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 = kmalloc(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         atomic_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         atomic_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 (atomic_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(atomic_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 = filemap_write_and_wait_range(inode->i_mapping, start, end);
461         if (err)
462                 goto out;
463
464         fuse_sync_writes(inode);
465
466         /*
467          * Due to implementation of fuse writeback
468          * filemap_write_and_wait_range() does not catch errors.
469          * We have to do this directly after fuse_sync_writes()
470          */
471         err = filemap_check_errors(file->f_mapping);
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->write);
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->file;
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 fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
711         struct inode *inode = page->mapping->host;
712         struct fuse_conn *fc = get_fuse_conn(inode);
713         struct fuse_req *req;
714         size_t num_read;
715         loff_t pos = page_offset(page);
716         size_t count = PAGE_SIZE;
717         u64 attr_ver;
718         int err;
719
720         /*
721          * Page writeback can extend beyond the lifetime of the
722          * page-cache page, so make sure we read a properly synced
723          * page.
724          */
725         fuse_wait_on_page_writeback(inode, page->index);
726
727         req = fuse_get_req(fc, 1);
728         if (IS_ERR(req))
729                 return PTR_ERR(req);
730
731         attr_ver = fuse_get_attr_version(fc);
732
733         req->out.page_zeroing = 1;
734         req->out.argpages = 1;
735         req->num_pages = 1;
736         req->pages[0] = page;
737         req->page_descs[0].length = count;
738         num_read = fuse_send_read(req, &io, pos, count, NULL);
739         err = req->out.h.error;
740
741         if (!err) {
742                 /*
743                  * Short read means EOF.  If file size is larger, truncate it
744                  */
745                 if (num_read < count)
746                         fuse_short_read(req, inode, attr_ver);
747
748                 SetPageUptodate(page);
749         }
750
751         fuse_put_request(fc, req);
752
753         return err;
754 }
755
756 static int fuse_readpage(struct file *file, struct page *page)
757 {
758         struct inode *inode = page->mapping->host;
759         int err;
760
761         err = -EIO;
762         if (is_bad_inode(inode))
763                 goto out;
764
765         err = fuse_do_readpage(file, page);
766         fuse_invalidate_atime(inode);
767  out:
768         unlock_page(page);
769         return err;
770 }
771
772 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
773 {
774         int i;
775         size_t count = req->misc.read.in.size;
776         size_t num_read = req->out.args[0].size;
777         struct address_space *mapping = NULL;
778
779         for (i = 0; mapping == NULL && i < req->num_pages; i++)
780                 mapping = req->pages[i]->mapping;
781
782         if (mapping) {
783                 struct inode *inode = mapping->host;
784
785                 /*
786                  * Short read means EOF. If file size is larger, truncate it
787                  */
788                 if (!req->out.h.error && num_read < count)
789                         fuse_short_read(req, inode, req->misc.read.attr_ver);
790
791                 fuse_invalidate_atime(inode);
792         }
793
794         for (i = 0; i < req->num_pages; i++) {
795                 struct page *page = req->pages[i];
796                 if (!req->out.h.error)
797                         SetPageUptodate(page);
798                 else
799                         SetPageError(page);
800                 unlock_page(page);
801                 put_page(page);
802         }
803         if (req->ff)
804                 fuse_file_put(req->ff, false);
805 }
806
807 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
808 {
809         struct fuse_file *ff = file->private_data;
810         struct fuse_conn *fc = ff->fc;
811         loff_t pos = page_offset(req->pages[0]);
812         size_t count = req->num_pages << PAGE_SHIFT;
813
814         req->out.argpages = 1;
815         req->out.page_zeroing = 1;
816         req->out.page_replace = 1;
817         fuse_read_fill(req, file, pos, count, FUSE_READ);
818         req->misc.read.attr_ver = fuse_get_attr_version(fc);
819         if (fc->async_read) {
820                 req->ff = fuse_file_get(ff);
821                 req->end = fuse_readpages_end;
822                 fuse_request_send_background(fc, req);
823         } else {
824                 fuse_request_send(fc, req);
825                 fuse_readpages_end(fc, req);
826                 fuse_put_request(fc, req);
827         }
828 }
829
830 struct fuse_fill_data {
831         struct fuse_req *req;
832         struct file *file;
833         struct inode *inode;
834         unsigned nr_pages;
835 };
836
837 static int fuse_readpages_fill(void *_data, struct page *page)
838 {
839         struct fuse_fill_data *data = _data;
840         struct fuse_req *req = data->req;
841         struct inode *inode = data->inode;
842         struct fuse_conn *fc = get_fuse_conn(inode);
843
844         fuse_wait_on_page_writeback(inode, page->index);
845
846         if (req->num_pages &&
847             (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
848              (req->num_pages + 1) * PAGE_SIZE > fc->max_read ||
849              req->pages[req->num_pages - 1]->index + 1 != page->index)) {
850                 int nr_alloc = min_t(unsigned, data->nr_pages,
851                                      FUSE_MAX_PAGES_PER_REQ);
852                 fuse_send_readpages(req, data->file);
853                 if (fc->async_read)
854                         req = fuse_get_req_for_background(fc, nr_alloc);
855                 else
856                         req = fuse_get_req(fc, nr_alloc);
857
858                 data->req = req;
859                 if (IS_ERR(req)) {
860                         unlock_page(page);
861                         return PTR_ERR(req);
862                 }
863         }
864
865         if (WARN_ON(req->num_pages >= req->max_pages)) {
866                 fuse_put_request(fc, req);
867                 return -EIO;
868         }
869
870         get_page(page);
871         req->pages[req->num_pages] = page;
872         req->page_descs[req->num_pages].length = PAGE_SIZE;
873         req->num_pages++;
874         data->nr_pages--;
875         return 0;
876 }
877
878 static int fuse_readpages(struct file *file, struct address_space *mapping,
879                           struct list_head *pages, unsigned nr_pages)
880 {
881         struct inode *inode = mapping->host;
882         struct fuse_conn *fc = get_fuse_conn(inode);
883         struct fuse_fill_data data;
884         int err;
885         int nr_alloc = min_t(unsigned, nr_pages, FUSE_MAX_PAGES_PER_REQ);
886
887         err = -EIO;
888         if (is_bad_inode(inode))
889                 goto out;
890
891         data.file = file;
892         data.inode = inode;
893         if (fc->async_read)
894                 data.req = fuse_get_req_for_background(fc, nr_alloc);
895         else
896                 data.req = fuse_get_req(fc, nr_alloc);
897         data.nr_pages = nr_pages;
898         err = PTR_ERR(data.req);
899         if (IS_ERR(data.req))
900                 goto out;
901
902         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
903         if (!err) {
904                 if (data.req->num_pages)
905                         fuse_send_readpages(data.req, file);
906                 else
907                         fuse_put_request(fc, data.req);
908         }
909 out:
910         return err;
911 }
912
913 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
914 {
915         struct inode *inode = iocb->ki_filp->f_mapping->host;
916         struct fuse_conn *fc = get_fuse_conn(inode);
917
918         /*
919          * In auto invalidate mode, always update attributes on read.
920          * Otherwise, only update if we attempt to read past EOF (to ensure
921          * i_size is up to date).
