Merge branch 'overlayfs-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mszer...
[sfrench/cifs-2.6.git] / fs / fuse / file.c
1 /*
2   FUSE: Filesystem in Userspace
3   Copyright (C) 2001-2008  Miklos Szeredi <miklos@szeredi.hu>
4
5   This program can be distributed under the terms of the GNU GPL.
6   See the file COPYING.
7 */
8
9 #include "fuse_i.h"
10
11 #include <linux/pagemap.h>
12 #include <linux/slab.h>
13 #include <linux/kernel.h>
14 #include <linux/sched.h>
15 #include <linux/module.h>
16 #include <linux/compat.h>
17 #include <linux/swap.h>
18 #include <linux/falloc.h>
19 #include <linux/uio.h>
20
21 static const struct file_operations fuse_direct_io_file_operations;
22
23 static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
24                           int opcode, struct fuse_open_out *outargp)
25 {
26         struct fuse_open_in inarg;
27         FUSE_ARGS(args);
28
29         memset(&inarg, 0, sizeof(inarg));
30         inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
31         if (!fc->atomic_o_trunc)
32                 inarg.flags &= ~O_TRUNC;
33         args.in.h.opcode = opcode;
34         args.in.h.nodeid = nodeid;
35         args.in.numargs = 1;
36         args.in.args[0].size = sizeof(inarg);
37         args.in.args[0].value = &inarg;
38         args.out.numargs = 1;
39         args.out.args[0].size = sizeof(*outargp);
40         args.out.args[0].value = outargp;
41
42         return fuse_simple_request(fc, &args);
43 }
44
45 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
46 {
47         struct fuse_file *ff;
48
49         ff = kzalloc(sizeof(struct fuse_file), GFP_KERNEL);
50         if (unlikely(!ff))
51                 return NULL;
52
53         ff->fc = fc;
54         ff->reserved_req = fuse_request_alloc(0);
55         if (unlikely(!ff->reserved_req)) {
56                 kfree(ff);
57                 return NULL;
58         }
59
60         INIT_LIST_HEAD(&ff->write_entry);
61         refcount_set(&ff->count, 1);
62         RB_CLEAR_NODE(&ff->polled_node);
63         init_waitqueue_head(&ff->poll_wait);
64
65         spin_lock(&fc->lock);
66         ff->kh = ++fc->khctr;
67         spin_unlock(&fc->lock);
68
69         return ff;
70 }
71
72 void fuse_file_free(struct fuse_file *ff)
73 {
74         fuse_request_free(ff->reserved_req);
75         kfree(ff);
76 }
77
78 static struct fuse_file *fuse_file_get(struct fuse_file *ff)
79 {
80         refcount_inc(&ff->count);
81         return ff;
82 }
83
84 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
85 {
86         iput(req->misc.release.inode);
87 }
88
89 static void fuse_file_put(struct fuse_file *ff, bool sync)
90 {
91         if (refcount_dec_and_test(&ff->count)) {
92                 struct fuse_req *req = ff->reserved_req;
93
94                 if (ff->fc->no_open) {
95                         /*
96                          * Drop the release request when client does not
97                          * implement 'open'
98                          */
99                         __clear_bit(FR_BACKGROUND, &req->flags);
100                         iput(req->misc.release.inode);
101                         fuse_put_request(ff->fc, req);
102                 } else if (sync) {
103                         __set_bit(FR_FORCE, &req->flags);
104                         __clear_bit(FR_BACKGROUND, &req->flags);
105                         fuse_request_send(ff->fc, req);
106                         iput(req->misc.release.inode);
107                         fuse_put_request(ff->fc, req);
108                 } else {
109                         req->end = fuse_release_end;
110                         __set_bit(FR_BACKGROUND, &req->flags);
111                         fuse_request_send_background(ff->fc, req);
112                 }
113                 kfree(ff);
114         }
115 }
116
117 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
118                  bool isdir)
119 {
120         struct fuse_file *ff;
121         int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
122
123         ff = fuse_file_alloc(fc);
124         if (!ff)
125                 return -ENOMEM;
126
127         ff->fh = 0;
128         ff->open_flags = FOPEN_KEEP_CACHE; /* Default for no-open */
129         if (!fc->no_open || isdir) {
130                 struct fuse_open_out outarg;
131                 int err;
132
133                 err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
134                 if (!err) {
135                         ff->fh = outarg.fh;
136                         ff->open_flags = outarg.open_flags;
137
138                 } else if (err != -ENOSYS || isdir) {
139                         fuse_file_free(ff);
140                         return err;
141                 } else {
