Merge tag 'pull-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs
[platform/kernel/linux-starfive.git] / fs / fuse / file.c
1 /*
2   FUSE: Filesystem in Userspace
3   Copyright (C) 2001-2008  Miklos Szeredi <miklos@szeredi.hu>
4
5   This program can be distributed under the terms of the GNU GPL.
6   See the file COPYING.
7 */
8
9 #include "fuse_i.h"
10
11 #include <linux/pagemap.h>
12 #include <linux/slab.h>
13 #include <linux/kernel.h>
14 #include <linux/sched.h>
15 #include <linux/sched/signal.h>
16 #include <linux/module.h>
17 #include <linux/swap.h>
18 #include <linux/falloc.h>
19 #include <linux/uio.h>
20 #include <linux/fs.h>
21
22 static int fuse_send_open(struct fuse_mount *fm, u64 nodeid,
23                           unsigned int open_flags, int opcode,
24                           struct fuse_open_out *outargp)
25 {
26         struct fuse_open_in inarg;
27         FUSE_ARGS(args);
28
29         memset(&inarg, 0, sizeof(inarg));
30         inarg.flags = open_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
31         if (!fm->fc->atomic_o_trunc)
32                 inarg.flags &= ~O_TRUNC;
33
34         if (fm->fc->handle_killpriv_v2 &&
35             (inarg.flags & O_TRUNC) && !capable(CAP_FSETID)) {
36                 inarg.open_flags |= FUSE_OPEN_KILL_SUIDGID;
37         }
38
39         args.opcode = opcode;
40         args.nodeid = nodeid;
41         args.in_numargs = 1;
42         args.in_args[0].size = sizeof(inarg);
43         args.in_args[0].value = &inarg;
44         args.out_numargs = 1;
45         args.out_args[0].size = sizeof(*outargp);
46         args.out_args[0].value = outargp;
47
48         return fuse_simple_request(fm, &args);
49 }
50
51 struct fuse_release_args {
52         struct fuse_args args;
53         struct fuse_release_in inarg;
54         struct inode *inode;
55 };
56
57 struct fuse_file *fuse_file_alloc(struct fuse_mount *fm)
58 {
59         struct fuse_file *ff;
60
61         ff = kzalloc(sizeof(struct fuse_file), GFP_KERNEL_ACCOUNT);
62         if (unlikely(!ff))
63                 return NULL;
64
65         ff->fm = fm;
66         ff->release_args = kzalloc(sizeof(*ff->release_args),
67                                    GFP_KERNEL_ACCOUNT);
68         if (!ff->release_args) {
69                 kfree(ff);
70                 return NULL;
71         }
72
73         INIT_LIST_HEAD(&ff->write_entry);
74         mutex_init(&ff->readdir.lock);
75         refcount_set(&ff->count, 1);
76         RB_CLEAR_NODE(&ff->polled_node);
77         init_waitqueue_head(&ff->poll_wait);
78
79         ff->kh = atomic64_inc_return(&fm->fc->khctr);
80
81         return ff;
82 }
83
84 void fuse_file_free(struct fuse_file *ff)
85 {
86         kfree(ff->release_args);
87         mutex_destroy(&ff->readdir.lock);
88         kfree(ff);
89 }
90
91 static struct fuse_file *fuse_file_get(struct fuse_file *ff)
92 {
93         refcount_inc(&ff->count);
94         return ff;
95 }
96
97 static void fuse_release_end(struct fuse_mount *fm, struct fuse_args *args,
98                              int error)
99 {
100         struct fuse_release_args *ra = container_of(args, typeof(*ra), args);
101
102         iput(ra->inode);
103         kfree(ra);
104 }
105
106 static void fuse_file_put(struct fuse_file *ff, bool sync, bool isdir)
107 {
108         if (refcount_dec_and_test(&ff->count)) {
109                 struct fuse_args *args = &ff->release_args->args;
110
111                 if (isdir ? ff->fm->fc->no_opendir : ff->fm->fc->no_open) {
112                         /* Do nothing when client does not implement 'open' */
113                         fuse_release_end(ff->fm, args, 0);
114                 } else if (sync) {
115                         fuse_simple_request(ff->fm, args);
116                         fuse_release_end(ff->fm, args, 0);
117                 } else {
118                         args->end = fuse_release_end;
119                         if (fuse_simple_background(ff->fm, args,
120                                                    GFP_KERNEL | __GFP_NOFAIL))
121                                 fuse_release_end(ff->fm, args, -ENOTCONN);
122                 }
123                 kfree(ff);
124         }
125 }
126
127 struct fuse_file *fuse_file_open(struct fuse_mount *fm, u64 nodeid,
128                                  unsigned int open_flags, bool isdir)
129 {
130         struct fuse_conn *fc = fm->fc;
131         struct fuse_file *ff;
132         int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
133
134         ff = fuse_file_alloc(fm);
135         if (!ff)
136                 return ERR_PTR(-ENOMEM);
137
138         ff->fh = 0;
139         /* Default for no-open */
140         ff->open_flags = FOPEN_KEEP_CACHE | (isdir ? FOPEN_CACHE_DIR : 0);
141         if (isdir ? !fc->no_opendir : !fc->no_open) {
142                 struct fuse_open_out outarg;
143                 int err;
144
145                 err = fuse_send_open(fm, nodeid, open_flags, opcode, &outarg);
146                 if (!err) {
147                         ff->fh = outarg.fh;
148                         ff->open_flags = outarg.open_flags;
149
150                 } else if (err != -ENOSYS) {
151                         fuse_file_free(ff);
152                         return ERR_PTR(err);
153                 } else {
154                         if (isdir)
155                                 fc->no_opendir = 1;
156                         else
157                                 fc->no_open = 1;
158                 }
159         }
160
161         if (isdir)
162                 ff->open_flags &= ~FOPEN_DIRECT_IO;
163
164         ff->nodeid = nodeid;
165
166         return ff;
167 }
168
169 int fuse_do_open(struct fuse_mount *fm, u64 nodeid, struct file *file,
170                  bool isdir)
171 {
172         struct fuse_file *ff = fuse_file_open(fm, nodeid, file->f_flags, isdir);
173
174         if (!IS_ERR(ff))
175                 file->private_data = ff;
176
177         return PTR_ERR_OR_ZERO(ff);
178 }
179 EXPORT_SYMBOL_GPL(fuse_do_open);
180
181 static void fuse_link_write_file(struct file *file)
182 {
183         struct inode *inode = file_inode(file);
184         struct fuse_inode *fi = get_fuse_inode(inode);
185         struct fuse_file *ff = file->private_data;
186         /*
187          * file may be written through mmap, so chain it onto the
188          * inodes's write_file list
189          */
190         spin_lock(&fi->lock);
191         if (list_empty(&ff->write_entry))
192                 list_add(&ff->write_entry, &fi->write_files);
193         spin_unlock(&fi->lock);
194 }
195
196 void fuse_finish_open(struct inode *inode, struct file *file)
197 {
198         struct fuse_file *ff = file->private_data;
199         struct fuse_conn *fc = get_fuse_conn(inode);
200
201         if (ff->open_flags & FOPEN_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
220 int 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);
267 out_inode_unlock:
268         if (is_wb_truncate || dax_truncate)
269                 inode_unlock(inode);
270
271         return err;
272 }
273
274 static 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
304 void 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
332 void 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
338 static int fuse_open(struct inode *inode, struct file *file)
339 {
340         return fuse_open_common(inode, file, false);
341 }
342
343 static 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
360 void 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 }
371 EXPORT_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  */
377 u64 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
395 struct 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
404 static 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  */
434 static 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
447 static 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  */
458 static 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  */
474 static void fuse_sync_writes(struct inode *inode)
475 {
476         fuse_set_nowrite(inode);
477         fuse_release_nowrite(inode);
478 }
479
480 static 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
527 inval_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
537 int 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
557 static 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         }
601 out:
602         inode_unlock(inode);
603
604         return err;
605 }
606
607 void 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
627 static 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
639 static void fuse_io_release(struct kref *kref)
640 {
641         kfree(container_of(kref, struct fuse_io_priv, refcnt));
642 }
643
644 static 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  */
671 static 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
705 static 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
723 static void fuse_io_free(struct fuse_io_args *ia)
724 {
725         kfree(ia->ap.pages);
726         kfree(ia);
727 }
728
729 static 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
758 static 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
779 static 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
798 static 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
813 static 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
829 static 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
873 static 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
890 static 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
931 static 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
967 static 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
1011 static 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
1032 static 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