922          */
923         if (fc->auto_inval_data ||
924             (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
925                 int err;
926                 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
927                 if (err)
928                         return err;
929         }
930
931         return generic_file_read_iter(iocb, to);
932 }
933
934 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
935                             loff_t pos, size_t count)
936 {
937         struct fuse_write_in *inarg = &req->misc.write.in;
938         struct fuse_write_out *outarg = &req->misc.write.out;
939
940         inarg->fh = ff->fh;
941         inarg->offset = pos;
942         inarg->size = count;
943         req->in.h.opcode = FUSE_WRITE;
944         req->in.h.nodeid = ff->nodeid;
945         req->in.numargs = 2;
946         if (ff->fc->minor < 9)
947                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
948         else
949                 req->in.args[0].size = sizeof(struct fuse_write_in);
950         req->in.args[0].value = inarg;
951         req->in.args[1].size = count;
952         req->out.numargs = 1;
953         req->out.args[0].size = sizeof(struct fuse_write_out);
954         req->out.args[0].value = outarg;
955 }
956
957 static size_t fuse_send_write(struct fuse_req *req, struct fuse_io_priv *io,
958                               loff_t pos, size_t count, fl_owner_t owner)
959 {
960         struct file *file = io->file;
961         struct fuse_file *ff = file->private_data;
962         struct fuse_conn *fc = ff->fc;
963         struct fuse_write_in *inarg = &req->misc.write.in;
964
965         fuse_write_fill(req, ff, pos, count);
966         inarg->flags = file->f_flags;
967         if (owner != NULL) {
968                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
969                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
970         }
971
972         if (io->async)
973                 return fuse_async_req_send(fc, req, count, io);
974
975         fuse_request_send(fc, req);
976         return req->misc.write.out.size;
977 }
978
979 bool fuse_write_update_size(struct inode *inode, loff_t pos)
980 {
981         struct fuse_conn *fc = get_fuse_conn(inode);
982         struct fuse_inode *fi = get_fuse_inode(inode);
983         bool ret = false;
984
985         spin_lock(&fc->lock);
986         fi->attr_version = ++fc->attr_version;
987         if (pos > inode->i_size) {
988                 i_size_write(inode, pos);
989                 ret = true;
990         }
991         spin_unlock(&fc->lock);
992
993         return ret;
994 }
995
996 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
997                                     struct inode *inode, loff_t pos,
998                                     size_t count)
999 {
1000         size_t res;
1001         unsigned offset;
1002         unsigned i;
1003         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
1004
1005         for (i = 0; i < req->num_pages; i++)
1006                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
1007
1008         res = fuse_send_write(req, &io, pos, count, NULL);
1009
1010         offset = req->page_descs[0].offset;
1011         count = res;
1012         for (i = 0; i < req->num_pages; i++) {
1013                 struct page *page = req->pages[i];
1014
1015                 if (!req->out.h.error && !offset && count >= PAGE_SIZE)
1016                         SetPageUptodate(page);
1017
1018                 if (count > PAGE_SIZE - offset)
1019                         count -= PAGE_SIZE - offset;
1020                 else
1021                         count = 0;
1022                 offset = 0;
1023
1024                 unlock_page(page);
1025                 put_page(page);
1026         }
1027
1028         return res;
1029 }
1030
1031 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
1032                                struct address_space *mapping,
1033                                struct iov_iter *ii, loff_t pos)
1034 {
1035         struct fuse_conn *fc = get_fuse_conn(mapping->host);
1036         unsigned offset = pos & (PAGE_SIZE - 1);
1037         size_t count = 0;
1038         int err;
1039
1040         req->in.argpages = 1;
1041         req->page_descs[0].offset = offset;
1042
1043         do {
1044                 size_t tmp;
1045                 struct page *page;
1046                 pgoff_t index = pos >> PAGE_SHIFT;
1047                 size_t bytes = min_t(size_t, PAGE_SIZE - offset,
1048                                      iov_iter_count(ii));
1049
1050                 bytes = min_t(size_t, bytes, fc->max_write - count);
1051
1052  again:
1053                 err = -EFAULT;
1054                 if (iov_iter_fault_in_readable(ii, bytes))
1055                         break;
1056
1057                 err = -ENOMEM;
1058                 page = grab_cache_page_write_begin(mapping, index, 0);
1059                 if (!page)
1060                         break;
1061
1062                 if (mapping_writably_mapped(mapping))
1063                         flush_dcache_page(page);
1064
1065                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
1066                 flush_dcache_page(page);
1067
1068                 iov_iter_advance(ii, tmp);
1069                 if (!tmp) {
1070                         unlock_page(page);
1071                         put_page(page);
1072                         bytes = min(bytes, iov_iter_single_seg_count(ii));
1073                         goto again;
1074                 }
1075
1076                 err = 0;
1077                 req->pages[req->num_pages] = page;
1078                 req->page_descs[req->num_pages].length = tmp;
1079                 req->num_pages++;
1080
1081                 count += tmp;
1082                 pos += tmp;
1083                 offset += tmp;
1084                 if (offset == PAGE_SIZE)
1085                         offset = 0;
1086
1087                 if (!fc->big_writes)
1088                         break;
1089         } while (iov_iter_count(ii) && count < fc->max_write &&
1090                  req->num_pages < req->max_pages && offset == 0);
1091
1092         return count > 0 ? count : err;
1093 }
1094
1095 static inline unsigned fuse_wr_pages(loff_t pos, size_t len)
1096 {
1097         return min_t(unsigned,
1098                      ((pos + len - 1) >> PAGE_SHIFT) -
1099                      (pos >> PAGE_SHIFT) + 1,
1100                      FUSE_MAX_PAGES_PER_REQ);
1101 }
1102
1103 static ssize_t fuse_perform_write(struct file *file,
1104                                   struct address_space *mapping,
1105                                   struct iov_iter *ii, loff_t pos)
1106 {
1107         struct inode *inode = mapping->host;
1108         struct fuse_conn *fc = get_fuse_conn(inode);
1109         struct fuse_inode *fi = get_fuse_inode(inode);
1110         int err = 0;
1111         ssize_t res = 0;
1112
1113         if (is_bad_inode(inode))
1114                 return -EIO;
1115
1116         if (inode->i_size < pos + iov_iter_count(ii))
1117                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1118
1119         do {
1120                 struct fuse_req *req;
1121                 ssize_t count;
1122                 unsigned nr_pages = fuse_wr_pages(pos, iov_iter_count(ii));
1123
1124                 req = fuse_get_req(fc, nr_pages);
1125                 if (IS_ERR(req)) {
1126                         err = PTR_ERR(req);
1127                         break;
1128                 }
1129
1130                 count = fuse_fill_write_pages(req, mapping, ii, pos);
1131                 if (count <= 0) {
1132                         err = count;
1133                 } else {
1134                         size_t num_written;
1135
1136                         num_written = fuse_send_write_pages(req, file, inode,
1137                                                             pos, count);
1138                         err = req->out.h.error;
1139                         if (!err) {
1140                                 res += num_written;
1141                                 pos += num_written;
1142
1143                                 /* break out of the loop on short write */
1144                                 if (num_written != count)
1145                                         err = -EIO;
1146                         }
1147                 }
1148                 fuse_put_request(fc, req);
1149         } while (!err && iov_iter_count(ii));
1150
1151         if (res > 0)
1152                 fuse_write_update_size(inode, pos);
1153
1154         clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1155         fuse_invalidate_attr(inode);
1156
1157         return res > 0 ? res : err;
1158 }
1159
1160 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1161 {
1162         struct file *file = iocb->ki_filp;
1163         struct address_space *mapping = file->f_mapping;
1164         ssize_t written = 0;
1165         ssize_t written_buffered = 0;
1166         struct inode *inode = mapping->host;
1167         ssize_t err;
1168         loff_t endbyte = 0;
1169
1170         if (get_fuse_conn(inode)->writeback_cache) {
1171                 /* Update size (EOF optimization) and mode (SUID clearing) */
1172                 err = fuse_update_attributes(mapping->host, NULL, file, NULL);
1173                 if (err)
1174                         return err;
1175
1176                 return generic_file_write_iter(iocb, from);
1177         }
1178
1179         inode_lock(inode);
1180
1181         /* We can write back this queue in page reclaim */
1182         current->backing_dev_info = inode_to_bdi(inode);
1183
1184         err = generic_write_checks(iocb, from);
1185         if (err <= 0)
1186                 goto out;
1187
1188         err = file_remove_privs(file);
1189         if (err)
1190                 goto out;
1191
1192         err = file_update_time(file);
1193         if (err)
1194                 goto out;
1195
1196         if (iocb->ki_flags & IOCB_DIRECT) {
1197                 loff_t pos = iocb->ki_pos;
1198                 written = generic_file_direct_write(iocb, from);
1199                 if (written < 0 || !iov_iter_count(from))
1200                         goto out;
1201
1202                 pos += written;
1203
1204                 written_buffered = fuse_perform_write(file, mapping, from, pos);
1205                 if (written_buffered < 0) {
1206                         err = written_buffered;
1207                         goto out;
1208                 }
1209                 endbyte = pos + written_buffered - 1;
1210
1211                 err = filemap_write_and_wait_range(file->f_mapping, pos,
1212                                                    endbyte);
1213                 if (err)
1214                         goto out;
1215
1216                 invalidate_mapping_pages(file->f_mapping,
1217                                          pos >> PAGE_SHIFT,
1218                                          endbyte >> PAGE_SHIFT);
1219
1220                 written += written_buffered;
1221                 iocb->ki_pos = pos + written_buffered;
1222         } else {
1223                 written = fuse_perform_write(file, mapping, from, iocb->ki_pos);
1224                 if (written >= 0)
1225                         iocb->ki_pos += written;
1226         }
1227 out:
1228         current->backing_dev_info = NULL;
1229         inode_unlock(inode);
1230
1231         return written ? written : err;
1232 }
1233
1234 static inline void fuse_page_descs_length_init(struct fuse_req *req,
1235                 unsigned index, unsigned nr_pages)
1236 {
1237         int i;
1238
1239         for (i = index; i < index + nr_pages; i++)
1240                 req->page_descs[i].length = PAGE_SIZE -
1241                         req->page_descs[i].offset;
1242 }
1243
1244 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1245 {
1246         return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1247 }
1248
1249 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1250                                         size_t max_size)
1251 {
1252         return min(iov_iter_single_seg_count(ii), max_size);
1253 }
1254