142                         fc->no_open = 1;
143                 }
144         }
145
146         if (isdir)
147                 ff->open_flags &= ~FOPEN_DIRECT_IO;
148
149         ff->nodeid = nodeid;
150         file->private_data = ff;
151
152         return 0;
153 }
154 EXPORT_SYMBOL_GPL(fuse_do_open);
155
156 static void fuse_link_write_file(struct file *file)
157 {
158         struct inode *inode = file_inode(file);
159         struct fuse_conn *fc = get_fuse_conn(inode);
160         struct fuse_inode *fi = get_fuse_inode(inode);
161         struct fuse_file *ff = file->private_data;
162         /*
163          * file may be written through mmap, so chain it onto the
164          * inodes's write_file list
165          */
166         spin_lock(&fc->lock);
167         if (list_empty(&ff->write_entry))
168                 list_add(&ff->write_entry, &fi->write_files);
169         spin_unlock(&fc->lock);
170 }
171
172 void fuse_finish_open(struct inode *inode, struct file *file)
173 {
174         struct fuse_file *ff = file->private_data;
175         struct fuse_conn *fc = get_fuse_conn(inode);
176
177         if (ff->open_flags & FOPEN_DIRECT_IO)
178                 file->f_op = &fuse_direct_io_file_operations;
179         if (!(ff->open_flags & FOPEN_KEEP_CACHE))
180                 invalidate_inode_pages2(inode->i_mapping);
181         if (ff->open_flags & FOPEN_NONSEEKABLE)
182                 nonseekable_open(inode, file);
183         if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
184                 struct fuse_inode *fi = get_fuse_inode(inode);
185
186                 spin_lock(&fc->lock);
187                 fi->attr_version = ++fc->attr_version;
188                 i_size_write(inode, 0);
189                 spin_unlock(&fc->lock);
190                 fuse_invalidate_attr(inode);
191                 if (fc->writeback_cache)
192                         file_update_time(file);
193         }
194         if ((file->f_mode & FMODE_WRITE) && fc->writeback_cache)
195                 fuse_link_write_file(file);
196 }
197
198 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
199 {
200         struct fuse_conn *fc = get_fuse_conn(inode);
201         int err;
202         bool lock_inode = (file->f_flags & O_TRUNC) &&
203                           fc->atomic_o_trunc &&
204                           fc->writeback_cache;
205
206         err = generic_file_open(inode, file);
207         if (err)
208                 return err;
209
210         if (lock_inode)
211                 inode_lock(inode);
212
213         err = fuse_do_open(fc, get_node_id(inode), file, isdir);
214
215         if (!err)
216                 fuse_finish_open(inode, file);
217
218         if (lock_inode)
219                 inode_unlock(inode);
220
221         return err;
222 }
223
224 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
225 {
226         struct fuse_conn *fc = ff->fc;
227         struct fuse_req *req = ff->reserved_req;
228         struct fuse_release_in *inarg = &req->misc.release.in;
229
230         spin_lock(&fc->lock);
231         list_del(&ff->write_entry);
232         if (!RB_EMPTY_NODE(&ff->polled_node))
233                 rb_erase(&ff->polled_node, &fc->polled_files);
234         spin_unlock(&fc->lock);
235
236         wake_up_interruptible_all(&ff->poll_wait);
237
238         inarg->fh = ff->fh;
239         inarg->flags = flags;
240         req->in.h.opcode = opcode;
241         req->in.h.nodeid = ff->nodeid;
242         req->in.numargs = 1;
243         req->in.args[0].size = sizeof(struct fuse_release_in);
244         req->in.args[0].value = inarg;
245 }
246
247 void fuse_release_common(struct file *file, int opcode)
248 {
249         struct fuse_file *ff = file->private_data;
250         struct fuse_req *req = ff->reserved_req;
251
252         fuse_prepare_release(ff, file->f_flags, opcode);
253
254         if (ff->flock) {
255                 struct fuse_release_in *inarg = &req->misc.release.in;
256                 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
257                 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
258                                                        (fl_owner_t) file);
259         }
260         /* Hold inode until release is finished */
261         req->misc.release.inode = igrab(file_inode(file));
262
263         /*
264          * Normally this will send the RELEASE request, however if
265          * some asynchronous READ or WRITE requests are outstanding,
266          * the sending will be delayed.
267          *
268          * Make the release synchronous if this is a fuseblk mount,
269          * synchronous RELEASE is allowed (and desirable) in this case
270          * because the server can be trusted not to screw up.
