
612
613 ia->read.in.fh = ff->fh;
614 ia->read.in.offset = pos;
615 ia->read.in.size = count;
616 ia->read.in.flags = file->f_flags;
617 args->opcode = opcode;
618 args->nodeid = ff->nodeid;
619 args->in_numargs = 1;
620 args->in_args[0].size = sizeof(ia->read.in);
621 args->in_args[0].value = &ia->read.in;
622 args->out_argvar = true;
623 args->out_numargs = 1;
624 args->out_args[0].size = count;
625}
626
627static void fuse_release_user_pages(struct fuse_args_pages *ap,
628 bool should_dirty)
629{
630 unsigned int i;
631
632 for (i = 0; i < ap->num_pages; i++) {
633 if (should_dirty)
634 set_page_dirty_lock(ap->pages[i]);
635 put_page(ap->pages[i]);
636 }
637}
638
639static void fuse_io_release(struct kref *kref)
640{
641 kfree(container_of(kref, struct fuse_io_priv, refcnt));
642}
643
644static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
645{
646 if (io->err)
647 return io->err;
648
649 if (io->bytes >= 0 && io->write)
650 return -EIO;
651
652 return io->bytes < 0 ? io->size : io->bytes;
653}
654
655/**
656 * In case of short read, the caller sets 'pos' to the position of
657 * actual end of fuse request in IO request. Otherwise, if bytes_requested
658 * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
659 *
660 * An example:
661 * User requested DIO read of 64K. It was split into two 32K fuse requests,
662 * both submitted asynchronously. The first of them was ACKed by userspace as
663 * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
664 * second request was ACKed as short, e.g. only 1K was read, resulting in
665 * pos == 33K.
666 *
667 * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
668 * will be equal to the length of the longest contiguous fragment of
669 * transferred data starting from the beginning of IO request.
670 */
671static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
672{
673 int left;
674
675 spin_lock(&io->lock);
676 if (err)
677 io->err = io->err ? : err;
678 else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
679 io->bytes = pos;
680
681 left = --io->reqs;
682 if (!left && io->blocking)
683 complete(io->done);
684 spin_unlock(&io->lock);
685
686 if (!left && !io->blocking) {
687 ssize_t res = fuse_get_res_by_io(io);
688
689 if (res >= 0) {
690 struct inode *inode = file_inode(io->iocb->ki_filp);
691 struct fuse_conn *fc = get_fuse_conn(inode);
692 struct fuse_inode *fi = get_fuse_inode(inode);
693
694 spin_lock(&fi->lock);
695 fi->attr_version = atomic64_inc_return(&fc->attr_version);
696 spin_unlock(&fi->lock);
697 }
698
699 io->iocb->ki_complete(io->iocb, res);
700 }
701
702 kref_put(&io->refcnt, fuse_io_release);
703}
704
705static struct fuse_io_args *fuse_io_alloc(struct fuse_io_priv *io,
706 unsigned int npages)
707{
708 struct fuse_io_args *ia;
709
710 ia = kzalloc(sizeof(*ia), GFP_KERNEL);
711 if (ia) {
712 ia->io = io;
713 ia->ap.pages = fuse_pages_alloc(npages, GFP_KERNEL,
714 &ia->ap.descs);
715 if (!ia->ap.pages) {
716 kfree(ia);
717 ia = NULL;
718 }
719 }
720 return ia;
721}
722
723static void fuse_io_free(struct fuse_io_args *ia)
724{
725 kfree(ia->ap.pages);
726 kfree(ia);
727}
728
729static void fuse_aio_complete_req(struct fuse_mount *fm, struct fuse_args *args,
730 int err)
731{
732 struct fuse_io_args *ia = container_of(args, typeof(*ia), ap.args);
733 struct fuse_io_priv *io = ia->io;
734 ssize_t pos = -1;
735
736 fuse_release_user_pages(&ia->ap, io->should_dirty);
737
738 if (err) {
739 /* Nothing */
740 } else if (io->write) {
741 if (ia->write.out.size > ia->write.in.size) {
742 err = -EIO;
743 } else if (ia->write.in.size != ia->write.out.size) {
744 pos = ia->write.in.offset - io->offset +
745 ia->write.out.size;
746 }
747 } else {
748 u32 outsize = args->out_args[0].size;
749
750 if (ia->read.in.size != outsize)
751 pos = ia->read.in.offset - io->offset + outsize;
752 }
753
754 fuse_aio_complete(io, err, pos);
755 fuse_io_free(ia);
756}
757
758static ssize_t fuse_async_req_send(struct fuse_mount *fm,
759 struct fuse_io_args *ia, size_t num_bytes)
760{
761 ssize_t err;
762 struct fuse_io_priv *io = ia->io;
763
764 spin_lock(&io->lock);
765 kref_get(&io->refcnt);
766 io->size += num_bytes;
767 io->reqs++;
768 spin_unlock(&io->lock);
769
770 ia->ap.args.end = fuse_aio_complete_req;
771 ia->ap.args.may_block = io->should_dirty;
772 err = fuse_simple_background(fm, &ia->ap.args, GFP_KERNEL);
773 if (err)
774 fuse_aio_complete_req(fm, &ia->ap.args, err);
775
776 return num_bytes;
777}
778
779static ssize_t fuse_send_read(struct fuse_io_args *ia, loff_t pos, size_t count,
780 fl_owner_t owner)
781{
782 struct file *file = ia->io->iocb->ki_filp;
783 struct fuse_file *ff = file->private_data;
784 struct fuse_mount *fm = ff->fm;
785
786 fuse_read_args_fill(ia, file, pos, count, FUSE_READ);
787 if (owner != NULL) {
788 ia->read.in.read_flags |= FUSE_READ_LOCKOWNER;
789 ia->read.in.lock_owner = fuse_lock_owner_id(fm->fc, owner);
790 }
791
792 if (ia->io->async)
793 return fuse_async_req_send(fm, ia, count);
794
795 return fuse_simple_request(fm, &ia->ap.args);
796}
797
798static void fuse_read_update_size(struct inode *inode, loff_t size,
799 u64 attr_ver)
800{
801 struct fuse_conn *fc = get_fuse_conn(inode);
802 struct fuse_inode *fi = get_fuse_inode(inode);
803
804 spin_lock(&fi->lock);
805 if (attr_ver >= fi->attr_version && size < inode->i_size &&
806 !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
807 fi->attr_version = atomic64_inc_return(&fc->attr_version);
808 i_size_write(inode, size);
809 }
810 spin_unlock(&fi->lock);
811}
812
813static void fuse_short_read(struct inode *inode, u64 attr_ver, size_t num_read,
814 struct fuse_args_pages *ap)
815{
816 struct fuse_conn *fc = get_fuse_conn(inode);
817
818 /*
819 * If writeback_cache is enabled, a short read means there's a hole in
820 * the file. Some data after the hole is in page cache, but has not
821 * reached the client fs yet. So the hole is not present there.
822 */
823 if (!fc->writeback_cache) {
824 loff_t pos = page_offset(ap->pages[0]) + num_read;
825 fuse_read_update_size(inode, pos, attr_ver);
826 }
827}
828
829static int fuse_do_readpage(struct file *file, struct page *page)
830{
831 struct inode *inode = page->mapping->host;
832 struct fuse_mount *fm = get_fuse_mount(inode);
833 loff_t pos = page_offset(page);
834 struct fuse_page_desc desc = { .length = PAGE_SIZE };
835 struct fuse_io_args ia = {
836 .ap.args.page_zeroing = true,
837 .ap.args.out_pages = true,
838 .ap.num_pages = 1,
839 .ap.pages = &page,
840 .ap.descs = &desc,
841 };
842 ssize_t res;
843 u64 attr_ver;
844
845 /*
846 * Page writeback can extend beyond the lifetime of the
847 * page-cache page, so make sure we read a properly synced
848 * page.
849 */
850 fuse_wait_on_page_writeback(inode, page->index);
851
852 attr_ver = fuse_get_attr_version(fm->fc);
853
854 /* Don't overflow end offset */
855 if (pos + (desc.length - 1) == LLONG_MAX)
856 desc.length--;
857
858 fuse_read_args_fill(&ia, file, pos, desc.length, FUSE_READ);
859 res = fuse_simple_request(fm, &ia.ap.args);
860 if (res < 0)
861 return res;
862 /*
863 * Short read means EOF. If file size is larger, truncate it
864 */
865 if (res < desc.length)
866 fuse_short_read(inode, attr_ver, res, &ia.ap);
867
868 SetPageUptodate(page);
869
870 return 0;
871}
872
873static int fuse_read_folio(struct file *file, struct folio *folio)
874{
875 struct page *page = &folio->page;
876 struct inode *inode = page->mapping->host;
877 int err;
878
879 err = -EIO;
880 if (fuse_is_bad(inode))
881 goto out;
882
883 err = fuse_do_readpage(file, page);
884 fuse_invalidate_atime(inode);
885 out:
886 unlock_page(page);
887 return err;
888}
889
890static void fuse_readpages_end(struct fuse_mount *fm, struct fuse_args *args,
891 int err)
892{
893 int i;
894 struct fuse_io_args *ia = container_of(args, typeof(*ia), ap.args);
895 struct fuse_args_pages *ap = &ia->ap;
896 size_t count = ia->read.in.size;
897 size_t num_read = args->out_args[0].size;
898 struct address_space *mapping = NULL;
899
900 for (i = 0; mapping == NULL && i < ap->num_pages; i++)
901 mapping = ap->pages[i]->mapping;
902
903 if (mapping) {
904 struct inode *inode = mapping->host;
905
906 /*
907 * Short read means EOF. If file size is larger, truncate it
908 */
909 if (!err && num_read < count)
910 fuse_short_read(inode, ia->read.attr_ver, num_read, ap);
911
912 fuse_invalidate_atime(inode);
913 }
914
915 for (i = 0; i < ap->num_pages; i++) {
916 struct page *page = ap->pages[i];
917
918 if (!err)
919 SetPageUptodate(page);
920 else
921 SetPageError(page);
922 unlock_page(page);
923 put_page(page);
924 }
925 if (ia->ff)
926 fuse_file_put(ia->ff, false, false);
927
928 fuse_io_free(ia);
929}
930
931static void fuse_send_readpages(struct fuse_io_args *ia, struct file *file)
932{
933 struct fuse_file *ff = file->private_data;
934 struct fuse_mount *fm = ff->fm;
935 struct fuse_args_pages *ap = &ia->ap;
936 loff_t pos = page_offset(ap->pages[0]);
937 size_t count = ap->num_pages << PAGE_SHIFT;
938 ssize_t res;
939 int err;
940
941 ap->args.out_pages = true;
942 ap->args.page_zeroing = true;
943 ap->args.page_replace = true;
944
945 /* Don't overflow end offset */
946 if (pos + (count - 1) == LLONG_MAX) {
947 count--;
948 ap->descs[ap->num_pages - 1].length--;
949 }
950 WARN_ON((loff_t) (pos + count) < 0);
951
952 fuse_read_args_fill(ia, file, pos, count, FUSE_READ);
953 ia->read.attr_ver = fuse_get_attr_version(fm->fc);
954 if (fm->fc->async_read) {
955 ia->ff = fuse_file_get(ff);
956 ap->args.end = fuse_readpages_end;
957 err = fuse_simple_background(fm, &ap->args, GFP_KERNEL);
958 if (!err)
959 return;
960 } else {
961 res = fuse_simple_request(fm, &ap->args);
962 err = res < 0 ? res : 0;
963 }
964 fuse_readpages_end(fm, &ap->args, err);
965}
966
967static void fuse_readahead(struct readahead_control *rac)
968{
969 struct inode *inode = rac->mapping->host;
970 struct fuse_conn *fc = get_fuse_conn(inode);
971 unsigned int i, max_pages, nr_pages = 0;
972
973 if (fuse_is_bad(inode))
974 return;
975
976 max_pages = min_t(unsigned int, fc->max_pages,
977 fc->max_read / PAGE_SIZE);
978
979 for (;;) {
980 struct fuse_io_args *ia;
981 struct fuse_args_pages *ap;
982
983 if (fc->num_background >= fc->congestion_threshold &&
984 rac->ra->async_size >= readahead_count(rac))
985 /*
986 * Congested and only async pages left, so skip the
987 * rest.
988 */
989 break;
990
991 nr_pages = readahead_count(rac) - nr_pages;
992 if (nr_pages > max_pages)
993 nr_pages = max_pages;
994 if (nr_pages == 0)
995 break;
996 ia = fuse_io_alloc(NULL, nr_pages);
997 if (!ia)
998 return;
999 ap = &ia->ap;
1000 nr_pages = __readahead_batch(rac, ap->pages, nr_pages);
1001 for (i = 0; i < nr_pages; i++) {
1002 fuse_wait_on_page_writeback(inode,
1003 readahead_index(rac) + i);
1004 ap->descs[i].length = PAGE_SIZE;
1005 }
1006 ap->num_pages = nr_pages;
1007 fuse_send_readpages(ia, rac->file);
1008 }
1009}
1010
1011static ssize_t fuse_cache_read_iter(struct kiocb *iocb, struct iov_iter *to)
1012{
1013 struct inode *inode = iocb->ki_filp->f_mapping->host;
1014 struct fuse_conn *fc = get_fuse_conn(inode);
1015
1016 /*
1017 * In auto invalidate mode, always update attributes on read.
1018 * Otherwise, only update if we attempt to read past EOF (to ensure
1019 * i_size is up to date).
1020 */
1021 if (fc->auto_inval_data ||
1022 (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
1023 int err;
1024 err = fuse_update_attributes(inode, iocb->ki_filp, STATX_SIZE);
1025 if (err)
1026 return err;
1027 }
1028
1029 return generic_file_read_iter(iocb, to);
1030}
1031
1032static void fuse_write_args_fill(struct fuse_io_args *ia, struct fuse_file *ff,
1033 loff_t pos, size_t count)
1034{
1035 struct fuse_args *args = &ia->ap.args;
1036
1037 ia->write.in.fh = ff->fh;
1038 ia->write.in.offset = pos;
1039 ia->write.in.size = count;
1040 args->opcode = FUSE_WRITE;
1041 args->nodeid = ff->nodeid;
1042 args->in_numargs = 2;
1043 if (ff->fm->fc->minor < 9)
1044 args->in_args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
1045 else
1046 args->in_args[0].size = sizeof(ia->write.in);
1047 args->in_args[0].value = &ia->write.in;
1048 args->in_args[1].size = count;
1049 args->out_numargs = 1;
1050 args->out_args[0].size = sizeof(ia->write.out);
1051 args->out_args[0].value = &ia->write.out;
1052}
1053
1054static unsigned int fuse_write_flags(struct kiocb *iocb)
1055{
1056 unsigned int flags = iocb->ki_filp->f_flags;
1057
1058 if (iocb_is_dsync(iocb))
1059 flags |= O_DSYNC;
1060 if (iocb->ki_flags & IOCB_SYNC)
1061 flags |= O_SYNC;
1062
1063 return flags;
1064}
1065
1066static ssize_t fuse_send_write(struct fuse_io_args *ia, loff_t pos,
1067 size_t count, fl_owner_t owner)
1068{
1069 struct kiocb *iocb = ia->io->iocb;
1070 struct file *file = iocb->ki_filp;
1071 struct fuse_file *ff = file->private_data;
1072 struct fuse_mount *fm = ff->fm;
1073 struct fuse_write_in *inarg = &ia->write.in;
1074 ssize_t err;
1075
1076 fuse_write_args_fill(ia, ff, pos, count);
1077 inarg->flags = fuse_write_flags(iocb);
1078 if (owner != NULL) {
1079 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
1080 inarg->lock_owner = fuse_lock_owner_id(fm->fc, owner);
1081 }
1082
1083 if (ia->io->async)
1084 return fuse_async_req_send(fm, ia, count);
1085
1086 err = fuse_simple_request(fm, &ia->ap.args);
1087 if (!err && ia->write.out.size > count)
1088 err = -EIO;
1089
1090 return err ?: ia->write.out.size;
1091}
1092
1093bool fuse_write_update_attr(struct inode *inode, loff_t pos, ssize_t written)
1094{
1095 struct fuse_conn *fc = get_fuse_conn(inode);
1096 struct fuse_inode *fi = get_fuse_inode(inode);
1097 bool ret = false;
1098
1099 spin_lock(&fi->lock);
1100 fi->attr_version = atomic64_inc_return(&fc->attr_version);
1101 if (written > 0 && pos > inode->i_size) {
1102 i_size_write(inode, pos);
1103 ret = true;
1104 }
1105 spin_unlock(&fi->lock);
1106
1107 fuse_invalidate_attr_mask(inode, FUSE_STATX_MODSIZE);
1108
1109 return ret;
1110}
1111
1112static ssize_t fuse_send_write_pages(struct fuse_io_args *ia,
1113 struct kiocb *iocb, struct inode *inode,
1114 loff_t pos, size_t count)
1115{
1116 struct fuse_args_pages *ap = &ia->ap;
1117 struct file *file = iocb->ki_filp;
1118 struct fuse_file *ff = file->private_data;
1119 struct fuse_mount *fm = ff->fm;
1120 unsigned int offset, i;
1121 bool short_write;
1122 int err;
1123
1124 for (i = 0; i < ap->num_pages; i++)
1125 fuse_wait_on_page_writeback(inode, ap->pages[i]->index);
1126
1127 fuse_write_args_fill(ia, ff, pos, count);
1128 ia->write.in.flags = fuse_write_flags(iocb);
1129 if (fm->fc->handle_killpriv_v2 && !capable(CAP_FSETID))
1130 ia->write.in.write_flags |= FUSE_WRITE_KILL_SUIDGID;
1131
1132 err = fuse_simple_request(fm, &ap->args);
1133 if (!err && ia->write.out.size > count)
1134 err = -EIO;
1135
1136 short_write = ia->write.out.size < count;
1137 offset = ap->descs[0].offset;
1138 count = ia->write.out.size;
1139 for (i = 0; i < ap->num_pages; i++) {
1140 struct page *page = ap->pages[i];
1141
1142 if (err) {
1143 ClearPageUptodate(page);
1144 } else {
1145 if (count >= PAGE_SIZE - offset)
1146 count -= PAGE_SIZE - offset;
1147 else {
1148 if (short_write)
1149 ClearPageUptodate(page);
1150 count = 0;
1151 }
1152 offset = 0;
1153 }
1154 if (ia->write.page_locked && (i == ap->num_pages - 1))
1155 unlock_page(page);
1156 put_page(page);
1157 }
1158
1159 return err;
1160}
1161
1162static ssize_t fuse_fill_write_pages(struct fuse_io_args *ia,
1163 struct address_space *mapping,
1164 struct iov_iter *ii, loff_t pos,
1165 unsigned int max_pages)
1166{
1167 struct fuse_args_pages *ap = &ia->ap;
1168 struct fuse_conn *fc = get_fuse_conn(mapping->host);
1169 unsigned offset = pos & (PAGE_SIZE - 1);
1170 size_t count = 0;
1171 int err;
1172
1173 ap->args.in_pages = true;
1174 ap->descs[0].offset = offset;
1175
1176 do {
1177 size_t tmp;
1178 struct page *page;
1179 pgoff_t index = pos >> PAGE_SHIFT;
1180 size_t bytes = min_t(size_t, PAGE_SIZE - offset,
1181 iov_iter_count(ii));
1182
1183 bytes = min_t(size_t, bytes, fc->max_write - count);
1184
1185 again:
1186 err = -EFAULT;
1187 if (fault_in_iov_iter_readable(ii, bytes))
1188 break;
1189
1190 err = -ENOMEM;
1191 page = grab_cache_page_write_begin(mapping, index);
1192 if (!page)
1193 break;
1194
1195 if (mapping_writably_mapped(mapping))
1196 flush_dcache_page(page);
1197
1198 tmp = copy_page_from_iter_atomic(page, offset, bytes, ii);
1199 flush_dcache_page(page);
1200
1201 if (!tmp) {
1202 unlock_page(page);
1203 put_page(page);
1204 goto again;
1205 }
1206
1207 err = 0;
1208 ap->pages[ap->num_pages] = page;
1209 ap->descs[ap->num_pages].length = tmp;
1210 ap->num_pages++;
1211
1212 count += tmp;
1213 pos += tmp;
1214 offset += tmp;
1215 if (offset == PAGE_SIZE)
1216 offset = 0;
1217
1218 /* If we copied full page, mark it uptodate */
1219 if (tmp == PAGE_SIZE)
1220 SetPageUptodate(page);
1221
1222 if (PageUptodate(page)) {
1223 unlock_page(page);
1224 } else {
1225 ia->write.page_locked = true;
1226 break;
1227 }
1228 if (!fc->big_writes)
1229 break;
1230 } while (iov_iter_count(ii) && count < fc->max_write &&
1231 ap->num_pages < max_pages && offset == 0);
1232
1233 return count > 0 ? count : err;
1234}
1235
1236static inline unsigned int fuse_wr_pages(loff_t pos, size_t len,
1237 unsigned int max_pages)
1238{
1239 return min_t(unsigned int,
1240 ((pos + len - 1) >> PAGE_SHIFT) -
1241 (pos >> PAGE_SHIFT) + 1,
1242 max_pages);
1243}
1244
1245static ssize_t fuse_perform_write(struct kiocb *iocb,
1246 struct address_space *mapping,
1247 struct iov_iter *ii, loff_t pos)
1248{
1249 struct inode *inode = mapping->host;
1250 struct fuse_conn *fc = get_fuse_conn(inode);
1251 struct fuse_inode *fi = get_fuse_inode(inode);
1252 int err = 0;
1253 ssize_t res = 0;
1254
1255 if (inode->i_size < pos + iov_iter_count(ii))
1256 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1257
1258 do {
1259 ssize_t count;
1260 struct fuse_io_args ia = {};
1261 struct fuse_args_pages *ap = &ia.ap;
1262 unsigned int nr_pages = fuse_wr_pages(pos, iov_iter_count(ii),
1263 fc->max_pages);
1264
1265 ap->pages = fuse_pages_alloc(nr_pages, GFP_KERNEL, &ap->descs);
1266 if (!ap->pages) {
1267 err = -ENOMEM;
1268 break;
1269 }
1270
1271 count = fuse_fill_write_pages(&ia, mapping, ii, pos, nr_pages);
1272 if (count <= 0) {
1273 err = count;
1274 } else {
1275 err = fuse_send_write_pages(&ia, iocb, inode,
1276 pos, count);
1277 if (!err) {
1278 size_t num_written = ia.write.out.size;
1279
1280 res += num_written;
1281 pos += num_written;
1282
1283 /* break out of the loop on short write */
1284 if (num_written != count)
1285 err = -EIO;
1286 }
1287 }
1288 kfree(ap->pages);
1289 } while (!err && iov_iter_count(ii));
1290
1291 fuse_write_update_attr(inode, pos, res);
1292 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1293
1294 return res > 0 ? res : err;
1295}
1296
1297static ssize_t fuse_cache_write_iter(struct kiocb *iocb, struct iov_iter *from)
1298{
1299 struct file *file = iocb->ki_filp;
1300 struct address_space *mapping = file->f_mapping;
1301 ssize_t written = 0;
1302 ssize_t written_buffered = 0;
1303 struct inode *inode = mapping->host;
1304 ssize_t err;
1305 struct fuse_conn *fc = get_fuse_conn(inode);
1306 loff_t endbyte = 0;
1307
1308 if (fc->writeback_cache) {
1309 /* Update size (EOF optimization) and mode (SUID clearing) */
1310 err = fuse_update_attributes(mapping->host, file,
1311 STATX_SIZE | STATX_MODE);
1312 if (err)
1313 return err;
1314
1315 if (fc->handle_killpriv_v2 &&
1316 setattr_should_drop_suidgid(&init_user_ns, file_inode(file))) {
1317 goto writethrough;
1318 }
1319
1320 return generic_file_write_iter(iocb, from);
1321 }
1322
1323writethrough:
1324 inode_lock(inode);
1325
1326 /* We can write back this queue in page reclaim */
1327 current->backing_dev_info = inode_to_bdi(inode);
1328
1329 err = generic_write_checks(iocb, from);
1330 if (err <= 0)
1331 goto out;
1332
1333 err = file_remove_privs(file);
1334 if (err)
1335 goto out;
1336
1337 err = file_update_time(file);
1338 if (err)
1339 goto out;
1340
1341 if (iocb->ki_flags & IOCB_DIRECT) {
1342 loff_t pos = iocb->ki_pos;
1343 written = generic_file_direct_write(iocb, from);
1344 if (written < 0 || !iov_iter_count(from))
1345 goto out;
1346
1347 pos += written;
1348
1349 written_buffered = fuse_perform_write(iocb, mapping, from, pos);
1350 if (written_buffered < 0) {
1351 err = written_buffered;
1352 goto out;
1353 }
1354 endbyte = pos + written_buffered - 1;
1355
1356 err = filemap_write_and_wait_range(file->f_mapping, pos,
1357 endbyte);
1358 if (err)
1359 goto out;
1360
1361 invalidate_mapping_pages(file->f_mapping,
1362 pos >> PAGE_SHIFT,
1363 endbyte >> PAGE_SHIFT);
1364
1365 written += written_buffered;
1366 iocb->ki_pos = pos + written_buffered;
1367 } else {
1368 written = fuse_perform_write(iocb, mapping, from, iocb->ki_pos);
1369 if (written >= 0)
1370 iocb->ki_pos += written;
1371 }
1372out:
1373 current->backing_dev_info = NULL;
1374 inode_unlock(inode);
1375 if (written > 0)
1376 written = generic_write_sync(iocb, written);
1377
1378 return written ? written : err;
1379}
1380
1381static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1382{
1383 return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1384}
1385
1386static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1387 size_t max_size)
1388{
1389 return min(iov_iter_single_seg_count(ii), max_size);
1390}
1391
1392static int fuse_get_user_pages(struct fuse_args_pages *ap, struct iov_iter *ii,
1393 size_t *nbytesp, int write,
1394 unsigned int max_pages)
1395{
1396 size_t nbytes = 0; /* # bytes already packed in req */
1397 ssize_t ret = 0;
1398
1399 /* Special case for kernel I/O: can copy directly into the buffer */
1400 if (iov_iter_is_kvec(ii)) {
1401 unsigned long user_addr = fuse_get_user_addr(ii);
1402 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1403
1404 if (write)
1405 ap->args.in_args[1].value = (void *) user_addr;
1406 else
1407 ap->args.out_args[0].value = (void *) user_addr;
1408
1409 iov_iter_advance(ii, frag_size);
1410 *nbytesp = frag_size;
1411 return 0;
1412 }
1413
1414 while (nbytes < *nbytesp && ap->num_pages < max_pages) {
1415 unsigned npages;
1416 size_t start;
1417 ret = iov_iter_get_pages2(ii, &ap->pages[ap->num_pages],
1418 *nbytesp - nbytes,
1419 max_pages - ap->num_pages,
1420 &start);
1421 if (ret < 0)
1422 break;
1423
1424 nbytes += ret;
1425
1426 ret += start;
1427 npages = DIV_ROUND_UP(ret, PAGE_SIZE);
1428
1429 ap->descs[ap->num_pages].offset = start;
1430 fuse_page_descs_length_init(ap->descs, ap->num_pages, npages);
1431
1432 ap->num_pages += npages;
1433 ap->descs[ap->num_pages - 1].length -=
1434 (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1435 }
1436
1437 ap->args.user_pages = true;
1438 if (write)
1439 ap->args.in_pages = true;
1440 else
1441 ap->args.out_pages = true;
1442
1443 *nbytesp = nbytes;
1444
1445 return ret < 0 ? ret : 0;
1446}
1447
1448ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1449 loff_t *ppos, int flags)
1450{
1451 int write = flags & FUSE_DIO_WRITE;
1452 int cuse = flags & FUSE_DIO_CUSE;
1453 struct file *file = io->iocb->ki_filp;
1454 struct inode *inode = file->f_mapping->host;
1455 struct fuse_file *ff = file->private_data;
1456 struct fuse_conn *fc = ff->fm->fc;
1457 size_t nmax = write ? fc->max_write : fc->max_read;
1458 loff_t pos = *ppos;
1459 size_t count = iov_iter_count(iter);
1460 pgoff_t idx_from = pos >> PAGE_SHIFT;
1461 pgoff_t idx_to = (pos + count - 1) >> PAGE_SHIFT;
1462 ssize_t res = 0;
1463 int err = 0;
1464 struct fuse_io_args *ia;
1465 unsigned int max_pages;
1466
1467 max_pages = iov_iter_npages(iter, fc->max_pages);
1468 ia = fuse_io_alloc(io, max_pages);
1469 if (!ia)
1470 return -ENOMEM;
1471
1472 if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1473 if (!write)
1474 inode_lock(inode);
1475 fuse_sync_writes(inode);
1476 if (!write)
1477 inode_unlock(inode);
1478 }
1479
1480 io->should_dirty = !write && user_backed_iter(iter);
1481 while (count) {
1482 ssize_t nres;
1483 fl_owner_t owner = current->files;
1484 size_t nbytes = min(count, nmax);
1485
1486 err = fuse_get_user_pages(&ia->ap, iter, &nbytes, write,
1487 max_pages);
1488 if (err && !nbytes)
1489 break;
1490
1491 if (write) {
1492 if (!capable(CAP_FSETID))
1493 ia->write.in.write_flags |= FUSE_WRITE_KILL_SUIDGID;
1494
1495 nres = fuse_send_write(ia, pos, nbytes, owner);
1496 } else {
1497 nres = fuse_send_read(ia, pos, nbytes, owner);
1498 }
1499
1500 if (!io->async || nres < 0) {
1501 fuse_release_user_pages(&ia->ap, io->should_dirty);
1502 fuse_io_free(ia);
1503 }
1504 ia = NULL;
1505 if (nres < 0) {
1506 iov_iter_revert(iter, nbytes);
1507 err = nres;
1508 break;
1509 }
1510 WARN_ON(nres > nbytes);
1511
1512 count -= nres;
1513 res += nres;
1514 pos += nres;
1515 if (nres != nbytes) {
1516 iov_iter_revert(iter, nbytes - nres);
1517 break;
1518 }
1519 if (count) {
1520 max_pages = iov_iter_npages(iter, fc->max_pages);
1521 ia = fuse_io_alloc(io, max_pages);
1522 if (!ia)
1523 break;
1524 }
1525 }
1526 if (ia)
1527 fuse_io_free(ia);
1528 if (res > 0)
1529 *ppos = pos;
1530
1531 return res > 0 ? res : err;
1532}
1533EXPORT_SYMBOL_GPL(fuse_direct_io);
1534
1535static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1536 struct iov_iter *iter,
1537 loff_t *ppos)
1538{
1539 ssize_t res;
1540 struct inode *inode = file_inode(io->iocb->ki_filp);
1541
1542 res = fuse_direct_io(io, iter, ppos, 0);
1543
1544 fuse_invalidate_atime(inode);
1545
1546 return res;
1547}
1548
1549static ssize_t fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter);
1550
1551static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1552{
1553 ssize_t res;
1554
1555 if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1556 res = fuse_direct_IO(iocb, to);
1557 } else {
1558 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1559
1560 res = __fuse_direct_read(&io, to, &iocb->ki_pos);
1561 }
1562
1563 return res;
1564}
1565
1566static bool fuse_direct_write_extending_i_size(struct kiocb *iocb,
1567 struct iov_iter *iter)
1568{
1569 struct inode *inode = file_inode(iocb->ki_filp);
1570
1571 return iocb->ki_pos + iov_iter_count(iter) > i_size_read(inode);
1572}
1573
1574static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1575{
1576 struct inode *inode = file_inode(iocb->ki_filp);
1577 struct file *file = iocb->ki_filp;
1578 struct fuse_file *ff = file->private_data;
1579 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1580 ssize_t res;
1581 bool exclusive_lock =
1582 !(ff->open_flags & FOPEN_PARALLEL_DIRECT_WRITES) ||
1583 iocb->ki_flags & IOCB_APPEND ||
1584 fuse_direct_write_extending_i_size(iocb, from);
1585
1586 /*
1587 * Take exclusive lock if
1588 * - Parallel direct writes are disabled - a user space decision
1589 * - Parallel direct writes are enabled and i_size is being extended.