1054 static 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
1066 static 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
1093 bool 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
1112 static 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
1162 static 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
1236 static 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
1245 static 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
1297 static 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                     should_remove_suid(file_dentry(file))) {
1317                         goto writethrough;
1318                 }
1319
1320                 return generic_file_write_iter(iocb, from);
1321         }
1322
1323 writethrough:
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         }
1372 out:
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
1381 static 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
1386 static 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
1392 static 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
1448 ssize_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 }
1533 EXPORT_SYMBOL_GPL(fuse_direct_io);
1534
1535 static 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
1549 static ssize_t fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter);
1550
1551 static 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
1566 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1567 {
1568         struct inode *inode = file_inode(iocb->ki_filp);
1569         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1570         ssize_t res;
1571
1572         /* Don't allow parallel writes to the same file */
1573         inode_lock(inode);
1574         res = generic_write_checks(iocb, from);
1575         if (res > 0) {
1576                 if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1577                         res = fuse_direct_IO(iocb, from);
1578                 } else {
1579                         res = fuse_direct_io(&io, from, &iocb->ki_pos,
1580                                              FUSE_DIO_WRITE);
1581                         fuse_write_update_attr(inode, iocb->ki_pos, res);
1582                 }
1583         }
1584         inode_unlock(inode);
1585
1586         return res;
1587 }
1588
1589 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
1590 {
1591         struct file *file = iocb->ki_filp;
1592         struct fuse_file *ff = file->private_data;
1593         struct inode *inode = file_inode(file);
1594
1595         if (fuse_is_bad(inode))
1596                 return -EIO;
1597
1598         if (FUSE_IS_DAX(inode))
1599                 return fuse_dax_read_iter(iocb, to);
1600
1601         if (!(ff->open_flags & FOPEN_DIRECT_IO))
1602                 return fuse_cache_read_iter(iocb, to);
1603         else
1604                 return fuse_direct_read_iter(iocb, to);
1605 }
1606
1607 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1608 {
1609         struct file *file = iocb->ki_filp;
1610         struct fuse_file *ff = file->private_data;
1611         struct inode *inode = file_inode(file);
1612
1613         if (fuse_is_bad(inode))
1614                 return -EIO;
1615
1616         if (FUSE_IS_DAX(inode))
1617                 return fuse_dax_write_iter(iocb, from);
1618
1619         if (!(ff->open_flags & FOPEN_DIRECT_IO))
1620                 return fuse_cache_write_iter(iocb, from);
1621         else
1622                 return fuse_direct_write_iter(iocb, from);
1623 }
1624
1625 static void fuse_writepage_free(struct fuse_writepage_args *wpa)
1626 {
1627         struct fuse_args_pages *ap = &wpa->ia.ap;
1628         int i;
1629
1630         if (wpa->bucket)
1631                 fuse_sync_bucket_dec(wpa->bucket);
1632
1633         for (i = 0; i < ap->num_pages; i++)
1634                 __free_page(ap->pages[i]);
1635
1636         if (wpa->ia.ff)
1637                 fuse_file_put(wpa->ia.ff, false, false);
1638
1639         kfree(ap->pages);
1640         kfree(wpa);
1641 }
1642
1643 static void fuse_writepage_finish(struct fuse_mount *fm,
1644                                   struct fuse_writepage_args *wpa)
1645 {
1646         struct fuse_args_pages *ap = &wpa->ia.ap;
1647         struct inode *inode = wpa->inode;
1648         struct fuse_inode *fi = get_fuse_inode(inode);
1649         struct backing_dev_info *bdi = inode_to_bdi(inode);
1650         int i;
1651
1652         for (i = 0; i < ap->num_pages; i++) {
1653                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1654                 dec_node_page_state(ap->pages[i], NR_WRITEBACK_TEMP);
1655                 wb_writeout_inc(&bdi->wb);
1656         }
1657         wake_up(&fi->page_waitq);
1658 }
1659
1660 /* Called under fi->lock, may release and reacquire it */
1661 static void fuse_send_writepage(struct fuse_mount *fm,
1662                                 struct fuse_writepage_args *wpa, loff_t size)
1663 __releases(fi->lock)
1664 __acquires(fi->lock)
1665 {
1666         struct fuse_writepage_args *aux, *next;
1667         struct fuse_inode *fi = get_fuse_inode(wpa->inode);
1668         struct fuse_write_in *inarg = &wpa->ia.write.in;
1669         struct fuse_args *args = &wpa->ia.ap.args;
1670         __u64 data_size = wpa->ia.ap.num_pages * PAGE_SIZE;
1671         int err;
1672
1673         fi->writectr++;
1674         if (inarg->offset + data_size <= size) {
1675                 inarg->size = data_size;
1676         } else if (inarg->offset < size) {
1677                 inarg->size = size - inarg->offset;
1678         } else {
1679                 /* Got truncated off completely */
1680                 goto out_free;
1681         }
1682
1683         args->in_args[1].size = inarg->size;
1684         args->force = true;
1685         args->nocreds = true;
1686
1687         err = fuse_simple_background(fm, args, GFP_ATOMIC);
1688         if (err == -ENOMEM) {
1689                 spin_unlock(&fi->lock);
1690                 err = fuse_simple_background(fm, args, GFP_NOFS | __GFP_NOFAIL);
1691                 spin_lock(&fi->lock);
1692         }
1693
1694         /* Fails on broken connection only */
1695         if (unlikely(err))
1696                 goto out_free;
1697
1698         return;
1699
1700  out_free:
1701         fi->writectr--;
1702         rb_erase(&wpa->writepages_entry, &fi->writepages);
1703         fuse_writepage_finish(fm, wpa);
1704         spin_unlock(&fi->lock);
1705
1706         /* After fuse_writepage_finish() aux request list is private */
1707         for (aux = wpa->next; aux; aux = next) {
1708                 next = aux->next;
1709                 aux->next = NULL;
1710                 fuse_writepage_free(aux);
1711         }
1712
1713         fuse_writepage_free(wpa);
1714         spin_lock(&fi->lock);
1715 }
1716
1717 /*
1718  * If fi->writectr is positive (no truncate or fsync going on) send
1719  * all queued writepage requests.
1720  *
1721  * Called with fi->lock
1722  */
1723 void fuse_flush_writepages(struct inode *inode)
1724 __releases(fi->lock)
1725 __acquires(fi->lock)
1726 {
1727         struct fuse_mount *fm = get_fuse_mount(inode);
1728         struct fuse_inode *fi = get_fuse_inode(inode);
1729         loff_t crop = i_size_read(inode);
1730         struct fuse_writepage_args *wpa;
1731
1732         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1733                 wpa = list_entry(fi->queued_writes.next,
1734                                  struct fuse_writepage_args, queue_entry);
1735                 list_del_init(&wpa->queue_entry);
1736                 fuse_send_writepage(fm, wpa, crop);
1737         }
1738 }
1739
1740 static struct fuse_writepage_args *fuse_insert_writeback(struct rb_root *root,
1741                                                 struct fuse_writepage_args *wpa)
1742 {
1743         pgoff_t idx_from = wpa->ia.write.in.offset >> PAGE_SHIFT;
1744         pgoff_t idx_to = idx_from + wpa->ia.ap.num_pages - 1;
1745         struct rb_node **p = &root->rb_node;
1746         struct rb_node  *parent = NULL;
1747
1748         WARN_ON(!wpa->ia.ap.num_pages);
1749         while (*p) {
1750                 struct fuse_writepage_args *curr;
1751                 pgoff_t curr_index;
1752
1753                 parent = *p;
1754                 curr = rb_entry(parent, struct fuse_writepage_args,
1755                                 writepages_entry);
1756                 WARN_ON(curr->inode != wpa->inode);
1757                 curr_index = curr->ia.write.in.offset >> PAGE_SHIFT;
1758
1759                 if (idx_from >= curr_index + curr->ia.ap.num_pages)
1760                         p = &(*p)->rb_right;
1761                 else if (idx_to < curr_index)
1762                         p = &(*p)->rb_left;
1763                 else
1764                         return curr;
1765         }
1766
1767         rb_link_node(&wpa->writepages_entry, parent, p);
1768         rb_insert_color(&wpa->writepages_entry, root);
1769         return NULL;
1770 }
1771
1772 static void tree_insert(struct rb_root *root, struct fuse_writepage_args *wpa)
1773 {
1774         WARN_ON(fuse_insert_writeback(root, wpa));
1775 }
1776
1777 static void fuse_writepage_end(struct fuse_mount *fm, struct fuse_args *args,
1778                                int error)
1779 {
1780         struct fuse_writepage_args *wpa =
1781                 container_of(args, typeof(*wpa), ia.ap.args);
1782         struct inode *inode = wpa->inode;
1783         struct fuse_inode *fi = get_fuse_inode(inode);
1784         struct fuse_conn *fc = get_fuse_conn(inode);
1785
1786         mapping_set_error(inode->i_mapping, error);
1787         /*
1788          * A writeback finished and this might have updated mtime/ctime on
1789          * server making local mtime/ctime stale.  Hence invalidate attrs.