271          */
272         fuse_file_put(ff, ff->fc->destroy_req != NULL);
273 }
274
275 static int fuse_open(struct inode *inode, struct file *file)
276 {
277         return fuse_open_common(inode, file, false);
278 }
279
280 static int fuse_release(struct inode *inode, struct file *file)
281 {
282         struct fuse_conn *fc = get_fuse_conn(inode);
283
284         /* see fuse_vma_close() for !writeback_cache case */
285         if (fc->writeback_cache)
286                 write_inode_now(inode, 1);
287
288         fuse_release_common(file, FUSE_RELEASE);
289
290         /* return value is ignored by VFS */
291         return 0;
292 }
293
294 void fuse_sync_release(struct fuse_file *ff, int flags)
295 {
296         WARN_ON(refcount_read(&ff->count) > 1);
297         fuse_prepare_release(ff, flags, FUSE_RELEASE);
298         /*
299          * iput(NULL) is a no-op and since the refcount is 1 and everything's
300          * synchronous, we are fine with not doing igrab() here"
301          */
302         fuse_file_put(ff, true);
303 }
304 EXPORT_SYMBOL_GPL(fuse_sync_release);
305
306 /*
307  * Scramble the ID space with XTEA, so that the value of the files_struct
308  * pointer is not exposed to userspace.
309  */
310 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
311 {
312         u32 *k = fc->scramble_key;
313         u64 v = (unsigned long) id;
314         u32 v0 = v;
315         u32 v1 = v >> 32;
316         u32 sum = 0;
317         int i;
318
319         for (i = 0; i < 32; i++) {
320                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
321                 sum += 0x9E3779B9;
322                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
323         }
324
325         return (u64) v0 + ((u64) v1 << 32);
326 }
327
328 /*
329  * Check if any page in a range is under writeback
330  *
331  * This is currently done by walking the list of writepage requests
332  * for the inode, which can be pretty inefficient.
333  */
334 static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
335                                    pgoff_t idx_to)
336 {
337         struct fuse_conn *fc = get_fuse_conn(inode);
338         struct fuse_inode *fi = get_fuse_inode(inode);
339         struct fuse_req *req;
340         bool found = false;
341
342         spin_lock(&fc->lock);
343         list_for_each_entry(req, &fi->writepages, writepages_entry) {
344                 pgoff_t curr_index;
345
346                 BUG_ON(req->inode != inode);
347                 curr_index = req->misc.write.in.offset >> PAGE_SHIFT;
348                 if (idx_from < curr_index + req->num_pages &&
349                     curr_index <= idx_to) {
350                         found = true;
351                         break;
352                 }
353         }
354         spin_unlock(&fc->lock);
355
356         return found;
357 }
358
359 static inline bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
360 {
361         return fuse_range_is_writeback(inode, index, index);
362 }
363
364 /*
365  * Wait for page writeback to be completed.
366  *
367  * Since fuse doesn't rely on the VM writeback tracking, this has to
368  * use some other means.
369  */
370 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
371 {
372         struct fuse_inode *fi = get_fuse_inode(inode);
373
374         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
375         return 0;
376 }
377
378 /*
379  * Wait for all pending writepages on the inode to finish.
380  *
381  * This is currently done by blocking further writes with FUSE_NOWRITE
382  * and waiting for all sent writes to complete.
383  *
384  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
385  * could conflict with truncation.
386  */
387 static void fuse_sync_writes(struct inode *inode)
388 {
389         fuse_set_nowrite(inode);
390         fuse_release_nowrite(inode);
391 }
392
393 static int fuse_flush(struct file *file, fl_owner_t id)
394 {
395         struct inode *inode = file_inode(file);
396         struct fuse_conn *fc = get_fuse_conn(inode);
397         struct fuse_file *ff = file->private_data;
398         struct fuse_req *req;
399         struct fuse_flush_in inarg;
400         int err;
401
402         if (is_bad_inode(inode))
403                 return -EIO;
404
405         if (fc->no_flush)
406                 return 0;
407
408         err = write_inode_now(inode, 1);
409         if (err)
410                 return err;
411
412         inode_lock(inode);
413         fuse_sync_writes(inode);
414         inode_unlock(inode);
415
416         err = filemap_check_errors(file->f_mapping);
417         if (err)
418                 return err;
419
420         req = fuse_get_req_nofail_nopages(fc, file);
421         memset(&inarg, 0, sizeof(inarg));
422         inarg.fh = ff->fh;
423         inarg.lock_owner = fuse_lock_owner_id(fc, id);
424         req->in.h.opcode = FUSE_FLUSH;
425         req->in.h.nodeid = get_node_id(inode);
426         req->in.numargs = 1;
427         req->in.args[0].size = sizeof(inarg);
428         req->in.args[0].value = &inarg;
429         __set_bit(FR_FORCE, &req->flags);
430         fuse_request_send(fc, req);
431         err = req->out.h.error;
432         fuse_put_request(fc, req);
433         if (err == -ENOSYS) {
434                 fc->no_flush = 1;
435                 err = 0;
436         }
437         return err;
438 }
439
440 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
441                       int datasync, int isdir)
442 {
443         struct inode *inode = file->f_mapping->host;
444         struct fuse_conn *fc = get_fuse_conn(inode);
445         struct fuse_file *ff = file->private_data;
446         FUSE_ARGS(args);
447         struct fuse_fsync_in inarg;
448         int err;
449
450         if (is_bad_inode(inode))
451                 return -EIO;
452
453         inode_lock(inode);
454
455         /*
456          * Start writeback against all dirty pages of the inode, then
457          * wait for all outstanding writes, before sending the FSYNC
458          * request.