1590 * This might not be needed at all, but needs further investigation.
1591 */
1592 if (exclusive_lock)
1593 inode_lock(inode);
1594 else {
1595 inode_lock_shared(inode);
1596
1597 /* A race with truncate might have come up as the decision for
1598 * the lock type was done without holding the lock, check again.
1599 */
1600 if (fuse_direct_write_extending_i_size(iocb, from)) {
1601 inode_unlock_shared(inode);
1602 inode_lock(inode);
1603 exclusive_lock = true;
1604 }
1605 }
1606
1607 res = generic_write_checks(iocb, from);
1608 if (res > 0) {
1609 if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1610 res = fuse_direct_IO(iocb, from);
1611 } else {
1612 res = fuse_direct_io(&io, from, &iocb->ki_pos,
1613 FUSE_DIO_WRITE);
1614 fuse_write_update_attr(inode, iocb->ki_pos, res);
1615 }
1616 }
1617 if (exclusive_lock)
1618 inode_unlock(inode);
1619 else
1620 inode_unlock_shared(inode);
1621
1622 return res;
1623}
1624
1625static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
1626{
1627 struct file *file = iocb->ki_filp;
1628 struct fuse_file *ff = file->private_data;
1629 struct inode *inode = file_inode(file);
1630
1631 if (fuse_is_bad(inode))
1632 return -EIO;
1633
1634 if (FUSE_IS_DAX(inode))
1635 return fuse_dax_read_iter(iocb, to);
1636
1637 if (!(ff->open_flags & FOPEN_DIRECT_IO))
1638 return fuse_cache_read_iter(iocb, to);
1639 else
1640 return fuse_direct_read_iter(iocb, to);
1641}
1642
1643static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1644{
1645 struct file *file = iocb->ki_filp;
1646 struct fuse_file *ff = file->private_data;
1647 struct inode *inode = file_inode(file);
1648
1649 if (fuse_is_bad(inode))
1650 return -EIO;
1651
1652 if (FUSE_IS_DAX(inode))
1653 return fuse_dax_write_iter(iocb, from);
1654
1655 if (!(ff->open_flags & FOPEN_DIRECT_IO))
1656 return fuse_cache_write_iter(iocb, from);
1657 else
1658 return fuse_direct_write_iter(iocb, from);
1659}
1660
1661static void fuse_writepage_free(struct fuse_writepage_args *wpa)
1662{
1663 struct fuse_args_pages *ap = &wpa->ia.ap;
1664 int i;
1665
1666 if (wpa->bucket)
1667 fuse_sync_bucket_dec(wpa->bucket);
1668
1669 for (i = 0; i < ap->num_pages; i++)
1670 __free_page(ap->pages[i]);
1671
1672 if (wpa->ia.ff)
1673 fuse_file_put(wpa->ia.ff, false, false);
1674
1675 kfree(ap->pages);
1676 kfree(wpa);
1677}
1678
1679static void fuse_writepage_finish(struct fuse_mount *fm,
1680 struct fuse_writepage_args *wpa)
1681{
1682 struct fuse_args_pages *ap = &wpa->ia.ap;
1683 struct inode *inode = wpa->inode;
1684 struct fuse_inode *fi = get_fuse_inode(inode);
1685 struct backing_dev_info *bdi = inode_to_bdi(inode);
1686 int i;
1687
1688 for (i = 0; i < ap->num_pages; i++) {
1689 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1690 dec_node_page_state(ap->pages[i], NR_WRITEBACK_TEMP);
1691 wb_writeout_inc(&bdi->wb);
1692 }
1693 wake_up(&fi->page_waitq);
1694}
1695
1696/* Called under fi->lock, may release and reacquire it */
1697static void fuse_send_writepage(struct fuse_mount *fm,
1698 struct fuse_writepage_args *wpa, loff_t size)
1699__releases(fi->lock)
1700__acquires(fi->lock)
1701{
1702 struct fuse_writepage_args *aux, *next;
1703 struct fuse_inode *fi = get_fuse_inode(wpa->inode);
1704 struct fuse_write_in *inarg = &wpa->ia.write.in;
1705 struct fuse_args *args = &wpa->ia.ap.args;
1706 __u64 data_size = wpa->ia.ap.num_pages * PAGE_SIZE;
1707 int err;
1708
1709 fi->writectr++;
1710 if (inarg->offset + data_size <= size) {
1711 inarg->size = data_size;
1712 } else if (inarg->offset < size) {
1713 inarg->size = size - inarg->offset;
1714 } else {
1715 /* Got truncated off completely */
1716 goto out_free;
1717 }
1718
1719 args->in_args[1].size = inarg->size;
1720 args->force = true;
1721 args->nocreds = true;
1722
1723 err = fuse_simple_background(fm, args, GFP_ATOMIC);
1724 if (err == -ENOMEM) {
1725 spin_unlock(&fi->lock);
1726 err = fuse_simple_background(fm, args, GFP_NOFS | __GFP_NOFAIL);
1727 spin_lock(&fi->lock);
1728 }
1729
1730 /* Fails on broken connection only */
1731 if (unlikely(err))
1732 goto out_free;
1733
1734 return;
1735
1736 out_free:
1737 fi->writectr--;
1738 rb_erase(&wpa->writepages_entry, &fi->writepages);
1739 fuse_writepage_finish(fm, wpa);
1740 spin_unlock(&fi->lock);
1741
1742 /* After fuse_writepage_finish() aux request list is private */
1743 for (aux = wpa->next; aux; aux = next) {
1744 next = aux->next;
1745 aux->next = NULL;
1746 fuse_writepage_free(aux);
1747 }
1748
1749 fuse_writepage_free(wpa);
1750 spin_lock(&fi->lock);
1751}
1752
1753/*
1754 * If fi->writectr is positive (no truncate or fsync going on) send
1755 * all queued writepage requests.
1756 *
1757 * Called with fi->lock
1758 */
1759void fuse_flush_writepages(struct inode *inode)
1760__releases(fi->lock)
1761__acquires(fi->lock)
1762{
1763 struct fuse_mount *fm = get_fuse_mount(inode);
1764 struct fuse_inode *fi = get_fuse_inode(inode);
1765 loff_t crop = i_size_read(inode);
1766 struct fuse_writepage_args *wpa;
1767
1768 while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1769 wpa = list_entry(fi->queued_writes.next,
1770 struct fuse_writepage_args, queue_entry);
1771 list_del_init(&wpa->queue_entry);
1772 fuse_send_writepage(fm, wpa, crop);
1773 }
1774}
1775
1776static struct fuse_writepage_args *fuse_insert_writeback(struct rb_root *root,
1777 struct fuse_writepage_args *wpa)
1778{
1779 pgoff_t idx_from = wpa->ia.write.in.offset >> PAGE_SHIFT;
1780 pgoff_t idx_to = idx_from + wpa->ia.ap.num_pages - 1;
1781 struct rb_node **p = &root->rb_node;
1782 struct rb_node *parent = NULL;
1783
1784 WARN_ON(!wpa->ia.ap.num_pages);
1785 while (*p) {
1786 struct fuse_writepage_args *curr;
1787 pgoff_t curr_index;
1788
1789 parent = *p;
1790 curr = rb_entry(parent, struct fuse_writepage_args,
1791 writepages_entry);
1792 WARN_ON(curr->inode != wpa->inode);
1793 curr_index = curr->ia.write.in.offset >> PAGE_SHIFT;
1794
1795 if (idx_from >= curr_index + curr->ia.ap.num_pages)
1796 p = &(*p)->rb_right;
1797 else if (idx_to < curr_index)
1798 p = &(*p)->rb_left;
1799 else
1800 return curr;
1801 }
1802
1803 rb_link_node(&wpa->writepages_entry, parent, p);
1804 rb_insert_color(&wpa->writepages_entry, root);
1805 return NULL;
1806}
1807
1808static void tree_insert(struct rb_root *root, struct fuse_writepage_args *wpa)
1809{
1810 WARN_ON(fuse_insert_writeback(root, wpa));
1811}
1812
1813static void fuse_writepage_end(struct fuse_mount *fm, struct fuse_args *args,
1814 int error)
1815{
1816 struct fuse_writepage_args *wpa =
1817 container_of(args, typeof(*wpa), ia.ap.args);
1818 struct inode *inode = wpa->inode;
1819 struct fuse_inode *fi = get_fuse_inode(inode);
1820 struct fuse_conn *fc = get_fuse_conn(inode);
1821
1822 mapping_set_error(inode->i_mapping, error);
1823 /*
1824 * A writeback finished and this might have updated mtime/ctime on
1825 * server making local mtime/ctime stale. Hence invalidate attrs.
1826 * Do this only if writeback_cache is not enabled. If writeback_cache
1827 * is enabled, we trust local ctime/mtime.
1828 */
1829 if (!fc->writeback_cache)
1830 fuse_invalidate_attr_mask(inode, FUSE_STATX_MODIFY);
1831 spin_lock(&fi->lock);
1832 rb_erase(&wpa->writepages_entry, &fi->writepages);
1833 while (wpa->next) {
1834 struct fuse_mount *fm = get_fuse_mount(inode);
1835 struct fuse_write_in *inarg = &wpa->ia.write.in;
1836 struct fuse_writepage_args *next = wpa->next;
1837
1838 wpa->next = next->next;
1839 next->next = NULL;
1840 next->ia.ff = fuse_file_get(wpa->ia.ff);
1841 tree_insert(&fi->writepages, next);
1842
1843 /*
1844 * Skip fuse_flush_writepages() to make it easy to crop requests
1845 * based on primary request size.
1846 *
1847 * 1st case (trivial): there are no concurrent activities using
1848 * fuse_set/release_nowrite. Then we're on safe side because
1849 * fuse_flush_writepages() would call fuse_send_writepage()
1850 * anyway.
1851 *
1852 * 2nd case: someone called fuse_set_nowrite and it is waiting
1853 * now for completion of all in-flight requests. This happens
1854 * rarely and no more than once per page, so this should be
1855 * okay.
1856 *
1857 * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1858 * of fuse_set_nowrite..fuse_release_nowrite section. The fact
1859 * that fuse_set_nowrite returned implies that all in-flight
1860 * requests were completed along with all of their secondary
1861 * requests. Further primary requests are blocked by negative
1862 * writectr. Hence there cannot be any in-flight requests and
1863 * no invocations of fuse_writepage_end() while we're in
1864 * fuse_set_nowrite..fuse_release_nowrite section.
1865 */
1866 fuse_send_writepage(fm, next, inarg->offset + inarg->size);
1867 }
1868 fi->writectr--;
1869 fuse_writepage_finish(fm, wpa);
1870 spin_unlock(&fi->lock);
1871 fuse_writepage_free(wpa);
1872}
1873
1874static struct fuse_file *__fuse_write_file_get(struct fuse_inode *fi)
1875{
1876 struct fuse_file *ff;
1877
1878 spin_lock(&fi->lock);
1879 ff = list_first_entry_or_null(&fi->write_files, struct fuse_file,
1880 write_entry);
1881 if (ff)
1882 fuse_file_get(ff);
1883 spin_unlock(&fi->lock);
1884
1885 return ff;
1886}
1887
1888static struct fuse_file *fuse_write_file_get(struct fuse_inode *fi)
1889{
1890 struct fuse_file *ff = __fuse_write_file_get(fi);
1891 WARN_ON(!ff);
1892 return ff;
1893}
1894
1895int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1896{
1897 struct fuse_inode *fi = get_fuse_inode(inode);
1898 struct fuse_file *ff;
1899 int err;
1900
1901 /*
1902 * Inode is always written before the last reference is dropped and
1903 * hence this should not be reached from reclaim.
1904 *
1905 * Writing back the inode from reclaim can deadlock if the request
1906 * processing itself needs an allocation. Allocations triggering
1907 * reclaim while serving a request can't be prevented, because it can
1908 * involve any number of unrelated userspace processes.
1909 */
1910 WARN_ON(wbc->for_reclaim);
1911
1912 ff = __fuse_write_file_get(fi);
1913 err = fuse_flush_times(inode, ff);
1914 if (ff)
1915 fuse_file_put(ff, false, false);
1916
1917 return err;
1918}
1919
1920static struct fuse_writepage_args *fuse_writepage_args_alloc(void)
1921{
1922 struct fuse_writepage_args *wpa;
1923 struct fuse_args_pages *ap;
1924
1925 wpa = kzalloc(sizeof(*wpa), GFP_NOFS);
1926 if (wpa) {
1927 ap = &wpa->ia.ap;
1928 ap->num_pages = 0;
1929 ap->pages = fuse_pages_alloc(1, GFP_NOFS, &ap->descs);
1930 if (!ap->pages) {
1931 kfree(wpa);
1932 wpa = NULL;
1933 }
1934 }
1935 return wpa;
1936
1937}
1938
1939static void fuse_writepage_add_to_bucket(struct fuse_conn *fc,
1940 struct fuse_writepage_args *wpa)
1941{
1942 if (!fc->sync_fs)
1943 return;
1944
1945 rcu_read_lock();
1946 /* Prevent resurrection of dead bucket in unlikely race with syncfs */
1947 do {
1948 wpa->bucket = rcu_dereference(fc->curr_bucket);
1949 } while (unlikely(!atomic_inc_not_zero(&wpa->bucket->count)));
1950 rcu_read_unlock();
1951}
1952
1953static int fuse_writepage_locked(struct page *page)
1954{
1955 struct address_space *mapping = page->mapping;
1956 struct inode *inode = mapping->host;
1957 struct fuse_conn *fc = get_fuse_conn(inode);
1958 struct fuse_inode *fi = get_fuse_inode(inode);
1959 struct fuse_writepage_args *wpa;
1960 struct fuse_args_pages *ap;
1961 struct page *tmp_page;
1962 int error = -ENOMEM;
1963
1964 set_page_writeback(page);
1965
1966 wpa = fuse_writepage_args_alloc();
1967 if (!wpa)
1968 goto err;
1969 ap = &wpa->ia.ap;
1970
1971 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1972 if (!tmp_page)
1973 goto err_free;
1974
1975 error = -EIO;
1976 wpa->ia.ff = fuse_write_file_get(fi);
1977 if (!wpa->ia.ff)
1978 goto err_nofile;
1979
1980 fuse_writepage_add_to_bucket(fc, wpa);
1981 fuse_write_args_fill(&wpa->ia, wpa->ia.ff, page_offset(page), 0);
1982
1983 copy_highpage(tmp_page, page);
1984 wpa->ia.write.in.write_flags |= FUSE_WRITE_CACHE;
1985 wpa->next = NULL;
1986 ap->args.in_pages = true;
1987 ap->num_pages = 1;
1988 ap->pages[0] = tmp_page;
1989 ap->descs[0].offset = 0;
1990 ap->descs[0].length = PAGE_SIZE;
1991 ap->args.end = fuse_writepage_end;
1992 wpa->inode = inode;
1993
1994 inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1995 inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1996
1997 spin_lock(&fi->lock);
1998 tree_insert(&fi->writepages, wpa);
1999 list_add_tail(&wpa->queue_entry, &fi->queued_writes);
2000 fuse_flush_writepages(inode);
2001 spin_unlock(&fi->lock);
2002
2003 end_page_writeback(page);
2004
2005 return 0;
2006
2007err_nofile:
2008 __free_page(tmp_page);
2009err_free:
2010 kfree(wpa);
2011err:
2012 mapping_set_error(page->mapping, error);
2013 end_page_writeback(page);
2014 return error;
2015}
2016
2017static int fuse_writepage(struct page *page, struct writeback_control *wbc)
2018{
2019 struct fuse_conn *fc = get_fuse_conn(page->mapping->host);
2020 int err;
2021
2022 if (fuse_page_is_writeback(page->mapping->host, page->index)) {
2023 /*
2024 * ->writepages() should be called for sync() and friends. We
2025 * should only get here on direct reclaim and then we are
2026 * allowed to skip a page which is already in flight
2027 */
2028 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
2029
2030 redirty_page_for_writepage(wbc, page);
2031 unlock_page(page);
2032 return 0;
2033 }
2034
2035 if (wbc->sync_mode == WB_SYNC_NONE &&
2036 fc->num_background >= fc->congestion_threshold)
2037 return AOP_WRITEPAGE_ACTIVATE;
2038
2039 err = fuse_writepage_locked(page);
2040 unlock_page(page);
2041
2042 return err;
2043}
2044
2045struct fuse_fill_wb_data {
2046 struct fuse_writepage_args *wpa;
2047 struct fuse_file *ff;
2048 struct inode *inode;
2049 struct page **orig_pages;
2050 unsigned int max_pages;
2051};
2052
2053static bool fuse_pages_realloc(struct fuse_fill_wb_data *data)
2054{
2055 struct fuse_args_pages *ap = &data->wpa->ia.ap;
2056 struct fuse_conn *fc = get_fuse_conn(data->inode);
2057 struct page **pages;
2058 struct fuse_page_desc *descs;
2059 unsigned int npages = min_t(unsigned int,
2060 max_t(unsigned int, data->max_pages * 2,
2061 FUSE_DEFAULT_MAX_PAGES_PER_REQ),
2062 fc->max_pages);
2063 WARN_ON(npages <= data->max_pages);
2064
2065 pages = fuse_pages_alloc(npages, GFP_NOFS, &descs);
2066 if (!pages)
2067 return false;
2068
2069 memcpy(pages, ap->pages, sizeof(struct page *) * ap->num_pages);
2070 memcpy(descs, ap->descs, sizeof(struct fuse_page_desc) * ap->num_pages);
2071 kfree(ap->pages);
2072 ap->pages = pages;
2073 ap->descs = descs;
2074 data->max_pages = npages;
2075
2076 return true;
2077}
2078
2079static void fuse_writepages_send(struct fuse_fill_wb_data *data)
2080{
2081 struct fuse_writepage_args *wpa = data->wpa;
2082 struct inode *inode = data->inode;
2083 struct fuse_inode *fi = get_fuse_inode(inode);
2084 int num_pages = wpa->ia.ap.num_pages;
2085 int i;
2086
2087 wpa->ia.ff = fuse_file_get(data->ff);
2088 spin_lock(&fi->lock);
2089 list_add_tail(&wpa->queue_entry, &fi->queued_writes);
2090 fuse_flush_writepages(inode);
2091 spin_unlock(&fi->lock);
2092
2093 for (i = 0; i < num_pages; i++)
2094 end_page_writeback(data->orig_pages[i]);
2095}
2096
2097/*
2098 * Check under fi->lock if the page is under writeback, and insert it onto the
2099 * rb_tree if not. Otherwise iterate auxiliary write requests, to see if there's
2100 * one already added for a page at this offset. If there's none, then insert
2101 * this new request onto the auxiliary list, otherwise reuse the existing one by
2102 * swapping the new temp page with the old one.
