Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mszeredi...
[sfrench/cifs-2.6.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/module.h>
16 #include <linux/compat.h>
17 #include <linux/swap.h>
18 #include <linux/falloc.h>
19 #include <linux/uio.h>
20
21 static const struct file_operations fuse_direct_io_file_operations;
22
23 static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
24                           int opcode, 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 = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
31         if (!fc->atomic_o_trunc)
32                 inarg.flags &= ~O_TRUNC;
33         args.in.h.opcode = opcode;
34         args.in.h.nodeid = nodeid;
35         args.in.numargs = 1;
36         args.in.args[0].size = sizeof(inarg);
37         args.in.args[0].value = &inarg;
38         args.out.numargs = 1;
39         args.out.args[0].size = sizeof(*outargp);
40         args.out.args[0].value = outargp;
41
42         return fuse_simple_request(fc, &args);
43 }
44
45 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
46 {
47         struct fuse_file *ff;
48
49         ff = kzalloc(sizeof(struct fuse_file), GFP_KERNEL);
50         if (unlikely(!ff))
51                 return NULL;
52
53         ff->fc = fc;
54         ff->reserved_req = fuse_request_alloc(0);
55         if (unlikely(!ff->reserved_req)) {
56                 kfree(ff);
57                 return NULL;
58         }
59
60         INIT_LIST_HEAD(&ff->write_entry);
61         refcount_set(&ff->count, 1);
62         RB_CLEAR_NODE(&ff->polled_node);
63         init_waitqueue_head(&ff->poll_wait);
64
65         spin_lock(&fc->lock);
66         ff->kh = ++fc->khctr;
67         spin_unlock(&fc->lock);
68
69         return ff;
70 }
71
72 void fuse_file_free(struct fuse_file *ff)
73 {
74         fuse_request_free(ff->reserved_req);
75         kfree(ff);
76 }
77
78 static struct fuse_file *fuse_file_get(struct fuse_file *ff)
79 {
80         refcount_inc(&ff->count);
81         return ff;
82 }
83
84 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
85 {
86         iput(req->misc.release.inode);
87 }
88
89 static void fuse_file_put(struct fuse_file *ff, bool sync)
90 {
91         if (refcount_dec_and_test(&ff->count)) {
92                 struct fuse_req *req = ff->reserved_req;
93
94                 if (ff->fc->no_open) {
95                         /*
96                          * Drop the release request when client does not
97                          * implement 'open'
98                          */
99                         __clear_bit(FR_BACKGROUND, &req->flags);
100                         iput(req->misc.release.inode);
101                         fuse_put_request(ff->fc, req);
102                 } else if (sync) {
103                         __set_bit(FR_FORCE, &req->flags);
104                         __clear_bit(FR_BACKGROUND, &req->flags);
105                         fuse_request_send(ff->fc, req);
106                         iput(req->misc.release.inode);
107                         fuse_put_request(ff->fc, req);
108                 } else {
109                         req->end = fuse_release_end;
110                         __set_bit(FR_BACKGROUND, &req->flags);
111                         fuse_request_send_background(ff->fc, req);
112                 }
113                 kfree(ff);
114         }
115 }
116
117 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
118                  bool isdir)
119 {
120         struct fuse_file *ff;
121         int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
122
123         ff = fuse_file_alloc(fc);
124         if (!ff)
125                 return -ENOMEM;
126
127         ff->fh = 0;
128         ff->open_flags = FOPEN_KEEP_CACHE; /* Default for no-open */
129         if (!fc->no_open || isdir) {
130                 struct fuse_open_out outarg;
131                 int err;
132
133                 err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
134                 if (!err) {
135                         ff->fh = outarg.fh;
136                         ff->open_flags = outarg.open_flags;
137
138                 } else if (err != -ENOSYS || isdir) {
139                         fuse_file_free(ff);
140                         return err;
141                 } else {
142                         fc->no_open = 1;
143                 }
144         }
145
146         if (isdir)
147                 ff->open_flags &= ~FOPEN_DIRECT_IO;
148
149         ff->nodeid = nodeid;
150         file->private_data = ff;
151
152         return 0;
153 }
154 EXPORT_SYMBOL_GPL(fuse_do_open);
155
156 static void fuse_link_write_file(struct file *file)
157 {
158         struct inode *inode = file_inode(file);
159         struct fuse_conn *fc = get_fuse_conn(inode);
160         struct fuse_inode *fi = get_fuse_inode(inode);
161         struct fuse_file *ff = file->private_data;
162         /*
163          * file may be written through mmap, so chain it onto the
164          * inodes's write_file list
165          */
166         spin_lock(&fc->lock);
167         if (list_empty(&ff->write_entry))
168                 list_add(&ff->write_entry, &fi->write_files);
169         spin_unlock(&fc->lock);
170 }
171
172 void fuse_finish_open(struct inode *inode, struct file *file)
173 {
174         struct fuse_file *ff = file->private_data;
175         struct fuse_conn *fc = get_fuse_conn(inode);
176
177         if (ff->open_flags & FOPEN_DIRECT_IO)
178                 file->f_op = &fuse_direct_io_file_operations;
179         if (!(ff->open_flags & FOPEN_KEEP_CACHE))
180                 invalidate_inode_pages2(inode->i_mapping);
181         if (ff->open_flags & FOPEN_NONSEEKABLE)
182                 nonseekable_open(inode, file);
183         if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
184                 struct fuse_inode *fi = get_fuse_inode(inode);
185
186                 spin_lock(&fc->lock);
187                 fi->attr_version = ++fc->attr_version;
188                 i_size_write(inode, 0);
189                 spin_unlock(&fc->lock);
190                 fuse_invalidate_attr(inode);
191                 if (fc->writeback_cache)
192                         file_update_time(file);
193         }
194         if ((file->f_mode & FMODE_WRITE) && fc->writeback_cache)
195                 fuse_link_write_file(file);
196 }
197
198 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
199 {
200         struct fuse_conn *fc = get_fuse_conn(inode);
201         int err;
202         bool lock_inode = (file->f_flags & O_TRUNC) &&
203                           fc->atomic_o_trunc &&
204                           fc->writeback_cache;
205
206         err = generic_file_open(inode, file);
207         if (err)
208                 return err;
209
210         if (lock_inode)
211                 inode_lock(inode);
212
213         err = fuse_do_open(fc, get_node_id(inode), file, isdir);
214
215         if (!err)
216                 fuse_finish_open(inode, file);
217
218         if (lock_inode)
219                 inode_unlock(inode);
220
221         return err;
222 }
223
224 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
225 {
226         struct fuse_conn *fc = ff->fc;
227         struct fuse_req *req = ff->reserved_req;
228         struct fuse_release_in *inarg = &req->misc.release.in;
229
230         spin_lock(&fc->lock);
231         list_del(&ff->write_entry);
232         if (!RB_EMPTY_NODE(&ff->polled_node))
233                 rb_erase(&ff->polled_node, &fc->polled_files);
234         spin_unlock(&fc->lock);
235
236         wake_up_interruptible_all(&ff->poll_wait);
237
238         inarg->fh = ff->fh;
239         inarg->flags = flags;
240         req->in.h.opcode = opcode;
241         req->in.h.nodeid = ff->nodeid;
242         req->in.numargs = 1;
243         req->in.args[0].size = sizeof(struct fuse_release_in);
244         req->in.args[0].value = inarg;
245 }
246
247 void fuse_release_common(struct file *file, int opcode)
248 {
249         struct fuse_file *ff = file->private_data;
250         struct fuse_req *req = ff->reserved_req;
251
252         fuse_prepare_release(ff, file->f_flags, opcode);
253
254         if (ff->flock) {
255                 struct fuse_release_in *inarg = &req->misc.release.in;
256                 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
257                 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
258                                                        (fl_owner_t) file);
259         }
260         /* Hold inode until release is finished */
261         req->misc.release.inode = igrab(file_inode(file));
262
263         /*
264          * Normally this will send the RELEASE request, however if
265          * some asynchronous READ or WRITE requests are outstanding,
266          * the sending will be delayed.
267          *
268          * Make the release synchronous if this is a fuseblk mount,
269          * synchronous RELEASE is allowed (and desirable) in this case
270          * because the server can be trusted not to screw up.
271          */
272         fuse_file_put(ff, ff->fc->destroy_req != NULL);
273 }
274
275 static int fuse_open(struct inode *inode, struct file *file)
276 {
277         return fuse_open_common(inode, file, false);
278 }
279
280 static int fuse_release(struct inode *inode, struct file *file)
281 {
282         struct fuse_conn *fc = get_fuse_conn(inode);
283
284         /* see fuse_vma_close() for !writeback_cache case */
285         if (fc->writeback_cache)
286                 write_inode_now(inode, 1);
287
288         fuse_release_common(file, FUSE_RELEASE);
289
290         /* return value is ignored by VFS */
291         return 0;
292 }
293
294 void fuse_sync_release(struct fuse_file *ff, int flags)
295 {
296         WARN_ON(refcount_read(&ff->count) > 1);
297         fuse_prepare_release(ff, flags, FUSE_RELEASE);
298         /*
299          * iput(NULL) is a no-op and since the refcount is 1 and everything's
300          * synchronous, we are fine with not doing igrab() here"
301          */
302         fuse_file_put(ff, true);
303 }
304 EXPORT_SYMBOL_GPL(fuse_sync_release);
305
306 /*
307  * Scramble the ID space with XTEA, so that the value of the files_struct
308  * pointer is not exposed to userspace.
309  */
310 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
311 {
312         u32 *k = fc->scramble_key;
313         u64 v = (unsigned long) id;
314         u32 v0 = v;
315         u32 v1 = v >> 32;
316         u32 sum = 0;
317         int i;
318
319         for (i = 0; i < 32; i++) {
320                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
321                 sum += 0x9E3779B9;
322                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
323         }
324
325         return (u64) v0 + ((u64) v1 << 32);
326 }
327
328 /*
329  * Check if any page in a range is under writeback
330  *
331  * This is currently done by walking the list of writepage requests
332  * for the inode, which can be pretty inefficient.
333  */
334 static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
335                                    pgoff_t idx_to)
336 {
337         struct fuse_conn *fc = get_fuse_conn(inode);
338         struct fuse_inode *fi = get_fuse_inode(inode);
339         struct fuse_req *req;
340         bool found = false;
341
342         spin_lock(&fc->lock);
343         list_for_each_entry(req, &fi->writepages, writepages_entry) {
344                 pgoff_t curr_index;
345
346                 BUG_ON(req->inode != inode);
347                 curr_index = req->misc.write.in.offset >> PAGE_SHIFT;
348                 if (idx_from < curr_index + req->num_pages &&
349                     curr_index <= idx_to) {
350                         found = true;
351                         break;
352                 }
353         }
354         spin_unlock(&fc->lock);
355
356         return found;
357 }
358
359 static inline bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
360 {
361         return fuse_range_is_writeback(inode, index, index);
362 }
363
364 /*
365  * Wait for page writeback to be completed.
366  *
367  * Since fuse doesn't rely on the VM writeback tracking, this has to
368  * use some other means.
369  */
370 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
371 {
372         struct fuse_inode *fi = get_fuse_inode(inode);
373
374         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
375         return 0;
376 }
377
378 /*
379  * Wait for all pending writepages on the inode to finish.
380  *
381  * This is currently done by blocking further writes with FUSE_NOWRITE
382  * and waiting for all sent writes to complete.
383  *
384  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
385  * could conflict with truncation.