1790          * Do this only if writeback_cache is not enabled.  If writeback_cache
1791          * is enabled, we trust local ctime/mtime.
1792          */
1793         if (!fc->writeback_cache)
1794                 fuse_invalidate_attr_mask(inode, FUSE_STATX_MODIFY);
1795         spin_lock(&fi->lock);
1796         rb_erase(&wpa->writepages_entry, &fi->writepages);
1797         while (wpa->next) {
1798                 struct fuse_mount *fm = get_fuse_mount(inode);
1799                 struct fuse_write_in *inarg = &wpa->ia.write.in;
1800                 struct fuse_writepage_args *next = wpa->next;
1801
1802                 wpa->next = next->next;
1803                 next->next = NULL;
1804                 next->ia.ff = fuse_file_get(wpa->ia.ff);
1805                 tree_insert(&fi->writepages, next);
1806
1807                 /*
1808                  * Skip fuse_flush_writepages() to make it easy to crop requests
1809                  * based on primary request size.
1810                  *
1811                  * 1st case (trivial): there are no concurrent activities using
1812                  * fuse_set/release_nowrite.  Then we're on safe side because
1813                  * fuse_flush_writepages() would call fuse_send_writepage()
1814                  * anyway.
1815                  *
1816                  * 2nd case: someone called fuse_set_nowrite and it is waiting
1817                  * now for completion of all in-flight requests.  This happens
1818                  * rarely and no more than once per page, so this should be
1819                  * okay.
1820                  *
1821                  * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1822                  * of fuse_set_nowrite..fuse_release_nowrite section.  The fact
1823                  * that fuse_set_nowrite returned implies that all in-flight
1824                  * requests were completed along with all of their secondary
1825                  * requests.  Further primary requests are blocked by negative
1826                  * writectr.  Hence there cannot be any in-flight requests and
1827                  * no invocations of fuse_writepage_end() while we're in
1828                  * fuse_set_nowrite..fuse_release_nowrite section.
1829                  */
1830                 fuse_send_writepage(fm, next, inarg->offset + inarg->size);
1831         }
1832         fi->writectr--;
1833         fuse_writepage_finish(fm, wpa);
1834         spin_unlock(&fi->lock);
1835         fuse_writepage_free(wpa);
1836 }
1837
1838 static struct fuse_file *__fuse_write_file_get(struct fuse_inode *fi)
1839 {
1840         struct fuse_file *ff;
1841
1842         spin_lock(&fi->lock);
1843         ff = list_first_entry_or_null(&fi->write_files, struct fuse_file,
1844                                       write_entry);
1845         if (ff)
1846                 fuse_file_get(ff);
1847         spin_unlock(&fi->lock);
1848
1849         return ff;
1850 }
1851
1852 static struct fuse_file *fuse_write_file_get(struct fuse_inode *fi)
1853 {
1854         struct fuse_file *ff = __fuse_write_file_get(fi);
1855         WARN_ON(!ff);
1856         return ff;
1857 }
1858
1859 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1860 {
1861         struct fuse_inode *fi = get_fuse_inode(inode);
1862         struct fuse_file *ff;
1863         int err;
1864
1865         /*
1866          * Inode is always written before the last reference is dropped and
1867          * hence this should not be reached from reclaim.
1868          *
1869          * Writing back the inode from reclaim can deadlock if the request
1870          * processing itself needs an allocation.  Allocations triggering
1871          * reclaim while serving a request can't be prevented, because it can
1872          * involve any number of unrelated userspace processes.
1873          */
1874         WARN_ON(wbc->for_reclaim);
1875
1876         ff = __fuse_write_file_get(fi);
1877         err = fuse_flush_times(inode, ff);
1878         if (ff)
1879                 fuse_file_put(ff, false, false);
1880
1881         return err;
1882 }
1883
1884 static struct fuse_writepage_args *fuse_writepage_args_alloc(void)
1885 {
1886         struct fuse_writepage_args *wpa;
1887         struct fuse_args_pages *ap;
1888
1889         wpa = kzalloc(sizeof(*wpa), GFP_NOFS);
1890         if (wpa) {
1891                 ap = &wpa->ia.ap;
1892                 ap->num_pages = 0;
1893                 ap->pages = fuse_pages_alloc(1, GFP_NOFS, &ap->descs);
1894                 if (!ap->pages) {
1895                         kfree(wpa);
1896                         wpa = NULL;
1897                 }
1898         }
1899         return wpa;
1900
1901 }
1902
1903 static void fuse_writepage_add_to_bucket(struct fuse_conn *fc,
1904                                          struct fuse_writepage_args *wpa)
1905 {
1906         if (!fc->sync_fs)
1907                 return;
1908
1909         rcu_read_lock();
1910         /* Prevent resurrection of dead bucket in unlikely race with syncfs */
1911         do {
1912                 wpa->bucket = rcu_dereference(fc->curr_bucket);
1913         } while (unlikely(!atomic_inc_not_zero(&wpa->bucket->count)));
1914         rcu_read_unlock();
1915 }
1916
1917 static int fuse_writepage_locked(struct page *page)
1918 {
1919         struct address_space *mapping = page->mapping;
1920         struct inode *inode = mapping->host;
1921         struct fuse_conn *fc = get_fuse_conn(inode);
1922         struct fuse_inode *fi = get_fuse_inode(inode);
1923         struct fuse_writepage_args *wpa;
1924         struct fuse_args_pages *ap;
1925         struct page *tmp_page;
1926         int error = -ENOMEM;
1927
1928         set_page_writeback(page);
1929
1930         wpa = fuse_writepage_args_alloc();
1931         if (!wpa)
1932                 goto err;
1933         ap = &wpa->ia.ap;
1934
1935         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1936         if (!tmp_page)
1937                 goto err_free;
1938
1939         error = -EIO;
1940         wpa->ia.ff = fuse_write_file_get(fi);
1941         if (!wpa->ia.ff)
1942                 goto err_nofile;
1943
1944         fuse_writepage_add_to_bucket(fc, wpa);
1945         fuse_write_args_fill(&wpa->ia, wpa->ia.ff, page_offset(page), 0);
1946
1947         copy_highpage(tmp_page, page);
1948         wpa->ia.write.in.write_flags |= FUSE_WRITE_CACHE;
1949         wpa->next = NULL;
1950         ap->args.in_pages = true;
1951         ap->num_pages = 1;
1952         ap->pages[0] = tmp_page;
1953         ap->descs[0].offset = 0;
1954         ap->descs[0].length = PAGE_SIZE;
1955         ap->args.end = fuse_writepage_end;
1956         wpa->inode = inode;
1957
1958         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1959         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1960
1961         spin_lock(&fi->lock);
1962         tree_insert(&fi->writepages, wpa);
1963         list_add_tail(&wpa->queue_entry, &fi->queued_writes);
1964         fuse_flush_writepages(inode);
1965         spin_unlock(&fi->lock);
1966
1967         end_page_writeback(page);
1968
1969         return 0;
1970
1971 err_nofile:
1972         __free_page(tmp_page);
1973 err_free:
1974         kfree(wpa);
1975 err:
1976         mapping_set_error(page->mapping, error);
1977         end_page_writeback(page);
1978         return error;
1979 }
1980
1981 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1982 {
1983         struct fuse_conn *fc = get_fuse_conn(page->mapping->host);
1984         int err;
1985
1986         if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1987                 /*
1988                  * ->writepages() should be called for sync() and friends.  We
1989                  * should only get here on direct reclaim and then we are
1990                  * allowed to skip a page which is already in flight
1991                  */
1992                 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1993
1994                 redirty_page_for_writepage(wbc, page);
1995                 unlock_page(page);
1996                 return 0;
1997         }
1998
1999         if (wbc->sync_mode == WB_SYNC_NONE &&
2000             fc->num_background >= fc->congestion_threshold)
2001                 return AOP_WRITEPAGE_ACTIVATE;
2002
2003         err = fuse_writepage_locked(page);
2004         unlock_page(page);
2005
2006         return err;
2007 }
2008
2009 struct fuse_fill_wb_data {
2010         struct fuse_writepage_args *wpa;
2011         struct fuse_file *ff;
2012         struct inode *inode;
2013         struct page **orig_pages;
2014         unsigned int max_pages;
2015 };
2016
2017 static bool fuse_pages_realloc(struct fuse_fill_wb_data *data)
2018 {
2019         struct fuse_args_pages *ap = &data->wpa->ia.ap;
2020         struct fuse_conn *fc = get_fuse_conn(data->inode);
2021         struct page **pages;
2022         struct fuse_page_desc *descs;
2023         unsigned int npages = min_t(unsigned int,
2024                                     max_t(unsigned int, data->max_pages * 2,
2025                                           FUSE_DEFAULT_MAX_PAGES_PER_REQ),
2026                                     fc->max_pages);
2027         WARN_ON(npages <= data->max_pages);
2028
2029         pages = fuse_pages_alloc(npages, GFP_NOFS, &descs);
2030         if (!pages)
2031                 return false;
2032
2033         memcpy(pages, ap->pages, sizeof(struct page *) * ap->num_pages);
2034         memcpy(descs, ap->descs, sizeof(struct fuse_page_desc) * ap->num_pages);
2035         kfree(ap->pages);
2036         ap->pages = pages;
2037         ap->descs = descs;
2038         data->max_pages = npages;
2039
2040         return true;
2041 }
2042
2043 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
2044 {
2045         struct fuse_writepage_args *wpa = data->wpa;
2046         struct inode *inode = data->inode;
2047         struct fuse_inode *fi = get_fuse_inode(inode);
2048         int num_pages = wpa->ia.ap.num_pages;
2049         int i;
2050
2051         wpa->ia.ff = fuse_file_get(data->ff);
2052         spin_lock(&fi->lock);
2053         list_add_tail(&wpa->queue_entry, &fi->queued_writes);
2054         fuse_flush_writepages(inode);
2055         spin_unlock(&fi->lock);
2056
2057         for (i = 0; i < num_pages; i++)
2058                 end_page_writeback(data->orig_pages[i]);
2059 }
2060
2061 /*
2062  * Check under fi->lock if the page is under writeback, and insert it onto the
2063  * rb_tree if not. Otherwise iterate auxiliary write requests, to see if there's
2064  * one already added for a page at this offset.  If there's none, then insert
2065  * this new request onto the auxiliary list, otherwise reuse the existing one by
2066  * swapping the new temp page with the old one.