2103 */
2104static bool fuse_writepage_add(struct fuse_writepage_args *new_wpa,
2105 struct page *page)
2106{
2107 struct fuse_inode *fi = get_fuse_inode(new_wpa->inode);
2108 struct fuse_writepage_args *tmp;
2109 struct fuse_writepage_args *old_wpa;
2110 struct fuse_args_pages *new_ap = &new_wpa->ia.ap;
2111
2112 WARN_ON(new_ap->num_pages != 0);
2113 new_ap->num_pages = 1;
2114
2115 spin_lock(&fi->lock);
2116 old_wpa = fuse_insert_writeback(&fi->writepages, new_wpa);
2117 if (!old_wpa) {
2118 spin_unlock(&fi->lock);
2119 return true;
2120 }
2121
2122 for (tmp = old_wpa->next; tmp; tmp = tmp->next) {
2123 pgoff_t curr_index;
2124
2125 WARN_ON(tmp->inode != new_wpa->inode);
2126 curr_index = tmp->ia.write.in.offset >> PAGE_SHIFT;
2127 if (curr_index == page->index) {
2128 WARN_ON(tmp->ia.ap.num_pages != 1);
2129 swap(tmp->ia.ap.pages[0], new_ap->pages[0]);
2130 break;
2131 }
2132 }
2133
2134 if (!tmp) {
2135 new_wpa->next = old_wpa->next;
2136 old_wpa->next = new_wpa;
2137 }
2138
2139 spin_unlock(&fi->lock);
2140
2141 if (tmp) {
2142 struct backing_dev_info *bdi = inode_to_bdi(new_wpa->inode);
2143
2144 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
2145 dec_node_page_state(new_ap->pages[0], NR_WRITEBACK_TEMP);
2146 wb_writeout_inc(&bdi->wb);
2147 fuse_writepage_free(new_wpa);
2148 }
2149
2150 return false;
2151}
2152
2153static bool fuse_writepage_need_send(struct fuse_conn *fc, struct page *page,
2154 struct fuse_args_pages *ap,
2155 struct fuse_fill_wb_data *data)
2156{
2157 WARN_ON(!ap->num_pages);
2158
2159 /*
2160 * Being under writeback is unlikely but possible. For example direct
2161 * read to an mmaped fuse file will set the page dirty twice; once when
2162 * the pages are faulted with get_user_pages(), and then after the read
2163 * completed.
2164 */
2165 if (fuse_page_is_writeback(data->inode, page->index))
2166 return true;
2167
2168 /* Reached max pages */
2169 if (ap->num_pages == fc->max_pages)
2170 return true;
2171
2172 /* Reached max write bytes */
2173 if ((ap->num_pages + 1) * PAGE_SIZE > fc->max_write)
2174 return true;
2175
2176 /* Discontinuity */
2177 if (data->orig_pages[ap->num_pages - 1]->index + 1 != page->index)
2178 return true;
2179
2180 /* Need to grow the pages array? If so, did the expansion fail? */
2181 if (ap->num_pages == data->max_pages && !fuse_pages_realloc(data))
2182 return true;
2183
2184 return false;
2185}
2186
2187static int fuse_writepages_fill(struct page *page,
2188 struct writeback_control *wbc, void *_data)
2189{
2190 struct fuse_fill_wb_data *data = _data;
2191 struct fuse_writepage_args *wpa = data->wpa;
2192 struct fuse_args_pages *ap = &wpa->ia.ap;
2193 struct inode *inode = data->inode;
2194 struct fuse_inode *fi = get_fuse_inode(inode);
2195 struct fuse_conn *fc = get_fuse_conn(inode);
2196 struct page *tmp_page;
2197 int err;
2198
2199 if (!data->ff) {
2200 err = -EIO;
2201 data->ff = fuse_write_file_get(fi);
2202 if (!data->ff)
2203 goto out_unlock;
2204 }
2205
2206 if (wpa && fuse_writepage_need_send(fc, page, ap, data)) {
2207 fuse_writepages_send(data);
2208 data->wpa = NULL;
2209 }
2210
2211 err = -ENOMEM;
2212 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2213 if (!tmp_page)
2214 goto out_unlock;
2215
2216 /*
2217 * The page must not be redirtied until the writeout is completed
2218 * (i.e. userspace has sent a reply to the write request). Otherwise
2219 * there could be more than one temporary page instance for each real
2220 * page.
2221 *
2222 * This is ensured by holding the page lock in page_mkwrite() while
2223 * checking fuse_page_is_writeback(). We already hold the page lock
2224 * since clear_page_dirty_for_io() and keep it held until we add the
2225 * request to the fi->writepages list and increment ap->num_pages.
2226 * After this fuse_page_is_writeback() will indicate that the page is
2227 * under writeback, so we can release the page lock.
2228 */
2229 if (data->wpa == NULL) {
2230 err = -ENOMEM;
2231 wpa = fuse_writepage_args_alloc();
2232 if (!wpa) {
2233 __free_page(tmp_page);
2234 goto out_unlock;
2235 }
2236 fuse_writepage_add_to_bucket(fc, wpa);
2237
2238 data->max_pages = 1;
2239
2240 ap = &wpa->ia.ap;
2241 fuse_write_args_fill(&wpa->ia, data->ff, page_offset(page), 0);
2242 wpa->ia.write.in.write_flags |= FUSE_WRITE_CACHE;
2243 wpa->next = NULL;
2244 ap->args.in_pages = true;
2245 ap->args.end = fuse_writepage_end;
2246 ap->num_pages = 0;
2247 wpa->inode = inode;
2248 }
2249 set_page_writeback(page);
2250
2251 copy_highpage(tmp_page, page);
2252 ap->pages[ap->num_pages] = tmp_page;
2253 ap->descs[ap->num_pages].offset = 0;
2254 ap->descs[ap->num_pages].length = PAGE_SIZE;
2255 data->orig_pages[ap->num_pages] = page;
2256
2257 inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
2258 inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
2259
2260 err = 0;
2261 if (data->wpa) {
2262 /*
2263 * Protected by fi->lock against concurrent access by
2264 * fuse_page_is_writeback().
2265 */
2266 spin_lock(&fi->lock);
2267 ap->num_pages++;
2268 spin_unlock(&fi->lock);
2269 } else if (fuse_writepage_add(wpa, page)) {
2270 data->wpa = wpa;
2271 } else {
2272 end_page_writeback(page);
2273 }
2274out_unlock:
2275 unlock_page(page);
2276
2277 return err;
2278}
2279
2280static int fuse_writepages(struct address_space *mapping,
2281 struct writeback_control *wbc)
2282{
2283 struct inode *inode = mapping->host;
2284 struct fuse_conn *fc = get_fuse_conn(inode);
2285 struct fuse_fill_wb_data data;
2286 int err;
2287
2288 err = -EIO;
2289 if (fuse_is_bad(inode))
2290 goto out;
2291
2292 if (wbc->sync_mode == WB_SYNC_NONE &&
2293 fc->num_background >= fc->congestion_threshold)
2294 return 0;
2295
2296 data.inode = inode;
2297 data.wpa = NULL;
2298 data.ff = NULL;
2299
2300 err = -ENOMEM;
2301 data.orig_pages = kcalloc(fc->max_pages,
2302 sizeof(struct page *),
2303 GFP_NOFS);
2304 if (!data.orig_pages)
2305 goto out;
2306
2307 err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
2308 if (data.wpa) {
2309 WARN_ON(!data.wpa->ia.ap.num_pages);
2310 fuse_writepages_send(&data);
2311 }
2312 if (data.ff)
2313 fuse_file_put(data.ff, false, false);
2314
2315 kfree(data.orig_pages);
2316out:
2317 return err;
2318}
2319
2320/*
2321 * It's worthy to make sure that space is reserved on disk for the write,
2322 * but how to implement it without killing performance need more thinking.
2323 */
2324static int fuse_write_begin(struct file *file, struct address_space *mapping,
2325 loff_t pos, unsigned len, struct page **pagep, void **fsdata)
2326{
2327 pgoff_t index = pos >> PAGE_SHIFT;
2328 struct fuse_conn *fc = get_fuse_conn(file_inode(file));
2329 struct page *page;
2330 loff_t fsize;
2331 int err = -ENOMEM;
2332
2333 WARN_ON(!fc->writeback_cache);
2334
2335 page = grab_cache_page_write_begin(mapping, index);
2336 if (!page)
2337 goto error;
2338
2339 fuse_wait_on_page_writeback(mapping->host, page->index);
2340
2341 if (PageUptodate(page) || len == PAGE_SIZE)
2342 goto success;
2343 /*
2344 * Check if the start this page comes after the end of file, in which
2345 * case the readpage can be optimized away.
2346 */
2347 fsize = i_size_read(mapping->host);
2348 if (fsize <= (pos & PAGE_MASK)) {
2349 size_t off = pos & ~PAGE_MASK;
2350 if (off)
2351 zero_user_segment(page, 0, off);
2352 goto success;
2353 }
2354 err = fuse_do_readpage(file, page);
2355 if (err)
2356 goto cleanup;
2357success:
2358 *pagep = page;
2359 return 0;
2360
2361cleanup:
2362 unlock_page(page);
2363 put_page(page);
2364error:
2365 return err;
2366}
2367
2368static int fuse_write_end(struct file *file, struct address_space *mapping,
2369 loff_t pos, unsigned len, unsigned copied,
2370 struct page *page, void *fsdata)
2371{
2372 struct inode *inode = page->mapping->host;
2373
2374 /* Haven't copied anything? Skip zeroing, size extending, dirtying. */
2375 if (!copied)
2376 goto unlock;
2377
2378 pos += copied;
2379 if (!PageUptodate(page)) {
2380 /* Zero any unwritten bytes at the end of the page */
2381 size_t endoff = pos & ~PAGE_MASK;
2382 if (endoff)
2383 zero_user_segment(page, endoff, PAGE_SIZE);
2384 SetPageUptodate(page);
2385 }
2386
2387 if (pos > inode->i_size)
2388 i_size_write(inode, pos);
2389
2390 set_page_dirty(page);
2391
2392unlock:
2393 unlock_page(page);
2394 put_page(page);
2395
2396 return copied;
2397}
2398
2399static int fuse_launder_folio(struct folio *folio)
2400{
2401 int err = 0;
2402 if (folio_clear_dirty_for_io(folio)) {
2403 struct inode *inode = folio->mapping->host;
2404
2405 /* Serialize with pending writeback for the same page */
2406 fuse_wait_on_page_writeback(inode, folio->index);
2407 err = fuse_writepage_locked(&folio->page);
2408 if (!err)
2409 fuse_wait_on_page_writeback(inode, folio->index);
2410 }
2411 return err;
2412}
2413
2414/*
2415 * Write back dirty data/metadata now (there may not be any suitable
2416 * open files later for data)
2417 */
2418static void fuse_vma_close(struct vm_area_struct *vma)
2419{
2420 int err;
2421
2422 err = write_inode_now(vma->vm_file->f_mapping->host, 1);
2423 mapping_set_error(vma->vm_file->f_mapping, err);
2424}
2425
2426/*
2427 * Wait for writeback against this page to complete before allowing it
2428 * to be marked dirty again, and hence written back again, possibly
2429 * before the previous writepage completed.
2430 *
2431 * Block here, instead of in ->writepage(), so that the userspace fs
2432 * can only block processes actually operating on the filesystem.
2433 *
2434 * Otherwise unprivileged userspace fs would be able to block
2435 * unrelated:
2436 *
2437 * - page migration
2438 * - sync(2)
2439 * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2440 */
2441static vm_fault_t fuse_page_mkwrite(struct vm_fault *vmf)
2442{
2443 struct page *page = vmf->page;
2444 struct inode *inode = file_inode(vmf->vma->vm_file);
2445
2446 file_update_time(vmf->vma->vm_file);
2447 lock_page(page);
2448 if (page->mapping != inode->i_mapping) {
2449 unlock_page(page);
2450 return VM_FAULT_NOPAGE;
2451 }
2452
2453 fuse_wait_on_page_writeback(inode, page->index);
2454 return VM_FAULT_LOCKED;
2455}
2456
2457static const struct vm_operations_struct fuse_file_vm_ops = {
2458 .close = fuse_vma_close,
2459 .fault = filemap_fault,
2460 .map_pages = filemap_map_pages,
2461 .page_mkwrite = fuse_page_mkwrite,
2462};
2463
2464static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2465{
2466 struct fuse_file *ff = file->private_data;
2467
2468 /* DAX mmap is superior to direct_io mmap */
2469 if (FUSE_IS_DAX(file_inode(file)))
2470 return fuse_dax_mmap(file, vma);
2471
2472 if (ff->open_flags & FOPEN_DIRECT_IO) {
2473 /* Can't provide the coherency needed for MAP_SHARED */
2474 if (vma->vm_flags & VM_MAYSHARE)
2475 return -ENODEV;
2476
2477 invalidate_inode_pages2(file->f_mapping);
2478
2479 return generic_file_mmap(file, vma);
2480 }
2481
2482 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2483 fuse_link_write_file(file);
2484
2485 file_accessed(file);
2486 vma->vm_ops = &fuse_file_vm_ops;
2487 return 0;
2488}
2489
2490static int convert_fuse_file_lock(struct fuse_conn *fc,
2491 const struct fuse_file_lock *ffl,
2492 struct file_lock *fl)
2493{
2494 switch (ffl->type) {
2495 case F_UNLCK:
2496 break;
2497
2498 case F_RDLCK:
2499 case F_WRLCK:
2500 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2501 ffl->end < ffl->start)
2502 return -EIO;
2503
2504 fl->fl_start = ffl->start;
2505 fl->fl_end = ffl->end;
2506
2507 /*
2508 * Convert pid into init's pid namespace. The locks API will
2509 * translate it into the caller's pid namespace.
2510 */
2511 rcu_read_lock();
2512 fl->fl_pid = pid_nr_ns(find_pid_ns(ffl->pid, fc->pid_ns), &init_pid_ns);
2513 rcu_read_unlock();
2514 break;
2515
2516 default:
2517 return -EIO;
2518 }
2519 fl->fl_type = ffl->type;
2520 return 0;
2521}
2522
2523static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2524 const struct file_lock *fl, int opcode, pid_t pid,
2525 int flock, struct fuse_lk_in *inarg)
2526{
2527 struct inode *inode = file_inode(file);
2528 struct fuse_conn *fc = get_fuse_conn(inode);
2529 struct fuse_file *ff = file->private_data;
2530
2531 memset(inarg, 0, sizeof(*inarg));
2532 inarg->fh = ff->fh;
2533 inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2534 inarg->lk.start = fl->fl_start;
2535 inarg->lk.end = fl->fl_end;
2536 inarg->lk.type = fl->fl_type;
2537 inarg->lk.pid = pid;
2538 if (flock)
2539 inarg->lk_flags |= FUSE_LK_FLOCK;
2540 args->opcode = opcode;
2541 args->nodeid = get_node_id(inode);
2542 args->in_numargs = 1;
2543 args->in_args[0].size = sizeof(*inarg);
2544 args->in_args[0].value = inarg;
2545}
2546
2547static int fuse_getlk(struct file *file, struct file_lock *fl)
2548{
2549 struct inode *inode = file_inode(file);
2550 struct fuse_mount *fm = get_fuse_mount(inode);
2551 FUSE_ARGS(args);
2552 struct fuse_lk_in inarg;
2553 struct fuse_lk_out outarg;
2554 int err;
2555
2556 fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2557 args.out_numargs = 1;
2558 args.out_args[0].size = sizeof(outarg);
2559 args.out_args[0].value = &outarg;
2560 err = fuse_simple_request(fm, &args);
2561 if (!err)
2562 err = convert_fuse_file_lock(fm->fc, &outarg.lk, fl);
2563
2564 return err;
2565}
2566
2567static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2568{
2569 struct inode *inode = file_inode(file);
2570 struct fuse_mount *fm = get_fuse_mount(inode);
2571 FUSE_ARGS(args);
2572 struct fuse_lk_in inarg;
2573 int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2574 struct pid *pid = fl->fl_type != F_UNLCK ? task_tgid(current) : NULL;
2575 pid_t pid_nr = pid_nr_ns(pid, fm->fc->pid_ns);
2576 int err;
2577
2578 if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2579 /* NLM needs asynchronous locks, which we don't support yet */
2580 return -ENOLCK;
2581 }
2582
2583 /* Unlock on close is handled by the flush method */
2584 if ((fl->fl_flags & FL_CLOSE_POSIX) == FL_CLOSE_POSIX)
2585 return 0;
2586
2587 fuse_lk_fill(&args, file, fl, opcode, pid_nr, flock, &inarg);
2588 err = fuse_simple_request(fm, &args);
2589
2590 /* locking is restartable */
2591 if (err == -EINTR)
2592 err = -ERESTARTSYS;
2593
2594 return err;
2595}
2596
2597static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2598{
2599 struct inode *inode = file_inode(file);
2600 struct fuse_conn *fc = get_fuse_conn(inode);
2601 int err;
2602
2603 if (cmd == F_CANCELLK) {
2604 err = 0;
2605 } else if (cmd == F_GETLK) {
2606 if (fc->no_lock) {
2607 posix_test_lock(file, fl);
2608 err = 0;
2609 } else
2610 err = fuse_getlk(file, fl);
2611 } else {
2612 if (fc->no_lock)
2613 err = posix_lock_file(file, fl, NULL);
2614 else
2615 err = fuse_setlk(file, fl, 0);
2616 }
2617 return err;
2618}
2619
2620static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2621{
2622 struct inode *inode = file_inode(file);
2623 struct fuse_conn *fc = get_fuse_conn(inode);
2624 int err;
2625
2626 if (fc->no_flock) {
2627 err = locks_lock_file_wait(file, fl);
2628 } else {
2629 struct fuse_file *ff = file->private_data;
2630
2631 /* emulate flock with POSIX locks */
2632 ff->flock = true;
2633 err = fuse_setlk(file, fl, 1);
2634 }
2635
2636 return err;
2637}
2638
2639static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2640{
2641 struct inode *inode = mapping->host;
2642 struct fuse_mount *fm = get_fuse_mount(inode);
2643 FUSE_ARGS(args);
2644 struct fuse_bmap_in inarg;
2645 struct fuse_bmap_out outarg;
2646 int err;
2647
2648 if (!inode->i_sb->s_bdev || fm->fc->no_bmap)
2649 return 0;
2650
2651 memset(&inarg, 0, sizeof(inarg));
2652 inarg.block = block;
2653 inarg.blocksize = inode->i_sb->s_blocksize;
2654 args.opcode = FUSE_BMAP;
2655 args.nodeid = get_node_id(inode);
2656 args.in_numargs = 1;
2657 args.in_args[0].size = sizeof(inarg);
2658 args.in_args[0].value = &inarg;
2659 args.out_numargs = 1;
2660 args.out_args[0].size = sizeof(outarg);
2661 args.out_args[0].value = &outarg;
2662 err = fuse_simple_request(fm, &args);
2663 if (err == -ENOSYS)
2664 fm->fc->no_bmap = 1;
2665
2666 return err ? 0 : outarg.block;
2667}
2668
2669static loff_t fuse_lseek(struct file *file, loff_t offset, int whence)
2670{
2671 struct inode *inode = file->f_mapping->host;
2672 struct fuse_mount *fm = get_fuse_mount(inode);
2673 struct fuse_file *ff = file->private_data;
2674 FUSE_ARGS(args);
2675 struct fuse_lseek_in inarg = {
2676 .fh = ff->fh,
2677 .offset = offset,
2678 .whence = whence
2679 };
2680 struct fuse_lseek_out outarg;
2681 int err;
2682
2683 if (fm->fc->no_lseek)
2684 goto fallback;
2685
2686 args.opcode = FUSE_LSEEK;
2687 args.nodeid = ff->nodeid;
2688 args.in_numargs = 1;
2689 args.in_args[0].size = sizeof(inarg);
2690 args.in_args[0].value = &inarg;
2691 args.out_numargs = 1;
2692 args.out_args[0].size = sizeof(outarg);
2693 args.out_args[0].value = &outarg;
2694 err = fuse_simple_request(fm, &args);
2695 if (err) {
2696 if (err == -ENOSYS) {
2697 fm->fc->no_lseek = 1;
2698 goto fallback;
2699 }
2700 return err;
2701 }
2702
2703 return vfs_setpos(file, outarg.offset, inode->i_sb->s_maxbytes);
2704
2705fallback:
2706 err = fuse_update_attributes(inode, file, STATX_SIZE);
2707 if (!err)
2708 return generic_file_llseek(file, offset, whence);
2709 else
2710 return err;
2711}
2712
2713static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2714{
2715 loff_t retval;
2716 struct inode *inode = file_inode(file);
2717
2718 switch (whence) {
2719 case SEEK_SET:
2720 case SEEK_CUR:
2721 /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2722 retval = generic_file_llseek(file, offset, whence);
2723 break;
2724 case SEEK_END:
2725 inode_lock(inode);
2726 retval = fuse_update_attributes(inode, file, STATX_SIZE);
2727 if (!retval)
2728 retval = generic_file_llseek(file, offset, whence);
2729 inode_unlock(inode);
2730 break;
2731 case SEEK_HOLE:
2732 case SEEK_DATA:
2733 inode_lock(inode);
2734 retval = fuse_lseek(file, offset, whence);
2735 inode_unlock(inode);
2736 break;
2737 default:
2738 retval = -EINVAL;
2739 }
2740
2741 return retval;
2742}
2743
2744/*
2745 * All files which have been polled are linked to RB tree
2746 * fuse_conn->polled_files which is indexed by kh. Walk the tree and
2747 * find the matching one.
2748 */
2749static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2750 struct rb_node **parent_out)
2751{
2752 struct rb_node **link = &fc->polled_files.rb_node;
2753 struct rb_node *last = NULL;
2754
2755 while (*link) {
2756 struct fuse_file *ff;
2757
2758 last = *link;
2759 ff = rb_entry(last, struct fuse_file, polled_node);
2760
2761 if (kh < ff->kh)
2762 link = &last->rb_left;
2763 else if (kh > ff->kh)
2764 link = &last->rb_right;
2765 else
2766 return link;
2767 }
2768
2769 if (parent_out)
2770 *parent_out = last;
2771 return link;
2772}
2773
2774/*
2775 * The file is about to be polled. Make sure it's on the polled_files
2776 * RB tree. Note that files once added to the polled_files tree are
2777 * not removed before the file is released. This is because a file
2778 * polled once is likely to be polled again.
2779 */
2780static void fuse_register_polled_file(struct fuse_conn *fc,
2781 struct fuse_file *ff)
2782{
2783 spin_lock(&fc->lock);
2784 if (RB_EMPTY_NODE(&ff->polled_node)) {
2785 struct rb_node **link, *parent;
2786
2787 link = fuse_find_polled_node(fc, ff->kh, &parent);
2788 BUG_ON(*link);
2789 rb_link_node(&ff->polled_node, parent, link);
2790 rb_insert_color(&ff->polled_node, &fc->polled_files);
2791 }
2792 spin_unlock(&fc->lock);
2793}
2794
2795__poll_t fuse_file_poll(struct file *file, poll_table *wait)
2796{
2797 struct fuse_file *ff = file->private_data;
2798 struct fuse_mount *fm = ff->fm;
2799 struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2800 struct fuse_poll_out outarg;
2801 FUSE_ARGS(args);
2802 int err;
2803
2804 if (fm->fc->no_poll)
2805 return DEFAULT_POLLMASK;
2806
2807 poll_wait(file, &ff->poll_wait, wait);
2808 inarg.events = mangle_poll(poll_requested_events(wait));
2809
2810 /*
2811 * Ask for notification iff there's someone waiting for it.
2812 * The client may ignore the flag and always notify.