1255 static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1256                                size_t *nbytesp, int write)
1257 {
1258         size_t nbytes = 0;  /* # bytes already packed in req */
1259         ssize_t ret = 0;
1260
1261         /* Special case for kernel I/O: can copy directly into the buffer */
1262         if (ii->type & ITER_KVEC) {
1263                 unsigned long user_addr = fuse_get_user_addr(ii);
1264                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1265
1266                 if (write)
1267                         req->in.args[1].value = (void *) user_addr;
1268                 else
1269                         req->out.args[0].value = (void *) user_addr;
1270
1271                 iov_iter_advance(ii, frag_size);
1272                 *nbytesp = frag_size;
1273                 return 0;
1274         }
1275
1276         while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1277                 unsigned npages;
1278                 size_t start;
1279                 ret = iov_iter_get_pages(ii, &req->pages[req->num_pages],
1280                                         *nbytesp - nbytes,
1281                                         req->max_pages - req->num_pages,
1282                                         &start);
1283                 if (ret < 0)
1284                         break;
1285
1286                 iov_iter_advance(ii, ret);
1287                 nbytes += ret;
1288
1289                 ret += start;
1290                 npages = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
1291
1292                 req->page_descs[req->num_pages].offset = start;
1293                 fuse_page_descs_length_init(req, req->num_pages, npages);
1294
1295                 req->num_pages += npages;
1296                 req->page_descs[req->num_pages - 1].length -=
1297                         (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1298         }
1299
1300         if (write)
1301                 req->in.argpages = 1;
1302         else
1303                 req->out.argpages = 1;
1304
1305         *nbytesp = nbytes;
1306
1307         return ret < 0 ? ret : 0;
1308 }
1309
1310 static inline int fuse_iter_npages(const struct iov_iter *ii_p)
1311 {
1312         return iov_iter_npages(ii_p, FUSE_MAX_PAGES_PER_REQ);
1313 }
1314
1315 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1316                        loff_t *ppos, int flags)
1317 {
1318         int write = flags & FUSE_DIO_WRITE;
1319         bool should_dirty = !write && iter_is_iovec(iter);
1320         int cuse = flags & FUSE_DIO_CUSE;
1321         struct file *file = io->file;
1322         struct inode *inode = file->f_mapping->host;
1323         struct fuse_file *ff = file->private_data;
1324         struct fuse_conn *fc = ff->fc;
1325         size_t nmax = write ? fc->max_write : fc->max_read;
1326         loff_t pos = *ppos;
1327         size_t count = iov_iter_count(iter);
1328         pgoff_t idx_from = pos >> PAGE_SHIFT;
1329         pgoff_t idx_to = (pos + count - 1) >> PAGE_SHIFT;
1330         ssize_t res = 0;
1331         struct fuse_req *req;
1332         int err = 0;
1333
1334         if (io->async)
1335                 req = fuse_get_req_for_background(fc, fuse_iter_npages(iter));
1336         else
1337                 req = fuse_get_req(fc, fuse_iter_npages(iter));
1338         if (IS_ERR(req))
1339                 return PTR_ERR(req);
1340
1341         if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1342                 if (!write)
1343                         inode_lock(inode);
1344                 fuse_sync_writes(inode);
1345                 if (!write)
1346                         inode_unlock(inode);
1347         }
1348
1349         while (count) {
1350                 size_t nres;
1351                 fl_owner_t owner = current->files;
1352                 size_t nbytes = min(count, nmax);
1353                 err = fuse_get_user_pages(req, iter, &nbytes, write);
1354                 if (err && !nbytes)
1355                         break;
1356
1357                 if (write)
1358                         nres = fuse_send_write(req, io, pos, nbytes, owner);
1359                 else
1360                         nres = fuse_send_read(req, io, pos, nbytes, owner);
1361
1362                 if (!io->async)
1363                         fuse_release_user_pages(req, should_dirty);
1364                 if (req->out.h.error) {
1365                         err = req->out.h.error;
1366                         break;
1367                 } else if (nres > nbytes) {
1368                         res = 0;
1369                         err = -EIO;
1370                         break;
1371                 }
1372                 count -= nres;
1373                 res += nres;
1374                 pos += nres;
1375                 if (nres != nbytes)
1376                         break;
1377                 if (count) {
1378                         fuse_put_request(fc, req);
1379                         if (io->async)
1380                                 req = fuse_get_req_for_background(fc,
1381                                         fuse_iter_npages(iter));
1382                         else
1383                                 req = fuse_get_req(fc, fuse_iter_npages(iter));
1384                         if (IS_ERR(req))
1385                                 break;
1386                 }
1387         }
1388         if (!IS_ERR(req))
1389                 fuse_put_request(fc, req);
1390         if (res > 0)
1391                 *ppos = pos;
1392
1393         return res > 0 ? res : err;
1394 }
1395 EXPORT_SYMBOL_GPL(fuse_direct_io);
1396
1397 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1398                                   struct iov_iter *iter,
1399                                   loff_t *ppos)
1400 {
1401         ssize_t res;
1402         struct file *file = io->file;
1403         struct inode *inode = file_inode(file);
1404
1405         if (is_bad_inode(inode))
1406                 return -EIO;
1407
1408         res = fuse_direct_io(io, iter, ppos, 0);
1409
1410         fuse_invalidate_attr(inode);
1411
1412         return res;
1413 }
1414
1415 static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1416 {
1417         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb->ki_filp);
1418         return __fuse_direct_read(&io, to, &iocb->ki_pos);
1419 }
1420
1421 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1422 {
1423         struct file *file = iocb->ki_filp;
1424         struct inode *inode = file_inode(file);
1425         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
1426         ssize_t res;
1427
1428         if (is_bad_inode(inode))
1429                 return -EIO;
1430
1431         /* Don't allow parallel writes to the same file */
1432         inode_lock(inode);
1433         res = generic_write_checks(iocb, from);
1434         if (res > 0)
1435                 res = fuse_direct_io(&io, from, &iocb->ki_pos, FUSE_DIO_WRITE);
1436         fuse_invalidate_attr(inode);
1437         if (res > 0)
1438                 fuse_write_update_size(inode, iocb->ki_pos);
1439         inode_unlock(inode);
1440
1441         return res;
1442 }
1443
1444 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1445 {
1446         int i;
1447
1448         for (i = 0; i < req->num_pages; i++)
1449                 __free_page(req->pages[i]);
1450
1451         if (req->ff)
1452                 fuse_file_put(req->ff, false);
1453 }
1454
1455 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1456 {
1457         struct inode *inode = req->inode;
1458         struct fuse_inode *fi = get_fuse_inode(inode);
1459         struct backing_dev_info *bdi = inode_to_bdi(inode);
1460         int i;
1461
1462         list_del(&req->writepages_entry);
1463         for (i = 0; i < req->num_pages; i++) {
1464                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1465                 dec_node_page_state(req->pages[i], NR_WRITEBACK_TEMP);
1466                 wb_writeout_inc(&bdi->wb);
1467         }
1468         wake_up(&fi->page_waitq);
1469 }
1470
1471 /* Called under fc->lock, may release and reacquire it */
1472 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req,
1473                                 loff_t size)
1474 __releases(fc->lock)
1475 __acquires(fc->lock)
1476 {
1477         struct fuse_inode *fi = get_fuse_inode(req->inode);
1478         struct fuse_write_in *inarg = &req->misc.write.in;
1479         __u64 data_size = req->num_pages * PAGE_SIZE;
1480
1481         if (!fc->connected)
1482                 goto out_free;
1483
1484         if (inarg->offset + data_size <= size) {
1485                 inarg->size = data_size;
1486         } else if (inarg->offset < size) {
1487                 inarg->size = size - inarg->offset;
1488         } else {
1489                 /* Got truncated off completely */
1490                 goto out_free;
1491         }
1492
1493         req->in.args[1].size = inarg->size;
1494         fi->writectr++;
1495         fuse_request_send_background_locked(fc, req);
1496         return;
1497
1498  out_free:
1499         fuse_writepage_finish(fc, req);
1500         spin_unlock(&fc->lock);
1501         fuse_writepage_free(fc, req);
1502         fuse_put_request(fc, req);
1503         spin_lock(&fc->lock);
1504 }
1505
1506 /*
1507  * If fi->writectr is positive (no truncate or fsync going on) send
1508  * all queued writepage requests.
1509  *
1510  * Called with fc->lock
1511  */
1512 void fuse_flush_writepages(struct inode *inode)
1513 __releases(fc->lock)
1514 __acquires(fc->lock)
1515 {
1516         struct fuse_conn *fc = get_fuse_conn(inode);
1517         struct fuse_inode *fi = get_fuse_inode(inode);
1518         size_t crop = i_size_read(inode);
1519         struct fuse_req *req;
1520
1521         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1522                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1523                 list_del_init(&req->list);
1524                 fuse_send_writepage(fc, req, crop);
1525         }
1526 }
1527
1528 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1529 {
1530         struct inode *inode = req->inode;
1531         struct fuse_inode *fi = get_fuse_inode(inode);
1532
1533         mapping_set_error(inode->i_mapping, req->out.h.error);
1534         spin_lock(&fc->lock);
1535         while (req->misc.write.next) {
1536                 struct fuse_conn *fc = get_fuse_conn(inode);
1537                 struct fuse_write_in *inarg = &req->misc.write.in;
1538                 struct fuse_req *next = req->misc.write.next;
1539                 req->misc.write.next = next->misc.write.next;
1540                 next->misc.write.next = NULL;
1541                 next->ff = fuse_file_get(req->ff);
1542                 list_add(&next->writepages_entry, &fi->writepages);
1543
1544                 /*
1545                  * Skip fuse_flush_writepages() to make it easy to crop requests
1546                  * based on primary request size.
1547                  *
1548                  * 1st case (trivial): there are no concurrent activities using
1549                  * fuse_set/release_nowrite.  Then we're on safe side because
1550                  * fuse_flush_writepages() would call fuse_send_writepage()
1551                  * anyway.
1552                  *
1553                  * 2nd case: someone called fuse_set_nowrite and it is waiting
1554                  * now for completion of all in-flight requests.  This happens
1555                  * rarely and no more than once per page, so this should be
1556                  * okay.
1557                  *
1558                  * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1559                  * of fuse_set_nowrite..fuse_release_nowrite section.  The fact
1560                  * that fuse_set_nowrite returned implies that all in-flight
1561                  * requests were completed along with all of their secondary
1562                  * requests.  Further primary requests are blocked by negative
1563                  * writectr.  Hence there cannot be any in-flight requests and
1564                  * no invocations of fuse_writepage_end() while we're in
1565                  * fuse_set_nowrite..fuse_release_nowrite section.