386  */
387 static void fuse_sync_writes(struct inode *inode)
388 {
389         fuse_set_nowrite(inode);
390         fuse_release_nowrite(inode);
391 }
392
393 static int fuse_flush(struct file *file, fl_owner_t id)
394 {
395         struct inode *inode = file_inode(file);
396         struct fuse_conn *fc = get_fuse_conn(inode);
397         struct fuse_file *ff = file->private_data;
398         struct fuse_req *req;
399         struct fuse_flush_in inarg;
400         int err;
401
402         if (is_bad_inode(inode))
403                 return -EIO;
404
405         if (fc->no_flush)
406                 return 0;
407
408         err = write_inode_now(inode, 1);
409         if (err)
410                 return err;
411
412         inode_lock(inode);
413         fuse_sync_writes(inode);
414         inode_unlock(inode);
415
416         err = filemap_check_errors(file->f_mapping);
417         if (err)
418                 return err;
419
420         req = fuse_get_req_nofail_nopages(fc, file);
421         memset(&inarg, 0, sizeof(inarg));
422         inarg.fh = ff->fh;
423         inarg.lock_owner = fuse_lock_owner_id(fc, id);
424         req->in.h.opcode = FUSE_FLUSH;
425         req->in.h.nodeid = get_node_id(inode);
426         req->in.numargs = 1;
427         req->in.args[0].size = sizeof(inarg);
428         req->in.args[0].value = &inarg;
429         __set_bit(FR_FORCE, &req->flags);
430         fuse_request_send(fc, req);
431         err = req->out.h.error;
432         fuse_put_request(fc, req);
433         if (err == -ENOSYS) {
434                 fc->no_flush = 1;
435                 err = 0;
436         }
437         return err;
438 }
439
440 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
441                       int datasync, int isdir)
442 {
443         struct inode *inode = file->f_mapping->host;
444         struct fuse_conn *fc = get_fuse_conn(inode);
445         struct fuse_file *ff = file->private_data;
446         FUSE_ARGS(args);
447         struct fuse_fsync_in inarg;
448         int err;
449
450         if (is_bad_inode(inode))
451                 return -EIO;
452
453         inode_lock(inode);
454
455         /*
456          * Start writeback against all dirty pages of the inode, then
457          * wait for all outstanding writes, before sending the FSYNC
458          * request.
459          */
460         err = file_write_and_wait_range(file, start, end);
461         if (err)
462                 goto out;
463
464         fuse_sync_writes(inode);
465
466         /*
467          * Due to implementation of fuse writeback
468          * file_write_and_wait_range() does not catch errors.
469          * We have to do this directly after fuse_sync_writes()
470          */
471         err = file_check_and_advance_wb_err(file);
472         if (err)
473                 goto out;
474
475         err = sync_inode_metadata(inode, 1);
476         if (err)
477                 goto out;
478
479         if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
480                 goto out;
481
482         memset(&inarg, 0, sizeof(inarg));
483         inarg.fh = ff->fh;
484         inarg.fsync_flags = datasync ? 1 : 0;
485         args.in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
486         args.in.h.nodeid = get_node_id(inode);
487         args.in.numargs = 1;
488         args.in.args[0].size = sizeof(inarg);
489         args.in.args[0].value = &inarg;
490         err = fuse_simple_request(fc, &args);
491         if (err == -ENOSYS) {
492                 if (isdir)
493                         fc->no_fsyncdir = 1;
494                 else
495                         fc->no_fsync = 1;
496                 err = 0;
497         }
498 out:
499         inode_unlock(inode);
500         return err;
501 }
502
503 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
504                       int datasync)
505 {
506         return fuse_fsync_common(file, start, end, datasync, 0);
507 }
508
509 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
510                     size_t count, int opcode)
511 {
512         struct fuse_read_in *inarg = &req->misc.read.in;
513         struct fuse_file *ff = file->private_data;
514
515         inarg->fh = ff->fh;
516         inarg->offset = pos;
517         inarg->size = count;
518         inarg->flags = file->f_flags;
519         req->in.h.opcode = opcode;
520         req->in.h.nodeid = ff->nodeid;
521         req->in.numargs = 1;
522         req->in.args[0].size = sizeof(struct fuse_read_in);
523         req->in.args[0].value = inarg;
524         req->out.argvar = 1;
525         req->out.numargs = 1;
526         req->out.args[0].size = count;
527 }
528
529 static void fuse_release_user_pages(struct fuse_req *req, bool should_dirty)
530 {
531         unsigned i;
532
533         for (i = 0; i < req->num_pages; i++) {
534                 struct page *page = req->pages[i];
535                 if (should_dirty)
536                         set_page_dirty_lock(page);
537                 put_page(page);
538         }
539 }
540
541 static void fuse_io_release(struct kref *kref)
542 {
543         kfree(container_of(kref, struct fuse_io_priv, refcnt));
544 }
545
546 static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
547 {
548         if (io->err)
549                 return io->err;
550
551         if (io->bytes >= 0 && io->write)
552                 return -EIO;
553
554         return io->bytes < 0 ? io->size : io->bytes;
555 }
556
557 /**
558  * In case of short read, the caller sets 'pos' to the position of
559  * actual end of fuse request in IO request. Otherwise, if bytes_requested
560  * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
561  *
562  * An example:
563  * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
564  * both submitted asynchronously. The first of them was ACKed by userspace as
565  * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
566  * second request was ACKed as short, e.g. only 1K was read, resulting in
567  * pos == 33K.
568  *
569  * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
570  * will be equal to the length of the longest contiguous fragment of
571  * transferred data starting from the beginning of IO request.
572  */
573 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
574 {
575         int left;
576
577         spin_lock(&io->lock);
578         if (err)
579                 io->err = io->err ? : err;
580         else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
581                 io->bytes = pos;
582
583         left = --io->reqs;
584         if (!left && io->blocking)
585                 complete(io->done);
586         spin_unlock(&io->lock);
587
588         if (!left && !io->blocking) {
589                 ssize_t res = fuse_get_res_by_io(io);
590
591                 if (res >= 0) {
592                         struct inode *inode = file_inode(io->iocb->ki_filp);
593                         struct fuse_conn *fc = get_fuse_conn(inode);
594                         struct fuse_inode *fi = get_fuse_inode(inode);
595
596                         spin_lock(&fc->lock);
597                         fi->attr_version = ++fc->attr_version;
598                         spin_unlock(&fc->lock);
599                 }
600
601                 io->iocb->ki_complete(io->iocb, res, 0);
602         }
603
604         kref_put(&io->refcnt, fuse_io_release);
605 }
606
607 static void fuse_aio_complete_req(struct fuse_conn *fc, struct fuse_req *req)
608 {
609         struct fuse_io_priv *io = req->io;
610         ssize_t pos = -1;
611
612         fuse_release_user_pages(req, io->should_dirty);
613
614         if (io->write) {
615                 if (req->misc.write.in.size != req->misc.write.out.size)
616                         pos = req->misc.write.in.offset - io->offset +
617                                 req->misc.write.out.size;
618         } else {
619                 if (req->misc.read.in.size != req->out.args[0].size)
620                         pos = req->misc.read.in.offset - io->offset +
621                                 req->out.args[0].size;
622         }
623
624         fuse_aio_complete(io, req->out.h.error, pos);
625 }
626
627 static size_t fuse_async_req_send(struct fuse_conn *fc, struct fuse_req *req,
628                 size_t num_bytes, struct fuse_io_priv *io)
629 {
630         spin_lock(&io->lock);
631         kref_get(&io->refcnt);
632         io->size += num_bytes;
633         io->reqs++;
634         spin_unlock(&io->lock);
635
636         req->io = io;
637         req->end = fuse_aio_complete_req;
638
639         __fuse_get_request(req);
640         fuse_request_send_background(fc, req);
641
642         return num_bytes;
643 }
644
645 static size_t fuse_send_read(struct fuse_req *req, struct fuse_io_priv *io,
646                              loff_t pos, size_t count, fl_owner_t owner)
647 {
648         struct file *file = io->iocb->ki_filp;
649         struct fuse_file *ff = file->private_data;
650         struct fuse_conn *fc = ff->fc;
651
652         fuse_read_fill(req, file, pos, count, FUSE_READ);
653         if (owner != NULL) {
654                 struct fuse_read_in *inarg = &req->misc.read.in;
655
656                 inarg->read_flags |= FUSE_READ_LOCKOWNER;
657                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
658         }
659
660         if (io->async)
661                 return fuse_async_req_send(fc, req, count, io);
662
663         fuse_request_send(fc, req);
664         return req->out.args[0].size;
665 }
666
667 static void fuse_read_update_size(struct inode *inode, loff_t size,
668                                   u64 attr_ver)
669 {
670         struct fuse_conn *fc = get_fuse_conn(inode);
671         struct fuse_inode *fi = get_fuse_inode(inode);
672
673         spin_lock(&fc->lock);
674         if (attr_ver == fi->attr_version && size < inode->i_size &&
675             !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
676                 fi->attr_version = ++fc->attr_version;
677                 i_size_write(inode, size);
678         }
679         spin_unlock(&fc->lock);
680 }
681
682 static void fuse_short_read(struct fuse_req *req, struct inode *inode,
683                             u64 attr_ver)
684 {
685         size_t num_read = req->out.args[0].size;
686         struct fuse_conn *fc = get_fuse_conn(inode);
687
688         if (fc->writeback_cache) {
689                 /*
690                  * A hole in a file. Some data after the hole are in page cache,
691                  * but have not reached the client fs yet. So, the hole is not
692                  * present there.
693                  */
694                 int i;
695                 int start_idx = num_read >> PAGE_SHIFT;
696                 size_t off = num_read & (PAGE_SIZE - 1);
697
698                 for (i = start_idx; i < req->num_pages; i++) {
699                         zero_user_segment(req->pages[i], off, PAGE_SIZE);
700                         off = 0;
701                 }
702         } else {
703                 loff_t pos = page_offset(req->pages[0]) + num_read;
704                 fuse_read_update_size(inode, pos, attr_ver);
705         }
706 }
707
708 static int fuse_do_readpage(struct file *file, struct page *page)
709 {
710         struct kiocb iocb;
711         struct fuse_io_priv io;
712         struct inode *inode = page->mapping->host;
713         struct fuse_conn *fc = get_fuse_conn(inode);
714         struct fuse_req *req;
715         size_t num_read;
716         loff_t pos = page_offset(page);
717         size_t count = PAGE_SIZE;
718         u64 attr_ver;
719         int err;
720
721         /*
722          * Page writeback can extend beyond the lifetime of the
723          * page-cache page, so make sure we read a properly synced
724          * page.