2813 */
2814 if (waitqueue_active(&ff->poll_wait)) {
2815 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2816 fuse_register_polled_file(fm->fc, ff);
2817 }
2818
2819 args.opcode = FUSE_POLL;
2820 args.nodeid = ff->nodeid;
2821 args.in_numargs = 1;
2822 args.in_args[0].size = sizeof(inarg);
2823 args.in_args[0].value = &inarg;
2824 args.out_numargs = 1;
2825 args.out_args[0].size = sizeof(outarg);
2826 args.out_args[0].value = &outarg;
2827 err = fuse_simple_request(fm, &args);
2828
2829 if (!err)
2830 return demangle_poll(outarg.revents);
2831 if (err == -ENOSYS) {
2832 fm->fc->no_poll = 1;
2833 return DEFAULT_POLLMASK;
2834 }
2835 return EPOLLERR;
2836}
2837EXPORT_SYMBOL_GPL(fuse_file_poll);
2838
2839/*
2840 * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2841 * wakes up the poll waiters.
2842 */
2843int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2844 struct fuse_notify_poll_wakeup_out *outarg)
2845{
2846 u64 kh = outarg->kh;
2847 struct rb_node **link;
2848
2849 spin_lock(&fc->lock);
2850
2851 link = fuse_find_polled_node(fc, kh, NULL);
2852 if (*link) {
2853 struct fuse_file *ff;
2854
2855 ff = rb_entry(*link, struct fuse_file, polled_node);
2856 wake_up_interruptible_sync(&ff->poll_wait);
2857 }
2858
2859 spin_unlock(&fc->lock);
2860 return 0;
2861}
2862
2863static void fuse_do_truncate(struct file *file)
2864{
2865 struct inode *inode = file->f_mapping->host;
2866 struct iattr attr;
2867
2868 attr.ia_valid = ATTR_SIZE;
2869 attr.ia_size = i_size_read(inode);
2870
2871 attr.ia_file = file;
2872 attr.ia_valid |= ATTR_FILE;
2873
2874 fuse_do_setattr(file_dentry(file), &attr, file);
2875}
2876
2877static inline loff_t fuse_round_up(struct fuse_conn *fc, loff_t off)
2878{
2879 return round_up(off, fc->max_pages << PAGE_SHIFT);
2880}
2881
2882static ssize_t
2883fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
2884{
2885 DECLARE_COMPLETION_ONSTACK(wait);
2886 ssize_t ret = 0;
2887 struct file *file = iocb->ki_filp;
2888 struct fuse_file *ff = file->private_data;
2889 loff_t pos = 0;
2890 struct inode *inode;
2891 loff_t i_size;
2892 size_t count = iov_iter_count(iter), shortened = 0;
2893 loff_t offset = iocb->ki_pos;
2894 struct fuse_io_priv *io;
2895
2896 pos = offset;
2897 inode = file->f_mapping->host;
2898 i_size = i_size_read(inode);
2899
2900 if ((iov_iter_rw(iter) == READ) && (offset >= i_size))
2901 return 0;
2902
2903 io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2904 if (!io)
2905 return -ENOMEM;
2906 spin_lock_init(&io->lock);
2907 kref_init(&io->refcnt);
2908 io->reqs = 1;
2909 io->bytes = -1;
2910 io->size = 0;
2911 io->offset = offset;
2912 io->write = (iov_iter_rw(iter) == WRITE);
2913 io->err = 0;
2914 /*
2915 * By default, we want to optimize all I/Os with async request
2916 * submission to the client filesystem if supported.
2917 */
2918 io->async = ff->fm->fc->async_dio;
2919 io->iocb = iocb;
2920 io->blocking = is_sync_kiocb(iocb);
2921
2922 /* optimization for short read */
2923 if (io->async && !io->write && offset + count > i_size) {
2924 iov_iter_truncate(iter, fuse_round_up(ff->fm->fc, i_size - offset));
2925 shortened = count - iov_iter_count(iter);
2926 count -= shortened;
2927 }
2928
2929 /*
2930 * We cannot asynchronously extend the size of a file.
2931 * In such case the aio will behave exactly like sync io.
2932 */
2933 if ((offset + count > i_size) && io->write)
2934 io->blocking = true;
2935
2936 if (io->async && io->blocking) {
2937 /*
2938 * Additional reference to keep io around after
2939 * calling fuse_aio_complete()
2940 */
2941 kref_get(&io->refcnt);
2942 io->done = &wait;
2943 }
2944
2945 if (iov_iter_rw(iter) == WRITE) {
2946 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
2947 fuse_invalidate_attr_mask(inode, FUSE_STATX_MODSIZE);
2948 } else {
2949 ret = __fuse_direct_read(io, iter, &pos);
2950 }
2951 iov_iter_reexpand(iter, iov_iter_count(iter) + shortened);
2952
2953 if (io->async) {
2954 bool blocking = io->blocking;
2955
2956 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2957
2958 /* we have a non-extending, async request, so return */
2959 if (!blocking)
2960 return -EIOCBQUEUED;
2961
2962 wait_for_completion(&wait);
2963 ret = fuse_get_res_by_io(io);
2964 }
2965
2966 kref_put(&io->refcnt, fuse_io_release);
2967
2968 if (iov_iter_rw(iter) == WRITE) {
2969 fuse_write_update_attr(inode, pos, ret);
2970 /* For extending writes we already hold exclusive lock */
2971 if (ret < 0 && offset + count > i_size)
2972 fuse_do_truncate(file);
2973 }
2974
2975 return ret;
2976}
2977
2978static int fuse_writeback_range(struct inode *inode, loff_t start, loff_t end)
2979{
2980 int err = filemap_write_and_wait_range(inode->i_mapping, start, LLONG_MAX);
2981
2982 if (!err)
2983 fuse_sync_writes(inode);
2984
2985 return err;
2986}
2987
2988static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2989 loff_t length)
2990{
2991 struct fuse_file *ff = file->private_data;
2992 struct inode *inode = file_inode(file);
2993 struct fuse_inode *fi = get_fuse_inode(inode);
2994 struct fuse_mount *fm = ff->fm;
2995 FUSE_ARGS(args);
2996 struct fuse_fallocate_in inarg = {
2997 .fh = ff->fh,
2998 .offset = offset,
2999 .length = length,
3000 .mode = mode
3001 };
3002 int err;
3003 bool block_faults = FUSE_IS_DAX(inode) &&
3004 (!(mode & FALLOC_FL_KEEP_SIZE) ||
3005 (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE)));
3006
3007 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3008 FALLOC_FL_ZERO_RANGE))
3009 return -EOPNOTSUPP;
3010
3011 if (fm->fc->no_fallocate)
3012 return -EOPNOTSUPP;
3013
3014 inode_lock(inode);
3015 if (block_faults) {
3016 filemap_invalidate_lock(inode->i_mapping);
3017 err = fuse_dax_break_layouts(inode, 0, 0);
3018 if (err)
3019 goto out;
3020 }
3021
3022 if (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE)) {
3023 loff_t endbyte = offset + length - 1;
3024
3025 err = fuse_writeback_range(inode, offset, endbyte);
3026 if (err)
3027 goto out;
3028 }
3029
3030 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
3031 offset + length > i_size_read(inode)) {
3032 err = inode_newsize_ok(inode, offset + length);
3033 if (err)
3034 goto out;
3035 }
3036
3037 err = file_modified(file);
3038 if (err)
3039 goto out;
3040
3041 if (!(mode & FALLOC_FL_KEEP_SIZE))
3042 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3043
3044 args.opcode = FUSE_FALLOCATE;
3045 args.nodeid = ff->nodeid;
3046 args.in_numargs = 1;
3047 args.in_args[0].size = sizeof(inarg);
3048 args.in_args[0].value = &inarg;
3049 err = fuse_simple_request(fm, &args);
3050 if (err == -ENOSYS) {
3051 fm->fc->no_fallocate = 1;
3052 err = -EOPNOTSUPP;
3053 }
3054 if (err)
3055 goto out;
3056
3057 /* we could have extended the file */
3058 if (!(mode & FALLOC_FL_KEEP_SIZE)) {
3059 if (fuse_write_update_attr(inode, offset + length, length))
3060 file_update_time(file);
3061 }
3062
3063 if (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE))
3064 truncate_pagecache_range(inode, offset, offset + length - 1);
3065
3066 fuse_invalidate_attr_mask(inode, FUSE_STATX_MODSIZE);
3067
3068out:
3069 if (!(mode & FALLOC_FL_KEEP_SIZE))
3070 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3071
3072 if (block_faults)
3073 filemap_invalidate_unlock(inode->i_mapping);
3074
3075 inode_unlock(inode);
3076
3077 fuse_flush_time_update(inode);
3078
3079 return err;
3080}
3081
3082static ssize_t __fuse_copy_file_range(struct file *file_in, loff_t pos_in,
3083 struct file *file_out, loff_t pos_out,
3084 size_t len, unsigned int flags)
3085{
3086 struct fuse_file *ff_in = file_in->private_data;
3087 struct fuse_file *ff_out = file_out->private_data;
3088 struct inode *inode_in = file_inode(file_in);
3089 struct inode *inode_out = file_inode(file_out);
3090 struct fuse_inode *fi_out = get_fuse_inode(inode_out);
3091 struct fuse_mount *fm = ff_in->fm;
3092 struct fuse_conn *fc = fm->fc;
3093 FUSE_ARGS(args);
3094 struct fuse_copy_file_range_in inarg = {
3095 .fh_in = ff_in->fh,
3096 .off_in = pos_in,
3097 .nodeid_out = ff_out->nodeid,
3098 .fh_out = ff_out->fh,
3099 .off_out = pos_out,
3100 .len = len,
3101 .flags = flags
3102 };
3103 struct fuse_write_out outarg;
3104 ssize_t err;
3105 /* mark unstable when write-back is not used, and file_out gets
3106 * extended */
3107 bool is_unstable = (!fc->writeback_cache) &&
3108 ((pos_out + len) > inode_out->i_size);
3109
3110 if (fc->no_copy_file_range)
3111 return -EOPNOTSUPP;
3112
3113 if (file_inode(file_in)->i_sb != file_inode(file_out)->i_sb)
3114 return -EXDEV;
3115
3116 inode_lock(inode_in);
3117 err = fuse_writeback_range(inode_in, pos_in, pos_in + len - 1);
3118 inode_unlock(inode_in);
3119 if (err)
3120 return err;
3121
3122 inode_lock(inode_out);
3123
3124 err = file_modified(file_out);
3125 if (err)
3126 goto out;
3127
3128 /*
3129 * Write out dirty pages in the destination file before sending the COPY
3130 * request to userspace. After the request is completed, truncate off
3131 * pages (including partial ones) from the cache that have been copied,
3132 * since these contain stale data at that point.
3133 *
3134 * This should be mostly correct, but if the COPY writes to partial
3135 * pages (at the start or end) and the parts not covered by the COPY are
3136 * written through a memory map after calling fuse_writeback_range(),
3137 * then these partial page modifications will be lost on truncation.
3138 *
3139 * It is unlikely that someone would rely on such mixed style
3140 * modifications. Yet this does give less guarantees than if the
3141 * copying was performed with write(2).
3142 *
3143 * To fix this a mapping->invalidate_lock could be used to prevent new
3144 * faults while the copy is ongoing.
3145 */
3146 err = fuse_writeback_range(inode_out, pos_out, pos_out + len - 1);
3147 if (err)
3148 goto out;
3149
3150 if (is_unstable)
3151 set_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3152
3153 args.opcode = FUSE_COPY_FILE_RANGE;
3154 args.nodeid = ff_in->nodeid;
3155 args.in_numargs = 1;
3156 args.in_args[0].size = sizeof(inarg);
3157 args.in_args[0].value = &inarg;
3158 args.out_numargs = 1;
3159 args.out_args[0].size = sizeof(outarg);
3160 args.out_args[0].value = &outarg;
3161 err = fuse_simple_request(fm, &args);
3162 if (err == -ENOSYS) {
3163 fc->no_copy_file_range = 1;
3164 err = -EOPNOTSUPP;
3165 }
3166 if (err)
3167 goto out;
3168
3169 truncate_inode_pages_range(inode_out->i_mapping,
3170 ALIGN_DOWN(pos_out, PAGE_SIZE),
3171 ALIGN(pos_out + outarg.size, PAGE_SIZE) - 1);
3172
3173 file_update_time(file_out);
3174 fuse_write_update_attr(inode_out, pos_out + outarg.size, outarg.size);
3175
3176 err = outarg.size;
3177out:
3178 if (is_unstable)
3179 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3180
3181 inode_unlock(inode_out);
3182 file_accessed(file_in);
3183
3184 fuse_flush_time_update(inode_out);
3185
3186 return err;
3187}
3188
3189static ssize_t fuse_copy_file_range(struct file *src_file, loff_t src_off,
3190 struct file *dst_file, loff_t dst_off,
3191 size_t len, unsigned int flags)
3192{
3193 ssize_t ret;
3194
3195 ret = __fuse_copy_file_range(src_file, src_off, dst_file, dst_off,
3196 len, flags);
3197
3198 if (ret == -EOPNOTSUPP || ret == -EXDEV)
3199 ret = generic_copy_file_range(src_file, src_off, dst_file,
3200 dst_off, len, flags);
3201 return ret;
3202}
3203
3204static const struct file_operations fuse_file_operations = {
3205 .llseek = fuse_file_llseek,
3206 .read_iter = fuse_file_read_iter,
3207 .write_iter = fuse_file_write_iter,
3208 .mmap = fuse_file_mmap,
3209 .open = fuse_open,
3210 .flush = fuse_flush,
3211 .release = fuse_release,
3212 .fsync = fuse_fsync,
3213 .lock = fuse_file_lock,
3214 .get_unmapped_area = thp_get_unmapped_area,
3215 .flock = fuse_file_flock,
3216 .splice_read = generic_file_splice_read,
3217 .splice_write = iter_file_splice_write,
3218 .unlocked_ioctl = fuse_file_ioctl,
3219 .compat_ioctl = fuse_file_compat_ioctl,
3220 .poll = fuse_file_poll,
3221 .fallocate = fuse_file_fallocate,
3222 .copy_file_range = fuse_copy_file_range,
3223};
3224
3225static const struct address_space_operations fuse_file_aops = {
3226 .read_folio = fuse_read_folio,
3227 .readahead = fuse_readahead,
3228 .writepage = fuse_writepage,
3229 .writepages = fuse_writepages,
3230 .launder_folio = fuse_launder_folio,
3231 .dirty_folio = filemap_dirty_folio,
3232 .bmap = fuse_bmap,
3233 .direct_IO = fuse_direct_IO,
3234 .write_begin = fuse_write_begin,
3235 .write_end = fuse_write_end,
3236};
3237
3238void fuse_init_file_inode(struct inode *inode, unsigned int flags)
3239{
3240 struct fuse_inode *fi = get_fuse_inode(inode);
3241
3242 inode->i_fop = &fuse_file_operations;
3243 inode->i_data.a_ops = &fuse_file_aops;
3244
3245 INIT_LIST_HEAD(&fi->write_files);
3246 INIT_LIST_HEAD(&fi->queued_writes);
3247 fi->writectr = 0;
3248 init_waitqueue_head(&fi->page_waitq);
3249 fi->writepages = RB_ROOT;
3250
3251 if (IS_ENABLED(CONFIG_FUSE_DAX))
3252 fuse_dax_inode_init(inode, flags);
3253}
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/compat.h>
18#include <linux/swap.h>
19#include <linux/falloc.h>
20#include <linux/uio.h>
21
22static struct page **fuse_pages_alloc(unsigned int npages, gfp_t flags,
23 struct fuse_page_desc **desc)
24{
25 struct page **pages;
26
27 pages = kzalloc(npages * (sizeof(struct page *) +
28 sizeof(struct fuse_page_desc)), flags);
29 *desc = (void *) (pages + npages);
30
31 return pages;
32}
33
34static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
35 int opcode, struct fuse_open_out *outargp)
36{
37 struct fuse_open_in inarg;
38 FUSE_ARGS(args);
39
40 memset(&inarg, 0, sizeof(inarg));
41 inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
42 if (!fc->atomic_o_trunc)
43 inarg.flags &= ~O_TRUNC;
44 args.opcode = opcode;
45 args.nodeid = nodeid;
46 args.in_numargs = 1;
47 args.in_args[0].size = sizeof(inarg);
48 args.in_args[0].value = &inarg;
49 args.out_numargs = 1;
50 args.out_args[0].size = sizeof(*outargp);
51 args.out_args[0].value = outargp;
52
53 return fuse_simple_request(fc, &args);
54}
55
56struct fuse_release_args {
57 struct fuse_args args;
58 struct fuse_release_in inarg;
59 struct inode *inode;
60};
61
62struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
63{
64 struct fuse_file *ff;
65
66 ff = kzalloc(sizeof(struct fuse_file), GFP_KERNEL_ACCOUNT);
67 if (unlikely(!ff))
68 return NULL;
69
70 ff->fc = fc;
71 ff->release_args = kzalloc(sizeof(*ff->release_args),
72 GFP_KERNEL_ACCOUNT);
73 if (!ff->release_args) {
74 kfree(ff);
75 return NULL;
76 }
77
78 INIT_LIST_HEAD(&ff->write_entry);
79 mutex_init(&ff->readdir.lock);
80 refcount_set(&ff->count, 1);
81 RB_CLEAR_NODE(&ff->polled_node);
82 init_waitqueue_head(&ff->poll_wait);
83
84 ff->kh = atomic64_inc_return(&fc->khctr);
85
86 return ff;
87}
88
89void fuse_file_free(struct fuse_file *ff)
90{
91 kfree(ff->release_args);
92 mutex_destroy(&ff->readdir.lock);
93 kfree(ff);
94}
95
96static struct fuse_file *fuse_file_get(struct fuse_file *ff)
97{
98 refcount_inc(&ff->count);
99 return ff;
100}
101
102static void fuse_release_end(struct fuse_conn *fc, struct fuse_args *args,
103 int error)
104{
105 struct fuse_release_args *ra = container_of(args, typeof(*ra), args);
106
107 iput(ra->inode);
108 kfree(ra);
109}
110
111static void fuse_file_put(struct fuse_file *ff, bool sync, bool isdir)
112{
113 if (refcount_dec_and_test(&ff->count)) {
114 struct fuse_args *args = &ff->release_args->args;
115
116 if (isdir ? ff->fc->no_opendir : ff->fc->no_open) {
117 /* Do nothing when client does not implement 'open' */
118 fuse_release_end(ff->fc, args, 0);
119 } else if (sync) {
120 fuse_simple_request(ff->fc, args);
121 fuse_release_end(ff->fc, args, 0);
122 } else {
123 args->end = fuse_release_end;
124 if (fuse_simple_background(ff->fc, args,
125 GFP_KERNEL | __GFP_NOFAIL))
126 fuse_release_end(ff->fc, args, -ENOTCONN);
127 }
128 kfree(ff);
129 }
130}
131
132int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
133 bool isdir)
134{
135 struct fuse_file *ff;
136 int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
137
138 ff = fuse_file_alloc(fc);
139 if (!ff)
140 return -ENOMEM;
141
142 ff->fh = 0;
143 /* Default for no-open */
144 ff->open_flags = FOPEN_KEEP_CACHE | (isdir ? FOPEN_CACHE_DIR : 0);
145 if (isdir ? !fc->no_opendir : !fc->no_open) {
146 struct fuse_open_out outarg;
147 int err;
148
149 err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
150 if (!err) {
151 ff->fh = outarg.fh;
152 ff->open_flags = outarg.open_flags;
153
154 } else if (err != -ENOSYS) {
155 fuse_file_free(ff);
156 return err;
157 } else {
158 if (isdir)
159 fc->no_opendir = 1;
160 else
161 fc->no_open = 1;
162 }
163 }
164
165 if (isdir)
166 ff->open_flags &= ~FOPEN_DIRECT_IO;
167
168 ff->nodeid = nodeid;
169 file->private_data = ff;
170
171 return 0;
172}
173EXPORT_SYMBOL_GPL(fuse_do_open);
174
175static void fuse_link_write_file(struct file *file)
176{
177 struct inode *inode = file_inode(file);
178 struct fuse_inode *fi = get_fuse_inode(inode);
179 struct fuse_file *ff = file->private_data;
180 /*
181 * file may be written through mmap, so chain it onto the
182 * inodes's write_file list
183 */
184 spin_lock(&fi->lock);
185 if (list_empty(&ff->write_entry))
186 list_add(&ff->write_entry, &fi->write_files);
187 spin_unlock(&fi->lock);
188}
189
190void fuse_finish_open(struct inode *inode, struct file *file)
191{
192 struct fuse_file *ff = file->private_data;
193 struct fuse_conn *fc = get_fuse_conn(inode);
194
195 if (!(ff->open_flags & FOPEN_KEEP_CACHE))
196 invalidate_inode_pages2(inode->i_mapping);
197 if (ff->open_flags & FOPEN_STREAM)
198 stream_open(inode, file);
199 else if (ff->open_flags & FOPEN_NONSEEKABLE)
200 nonseekable_open(inode, file);
201 if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
202 struct fuse_inode *fi = get_fuse_inode(inode);
203
204 spin_lock(&fi->lock);
205 fi->attr_version = atomic64_inc_return(&fc->attr_version);
206 i_size_write(inode, 0);
207 spin_unlock(&fi->lock);
208 fuse_invalidate_attr(inode);
209 if (fc->writeback_cache)
210 file_update_time(file);
211 }
212 if ((file->f_mode & FMODE_WRITE) && fc->writeback_cache)
213 fuse_link_write_file(file);
214}
215
216int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
217{
218 struct fuse_conn *fc = get_fuse_conn(inode);
219 int err;
220 bool is_wb_truncate = (file->f_flags & O_TRUNC) &&
221 fc->atomic_o_trunc &&
222 fc->writeback_cache;
223
224 err = generic_file_open(inode, file);
225 if (err)
226 return err;
227
228 if (is_wb_truncate) {
229 inode_lock(inode);
230 fuse_set_nowrite(inode);
231 }
232
233 err = fuse_do_open(fc, get_node_id(inode), file, isdir);
234
235 if (!err)
236 fuse_finish_open(inode, file);
237
238 if (is_wb_truncate) {
239 fuse_release_nowrite(inode);
240 inode_unlock(inode);
241 }
242
243 return err;
244}
245
246static void fuse_prepare_release(struct fuse_inode *fi, struct fuse_file *ff,
247 int flags, int opcode)
248{
249 struct fuse_conn *fc = ff->fc;
250 struct fuse_release_args *ra = ff->release_args;
251
252 /* Inode is NULL on error path of fuse_create_open() */
253 if (likely(fi)) {
254 spin_lock(&fi->lock);
255 list_del(&ff->write_entry);
256 spin_unlock(&fi->lock);
257 }
258 spin_lock(&fc->lock);
259 if (!RB_EMPTY_NODE(&ff->polled_node))
260 rb_erase(&ff->polled_node, &fc->polled_files);
261 spin_unlock(&fc->lock);
262
263 wake_up_interruptible_all(&ff->poll_wait);
264
265 ra->inarg.fh = ff->fh;
266 ra->inarg.flags = flags;
267 ra->args.in_numargs = 1;
268 ra->args.in_args[0].size = sizeof(struct fuse_release_in);
269 ra->args.in_args[0].value = &ra->inarg;
270 ra->args.opcode = opcode;
271 ra->args.nodeid = ff->nodeid;
272 ra->args.force = true;
273 ra->args.nocreds = true;
274}
275
276void fuse_release_common(struct file *file, bool isdir)
277{
278 struct fuse_inode *fi = get_fuse_inode(file_inode(file));
279 struct fuse_file *ff = file->private_data;
280 struct fuse_release_args *ra = ff->release_args;
281 int opcode = isdir ? FUSE_RELEASEDIR : FUSE_RELEASE;
282
283 fuse_prepare_release(fi, ff, file->f_flags, opcode);
284
285 if (ff->flock) {
286 ra->inarg.release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
287 ra->inarg.lock_owner = fuse_lock_owner_id(ff->fc,
288 (fl_owner_t) file);
289 }
290 /* Hold inode until release is finished */
291 ra->inode = igrab(file_inode(file));
292
293 /*
294 * Normally this will send the RELEASE request, however if
295 * some asynchronous READ or WRITE requests are outstanding,
296 * the sending will be delayed.
297 *
298 * Make the release synchronous if this is a fuseblk mount,
299 * synchronous RELEASE is allowed (and desirable) in this case
300 * because the server can be trusted not to screw up.
301 */
302 fuse_file_put(ff, ff->fc->destroy, isdir);
303}
304
305static int fuse_open(struct inode *inode, struct file *file)
306{
307 return fuse_open_common(inode, file, false);
308}
309
310static int fuse_release(struct inode *inode, struct file *file)
311{
312 struct fuse_conn *fc = get_fuse_conn(inode);
313
314 /* see fuse_vma_close() for !writeback_cache case */
315 if (fc->writeback_cache)
316 write_inode_now(inode, 1);
317
318 fuse_release_common(file, false);
319
320 /* return value is ignored by VFS */
321 return 0;
322}
323
324void fuse_sync_release(struct fuse_inode *fi, struct fuse_file *ff, int flags)
325{
326 WARN_ON(refcount_read(&ff->count) > 1);
327 fuse_prepare_release(fi, ff, flags, FUSE_RELEASE);
328 /*
329 * iput(NULL) is a no-op and since the refcount is 1 and everything's
330 * synchronous, we are fine with not doing igrab() here"
331 */
332 fuse_file_put(ff, true, false);
333}
334EXPORT_SYMBOL_GPL(fuse_sync_release);
335
336/*
337 * Scramble the ID space with XTEA, so that the value of the files_struct
338 * pointer is not exposed to userspace.