1566                  */
1567                 fuse_send_writepage(fc, next, inarg->offset + inarg->size);
1568         }
1569         fi->writectr--;
1570         fuse_writepage_finish(fc, req);
1571         spin_unlock(&fc->lock);
1572         fuse_writepage_free(fc, req);
1573 }
1574
1575 static struct fuse_file *__fuse_write_file_get(struct fuse_conn *fc,
1576                                                struct fuse_inode *fi)
1577 {
1578         struct fuse_file *ff = NULL;
1579
1580         spin_lock(&fc->lock);
1581         if (!list_empty(&fi->write_files)) {
1582                 ff = list_entry(fi->write_files.next, struct fuse_file,
1583                                 write_entry);
1584                 fuse_file_get(ff);
1585         }
1586         spin_unlock(&fc->lock);
1587
1588         return ff;
1589 }
1590
1591 static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1592                                              struct fuse_inode *fi)
1593 {
1594         struct fuse_file *ff = __fuse_write_file_get(fc, fi);
1595         WARN_ON(!ff);
1596         return ff;
1597 }
1598
1599 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1600 {
1601         struct fuse_conn *fc = get_fuse_conn(inode);
1602         struct fuse_inode *fi = get_fuse_inode(inode);
1603         struct fuse_file *ff;
1604         int err;
1605
1606         ff = __fuse_write_file_get(fc, fi);
1607         err = fuse_flush_times(inode, ff);
1608         if (ff)
1609                 fuse_file_put(ff, 0);
1610
1611         return err;
1612 }
1613
1614 static int fuse_writepage_locked(struct page *page)
1615 {
1616         struct address_space *mapping = page->mapping;
1617         struct inode *inode = mapping->host;
1618         struct fuse_conn *fc = get_fuse_conn(inode);
1619         struct fuse_inode *fi = get_fuse_inode(inode);
1620         struct fuse_req *req;
1621         struct page *tmp_page;
1622         int error = -ENOMEM;
1623
1624         set_page_writeback(page);
1625
1626         req = fuse_request_alloc_nofs(1);
1627         if (!req)
1628                 goto err;
1629
1630         /* writeback always goes to bg_queue */
1631         __set_bit(FR_BACKGROUND, &req->flags);
1632         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1633         if (!tmp_page)
1634                 goto err_free;
1635
1636         error = -EIO;
1637         req->ff = fuse_write_file_get(fc, fi);
1638         if (!req->ff)
1639                 goto err_nofile;
1640
1641         fuse_write_fill(req, req->ff, page_offset(page), 0);
1642
1643         copy_highpage(tmp_page, page);
1644         req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1645         req->misc.write.next = NULL;
1646         req->in.argpages = 1;
1647         req->num_pages = 1;
1648         req->pages[0] = tmp_page;
1649         req->page_descs[0].offset = 0;
1650         req->page_descs[0].length = PAGE_SIZE;
1651         req->end = fuse_writepage_end;
1652         req->inode = inode;
1653
1654         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1655         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1656
1657         spin_lock(&fc->lock);
1658         list_add(&req->writepages_entry, &fi->writepages);
1659         list_add_tail(&req->list, &fi->queued_writes);
1660         fuse_flush_writepages(inode);
1661         spin_unlock(&fc->lock);
1662
1663         end_page_writeback(page);
1664
1665         return 0;
1666
1667 err_nofile:
1668         __free_page(tmp_page);
1669 err_free:
1670         fuse_request_free(req);
1671 err:
1672         end_page_writeback(page);
1673         return error;
1674 }
1675
1676 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1677 {
1678         int err;
1679
1680         if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1681                 /*
1682                  * ->writepages() should be called for sync() and friends.  We
1683                  * should only get here on direct reclaim and then we are
1684                  * allowed to skip a page which is already in flight
1685                  */
1686                 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1687
1688                 redirty_page_for_writepage(wbc, page);
1689                 return 0;
1690         }
1691
1692         err = fuse_writepage_locked(page);
1693         unlock_page(page);
1694
1695         return err;
1696 }
1697
1698 struct fuse_fill_wb_data {
1699         struct fuse_req *req;
1700         struct fuse_file *ff;
1701         struct inode *inode;
1702         struct page **orig_pages;
1703 };
1704
1705 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1706 {
1707         struct fuse_req *req = data->req;
1708         struct inode *inode = data->inode;
1709         struct fuse_conn *fc = get_fuse_conn(inode);
1710         struct fuse_inode *fi = get_fuse_inode(inode);
1711         int num_pages = req->num_pages;
1712         int i;
1713
1714         req->ff = fuse_file_get(data->ff);
1715         spin_lock(&fc->lock);
1716         list_add_tail(&req->list, &fi->queued_writes);
1717         fuse_flush_writepages(inode);
1718         spin_unlock(&fc->lock);
1719
1720         for (i = 0; i < num_pages; i++)
1721                 end_page_writeback(data->orig_pages[i]);
1722 }
1723
1724 static bool fuse_writepage_in_flight(struct fuse_req *new_req,
1725                                      struct page *page)
1726 {
1727         struct fuse_conn *fc = get_fuse_conn(new_req->inode);
1728         struct fuse_inode *fi = get_fuse_inode(new_req->inode);
1729         struct fuse_req *tmp;
1730         struct fuse_req *old_req;
1731         bool found = false;
1732         pgoff_t curr_index;
1733
1734         BUG_ON(new_req->num_pages != 0);
1735
1736         spin_lock(&fc->lock);
1737         list_del(&new_req->writepages_entry);
1738         list_for_each_entry(old_req, &fi->writepages, writepages_entry) {
1739                 BUG_ON(old_req->inode != new_req->inode);
1740                 curr_index = old_req->misc.write.in.offset >> PAGE_SHIFT;
1741                 if (curr_index <= page->index &&
1742                     page->index < curr_index + old_req->num_pages) {
1743                         found = true;
1744                         break;
1745                 }
1746         }
1747         if (!found) {
1748                 list_add(&new_req->writepages_entry, &fi->writepages);
1749                 goto out_unlock;
1750         }
1751
1752         new_req->num_pages = 1;
1753         for (tmp = old_req; tmp != NULL; tmp = tmp->misc.write.next) {
1754                 BUG_ON(tmp->inode != new_req->inode);
1755                 curr_index = tmp->misc.write.in.offset >> PAGE_SHIFT;
1756                 if (tmp->num_pages == 1 &&
1757                     curr_index == page->index) {
1758                         old_req = tmp;
1759                 }
1760         }
1761
1762         if (old_req->num_pages == 1 && test_bit(FR_PENDING, &old_req->flags)) {
1763                 struct backing_dev_info *bdi = inode_to_bdi(page->mapping->host);
1764
1765                 copy_highpage(old_req->pages[0], page);
1766                 spin_unlock(&fc->lock);
1767
1768                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1769                 dec_node_page_state(page, NR_WRITEBACK_TEMP);
1770                 wb_writeout_inc(&bdi->wb);
1771                 fuse_writepage_free(fc, new_req);
1772                 fuse_request_free(new_req);
1773                 goto out;
1774         } else {
1775                 new_req->misc.write.next = old_req->misc.write.next;
1776                 old_req->misc.write.next = new_req;
1777         }
1778 out_unlock:
1779         spin_unlock(&fc->lock);
1780 out:
1781         return found;
1782 }
1783
1784 static int fuse_writepages_fill(struct page *page,
1785                 struct writeback_control *wbc, void *_data)
1786 {
1787         struct fuse_fill_wb_data *data = _data;
1788         struct fuse_req *req = data->req;
1789         struct inode *inode = data->inode;
1790         struct fuse_conn *fc = get_fuse_conn(inode);
1791         struct page *tmp_page;
1792         bool is_writeback;
1793         int err;
1794
1795         if (!data->ff) {
1796                 err = -EIO;
1797                 data->ff = fuse_write_file_get(fc, get_fuse_inode(inode));
1798                 if (!data->ff)
1799                         goto out_unlock;
1800         }
1801
1802         /*
1803          * Being under writeback is unlikely but possible.  For example direct
1804          * read to an mmaped fuse file will set the page dirty twice; once when
1805          * the pages are faulted with get_user_pages(), and then after the read
1806          * completed.
1807          */
1808         is_writeback = fuse_page_is_writeback(inode, page->index);
1809
1810         if (req && req->num_pages &&
1811             (is_writeback || req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
1812              (req->num_pages + 1) * PAGE_SIZE > fc->max_write ||
1813              data->orig_pages[req->num_pages - 1]->index + 1 != page->index)) {
1814                 fuse_writepages_send(data);
1815                 data->req = NULL;
1816         }
1817         err = -ENOMEM;
1818         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1819         if (!tmp_page)
1820                 goto out_unlock;
1821
1822         /*
1823          * The page must not be redirtied until the writeout is completed
1824          * (i.e. userspace has sent a reply to the write request).  Otherwise
1825          * there could be more than one temporary page instance for each real
1826          * page.
1827          *
1828          * This is ensured by holding the page lock in page_mkwrite() while
1829          * checking fuse_page_is_writeback().  We already hold the page lock
1830          * since clear_page_dirty_for_io() and keep it held until we add the
1831          * request to the fi->writepages list and increment req->num_pages.
1832          * After this fuse_page_is_writeback() will indicate that the page is
1833          * under writeback, so we can release the page lock.
1834          */
1835         if (data->req == NULL) {
1836                 struct fuse_inode *fi = get_fuse_inode(inode);
1837
1838                 err = -ENOMEM;
1839                 req = fuse_request_alloc_nofs(FUSE_MAX_PAGES_PER_REQ);
1840                 if (!req) {
1841                         __free_page(tmp_page);
1842                         goto out_unlock;
1843                 }
1844
1845                 fuse_write_fill(req, data->ff, page_offset(page), 0);
1846                 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1847                 req->misc.write.next = NULL;
1848                 req->in.argpages = 1;
1849                 __set_bit(FR_BACKGROUND, &req->flags);
1850                 req->num_pages = 0;
1851                 req->end = fuse_writepage_end;
1852                 req->inode = inode;
1853
1854                 spin_lock(&fc->lock);
1855                 list_add(&req->writepages_entry, &fi->writepages);
1856                 spin_unlock(&fc->lock);
1857
1858                 data->req = req;
1859         }
1860         set_page_writeback(page);
1861
1862         copy_highpage(tmp_page, page);
1863         req->pages[req->num_pages] = tmp_page;
1864         req->page_descs[req->num_pages].offset = 0;
1865         req->page_descs[req->num_pages].length = PAGE_SIZE;
1866
1867         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1868         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1869
1870         err = 0;
1871         if (is_writeback && fuse_writepage_in_flight(req, page)) {
1872                 end_page_writeback(page);
1873                 data->req = NULL;
1874                 goto out_unlock;
1875         }
1876         data->orig_pages[req->num_pages] = page;
1877
1878         /*
1879          * Protected by fc->lock against concurrent access by
1880          * fuse_page_is_writeback().