2067  */
2068 static bool fuse_writepage_add(struct fuse_writepage_args *new_wpa,
2069                                struct page *page)
2070 {
2071         struct fuse_inode *fi = get_fuse_inode(new_wpa->inode);
2072         struct fuse_writepage_args *tmp;
2073         struct fuse_writepage_args *old_wpa;
2074         struct fuse_args_pages *new_ap = &new_wpa->ia.ap;
2075
2076         WARN_ON(new_ap->num_pages != 0);
2077         new_ap->num_pages = 1;
2078
2079         spin_lock(&fi->lock);
2080         old_wpa = fuse_insert_writeback(&fi->writepages, new_wpa);
2081         if (!old_wpa) {
2082                 spin_unlock(&fi->lock);
2083                 return true;
2084         }
2085
2086         for (tmp = old_wpa->next; tmp; tmp = tmp->next) {
2087                 pgoff_t curr_index;
2088
2089                 WARN_ON(tmp->inode != new_wpa->inode);
2090                 curr_index = tmp->ia.write.in.offset >> PAGE_SHIFT;
2091                 if (curr_index == page->index) {
2092                         WARN_ON(tmp->ia.ap.num_pages != 1);
2093                         swap(tmp->ia.ap.pages[0], new_ap->pages[0]);
2094                         break;
2095                 }
2096         }
2097
2098         if (!tmp) {
2099                 new_wpa->next = old_wpa->next;
2100                 old_wpa->next = new_wpa;
2101         }
2102
2103         spin_unlock(&fi->lock);
2104
2105         if (tmp) {
2106                 struct backing_dev_info *bdi = inode_to_bdi(new_wpa->inode);
2107
2108                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
2109                 dec_node_page_state(new_ap->pages[0], NR_WRITEBACK_TEMP);
2110                 wb_writeout_inc(&bdi->wb);
2111                 fuse_writepage_free(new_wpa);
2112         }
2113
2114         return false;
2115 }
2116
2117 static bool fuse_writepage_need_send(struct fuse_conn *fc, struct page *page,
2118                                      struct fuse_args_pages *ap,
2119                                      struct fuse_fill_wb_data *data)
2120 {
2121         WARN_ON(!ap->num_pages);
2122
2123         /*
2124          * Being under writeback is unlikely but possible.  For example direct
2125          * read to an mmaped fuse file will set the page dirty twice; once when
2126          * the pages are faulted with get_user_pages(), and then after the read
2127          * completed.
2128          */
2129         if (fuse_page_is_writeback(data->inode, page->index))
2130                 return true;
2131
2132         /* Reached max pages */
2133         if (ap->num_pages == fc->max_pages)
2134                 return true;
2135
2136         /* Reached max write bytes */
2137         if ((ap->num_pages + 1) * PAGE_SIZE > fc->max_write)
2138                 return true;
2139
2140         /* Discontinuity */
2141         if (data->orig_pages[ap->num_pages - 1]->index + 1 != page->index)
2142                 return true;
2143
2144         /* Need to grow the pages array?  If so, did the expansion fail? */
2145         if (ap->num_pages == data->max_pages && !fuse_pages_realloc(data))
2146                 return true;
2147
2148         return false;
2149 }
2150
2151 static int fuse_writepages_fill(struct page *page,
2152                 struct writeback_control *wbc, void *_data)
2153 {
2154         struct fuse_fill_wb_data *data = _data;
2155         struct fuse_writepage_args *wpa = data->wpa;
2156         struct fuse_args_pages *ap = &wpa->ia.ap;
2157         struct inode *inode = data->inode;
2158         struct fuse_inode *fi = get_fuse_inode(inode);
2159         struct fuse_conn *fc = get_fuse_conn(inode);
2160         struct page *tmp_page;
2161         int err;
2162
2163         if (!data->ff) {
2164                 err = -EIO;
2165                 data->ff = fuse_write_file_get(fi);
2166                 if (!data->ff)
2167                         goto out_unlock;
2168         }
2169
2170         if (wpa && fuse_writepage_need_send(fc, page, ap, data)) {
2171                 fuse_writepages_send(data);
2172                 data->wpa = NULL;
2173         }
2174
2175         err = -ENOMEM;
2176         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2177         if (!tmp_page)
2178                 goto out_unlock;
2179
2180         /*
2181          * The page must not be redirtied until the writeout is completed
2182          * (i.e. userspace has sent a reply to the write request).  Otherwise
2183          * there could be more than one temporary page instance for each real
2184          * page.
2185          *
2186          * This is ensured by holding the page lock in page_mkwrite() while
2187          * checking fuse_page_is_writeback().  We already hold the page lock
2188          * since clear_page_dirty_for_io() and keep it held until we add the
2189          * request to the fi->writepages list and increment ap->num_pages.
2190          * After this fuse_page_is_writeback() will indicate that the page is
2191          * under writeback, so we can release the page lock.
2192          */
2193         if (data->wpa == NULL) {
2194                 err = -ENOMEM;
2195                 wpa = fuse_writepage_args_alloc();
2196                 if (!wpa) {
2197                         __free_page(tmp_page);
2198                         goto out_unlock;
2199                 }
2200                 fuse_writepage_add_to_bucket(fc, wpa);
2201
2202                 data->max_pages = 1;
2203
2204                 ap = &wpa->ia.ap;
2205                 fuse_write_args_fill(&wpa->ia, data->ff, page_offset(page), 0);
2206                 wpa->ia.write.in.write_flags |= FUSE_WRITE_CACHE;
2207                 wpa->next = NULL;
2208                 ap->args.in_pages = true;
2209                 ap->args.end = fuse_writepage_end;
2210                 ap->num_pages = 0;
2211                 wpa->inode = inode;
2212         }
2213         set_page_writeback(page);
2214
2215         copy_highpage(tmp_page, page);
2216         ap->pages[ap->num_pages] = tmp_page;
2217         ap->descs[ap->num_pages].offset = 0;
2218         ap->descs[ap->num_pages].length = PAGE_SIZE;
2219         data->orig_pages[ap->num_pages] = page;
2220
2221         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
2222         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
2223
2224         err = 0;
2225         if (data->wpa) {
2226                 /*
2227                  * Protected by fi->lock against concurrent access by
2228                  * fuse_page_is_writeback().