725          */
726         fuse_wait_on_page_writeback(inode, page->index);
727
728         req = fuse_get_req(fc, 1);
729         if (IS_ERR(req))
730                 return PTR_ERR(req);
731
732         attr_ver = fuse_get_attr_version(fc);
733
734         req->out.page_zeroing = 1;
735         req->out.argpages = 1;
736         req->num_pages = 1;
737         req->pages[0] = page;
738         req->page_descs[0].length = count;
739         init_sync_kiocb(&iocb, file);
740         io = (struct fuse_io_priv) FUSE_IO_PRIV_SYNC(&iocb);
741         num_read = fuse_send_read(req, &io, pos, count, NULL);
742         err = req->out.h.error;
743
744         if (!err) {
745                 /*
746                  * Short read means EOF.  If file size is larger, truncate it
747                  */
748                 if (num_read < count)
749                         fuse_short_read(req, inode, attr_ver);
750
751                 SetPageUptodate(page);
752         }
753
754         fuse_put_request(fc, req);
755
756         return err;
757 }
758
759 static int fuse_readpage(struct file *file, struct page *page)
760 {
761         struct inode *inode = page->mapping->host;
762         int err;
763
764         err = -EIO;
765         if (is_bad_inode(inode))
766                 goto out;
767
768         err = fuse_do_readpage(file, page);
769         fuse_invalidate_atime(inode);
770  out:
771         unlock_page(page);
772         return err;
773 }
774
775 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
776 {
777         int i;
778         size_t count = req->misc.read.in.size;
779         size_t num_read = req->out.args[0].size;
780         struct address_space *mapping = NULL;
781
782         for (i = 0; mapping == NULL && i < req->num_pages; i++)
783                 mapping = req->pages[i]->mapping;
784
785         if (mapping) {
786                 struct inode *inode = mapping->host;
787
788                 /*
789                  * Short read means EOF. If file size is larger, truncate it
790                  */
791                 if (!req->out.h.error && num_read < count)
792                         fuse_short_read(req, inode, req->misc.read.attr_ver);
793
794                 fuse_invalidate_atime(inode);
795         }
796
797         for (i = 0; i < req->num_pages; i++) {
798                 struct page *page = req->pages[i];
799                 if (!req->out.h.error)
800                         SetPageUptodate(page);
801                 else
802                         SetPageError(page);
803                 unlock_page(page);
804                 put_page(page);
805         }
806         if (req->ff)
807                 fuse_file_put(req->ff, false);
808 }
809
810 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
811 {
812         struct fuse_file *ff = file->private_data;
813         struct fuse_conn *fc = ff->fc;
814         loff_t pos = page_offset(req->pages[0]);
815         size_t count = req->num_pages << PAGE_SHIFT;
816
817         req->out.argpages = 1;
818         req->out.page_zeroing = 1;
819         req->out.page_replace = 1;
820         fuse_read_fill(req, file, pos, count, FUSE_READ);
821         req->misc.read.attr_ver = fuse_get_attr_version(fc);
822         if (fc->async_read) {
823                 req->ff = fuse_file_get(ff);
824                 req->end = fuse_readpages_end;
825                 fuse_request_send_background(fc, req);
826         } else {
827                 fuse_request_send(fc, req);
828                 fuse_readpages_end(fc, req);
829                 fuse_put_request(fc, req);
830         }
831 }
832
833 struct fuse_fill_data {
834         struct fuse_req *req;
835         struct file *file;
836         struct inode *inode;
837         unsigned nr_pages;
838 };
839
840 static int fuse_readpages_fill(void *_data, struct page *page)
841 {
842         struct fuse_fill_data *data = _data;
843         struct fuse_req *req = data->req;
844         struct inode *inode = data->inode;
845         struct fuse_conn *fc = get_fuse_conn(inode);
846
847         fuse_wait_on_page_writeback(inode, page->index);
848
849         if (req->num_pages &&
850             (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
851              (req->num_pages + 1) * PAGE_SIZE > fc->max_read ||
852              req->pages[req->num_pages - 1]->index + 1 != page->index)) {
853                 int nr_alloc = min_t(unsigned, data->nr_pages,
854                                      FUSE_MAX_PAGES_PER_REQ);
855                 fuse_send_readpages(req, data->file);
856                 if (fc->async_read)
857                         req = fuse_get_req_for_background(fc, nr_alloc);
858                 else
859                         req = fuse_get_req(fc, nr_alloc);
860
861                 data->req = req;
862                 if (IS_ERR(req)) {
863                         unlock_page(page);
864                         return PTR_ERR(req);
865                 }
866         }
867
868         if (WARN_ON(req->num_pages >= req->max_pages)) {
869                 fuse_put_request(fc, req);
870                 return -EIO;
871         }
872
873         get_page(page);
874         req->pages[req->num_pages] = page;
875         req->page_descs[req->num_pages].length = PAGE_SIZE;
876         req->num_pages++;
877         data->nr_pages--;
878         return 0;
879 }
880
881 static int fuse_readpages(struct file *file, struct address_space *mapping,
882                           struct list_head *pages, unsigned nr_pages)
883 {
884         struct inode *inode = mapping->host;
885         struct fuse_conn *fc = get_fuse_conn(inode);
886         struct fuse_fill_data data;
887         int err;
888         int nr_alloc = min_t(unsigned, nr_pages, FUSE_MAX_PAGES_PER_REQ);
889
890         err = -EIO;
891         if (is_bad_inode(inode))
892                 goto out;
893
894         data.file = file;
895         data.inode = inode;
896         if (fc->async_read)
897                 data.req = fuse_get_req_for_background(fc, nr_alloc);
898         else
899                 data.req = fuse_get_req(fc, nr_alloc);
900         data.nr_pages = nr_pages;
901         err = PTR_ERR(data.req);
902         if (IS_ERR(data.req))
903                 goto out;
904
905         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
906         if (!err) {
907                 if (data.req->num_pages)
908                         fuse_send_readpages(data.req, file);
909                 else
910                         fuse_put_request(fc, data.req);
911         }
912 out:
913         return err;
914 }
915
916 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
917 {
918         struct inode *inode = iocb->ki_filp->f_mapping->host;
919         struct fuse_conn *fc = get_fuse_conn(inode);
920
921         /*
922          * In auto invalidate mode, always update attributes on read.
923          * Otherwise, only update if we attempt to read past EOF (to ensure
924          * i_size is up to date).
925          */
926         if (fc->auto_inval_data ||
927             (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
928                 int err;
929                 err = fuse_update_attributes(inode, iocb->ki_filp);
930                 if (err)
931                         return err;
932         }
933
934         return generic_file_read_iter(iocb, to);
935 }
936
937 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
938                             loff_t pos, size_t count)
939 {
940         struct fuse_write_in *inarg = &req->misc.write.in;
941         struct fuse_write_out *outarg = &req->misc.write.out;
942
943         inarg->fh = ff->fh;
944         inarg->offset = pos;
945         inarg->size = count;
946         req->in.h.opcode = FUSE_WRITE;
947         req->in.h.nodeid = ff->nodeid;
948         req->in.numargs = 2;
949         if (ff->fc->minor < 9)
950                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
951         else
952                 req->in.args[0].size = sizeof(struct fuse_write_in);
953         req->in.args[0].value = inarg;
954         req->in.args[1].size = count;
955         req->out.numargs = 1;
956         req->out.args[0].size = sizeof(struct fuse_write_out);
957         req->out.args[0].value = outarg;
958 }
959
960 static size_t fuse_send_write(struct fuse_req *req, struct fuse_io_priv *io,
961                               loff_t pos, size_t count, fl_owner_t owner)
962 {
963         struct kiocb *iocb = io->iocb;
964         struct file *file = iocb->ki_filp;
965         struct fuse_file *ff = file->private_data;
966         struct fuse_conn *fc = ff->fc;
967         struct fuse_write_in *inarg = &req->misc.write.in;
968
969         fuse_write_fill(req, ff, pos, count);
970         inarg->flags = file->f_flags;
971         if (iocb->ki_flags & IOCB_DSYNC)
972                 inarg->flags |= O_DSYNC;
973         if (iocb->ki_flags & IOCB_SYNC)
974                 inarg->flags |= O_SYNC;
975         if (owner != NULL) {
976                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
977                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
978         }
979
980         if (io->async)
981                 return fuse_async_req_send(fc, req, count, io);
982
983         fuse_request_send(fc, req);
984         return req->misc.write.out.size;
985 }
986
987 bool fuse_write_update_size(struct inode *inode, loff_t pos)
988 {
989         struct fuse_conn *fc = get_fuse_conn(inode);
990         struct fuse_inode *fi = get_fuse_inode(inode);
991         bool ret = false;
992
993         spin_lock(&fc->lock);
994         fi->attr_version = ++fc->attr_version;
995         if (pos > inode->i_size) {
996                 i_size_write(inode, pos);
997                 ret = true;
998         }
999         spin_unlock(&fc->lock);
1000
1001         return ret;
1002 }
1003
1004 static size_t fuse_send_write_pages(struct fuse_req *req, struct kiocb *iocb,
1005                                     struct inode *inode, loff_t pos,
1006                                     size_t count)
1007 {
1008         size_t res;
1009         unsigned offset;
1010         unsigned i;
1011         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1012
1013         for (i = 0; i < req->num_pages; i++)
1014                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
1015
1016         res = fuse_send_write(req, &io, pos, count, NULL);
1017
1018         offset = req->page_descs[0].offset;
1019         count = res;
1020         for (i = 0; i < req->num_pages; i++) {
1021                 struct page *page = req->pages[i];
1022
1023                 if (!req->out.h.error && !offset && count >= PAGE_SIZE)
1024                         SetPageUptodate(page);
1025
1026                 if (count > PAGE_SIZE - offset)
1027                         count -= PAGE_SIZE - offset;
1028                 else
1029                         count = 0;
1030                 offset = 0;
1031
1032                 unlock_page(page);
1033                 put_page(page);
1034         }
1035
1036         return res;
1037 }
1038
1039 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
1040                                struct address_space *mapping,
1041                                struct iov_iter *ii, loff_t pos)
1042 {
1043         struct fuse_conn *fc = get_fuse_conn(mapping->host);
1044         unsigned offset = pos & (PAGE_SIZE - 1);
1045         size_t count = 0;
1046         int err;
1047
1048         req->in.argpages = 1;
1049         req->page_descs[0].offset = offset;
1050
1051         do {
1052                 size_t tmp;
1053                 struct page *page;
1054                 pgoff_t index = pos >> PAGE_SHIFT;
1055                 size_t bytes = min_t(size_t, PAGE_SIZE - offset,
1056                                      iov_iter_count(ii));
1057
1058                 bytes = min_t(size_t, bytes, fc->max_write - count);
1059
1060  again:
1061                 err = -EFAULT;
1062                 if (iov_iter_fault_in_readable(ii, bytes))
1063                         break;
1064
1065                 err = -ENOMEM;
1066                 page = grab_cache_page_write_begin(mapping, index, 0);
1067                 if (!page)
1068                         break;
1069
1070                 if (mapping_writably_mapped(mapping))
1071                         flush_dcache_page(page);
1072
1073                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
1074                 flush_dcache_page(page);
1075
1076                 iov_iter_advance(ii, tmp);
1077                 if (!tmp) {
1078                         unlock_page(page);
1079                         put_page(page);
1080                         bytes = min(bytes, iov_iter_single_seg_count(ii));
1081                         goto again;
1082                 }
1083
1084                 err = 0;
1085                 req->pages[req->num_pages] = page;
1086                 req->page_descs[req->num_pages].length = tmp;
1087                 req->num_pages++;
1088
1089                 count += tmp;
1090                 pos += tmp;
1091                 offset += tmp;
1092                 if (offset == PAGE_SIZE)
1093                         offset = 0;
1094
1095                 if (!fc->big_writes)
1096                         break;
1097         } while (iov_iter_count(ii) && count < fc->max_write &&
1098                  req->num_pages < req->max_pages && offset == 0);
1099
1100         return count > 0 ? count : err;
1101 }
1102
1103 static inline unsigned fuse_wr_pages(loff_t pos, size_t len)
1104 {
1105         return min_t(unsigned,
1106                      ((pos + len - 1) >> PAGE_SHIFT) -
1107                      (pos >> PAGE_SHIFT) + 1,
1108                      FUSE_MAX_PAGES_PER_REQ);
1109 }
1110
1111 static ssize_t fuse_perform_write(struct kiocb *iocb,
1112                                   struct address_space *mapping,
1113                                   struct iov_iter *ii, loff_t pos)
1114 {
1115         struct inode *inode = mapping->host;
1116         struct fuse_conn *fc = get_fuse_conn(inode);
1117         struct fuse_inode *fi = get_fuse_inode(inode);
1118         int err = 0;
1119         ssize_t res = 0;
1120
1121         if (is_bad_inode(inode))
1122                 return -EIO;
1123
1124         if (inode->i_size < pos + iov_iter_count(ii))
1125                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1126
1127         do {
1128                 struct fuse_req *req;
1129                 ssize_t count;
1130                 unsigned nr_pages = fuse_wr_pages(pos, iov_iter_count(ii));
1131
1132                 req = fuse_get_req(fc, nr_pages);
1133                 if (IS_ERR(req)) {
1134                         err = PTR_ERR(req);
1135                         break;
1136                 }
1137
1138                 count = fuse_fill_write_pages(req, mapping, ii, pos);
1139                 if (count <= 0) {
1140                         err = count;
1141                 } else {
1142                         size_t num_written;
1143
1144                         num_written = fuse_send_write_pages(req, iocb, inode,
1145                                                             pos, count);
1146                         err = req->out.h.error;
1147                         if (!err) {
1148                                 res += num_written;
1149                                 pos += num_written;
1150
1151                                 /* break out of the loop on short write */
1152                                 if (num_written != count)