1881          */
1882         spin_lock(&fc->lock);
1883         req->num_pages++;
1884         spin_unlock(&fc->lock);
1885
1886 out_unlock:
1887         unlock_page(page);
1888
1889         return err;
1890 }
1891
1892 static int fuse_writepages(struct address_space *mapping,
1893                            struct writeback_control *wbc)
1894 {
1895         struct inode *inode = mapping->host;
1896         struct fuse_fill_wb_data data;
1897         int err;
1898
1899         err = -EIO;
1900         if (is_bad_inode(inode))
1901                 goto out;
1902
1903         data.inode = inode;
1904         data.req = NULL;
1905         data.ff = NULL;
1906
1907         err = -ENOMEM;
1908         data.orig_pages = kcalloc(FUSE_MAX_PAGES_PER_REQ,
1909                                   sizeof(struct page *),
1910                                   GFP_NOFS);
1911         if (!data.orig_pages)
1912                 goto out;
1913
1914         err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
1915         if (data.req) {
1916                 /* Ignore errors if we can write at least one page */
1917                 BUG_ON(!data.req->num_pages);
1918                 fuse_writepages_send(&data);
1919                 err = 0;
1920         }
1921         if (data.ff)
1922                 fuse_file_put(data.ff, false);
1923
1924         kfree(data.orig_pages);
1925 out:
1926         return err;
1927 }
1928
1929 /*
1930  * It's worthy to make sure that space is reserved on disk for the write,
1931  * but how to implement it without killing performance need more thinking.
1932  */
1933 static int fuse_write_begin(struct file *file, struct address_space *mapping,
1934                 loff_t pos, unsigned len, unsigned flags,
1935                 struct page **pagep, void **fsdata)
1936 {
1937         pgoff_t index = pos >> PAGE_SHIFT;
1938         struct fuse_conn *fc = get_fuse_conn(file_inode(file));
1939         struct page *page;
1940         loff_t fsize;
1941         int err = -ENOMEM;
1942
1943         WARN_ON(!fc->writeback_cache);
1944
1945         page = grab_cache_page_write_begin(mapping, index, flags);
1946         if (!page)
1947                 goto error;
1948
1949         fuse_wait_on_page_writeback(mapping->host, page->index);
1950
1951         if (PageUptodate(page) || len == PAGE_SIZE)
1952                 goto success;
1953         /*
1954          * Check if the start this page comes after the end of file, in which
1955          * case the readpage can be optimized away.
1956          */
1957         fsize = i_size_read(mapping->host);
1958         if (fsize <= (pos & PAGE_MASK)) {
1959                 size_t off = pos & ~PAGE_MASK;
1960                 if (off)
1961                         zero_user_segment(page, 0, off);
1962                 goto success;
1963         }
1964         err = fuse_do_readpage(file, page);
1965         if (err)
1966                 goto cleanup;
1967 success:
1968         *pagep = page;
1969         return 0;
1970
1971 cleanup:
1972         unlock_page(page);
1973         put_page(page);
1974 error:
1975         return err;
1976 }
1977
1978 static int fuse_write_end(struct file *file, struct address_space *mapping,
1979                 loff_t pos, unsigned len, unsigned copied,
1980                 struct page *page, void *fsdata)
1981 {
1982         struct inode *inode = page->mapping->host;
1983
1984         /* Haven't copied anything?  Skip zeroing, size extending, dirtying. */
1985         if (!copied)
1986                 goto unlock;
1987
1988         if (!PageUptodate(page)) {
1989                 /* Zero any unwritten bytes at the end of the page */
1990                 size_t endoff = (pos + copied) & ~PAGE_MASK;
1991                 if (endoff)
1992                         zero_user_segment(page, endoff, PAGE_SIZE);
1993                 SetPageUptodate(page);
1994         }
1995
1996         fuse_write_update_size(inode, pos + copied);
1997         set_page_dirty(page);
1998
1999 unlock:
2000         unlock_page(page);
2001         put_page(page);
2002
2003         return copied;
2004 }
2005
2006 static int fuse_launder_page(struct page *page)
2007 {
2008         int err = 0;
2009         if (clear_page_dirty_for_io(page)) {
2010                 struct inode *inode = page->mapping->host;
2011                 err = fuse_writepage_locked(page);
2012                 if (!err)
2013                         fuse_wait_on_page_writeback(inode, page->index);
2014         }
2015         return err;
2016 }
2017
2018 /*
2019  * Write back dirty pages now, because there may not be any suitable
2020  * open files later
2021  */
2022 static void fuse_vma_close(struct vm_area_struct *vma)
2023 {
2024         filemap_write_and_wait(vma->vm_file->f_mapping);
2025 }
2026
2027 /*
2028  * Wait for writeback against this page to complete before allowing it
2029  * to be marked dirty again, and hence written back again, possibly
2030  * before the previous writepage completed.
2031  *
2032  * Block here, instead of in ->writepage(), so that the userspace fs
2033  * can only block processes actually operating on the filesystem.
2034  *
2035  * Otherwise unprivileged userspace fs would be able to block
2036  * unrelated:
2037  *
2038  * - page migration
2039  * - sync(2)
2040  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2041  */
2042 static int fuse_page_mkwrite(struct vm_fault *vmf)
2043 {
2044         struct page *page = vmf->page;
2045         struct inode *inode = file_inode(vmf->vma->vm_file);
2046
2047         file_update_time(vmf->vma->vm_file);
2048         lock_page(page);
2049         if (page->mapping != inode->i_mapping) {
2050                 unlock_page(page);
2051                 return VM_FAULT_NOPAGE;
2052         }
2053
2054         fuse_wait_on_page_writeback(inode, page->index);
2055         return VM_FAULT_LOCKED;
2056 }
2057
2058 static const struct vm_operations_struct fuse_file_vm_ops = {
2059         .close          = fuse_vma_close,
2060         .fault          = filemap_fault,
2061         .map_pages      = filemap_map_pages,
2062         .page_mkwrite   = fuse_page_mkwrite,
2063 };
2064
2065 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2066 {
2067         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2068                 fuse_link_write_file(file);
2069
2070         file_accessed(file);
2071         vma->vm_ops = &fuse_file_vm_ops;
2072         return 0;
2073 }
2074
2075 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
2076 {
2077         /* Can't provide the coherency needed for MAP_SHARED */
2078         if (vma->vm_flags & VM_MAYSHARE)
2079                 return -ENODEV;
2080
2081         invalidate_inode_pages2(file->f_mapping);
2082
2083         return generic_file_mmap(file, vma);
2084 }
2085
2086 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
2087                                   struct file_lock *fl)
2088 {
2089         switch (ffl->type) {
2090         case F_UNLCK:
2091                 break;
2092
2093         case F_RDLCK:
2094         case F_WRLCK:
2095                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2096                     ffl->end < ffl->start)
2097                         return -EIO;
2098
2099                 fl->fl_start = ffl->start;
2100                 fl->fl_end = ffl->end;
2101                 fl->fl_pid = ffl->pid;
2102                 break;
2103
2104         default:
2105                 return -EIO;
2106         }
2107         fl->fl_type = ffl->type;
2108         return 0;
2109 }
2110
2111 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2112                          const struct file_lock *fl, int opcode, pid_t pid,
2113                          int flock, struct fuse_lk_in *inarg)
2114 {
2115         struct inode *inode = file_inode(file);
2116         struct fuse_conn *fc = get_fuse_conn(inode);
2117         struct fuse_file *ff = file->private_data;
2118
2119         memset(inarg, 0, sizeof(*inarg));
2120         inarg->fh = ff->fh;
2121         inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2122         inarg->lk.start = fl->fl_start;
2123         inarg->lk.end = fl->fl_end;
2124         inarg->lk.type = fl->fl_type;
2125         inarg->lk.pid = pid;
2126         if (flock)
2127                 inarg->lk_flags |= FUSE_LK_FLOCK;
2128         args->in.h.opcode = opcode;
2129         args->in.h.nodeid = get_node_id(inode);
2130         args->in.numargs = 1;
2131         args->in.args[0].size = sizeof(*inarg);
2132         args->in.args[0].value = inarg;
2133 }
2134
2135 static int fuse_getlk(struct file *file, struct file_lock *fl)
2136 {
2137         struct inode *inode = file_inode(file);
2138         struct fuse_conn *fc = get_fuse_conn(inode);
2139         FUSE_ARGS(args);
2140         struct fuse_lk_in inarg;
2141         struct fuse_lk_out outarg;
2142         int err;
2143
2144         fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2145         args.out.numargs = 1;
2146         args.out.args[0].size = sizeof(outarg);
2147         args.out.args[0].value = &outarg;
2148         err = fuse_simple_request(fc, &args);
2149         if (!err)
2150                 err = convert_fuse_file_lock(&outarg.lk, fl);
2151
2152         return err;
2153 }
2154
2155 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2156 {
2157         struct inode *inode = file_inode(file);
2158         struct fuse_conn *fc = get_fuse_conn(inode);
2159         FUSE_ARGS(args);
2160         struct fuse_lk_in inarg;
2161         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2162         pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
2163         int err;
2164
2165         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2166                 /* NLM needs asynchronous locks, which we don't support yet */
2167                 return -ENOLCK;
2168         }
2169
2170         /* Unlock on close is handled by the flush method */
2171         if (fl->fl_flags & FL_CLOSE)
2172                 return 0;
2173
2174         fuse_lk_fill(&args, file, fl, opcode, pid, flock, &inarg);
2175         err = fuse_simple_request(fc, &args);
2176
2177         /* locking is restartable */
2178         if (err == -EINTR)
2179                 err = -ERESTARTSYS;
2180
2181         return err;
2182 }
2183
2184 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2185 {