2229                  */
2230                 spin_lock(&fi->lock);
2231                 ap->num_pages++;
2232                 spin_unlock(&fi->lock);
2233         } else if (fuse_writepage_add(wpa, page)) {
2234                 data->wpa = wpa;
2235         } else {
2236                 end_page_writeback(page);
2237         }
2238 out_unlock:
2239         unlock_page(page);
2240
2241         return err;
2242 }
2243
2244 static int fuse_writepages(struct address_space *mapping,
2245                            struct writeback_control *wbc)
2246 {
2247         struct inode *inode = mapping->host;
2248         struct fuse_conn *fc = get_fuse_conn(inode);
2249         struct fuse_fill_wb_data data;
2250         int err;
2251
2252         err = -EIO;
2253         if (fuse_is_bad(inode))
2254                 goto out;
2255
2256         if (wbc->sync_mode == WB_SYNC_NONE &&
2257             fc->num_background >= fc->congestion_threshold)
2258                 return 0;
2259
2260         data.inode = inode;
2261         data.wpa = NULL;
2262         data.ff = NULL;
2263
2264         err = -ENOMEM;
2265         data.orig_pages = kcalloc(fc->max_pages,
2266                                   sizeof(struct page *),
2267                                   GFP_NOFS);
2268         if (!data.orig_pages)
2269                 goto out;
2270
2271         err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
2272         if (data.wpa) {
2273                 WARN_ON(!data.wpa->ia.ap.num_pages);
2274                 fuse_writepages_send(&data);
2275         }
2276         if (data.ff)
2277                 fuse_file_put(data.ff, false, false);
2278
2279         kfree(data.orig_pages);
2280 out:
2281         return err;
2282 }
2283
2284 /*
2285  * It's worthy to make sure that space is reserved on disk for the write,
2286  * but how to implement it without killing performance need more thinking.
2287  */
2288 static int fuse_write_begin(struct file *file, struct address_space *mapping,
2289                 loff_t pos, unsigned len, struct page **pagep, void **fsdata)
2290 {
2291         pgoff_t index = pos >> PAGE_SHIFT;
2292         struct fuse_conn *fc = get_fuse_conn(file_inode(file));
2293         struct page *page;
2294         loff_t fsize;
2295         int err = -ENOMEM;
2296
2297         WARN_ON(!fc->writeback_cache);
2298
2299         page = grab_cache_page_write_begin(mapping, index);
2300         if (!page)
2301                 goto error;
2302
2303         fuse_wait_on_page_writeback(mapping->host, page->index);
2304
2305         if (PageUptodate(page) || len == PAGE_SIZE)
2306                 goto success;
2307         /*
2308          * Check if the start this page comes after the end of file, in which
2309          * case the readpage can be optimized away.
2310          */
2311         fsize = i_size_read(mapping->host);
2312         if (fsize <= (pos & PAGE_MASK)) {
2313                 size_t off = pos & ~PAGE_MASK;
2314                 if (off)
2315                         zero_user_segment(page, 0, off);
2316                 goto success;
2317         }
2318         err = fuse_do_readpage(file, page);
2319         if (err)
2320                 goto cleanup;
2321 success:
2322         *pagep = page;
2323         return 0;
2324
2325 cleanup:
2326         unlock_page(page);
2327         put_page(page);
2328 error:
2329         return err;
2330 }
2331
2332 static int fuse_write_end(struct file *file, struct address_space *mapping,
2333                 loff_t pos, unsigned len, unsigned copied,
2334                 struct page *page, void *fsdata)
2335 {
2336         struct inode *inode = page->mapping->host;
2337
2338         /* Haven't copied anything?  Skip zeroing, size extending, dirtying. */
2339         if (!copied)
2340                 goto unlock;
2341
2342         pos += copied;
2343         if (!PageUptodate(page)) {
2344                 /* Zero any unwritten bytes at the end of the page */
2345                 size_t endoff = pos & ~PAGE_MASK;
2346                 if (endoff)
2347                         zero_user_segment(page, endoff, PAGE_SIZE);
2348                 SetPageUptodate(page);
2349         }
2350
2351         if (pos > inode->i_size)
2352                 i_size_write(inode, pos);
2353
2354         set_page_dirty(page);
2355
2356 unlock:
2357         unlock_page(page);
2358         put_page(page);
2359
2360         return copied;
2361 }
2362
2363 static int fuse_launder_folio(struct folio *folio)
2364 {
2365         int err = 0;
2366         if (folio_clear_dirty_for_io(folio)) {
2367                 struct inode *inode = folio->mapping->host;
2368
2369                 /* Serialize with pending writeback for the same page */
2370                 fuse_wait_on_page_writeback(inode, folio->index);
2371                 err = fuse_writepage_locked(&folio->page);
2372                 if (!err)
2373                         fuse_wait_on_page_writeback(inode, folio->index);
2374         }
2375         return err;
2376 }
2377
2378 /*
2379  * Write back dirty data/metadata now (there may not be any suitable
2380  * open files later for data)
2381  */
2382 static void fuse_vma_close(struct vm_area_struct *vma)
2383 {
2384         int err;
2385
2386         err = write_inode_now(vma->vm_file->f_mapping->host, 1);
2387         mapping_set_error(vma->vm_file->f_mapping, err);
2388 }
2389
2390 /*
2391  * Wait for writeback against this page to complete before allowing it
2392  * to be marked dirty again, and hence written back again, possibly
2393  * before the previous writepage completed.
2394  *
2395  * Block here, instead of in ->writepage(), so that the userspace fs
2396  * can only block processes actually operating on the filesystem.
2397  *
2398  * Otherwise unprivileged userspace fs would be able to block
2399  * unrelated:
2400  *
2401  * - page migration
2402  * - sync(2)
2403  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2404  */
2405 static vm_fault_t fuse_page_mkwrite(struct vm_fault *vmf)
2406 {
2407         struct page *page = vmf->page;
2408         struct inode *inode = file_inode(vmf->vma->vm_file);
2409
2410         file_update_time(vmf->vma->vm_file);
2411         lock_page(page);
2412         if (page->mapping != inode->i_mapping) {
2413                 unlock_page(page);
2414                 return VM_FAULT_NOPAGE;
2415         }
2416
2417         fuse_wait_on_page_writeback(inode, page->index);
2418         return VM_FAULT_LOCKED;
2419 }
2420
2421 static const struct vm_operations_struct fuse_file_vm_ops = {
2422         .close          = fuse_vma_close,
2423         .fault          = filemap_fault,
2424         .map_pages      = filemap_map_pages,
2425         .page_mkwrite   = fuse_page_mkwrite,
2426 };
2427
2428 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2429 {
2430         struct fuse_file *ff = file->private_data;
2431
2432         /* DAX mmap is superior to direct_io mmap */
2433         if (FUSE_IS_DAX(file_inode(file)))
2434                 return fuse_dax_mmap(file, vma);
2435
2436         if (ff->open_flags & FOPEN_DIRECT_IO) {
2437                 /* Can't provide the coherency needed for MAP_SHARED */
2438                 if (vma->vm_flags & VM_MAYSHARE)
2439                         return -ENODEV;
2440
2441                 invalidate_inode_pages2(file->f_mapping);
2442
2443                 return generic_file_mmap(file, vma);
2444         }
2445
2446         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2447                 fuse_link_write_file(file);
2448
2449         file_accessed(file);
2450         vma->vm_ops = &fuse_file_vm_ops;
2451         return 0;
2452 }
2453
2454 static int convert_fuse_file_lock(struct fuse_conn *fc,
2455                                   const struct fuse_file_lock *ffl,
2456                                   struct file_lock *fl)
2457 {
2458         switch (ffl->type) {
2459         case F_UNLCK:
2460                 break;
2461
2462         case F_RDLCK:
2463         case F_WRLCK:
2464                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2465                     ffl->end < ffl->start)
2466                         return -EIO;
2467
2468                 fl->fl_start = ffl->start;
2469                 fl->fl_end = ffl->end;
2470
2471                 /*
2472                  * Convert pid into init's pid namespace.  The locks API will
2473                  * translate it into the caller's pid namespace.