1153                                         err = -EIO;
1154                         }
1155                 }
1156                 fuse_put_request(fc, req);
1157         } while (!err && iov_iter_count(ii));
1158
1159         if (res > 0)
1160                 fuse_write_update_size(inode, pos);
1161
1162         clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1163         fuse_invalidate_attr(inode);
1164
1165         return res > 0 ? res : err;
1166 }
1167
1168 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1169 {
1170         struct file *file = iocb->ki_filp;
1171         struct address_space *mapping = file->f_mapping;
1172         ssize_t written = 0;
1173         ssize_t written_buffered = 0;
1174         struct inode *inode = mapping->host;
1175         ssize_t err;
1176         loff_t endbyte = 0;
1177
1178         if (get_fuse_conn(inode)->writeback_cache) {
1179                 /* Update size (EOF optimization) and mode (SUID clearing) */
1180                 err = fuse_update_attributes(mapping->host, file);
1181                 if (err)
1182                         return err;
1183
1184                 return generic_file_write_iter(iocb, from);
1185         }
1186
1187         inode_lock(inode);
1188
1189         /* We can write back this queue in page reclaim */
1190         current->backing_dev_info = inode_to_bdi(inode);
1191
1192         err = generic_write_checks(iocb, from);
1193         if (err <= 0)
1194                 goto out;
1195
1196         err = file_remove_privs(file);
1197         if (err)
1198                 goto out;
1199
1200         err = file_update_time(file);
1201         if (err)
1202                 goto out;
1203
1204         if (iocb->ki_flags & IOCB_DIRECT) {
1205                 loff_t pos = iocb->ki_pos;
1206                 written = generic_file_direct_write(iocb, from);
1207                 if (written < 0 || !iov_iter_count(from))
1208                         goto out;
1209
1210                 pos += written;
1211
1212                 written_buffered = fuse_perform_write(iocb, mapping, from, pos);
1213                 if (written_buffered < 0) {
1214                         err = written_buffered;
1215                         goto out;
1216                 }
1217                 endbyte = pos + written_buffered - 1;
1218
1219                 err = filemap_write_and_wait_range(file->f_mapping, pos,
1220                                                    endbyte);
1221                 if (err)
1222                         goto out;
1223
1224                 invalidate_mapping_pages(file->f_mapping,
1225                                          pos >> PAGE_SHIFT,
1226                                          endbyte >> PAGE_SHIFT);
1227
1228                 written += written_buffered;
1229                 iocb->ki_pos = pos + written_buffered;
1230         } else {
1231                 written = fuse_perform_write(iocb, mapping, from, iocb->ki_pos);
1232                 if (written >= 0)
1233                         iocb->ki_pos += written;
1234         }
1235 out:
1236         current->backing_dev_info = NULL;
1237         inode_unlock(inode);
1238         if (written > 0)
1239                 written = generic_write_sync(iocb, written);
1240
1241         return written ? written : err;
1242 }
1243
1244 static inline void fuse_page_descs_length_init(struct fuse_req *req,
1245                 unsigned index, unsigned nr_pages)
1246 {
1247         int i;
1248
1249         for (i = index; i < index + nr_pages; i++)
1250                 req->page_descs[i].length = PAGE_SIZE -
1251                         req->page_descs[i].offset;
1252 }
1253
1254 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1255 {
1256         return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1257 }
1258
1259 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1260                                         size_t max_size)
1261 {
1262         return min(iov_iter_single_seg_count(ii), max_size);
1263 }
1264
1265 static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1266                                size_t *nbytesp, int write)
1267 {
1268         size_t nbytes = 0;  /* # bytes already packed in req */
1269         ssize_t ret = 0;
1270
1271         /* Special case for kernel I/O: can copy directly into the buffer */
1272         if (ii->type & ITER_KVEC) {
1273                 unsigned long user_addr = fuse_get_user_addr(ii);
1274                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1275
1276                 if (write)
1277                         req->in.args[1].value = (void *) user_addr;
1278                 else
1279                         req->out.args[0].value = (void *) user_addr;
1280
1281                 iov_iter_advance(ii, frag_size);
1282                 *nbytesp = frag_size;
1283                 return 0;
1284         }
1285
1286         while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1287                 unsigned npages;
1288                 size_t start;
1289                 ret = iov_iter_get_pages(ii, &req->pages[req->num_pages],
1290                                         *nbytesp - nbytes,
1291                                         req->max_pages - req->num_pages,
1292                                         &start);
1293                 if (ret < 0)
1294                         break;
1295
1296                 iov_iter_advance(ii, ret);
1297                 nbytes += ret;
1298
1299                 ret += start;
1300                 npages = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
1301
1302                 req->page_descs[req->num_pages].offset = start;
1303                 fuse_page_descs_length_init(req, req->num_pages, npages);
1304
1305                 req->num_pages += npages;
1306                 req->page_descs[req->num_pages - 1].length -=
1307                         (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1308         }
1309
1310         if (write)
1311                 req->in.argpages = 1;
1312         else
1313                 req->out.argpages = 1;
1314
1315         *nbytesp = nbytes;
1316
1317         return ret < 0 ? ret : 0;
1318 }
1319
1320 static inline int fuse_iter_npages(const struct iov_iter *ii_p)
1321 {
1322         return iov_iter_npages(ii_p, FUSE_MAX_PAGES_PER_REQ);
1323 }
1324
1325 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1326                        loff_t *ppos, int flags)
1327 {
1328         int write = flags & FUSE_DIO_WRITE;
1329         int cuse = flags & FUSE_DIO_CUSE;
1330         struct file *file = io->iocb->ki_filp;
1331         struct inode *inode = file->f_mapping->host;
1332         struct fuse_file *ff = file->private_data;
1333         struct fuse_conn *fc = ff->fc;
1334         size_t nmax = write ? fc->max_write : fc->max_read;
1335         loff_t pos = *ppos;
1336         size_t count = iov_iter_count(iter);
1337         pgoff_t idx_from = pos >> PAGE_SHIFT;
1338         pgoff_t idx_to = (pos + count - 1) >> PAGE_SHIFT;
1339         ssize_t res = 0;
1340         struct fuse_req *req;
1341         int err = 0;
1342
1343         if (io->async)
1344                 req = fuse_get_req_for_background(fc, fuse_iter_npages(iter));
1345         else
1346                 req = fuse_get_req(fc, fuse_iter_npages(iter));
1347         if (IS_ERR(req))
1348                 return PTR_ERR(req);
1349
1350         if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1351                 if (!write)
1352                         inode_lock(inode);
1353                 fuse_sync_writes(inode);
1354                 if (!write)
1355                         inode_unlock(inode);
1356         }
1357
1358         io->should_dirty = !write && iter_is_iovec(iter);
1359         while (count) {
1360                 size_t nres;
1361                 fl_owner_t owner = current->files;
1362                 size_t nbytes = min(count, nmax);
1363                 err = fuse_get_user_pages(req, iter, &nbytes, write);
1364                 if (err && !nbytes)
1365                         break;
1366
1367                 if (write)
1368                         nres = fuse_send_write(req, io, pos, nbytes, owner);
1369                 else
1370                         nres = fuse_send_read(req, io, pos, nbytes, owner);
1371
1372                 if (!io->async)
1373                         fuse_release_user_pages(req, io->should_dirty);
1374                 if (req->out.h.error) {
1375                         err = req->out.h.error;
1376                         break;
1377                 } else if (nres > nbytes) {
1378                         res = 0;
1379                         err = -EIO;
1380                         break;
1381                 }
1382                 count -= nres;
1383                 res += nres;
1384                 pos += nres;
1385                 if (nres != nbytes)
1386                         break;
1387                 if (count) {
1388                         fuse_put_request(fc, req);
1389                         if (io->async)
1390                                 req = fuse_get_req_for_background(fc,
1391                                         fuse_iter_npages(iter));
1392                         else
1393                                 req = fuse_get_req(fc, fuse_iter_npages(iter));
1394                         if (IS_ERR(req))
1395                                 break;
1396                 }
1397         }
1398         if (!IS_ERR(req))
1399                 fuse_put_request(fc, req);
1400         if (res > 0)
1401                 *ppos = pos;
1402
1403         return res > 0 ? res : err;
1404 }
1405 EXPORT_SYMBOL_GPL(fuse_direct_io);
1406
1407 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1408                                   struct iov_iter *iter,
1409                                   loff_t *ppos)
1410 {
1411         ssize_t res;
1412         struct inode *inode = file_inode(io->iocb->ki_filp);
1413
1414         if (is_bad_inode(inode))
1415                 return -EIO;
1416
1417         res = fuse_direct_io(io, iter, ppos, 0);
1418
1419         fuse_invalidate_attr(inode);
1420
1421         return res;
1422 }
1423
1424 static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1425 {
1426         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1427         return __fuse_direct_read(&io, to, &iocb->ki_pos);
1428 }
1429
1430 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1431 {
1432         struct inode *inode = file_inode(iocb->ki_filp);
1433         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1434         ssize_t res;
1435
1436         if (is_bad_inode(inode))
1437                 return -EIO;
1438
1439         /* Don't allow parallel writes to the same file */
1440         inode_lock(inode);
1441         res = generic_write_checks(iocb, from);
1442         if (res > 0)
1443                 res = fuse_direct_io(&io, from, &iocb->ki_pos, FUSE_DIO_WRITE);
1444         fuse_invalidate_attr(inode);
1445         if (res > 0)
1446                 fuse_write_update_size(inode, iocb->ki_pos);
1447         inode_unlock(inode);
1448
1449         return res;
1450 }
1451
1452 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1453 {
1454         int i;
1455
1456         for (i = 0; i < req->num_pages; i++)
1457                 __free_page(req->pages[i]);
1458
1459         if (req->ff)
1460                 fuse_file_put(req->ff, false);
1461 }
1462
1463 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1464 {
1465         struct inode *inode = req->inode;
1466         struct fuse_inode *fi = get_fuse_inode(inode);
1467         struct backing_dev_info *bdi = inode_to_bdi(inode);
1468         int i;
1469
1470         list_del(&req->writepages_entry);
1471         for (i = 0; i < req->num_pages; i++) {
1472                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1473                 dec_node_page_state(req->pages[i], NR_WRITEBACK_TEMP);
1474                 wb_writeout_inc(&bdi->wb);
1475         }
1476         wake_up(&fi->page_waitq);
1477 }
1478
1479 /* Called under fc->lock, may release and reacquire it */
1480 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req,
1481                                 loff_t size)
1482 __releases(fc->lock)
1483 __acquires(fc->lock)
1484 {
1485         struct fuse_inode *fi = get_fuse_inode(req->inode);
1486         struct fuse_write_in *inarg = &req->misc.write.in;
1487         __u64 data_size = req->num_pages * PAGE_SIZE;
1488
1489         if (!fc->connected)
1490                 goto out_free;
1491
1492         if (inarg->offset + data_size <= size) {
1493                 inarg->size = data_size;
1494         } else if (inarg->offset < size) {
1495                 inarg->size = size - inarg->offset;
1496         } else {
1497                 /* Got truncated off completely */
1498                 goto out_free;
1499         }
1500
1501         req->in.args[1].size = inarg->size;
1502         fi->writectr++;
1503         fuse_request_send_background_locked(fc, req);
1504         return;
1505
1506  out_free:
1507         fuse_writepage_finish(fc, req);
1508         spin_unlock(&fc->lock);
1509         fuse_writepage_free(fc, req);
1510         fuse_put_request(fc, req);
1511         spin_lock(&fc->lock);
1512 }
1513
1514 /*
1515  * If fi->writectr is positive (no truncate or fsync going on) send
1516  * all queued writepage requests.
1517  *
1518  * Called with fc->lock
1519  */
1520 void fuse_flush_writepages(struct inode *inode)
1521 __releases(fc->lock)
1522 __acquires(fc->lock)
1523 {
1524         struct fuse_conn *fc = get_fuse_conn(inode);
1525         struct fuse_inode *fi = get_fuse_inode(inode);
1526         size_t crop = i_size_read(inode);
1527         struct fuse_req *req;
1528
1529         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1530                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1531                 list_del_init(&req->list);
1532                 fuse_send_writepage(fc, req, crop);
1533         }
1534 }
1535
1536 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1537 {
1538         struct inode *inode = req->inode;
1539         struct fuse_inode *fi = get_fuse_inode(inode);
1540
1541         mapping_set_error(inode->i_mapping, req->out.h.error);
1542         spin_lock(&fc->lock);
1543         while (req->misc.write.next) {
1544                 struct fuse_conn *fc = get_fuse_conn(inode);
1545                 struct fuse_write_in *inarg = &req->misc.write.in;
1546                 struct fuse_req *next = req->misc.write.next;
1547                 req->misc.write.next = next->misc.write.next;
1548                 next->misc.write.next = NULL;
1549                 next->ff = fuse_file_get(req->ff);
1550                 list_add(&next->writepages_entry, &fi->writepages);
1551
1552                 /*
1553                  * Skip fuse_flush_writepages() to make it easy to crop requests
1554                  * based on primary request size.
1555                  *
1556                  * 1st case (trivial): there are no concurrent activities using
1557                  * fuse_set/release_nowrite.  Then we're on safe side because
1558                  * fuse_flush_writepages() would call fuse_send_writepage()
1559                  * anyway.
1560                  *
1561                  * 2nd case: someone called fuse_set_nowrite and it is waiting
1562                  * now for completion of all in-flight requests.  This happens
1563                  * rarely and no more than once per page, so this should be
1564                  * okay.