2186         struct inode *inode = file_inode(file);
2187         struct fuse_conn *fc = get_fuse_conn(inode);
2188         int err;
2189
2190         if (cmd == F_CANCELLK) {
2191                 err = 0;
2192         } else if (cmd == F_GETLK) {
2193                 if (fc->no_lock) {
2194                         posix_test_lock(file, fl);
2195                         err = 0;
2196                 } else
2197                         err = fuse_getlk(file, fl);
2198         } else {
2199                 if (fc->no_lock)
2200                         err = posix_lock_file(file, fl, NULL);
2201                 else
2202                         err = fuse_setlk(file, fl, 0);
2203         }
2204         return err;
2205 }
2206
2207 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2208 {
2209         struct inode *inode = file_inode(file);
2210         struct fuse_conn *fc = get_fuse_conn(inode);
2211         int err;
2212
2213         if (fc->no_flock) {
2214                 err = locks_lock_file_wait(file, fl);
2215         } else {
2216                 struct fuse_file *ff = file->private_data;
2217
2218                 /* emulate flock with POSIX locks */
2219                 ff->flock = true;
2220                 err = fuse_setlk(file, fl, 1);
2221         }
2222
2223         return err;
2224 }
2225
2226 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2227 {
2228         struct inode *inode = mapping->host;
2229         struct fuse_conn *fc = get_fuse_conn(inode);
2230         FUSE_ARGS(args);
2231         struct fuse_bmap_in inarg;
2232         struct fuse_bmap_out outarg;
2233         int err;
2234
2235         if (!inode->i_sb->s_bdev || fc->no_bmap)
2236                 return 0;
2237
2238         memset(&inarg, 0, sizeof(inarg));
2239         inarg.block = block;
2240         inarg.blocksize = inode->i_sb->s_blocksize;
2241         args.in.h.opcode = FUSE_BMAP;
2242         args.in.h.nodeid = get_node_id(inode);
2243         args.in.numargs = 1;
2244         args.in.args[0].size = sizeof(inarg);
2245         args.in.args[0].value = &inarg;
2246         args.out.numargs = 1;
2247         args.out.args[0].size = sizeof(outarg);
2248         args.out.args[0].value = &outarg;
2249         err = fuse_simple_request(fc, &args);
2250         if (err == -ENOSYS)
2251                 fc->no_bmap = 1;
2252
2253         return err ? 0 : outarg.block;
2254 }
2255
2256 static loff_t fuse_lseek(struct file *file, loff_t offset, int whence)
2257 {
2258         struct inode *inode = file->f_mapping->host;
2259         struct fuse_conn *fc = get_fuse_conn(inode);
2260         struct fuse_file *ff = file->private_data;
2261         FUSE_ARGS(args);
2262         struct fuse_lseek_in inarg = {
2263                 .fh = ff->fh,
2264                 .offset = offset,
2265                 .whence = whence
2266         };
2267         struct fuse_lseek_out outarg;
2268         int err;
2269
2270         if (fc->no_lseek)
2271                 goto fallback;
2272
2273         args.in.h.opcode = FUSE_LSEEK;
2274         args.in.h.nodeid = ff->nodeid;
2275         args.in.numargs = 1;
2276         args.in.args[0].size = sizeof(inarg);
2277         args.in.args[0].value = &inarg;
2278         args.out.numargs = 1;
2279         args.out.args[0].size = sizeof(outarg);
2280         args.out.args[0].value = &outarg;
2281         err = fuse_simple_request(fc, &args);
2282         if (err) {
2283                 if (err == -ENOSYS) {
2284                         fc->no_lseek = 1;
2285                         goto fallback;
2286                 }
2287                 return err;
2288         }
2289
2290         return vfs_setpos(file, outarg.offset, inode->i_sb->s_maxbytes);
2291
2292 fallback:
2293         err = fuse_update_attributes(inode, NULL, file, NULL);
2294         if (!err)
2295                 return generic_file_llseek(file, offset, whence);
2296         else
2297                 return err;
2298 }
2299
2300 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2301 {
2302         loff_t retval;
2303         struct inode *inode = file_inode(file);
2304
2305         switch (whence) {
2306         case SEEK_SET:
2307         case SEEK_CUR:
2308                  /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2309                 retval = generic_file_llseek(file, offset, whence);
2310                 break;
2311         case SEEK_END:
2312                 inode_lock(inode);
2313                 retval = fuse_update_attributes(inode, NULL, file, NULL);
2314                 if (!retval)
2315                         retval = generic_file_llseek(file, offset, whence);
2316                 inode_unlock(inode);
2317                 break;
2318         case SEEK_HOLE:
2319         case SEEK_DATA:
2320                 inode_lock(inode);
2321                 retval = fuse_lseek(file, offset, whence);
2322                 inode_unlock(inode);
2323                 break;
2324         default:
2325                 retval = -EINVAL;
2326         }
2327
2328         return retval;
2329 }
2330
2331 /*
2332  * CUSE servers compiled on 32bit broke on 64bit kernels because the
2333  * ABI was defined to be 'struct iovec' which is different on 32bit
2334  * and 64bit.  Fortunately we can determine which structure the server
2335  * used from the size of the reply.
2336  */
2337 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2338                                      size_t transferred, unsigned count,
2339                                      bool is_compat)
2340 {
2341 #ifdef CONFIG_COMPAT
2342         if (count * sizeof(struct compat_iovec) == transferred) {
2343                 struct compat_iovec *ciov = src;
2344                 unsigned i;
2345
2346                 /*
2347                  * With this interface a 32bit server cannot support
2348                  * non-compat (i.e. ones coming from 64bit apps) ioctl
2349                  * requests
2350                  */
2351                 if (!is_compat)
2352                         return -EINVAL;
2353
2354                 for (i = 0; i < count; i++) {
2355                         dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2356                         dst[i].iov_len = ciov[i].iov_len;
2357                 }
2358                 return 0;
2359         }
2360 #endif
2361
2362         if (count * sizeof(struct iovec) != transferred)
2363                 return -EIO;
2364
2365         memcpy(dst, src, transferred);
2366         return 0;
2367 }
2368
2369 /* Make sure iov_length() won't overflow */
2370 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
2371 {
2372         size_t n;
2373         u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
2374
2375         for (n = 0; n < count; n++, iov++) {
2376                 if (iov->iov_len > (size_t) max)
2377                         return -ENOMEM;
2378                 max -= iov->iov_len;
2379         }
2380         return 0;
2381 }
2382
2383 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2384                                  void *src, size_t transferred, unsigned count,
2385                                  bool is_compat)
2386 {
2387         unsigned i;
2388         struct fuse_ioctl_iovec *fiov = src;
2389
2390         if (fc->minor < 16) {
2391                 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2392                                                  count, is_compat);
2393         }
2394
2395         if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2396                 return -EIO;
2397
2398         for (i = 0; i < count; i++) {
2399                 /* Did the server supply an inappropriate value? */
2400                 if (fiov[i].base != (unsigned long) fiov[i].base ||
2401                     fiov[i].len != (unsigned long) fiov[i].len)
2402                         return -EIO;
2403
2404                 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2405                 dst[i].iov_len = (size_t) fiov[i].len;
2406
2407 #ifdef CONFIG_COMPAT
2408                 if (is_compat &&
2409                     (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2410                      (compat_size_t) dst[i].iov_len != fiov[i].len))
2411                         return -EIO;
2412 #endif
2413         }
2414
2415         return 0;
2416 }
2417
2418
2419 /*
2420  * For ioctls, there is no generic way to determine how much memory
2421  * needs to be read and/or written.  Furthermore, ioctls are allowed
2422  * to dereference the passed pointer, so the parameter requires deep
2423  * copying but FUSE has no idea whatsoever about what to copy in or
2424  * out.
2425  *
2426  * This is solved by allowing FUSE server to retry ioctl with
2427  * necessary in/out iovecs.  Let's assume the ioctl implementation
2428  * needs to read in the following structure.
2429  *
2430  * struct a {
2431  *      char    *buf;
2432  *      size_t  buflen;
2433  * }
2434  *
2435  * On the first callout to FUSE server, inarg->in_size and
2436  * inarg->out_size will be NULL; then, the server completes the ioctl
2437  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2438  * the actual iov array to
2439  *
2440  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
2441  *
2442  * which tells FUSE to copy in the requested area and retry the ioctl.
2443  * On the second round, the server has access to the structure and
2444  * from that it can tell what to look for next, so on the invocation,
2445  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2446  *
2447  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
2448  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
2449  *
2450  * FUSE will copy both struct a and the pointed buffer from the
2451  * process doing the ioctl and retry ioctl with both struct a and the
2452  * buffer.
2453  *
2454  * This time, FUSE server has everything it needs and completes ioctl
2455  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2456  *
2457  * Copying data out works the same way.
2458  *
2459  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2460  * automatically initializes in and out iovs by decoding @cmd with
2461  * _IOC_* macros and the server is not allowed to request RETRY.  This
2462  * limits ioctl data transfers to well-formed ioctls and is the forced
2463  * behavior for all FUSE servers.