2474                  */
2475                 rcu_read_lock();
2476                 fl->fl_pid = pid_nr_ns(find_pid_ns(ffl->pid, fc->pid_ns), &init_pid_ns);
2477                 rcu_read_unlock();
2478                 break;
2479
2480         default:
2481                 return -EIO;
2482         }
2483         fl->fl_type = ffl->type;
2484         return 0;
2485 }
2486
2487 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2488                          const struct file_lock *fl, int opcode, pid_t pid,
2489                          int flock, struct fuse_lk_in *inarg)
2490 {
2491         struct inode *inode = file_inode(file);
2492         struct fuse_conn *fc = get_fuse_conn(inode);
2493         struct fuse_file *ff = file->private_data;
2494
2495         memset(inarg, 0, sizeof(*inarg));
2496         inarg->fh = ff->fh;
2497         inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2498         inarg->lk.start = fl->fl_start;
2499         inarg->lk.end = fl->fl_end;
2500         inarg->lk.type = fl->fl_type;
2501         inarg->lk.pid = pid;
2502         if (flock)
2503                 inarg->lk_flags |= FUSE_LK_FLOCK;
2504         args->opcode = opcode;
2505         args->nodeid = get_node_id(inode);
2506         args->in_numargs = 1;
2507         args->in_args[0].size = sizeof(*inarg);
2508         args->in_args[0].value = inarg;
2509 }
2510
2511 static int fuse_getlk(struct file *file, struct file_lock *fl)
2512 {
2513         struct inode *inode = file_inode(file);
2514         struct fuse_mount *fm = get_fuse_mount(inode);
2515         FUSE_ARGS(args);
2516         struct fuse_lk_in inarg;
2517         struct fuse_lk_out outarg;
2518         int err;
2519
2520         fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2521         args.out_numargs = 1;
2522         args.out_args[0].size = sizeof(outarg);
2523         args.out_args[0].value = &outarg;
2524         err = fuse_simple_request(fm, &args);
2525         if (!err)
2526                 err = convert_fuse_file_lock(fm->fc, &outarg.lk, fl);
2527
2528         return err;
2529 }
2530
2531 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2532 {
2533         struct inode *inode = file_inode(file);
2534         struct fuse_mount *fm = get_fuse_mount(inode);
2535         FUSE_ARGS(args);
2536         struct fuse_lk_in inarg;
2537         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2538         struct pid *pid = fl->fl_type != F_UNLCK ? task_tgid(current) : NULL;
2539         pid_t pid_nr = pid_nr_ns(pid, fm->fc->pid_ns);
2540         int err;
2541
2542         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2543                 /* NLM needs asynchronous locks, which we don't support yet */
2544                 return -ENOLCK;
2545         }
2546
2547         /* Unlock on close is handled by the flush method */
2548         if ((fl->fl_flags & FL_CLOSE_POSIX) == FL_CLOSE_POSIX)
2549                 return 0;
2550
2551         fuse_lk_fill(&args, file, fl, opcode, pid_nr, flock, &inarg);
2552         err = fuse_simple_request(fm, &args);
2553
2554         /* locking is restartable */
2555         if (err == -EINTR)
2556                 err = -ERESTARTSYS;
2557
2558         return err;
2559 }
2560
2561 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2562 {
2563         struct inode *inode = file_inode(file);
2564         struct fuse_conn *fc = get_fuse_conn(inode);
2565         int err;
2566
2567         if (cmd == F_CANCELLK) {
2568                 err = 0;
2569         } else if (cmd == F_GETLK) {
2570                 if (fc->no_lock) {
2571                         posix_test_lock(file, fl);
2572                         err = 0;
2573                 } else
2574                         err = fuse_getlk(file, fl);
2575         } else {
2576                 if (fc->no_lock)
2577                         err = posix_lock_file(file, fl, NULL);
2578                 else
2579                         err = fuse_setlk(file, fl, 0);
2580         }
2581         return err;
2582 }
2583
2584 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2585 {
2586         struct inode *inode = file_inode(file);
2587         struct fuse_conn *fc = get_fuse_conn(inode);
2588         int err;
2589
2590         if (fc->no_flock) {
2591                 err = locks_lock_file_wait(file, fl);
2592         } else {
2593                 struct fuse_file *ff = file->private_data;
2594
2595                 /* emulate flock with POSIX locks */
2596                 ff->flock = true;
2597                 err = fuse_setlk(file, fl, 1);
2598         }
2599
2600         return err;
2601 }
2602
2603 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2604 {
2605         struct inode *inode = mapping->host;
2606         struct fuse_mount *fm = get_fuse_mount(inode);
2607         FUSE_ARGS(args);
2608         struct fuse_bmap_in inarg;
2609         struct fuse_bmap_out outarg;
2610         int err;
2611
2612         if (!inode->i_sb->s_bdev || fm->fc->no_bmap)
2613                 return 0;
2614
2615         memset(&inarg, 0, sizeof(inarg));
2616         inarg.block = block;
2617         inarg.blocksize = inode->i_sb->s_blocksize;
2618         args.opcode = FUSE_BMAP;
2619         args.nodeid = get_node_id(inode);
2620         args.in_numargs = 1;
2621         args.in_args[0].size = sizeof(inarg);
2622         args.in_args[0].value = &inarg;
2623         args.out_numargs = 1;
2624         args.out_args[0].size = sizeof(outarg);
2625         args.out_args[0].value = &outarg;
2626         err = fuse_simple_request(fm, &args);
2627         if (err == -ENOSYS)
2628                 fm->fc->no_bmap = 1;
2629
2630         return err ? 0 : outarg.block;
2631 }
2632
2633 static loff_t fuse_lseek(struct file *file, loff_t offset, int whence)
2634 {
2635         struct inode *inode = file->f_mapping->host;
2636         struct fuse_mount *fm = get_fuse_mount(inode);
2637         struct fuse_file *ff = file->private_data;
2638         FUSE_ARGS(args);
2639         struct fuse_lseek_in inarg = {
2640                 .fh = ff->fh,
2641                 .offset = offset,
2642                 .whence = whence
2643         };
2644         struct fuse_lseek_out outarg;
2645         int err;
2646
2647         if (fm->fc->no_lseek)
2648                 goto fallback;
2649
2650         args.opcode = FUSE_LSEEK;
2651         args.nodeid = ff->nodeid;
2652         args.in_numargs = 1;
2653         args.in_args[0].size = sizeof(inarg);
2654         args.in_args[0].value = &inarg;
2655         args.out_numargs = 1;
2656         args.out_args[0].size = sizeof(outarg);
2657         args.out_args[0].value = &outarg;
2658         err = fuse_simple_request(fm, &args);
2659         if (err) {
2660                 if (err == -ENOSYS) {
2661                         fm->fc->no_lseek = 1;
2662                         goto fallback;
2663                 }
2664                 return err;
2665         }
2666
2667         return vfs_setpos(file, outarg.offset, inode->i_sb->s_maxbytes);
2668
2669 fallback:
2670         err = fuse_update_attributes(inode, file, STATX_SIZE);
2671         if (!err)
2672                 return generic_file_llseek(file, offset, whence);
2673         else
2674                 return err;
2675 }
2676
2677 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2678 {
2679         loff_t retval;
2680         struct inode *inode = file_inode(file);
2681
2682         switch (whence) {
2683         case SEEK_SET:
2684         case SEEK_CUR:
2685                  /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2686                 retval = generic_file_llseek(file, offset, whence);
2687                 break;
2688         case SEEK_END:
2689                 inode_lock(inode);
2690                 retval = fuse_update_attributes(inode, file, STATX_SIZE);
2691                 if (!retval)
2692                         retval = generic_file_llseek(file, offset, whence);
2693                 inode_unlock(inode);
2694                 break;
2695         case SEEK_HOLE:
2696         case SEEK_DATA:
2697                 inode_lock(inode);
2698                 retval = fuse_lseek(file, offset, whence);
2699                 inode_unlock(inode);
2700                 break;
2701         default:
2702                 retval = -EINVAL;
2703         }
2704
2705         return retval;
2706 }
2707
2708 /*
2709  * All files which have been polled are linked to RB tree
2710  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2711  * find the matching one.
2712  */
2713 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2714                                               struct rb_node **parent_out)
2715 {
2716         struct rb_node **link = &fc->polled_files.rb_node;
2717         struct rb_node *last = NULL;
2718
2719         while (*link) {
2720                 struct fuse_file *ff;
2721
2722                 last = *link;
2723                 ff = rb_entry(last, struct fuse_file, polled_node);
2724
2725                 if (kh < ff->kh)
2726                         link = &last->rb_left;
2727                 else if (kh > ff->kh)
2728                         link = &last->rb_right;
2729                 else
2730                         return link;
2731         }
2732
2733         if (parent_out)
2734                 *parent_out = last;
2735         return link;
2736 }
2737
2738 /*
2739  * The file is about to be polled.  Make sure it's on the polled_files
2740  * RB tree.  Note that files once added to the polled_files tree are
2741  * not removed before the file is released.  This is because a file
2742  * polled once is likely to be polled again.