339 */
340u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
341{
342 u32 *k = fc->scramble_key;
343 u64 v = (unsigned long) id;
344 u32 v0 = v;
345 u32 v1 = v >> 32;
346 u32 sum = 0;
347 int i;
348
349 for (i = 0; i < 32; i++) {
350 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
351 sum += 0x9E3779B9;
352 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
353 }
354
355 return (u64) v0 + ((u64) v1 << 32);
356}
357
358struct fuse_writepage_args {
359 struct fuse_io_args ia;
360 struct list_head writepages_entry;
361 struct list_head queue_entry;
362 struct fuse_writepage_args *next;
363 struct inode *inode;
364};
365
366static struct fuse_writepage_args *fuse_find_writeback(struct fuse_inode *fi,
367 pgoff_t idx_from, pgoff_t idx_to)
368{
369 struct fuse_writepage_args *wpa;
370
371 list_for_each_entry(wpa, &fi->writepages, writepages_entry) {
372 pgoff_t curr_index;
373
374 WARN_ON(get_fuse_inode(wpa->inode) != fi);
375 curr_index = wpa->ia.write.in.offset >> PAGE_SHIFT;
376 if (idx_from < curr_index + wpa->ia.ap.num_pages &&
377 curr_index <= idx_to) {
378 return wpa;
379 }
380 }
381 return NULL;
382}
383
384/*
385 * Check if any page in a range is under writeback
386 *
387 * This is currently done by walking the list of writepage requests
388 * for the inode, which can be pretty inefficient.
389 */
390static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
391 pgoff_t idx_to)
392{
393 struct fuse_inode *fi = get_fuse_inode(inode);
394 bool found;
395
396 spin_lock(&fi->lock);
397 found = fuse_find_writeback(fi, idx_from, idx_to);
398 spin_unlock(&fi->lock);
399
400 return found;
401}
402
403static inline bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
404{
405 return fuse_range_is_writeback(inode, index, index);
406}
407
408/*
409 * Wait for page writeback to be completed.
410 *
411 * Since fuse doesn't rely on the VM writeback tracking, this has to
412 * use some other means.
413 */
414static void fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
415{
416 struct fuse_inode *fi = get_fuse_inode(inode);
417
418 wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
419}
420
421/*
422 * Wait for all pending writepages on the inode to finish.
423 *
424 * This is currently done by blocking further writes with FUSE_NOWRITE
425 * and waiting for all sent writes to complete.
426 *
427 * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
428 * could conflict with truncation.
429 */
430static void fuse_sync_writes(struct inode *inode)
431{
432 fuse_set_nowrite(inode);
433 fuse_release_nowrite(inode);
434}
435
436static int fuse_flush(struct file *file, fl_owner_t id)
437{
438 struct inode *inode = file_inode(file);
439 struct fuse_conn *fc = get_fuse_conn(inode);
440 struct fuse_file *ff = file->private_data;
441 struct fuse_flush_in inarg;
442 FUSE_ARGS(args);
443 int err;
444
445 if (is_bad_inode(inode))
446 return -EIO;
447
448 if (fc->no_flush)
449 return 0;
450
451 err = write_inode_now(inode, 1);
452 if (err)
453 return err;
454
455 inode_lock(inode);
456 fuse_sync_writes(inode);
457 inode_unlock(inode);
458
459 err = filemap_check_errors(file->f_mapping);
460 if (err)
461 return err;
462
463 memset(&inarg, 0, sizeof(inarg));
464 inarg.fh = ff->fh;
465 inarg.lock_owner = fuse_lock_owner_id(fc, id);
466 args.opcode = FUSE_FLUSH;
467 args.nodeid = get_node_id(inode);
468 args.in_numargs = 1;
469 args.in_args[0].size = sizeof(inarg);
470 args.in_args[0].value = &inarg;
471 args.force = true;
472
473 err = fuse_simple_request(fc, &args);
474 if (err == -ENOSYS) {
475 fc->no_flush = 1;
476 err = 0;
477 }
478 return err;
479}
480
481int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
482 int datasync, int opcode)
483{
484 struct inode *inode = file->f_mapping->host;
485 struct fuse_conn *fc = get_fuse_conn(inode);
486 struct fuse_file *ff = file->private_data;
487 FUSE_ARGS(args);
488 struct fuse_fsync_in inarg;
489
490 memset(&inarg, 0, sizeof(inarg));
491 inarg.fh = ff->fh;
492 inarg.fsync_flags = datasync ? FUSE_FSYNC_FDATASYNC : 0;
493 args.opcode = opcode;
494 args.nodeid = get_node_id(inode);
495 args.in_numargs = 1;
496 args.in_args[0].size = sizeof(inarg);
497 args.in_args[0].value = &inarg;
498 return fuse_simple_request(fc, &args);
499}
500
501static int fuse_fsync(struct file *file, loff_t start, loff_t end,
502 int datasync)
503{
504 struct inode *inode = file->f_mapping->host;
505 struct fuse_conn *fc = get_fuse_conn(inode);
506 int err;
507
508 if (is_bad_inode(inode))
509 return -EIO;
510
511 inode_lock(inode);
512
513 /*
514 * Start writeback against all dirty pages of the inode, then
515 * wait for all outstanding writes, before sending the FSYNC
516 * request.
517 */
518 err = file_write_and_wait_range(file, start, end);
519 if (err)
520 goto out;
521
522 fuse_sync_writes(inode);
523
524 /*
525 * Due to implementation of fuse writeback
526 * file_write_and_wait_range() does not catch errors.
527 * We have to do this directly after fuse_sync_writes()
528 */
529 err = file_check_and_advance_wb_err(file);
530 if (err)
531 goto out;
532
533 err = sync_inode_metadata(inode, 1);
534 if (err)
535 goto out;
536
537 if (fc->no_fsync)
538 goto out;
539
540 err = fuse_fsync_common(file, start, end, datasync, FUSE_FSYNC);
541 if (err == -ENOSYS) {
542 fc->no_fsync = 1;
543 err = 0;
544 }
545out:
546 inode_unlock(inode);
547
548 return err;
549}
550
551void fuse_read_args_fill(struct fuse_io_args *ia, struct file *file, loff_t pos,
552 size_t count, int opcode)
553{
554 struct fuse_file *ff = file->private_data;
555 struct fuse_args *args = &ia->ap.args;
556
557 ia->read.in.fh = ff->fh;
558 ia->read.in.offset = pos;
559 ia->read.in.size = count;
560 ia->read.in.flags = file->f_flags;
561 args->opcode = opcode;
562 args->nodeid = ff->nodeid;
563 args->in_numargs = 1;
564 args->in_args[0].size = sizeof(ia->read.in);
565 args->in_args[0].value = &ia->read.in;
566 args->out_argvar = true;
567 args->out_numargs = 1;
568 args->out_args[0].size = count;
569}
570
571static void fuse_release_user_pages(struct fuse_args_pages *ap,
572 bool should_dirty)
573{
574 unsigned int i;
575
576 for (i = 0; i < ap->num_pages; i++) {
577 if (should_dirty)
578 set_page_dirty_lock(ap->pages[i]);
579 put_page(ap->pages[i]);
580 }
581}
582
583static void fuse_io_release(struct kref *kref)
584{
585 kfree(container_of(kref, struct fuse_io_priv, refcnt));
586}
587
588static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
589{
590 if (io->err)
591 return io->err;
592
593 if (io->bytes >= 0 && io->write)
594 return -EIO;
595
596 return io->bytes < 0 ? io->size : io->bytes;
597}
598
599/**
600 * In case of short read, the caller sets 'pos' to the position of
601 * actual end of fuse request in IO request. Otherwise, if bytes_requested
602 * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
603 *
604 * An example:
605 * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
606 * both submitted asynchronously. The first of them was ACKed by userspace as
607 * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
608 * second request was ACKed as short, e.g. only 1K was read, resulting in
609 * pos == 33K.
610 *
611 * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
612 * will be equal to the length of the longest contiguous fragment of
613 * transferred data starting from the beginning of IO request.
614 */
615static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
616{
617 int left;
618
619 spin_lock(&io->lock);
620 if (err)
621 io->err = io->err ? : err;
622 else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
623 io->bytes = pos;
624
625 left = --io->reqs;
626 if (!left && io->blocking)
627 complete(io->done);
628 spin_unlock(&io->lock);
629
630 if (!left && !io->blocking) {
631 ssize_t res = fuse_get_res_by_io(io);
632
633 if (res >= 0) {
634 struct inode *inode = file_inode(io->iocb->ki_filp);
635 struct fuse_conn *fc = get_fuse_conn(inode);
636 struct fuse_inode *fi = get_fuse_inode(inode);
637
638 spin_lock(&fi->lock);
639 fi->attr_version = atomic64_inc_return(&fc->attr_version);
640 spin_unlock(&fi->lock);
641 }
642
643 io->iocb->ki_complete(io->iocb, res, 0);
644 }
645
646 kref_put(&io->refcnt, fuse_io_release);
647}
648
649static struct fuse_io_args *fuse_io_alloc(struct fuse_io_priv *io,
650 unsigned int npages)
651{
652 struct fuse_io_args *ia;
653
654 ia = kzalloc(sizeof(*ia), GFP_KERNEL);
655 if (ia) {
656 ia->io = io;
657 ia->ap.pages = fuse_pages_alloc(npages, GFP_KERNEL,
658 &ia->ap.descs);
659 if (!ia->ap.pages) {
660 kfree(ia);
661 ia = NULL;
662 }
663 }
664 return ia;
665}
666
667static void fuse_io_free(struct fuse_io_args *ia)
668{
669 kfree(ia->ap.pages);
670 kfree(ia);
671}
672
673static void fuse_aio_complete_req(struct fuse_conn *fc, struct fuse_args *args,
674 int err)
675{
676 struct fuse_io_args *ia = container_of(args, typeof(*ia), ap.args);
677 struct fuse_io_priv *io = ia->io;
678 ssize_t pos = -1;
679
680 fuse_release_user_pages(&ia->ap, io->should_dirty);
681
682 if (err) {
683 /* Nothing */
684 } else if (io->write) {
685 if (ia->write.out.size > ia->write.in.size) {
686 err = -EIO;
687 } else if (ia->write.in.size != ia->write.out.size) {
688 pos = ia->write.in.offset - io->offset +
689 ia->write.out.size;
690 }
691 } else {
692 u32 outsize = args->out_args[0].size;
693
694 if (ia->read.in.size != outsize)
695 pos = ia->read.in.offset - io->offset + outsize;
696 }
697
698 fuse_aio_complete(io, err, pos);
699 fuse_io_free(ia);
700}
701
702static ssize_t fuse_async_req_send(struct fuse_conn *fc,
703 struct fuse_io_args *ia, size_t num_bytes)
704{
705 ssize_t err;
706 struct fuse_io_priv *io = ia->io;
707
708 spin_lock(&io->lock);
709 kref_get(&io->refcnt);
710 io->size += num_bytes;
711 io->reqs++;
712 spin_unlock(&io->lock);
713
714 ia->ap.args.end = fuse_aio_complete_req;
715 err = fuse_simple_background(fc, &ia->ap.args, GFP_KERNEL);
716
717 return err ?: num_bytes;
718}
719
720static ssize_t fuse_send_read(struct fuse_io_args *ia, loff_t pos, size_t count,
721 fl_owner_t owner)
722{
723 struct file *file = ia->io->iocb->ki_filp;
724 struct fuse_file *ff = file->private_data;
725 struct fuse_conn *fc = ff->fc;
726
727 fuse_read_args_fill(ia, file, pos, count, FUSE_READ);
728 if (owner != NULL) {
729 ia->read.in.read_flags |= FUSE_READ_LOCKOWNER;
730 ia->read.in.lock_owner = fuse_lock_owner_id(fc, owner);
731 }
732
733 if (ia->io->async)
734 return fuse_async_req_send(fc, ia, count);
735
736 return fuse_simple_request(fc, &ia->ap.args);
737}
738
739static void fuse_read_update_size(struct inode *inode, loff_t size,
740 u64 attr_ver)
741{
742 struct fuse_conn *fc = get_fuse_conn(inode);
743 struct fuse_inode *fi = get_fuse_inode(inode);
744
745 spin_lock(&fi->lock);
746 if (attr_ver == fi->attr_version && size < inode->i_size &&
747 !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
748 fi->attr_version = atomic64_inc_return(&fc->attr_version);
749 i_size_write(inode, size);
750 }
751 spin_unlock(&fi->lock);
752}
753
754static void fuse_short_read(struct inode *inode, u64 attr_ver, size_t num_read,
755 struct fuse_args_pages *ap)
756{
757 struct fuse_conn *fc = get_fuse_conn(inode);
758
759 if (fc->writeback_cache) {
760 /*
761 * A hole in a file. Some data after the hole are in page cache,
762 * but have not reached the client fs yet. So, the hole is not
763 * present there.
764 */
765 int i;
766 int start_idx = num_read >> PAGE_SHIFT;
767 size_t off = num_read & (PAGE_SIZE - 1);
768
769 for (i = start_idx; i < ap->num_pages; i++) {
770 zero_user_segment(ap->pages[i], off, PAGE_SIZE);
771 off = 0;
772 }
773 } else {
774 loff_t pos = page_offset(ap->pages[0]) + num_read;
775 fuse_read_update_size(inode, pos, attr_ver);
776 }
777}
778
779static int fuse_do_readpage(struct file *file, struct page *page)
780{
781 struct inode *inode = page->mapping->host;
782 struct fuse_conn *fc = get_fuse_conn(inode);
783 loff_t pos = page_offset(page);
784 struct fuse_page_desc desc = { .length = PAGE_SIZE };
785 struct fuse_io_args ia = {
786 .ap.args.page_zeroing = true,
787 .ap.args.out_pages = true,
788 .ap.num_pages = 1,
789 .ap.pages = &page,
790 .ap.descs = &desc,
791 };
792 ssize_t res;
793 u64 attr_ver;
794
795 /*
796 * Page writeback can extend beyond the lifetime of the
797 * page-cache page, so make sure we read a properly synced
798 * page.
799 */
800 fuse_wait_on_page_writeback(inode, page->index);
801
802 attr_ver = fuse_get_attr_version(fc);
803
804 fuse_read_args_fill(&ia, file, pos, desc.length, FUSE_READ);
805 res = fuse_simple_request(fc, &ia.ap.args);
806 if (res < 0)
807 return res;
808 /*
809 * Short read means EOF. If file size is larger, truncate it
810 */
811 if (res < desc.length)
812 fuse_short_read(inode, attr_ver, res, &ia.ap);
813
814 SetPageUptodate(page);
815
816 return 0;
817}
818
819static int fuse_readpage(struct file *file, struct page *page)
820{
821 struct inode *inode = page->mapping->host;
822 int err;
823
824 err = -EIO;
825 if (is_bad_inode(inode))
826 goto out;
827
828 err = fuse_do_readpage(file, page);
829 fuse_invalidate_atime(inode);
830 out:
831 unlock_page(page);
832 return err;
833}
834
835static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_args *args,
836 int err)
837{
838 int i;
839 struct fuse_io_args *ia = container_of(args, typeof(*ia), ap.args);
840 struct fuse_args_pages *ap = &ia->ap;
841 size_t count = ia->read.in.size;
842 size_t num_read = args->out_args[0].size;
843 struct address_space *mapping = NULL;
844
845 for (i = 0; mapping == NULL && i < ap->num_pages; i++)
846 mapping = ap->pages[i]->mapping;
847
848 if (mapping) {
849 struct inode *inode = mapping->host;
850
851 /*
852 * Short read means EOF. If file size is larger, truncate it
853 */
854 if (!err && num_read < count)
855 fuse_short_read(inode, ia->read.attr_ver, num_read, ap);
856
857 fuse_invalidate_atime(inode);
858 }
859
860 for (i = 0; i < ap->num_pages; i++) {
861 struct page *page = ap->pages[i];
862
863 if (!err)
864 SetPageUptodate(page);
865 else
866 SetPageError(page);
867 unlock_page(page);
868 put_page(page);
869 }
870 if (ia->ff)
871 fuse_file_put(ia->ff, false, false);
872
873 fuse_io_free(ia);
874}
875
876static void fuse_send_readpages(struct fuse_io_args *ia, struct file *file)
877{
878 struct fuse_file *ff = file->private_data;
879 struct fuse_conn *fc = ff->fc;
880 struct fuse_args_pages *ap = &ia->ap;
881 loff_t pos = page_offset(ap->pages[0]);
882 size_t count = ap->num_pages << PAGE_SHIFT;
883 int err;
884
885 ap->args.out_pages = true;
886 ap->args.page_zeroing = true;
887 ap->args.page_replace = true;
888 fuse_read_args_fill(ia, file, pos, count, FUSE_READ);
889 ia->read.attr_ver = fuse_get_attr_version(fc);
890 if (fc->async_read) {
891 ia->ff = fuse_file_get(ff);
892 ap->args.end = fuse_readpages_end;
893 err = fuse_simple_background(fc, &ap->args, GFP_KERNEL);
894 if (!err)
895 return;
896 } else {
897 err = fuse_simple_request(fc, &ap->args);
898 }
899 fuse_readpages_end(fc, &ap->args, err);
900}
901
902struct fuse_fill_data {
903 struct fuse_io_args *ia;
904 struct file *file;
905 struct inode *inode;
906 unsigned int nr_pages;
907 unsigned int max_pages;
908};
909
910static int fuse_readpages_fill(void *_data, struct page *page)
911{
912 struct fuse_fill_data *data = _data;
913 struct fuse_io_args *ia = data->ia;
914 struct fuse_args_pages *ap = &ia->ap;
915 struct inode *inode = data->inode;
916 struct fuse_conn *fc = get_fuse_conn(inode);
917
918 fuse_wait_on_page_writeback(inode, page->index);
919
920 if (ap->num_pages &&
921 (ap->num_pages == fc->max_pages ||
922 (ap->num_pages + 1) * PAGE_SIZE > fc->max_read ||
923 ap->pages[ap->num_pages - 1]->index + 1 != page->index)) {
924 data->max_pages = min_t(unsigned int, data->nr_pages,
925 fc->max_pages);
926 fuse_send_readpages(ia, data->file);
927 data->ia = ia = fuse_io_alloc(NULL, data->max_pages);
928 if (!ia) {
929 unlock_page(page);
930 return -ENOMEM;
931 }
932 ap = &ia->ap;
933 }
934
935 if (WARN_ON(ap->num_pages >= data->max_pages)) {
936 unlock_page(page);
937 fuse_io_free(ia);
938 return -EIO;
939 }
940
941 get_page(page);
942 ap->pages[ap->num_pages] = page;
943 ap->descs[ap->num_pages].length = PAGE_SIZE;
944 ap->num_pages++;
945 data->nr_pages--;
946 return 0;
947}
948
949static int fuse_readpages(struct file *file, struct address_space *mapping,
950 struct list_head *pages, unsigned nr_pages)
951{
952 struct inode *inode = mapping->host;
953 struct fuse_conn *fc = get_fuse_conn(inode);
954 struct fuse_fill_data data;
955 int err;
956
957 err = -EIO;
958 if (is_bad_inode(inode))
959 goto out;
960
961 data.file = file;
962 data.inode = inode;
963 data.nr_pages = nr_pages;
964 data.max_pages = min_t(unsigned int, nr_pages, fc->max_pages);
965;
966 data.ia = fuse_io_alloc(NULL, data.max_pages);
967 err = -ENOMEM;
968 if (!data.ia)
969 goto out;
970
971 err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
972 if (!err) {
973 if (data.ia->ap.num_pages)
974 fuse_send_readpages(data.ia, file);
975 else
976 fuse_io_free(data.ia);
977 }
978out:
979 return err;
980}
981
982static ssize_t fuse_cache_read_iter(struct kiocb *iocb, struct iov_iter *to)
983{
984 struct inode *inode = iocb->ki_filp->f_mapping->host;
985 struct fuse_conn *fc = get_fuse_conn(inode);
986
987 /*
988 * In auto invalidate mode, always update attributes on read.
989 * Otherwise, only update if we attempt to read past EOF (to ensure
990 * i_size is up to date).