1565                  *
1566                  * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1567                  * of fuse_set_nowrite..fuse_release_nowrite section.  The fact
1568                  * that fuse_set_nowrite returned implies that all in-flight
1569                  * requests were completed along with all of their secondary
1570                  * requests.  Further primary requests are blocked by negative
1571                  * writectr.  Hence there cannot be any in-flight requests and
1572                  * no invocations of fuse_writepage_end() while we're in
1573                  * fuse_set_nowrite..fuse_release_nowrite section.
1574                  */
1575                 fuse_send_writepage(fc, next, inarg->offset + inarg->size);
1576         }
1577         fi->writectr--;
1578         fuse_writepage_finish(fc, req);
1579         spin_unlock(&fc->lock);
1580         fuse_writepage_free(fc, req);
1581 }
1582
1583 static struct fuse_file *__fuse_write_file_get(struct fuse_conn *fc,
1584                                                struct fuse_inode *fi)
1585 {
1586         struct fuse_file *ff = NULL;
1587
1588         spin_lock(&fc->lock);
1589         if (!list_empty(&fi->write_files)) {
1590                 ff = list_entry(fi->write_files.next, struct fuse_file,
1591                                 write_entry);
1592                 fuse_file_get(ff);
1593         }
1594         spin_unlock(&fc->lock);
1595
1596         return ff;
1597 }
1598
1599 static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1600                                              struct fuse_inode *fi)
1601 {
1602         struct fuse_file *ff = __fuse_write_file_get(fc, fi);
1603         WARN_ON(!ff);
1604         return ff;
1605 }
1606
1607 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1608 {
1609         struct fuse_conn *fc = get_fuse_conn(inode);
1610         struct fuse_inode *fi = get_fuse_inode(inode);
1611         struct fuse_file *ff;
1612         int err;
1613
1614         ff = __fuse_write_file_get(fc, fi);
1615         err = fuse_flush_times(inode, ff);
1616         if (ff)
1617                 fuse_file_put(ff, 0);
1618
1619         return err;
1620 }
1621
1622 static int fuse_writepage_locked(struct page *page)
1623 {
1624         struct address_space *mapping = page->mapping;
1625         struct inode *inode = mapping->host;
1626         struct fuse_conn *fc = get_fuse_conn(inode);
1627         struct fuse_inode *fi = get_fuse_inode(inode);
1628         struct fuse_req *req;
1629         struct page *tmp_page;
1630         int error = -ENOMEM;
1631
1632         set_page_writeback(page);
1633
1634         req = fuse_request_alloc_nofs(1);
1635         if (!req)
1636                 goto err;
1637
1638         /* writeback always goes to bg_queue */
1639         __set_bit(FR_BACKGROUND, &req->flags);
1640         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1641         if (!tmp_page)
1642                 goto err_free;
1643
1644         error = -EIO;
1645         req->ff = fuse_write_file_get(fc, fi);
1646         if (!req->ff)
1647                 goto err_nofile;
1648
1649         fuse_write_fill(req, req->ff, page_offset(page), 0);
1650
1651         copy_highpage(tmp_page, page);
1652         req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1653         req->misc.write.next = NULL;
1654         req->in.argpages = 1;
1655         req->num_pages = 1;
1656         req->pages[0] = tmp_page;
1657         req->page_descs[0].offset = 0;
1658         req->page_descs[0].length = PAGE_SIZE;
1659         req->end = fuse_writepage_end;
1660         req->inode = inode;
1661
1662         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1663         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1664
1665         spin_lock(&fc->lock);
1666         list_add(&req->writepages_entry, &fi->writepages);
1667         list_add_tail(&req->list, &fi->queued_writes);
1668         fuse_flush_writepages(inode);
1669         spin_unlock(&fc->lock);
1670
1671         end_page_writeback(page);
1672
1673         return 0;
1674
1675 err_nofile:
1676         __free_page(tmp_page);
1677 err_free:
1678         fuse_request_free(req);
1679 err:
1680         mapping_set_error(page->mapping, error);
1681         end_page_writeback(page);
1682         return error;
1683 }
1684
1685 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1686 {
1687         int err;
1688
1689         if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1690                 /*
1691                  * ->writepages() should be called for sync() and friends.  We
1692                  * should only get here on direct reclaim and then we are
1693                  * allowed to skip a page which is already in flight
1694                  */
1695                 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1696
1697                 redirty_page_for_writepage(wbc, page);
1698                 return 0;
1699         }
1700
1701         err = fuse_writepage_locked(page);
1702         unlock_page(page);
1703
1704         return err;
1705 }
1706
1707 struct fuse_fill_wb_data {
1708         struct fuse_req *req;
1709         struct fuse_file *ff;
1710         struct inode *inode;
1711         struct page **orig_pages;
1712 };
1713
1714 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1715 {
1716         struct fuse_req *req = data->req;
1717         struct inode *inode = data->inode;
1718         struct fuse_conn *fc = get_fuse_conn(inode);
1719         struct fuse_inode *fi = get_fuse_inode(inode);
1720         int num_pages = req->num_pages;
1721         int i;
1722
1723         req->ff = fuse_file_get(data->ff);
1724         spin_lock(&fc->lock);
1725         list_add_tail(&req->list, &fi->queued_writes);
1726         fuse_flush_writepages(inode);
1727         spin_unlock(&fc->lock);
1728
1729         for (i = 0; i < num_pages; i++)
1730                 end_page_writeback(data->orig_pages[i]);
1731 }
1732
1733 static bool fuse_writepage_in_flight(struct fuse_req *new_req,
1734                                      struct page *page)
1735 {
1736         struct fuse_conn *fc = get_fuse_conn(new_req->inode);
1737         struct fuse_inode *fi = get_fuse_inode(new_req->inode);
1738         struct fuse_req *tmp;
1739         struct fuse_req *old_req;
1740         bool found = false;
1741         pgoff_t curr_index;
1742
1743         BUG_ON(new_req->num_pages != 0);
1744
1745         spin_lock(&fc->lock);
1746         list_del(&new_req->writepages_entry);
1747         list_for_each_entry(old_req, &fi->writepages, writepages_entry) {
1748                 BUG_ON(old_req->inode != new_req->inode);
1749                 curr_index = old_req->misc.write.in.offset >> PAGE_SHIFT;
1750                 if (curr_index <= page->index &&
1751                     page->index < curr_index + old_req->num_pages) {
1752                         found = true;
1753                         break;
1754                 }
1755         }
1756         if (!found) {
1757                 list_add(&new_req->writepages_entry, &fi->writepages);
1758                 goto out_unlock;
1759         }
1760
1761         new_req->num_pages = 1;
1762         for (tmp = old_req; tmp != NULL; tmp = tmp->misc.write.next) {
1763                 BUG_ON(tmp->inode != new_req->inode);
1764                 curr_index = tmp->misc.write.in.offset >> PAGE_SHIFT;
1765                 if (tmp->num_pages == 1 &&
1766                     curr_index == page->index) {
1767                         old_req = tmp;
1768                 }
1769         }
1770
1771         if (old_req->num_pages == 1 && test_bit(FR_PENDING, &old_req->flags)) {
1772                 struct backing_dev_info *bdi = inode_to_bdi(page->mapping->host);
1773
1774                 copy_highpage(old_req->pages[0], page);
1775                 spin_unlock(&fc->lock);
1776
1777                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1778                 dec_node_page_state(page, NR_WRITEBACK_TEMP);
1779                 wb_writeout_inc(&bdi->wb);
1780                 fuse_writepage_free(fc, new_req);
1781                 fuse_request_free(new_req);
1782                 goto out;
1783         } else {
1784                 new_req->misc.write.next = old_req->misc.write.next;
1785                 old_req->misc.write.next = new_req;
1786         }
1787 out_unlock:
1788         spin_unlock(&fc->lock);
1789 out:
1790         return found;
1791 }
1792
1793 static int fuse_writepages_fill(struct page *page,
1794                 struct writeback_control *wbc, void *_data)
1795 {
1796         struct fuse_fill_wb_data *data = _data;
1797         struct fuse_req *req = data->req;
1798         struct inode *inode = data->inode;
1799         struct fuse_conn *fc = get_fuse_conn(inode);
1800         struct page *tmp_page;
1801         bool is_writeback;
1802         int err;
1803
1804         if (!data->ff) {
1805                 err = -EIO;
1806                 data->ff = fuse_write_file_get(fc, get_fuse_inode(inode));
1807                 if (!data->ff)
1808                         goto out_unlock;
1809         }
1810
1811         /*
1812          * Being under writeback is unlikely but possible.  For example direct
1813          * read to an mmaped fuse file will set the page dirty twice; once when
1814          * the pages are faulted with get_user_pages(), and then after the read
1815          * completed.
1816          */
1817         is_writeback = fuse_page_is_writeback(inode, page->index);
1818
1819         if (req && req->num_pages &&
1820             (is_writeback || req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
1821              (req->num_pages + 1) * PAGE_SIZE > fc->max_write ||
1822              data->orig_pages[req->num_pages - 1]->index + 1 != page->index)) {
1823                 fuse_writepages_send(data);
1824                 data->req = NULL;
1825         }
1826         err = -ENOMEM;
1827         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1828         if (!tmp_page)
1829                 goto out_unlock;
1830
1831         /*
1832          * The page must not be redirtied until the writeout is completed
1833          * (i.e. userspace has sent a reply to the write request).  Otherwise
1834          * there could be more than one temporary page instance for each real
1835          * page.
1836          *
1837          * This is ensured by holding the page lock in page_mkwrite() while
1838          * checking fuse_page_is_writeback().  We already hold the page lock
1839          * since clear_page_dirty_for_io() and keep it held until we add the
1840          * request to the fi->writepages list and increment req->num_pages.
1841          * After this fuse_page_is_writeback() will indicate that the page is
1842          * under writeback, so we can release the page lock.
1843          */
1844         if (data->req == NULL) {
1845                 struct fuse_inode *fi = get_fuse_inode(inode);
1846
1847                 err = -ENOMEM;
1848                 req = fuse_request_alloc_nofs(FUSE_MAX_PAGES_PER_REQ);
1849                 if (!req) {
1850                         __free_page(tmp_page);
1851                         goto out_unlock;
1852                 }
1853
1854                 fuse_write_fill(req, data->ff, page_offset(page), 0);
1855                 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1856                 req->misc.write.next = NULL;
1857                 req->in.argpages = 1;
1858                 __set_bit(FR_BACKGROUND, &req->flags);
1859                 req->num_pages = 0;
1860                 req->end = fuse_writepage_end;
1861                 req->inode = inode;
1862
1863                 spin_lock(&fc->lock);
1864                 list_add(&req->writepages_entry, &fi->writepages);
1865                 spin_unlock(&fc->lock);
1866
1867                 data->req = req;
1868         }
1869         set_page_writeback(page);
1870
1871         copy_highpage(tmp_page, page);
1872         req->pages[req->num_pages] = tmp_page;
1873         req->page_descs[req->num_pages].offset = 0;
1874         req->page_descs[req->num_pages].length = PAGE_SIZE;
1875
1876         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1877         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1878
1879         err = 0;
1880         if (is_writeback && fuse_writepage_in_flight(req, page)) {
1881                 end_page_writeback(page);
1882                 data->req = NULL;
1883                 goto out_unlock;
1884         }
1885         data->orig_pages[req->num_pages] = page;
1886
1887         /*
1888          * Protected by fc->lock against concurrent access by
1889          * fuse_page_is_writeback().
1890          */
1891         spin_lock(&fc->lock);
1892         req->num_pages++;
1893         spin_unlock(&fc->lock);
1894
1895 out_unlock:
1896         unlock_page(page);
1897
1898         return err;
1899 }
1900
1901 static int fuse_writepages(struct address_space *mapping,
1902                            struct writeback_control *wbc)
1903 {
1904         struct inode *inode = mapping->host;
1905         struct fuse_fill_wb_data data;
1906         int err;
1907
1908         err = -EIO;
1909         if (is_bad_inode(inode))
1910                 goto out;
1911
1912         data.inode = inode;
1913         data.req = NULL;
1914         data.ff = NULL;
1915
1916         err = -ENOMEM;
1917         data.orig_pages = kcalloc(FUSE_MAX_PAGES_PER_REQ,
1918                                   sizeof(struct page *),
1919                                   GFP_NOFS);
1920         if (!data.orig_pages)
1921                 goto out;
1922
1923         err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
1924         if (data.req) {
1925                 /* Ignore errors if we can write at least one page */
1926                 BUG_ON(!data.req->num_pages);
1927                 fuse_writepages_send(&data);
1928                 err = 0;
1929         }
1930         if (data.ff)
1931                 fuse_file_put(data.ff, false);
1932
1933         kfree(data.orig_pages);
1934 out:
1935         return err;
1936 }
1937
1938 /*
1939  * It's worthy to make sure that space is reserved on disk for the write,
1940  * but how to implement it without killing performance need more thinking.