2743  */
2744 static void fuse_register_polled_file(struct fuse_conn *fc,
2745                                       struct fuse_file *ff)
2746 {
2747         spin_lock(&fc->lock);
2748         if (RB_EMPTY_NODE(&ff->polled_node)) {
2749                 struct rb_node **link, *parent;
2750
2751                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2752                 BUG_ON(*link);
2753                 rb_link_node(&ff->polled_node, parent, link);
2754                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2755         }
2756         spin_unlock(&fc->lock);
2757 }
2758
2759 __poll_t fuse_file_poll(struct file *file, poll_table *wait)
2760 {
2761         struct fuse_file *ff = file->private_data;
2762         struct fuse_mount *fm = ff->fm;
2763         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2764         struct fuse_poll_out outarg;
2765         FUSE_ARGS(args);
2766         int err;
2767
2768         if (fm->fc->no_poll)
2769                 return DEFAULT_POLLMASK;
2770
2771         poll_wait(file, &ff->poll_wait, wait);
2772         inarg.events = mangle_poll(poll_requested_events(wait));
2773
2774         /*
2775          * Ask for notification iff there's someone waiting for it.
2776          * The client may ignore the flag and always notify.
2777          */
2778         if (waitqueue_active(&ff->poll_wait)) {
2779                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2780                 fuse_register_polled_file(fm->fc, ff);
2781         }
2782
2783         args.opcode = FUSE_POLL;
2784         args.nodeid = ff->nodeid;
2785         args.in_numargs = 1;
2786         args.in_args[0].size = sizeof(inarg);
2787         args.in_args[0].value = &inarg;
2788         args.out_numargs = 1;
2789         args.out_args[0].size = sizeof(outarg);
2790         args.out_args[0].value = &outarg;
2791         err = fuse_simple_request(fm, &args);
2792
2793         if (!err)
2794                 return demangle_poll(outarg.revents);
2795         if (err == -ENOSYS) {
2796                 fm->fc->no_poll = 1;
2797                 return DEFAULT_POLLMASK;
2798         }
2799         return EPOLLERR;
2800 }
2801 EXPORT_SYMBOL_GPL(fuse_file_poll);
2802
2803 /*
2804  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2805  * wakes up the poll waiters.
2806  */
2807 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2808                             struct fuse_notify_poll_wakeup_out *outarg)
2809 {
2810         u64 kh = outarg->kh;
2811         struct rb_node **link;
2812
2813         spin_lock(&fc->lock);
2814
2815         link = fuse_find_polled_node(fc, kh, NULL);
2816         if (*link) {
2817                 struct fuse_file *ff;
2818
2819                 ff = rb_entry(*link, struct fuse_file, polled_node);
2820                 wake_up_interruptible_sync(&ff->poll_wait);
2821         }
2822
2823         spin_unlock(&fc->lock);
2824         return 0;
2825 }
2826
2827 static void fuse_do_truncate(struct file *file)
2828 {
2829         struct inode *inode = file->f_mapping->host;
2830         struct iattr attr;
2831
2832         attr.ia_valid = ATTR_SIZE;
2833         attr.ia_size = i_size_read(inode);
2834
2835         attr.ia_file = file;
2836         attr.ia_valid |= ATTR_FILE;
2837
2838         fuse_do_setattr(file_dentry(file), &attr, file);
2839 }
2840
2841 static inline loff_t fuse_round_up(struct fuse_conn *fc, loff_t off)
2842 {
2843         return round_up(off, fc->max_pages << PAGE_SHIFT);
2844 }
2845
2846 static ssize_t
2847 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
2848 {
2849         DECLARE_COMPLETION_ONSTACK(wait);
2850         ssize_t ret = 0;
2851         struct file *file = iocb->ki_filp;
2852         struct fuse_file *ff = file->private_data;
2853         loff_t pos = 0;
2854         struct inode *inode;
2855         loff_t i_size;
2856         size_t count = iov_iter_count(iter), shortened = 0;
2857         loff_t offset = iocb->ki_pos;
2858         struct fuse_io_priv *io;
2859
2860         pos = offset;
2861         inode = file->f_mapping->host;
2862         i_size = i_size_read(inode);
2863
2864         if ((iov_iter_rw(iter) == READ) && (offset >= i_size))
2865                 return 0;
2866
2867         io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2868         if (!io)
2869                 return -ENOMEM;
2870         spin_lock_init(&io->lock);
2871         kref_init(&io->refcnt);
2872         io->reqs = 1;
2873         io->bytes = -1;
2874         io->size = 0;
2875         io->offset = offset;
2876         io->write = (iov_iter_rw(iter) == WRITE);
2877         io->err = 0;
2878         /*
2879          * By default, we want to optimize all I/Os with async request
2880          * submission to the client filesystem if supported.
2881          */
2882         io->async = ff->fm->fc->async_dio;
2883         io->iocb = iocb;
2884         io->blocking = is_sync_kiocb(iocb);
2885
2886         /* optimization for short read */
2887         if (io->async && !io->write && offset + count > i_size) {
2888                 iov_iter_truncate(iter, fuse_round_up(ff->fm->fc, i_size - offset));
2889                 shortened = count - iov_iter_count(iter);
2890                 count -= shortened;
2891         }
2892
2893         /*
2894          * We cannot asynchronously extend the size of a file.
2895          * In such case the aio will behave exactly like sync io.
2896          */
2897         if ((offset + count > i_size) && io->write)
2898                 io->blocking = true;
2899
2900         if (io->async && io->blocking) {
2901                 /*
2902                  * Additional reference to keep io around after
2903                  * calling fuse_aio_complete()
2904                  */
2905                 kref_get(&io->refcnt);
2906                 io->done = &wait;
2907         }
2908
2909         if (iov_iter_rw(iter) == WRITE) {
2910                 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
2911                 fuse_invalidate_attr_mask(inode, FUSE_STATX_MODSIZE);
2912         } else {
2913                 ret = __fuse_direct_read(io, iter, &pos);
2914         }
2915         iov_iter_reexpand(iter, iov_iter_count(iter) + shortened);
2916
2917         if (io->async) {
2918                 bool blocking = io->blocking;
2919
2920                 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2921
2922                 /* we have a non-extending, async request, so return */
2923                 if (!blocking)
2924                         return -EIOCBQUEUED;
2925
2926                 wait_for_completion(&wait);
2927                 ret = fuse_get_res_by_io(io);
2928         }
2929
2930         kref_put(&io->refcnt, fuse_io_release);
2931
2932         if (iov_iter_rw(iter) == WRITE) {
2933                 fuse_write_update_attr(inode, pos, ret);
2934                 if (ret < 0 && offset + count > i_size)
2935                         fuse_do_truncate(file);
2936         }
2937
2938         return ret;
2939 }
2940
2941 static int fuse_writeback_range(struct inode *inode, loff_t start, loff_t end)
2942 {
2943         int err = filemap_write_and_wait_range(inode->i_mapping, start, LLONG_MAX);
2944
2945         if (!err)
2946                 fuse_sync_writes(inode);
2947
2948         return err;
2949 }
2950
2951 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2952                                 loff_t length)
2953 {
2954         struct fuse_file *ff = file->private_data;
2955         struct inode *inode = file_inode(file);
2956         struct fuse_inode *fi = get_fuse_inode(inode);
2957         struct fuse_mount *fm = ff->fm;
2958         FUSE_ARGS(args);
2959         struct fuse_fallocate_in inarg = {
2960                 .fh = ff->fh,
2961                 .offset = offset,
2962                 .length = length,
2963                 .mode = mode
2964         };
2965         int err;
2966         bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
2967                            (mode & (FALLOC_FL_PUNCH_HOLE |
2968                                     FALLOC_FL_ZERO_RANGE));
2969
2970         bool block_faults = FUSE_IS_DAX(inode) && lock_inode;
2971
2972         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
2973                      FALLOC_FL_ZERO_RANGE))
2974                 return -EOPNOTSUPP;
2975
2976         if (fm->fc->no_fallocate)
2977                 return -EOPNOTSUPP;
2978
2979         if (lock_inode) {
2980                 inode_lock(inode);
2981                 if (block_faults) {
2982                         filemap_invalidate_lock(inode->i_mapping);
2983                         err = fuse_dax_break_layouts(inode, 0, 0);
2984                         if (err)
2985                                 goto out;
2986                 }
2987
2988                 if (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE)) {
2989                         loff_t endbyte = offset + length - 1;
2990
2991                         err = fuse_writeback_range(inode, offset, endbyte);
2992                         if (err)
2993                                 goto out;