991 */
992 if (fc->auto_inval_data ||
993 (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
994 int err;
995 err = fuse_update_attributes(inode, iocb->ki_filp);
996 if (err)
997 return err;
998 }
999
1000 return generic_file_read_iter(iocb, to);
1001}
1002
1003static void fuse_write_args_fill(struct fuse_io_args *ia, struct fuse_file *ff,
1004 loff_t pos, size_t count)
1005{
1006 struct fuse_args *args = &ia->ap.args;
1007
1008 ia->write.in.fh = ff->fh;
1009 ia->write.in.offset = pos;
1010 ia->write.in.size = count;
1011 args->opcode = FUSE_WRITE;
1012 args->nodeid = ff->nodeid;
1013 args->in_numargs = 2;
1014 if (ff->fc->minor < 9)
1015 args->in_args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
1016 else
1017 args->in_args[0].size = sizeof(ia->write.in);
1018 args->in_args[0].value = &ia->write.in;
1019 args->in_args[1].size = count;
1020 args->out_numargs = 1;
1021 args->out_args[0].size = sizeof(ia->write.out);
1022 args->out_args[0].value = &ia->write.out;
1023}
1024
1025static unsigned int fuse_write_flags(struct kiocb *iocb)
1026{
1027 unsigned int flags = iocb->ki_filp->f_flags;
1028
1029 if (iocb->ki_flags & IOCB_DSYNC)
1030 flags |= O_DSYNC;
1031 if (iocb->ki_flags & IOCB_SYNC)
1032 flags |= O_SYNC;
1033
1034 return flags;
1035}
1036
1037static ssize_t fuse_send_write(struct fuse_io_args *ia, loff_t pos,
1038 size_t count, fl_owner_t owner)
1039{
1040 struct kiocb *iocb = ia->io->iocb;
1041 struct file *file = iocb->ki_filp;
1042 struct fuse_file *ff = file->private_data;
1043 struct fuse_conn *fc = ff->fc;
1044 struct fuse_write_in *inarg = &ia->write.in;
1045 ssize_t err;
1046
1047 fuse_write_args_fill(ia, ff, pos, count);
1048 inarg->flags = fuse_write_flags(iocb);
1049 if (owner != NULL) {
1050 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
1051 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
1052 }
1053
1054 if (ia->io->async)
1055 return fuse_async_req_send(fc, ia, count);
1056
1057 err = fuse_simple_request(fc, &ia->ap.args);
1058 if (!err && ia->write.out.size > count)
1059 err = -EIO;
1060
1061 return err ?: ia->write.out.size;
1062}
1063
1064bool fuse_write_update_size(struct inode *inode, loff_t pos)
1065{
1066 struct fuse_conn *fc = get_fuse_conn(inode);
1067 struct fuse_inode *fi = get_fuse_inode(inode);
1068 bool ret = false;
1069
1070 spin_lock(&fi->lock);
1071 fi->attr_version = atomic64_inc_return(&fc->attr_version);
1072 if (pos > inode->i_size) {
1073 i_size_write(inode, pos);
1074 ret = true;
1075 }
1076 spin_unlock(&fi->lock);
1077
1078 return ret;
1079}
1080
1081static ssize_t fuse_send_write_pages(struct fuse_io_args *ia,
1082 struct kiocb *iocb, struct inode *inode,
1083 loff_t pos, size_t count)
1084{
1085 struct fuse_args_pages *ap = &ia->ap;
1086 struct file *file = iocb->ki_filp;
1087 struct fuse_file *ff = file->private_data;
1088 struct fuse_conn *fc = ff->fc;
1089 unsigned int offset, i;
1090 int err;
1091
1092 for (i = 0; i < ap->num_pages; i++)
1093 fuse_wait_on_page_writeback(inode, ap->pages[i]->index);
1094
1095 fuse_write_args_fill(ia, ff, pos, count);
1096 ia->write.in.flags = fuse_write_flags(iocb);
1097
1098 err = fuse_simple_request(fc, &ap->args);
1099
1100 offset = ap->descs[0].offset;
1101 count = ia->write.out.size;
1102 for (i = 0; i < ap->num_pages; i++) {
1103 struct page *page = ap->pages[i];
1104
1105 if (!err && !offset && count >= PAGE_SIZE)
1106 SetPageUptodate(page);
1107
1108 if (count > PAGE_SIZE - offset)
1109 count -= PAGE_SIZE - offset;
1110 else
1111 count = 0;
1112 offset = 0;
1113
1114 unlock_page(page);
1115 put_page(page);
1116 }
1117
1118 return err;
1119}
1120
1121static ssize_t fuse_fill_write_pages(struct fuse_args_pages *ap,
1122 struct address_space *mapping,
1123 struct iov_iter *ii, loff_t pos,
1124 unsigned int max_pages)
1125{
1126 struct fuse_conn *fc = get_fuse_conn(mapping->host);
1127 unsigned offset = pos & (PAGE_SIZE - 1);
1128 size_t count = 0;
1129 int err;
1130
1131 ap->args.in_pages = true;
1132 ap->descs[0].offset = offset;
1133
1134 do {
1135 size_t tmp;
1136 struct page *page;
1137 pgoff_t index = pos >> PAGE_SHIFT;
1138 size_t bytes = min_t(size_t, PAGE_SIZE - offset,
1139 iov_iter_count(ii));
1140
1141 bytes = min_t(size_t, bytes, fc->max_write - count);
1142
1143 again:
1144 err = -EFAULT;
1145 if (iov_iter_fault_in_readable(ii, bytes))
1146 break;
1147
1148 err = -ENOMEM;
1149 page = grab_cache_page_write_begin(mapping, index, 0);
1150 if (!page)
1151 break;
1152
1153 if (mapping_writably_mapped(mapping))
1154 flush_dcache_page(page);
1155
1156 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
1157 flush_dcache_page(page);
1158
1159 iov_iter_advance(ii, tmp);
1160 if (!tmp) {
1161 unlock_page(page);
1162 put_page(page);
1163 bytes = min(bytes, iov_iter_single_seg_count(ii));
1164 goto again;
1165 }
1166
1167 err = 0;
1168 ap->pages[ap->num_pages] = page;
1169 ap->descs[ap->num_pages].length = tmp;
1170 ap->num_pages++;
1171
1172 count += tmp;
1173 pos += tmp;
1174 offset += tmp;
1175 if (offset == PAGE_SIZE)
1176 offset = 0;
1177
1178 if (!fc->big_writes)
1179 break;
1180 } while (iov_iter_count(ii) && count < fc->max_write &&
1181 ap->num_pages < max_pages && offset == 0);
1182
1183 return count > 0 ? count : err;
1184}
1185
1186static inline unsigned int fuse_wr_pages(loff_t pos, size_t len,
1187 unsigned int max_pages)
1188{
1189 return min_t(unsigned int,
1190 ((pos + len - 1) >> PAGE_SHIFT) -
1191 (pos >> PAGE_SHIFT) + 1,
1192 max_pages);
1193}
1194
1195static ssize_t fuse_perform_write(struct kiocb *iocb,
1196 struct address_space *mapping,
1197 struct iov_iter *ii, loff_t pos)
1198{
1199 struct inode *inode = mapping->host;
1200 struct fuse_conn *fc = get_fuse_conn(inode);
1201 struct fuse_inode *fi = get_fuse_inode(inode);
1202 int err = 0;
1203 ssize_t res = 0;
1204
1205 if (inode->i_size < pos + iov_iter_count(ii))
1206 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1207
1208 do {
1209 ssize_t count;
1210 struct fuse_io_args ia = {};
1211 struct fuse_args_pages *ap = &ia.ap;
1212 unsigned int nr_pages = fuse_wr_pages(pos, iov_iter_count(ii),
1213 fc->max_pages);
1214
1215 ap->pages = fuse_pages_alloc(nr_pages, GFP_KERNEL, &ap->descs);
1216 if (!ap->pages) {
1217 err = -ENOMEM;
1218 break;
1219 }
1220
1221 count = fuse_fill_write_pages(ap, mapping, ii, pos, nr_pages);
1222 if (count <= 0) {
1223 err = count;
1224 } else {
1225 err = fuse_send_write_pages(&ia, iocb, inode,
1226 pos, count);
1227 if (!err) {
1228 size_t num_written = ia.write.out.size;
1229
1230 res += num_written;
1231 pos += num_written;
1232
1233 /* break out of the loop on short write */
1234 if (num_written != count)
1235 err = -EIO;
1236 }
1237 }
1238 kfree(ap->pages);
1239 } while (!err && iov_iter_count(ii));
1240
1241 if (res > 0)
1242 fuse_write_update_size(inode, pos);
1243
1244 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1245 fuse_invalidate_attr(inode);
1246
1247 return res > 0 ? res : err;
1248}
1249
1250static ssize_t fuse_cache_write_iter(struct kiocb *iocb, struct iov_iter *from)
1251{
1252 struct file *file = iocb->ki_filp;
1253 struct address_space *mapping = file->f_mapping;
1254 ssize_t written = 0;
1255 ssize_t written_buffered = 0;
1256 struct inode *inode = mapping->host;
1257 ssize_t err;
1258 loff_t endbyte = 0;
1259
1260 if (get_fuse_conn(inode)->writeback_cache) {
1261 /* Update size (EOF optimization) and mode (SUID clearing) */
1262 err = fuse_update_attributes(mapping->host, file);
1263 if (err)
1264 return err;
1265
1266 return generic_file_write_iter(iocb, from);
1267 }
1268
1269 inode_lock(inode);
1270
1271 /* We can write back this queue in page reclaim */
1272 current->backing_dev_info = inode_to_bdi(inode);
1273
1274 err = generic_write_checks(iocb, from);
1275 if (err <= 0)
1276 goto out;
1277
1278 err = file_remove_privs(file);
1279 if (err)
1280 goto out;
1281
1282 err = file_update_time(file);
1283 if (err)
1284 goto out;
1285
1286 if (iocb->ki_flags & IOCB_DIRECT) {
1287 loff_t pos = iocb->ki_pos;
1288 written = generic_file_direct_write(iocb, from);
1289 if (written < 0 || !iov_iter_count(from))
1290 goto out;
1291
1292 pos += written;
1293
1294 written_buffered = fuse_perform_write(iocb, mapping, from, pos);
1295 if (written_buffered < 0) {
1296 err = written_buffered;
1297 goto out;
1298 }
1299 endbyte = pos + written_buffered - 1;
1300
1301 err = filemap_write_and_wait_range(file->f_mapping, pos,
1302 endbyte);
1303 if (err)
1304 goto out;
1305
1306 invalidate_mapping_pages(file->f_mapping,
1307 pos >> PAGE_SHIFT,
1308 endbyte >> PAGE_SHIFT);
1309
1310 written += written_buffered;
1311 iocb->ki_pos = pos + written_buffered;
1312 } else {
1313 written = fuse_perform_write(iocb, mapping, from, iocb->ki_pos);
1314 if (written >= 0)
1315 iocb->ki_pos += written;
1316 }
1317out:
1318 current->backing_dev_info = NULL;
1319 inode_unlock(inode);
1320 if (written > 0)
1321 written = generic_write_sync(iocb, written);
1322
1323 return written ? written : err;
1324}
1325
1326static inline void fuse_page_descs_length_init(struct fuse_page_desc *descs,
1327 unsigned int index,
1328 unsigned int nr_pages)
1329{
1330 int i;
1331
1332 for (i = index; i < index + nr_pages; i++)
1333 descs[i].length = PAGE_SIZE - descs[i].offset;
1334}
1335
1336static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1337{
1338 return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1339}
1340
1341static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1342 size_t max_size)
1343{
1344 return min(iov_iter_single_seg_count(ii), max_size);
1345}
1346
1347static int fuse_get_user_pages(struct fuse_args_pages *ap, struct iov_iter *ii,
1348 size_t *nbytesp, int write,
1349 unsigned int max_pages)
1350{
1351 size_t nbytes = 0; /* # bytes already packed in req */
1352 ssize_t ret = 0;
1353
1354 /* Special case for kernel I/O: can copy directly into the buffer */
1355 if (iov_iter_is_kvec(ii)) {
1356 unsigned long user_addr = fuse_get_user_addr(ii);
1357 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1358
1359 if (write)
1360 ap->args.in_args[1].value = (void *) user_addr;
1361 else
1362 ap->args.out_args[0].value = (void *) user_addr;
1363
1364 iov_iter_advance(ii, frag_size);
1365 *nbytesp = frag_size;
1366 return 0;
1367 }
1368
1369 while (nbytes < *nbytesp && ap->num_pages < max_pages) {
1370 unsigned npages;
1371 size_t start;
1372 ret = iov_iter_get_pages(ii, &ap->pages[ap->num_pages],
1373 *nbytesp - nbytes,
1374 max_pages - ap->num_pages,
1375 &start);
1376 if (ret < 0)
1377 break;
1378
1379 iov_iter_advance(ii, ret);
1380 nbytes += ret;
1381
1382 ret += start;
1383 npages = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
1384
1385 ap->descs[ap->num_pages].offset = start;
1386 fuse_page_descs_length_init(ap->descs, ap->num_pages, npages);
1387
1388 ap->num_pages += npages;
1389 ap->descs[ap->num_pages - 1].length -=
1390 (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1391 }
1392
1393 if (write)
1394 ap->args.in_pages = 1;
1395 else
1396 ap->args.out_pages = 1;
1397
1398 *nbytesp = nbytes;
1399
1400 return ret < 0 ? ret : 0;
1401}
1402
1403ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1404 loff_t *ppos, int flags)
1405{
1406 int write = flags & FUSE_DIO_WRITE;
1407 int cuse = flags & FUSE_DIO_CUSE;
1408 struct file *file = io->iocb->ki_filp;
1409 struct inode *inode = file->f_mapping->host;
1410 struct fuse_file *ff = file->private_data;
1411 struct fuse_conn *fc = ff->fc;
1412 size_t nmax = write ? fc->max_write : fc->max_read;
1413 loff_t pos = *ppos;
1414 size_t count = iov_iter_count(iter);
1415 pgoff_t idx_from = pos >> PAGE_SHIFT;
1416 pgoff_t idx_to = (pos + count - 1) >> PAGE_SHIFT;
1417 ssize_t res = 0;
1418 int err = 0;
1419 struct fuse_io_args *ia;
1420 unsigned int max_pages;
1421
1422 max_pages = iov_iter_npages(iter, fc->max_pages);
1423 ia = fuse_io_alloc(io, max_pages);
1424 if (!ia)
1425 return -ENOMEM;
1426
1427 ia->io = io;
1428 if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1429 if (!write)
1430 inode_lock(inode);
1431 fuse_sync_writes(inode);
1432 if (!write)
1433 inode_unlock(inode);
1434 }
1435
1436 io->should_dirty = !write && iter_is_iovec(iter);
1437 while (count) {
1438 ssize_t nres;
1439 fl_owner_t owner = current->files;
1440 size_t nbytes = min(count, nmax);
1441
1442 err = fuse_get_user_pages(&ia->ap, iter, &nbytes, write,
1443 max_pages);
1444 if (err && !nbytes)
1445 break;
1446
1447 if (write) {
1448 if (!capable(CAP_FSETID))
1449 ia->write.in.write_flags |= FUSE_WRITE_KILL_PRIV;
1450
1451 nres = fuse_send_write(ia, pos, nbytes, owner);
1452 } else {
1453 nres = fuse_send_read(ia, pos, nbytes, owner);
1454 }
1455
1456 if (!io->async || nres < 0) {
1457 fuse_release_user_pages(&ia->ap, io->should_dirty);
1458 fuse_io_free(ia);
1459 }
1460 ia = NULL;
1461 if (nres < 0) {
1462 err = nres;
1463 break;
1464 }
1465 WARN_ON(nres > nbytes);
1466
1467 count -= nres;
1468 res += nres;
1469 pos += nres;
1470 if (nres != nbytes)
1471 break;
1472 if (count) {
1473 max_pages = iov_iter_npages(iter, fc->max_pages);
1474 ia = fuse_io_alloc(io, max_pages);
1475 if (!ia)
1476 break;
1477 }
1478 }
1479 if (ia)
1480 fuse_io_free(ia);
1481 if (res > 0)
1482 *ppos = pos;
1483
1484 return res > 0 ? res : err;
1485}
1486EXPORT_SYMBOL_GPL(fuse_direct_io);
1487
1488static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1489 struct iov_iter *iter,
1490 loff_t *ppos)
1491{
1492 ssize_t res;
1493 struct inode *inode = file_inode(io->iocb->ki_filp);
1494
1495 res = fuse_direct_io(io, iter, ppos, 0);
1496
1497 fuse_invalidate_atime(inode);
1498
1499 return res;
1500}
1501
1502static ssize_t fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter);
1503
1504static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1505{
1506 ssize_t res;
1507
1508 if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1509 res = fuse_direct_IO(iocb, to);
1510 } else {
1511 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1512
1513 res = __fuse_direct_read(&io, to, &iocb->ki_pos);
1514 }
1515
1516 return res;
1517}
1518
1519static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1520{
1521 struct inode *inode = file_inode(iocb->ki_filp);
1522 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1523 ssize_t res;
1524
1525 /* Don't allow parallel writes to the same file */
1526 inode_lock(inode);
1527 res = generic_write_checks(iocb, from);
1528 if (res > 0) {
1529 if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1530 res = fuse_direct_IO(iocb, from);
1531 } else {
1532 res = fuse_direct_io(&io, from, &iocb->ki_pos,
1533 FUSE_DIO_WRITE);
1534 }
1535 }
1536 fuse_invalidate_attr(inode);
1537 if (res > 0)
1538 fuse_write_update_size(inode, iocb->ki_pos);
1539 inode_unlock(inode);
1540
1541 return res;
1542}
1543
1544static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
1545{
1546 struct file *file = iocb->ki_filp;
1547 struct fuse_file *ff = file->private_data;
1548
1549 if (is_bad_inode(file_inode(file)))
1550 return -EIO;
1551
1552 if (!(ff->open_flags & FOPEN_DIRECT_IO))
1553 return fuse_cache_read_iter(iocb, to);
1554 else
1555 return fuse_direct_read_iter(iocb, to);
1556}
1557
1558static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1559{
1560 struct file *file = iocb->ki_filp;
1561 struct fuse_file *ff = file->private_data;
1562
1563 if (is_bad_inode(file_inode(file)))
1564 return -EIO;
1565
1566 if (!(ff->open_flags & FOPEN_DIRECT_IO))
1567 return fuse_cache_write_iter(iocb, from);
1568 else
1569 return fuse_direct_write_iter(iocb, from);
1570}
1571
1572static void fuse_writepage_free(struct fuse_writepage_args *wpa)
1573{
1574 struct fuse_args_pages *ap = &wpa->ia.ap;
1575 int i;
1576
1577 for (i = 0; i < ap->num_pages; i++)
1578 __free_page(ap->pages[i]);
1579
1580 if (wpa->ia.ff)
1581 fuse_file_put(wpa->ia.ff, false, false);
1582
1583 kfree(ap->pages);
1584 kfree(wpa);
1585}
1586
1587static void fuse_writepage_finish(struct fuse_conn *fc,
1588 struct fuse_writepage_args *wpa)
1589{
1590 struct fuse_args_pages *ap = &wpa->ia.ap;
1591 struct inode *inode = wpa->inode;
1592 struct fuse_inode *fi = get_fuse_inode(inode);
1593 struct backing_dev_info *bdi = inode_to_bdi(inode);
1594 int i;
1595
1596 list_del(&wpa->writepages_entry);
1597 for (i = 0; i < ap->num_pages; i++) {
1598 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1599 dec_node_page_state(ap->pages[i], NR_WRITEBACK_TEMP);
1600 wb_writeout_inc(&bdi->wb);
1601 }
1602 wake_up(&fi->page_waitq);
1603}
1604
1605/* Called under fi->lock, may release and reacquire it */
1606static void fuse_send_writepage(struct fuse_conn *fc,
1607 struct fuse_writepage_args *wpa, loff_t size)
1608__releases(fi->lock)
1609__acquires(fi->lock)
1610{
1611 struct fuse_writepage_args *aux, *next;
1612 struct fuse_inode *fi = get_fuse_inode(wpa->inode);
1613 struct fuse_write_in *inarg = &wpa->ia.write.in;
1614 struct fuse_args *args = &wpa->ia.ap.args;
1615 __u64 data_size = wpa->ia.ap.num_pages * PAGE_SIZE;
1616 int err;
1617
1618 fi->writectr++;
1619 if (inarg->offset + data_size <= size) {
1620 inarg->size = data_size;
1621 } else if (inarg->offset < size) {
1622 inarg->size = size - inarg->offset;
1623 } else {
1624 /* Got truncated off completely */
1625 goto out_free;
1626 }
1627
1628 args->in_args[1].size = inarg->size;
1629 args->force = true;
1630 args->nocreds = true;
1631
1632 err = fuse_simple_background(fc, args, GFP_ATOMIC);
1633 if (err == -ENOMEM) {
1634 spin_unlock(&fi->lock);
1635 err = fuse_simple_background(fc, args, GFP_NOFS | __GFP_NOFAIL);
1636 spin_lock(&fi->lock);
1637 }
1638
1639 /* Fails on broken connection only */
1640 if (unlikely(err))
1641 goto out_free;
1642
1643 return;
1644
1645 out_free:
1646 fi->writectr--;
1647 fuse_writepage_finish(fc, wpa);
1648 spin_unlock(&fi->lock);
1649
1650 /* After fuse_writepage_finish() aux request list is private */
1651 for (aux = wpa->next; aux; aux = next) {
1652 next = aux->next;
1653 aux->next = NULL;
1654 fuse_writepage_free(aux);
1655 }
1656
1657 fuse_writepage_free(wpa);
1658 spin_lock(&fi->lock);
1659}
1660
1661/*
1662 * If fi->writectr is positive (no truncate or fsync going on) send
1663 * all queued writepage requests.
1664 *
1665 * Called with fi->lock
1666 */
1667void fuse_flush_writepages(struct inode *inode)
1668__releases(fi->lock)
1669__acquires(fi->lock)
1670{
1671 struct fuse_conn *fc = get_fuse_conn(inode);
1672 struct fuse_inode *fi = get_fuse_inode(inode);
1673 loff_t crop = i_size_read(inode);
1674 struct fuse_writepage_args *wpa;
1675
1676 while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1677 wpa = list_entry(fi->queued_writes.next,
1678 struct fuse_writepage_args, queue_entry);
1679 list_del_init(&wpa->queue_entry);
1680 fuse_send_writepage(fc, wpa, crop);
1681 }
1682}
1683
1684static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_args *args,
1685 int error)
1686{
1687 struct fuse_writepage_args *wpa =
1688 container_of(args, typeof(*wpa), ia.ap.args);
1689 struct inode *inode = wpa->inode;
1690 struct fuse_inode *fi = get_fuse_inode(inode);
1691
1692 mapping_set_error(inode->i_mapping, error);
1693 spin_lock(&fi->lock);
1694 while (wpa->next) {
1695 struct fuse_conn *fc = get_fuse_conn(inode);
1696 struct fuse_write_in *inarg = &wpa->ia.write.in;
1697 struct fuse_writepage_args *next = wpa->next;
1698
1699 wpa->next = next->next;
1700 next->next = NULL;
1701 next->ia.ff = fuse_file_get(wpa->ia.ff);
1702 list_add(&next->writepages_entry, &fi->writepages);
1703
1704 /*
1705 * Skip fuse_flush_writepages() to make it easy to crop requests
1706 * based on primary request size.
1707 *
1708 * 1st case (trivial): there are no concurrent activities using
1709 * fuse_set/release_nowrite. Then we're on safe side because
1710 * fuse_flush_writepages() would call fuse_send_writepage()
1711 * anyway.
1712 *
1713 * 2nd case: someone called fuse_set_nowrite and it is waiting
1714 * now for completion of all in-flight requests. This happens
1715 * rarely and no more than once per page, so this should be
1716 * okay.
1717 *
1718 * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1719 * of fuse_set_nowrite..fuse_release_nowrite section. The fact
1720 * that fuse_set_nowrite returned implies that all in-flight
1721 * requests were completed along with all of their secondary
1722 * requests. Further primary requests are blocked by negative
1723 * writectr. Hence there cannot be any in-flight requests and
1724 * no invocations of fuse_writepage_end() while we're in
1725 * fuse_set_nowrite..fuse_release_nowrite section.
1726 */
1727 fuse_send_writepage(fc, next, inarg->offset + inarg->size);
1728 }
1729 fi->writectr--;
1730 fuse_writepage_finish(fc, wpa);
1731 spin_unlock(&fi->lock);
1732 fuse_writepage_free(wpa);
1733}
1734
1735static struct fuse_file *__fuse_write_file_get(struct fuse_conn *fc,
1736 struct fuse_inode *fi)
1737{
1738 struct fuse_file *ff = NULL;
1739
1740 spin_lock(&fi->lock);
1741 if (!list_empty(&fi->write_files)) {
1742 ff = list_entry(fi->write_files.next, struct fuse_file,
1743 write_entry);
1744 fuse_file_get(ff);
1745 }
1746 spin_unlock(&fi->lock);
1747
1748 return ff;
1749}
1750
1751static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1752 struct fuse_inode *fi)
1753{
1754 struct fuse_file *ff = __fuse_write_file_get(fc, fi);
1755 WARN_ON(!ff);
1756 return ff;
1757}
1758
1759int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1760{
1761 struct fuse_conn *fc = get_fuse_conn(inode);
1762 struct fuse_inode *fi = get_fuse_inode(inode);
1763 struct fuse_file *ff;
1764 int err;
1765
1766 ff = __fuse_write_file_get(fc, fi);
1767 err = fuse_flush_times(inode, ff);
1768 if (ff)
1769 fuse_file_put(ff, false, false);
1770
1771 return err;
1772}
1773
1774static struct fuse_writepage_args *fuse_writepage_args_alloc(void)
1775{
1776 struct fuse_writepage_args *wpa;
1777 struct fuse_args_pages *ap;
1778
1779 wpa = kzalloc(sizeof(*wpa), GFP_NOFS);
1780 if (wpa) {
1781 ap = &wpa->ia.ap;
1782 ap->num_pages = 0;
1783 ap->pages = fuse_pages_alloc(1, GFP_NOFS, &ap->descs);
1784 if (!ap->pages) {
1785 kfree(wpa);
1786 wpa = NULL;
1787 }
1788 }
1789 return wpa;
1790
1791}
1792
1793static int fuse_writepage_locked(struct page *page)
1794{
1795 struct address_space *mapping = page->mapping;
1796 struct inode *inode = mapping->host;
1797 struct fuse_conn *fc = get_fuse_conn(inode);
1798 struct fuse_inode *fi = get_fuse_inode(inode);
1799 struct fuse_writepage_args *wpa;
1800 struct fuse_args_pages *ap;
1801 struct page *tmp_page;
1802 int error = -ENOMEM;
1803
1804 set_page_writeback(page);
1805
1806 wpa = fuse_writepage_args_alloc();
1807 if (!wpa)
1808 goto err;
1809 ap = &wpa->ia.ap;
1810
1811 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1812 if (!tmp_page)
1813 goto err_free;
1814
1815 error = -EIO;
1816 wpa->ia.ff = fuse_write_file_get(fc, fi);
1817 if (!wpa->ia.ff)
1818 goto err_nofile;
1819
1820 fuse_write_args_fill(&wpa->ia, wpa->ia.ff, page_offset(page), 0);
1821
1822 copy_highpage(tmp_page, page);
1823 wpa->ia.write.in.write_flags |= FUSE_WRITE_CACHE;
1824 wpa->next = NULL;
1825 ap->args.in_pages = true;
1826 ap->num_pages = 1;
1827 ap->pages[0] = tmp_page;
1828 ap->descs[0].offset = 0;
1829 ap->descs[0].length = PAGE_SIZE;
1830 ap->args.end = fuse_writepage_end;
1831 wpa->inode = inode;
1832
1833 inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1834 inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1835
1836 spin_lock(&fi->lock);
1837 list_add(&wpa->writepages_entry, &fi->writepages);
1838 list_add_tail(&wpa->queue_entry, &fi->queued_writes);
1839 fuse_flush_writepages(inode);
1840 spin_unlock(&fi->lock);
1841
1842 end_page_writeback(page);
1843
1844 return 0;
1845
1846err_nofile:
1847 __free_page(tmp_page);
1848err_free:
1849 kfree(wpa);
1850err:
1851 mapping_set_error(page->mapping, error);
1852 end_page_writeback(page);
1853 return error;
1854}
1855
1856static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1857{
1858 int err;
1859
1860 if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1861 /*
1862 * ->writepages() should be called for sync() and friends. We
1863 * should only get here on direct reclaim and then we are
1864 * allowed to skip a page which is already in flight
1865 */
1866 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1867
1868 redirty_page_for_writepage(wbc, page);
1869 unlock_page(page);
1870 return 0;
1871 }
1872
1873 err = fuse_writepage_locked(page);
1874 unlock_page(page);
1875
1876 return err;
1877}
1878
1879struct fuse_fill_wb_data {
1880 struct fuse_writepage_args *wpa;
1881 struct fuse_file *ff;
1882 struct inode *inode;
1883 struct page **orig_pages;
1884 unsigned int max_pages;
1885};
1886
1887static bool fuse_pages_realloc(struct fuse_fill_wb_data *data)
1888{
1889 struct fuse_args_pages *ap = &data->wpa->ia.ap;
1890 struct fuse_conn *fc = get_fuse_conn(data->inode);
1891 struct page **pages;
1892 struct fuse_page_desc *descs;
1893 unsigned int npages = min_t(unsigned int,
1894 max_t(unsigned int, data->max_pages * 2,
1895 FUSE_DEFAULT_MAX_PAGES_PER_REQ),
1896 fc->max_pages);
1897 WARN_ON(npages <= data->max_pages);
1898
1899 pages = fuse_pages_alloc(npages, GFP_NOFS, &descs);
1900 if (!pages)
1901 return false;
1902
1903 memcpy(pages, ap->pages, sizeof(struct page *) * ap->num_pages);
1904 memcpy(descs, ap->descs, sizeof(struct fuse_page_desc) * ap->num_pages);
1905 kfree(ap->pages);
1906 ap->pages = pages;
1907 ap->descs = descs;
1908 data->max_pages = npages;
1909
1910 return true;
1911}
1912
1913static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1914{
1915 struct fuse_writepage_args *wpa = data->wpa;
1916 struct inode *inode = data->inode;
1917 struct fuse_inode *fi = get_fuse_inode(inode);
1918 int num_pages = wpa->ia.ap.num_pages;
1919 int i;
1920
1921 wpa->ia.ff = fuse_file_get(data->ff);
1922 spin_lock(&fi->lock);
1923 list_add_tail(&wpa->queue_entry, &fi->queued_writes);
1924 fuse_flush_writepages(inode);
1925 spin_unlock(&fi->lock);
1926
1927 for (i = 0; i < num_pages; i++)
1928 end_page_writeback(data->orig_pages[i]);
1929}
1930
1931/*
1932 * First recheck under fi->lock if the offending offset is still under
1933 * writeback. If yes, then iterate auxiliary write requests, to see if there's
1934 * one already added for a page at this offset. If there's none, then insert
1935 * this new request onto the auxiliary list, otherwise reuse the existing one by
1936 * copying the new page contents over to the old temporary page.