1941  */
1942 static int fuse_write_begin(struct file *file, struct address_space *mapping,
1943                 loff_t pos, unsigned len, unsigned flags,
1944                 struct page **pagep, void **fsdata)
1945 {
1946         pgoff_t index = pos >> PAGE_SHIFT;
1947         struct fuse_conn *fc = get_fuse_conn(file_inode(file));
1948         struct page *page;
1949         loff_t fsize;
1950         int err = -ENOMEM;
1951
1952         WARN_ON(!fc->writeback_cache);
1953
1954         page = grab_cache_page_write_begin(mapping, index, flags);
1955         if (!page)
1956                 goto error;
1957
1958         fuse_wait_on_page_writeback(mapping->host, page->index);
1959
1960         if (PageUptodate(page) || len == PAGE_SIZE)
1961                 goto success;
1962         /*
1963          * Check if the start this page comes after the end of file, in which
1964          * case the readpage can be optimized away.
1965          */
1966         fsize = i_size_read(mapping->host);
1967         if (fsize <= (pos & PAGE_MASK)) {
1968                 size_t off = pos & ~PAGE_MASK;
1969                 if (off)
1970                         zero_user_segment(page, 0, off);
1971                 goto success;
1972         }
1973         err = fuse_do_readpage(file, page);
1974         if (err)
1975                 goto cleanup;
1976 success:
1977         *pagep = page;
1978         return 0;
1979
1980 cleanup:
1981         unlock_page(page);
1982         put_page(page);
1983 error:
1984         return err;
1985 }
1986
1987 static int fuse_write_end(struct file *file, struct address_space *mapping,
1988                 loff_t pos, unsigned len, unsigned copied,
1989                 struct page *page, void *fsdata)
1990 {
1991         struct inode *inode = page->mapping->host;
1992
1993         /* Haven't copied anything?  Skip zeroing, size extending, dirtying. */
1994         if (!copied)
1995                 goto unlock;
1996
1997         if (!PageUptodate(page)) {
1998                 /* Zero any unwritten bytes at the end of the page */
1999                 size_t endoff = (pos + copied) & ~PAGE_MASK;
2000                 if (endoff)
2001                         zero_user_segment(page, endoff, PAGE_SIZE);
2002                 SetPageUptodate(page);
2003         }
2004
2005         fuse_write_update_size(inode, pos + copied);
2006         set_page_dirty(page);
2007
2008 unlock:
2009         unlock_page(page);
2010         put_page(page);
2011
2012         return copied;
2013 }
2014
2015 static int fuse_launder_page(struct page *page)
2016 {
2017         int err = 0;
2018         if (clear_page_dirty_for_io(page)) {
2019                 struct inode *inode = page->mapping->host;
2020                 err = fuse_writepage_locked(page);
2021                 if (!err)
2022                         fuse_wait_on_page_writeback(inode, page->index);
2023         }
2024         return err;
2025 }
2026
2027 /*
2028  * Write back dirty pages now, because there may not be any suitable
2029  * open files later
2030  */
2031 static void fuse_vma_close(struct vm_area_struct *vma)
2032 {
2033         filemap_write_and_wait(vma->vm_file->f_mapping);
2034 }
2035
2036 /*
2037  * Wait for writeback against this page to complete before allowing it
2038  * to be marked dirty again, and hence written back again, possibly
2039  * before the previous writepage completed.
2040  *
2041  * Block here, instead of in ->writepage(), so that the userspace fs
2042  * can only block processes actually operating on the filesystem.
2043  *
2044  * Otherwise unprivileged userspace fs would be able to block
2045  * unrelated:
2046  *
2047  * - page migration
2048  * - sync(2)
2049  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2050  */
2051 static int fuse_page_mkwrite(struct vm_fault *vmf)
2052 {
2053         struct page *page = vmf->page;
2054         struct inode *inode = file_inode(vmf->vma->vm_file);
2055
2056         file_update_time(vmf->vma->vm_file);
2057         lock_page(page);
2058         if (page->mapping != inode->i_mapping) {
2059                 unlock_page(page);
2060                 return VM_FAULT_NOPAGE;
2061         }
2062
2063         fuse_wait_on_page_writeback(inode, page->index);
2064         return VM_FAULT_LOCKED;
2065 }
2066
2067 static const struct vm_operations_struct fuse_file_vm_ops = {
2068         .close          = fuse_vma_close,
2069         .fault          = filemap_fault,
2070         .map_pages      = filemap_map_pages,
2071         .page_mkwrite   = fuse_page_mkwrite,
2072 };
2073
2074 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2075 {
2076         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2077                 fuse_link_write_file(file);
2078
2079         file_accessed(file);
2080         vma->vm_ops = &fuse_file_vm_ops;
2081         return 0;
2082 }
2083
2084 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
2085 {
2086         /* Can't provide the coherency needed for MAP_SHARED */
2087         if (vma->vm_flags & VM_MAYSHARE)
2088                 return -ENODEV;
2089
2090         invalidate_inode_pages2(file->f_mapping);
2091
2092         return generic_file_mmap(file, vma);
2093 }
2094
2095 static int convert_fuse_file_lock(struct fuse_conn *fc,
2096                                   const struct fuse_file_lock *ffl,
2097                                   struct file_lock *fl)
2098 {
2099         switch (ffl->type) {
2100         case F_UNLCK:
2101                 break;
2102
2103         case F_RDLCK:
2104         case F_WRLCK:
2105                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2106                     ffl->end < ffl->start)
2107                         return -EIO;
2108
2109                 fl->fl_start = ffl->start;
2110                 fl->fl_end = ffl->end;
2111
2112                 /*
2113                  * Convert pid into init's pid namespace.  The locks API will
2114                  * translate it into the caller's pid namespace.
2115                  */
2116                 rcu_read_lock();
2117                 fl->fl_pid = pid_nr_ns(find_pid_ns(ffl->pid, fc->pid_ns), &init_pid_ns);
2118                 rcu_read_unlock();
2119                 break;
2120
2121         default:
2122                 return -EIO;
2123         }
2124         fl->fl_type = ffl->type;
2125         return 0;
2126 }
2127
2128 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2129                          const struct file_lock *fl, int opcode, pid_t pid,
2130                          int flock, struct fuse_lk_in *inarg)
2131 {
2132         struct inode *inode = file_inode(file);
2133         struct fuse_conn *fc = get_fuse_conn(inode);
2134         struct fuse_file *ff = file->private_data;
2135
2136         memset(inarg, 0, sizeof(*inarg));
2137         inarg->fh = ff->fh;
2138         inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2139         inarg->lk.start = fl->fl_start;
2140         inarg->lk.end = fl->fl_end;
2141         inarg->lk.type = fl->fl_type;
2142         inarg->lk.pid = pid;
2143         if (flock)
2144                 inarg->lk_flags |= FUSE_LK_FLOCK;
2145         args->in.h.opcode = opcode;
2146         args->in.h.nodeid = get_node_id(inode);
2147         args->in.numargs = 1;
2148         args->in.args[0].size = sizeof(*inarg);
2149         args->in.args[0].value = inarg;
2150 }
2151
2152 static int fuse_getlk(struct file *file, struct file_lock *fl)
2153 {
2154         struct inode *inode = file_inode(file);
2155         struct fuse_conn *fc = get_fuse_conn(inode);
2156         FUSE_ARGS(args);
2157         struct fuse_lk_in inarg;
2158         struct fuse_lk_out outarg;
2159         int err;
2160
2161         fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2162         args.out.numargs = 1;
2163         args.out.args[0].size = sizeof(outarg);
2164         args.out.args[0].value = &outarg;
2165         err = fuse_simple_request(fc, &args);
2166         if (!err)
2167                 err = convert_fuse_file_lock(fc, &outarg.lk, fl);
2168
2169         return err;
2170 }
2171
2172 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2173 {
2174         struct inode *inode = file_inode(file);
2175         struct fuse_conn *fc = get_fuse_conn(inode);
2176         FUSE_ARGS(args);
2177         struct fuse_lk_in inarg;
2178         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2179         struct pid *pid = fl->fl_type != F_UNLCK ? task_tgid(current) : NULL;
2180         pid_t pid_nr = pid_nr_ns(pid, fc->pid_ns);
2181         int err;
2182
2183         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2184                 /* NLM needs asynchronous locks, which we don't support yet */
2185                 return -ENOLCK;
2186         }
2187
2188         /* Unlock on close is handled by the flush method */
2189         if ((fl->fl_flags & FL_CLOSE_POSIX) == FL_CLOSE_POSIX)
2190                 return 0;
2191
2192         fuse_lk_fill(&args, file, fl, opcode, pid_nr, flock, &inarg);
2193         err = fuse_simple_request(fc, &args);
2194
2195         /* locking is restartable */
2196         if (err == -EINTR)
2197                 err = -ERESTARTSYS;
2198
2199         return err;
2200 }
2201
2202 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2203 {
2204         struct inode *inode = file_inode(file);
2205         struct fuse_conn *fc = get_fuse_conn(inode);
2206         int err;
2207
2208         if (cmd == F_CANCELLK) {
2209                 err = 0;
2210         } else if (cmd == F_GETLK) {
2211                 if (fc->no_lock) {
2212                         posix_test_lock(file, fl);
2213                         err = 0;
2214                 } else
2215                         err = fuse_getlk(file, fl);
2216         } else {
2217                 if (fc->no_lock)
2218                         err = posix_lock_file(file, fl, NULL);
2219                 else
2220                         err = fuse_setlk(file, fl, 0);
2221         }
2222         return err;
2223 }
2224
2225 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2226 {
2227         struct inode *inode = file_inode(file);
2228         struct fuse_conn *fc = get_fuse_conn(inode);
2229         int err;
2230
2231         if (fc->no_flock) {
2232                 err = locks_lock_file_wait(file, fl);
2233         } else {
2234                 struct fuse_file *ff = file->private_data;
2235
2236                 /* emulate flock with POSIX locks */
2237                 ff->flock = true;
2238                 err = fuse_setlk(file, fl, 1);
2239         }
2240
2241         return err;
2242 }
2243
2244 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2245 {
2246         struct inode *inode = mapping->host;
2247         struct fuse_conn *fc = get_fuse_conn(inode);
2248         FUSE_ARGS(args);
2249         struct fuse_bmap_in inarg;
2250         struct fuse_bmap_out outarg;
2251         int err;
2252
2253         if (!inode->i_sb->s_bdev || fc->no_bmap)
2254                 return 0;
2255
2256         memset(&inarg, 0, sizeof(inarg));
2257         inarg.block = block;
2258         inarg.blocksize = inode->i_sb->s_blocksize;
2259         args.in.h.opcode = FUSE_BMAP;
2260         args.in.h.nodeid = get_node_id(inode);
2261         args.in.numargs = 1;
2262         args.in.args[0].size = sizeof(inarg);
2263         args.in.args[0].value = &inarg;
2264         args.out.numargs = 1;
2265         args.out.args[0].size = sizeof(outarg);
2266         args.out.args[0].value = &outarg;
2267         err = fuse_simple_request(fc, &args);
2268         if (err == -ENOSYS)
2269                 fc->no_bmap = 1;
2270
2271         return err ? 0 : outarg.block;
2272 }
2273
2274 static loff_t fuse_lseek(struct file *file, loff_t offset, int whence)
2275 {
2276         struct inode *inode = file->f_mapping->host;
2277         struct fuse_conn *fc = get_fuse_conn(inode);
2278         struct fuse_file *ff = file->private_data;
2279         FUSE_ARGS(args);
2280         struct fuse_lseek_in inarg = {
2281                 .fh = ff->fh,
2282                 .offset = offset,
2283                 .whence = whence
2284         };
2285         struct fuse_lseek_out outarg;
2286         int err;
2287
2288         if (fc->no_lseek)
2289                 goto fallback;
2290
2291         args.in.h.opcode = FUSE_LSEEK;
2292         args.in.h.nodeid = ff->nodeid;
2293         args.in.numargs = 1;
2294         args.in.args[0].size = sizeof(inarg);
2295         args.in.args[0].value = &inarg;
2296         args.out.numargs = 1;
2297         args.out.args[0].size = sizeof(outarg);
2298         args.out.args[0].value = &outarg;
2299         err = fuse_simple_request(fc, &args);
2300         if (err) {
2301                 if (err == -ENOSYS) {
2302                         fc->no_lseek = 1;
2303                         goto fallback;
2304                 }
2305                 return err;
2306         }
2307
2308         return vfs_setpos(file, outarg.offset, inode->i_sb->s_maxbytes);
2309
2310 fallback:
2311         err = fuse_update_attributes(inode, file);
2312         if (!err)
2313                 return generic_file_llseek(file, offset, whence);
2314         else
2315                 return err;
2316 }
2317
2318 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2319 {
2320         loff_t retval;
2321         struct inode *inode = file_inode(file);
2322
2323         switch (whence) {
2324         case SEEK_SET:
2325         case SEEK_CUR:
2326                  /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2327                 retval = generic_file_llseek(file, offset, whence);
2328                 break;
2329         case SEEK_END:
2330                 inode_lock(inode);
2331                 retval = fuse_update_attributes(inode, file);
2332                 if (!retval)
2333                         retval = generic_file_llseek(file, offset, whence);
2334                 inode_unlock(inode);
2335                 break;
2336         case SEEK_HOLE:
2337         case SEEK_DATA:
2338                 inode_lock(inode);
2339                 retval = fuse_lseek(file, offset, whence);
2340                 inode_unlock(inode);
2341                 break;
2342         default:
2343                 retval = -EINVAL;
2344         }
2345
2346         return retval;
2347 }
2348
2349 /*
2350  * CUSE servers compiled on 32bit broke on 64bit kernels because the
2351  * ABI was defined to be 'struct iovec' which is different on 32bit
2352  * and 64bit.  Fortunately we can determine which structure the server
2353  * used from the size of the reply.