2994                 }
2995         }
2996
2997         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
2998             offset + length > i_size_read(inode)) {
2999                 err = inode_newsize_ok(inode, offset + length);
3000                 if (err)
3001                         goto out;
3002         }
3003
3004         err = file_modified(file);
3005         if (err)
3006                 goto out;
3007
3008         if (!(mode & FALLOC_FL_KEEP_SIZE))
3009                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3010
3011         args.opcode = FUSE_FALLOCATE;
3012         args.nodeid = ff->nodeid;
3013         args.in_numargs = 1;
3014         args.in_args[0].size = sizeof(inarg);
3015         args.in_args[0].value = &inarg;
3016         err = fuse_simple_request(fm, &args);
3017         if (err == -ENOSYS) {
3018                 fm->fc->no_fallocate = 1;
3019                 err = -EOPNOTSUPP;
3020         }
3021         if (err)
3022                 goto out;
3023
3024         /* we could have extended the file */
3025         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
3026                 if (fuse_write_update_attr(inode, offset + length, length))
3027                         file_update_time(file);
3028         }
3029
3030         if (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE))
3031                 truncate_pagecache_range(inode, offset, offset + length - 1);
3032
3033         fuse_invalidate_attr_mask(inode, FUSE_STATX_MODSIZE);
3034
3035 out:
3036         if (!(mode & FALLOC_FL_KEEP_SIZE))
3037                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3038
3039         if (block_faults)
3040                 filemap_invalidate_unlock(inode->i_mapping);
3041
3042         if (lock_inode)
3043                 inode_unlock(inode);
3044
3045         fuse_flush_time_update(inode);
3046
3047         return err;
3048 }
3049
3050 static ssize_t __fuse_copy_file_range(struct file *file_in, loff_t pos_in,
3051                                       struct file *file_out, loff_t pos_out,
3052                                       size_t len, unsigned int flags)
3053 {
3054         struct fuse_file *ff_in = file_in->private_data;
3055         struct fuse_file *ff_out = file_out->private_data;
3056         struct inode *inode_in = file_inode(file_in);
3057         struct inode *inode_out = file_inode(file_out);
3058         struct fuse_inode *fi_out = get_fuse_inode(inode_out);
3059         struct fuse_mount *fm = ff_in->fm;
3060         struct fuse_conn *fc = fm->fc;
3061         FUSE_ARGS(args);
3062         struct fuse_copy_file_range_in inarg = {
3063                 .fh_in = ff_in->fh,
3064                 .off_in = pos_in,
3065                 .nodeid_out = ff_out->nodeid,
3066                 .fh_out = ff_out->fh,
3067                 .off_out = pos_out,
3068                 .len = len,
3069                 .flags = flags
3070         };
3071         struct fuse_write_out outarg;
3072         ssize_t err;
3073         /* mark unstable when write-back is not used, and file_out gets
3074          * extended */
3075         bool is_unstable = (!fc->writeback_cache) &&
3076                            ((pos_out + len) > inode_out->i_size);
3077
3078         if (fc->no_copy_file_range)
3079                 return -EOPNOTSUPP;
3080
3081         if (file_inode(file_in)->i_sb != file_inode(file_out)->i_sb)
3082                 return -EXDEV;
3083
3084         inode_lock(inode_in);
3085         err = fuse_writeback_range(inode_in, pos_in, pos_in + len - 1);
3086         inode_unlock(inode_in);
3087         if (err)
3088                 return err;
3089
3090         inode_lock(inode_out);
3091
3092         err = file_modified(file_out);
3093         if (err)
3094                 goto out;
3095
3096         /*
3097          * Write out dirty pages in the destination file before sending the COPY
3098          * request to userspace.  After the request is completed, truncate off
3099          * pages (including partial ones) from the cache that have been copied,
3100          * since these contain stale data at that point.
3101          *
3102          * This should be mostly correct, but if the COPY writes to partial
3103          * pages (at the start or end) and the parts not covered by the COPY are
3104          * written through a memory map after calling fuse_writeback_range(),
3105          * then these partial page modifications will be lost on truncation.
3106          *
3107          * It is unlikely that someone would rely on such mixed style
3108          * modifications.  Yet this does give less guarantees than if the
3109          * copying was performed with write(2).
3110          *
3111          * To fix this a mapping->invalidate_lock could be used to prevent new
3112          * faults while the copy is ongoing.
3113          */
3114         err = fuse_writeback_range(inode_out, pos_out, pos_out + len - 1);
3115         if (err)
3116                 goto out;
3117
3118         if (is_unstable)
3119                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3120
3121         args.opcode = FUSE_COPY_FILE_RANGE;
3122         args.nodeid = ff_in->nodeid;
3123         args.in_numargs = 1;
3124         args.in_args[0].size = sizeof(inarg);
3125         args.in_args[0].value = &inarg;
3126         args.out_numargs = 1;
3127         args.out_args[0].size = sizeof(outarg);
3128         args.out_args[0].value = &outarg;
3129         err = fuse_simple_request(fm, &args);
3130         if (err == -ENOSYS) {
3131                 fc->no_copy_file_range = 1;
3132                 err = -EOPNOTSUPP;
3133         }
3134         if (err)
3135                 goto out;
3136
3137         truncate_inode_pages_range(inode_out->i_mapping,
3138                                    ALIGN_DOWN(pos_out, PAGE_SIZE),
3139                                    ALIGN(pos_out + outarg.size, PAGE_SIZE) - 1);
3140
3141         file_update_time(file_out);
3142         fuse_write_update_attr(inode_out, pos_out + outarg.size, outarg.size);
3143
3144         err = outarg.size;
3145 out:
3146         if (is_unstable)
3147                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3148
3149         inode_unlock(inode_out);
3150         file_accessed(file_in);
3151
3152         fuse_flush_time_update(inode_out);
3153
3154         return err;
3155 }
3156
3157 static ssize_t fuse_copy_file_range(struct file *src_file, loff_t src_off,
3158                                     struct file *dst_file, loff_t dst_off,
3159                                     size_t len, unsigned int flags)
3160 {
3161         ssize_t ret;
3162
3163         ret = __fuse_copy_file_range(src_file, src_off, dst_file, dst_off,
3164                                      len, flags);
3165
3166         if (ret == -EOPNOTSUPP || ret == -EXDEV)
3167                 ret = generic_copy_file_range(src_file, src_off, dst_file,
3168                                               dst_off, len, flags);
3169         return ret;
3170 }
3171
3172 static const struct file_operations fuse_file_operations = {
3173         .llseek         = fuse_file_llseek,
3174         .read_iter      = fuse_file_read_iter,
3175         .write_iter     = fuse_file_write_iter,
3176         .mmap           = fuse_file_mmap,
3177         .open           = fuse_open,
3178         .flush          = fuse_flush,
3179         .release        = fuse_release,
3180         .fsync          = fuse_fsync,
3181         .lock           = fuse_file_lock,
3182         .get_unmapped_area = thp_get_unmapped_area,
3183         .flock          = fuse_file_flock,
3184         .splice_read    = generic_file_splice_read,
3185         .splice_write   = iter_file_splice_write,
3186         .unlocked_ioctl = fuse_file_ioctl,
3187         .compat_ioctl   = fuse_file_compat_ioctl,
3188         .poll           = fuse_file_poll,
3189         .fallocate      = fuse_file_fallocate,
3190         .copy_file_range = fuse_copy_file_range,
3191 };
3192
3193 static const struct address_space_operations fuse_file_aops  = {
3194         .read_folio     = fuse_read_folio,
3195         .readahead      = fuse_readahead,
3196         .writepage      = fuse_writepage,
3197         .writepages     = fuse_writepages,
3198         .launder_folio  = fuse_launder_folio,
3199         .dirty_folio    = filemap_dirty_folio,
3200         .bmap           = fuse_bmap,
3201         .direct_IO      = fuse_direct_IO,
3202         .write_begin    = fuse_write_begin,
3203         .write_end      = fuse_write_end,
3204 };
3205
3206 void fuse_init_file_inode(struct inode *inode, unsigned int flags)
3207 {
3208         struct fuse_inode *fi = get_fuse_inode(inode);
3209
3210         inode->i_fop = &fuse_file_operations;
3211         inode->i_data.a_ops = &fuse_file_aops;
3212
3213         INIT_LIST_HEAD(&fi->write_files);
3214         INIT_LIST_HEAD(&fi->queued_writes);
3215         fi->writectr = 0;
3216         init_waitqueue_head(&fi->page_waitq);
3217         fi->writepages = RB_ROOT;
3218
3219         if (IS_ENABLED(CONFIG_FUSE_DAX))
3220                 fuse_dax_inode_init(inode, flags);
3221 }