1937 */
1938static bool fuse_writepage_in_flight(struct fuse_writepage_args *new_wpa,
1939 struct page *page)
1940{
1941 struct fuse_inode *fi = get_fuse_inode(new_wpa->inode);
1942 struct fuse_writepage_args *tmp;
1943 struct fuse_writepage_args *old_wpa;
1944 struct fuse_args_pages *new_ap = &new_wpa->ia.ap;
1945
1946 WARN_ON(new_ap->num_pages != 0);
1947
1948 spin_lock(&fi->lock);
1949 list_del(&new_wpa->writepages_entry);
1950 old_wpa = fuse_find_writeback(fi, page->index, page->index);
1951 if (!old_wpa) {
1952 list_add(&new_wpa->writepages_entry, &fi->writepages);
1953 spin_unlock(&fi->lock);
1954 return false;
1955 }
1956
1957 new_ap->num_pages = 1;
1958 for (tmp = old_wpa->next; tmp; tmp = tmp->next) {
1959 pgoff_t curr_index;
1960
1961 WARN_ON(tmp->inode != new_wpa->inode);
1962 curr_index = tmp->ia.write.in.offset >> PAGE_SHIFT;
1963 if (curr_index == page->index) {
1964 WARN_ON(tmp->ia.ap.num_pages != 1);
1965 swap(tmp->ia.ap.pages[0], new_ap->pages[0]);
1966 break;
1967 }
1968 }
1969
1970 if (!tmp) {
1971 new_wpa->next = old_wpa->next;
1972 old_wpa->next = new_wpa;
1973 }
1974
1975 spin_unlock(&fi->lock);
1976
1977 if (tmp) {
1978 struct backing_dev_info *bdi = inode_to_bdi(new_wpa->inode);
1979
1980 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1981 dec_node_page_state(new_ap->pages[0], NR_WRITEBACK_TEMP);
1982 wb_writeout_inc(&bdi->wb);
1983 fuse_writepage_free(new_wpa);
1984 }
1985
1986 return true;
1987}
1988
1989static int fuse_writepages_fill(struct page *page,
1990 struct writeback_control *wbc, void *_data)
1991{
1992 struct fuse_fill_wb_data *data = _data;
1993 struct fuse_writepage_args *wpa = data->wpa;
1994 struct fuse_args_pages *ap = &wpa->ia.ap;
1995 struct inode *inode = data->inode;
1996 struct fuse_inode *fi = get_fuse_inode(inode);
1997 struct fuse_conn *fc = get_fuse_conn(inode);
1998 struct page *tmp_page;
1999 bool is_writeback;
2000 int err;
2001
2002 if (!data->ff) {
2003 err = -EIO;
2004 data->ff = fuse_write_file_get(fc, fi);
2005 if (!data->ff)
2006 goto out_unlock;
2007 }
2008
2009 /*
2010 * Being under writeback is unlikely but possible. For example direct
2011 * read to an mmaped fuse file will set the page dirty twice; once when
2012 * the pages are faulted with get_user_pages(), and then after the read
2013 * completed.
2014 */
2015 is_writeback = fuse_page_is_writeback(inode, page->index);
2016
2017 if (wpa && ap->num_pages &&
2018 (is_writeback || ap->num_pages == fc->max_pages ||
2019 (ap->num_pages + 1) * PAGE_SIZE > fc->max_write ||
2020 data->orig_pages[ap->num_pages - 1]->index + 1 != page->index)) {
2021 fuse_writepages_send(data);
2022 data->wpa = NULL;
2023 } else if (wpa && ap->num_pages == data->max_pages) {
2024 if (!fuse_pages_realloc(data)) {
2025 fuse_writepages_send(data);
2026 data->wpa = NULL;
2027 }
2028 }
2029
2030 err = -ENOMEM;
2031 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2032 if (!tmp_page)
2033 goto out_unlock;
2034
2035 /*
2036 * The page must not be redirtied until the writeout is completed
2037 * (i.e. userspace has sent a reply to the write request). Otherwise
2038 * there could be more than one temporary page instance for each real
2039 * page.
2040 *
2041 * This is ensured by holding the page lock in page_mkwrite() while
2042 * checking fuse_page_is_writeback(). We already hold the page lock
2043 * since clear_page_dirty_for_io() and keep it held until we add the
2044 * request to the fi->writepages list and increment ap->num_pages.
2045 * After this fuse_page_is_writeback() will indicate that the page is
2046 * under writeback, so we can release the page lock.
2047 */
2048 if (data->wpa == NULL) {
2049 err = -ENOMEM;
2050 wpa = fuse_writepage_args_alloc();
2051 if (!wpa) {
2052 __free_page(tmp_page);
2053 goto out_unlock;
2054 }
2055 data->max_pages = 1;
2056
2057 ap = &wpa->ia.ap;
2058 fuse_write_args_fill(&wpa->ia, data->ff, page_offset(page), 0);
2059 wpa->ia.write.in.write_flags |= FUSE_WRITE_CACHE;
2060 wpa->next = NULL;
2061 ap->args.in_pages = true;
2062 ap->args.end = fuse_writepage_end;
2063 ap->num_pages = 0;
2064 wpa->inode = inode;
2065
2066 spin_lock(&fi->lock);
2067 list_add(&wpa->writepages_entry, &fi->writepages);
2068 spin_unlock(&fi->lock);
2069
2070 data->wpa = wpa;
2071 }
2072 set_page_writeback(page);
2073
2074 copy_highpage(tmp_page, page);
2075 ap->pages[ap->num_pages] = tmp_page;
2076 ap->descs[ap->num_pages].offset = 0;
2077 ap->descs[ap->num_pages].length = PAGE_SIZE;
2078
2079 inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
2080 inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
2081
2082 err = 0;
2083 if (is_writeback && fuse_writepage_in_flight(wpa, page)) {
2084 end_page_writeback(page);
2085 data->wpa = NULL;
2086 goto out_unlock;
2087 }
2088 data->orig_pages[ap->num_pages] = page;
2089
2090 /*
2091 * Protected by fi->lock against concurrent access by
2092 * fuse_page_is_writeback().
2093 */
2094 spin_lock(&fi->lock);
2095 ap->num_pages++;
2096 spin_unlock(&fi->lock);
2097
2098out_unlock:
2099 unlock_page(page);
2100
2101 return err;
2102}
2103
2104static int fuse_writepages(struct address_space *mapping,
2105 struct writeback_control *wbc)
2106{
2107 struct inode *inode = mapping->host;
2108 struct fuse_conn *fc = get_fuse_conn(inode);
2109 struct fuse_fill_wb_data data;
2110 int err;
2111
2112 err = -EIO;
2113 if (is_bad_inode(inode))
2114 goto out;
2115
2116 data.inode = inode;
2117 data.wpa = NULL;
2118 data.ff = NULL;
2119
2120 err = -ENOMEM;
2121 data.orig_pages = kcalloc(fc->max_pages,
2122 sizeof(struct page *),
2123 GFP_NOFS);
2124 if (!data.orig_pages)
2125 goto out;
2126
2127 err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
2128 if (data.wpa) {
2129 /* Ignore errors if we can write at least one page */
2130 WARN_ON(!data.wpa->ia.ap.num_pages);
2131 fuse_writepages_send(&data);
2132 err = 0;
2133 }
2134 if (data.ff)
2135 fuse_file_put(data.ff, false, false);
2136
2137 kfree(data.orig_pages);
2138out:
2139 return err;
2140}
2141
2142/*
2143 * It's worthy to make sure that space is reserved on disk for the write,
2144 * but how to implement it without killing performance need more thinking.
2145 */
2146static int fuse_write_begin(struct file *file, struct address_space *mapping,
2147 loff_t pos, unsigned len, unsigned flags,
2148 struct page **pagep, void **fsdata)
2149{
2150 pgoff_t index = pos >> PAGE_SHIFT;
2151 struct fuse_conn *fc = get_fuse_conn(file_inode(file));
2152 struct page *page;
2153 loff_t fsize;
2154 int err = -ENOMEM;
2155
2156 WARN_ON(!fc->writeback_cache);
2157
2158 page = grab_cache_page_write_begin(mapping, index, flags);
2159 if (!page)
2160 goto error;
2161
2162 fuse_wait_on_page_writeback(mapping->host, page->index);
2163
2164 if (PageUptodate(page) || len == PAGE_SIZE)
2165 goto success;
2166 /*
2167 * Check if the start this page comes after the end of file, in which
2168 * case the readpage can be optimized away.
2169 */
2170 fsize = i_size_read(mapping->host);
2171 if (fsize <= (pos & PAGE_MASK)) {
2172 size_t off = pos & ~PAGE_MASK;
2173 if (off)
2174 zero_user_segment(page, 0, off);
2175 goto success;
2176 }
2177 err = fuse_do_readpage(file, page);
2178 if (err)
2179 goto cleanup;
2180success:
2181 *pagep = page;
2182 return 0;
2183
2184cleanup:
2185 unlock_page(page);
2186 put_page(page);
2187error:
2188 return err;
2189}
2190
2191static int fuse_write_end(struct file *file, struct address_space *mapping,
2192 loff_t pos, unsigned len, unsigned copied,
2193 struct page *page, void *fsdata)
2194{
2195 struct inode *inode = page->mapping->host;
2196
2197 /* Haven't copied anything? Skip zeroing, size extending, dirtying. */
2198 if (!copied)
2199 goto unlock;
2200
2201 if (!PageUptodate(page)) {
2202 /* Zero any unwritten bytes at the end of the page */
2203 size_t endoff = (pos + copied) & ~PAGE_MASK;
2204 if (endoff)
2205 zero_user_segment(page, endoff, PAGE_SIZE);
2206 SetPageUptodate(page);
2207 }
2208
2209 fuse_write_update_size(inode, pos + copied);
2210 set_page_dirty(page);
2211
2212unlock:
2213 unlock_page(page);
2214 put_page(page);
2215
2216 return copied;
2217}
2218
2219static int fuse_launder_page(struct page *page)
2220{
2221 int err = 0;
2222 if (clear_page_dirty_for_io(page)) {
2223 struct inode *inode = page->mapping->host;
2224 err = fuse_writepage_locked(page);
2225 if (!err)
2226 fuse_wait_on_page_writeback(inode, page->index);
2227 }
2228 return err;
2229}
2230
2231/*
2232 * Write back dirty pages now, because there may not be any suitable
2233 * open files later
2234 */
2235static void fuse_vma_close(struct vm_area_struct *vma)
2236{
2237 filemap_write_and_wait(vma->vm_file->f_mapping);
2238}
2239
2240/*
2241 * Wait for writeback against this page to complete before allowing it
2242 * to be marked dirty again, and hence written back again, possibly
2243 * before the previous writepage completed.
2244 *
2245 * Block here, instead of in ->writepage(), so that the userspace fs
2246 * can only block processes actually operating on the filesystem.
2247 *
2248 * Otherwise unprivileged userspace fs would be able to block
2249 * unrelated:
2250 *
2251 * - page migration
2252 * - sync(2)
2253 * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2254 */
2255static vm_fault_t fuse_page_mkwrite(struct vm_fault *vmf)
2256{
2257 struct page *page = vmf->page;
2258 struct inode *inode = file_inode(vmf->vma->vm_file);
2259
2260 file_update_time(vmf->vma->vm_file);
2261 lock_page(page);
2262 if (page->mapping != inode->i_mapping) {
2263 unlock_page(page);
2264 return VM_FAULT_NOPAGE;
2265 }
2266
2267 fuse_wait_on_page_writeback(inode, page->index);
2268 return VM_FAULT_LOCKED;
2269}
2270
2271static const struct vm_operations_struct fuse_file_vm_ops = {
2272 .close = fuse_vma_close,
2273 .fault = filemap_fault,
2274 .map_pages = filemap_map_pages,
2275 .page_mkwrite = fuse_page_mkwrite,
2276};
2277
2278static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2279{
2280 struct fuse_file *ff = file->private_data;
2281
2282 if (ff->open_flags & FOPEN_DIRECT_IO) {
2283 /* Can't provide the coherency needed for MAP_SHARED */
2284 if (vma->vm_flags & VM_MAYSHARE)
2285 return -ENODEV;
2286
2287 invalidate_inode_pages2(file->f_mapping);
2288
2289 return generic_file_mmap(file, vma);
2290 }
2291
2292 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2293 fuse_link_write_file(file);
2294
2295 file_accessed(file);
2296 vma->vm_ops = &fuse_file_vm_ops;
2297 return 0;
2298}
2299
2300static int convert_fuse_file_lock(struct fuse_conn *fc,
2301 const struct fuse_file_lock *ffl,
2302 struct file_lock *fl)
2303{
2304 switch (ffl->type) {
2305 case F_UNLCK:
2306 break;
2307
2308 case F_RDLCK:
2309 case F_WRLCK:
2310 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2311 ffl->end < ffl->start)
2312 return -EIO;
2313
2314 fl->fl_start = ffl->start;
2315 fl->fl_end = ffl->end;
2316
2317 /*
2318 * Convert pid into init's pid namespace. The locks API will
2319 * translate it into the caller's pid namespace.
2320 */
2321 rcu_read_lock();
2322 fl->fl_pid = pid_nr_ns(find_pid_ns(ffl->pid, fc->pid_ns), &init_pid_ns);
2323 rcu_read_unlock();
2324 break;
2325
2326 default:
2327 return -EIO;
2328 }
2329 fl->fl_type = ffl->type;
2330 return 0;
2331}
2332
2333static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2334 const struct file_lock *fl, int opcode, pid_t pid,
2335 int flock, struct fuse_lk_in *inarg)
2336{
2337 struct inode *inode = file_inode(file);
2338 struct fuse_conn *fc = get_fuse_conn(inode);
2339 struct fuse_file *ff = file->private_data;
2340
2341 memset(inarg, 0, sizeof(*inarg));
2342 inarg->fh = ff->fh;
2343 inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2344 inarg->lk.start = fl->fl_start;
2345 inarg->lk.end = fl->fl_end;
2346 inarg->lk.type = fl->fl_type;
2347 inarg->lk.pid = pid;
2348 if (flock)
2349 inarg->lk_flags |= FUSE_LK_FLOCK;
2350 args->opcode = opcode;
2351 args->nodeid = get_node_id(inode);
2352 args->in_numargs = 1;
2353 args->in_args[0].size = sizeof(*inarg);
2354 args->in_args[0].value = inarg;
2355}
2356
2357static int fuse_getlk(struct file *file, struct file_lock *fl)
2358{
2359 struct inode *inode = file_inode(file);
2360 struct fuse_conn *fc = get_fuse_conn(inode);
2361 FUSE_ARGS(args);
2362 struct fuse_lk_in inarg;
2363 struct fuse_lk_out outarg;
2364 int err;
2365
2366 fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2367 args.out_numargs = 1;
2368 args.out_args[0].size = sizeof(outarg);
2369 args.out_args[0].value = &outarg;
2370 err = fuse_simple_request(fc, &args);
2371 if (!err)
2372 err = convert_fuse_file_lock(fc, &outarg.lk, fl);
2373
2374 return err;
2375}
2376
2377static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2378{
2379 struct inode *inode = file_inode(file);
2380 struct fuse_conn *fc = get_fuse_conn(inode);
2381 FUSE_ARGS(args);
2382 struct fuse_lk_in inarg;
2383 int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2384 struct pid *pid = fl->fl_type != F_UNLCK ? task_tgid(current) : NULL;
2385 pid_t pid_nr = pid_nr_ns(pid, fc->pid_ns);
2386 int err;
2387
2388 if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2389 /* NLM needs asynchronous locks, which we don't support yet */
2390 return -ENOLCK;
2391 }
2392
2393 /* Unlock on close is handled by the flush method */
2394 if ((fl->fl_flags & FL_CLOSE_POSIX) == FL_CLOSE_POSIX)
2395 return 0;
2396
2397 fuse_lk_fill(&args, file, fl, opcode, pid_nr, flock, &inarg);
2398 err = fuse_simple_request(fc, &args);
2399
2400 /* locking is restartable */
2401 if (err == -EINTR)
2402 err = -ERESTARTSYS;
2403
2404 return err;
2405}
2406
2407static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2408{
2409 struct inode *inode = file_inode(file);
2410 struct fuse_conn *fc = get_fuse_conn(inode);
2411 int err;
2412
2413 if (cmd == F_CANCELLK) {
2414 err = 0;
2415 } else if (cmd == F_GETLK) {
2416 if (fc->no_lock) {
2417 posix_test_lock(file, fl);
2418 err = 0;
2419 } else
2420 err = fuse_getlk(file, fl);
2421 } else {
2422 if (fc->no_lock)
2423 err = posix_lock_file(file, fl, NULL);
2424 else
2425 err = fuse_setlk(file, fl, 0);
2426 }
2427 return err;
2428}
2429
2430static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2431{
2432 struct inode *inode = file_inode(file);
2433 struct fuse_conn *fc = get_fuse_conn(inode);
2434 int err;
2435
2436 if (fc->no_flock) {
2437 err = locks_lock_file_wait(file, fl);
2438 } else {
2439 struct fuse_file *ff = file->private_data;
2440
2441 /* emulate flock with POSIX locks */
2442 ff->flock = true;
2443 err = fuse_setlk(file, fl, 1);
2444 }
2445
2446 return err;
2447}
2448
2449static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2450{
2451 struct inode *inode = mapping->host;
2452 struct fuse_conn *fc = get_fuse_conn(inode);
2453 FUSE_ARGS(args);
2454 struct fuse_bmap_in inarg;
2455 struct fuse_bmap_out outarg;
2456 int err;
2457
2458 if (!inode->i_sb->s_bdev || fc->no_bmap)
2459 return 0;
2460
2461 memset(&inarg, 0, sizeof(inarg));
2462 inarg.block = block;
2463 inarg.blocksize = inode->i_sb->s_blocksize;
2464 args.opcode = FUSE_BMAP;
2465 args.nodeid = get_node_id(inode);
2466 args.in_numargs = 1;
2467 args.in_args[0].size = sizeof(inarg);
2468 args.in_args[0].value = &inarg;
2469 args.out_numargs = 1;
2470 args.out_args[0].size = sizeof(outarg);
2471 args.out_args[0].value = &outarg;
2472 err = fuse_simple_request(fc, &args);
2473 if (err == -ENOSYS)
2474 fc->no_bmap = 1;
2475
2476 return err ? 0 : outarg.block;
2477}
2478
2479static loff_t fuse_lseek(struct file *file, loff_t offset, int whence)
2480{
2481 struct inode *inode = file->f_mapping->host;
2482 struct fuse_conn *fc = get_fuse_conn(inode);
2483 struct fuse_file *ff = file->private_data;
2484 FUSE_ARGS(args);
2485 struct fuse_lseek_in inarg = {
2486 .fh = ff->fh,
2487 .offset = offset,
2488 .whence = whence
2489 };
2490 struct fuse_lseek_out outarg;
2491 int err;
2492
2493 if (fc->no_lseek)
2494 goto fallback;
2495
2496 args.opcode = FUSE_LSEEK;
2497 args.nodeid = ff->nodeid;
2498 args.in_numargs = 1;
2499 args.in_args[0].size = sizeof(inarg);
2500 args.in_args[0].value = &inarg;
2501 args.out_numargs = 1;
2502 args.out_args[0].size = sizeof(outarg);
2503 args.out_args[0].value = &outarg;
2504 err = fuse_simple_request(fc, &args);
2505 if (err) {
2506 if (err == -ENOSYS) {
2507 fc->no_lseek = 1;
2508 goto fallback;
2509 }
2510 return err;
2511 }
2512
2513 return vfs_setpos(file, outarg.offset, inode->i_sb->s_maxbytes);
2514
2515fallback:
2516 err = fuse_update_attributes(inode, file);
2517 if (!err)
2518 return generic_file_llseek(file, offset, whence);
2519 else
2520 return err;
2521}
2522
2523static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2524{
2525 loff_t retval;
2526 struct inode *inode = file_inode(file);
2527
2528 switch (whence) {
2529 case SEEK_SET:
2530 case SEEK_CUR:
2531 /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2532 retval = generic_file_llseek(file, offset, whence);
2533 break;
2534 case SEEK_END:
2535 inode_lock(inode);
2536 retval = fuse_update_attributes(inode, file);
2537 if (!retval)
2538 retval = generic_file_llseek(file, offset, whence);
2539 inode_unlock(inode);
2540 break;
2541 case SEEK_HOLE:
2542 case SEEK_DATA:
2543 inode_lock(inode);
2544 retval = fuse_lseek(file, offset, whence);
2545 inode_unlock(inode);
2546 break;
2547 default:
2548 retval = -EINVAL;
2549 }
2550
2551 return retval;
2552}
2553
2554/*
2555 * CUSE servers compiled on 32bit broke on 64bit kernels because the
2556 * ABI was defined to be 'struct iovec' which is different on 32bit
2557 * and 64bit. Fortunately we can determine which structure the server
2558 * used from the size of the reply.
2559 */
2560static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2561 size_t transferred, unsigned count,
2562 bool is_compat)
2563{
2564#ifdef CONFIG_COMPAT
2565 if (count * sizeof(struct compat_iovec) == transferred) {
2566 struct compat_iovec *ciov = src;
2567 unsigned i;
2568
2569 /*
2570 * With this interface a 32bit server cannot support
2571 * non-compat (i.e. ones coming from 64bit apps) ioctl
2572 * requests
2573 */
2574 if (!is_compat)
2575 return -EINVAL;
2576
2577 for (i = 0; i < count; i++) {
2578 dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2579 dst[i].iov_len = ciov[i].iov_len;
2580 }
2581 return 0;
2582 }
2583#endif
2584
2585 if (count * sizeof(struct iovec) != transferred)
2586 return -EIO;
2587
2588 memcpy(dst, src, transferred);
2589 return 0;
2590}
2591
2592/* Make sure iov_length() won't overflow */
2593static int fuse_verify_ioctl_iov(struct fuse_conn *fc, struct iovec *iov,
2594 size_t count)
2595{
2596 size_t n;
2597 u32 max = fc->max_pages << PAGE_SHIFT;
2598
2599 for (n = 0; n < count; n++, iov++) {
2600 if (iov->iov_len > (size_t) max)
2601 return -ENOMEM;
2602 max -= iov->iov_len;
2603 }
2604 return 0;
2605}
2606
2607static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2608 void *src, size_t transferred, unsigned count,
2609 bool is_compat)
2610{
2611 unsigned i;
2612 struct fuse_ioctl_iovec *fiov = src;
2613
2614 if (fc->minor < 16) {
2615 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2616 count, is_compat);
2617 }
2618
2619 if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2620 return -EIO;
2621
2622 for (i = 0; i < count; i++) {
2623 /* Did the server supply an inappropriate value? */
2624 if (fiov[i].base != (unsigned long) fiov[i].base ||
2625 fiov[i].len != (unsigned long) fiov[i].len)
2626 return -EIO;
2627
2628 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2629 dst[i].iov_len = (size_t) fiov[i].len;
2630
2631#ifdef CONFIG_COMPAT
2632 if (is_compat &&
2633 (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2634 (compat_size_t) dst[i].iov_len != fiov[i].len))
2635 return -EIO;
2636#endif
2637 }
2638
2639 return 0;
2640}
2641
2642
2643/*
2644 * For ioctls, there is no generic way to determine how much memory
2645 * needs to be read and/or written. Furthermore, ioctls are allowed
2646 * to dereference the passed pointer, so the parameter requires deep
2647 * copying but FUSE has no idea whatsoever about what to copy in or
2648 * out.