2354  */
2355 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2356                                      size_t transferred, unsigned count,
2357                                      bool is_compat)
2358 {
2359 #ifdef CONFIG_COMPAT
2360         if (count * sizeof(struct compat_iovec) == transferred) {
2361                 struct compat_iovec *ciov = src;
2362                 unsigned i;
2363
2364                 /*
2365                  * With this interface a 32bit server cannot support
2366                  * non-compat (i.e. ones coming from 64bit apps) ioctl
2367                  * requests
2368                  */
2369                 if (!is_compat)
2370                         return -EINVAL;
2371
2372                 for (i = 0; i < count; i++) {
2373                         dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2374                         dst[i].iov_len = ciov[i].iov_len;
2375                 }
2376                 return 0;
2377         }
2378 #endif
2379
2380         if (count * sizeof(struct iovec) != transferred)
2381                 return -EIO;
2382
2383         memcpy(dst, src, transferred);
2384         return 0;
2385 }
2386
2387 /* Make sure iov_length() won't overflow */
2388 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
2389 {
2390         size_t n;
2391         u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
2392
2393         for (n = 0; n < count; n++, iov++) {
2394                 if (iov->iov_len > (size_t) max)
2395                         return -ENOMEM;
2396                 max -= iov->iov_len;
2397         }
2398         return 0;
2399 }
2400
2401 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2402                                  void *src, size_t transferred, unsigned count,
2403                                  bool is_compat)
2404 {
2405         unsigned i;
2406         struct fuse_ioctl_iovec *fiov = src;
2407
2408         if (fc->minor < 16) {
2409                 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2410                                                  count, is_compat);
2411         }
2412
2413         if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2414                 return -EIO;
2415
2416         for (i = 0; i < count; i++) {
2417                 /* Did the server supply an inappropriate value? */
2418                 if (fiov[i].base != (unsigned long) fiov[i].base ||
2419                     fiov[i].len != (unsigned long) fiov[i].len)
2420                         return -EIO;
2421
2422                 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2423                 dst[i].iov_len = (size_t) fiov[i].len;
2424
2425 #ifdef CONFIG_COMPAT
2426                 if (is_compat &&
2427                     (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2428                      (compat_size_t) dst[i].iov_len != fiov[i].len))
2429                         return -EIO;
2430 #endif
2431         }
2432
2433         return 0;
2434 }
2435
2436
2437 /*
2438  * For ioctls, there is no generic way to determine how much memory
2439  * needs to be read and/or written.  Furthermore, ioctls are allowed
2440  * to dereference the passed pointer, so the parameter requires deep
2441  * copying but FUSE has no idea whatsoever about what to copy in or
2442  * out.
2443  *
2444  * This is solved by allowing FUSE server to retry ioctl with
2445  * necessary in/out iovecs.  Let's assume the ioctl implementation
2446  * needs to read in the following structure.
2447  *
2448  * struct a {
2449  *      char    *buf;
2450  *      size_t  buflen;
2451  * }
2452  *
2453  * On the first callout to FUSE server, inarg->in_size and
2454  * inarg->out_size will be NULL; then, the server completes the ioctl
2455  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2456  * the actual iov array to
2457  *
2458  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
2459  *
2460  * which tells FUSE to copy in the requested area and retry the ioctl.
2461  * On the second round, the server has access to the structure and
2462  * from that it can tell what to look for next, so on the invocation,
2463  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2464  *
2465  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
2466  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
2467  *
2468  * FUSE will copy both struct a and the pointed buffer from the
2469  * process doing the ioctl and retry ioctl with both struct a and the
2470  * buffer.
2471  *
2472  * This time, FUSE server has everything it needs and completes ioctl
2473  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2474  *
2475  * Copying data out works the same way.
2476  *
2477  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2478  * automatically initializes in and out iovs by decoding @cmd with
2479  * _IOC_* macros and the server is not allowed to request RETRY.  This
2480  * limits ioctl data transfers to well-formed ioctls and is the forced
2481  * behavior for all FUSE servers.
2482  */
2483 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2484                    unsigned int flags)
2485 {
2486         struct fuse_file *ff = file->private_data;
2487         struct fuse_conn *fc = ff->fc;
2488         struct fuse_ioctl_in inarg = {
2489                 .fh = ff->fh,
2490                 .cmd = cmd,
2491                 .arg = arg,
2492                 .flags = flags
2493         };
2494         struct fuse_ioctl_out outarg;
2495         struct fuse_req *req = NULL;
2496         struct page **pages = NULL;
2497         struct iovec *iov_page = NULL;
2498         struct iovec *in_iov = NULL, *out_iov = NULL;
2499         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
2500         size_t in_size, out_size, transferred, c;
2501         int err, i;
2502         struct iov_iter ii;
2503
2504 #if BITS_PER_LONG == 32
2505         inarg.flags |= FUSE_IOCTL_32BIT;
2506 #else
2507         if (flags & FUSE_IOCTL_COMPAT)
2508                 inarg.flags |= FUSE_IOCTL_32BIT;
2509 #endif
2510
2511         /* assume all the iovs returned by client always fits in a page */
2512         BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2513
2514         err = -ENOMEM;
2515         pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
2516         iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2517         if (!pages || !iov_page)
2518                 goto out;
2519
2520         /*
2521          * If restricted, initialize IO parameters as encoded in @cmd.
2522          * RETRY from server is not allowed.
2523          */
2524         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2525                 struct iovec *iov = iov_page;
2526
2527                 iov->iov_base = (void __user *)arg;
2528                 iov->iov_len = _IOC_SIZE(cmd);
2529
2530                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2531                         in_iov = iov;
2532                         in_iovs = 1;
2533                 }
2534
2535                 if (_IOC_DIR(cmd) & _IOC_READ) {
2536                         out_iov = iov;
2537                         out_iovs = 1;
2538                 }
2539         }
2540
2541  retry:
2542         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2543         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2544
2545         /*
2546          * Out data can be used either for actual out data or iovs,
2547          * make sure there always is at least one page.
2548          */
2549         out_size = max_t(size_t, out_size, PAGE_SIZE);
2550         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2551
2552         /* make sure there are enough buffer pages and init request with them */
2553         err = -ENOMEM;
2554         if (max_pages > FUSE_MAX_PAGES_PER_REQ)
2555                 goto out;
2556         while (num_pages < max_pages) {
2557                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2558                 if (!pages[num_pages])
2559                         goto out;
2560                 num_pages++;
2561         }
2562
2563         req = fuse_get_req(fc, num_pages);
2564         if (IS_ERR(req)) {
2565                 err = PTR_ERR(req);
2566                 req = NULL;
2567                 goto out;
2568         }
2569         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
2570         req->num_pages = num_pages;
2571         fuse_page_descs_length_init(req, 0, req->num_pages);
2572
2573         /* okay, let's send it to the client */
2574         req->in.h.opcode = FUSE_IOCTL;
2575         req->in.h.nodeid = ff->nodeid;
2576         req->in.numargs = 1;
2577         req->in.args[0].size = sizeof(inarg);
2578         req->in.args[0].value = &inarg;
2579         if (in_size) {
2580                 req->in.numargs++;
2581                 req->in.args[1].size = in_size;
2582                 req->in.argpages = 1;
2583
2584                 err = -EFAULT;
2585                 iov_iter_init(&ii, WRITE, in_iov, in_iovs, in_size);
2586                 for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= num_pages); i++) {
2587                         c = copy_page_from_iter(pages[i], 0, PAGE_SIZE, &ii);
2588                         if (c != PAGE_SIZE && iov_iter_count(&ii))
2589                                 goto out;
2590                 }
2591         }
2592
2593         req->out.numargs = 2;
2594         req->out.args[0].size = sizeof(outarg);
2595         req->out.args[0].value = &outarg;
2596         req->out.args[1].size = out_size;
2597         req->out.argpages = 1;
2598         req->out.argvar = 1;
2599
2600         fuse_request_send(fc, req);
2601         err = req->out.h.error;
2602         transferred = req->out.args[1].size;
2603         fuse_put_request(fc, req);
2604         req = NULL;
2605         if (err)
2606                 goto out;
2607
2608         /* did it ask for retry? */
2609         if (outarg.flags & FUSE_IOCTL_RETRY) {
2610                 void *vaddr;
2611
2612                 /* no retry if in restricted mode */
2613                 err = -EIO;
2614                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2615                         goto out;
2616
2617                 in_iovs = outarg.in_iovs;
2618                 out_iovs = outarg.out_iovs;
2619
2620                 /*
2621                  * Make sure things are in boundary, separate checks
2622                  * are to protect against overflow.
2623                  */
2624                 err = -ENOMEM;
2625                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2626                     out_iovs > FUSE_IOCTL_MAX_IOV ||
2627                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2628                         goto out;
2629
2630                 vaddr = kmap_atomic(pages[0]);
2631                 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2632                                             transferred, in_iovs + out_iovs,
2633                                             (flags & FUSE_IOCTL_COMPAT) != 0);
2634                 kunmap_atomic(vaddr);
2635                 if (err)
2636                         goto out;
2637
2638                 in_iov = iov_page;
2639                 out_iov = in_iov + in_iovs;
2640
2641                 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
2642                 if (err)
2643                         goto out;
2644
2645                 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
2646                 if (err)
2647                         goto out;
2648
2649                 goto retry;
2650         }
2651
2652         err = -EIO;
2653         if (transferred > inarg.out_size)
2654                 goto out;
2655
2656         err = -EFAULT;
2657         iov_iter_init(&ii, READ, out_iov, out_iovs, transferred);
2658         for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= num_pages); i++) {
2659                 c = copy_page_to_iter(pages[i], 0, PAGE_SIZE, &ii);
2660                 if (c != PAGE_SIZE && iov_iter_count(&ii))
2661                         goto out;
2662         }
2663         err = 0;
2664  out:
2665         if (req)
2666                 fuse_put_request(fc, req);
2667         free_page((unsigned long) iov_page);
2668         while (num_pages)
2669                 __free_page(pages[--num_pages]);
2670         kfree(pages);
2671
2672         return err ? err : outarg.result;
2673 }
2674 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2675
2676 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2677                        unsigned long arg, unsigned int flags)
2678 {
2679         struct inode *inode = file_inode(file);
2680         struct fuse_conn *fc = get_fuse_conn(inode);
2681
2682         if (!fuse_allow_current_process(fc))
2683                 return -EACCES;
2684
2685         if (is_bad_inode(inode))
2686                 return -EIO;
2687
2688         return fuse_do_ioctl(file, cmd, arg, flags);
2689 }
2690
2691 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2692                             unsigned long arg)
2693 {
2694         return fuse_ioctl_common(file, cmd, arg, 0);
2695 }
2696
2697 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2698                                    unsigned long arg)
2699 {
2700         return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2701 }
2702
2703 /*
2704  * All files which have been polled are linked to RB tree
2705  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2706  * find the matching one.