2464  */
2465 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2466                    unsigned int flags)
2467 {
2468         struct fuse_file *ff = file->private_data;
2469         struct fuse_conn *fc = ff->fc;
2470         struct fuse_ioctl_in inarg = {
2471                 .fh = ff->fh,
2472                 .cmd = cmd,
2473                 .arg = arg,
2474                 .flags = flags
2475         };
2476         struct fuse_ioctl_out outarg;
2477         struct fuse_req *req = NULL;
2478         struct page **pages = NULL;
2479         struct iovec *iov_page = NULL;
2480         struct iovec *in_iov = NULL, *out_iov = NULL;
2481         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
2482         size_t in_size, out_size, transferred, c;
2483         int err, i;
2484         struct iov_iter ii;
2485
2486 #if BITS_PER_LONG == 32
2487         inarg.flags |= FUSE_IOCTL_32BIT;
2488 #else
2489         if (flags & FUSE_IOCTL_COMPAT)
2490                 inarg.flags |= FUSE_IOCTL_32BIT;
2491 #endif
2492
2493         /* assume all the iovs returned by client always fits in a page */
2494         BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2495
2496         err = -ENOMEM;
2497         pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
2498         iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2499         if (!pages || !iov_page)
2500                 goto out;
2501
2502         /*
2503          * If restricted, initialize IO parameters as encoded in @cmd.
2504          * RETRY from server is not allowed.
2505          */
2506         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2507                 struct iovec *iov = iov_page;
2508
2509                 iov->iov_base = (void __user *)arg;
2510                 iov->iov_len = _IOC_SIZE(cmd);
2511
2512                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2513                         in_iov = iov;
2514                         in_iovs = 1;
2515                 }
2516
2517                 if (_IOC_DIR(cmd) & _IOC_READ) {
2518                         out_iov = iov;
2519                         out_iovs = 1;
2520                 }
2521         }
2522
2523  retry:
2524         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2525         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2526
2527         /*
2528          * Out data can be used either for actual out data or iovs,
2529          * make sure there always is at least one page.
2530          */
2531         out_size = max_t(size_t, out_size, PAGE_SIZE);
2532         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2533
2534         /* make sure there are enough buffer pages and init request with them */
2535         err = -ENOMEM;
2536         if (max_pages > FUSE_MAX_PAGES_PER_REQ)
2537                 goto out;
2538         while (num_pages < max_pages) {
2539                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2540                 if (!pages[num_pages])
2541                         goto out;
2542                 num_pages++;
2543         }
2544
2545         req = fuse_get_req(fc, num_pages);
2546         if (IS_ERR(req)) {
2547                 err = PTR_ERR(req);
2548                 req = NULL;
2549                 goto out;
2550         }
2551         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
2552         req->num_pages = num_pages;
2553         fuse_page_descs_length_init(req, 0, req->num_pages);
2554
2555         /* okay, let's send it to the client */
2556         req->in.h.opcode = FUSE_IOCTL;
2557         req->in.h.nodeid = ff->nodeid;
2558         req->in.numargs = 1;
2559         req->in.args[0].size = sizeof(inarg);
2560         req->in.args[0].value = &inarg;
2561         if (in_size) {
2562                 req->in.numargs++;
2563                 req->in.args[1].size = in_size;
2564                 req->in.argpages = 1;
2565
2566                 err = -EFAULT;
2567                 iov_iter_init(&ii, WRITE, in_iov, in_iovs, in_size);
2568                 for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= num_pages); i++) {
2569                         c = copy_page_from_iter(pages[i], 0, PAGE_SIZE, &ii);
2570                         if (c != PAGE_SIZE && iov_iter_count(&ii))
2571                                 goto out;
2572                 }
2573         }
2574
2575         req->out.numargs = 2;
2576         req->out.args[0].size = sizeof(outarg);
2577         req->out.args[0].value = &outarg;
2578         req->out.args[1].size = out_size;
2579         req->out.argpages = 1;
2580         req->out.argvar = 1;
2581
2582         fuse_request_send(fc, req);
2583         err = req->out.h.error;
2584         transferred = req->out.args[1].size;
2585         fuse_put_request(fc, req);
2586         req = NULL;
2587         if (err)
2588                 goto out;
2589
2590         /* did it ask for retry? */
2591         if (outarg.flags & FUSE_IOCTL_RETRY) {
2592                 void *vaddr;
2593
2594                 /* no retry if in restricted mode */
2595                 err = -EIO;
2596                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2597                         goto out;
2598
2599                 in_iovs = outarg.in_iovs;
2600                 out_iovs = outarg.out_iovs;
2601
2602                 /*
2603                  * Make sure things are in boundary, separate checks
2604                  * are to protect against overflow.
2605                  */
2606                 err = -ENOMEM;
2607                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2608                     out_iovs > FUSE_IOCTL_MAX_IOV ||
2609                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2610                         goto out;
2611
2612                 vaddr = kmap_atomic(pages[0]);
2613                 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2614                                             transferred, in_iovs + out_iovs,
2615                                             (flags & FUSE_IOCTL_COMPAT) != 0);
2616                 kunmap_atomic(vaddr);
2617                 if (err)
2618                         goto out;
2619
2620                 in_iov = iov_page;
2621                 out_iov = in_iov + in_iovs;
2622
2623                 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
2624                 if (err)
2625                         goto out;
2626
2627                 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
2628                 if (err)
2629                         goto out;
2630
2631                 goto retry;
2632         }
2633
2634         err = -EIO;
2635         if (transferred > inarg.out_size)
2636                 goto out;
2637
2638         err = -EFAULT;
2639         iov_iter_init(&ii, READ, out_iov, out_iovs, transferred);
2640         for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= num_pages); i++) {
2641                 c = copy_page_to_iter(pages[i], 0, PAGE_SIZE, &ii);
2642                 if (c != PAGE_SIZE && iov_iter_count(&ii))
2643                         goto out;
2644         }
2645         err = 0;
2646  out:
2647         if (req)
2648                 fuse_put_request(fc, req);
2649         free_page((unsigned long) iov_page);
2650         while (num_pages)
2651                 __free_page(pages[--num_pages]);
2652         kfree(pages);
2653
2654         return err ? err : outarg.result;
2655 }
2656 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2657
2658 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2659                        unsigned long arg, unsigned int flags)
2660 {
2661         struct inode *inode = file_inode(file);
2662         struct fuse_conn *fc = get_fuse_conn(inode);
2663
2664         if (!fuse_allow_current_process(fc))
2665                 return -EACCES;
2666
2667         if (is_bad_inode(inode))
2668                 return -EIO;
2669
2670         return fuse_do_ioctl(file, cmd, arg, flags);
2671 }
2672
2673 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2674                             unsigned long arg)
2675 {
2676         return fuse_ioctl_common(file, cmd, arg, 0);
2677 }
2678
2679 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2680                                    unsigned long arg)
2681 {
2682         return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2683 }
2684
2685 /*
2686  * All files which have been polled are linked to RB tree
2687  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2688  * find the matching one.
2689  */
2690 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2691                                               struct rb_node **parent_out)
2692 {
2693         struct rb_node **link = &fc->polled_files.rb_node;
2694         struct rb_node *last = NULL;
2695
2696         while (*link) {
2697                 struct fuse_file *ff;
2698
2699                 last = *link;
2700                 ff = rb_entry(last, struct fuse_file, polled_node);
2701
2702                 if (kh < ff->kh)
2703                         link = &last->rb_left;
2704                 else if (kh > ff->kh)
2705                         link = &last->rb_right;
2706                 else
2707                         return link;
2708         }
2709
2710         if (parent_out)
2711                 *parent_out = last;
2712         return link;
2713 }
2714
2715 /*
2716  * The file is about to be polled.  Make sure it's on the polled_files
2717  * RB tree.  Note that files once added to the polled_files tree are
2718  * not removed before the file is released.  This is because a file
2719  * polled once is likely to be polled again.
2720  */
2721 static void fuse_register_polled_file(struct fuse_conn *fc,
2722                                       struct fuse_file *ff)
2723 {
2724         spin_lock(&fc->lock);
2725         if (RB_EMPTY_NODE(&ff->polled_node)) {
2726                 struct rb_node **link, *uninitialized_var(parent);
2727
2728                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2729                 BUG_ON(*link);
2730                 rb_link_node(&ff->polled_node, parent, link);
2731                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2732         }
2733         spin_unlock(&fc->lock);
2734 }
2735
2736 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2737 {
2738         struct fuse_file *ff = file->private_data;
2739         struct fuse_conn *fc = ff->fc;
2740         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2741         struct fuse_poll_out outarg;
2742         FUSE_ARGS(args);
2743         int err;
2744
2745         if (fc->no_poll)
2746                 return DEFAULT_POLLMASK;
2747
2748         poll_wait(file, &ff->poll_wait, wait);
2749         inarg.events = (__u32)poll_requested_events(wait);
2750
2751         /*
2752          * Ask for notification iff there's someone waiting for it.
2753          * The client may ignore the flag and always notify.
2754          */
2755         if (waitqueue_active(&ff->poll_wait)) {
2756                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2757                 fuse_register_polled_file(fc, ff);
2758         }
2759
2760         args.in.h.opcode = FUSE_POLL;
2761         args.in.h.nodeid = ff->nodeid;
2762         args.in.numargs = 1;
2763         args.in.args[0].size = sizeof(inarg);
2764         args.in.args[0].value = &inarg;
2765         args.out.numargs = 1;
2766         args.out.args[0].size = sizeof(outarg);
2767         args.out.args[0].value = &outarg;
2768         err = fuse_simple_request(fc, &args);
2769
2770         if (!err)
2771                 return outarg.revents;
2772         if (err == -ENOSYS) {
2773                 fc->no_poll = 1;
2774                 return DEFAULT_POLLMASK;
2775         }
2776         return POLLERR;
2777 }
2778 EXPORT_SYMBOL_GPL(fuse_file_poll);
2779
2780 /*
2781  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2782  * wakes up the poll waiters.