2649 *
2650 * This is solved by allowing FUSE server to retry ioctl with
2651 * necessary in/out iovecs. Let's assume the ioctl implementation
2652 * needs to read in the following structure.
2653 *
2654 * struct a {
2655 * char *buf;
2656 * size_t buflen;
2657 * }
2658 *
2659 * On the first callout to FUSE server, inarg->in_size and
2660 * inarg->out_size will be NULL; then, the server completes the ioctl
2661 * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2662 * the actual iov array to
2663 *
2664 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) } }
2665 *
2666 * which tells FUSE to copy in the requested area and retry the ioctl.
2667 * On the second round, the server has access to the structure and
2668 * from that it can tell what to look for next, so on the invocation,
2669 * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2670 *
2671 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) },
2672 * { .iov_base = a.buf, .iov_len = a.buflen } }
2673 *
2674 * FUSE will copy both struct a and the pointed buffer from the
2675 * process doing the ioctl and retry ioctl with both struct a and the
2676 * buffer.
2677 *
2678 * This time, FUSE server has everything it needs and completes ioctl
2679 * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2680 *
2681 * Copying data out works the same way.
2682 *
2683 * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2684 * automatically initializes in and out iovs by decoding @cmd with
2685 * _IOC_* macros and the server is not allowed to request RETRY. This
2686 * limits ioctl data transfers to well-formed ioctls and is the forced
2687 * behavior for all FUSE servers.
2688 */
2689long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2690 unsigned int flags)
2691{
2692 struct fuse_file *ff = file->private_data;
2693 struct fuse_conn *fc = ff->fc;
2694 struct fuse_ioctl_in inarg = {
2695 .fh = ff->fh,
2696 .cmd = cmd,
2697 .arg = arg,
2698 .flags = flags
2699 };
2700 struct fuse_ioctl_out outarg;
2701 struct iovec *iov_page = NULL;
2702 struct iovec *in_iov = NULL, *out_iov = NULL;
2703 unsigned int in_iovs = 0, out_iovs = 0, max_pages;
2704 size_t in_size, out_size, c;
2705 ssize_t transferred;
2706 int err, i;
2707 struct iov_iter ii;
2708 struct fuse_args_pages ap = {};
2709
2710#if BITS_PER_LONG == 32
2711 inarg.flags |= FUSE_IOCTL_32BIT;
2712#else
2713 if (flags & FUSE_IOCTL_COMPAT) {
2714 inarg.flags |= FUSE_IOCTL_32BIT;
2715#ifdef CONFIG_X86_X32
2716 if (in_x32_syscall())
2717 inarg.flags |= FUSE_IOCTL_COMPAT_X32;
2718#endif
2719 }
2720#endif
2721
2722 /* assume all the iovs returned by client always fits in a page */
2723 BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2724
2725 err = -ENOMEM;
2726 ap.pages = fuse_pages_alloc(fc->max_pages, GFP_KERNEL, &ap.descs);
2727 iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2728 if (!ap.pages || !iov_page)
2729 goto out;
2730
2731 fuse_page_descs_length_init(ap.descs, 0, fc->max_pages);
2732
2733 /*
2734 * If restricted, initialize IO parameters as encoded in @cmd.
2735 * RETRY from server is not allowed.
2736 */
2737 if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2738 struct iovec *iov = iov_page;
2739
2740 iov->iov_base = (void __user *)arg;
2741 iov->iov_len = _IOC_SIZE(cmd);
2742
2743 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2744 in_iov = iov;
2745 in_iovs = 1;
2746 }
2747
2748 if (_IOC_DIR(cmd) & _IOC_READ) {
2749 out_iov = iov;
2750 out_iovs = 1;
2751 }
2752 }
2753
2754 retry:
2755 inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2756 inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2757
2758 /*
2759 * Out data can be used either for actual out data or iovs,
2760 * make sure there always is at least one page.
2761 */
2762 out_size = max_t(size_t, out_size, PAGE_SIZE);
2763 max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2764
2765 /* make sure there are enough buffer pages and init request with them */
2766 err = -ENOMEM;
2767 if (max_pages > fc->max_pages)
2768 goto out;
2769 while (ap.num_pages < max_pages) {
2770 ap.pages[ap.num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2771 if (!ap.pages[ap.num_pages])
2772 goto out;
2773 ap.num_pages++;
2774 }
2775
2776
2777 /* okay, let's send it to the client */
2778 ap.args.opcode = FUSE_IOCTL;
2779 ap.args.nodeid = ff->nodeid;
2780 ap.args.in_numargs = 1;
2781 ap.args.in_args[0].size = sizeof(inarg);
2782 ap.args.in_args[0].value = &inarg;
2783 if (in_size) {
2784 ap.args.in_numargs++;
2785 ap.args.in_args[1].size = in_size;
2786 ap.args.in_pages = true;
2787
2788 err = -EFAULT;
2789 iov_iter_init(&ii, WRITE, in_iov, in_iovs, in_size);
2790 for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= ap.num_pages); i++) {
2791 c = copy_page_from_iter(ap.pages[i], 0, PAGE_SIZE, &ii);
2792 if (c != PAGE_SIZE && iov_iter_count(&ii))
2793 goto out;
2794 }
2795 }
2796
2797 ap.args.out_numargs = 2;
2798 ap.args.out_args[0].size = sizeof(outarg);
2799 ap.args.out_args[0].value = &outarg;
2800 ap.args.out_args[1].size = out_size;
2801 ap.args.out_pages = true;
2802 ap.args.out_argvar = true;
2803
2804 transferred = fuse_simple_request(fc, &ap.args);
2805 err = transferred;
2806 if (transferred < 0)
2807 goto out;
2808
2809 /* did it ask for retry? */
2810 if (outarg.flags & FUSE_IOCTL_RETRY) {
2811 void *vaddr;
2812
2813 /* no retry if in restricted mode */
2814 err = -EIO;
2815 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2816 goto out;
2817
2818 in_iovs = outarg.in_iovs;
2819 out_iovs = outarg.out_iovs;
2820
2821 /*
2822 * Make sure things are in boundary, separate checks
2823 * are to protect against overflow.
2824 */
2825 err = -ENOMEM;
2826 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2827 out_iovs > FUSE_IOCTL_MAX_IOV ||
2828 in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2829 goto out;
2830
2831 vaddr = kmap_atomic(ap.pages[0]);
2832 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2833 transferred, in_iovs + out_iovs,
2834 (flags & FUSE_IOCTL_COMPAT) != 0);
2835 kunmap_atomic(vaddr);
2836 if (err)
2837 goto out;
2838
2839 in_iov = iov_page;
2840 out_iov = in_iov + in_iovs;
2841
2842 err = fuse_verify_ioctl_iov(fc, in_iov, in_iovs);
2843 if (err)
2844 goto out;
2845
2846 err = fuse_verify_ioctl_iov(fc, out_iov, out_iovs);
2847 if (err)
2848 goto out;
2849
2850 goto retry;
2851 }
2852
2853 err = -EIO;
2854 if (transferred > inarg.out_size)
2855 goto out;
2856
2857 err = -EFAULT;
2858 iov_iter_init(&ii, READ, out_iov, out_iovs, transferred);
2859 for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= ap.num_pages); i++) {
2860 c = copy_page_to_iter(ap.pages[i], 0, PAGE_SIZE, &ii);
2861 if (c != PAGE_SIZE && iov_iter_count(&ii))
2862 goto out;
2863 }
2864 err = 0;
2865 out:
2866 free_page((unsigned long) iov_page);
2867 while (ap.num_pages)
2868 __free_page(ap.pages[--ap.num_pages]);
2869 kfree(ap.pages);
2870
2871 return err ? err : outarg.result;
2872}
2873EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2874
2875long fuse_ioctl_common(struct file *file, unsigned int cmd,
2876 unsigned long arg, unsigned int flags)
2877{
2878 struct inode *inode = file_inode(file);
2879 struct fuse_conn *fc = get_fuse_conn(inode);
2880
2881 if (!fuse_allow_current_process(fc))
2882 return -EACCES;
2883
2884 if (is_bad_inode(inode))
2885 return -EIO;
2886
2887 return fuse_do_ioctl(file, cmd, arg, flags);
2888}
2889
2890static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2891 unsigned long arg)
2892{
2893 return fuse_ioctl_common(file, cmd, arg, 0);
2894}
2895
2896static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2897 unsigned long arg)
2898{
2899 return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2900}
2901
2902/*
2903 * All files which have been polled are linked to RB tree
2904 * fuse_conn->polled_files which is indexed by kh. Walk the tree and
2905 * find the matching one.
2906 */
2907static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2908 struct rb_node **parent_out)
2909{
2910 struct rb_node **link = &fc->polled_files.rb_node;
2911 struct rb_node *last = NULL;
2912
2913 while (*link) {
2914 struct fuse_file *ff;
2915
2916 last = *link;
2917 ff = rb_entry(last, struct fuse_file, polled_node);
2918
2919 if (kh < ff->kh)
2920 link = &last->rb_left;
2921 else if (kh > ff->kh)
2922 link = &last->rb_right;
2923 else
2924 return link;
2925 }
2926
2927 if (parent_out)
2928 *parent_out = last;
2929 return link;
2930}
2931
2932/*
2933 * The file is about to be polled. Make sure it's on the polled_files
2934 * RB tree. Note that files once added to the polled_files tree are
2935 * not removed before the file is released. This is because a file
2936 * polled once is likely to be polled again.
2937 */
2938static void fuse_register_polled_file(struct fuse_conn *fc,
2939 struct fuse_file *ff)
2940{
2941 spin_lock(&fc->lock);
2942 if (RB_EMPTY_NODE(&ff->polled_node)) {
2943 struct rb_node **link, *uninitialized_var(parent);
2944
2945 link = fuse_find_polled_node(fc, ff->kh, &parent);
2946 BUG_ON(*link);
2947 rb_link_node(&ff->polled_node, parent, link);
2948 rb_insert_color(&ff->polled_node, &fc->polled_files);
2949 }
2950 spin_unlock(&fc->lock);
2951}
2952
2953__poll_t fuse_file_poll(struct file *file, poll_table *wait)
2954{
2955 struct fuse_file *ff = file->private_data;
2956 struct fuse_conn *fc = ff->fc;
2957 struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2958 struct fuse_poll_out outarg;
2959 FUSE_ARGS(args);
2960 int err;
2961
2962 if (fc->no_poll)
2963 return DEFAULT_POLLMASK;
2964
2965 poll_wait(file, &ff->poll_wait, wait);
2966 inarg.events = mangle_poll(poll_requested_events(wait));
2967
2968 /*
2969 * Ask for notification iff there's someone waiting for it.
2970 * The client may ignore the flag and always notify.
2971 */
2972 if (waitqueue_active(&ff->poll_wait)) {
2973 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2974 fuse_register_polled_file(fc, ff);
2975 }
2976
2977 args.opcode = FUSE_POLL;
2978 args.nodeid = ff->nodeid;
2979 args.in_numargs = 1;
2980 args.in_args[0].size = sizeof(inarg);
2981 args.in_args[0].value = &inarg;
2982 args.out_numargs = 1;
2983 args.out_args[0].size = sizeof(outarg);
2984 args.out_args[0].value = &outarg;
2985 err = fuse_simple_request(fc, &args);
2986
2987 if (!err)
2988 return demangle_poll(outarg.revents);
2989 if (err == -ENOSYS) {
2990 fc->no_poll = 1;
2991 return DEFAULT_POLLMASK;
2992 }
2993 return EPOLLERR;
2994}
2995EXPORT_SYMBOL_GPL(fuse_file_poll);
2996
2997/*
2998 * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2999 * wakes up the poll waiters.
3000 */
3001int fuse_notify_poll_wakeup(struct fuse_conn *fc,
3002 struct fuse_notify_poll_wakeup_out *outarg)
3003{
3004 u64 kh = outarg->kh;
3005 struct rb_node **link;
3006
3007 spin_lock(&fc->lock);
3008
3009 link = fuse_find_polled_node(fc, kh, NULL);
3010 if (*link) {
3011 struct fuse_file *ff;
3012
3013 ff = rb_entry(*link, struct fuse_file, polled_node);
3014 wake_up_interruptible_sync(&ff->poll_wait);
3015 }
3016
3017 spin_unlock(&fc->lock);
3018 return 0;
3019}
3020
3021static void fuse_do_truncate(struct file *file)
3022{
3023 struct inode *inode = file->f_mapping->host;
3024 struct iattr attr;
3025
3026 attr.ia_valid = ATTR_SIZE;
3027 attr.ia_size = i_size_read(inode);
3028
3029 attr.ia_file = file;
3030 attr.ia_valid |= ATTR_FILE;
3031
3032 fuse_do_setattr(file_dentry(file), &attr, file);
3033}
3034
3035static inline loff_t fuse_round_up(struct fuse_conn *fc, loff_t off)
3036{
3037 return round_up(off, fc->max_pages << PAGE_SHIFT);
3038}
3039
3040static ssize_t
3041fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
3042{
3043 DECLARE_COMPLETION_ONSTACK(wait);
3044 ssize_t ret = 0;
3045 struct file *file = iocb->ki_filp;
3046 struct fuse_file *ff = file->private_data;
3047 bool async_dio = ff->fc->async_dio;
3048 loff_t pos = 0;
3049 struct inode *inode;
3050 loff_t i_size;
3051 size_t count = iov_iter_count(iter);
3052 loff_t offset = iocb->ki_pos;
3053 struct fuse_io_priv *io;
3054
3055 pos = offset;
3056 inode = file->f_mapping->host;
3057 i_size = i_size_read(inode);
3058
3059 if ((iov_iter_rw(iter) == READ) && (offset > i_size))
3060 return 0;
3061
3062 /* optimization for short read */
3063 if (async_dio && iov_iter_rw(iter) != WRITE && offset + count > i_size) {
3064 if (offset >= i_size)
3065 return 0;
3066 iov_iter_truncate(iter, fuse_round_up(ff->fc, i_size - offset));
3067 count = iov_iter_count(iter);
3068 }
3069
3070 io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
3071 if (!io)
3072 return -ENOMEM;
3073 spin_lock_init(&io->lock);
3074 kref_init(&io->refcnt);
3075 io->reqs = 1;
3076 io->bytes = -1;
3077 io->size = 0;
3078 io->offset = offset;
3079 io->write = (iov_iter_rw(iter) == WRITE);
3080 io->err = 0;
3081 /*
3082 * By default, we want to optimize all I/Os with async request
3083 * submission to the client filesystem if supported.
3084 */
3085 io->async = async_dio;
3086 io->iocb = iocb;
3087 io->blocking = is_sync_kiocb(iocb);
3088
3089 /*
3090 * We cannot asynchronously extend the size of a file.
3091 * In such case the aio will behave exactly like sync io.
3092 */
3093 if ((offset + count > i_size) && iov_iter_rw(iter) == WRITE)
3094 io->blocking = true;
3095
3096 if (io->async && io->blocking) {
3097 /*
3098 * Additional reference to keep io around after
3099 * calling fuse_aio_complete()
3100 */
3101 kref_get(&io->refcnt);
3102 io->done = &wait;
3103 }
3104
3105 if (iov_iter_rw(iter) == WRITE) {
3106 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
3107 fuse_invalidate_attr(inode);
3108 } else {
3109 ret = __fuse_direct_read(io, iter, &pos);
3110 }
3111
3112 if (io->async) {
3113 bool blocking = io->blocking;
3114
3115 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
3116
3117 /* we have a non-extending, async request, so return */
3118 if (!blocking)
3119 return -EIOCBQUEUED;
3120
3121 wait_for_completion(&wait);
3122 ret = fuse_get_res_by_io(io);
3123 }
3124
3125 kref_put(&io->refcnt, fuse_io_release);
3126
3127 if (iov_iter_rw(iter) == WRITE) {
3128 if (ret > 0)
3129 fuse_write_update_size(inode, pos);
3130 else if (ret < 0 && offset + count > i_size)
3131 fuse_do_truncate(file);
3132 }
3133
3134 return ret;
3135}
3136
3137static int fuse_writeback_range(struct inode *inode, loff_t start, loff_t end)
3138{
3139 int err = filemap_write_and_wait_range(inode->i_mapping, start, end);
3140
3141 if (!err)
3142 fuse_sync_writes(inode);
3143
3144 return err;
3145}
3146
3147static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
3148 loff_t length)
3149{
3150 struct fuse_file *ff = file->private_data;
3151 struct inode *inode = file_inode(file);
3152 struct fuse_inode *fi = get_fuse_inode(inode);
3153 struct fuse_conn *fc = ff->fc;
3154 FUSE_ARGS(args);
3155 struct fuse_fallocate_in inarg = {
3156 .fh = ff->fh,
3157 .offset = offset,
3158 .length = length,
3159 .mode = mode
3160 };
3161 int err;
3162 bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
3163 (mode & FALLOC_FL_PUNCH_HOLE);
3164
3165 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
3166 return -EOPNOTSUPP;
3167
3168 if (fc->no_fallocate)
3169 return -EOPNOTSUPP;
3170
3171 if (lock_inode) {
3172 inode_lock(inode);
3173 if (mode & FALLOC_FL_PUNCH_HOLE) {
3174 loff_t endbyte = offset + length - 1;
3175
3176 err = fuse_writeback_range(inode, offset, endbyte);
3177 if (err)
3178 goto out;
3179 }
3180 }
3181
3182 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
3183 offset + length > i_size_read(inode)) {
3184 err = inode_newsize_ok(inode, offset + length);
3185 if (err)
3186 goto out;
3187 }
3188
3189 if (!(mode & FALLOC_FL_KEEP_SIZE))
3190 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3191
3192 args.opcode = FUSE_FALLOCATE;
3193 args.nodeid = ff->nodeid;
3194 args.in_numargs = 1;
3195 args.in_args[0].size = sizeof(inarg);
3196 args.in_args[0].value = &inarg;
3197 err = fuse_simple_request(fc, &args);
3198 if (err == -ENOSYS) {
3199 fc->no_fallocate = 1;
3200 err = -EOPNOTSUPP;
3201 }
3202 if (err)
3203 goto out;
3204
3205 /* we could have extended the file */
3206 if (!(mode & FALLOC_FL_KEEP_SIZE)) {
3207 bool changed = fuse_write_update_size(inode, offset + length);
3208
3209 if (changed && fc->writeback_cache)
3210 file_update_time(file);
3211 }
3212
3213 if (mode & FALLOC_FL_PUNCH_HOLE)
3214 truncate_pagecache_range(inode, offset, offset + length - 1);
3215
3216 fuse_invalidate_attr(inode);
3217
3218out:
3219 if (!(mode & FALLOC_FL_KEEP_SIZE))
3220 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3221
3222 if (lock_inode)
3223 inode_unlock(inode);
3224
3225 return err;
3226}
3227
3228static ssize_t __fuse_copy_file_range(struct file *file_in, loff_t pos_in,
3229 struct file *file_out, loff_t pos_out,
3230 size_t len, unsigned int flags)
3231{
3232 struct fuse_file *ff_in = file_in->private_data;
3233 struct fuse_file *ff_out = file_out->private_data;
3234 struct inode *inode_in = file_inode(file_in);
3235 struct inode *inode_out = file_inode(file_out);
3236 struct fuse_inode *fi_out = get_fuse_inode(inode_out);
3237 struct fuse_conn *fc = ff_in->fc;
3238 FUSE_ARGS(args);
3239 struct fuse_copy_file_range_in inarg = {
3240 .fh_in = ff_in->fh,
3241 .off_in = pos_in,
3242 .nodeid_out = ff_out->nodeid,
3243 .fh_out = ff_out->fh,
3244 .off_out = pos_out,
3245 .len = len,
3246 .flags = flags
3247 };
3248 struct fuse_write_out outarg;
3249 ssize_t err;
3250 /* mark unstable when write-back is not used, and file_out gets
3251 * extended */
3252 bool is_unstable = (!fc->writeback_cache) &&
3253 ((pos_out + len) > inode_out->i_size);
3254
3255 if (fc->no_copy_file_range)
3256 return -EOPNOTSUPP;
3257
3258 if (file_inode(file_in)->i_sb != file_inode(file_out)->i_sb)
3259 return -EXDEV;
3260
3261 if (fc->writeback_cache) {
3262 inode_lock(inode_in);
3263 err = fuse_writeback_range(inode_in, pos_in, pos_in + len);
3264 inode_unlock(inode_in);
3265 if (err)
3266 return err;
3267 }
3268
3269 inode_lock(inode_out);
3270
3271 err = file_modified(file_out);
3272 if (err)
3273 goto out;
3274
3275 if (fc->writeback_cache) {
3276 err = fuse_writeback_range(inode_out, pos_out, pos_out + len);
3277 if (err)
3278 goto out;
3279 }
3280
3281 if (is_unstable)
3282 set_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3283
3284 args.opcode = FUSE_COPY_FILE_RANGE;
3285 args.nodeid = ff_in->nodeid;
3286 args.in_numargs = 1;
3287 args.in_args[0].size = sizeof(inarg);
3288 args.in_args[0].value = &inarg;
3289 args.out_numargs = 1;
3290 args.out_args[0].size = sizeof(outarg);
3291 args.out_args[0].value = &outarg;
3292 err = fuse_simple_request(fc, &args);
3293 if (err == -ENOSYS) {
3294 fc->no_copy_file_range = 1;
3295 err = -EOPNOTSUPP;
3296 }
3297 if (err)
3298 goto out;
3299
3300 if (fc->writeback_cache) {
3301 fuse_write_update_size(inode_out, pos_out + outarg.size);
3302 file_update_time(file_out);
3303 }
3304
3305 fuse_invalidate_attr(inode_out);
3306
3307 err = outarg.size;
3308out:
3309 if (is_unstable)
3310 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3311
3312 inode_unlock(inode_out);
3313 file_accessed(file_in);
3314
3315 return err;
3316}
3317
3318static ssize_t fuse_copy_file_range(struct file *src_file, loff_t src_off,
3319 struct file *dst_file, loff_t dst_off,
3320 size_t len, unsigned int flags)
3321{
3322 ssize_t ret;
3323
3324 ret = __fuse_copy_file_range(src_file, src_off, dst_file, dst_off,
3325 len, flags);
3326
3327 if (ret == -EOPNOTSUPP || ret == -EXDEV)
3328 ret = generic_copy_file_range(src_file, src_off, dst_file,
3329 dst_off, len, flags);
3330 return ret;
3331}
3332
3333static const struct file_operations fuse_file_operations = {
3334 .llseek = fuse_file_llseek,
3335 .read_iter = fuse_file_read_iter,
3336 .write_iter = fuse_file_write_iter,
3337 .mmap = fuse_file_mmap,
3338 .open = fuse_open,
3339 .flush = fuse_flush,
3340 .release = fuse_release,
3341 .fsync = fuse_fsync,
3342 .lock = fuse_file_lock,
3343 .flock = fuse_file_flock,
3344 .splice_read = generic_file_splice_read,
3345 .splice_write = iter_file_splice_write,
3346 .unlocked_ioctl = fuse_file_ioctl,
3347 .compat_ioctl = fuse_file_compat_ioctl,
3348 .poll = fuse_file_poll,
3349 .fallocate = fuse_file_fallocate,
3350 .copy_file_range = fuse_copy_file_range,
3351};
3352
3353static const struct address_space_operations fuse_file_aops = {
3354 .readpage = fuse_readpage,
3355 .writepage = fuse_writepage,
3356 .writepages = fuse_writepages,
3357 .launder_page = fuse_launder_page,
3358 .readpages = fuse_readpages,
3359 .set_page_dirty = __set_page_dirty_nobuffers,
3360 .bmap = fuse_bmap,
3361 .direct_IO = fuse_direct_IO,
3362 .write_begin = fuse_write_begin,
3363 .write_end = fuse_write_end,
3364};
3365
3366void fuse_init_file_inode(struct inode *inode)
3367{
3368 struct fuse_inode *fi = get_fuse_inode(inode);
3369
3370 inode->i_fop = &fuse_file_operations;
3371 inode->i_data.a_ops = &fuse_file_aops;
3372
3373 INIT_LIST_HEAD(&fi->write_files);
3374 INIT_LIST_HEAD(&fi->queued_writes);
3375 fi->writectr = 0;
3376 init_waitqueue_head(&fi->page_waitq);
3377 INIT_LIST_HEAD(&fi->writepages);
3378}