2707  */
2708 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2709                                               struct rb_node **parent_out)
2710 {
2711         struct rb_node **link = &fc->polled_files.rb_node;
2712         struct rb_node *last = NULL;
2713
2714         while (*link) {
2715                 struct fuse_file *ff;
2716
2717                 last = *link;
2718                 ff = rb_entry(last, struct fuse_file, polled_node);
2719
2720                 if (kh < ff->kh)
2721                         link = &last->rb_left;
2722                 else if (kh > ff->kh)
2723                         link = &last->rb_right;
2724                 else
2725                         return link;
2726         }
2727
2728         if (parent_out)
2729                 *parent_out = last;
2730         return link;
2731 }
2732
2733 /*
2734  * The file is about to be polled.  Make sure it's on the polled_files
2735  * RB tree.  Note that files once added to the polled_files tree are
2736  * not removed before the file is released.  This is because a file
2737  * polled once is likely to be polled again.
2738  */
2739 static void fuse_register_polled_file(struct fuse_conn *fc,
2740                                       struct fuse_file *ff)
2741 {
2742         spin_lock(&fc->lock);
2743         if (RB_EMPTY_NODE(&ff->polled_node)) {
2744                 struct rb_node **link, *uninitialized_var(parent);
2745
2746                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2747                 BUG_ON(*link);
2748                 rb_link_node(&ff->polled_node, parent, link);
2749                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2750         }
2751         spin_unlock(&fc->lock);
2752 }
2753
2754 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2755 {
2756         struct fuse_file *ff = file->private_data;
2757         struct fuse_conn *fc = ff->fc;
2758         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2759         struct fuse_poll_out outarg;
2760         FUSE_ARGS(args);
2761         int err;
2762
2763         if (fc->no_poll)
2764                 return DEFAULT_POLLMASK;
2765
2766         poll_wait(file, &ff->poll_wait, wait);
2767         inarg.events = (__u32)poll_requested_events(wait);
2768
2769         /*
2770          * Ask for notification iff there's someone waiting for it.
2771          * The client may ignore the flag and always notify.
2772          */
2773         if (waitqueue_active(&ff->poll_wait)) {
2774                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2775                 fuse_register_polled_file(fc, ff);
2776         }
2777
2778         args.in.h.opcode = FUSE_POLL;
2779         args.in.h.nodeid = ff->nodeid;
2780         args.in.numargs = 1;
2781         args.in.args[0].size = sizeof(inarg);
2782         args.in.args[0].value = &inarg;
2783         args.out.numargs = 1;
2784         args.out.args[0].size = sizeof(outarg);
2785         args.out.args[0].value = &outarg;
2786         err = fuse_simple_request(fc, &args);
2787
2788         if (!err)
2789                 return outarg.revents;
2790         if (err == -ENOSYS) {
2791                 fc->no_poll = 1;
2792                 return DEFAULT_POLLMASK;
2793         }
2794         return POLLERR;
2795 }
2796 EXPORT_SYMBOL_GPL(fuse_file_poll);
2797
2798 /*
2799  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2800  * wakes up the poll waiters.
2801  */
2802 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2803                             struct fuse_notify_poll_wakeup_out *outarg)
2804 {
2805         u64 kh = outarg->kh;
2806         struct rb_node **link;
2807
2808         spin_lock(&fc->lock);
2809
2810         link = fuse_find_polled_node(fc, kh, NULL);
2811         if (*link) {
2812                 struct fuse_file *ff;
2813
2814                 ff = rb_entry(*link, struct fuse_file, polled_node);
2815                 wake_up_interruptible_sync(&ff->poll_wait);
2816         }
2817
2818         spin_unlock(&fc->lock);
2819         return 0;
2820 }
2821
2822 static void fuse_do_truncate(struct file *file)
2823 {
2824         struct inode *inode = file->f_mapping->host;
2825         struct iattr attr;
2826
2827         attr.ia_valid = ATTR_SIZE;
2828         attr.ia_size = i_size_read(inode);
2829
2830         attr.ia_file = file;
2831         attr.ia_valid |= ATTR_FILE;
2832
2833         fuse_do_setattr(file_dentry(file), &attr, file);
2834 }
2835
2836 static inline loff_t fuse_round_up(loff_t off)
2837 {
2838         return round_up(off, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
2839 }
2840
2841 static ssize_t
2842 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
2843 {
2844         DECLARE_COMPLETION_ONSTACK(wait);
2845         ssize_t ret = 0;
2846         struct file *file = iocb->ki_filp;
2847         struct fuse_file *ff = file->private_data;
2848         bool async_dio = ff->fc->async_dio;
2849         loff_t pos = 0;
2850         struct inode *inode;
2851         loff_t i_size;
2852         size_t count = iov_iter_count(iter);
2853         loff_t offset = iocb->ki_pos;
2854         struct fuse_io_priv *io;
2855
2856         pos = offset;
2857         inode = file->f_mapping->host;
2858         i_size = i_size_read(inode);
2859
2860         if ((iov_iter_rw(iter) == READ) && (offset > i_size))
2861                 return 0;
2862
2863         /* optimization for short read */
2864         if (async_dio && iov_iter_rw(iter) != WRITE && offset + count > i_size) {
2865                 if (offset >= i_size)
2866                         return 0;
2867                 iov_iter_truncate(iter, fuse_round_up(i_size - offset));
2868                 count = iov_iter_count(iter);
2869         }
2870
2871         io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2872         if (!io)
2873                 return -ENOMEM;
2874         spin_lock_init(&io->lock);
2875         kref_init(&io->refcnt);
2876         io->reqs = 1;
2877         io->bytes = -1;
2878         io->size = 0;
2879         io->offset = offset;
2880         io->write = (iov_iter_rw(iter) == WRITE);
2881         io->err = 0;
2882         /*
2883          * By default, we want to optimize all I/Os with async request
2884          * submission to the client filesystem if supported.
2885          */
2886         io->async = async_dio;
2887         io->iocb = iocb;
2888         io->blocking = is_sync_kiocb(iocb);
2889
2890         /*
2891          * We cannot asynchronously extend the size of a file.
2892          * In such case the aio will behave exactly like sync io.
2893          */
2894         if ((offset + count > i_size) && iov_iter_rw(iter) == WRITE)
2895                 io->blocking = true;
2896
2897         if (io->async && io->blocking) {
2898                 /*
2899                  * Additional reference to keep io around after
2900                  * calling fuse_aio_complete()
2901                  */
2902                 kref_get(&io->refcnt);
2903                 io->done = &wait;
2904         }
2905
2906         if (iov_iter_rw(iter) == WRITE) {
2907                 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
2908                 fuse_invalidate_attr(inode);
2909         } else {
2910                 ret = __fuse_direct_read(io, iter, &pos);
2911         }
2912
2913         if (io->async) {
2914                 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2915
2916                 /* we have a non-extending, async request, so return */
2917                 if (!io->blocking)
2918                         return -EIOCBQUEUED;
2919
2920                 wait_for_completion(&wait);
2921                 ret = fuse_get_res_by_io(io);
2922         }
2923
2924         kref_put(&io->refcnt, fuse_io_release);
2925
2926         if (iov_iter_rw(iter) == WRITE) {
2927                 if (ret > 0)
2928                         fuse_write_update_size(inode, pos);
2929                 else if (ret < 0 && offset + count > i_size)
2930                         fuse_do_truncate(file);
2931         }
2932
2933         return ret;
2934 }
2935
2936 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2937                                 loff_t length)
2938 {
2939         struct fuse_file *ff = file->private_data;
2940         struct inode *inode = file_inode(file);
2941         struct fuse_inode *fi = get_fuse_inode(inode);
2942         struct fuse_conn *fc = ff->fc;
2943         FUSE_ARGS(args);
2944         struct fuse_fallocate_in inarg = {
2945                 .fh = ff->fh,
2946                 .offset = offset,
2947                 .length = length,
2948                 .mode = mode
2949         };
2950         int err;
2951         bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
2952                            (mode & FALLOC_FL_PUNCH_HOLE);
2953
2954         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2955                 return -EOPNOTSUPP;
2956
2957         if (fc->no_fallocate)
2958                 return -EOPNOTSUPP;
2959
2960         if (lock_inode) {
2961                 inode_lock(inode);
2962                 if (mode & FALLOC_FL_PUNCH_HOLE) {
2963                         loff_t endbyte = offset + length - 1;
2964                         err = filemap_write_and_wait_range(inode->i_mapping,
2965                                                            offset, endbyte);
2966                         if (err)
2967                                 goto out;
2968
2969                         fuse_sync_writes(inode);
2970                 }
2971         }
2972
2973         if (!(mode & FALLOC_FL_KEEP_SIZE))
2974                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2975
2976         args.in.h.opcode = FUSE_FALLOCATE;
2977         args.in.h.nodeid = ff->nodeid;
2978         args.in.numargs = 1;
2979         args.in.args[0].size = sizeof(inarg);
2980         args.in.args[0].value = &inarg;
2981         err = fuse_simple_request(fc, &args);
2982         if (err == -ENOSYS) {
2983                 fc->no_fallocate = 1;
2984                 err = -EOPNOTSUPP;
2985         }
2986         if (err)
2987                 goto out;
2988
2989         /* we could have extended the file */
2990         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
2991                 bool changed = fuse_write_update_size(inode, offset + length);
2992
2993                 if (changed && fc->writeback_cache)
2994                         file_update_time(file);
2995         }
2996
2997         if (mode & FALLOC_FL_PUNCH_HOLE)
2998                 truncate_pagecache_range(inode, offset, offset + length - 1);
2999
3000         fuse_invalidate_attr(inode);
3001
3002 out:
3003         if (!(mode & FALLOC_FL_KEEP_SIZE))
3004                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3005
3006         if (lock_inode)
3007                 inode_unlock(inode);
3008
3009         return err;
3010 }
3011
3012 static const struct file_operations fuse_file_operations = {
3013         .llseek         = fuse_file_llseek,
3014         .read_iter      = fuse_file_read_iter,
3015         .write_iter     = fuse_file_write_iter,
3016         .mmap           = fuse_file_mmap,
3017         .open           = fuse_open,
3018         .flush          = fuse_flush,
3019         .release        = fuse_release,
3020         .fsync          = fuse_fsync,
3021         .lock           = fuse_file_lock,
3022         .flock          = fuse_file_flock,
3023         .splice_read    = generic_file_splice_read,
3024         .unlocked_ioctl = fuse_file_ioctl,
3025         .compat_ioctl   = fuse_file_compat_ioctl,
3026         .poll           = fuse_file_poll,
3027         .fallocate      = fuse_file_fallocate,
3028 };
3029
3030 static const struct file_operations fuse_direct_io_file_operations = {
3031         .llseek         = fuse_file_llseek,
3032         .read_iter      = fuse_direct_read_iter,
3033         .write_iter     = fuse_direct_write_iter,
3034         .mmap           = fuse_direct_mmap,
3035         .open           = fuse_open,
3036         .flush          = fuse_flush,
3037         .release        = fuse_release,
3038         .fsync          = fuse_fsync,
3039         .lock           = fuse_file_lock,
3040         .flock          = fuse_file_flock,
3041         .unlocked_ioctl = fuse_file_ioctl,
3042         .compat_ioctl   = fuse_file_compat_ioctl,
3043         .poll           = fuse_file_poll,
3044         .fallocate      = fuse_file_fallocate,
3045         /* no splice_read */
3046 };
3047
3048 static const struct address_space_operations fuse_file_aops  = {
3049         .readpage       = fuse_readpage,
3050         .writepage      = fuse_writepage,
3051         .writepages     = fuse_writepages,
3052         .launder_page   = fuse_launder_page,
3053         .readpages      = fuse_readpages,
3054         .set_page_dirty = __set_page_dirty_nobuffers,
3055         .bmap           = fuse_bmap,
3056         .direct_IO      = fuse_direct_IO,
3057         .write_begin    = fuse_write_begin,
3058         .write_end      = fuse_write_end,
3059 };
3060
3061 void fuse_init_file_inode(struct inode *inode)
3062 {
3063         inode->i_fop = &fuse_file_operations;
3064         inode->i_data.a_ops = &fuse_file_aops;
3065 }