2783  */
2784 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2785                             struct fuse_notify_poll_wakeup_out *outarg)
2786 {
2787         u64 kh = outarg->kh;
2788         struct rb_node **link;
2789
2790         spin_lock(&fc->lock);
2791
2792         link = fuse_find_polled_node(fc, kh, NULL);
2793         if (*link) {
2794                 struct fuse_file *ff;
2795
2796                 ff = rb_entry(*link, struct fuse_file, polled_node);
2797                 wake_up_interruptible_sync(&ff->poll_wait);
2798         }
2799
2800         spin_unlock(&fc->lock);
2801         return 0;
2802 }
2803
2804 static void fuse_do_truncate(struct file *file)
2805 {
2806         struct inode *inode = file->f_mapping->host;
2807         struct iattr attr;
2808
2809         attr.ia_valid = ATTR_SIZE;
2810         attr.ia_size = i_size_read(inode);
2811
2812         attr.ia_file = file;
2813         attr.ia_valid |= ATTR_FILE;
2814
2815         fuse_do_setattr(file_dentry(file), &attr, file);
2816 }
2817
2818 static inline loff_t fuse_round_up(loff_t off)
2819 {
2820         return round_up(off, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
2821 }
2822
2823 static ssize_t
2824 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
2825 {
2826         DECLARE_COMPLETION_ONSTACK(wait);
2827         ssize_t ret = 0;
2828         struct file *file = iocb->ki_filp;
2829         struct fuse_file *ff = file->private_data;
2830         bool async_dio = ff->fc->async_dio;
2831         loff_t pos = 0;
2832         struct inode *inode;
2833         loff_t i_size;
2834         size_t count = iov_iter_count(iter);
2835         loff_t offset = iocb->ki_pos;
2836         struct fuse_io_priv *io;
2837
2838         pos = offset;
2839         inode = file->f_mapping->host;
2840         i_size = i_size_read(inode);
2841
2842         if ((iov_iter_rw(iter) == READ) && (offset > i_size))
2843                 return 0;
2844
2845         /* optimization for short read */
2846         if (async_dio && iov_iter_rw(iter) != WRITE && offset + count > i_size) {
2847                 if (offset >= i_size)
2848                         return 0;
2849                 iov_iter_truncate(iter, fuse_round_up(i_size - offset));
2850                 count = iov_iter_count(iter);
2851         }
2852
2853         io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2854         if (!io)
2855                 return -ENOMEM;
2856         spin_lock_init(&io->lock);
2857         kref_init(&io->refcnt);
2858         io->reqs = 1;
2859         io->bytes = -1;
2860         io->size = 0;
2861         io->offset = offset;
2862         io->write = (iov_iter_rw(iter) == WRITE);
2863         io->err = 0;
2864         io->file = file;
2865         /*
2866          * By default, we want to optimize all I/Os with async request
2867          * submission to the client filesystem if supported.
2868          */
2869         io->async = async_dio;
2870         io->iocb = iocb;
2871         io->blocking = is_sync_kiocb(iocb);
2872
2873         /*
2874          * We cannot asynchronously extend the size of a file.
2875          * In such case the aio will behave exactly like sync io.
2876          */
2877         if ((offset + count > i_size) && iov_iter_rw(iter) == WRITE)
2878                 io->blocking = true;
2879
2880         if (io->async && io->blocking) {
2881                 /*
2882                  * Additional reference to keep io around after
2883                  * calling fuse_aio_complete()
2884                  */
2885                 kref_get(&io->refcnt);
2886                 io->done = &wait;
2887         }
2888
2889         if (iov_iter_rw(iter) == WRITE) {
2890                 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
2891                 fuse_invalidate_attr(inode);
2892         } else {
2893                 ret = __fuse_direct_read(io, iter, &pos);
2894         }
2895
2896         if (io->async) {
2897                 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2898
2899                 /* we have a non-extending, async request, so return */
2900                 if (!io->blocking)
2901                         return -EIOCBQUEUED;
2902
2903                 wait_for_completion(&wait);
2904                 ret = fuse_get_res_by_io(io);
2905         }
2906
2907         kref_put(&io->refcnt, fuse_io_release);
2908
2909         if (iov_iter_rw(iter) == WRITE) {
2910                 if (ret > 0)
2911                         fuse_write_update_size(inode, pos);
2912                 else if (ret < 0 && offset + count > i_size)
2913                         fuse_do_truncate(file);
2914         }
2915
2916         return ret;
2917 }
2918
2919 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2920                                 loff_t length)
2921 {
2922         struct fuse_file *ff = file->private_data;
2923         struct inode *inode = file_inode(file);
2924         struct fuse_inode *fi = get_fuse_inode(inode);
2925         struct fuse_conn *fc = ff->fc;
2926         FUSE_ARGS(args);
2927         struct fuse_fallocate_in inarg = {
2928                 .fh = ff->fh,
2929                 .offset = offset,
2930                 .length = length,
2931                 .mode = mode
2932         };
2933         int err;
2934         bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
2935                            (mode & FALLOC_FL_PUNCH_HOLE);
2936
2937         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2938                 return -EOPNOTSUPP;
2939
2940         if (fc->no_fallocate)
2941                 return -EOPNOTSUPP;
2942
2943         if (lock_inode) {
2944                 inode_lock(inode);
2945                 if (mode & FALLOC_FL_PUNCH_HOLE) {
2946                         loff_t endbyte = offset + length - 1;
2947                         err = filemap_write_and_wait_range(inode->i_mapping,
2948                                                            offset, endbyte);
2949                         if (err)
2950                                 goto out;
2951
2952                         fuse_sync_writes(inode);
2953                 }
2954         }
2955
2956         if (!(mode & FALLOC_FL_KEEP_SIZE))
2957                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2958
2959         args.in.h.opcode = FUSE_FALLOCATE;
2960         args.in.h.nodeid = ff->nodeid;
2961         args.in.numargs = 1;
2962         args.in.args[0].size = sizeof(inarg);
2963         args.in.args[0].value = &inarg;
2964         err = fuse_simple_request(fc, &args);
2965         if (err == -ENOSYS) {
2966                 fc->no_fallocate = 1;
2967                 err = -EOPNOTSUPP;
2968         }
2969         if (err)
2970                 goto out;
2971
2972         /* we could have extended the file */
2973         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
2974                 bool changed = fuse_write_update_size(inode, offset + length);
2975
2976                 if (changed && fc->writeback_cache)
2977                         file_update_time(file);
2978         }
2979
2980         if (mode & FALLOC_FL_PUNCH_HOLE)
2981                 truncate_pagecache_range(inode, offset, offset + length - 1);
2982
2983         fuse_invalidate_attr(inode);
2984
2985 out:
2986         if (!(mode & FALLOC_FL_KEEP_SIZE))
2987                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2988
2989         if (lock_inode)
2990                 inode_unlock(inode);
2991
2992         return err;
2993 }
2994
2995 static const struct file_operations fuse_file_operations = {
2996         .llseek         = fuse_file_llseek,
2997         .read_iter      = fuse_file_read_iter,
2998         .write_iter     = fuse_file_write_iter,
2999         .mmap           = fuse_file_mmap,
3000         .open           = fuse_open,
3001         .flush          = fuse_flush,
3002         .release        = fuse_release,
3003         .fsync          = fuse_fsync,
3004         .lock           = fuse_file_lock,
3005         .flock          = fuse_file_flock,
3006         .splice_read    = generic_file_splice_read,
3007         .unlocked_ioctl = fuse_file_ioctl,
3008         .compat_ioctl   = fuse_file_compat_ioctl,
3009         .poll           = fuse_file_poll,
3010         .fallocate      = fuse_file_fallocate,
3011 };
3012
3013 static const struct file_operations fuse_direct_io_file_operations = {
3014         .llseek         = fuse_file_llseek,
3015         .read_iter      = fuse_direct_read_iter,
3016         .write_iter     = fuse_direct_write_iter,
3017         .mmap           = fuse_direct_mmap,
3018         .open           = fuse_open,
3019         .flush          = fuse_flush,
3020         .release        = fuse_release,
3021         .fsync          = fuse_fsync,
3022         .lock           = fuse_file_lock,
3023         .flock          = fuse_file_flock,
3024         .unlocked_ioctl = fuse_file_ioctl,
3025         .compat_ioctl   = fuse_file_compat_ioctl,
3026         .poll           = fuse_file_poll,
3027         .fallocate      = fuse_file_fallocate,
3028         /* no splice_read */
3029 };
3030
3031 static const struct address_space_operations fuse_file_aops  = {
3032         .readpage       = fuse_readpage,
3033         .writepage      = fuse_writepage,
3034         .writepages     = fuse_writepages,
3035         .launder_page   = fuse_launder_page,
3036         .readpages      = fuse_readpages,
3037         .set_page_dirty = __set_page_dirty_nobuffers,
3038         .bmap           = fuse_bmap,
3039         .direct_IO      = fuse_direct_IO,
3040         .write_begin    = fuse_write_begin,
3041         .write_end      = fuse_write_end,
3042 };
3043
3044 void fuse_init_file_inode(struct inode *inode)
3045 {
3046         inode->i_fop = &fuse_file_operations;
3047         inode->i_data.a_ops = &fuse_file_aops;
3048 }