459          */
460         err = file_write_and_wait_range(file, start, end);
461         if (err)
462                 goto out;
463
464         fuse_sync_writes(inode);
465
466         /*
467          * Due to implementation of fuse writeback
468          * file_write_and_wait_range() does not catch errors.
469          * We have to do this directly after fuse_sync_writes()
470          */
471         err = file_check_and_advance_wb_err(file);
472         if (err)
473                 goto out;
474
475         err = sync_inode_metadata(inode, 1);
476         if (err)
477                 goto out;
478
479         if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
480                 goto out;
481
482         memset(&inarg, 0, sizeof(inarg));
483         inarg.fh = ff->fh;
484         inarg.fsync_flags = datasync ? 1 : 0;
485         args.in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
486         args.in.h.nodeid = get_node_id(inode);
487         args.in.numargs = 1;
488         args.in.args[0].size = sizeof(inarg);
489         args.in.args[0].value = &inarg;
490         err = fuse_simple_request(fc, &args);
491         if (err == -ENOSYS) {
492                 if (isdir)
493                         fc->no_fsyncdir = 1;
494                 else
495                         fc->no_fsync = 1;
496                 err = 0;
497         }
498 out:
499         inode_unlock(inode);
500         return err;
501 }
502
503 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
504                       int datasync)
505 {
506         return fuse_fsync_common(file, start, end, datasync, 0);
507 }
508
509 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
510                     size_t count, int opcode)
511 {
512         struct fuse_read_in *inarg = &req->misc.read.in;
513         struct fuse_file *ff = file->private_data;
514
515         inarg->fh = ff->fh;
516         inarg->offset = pos;
517         inarg->size = count;
518         inarg->flags = file->f_flags;
519         req->in.h.opcode = opcode;
520         req->in.h.nodeid = ff->nodeid;
521         req->in.numargs = 1;
522         req->in.args[0].size = sizeof(struct fuse_read_in);
523         req->in.args[0].value = inarg;
524         req->out.argvar = 1;
525         req->out.numargs = 1;
526         req->out.args[0].size = count;
527 }
528
529 static void fuse_release_user_pages(struct fuse_req *req, bool should_dirty)
530 {
531         unsigned i;
532
533         for (i = 0; i < req->num_pages; i++) {
534                 struct page *page = req->pages[i];
535                 if (should_dirty)
536                         set_page_dirty_lock(page);
537                 put_page(page);
538         }
539 }
540
541 static void fuse_io_release(struct kref *kref)
542 {
543         kfree(container_of(kref, struct fuse_io_priv, refcnt));
544 }
545
546 static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
547 {
548         if (io->err)
549                 return io->err;
550
551         if (io->bytes >= 0 && io->write)
552                 return -EIO;
553
554         return io->bytes < 0 ? io->size : io->bytes;
555 }
556
557 /**
558  * In case of short read, the caller sets 'pos' to the position of
559  * actual end of fuse request in IO request. Otherwise, if bytes_requested
560  * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
561  *
562  * An example:
563  * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
564  * both submitted asynchronously. The first of them was ACKed by userspace as
565  * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
566  * second request was ACKed as short, e.g. only 1K was read, resulting in
567  * pos == 33K.
568  *
569  * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
570  * will be equal to the length of the longest contiguous fragment of
571  * transferred data starting from the beginning of IO request.
572  */
573 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
574 {
575         int left;
576
577         spin_lock(&io->lock);
578         if (err)
579                 io->err = io->err ? : err;
580         else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
581                 io->bytes = pos;
582
583         left = --io->reqs;
584         if (!left && io->blocking)
585                 complete(io->done);
586         spin_unlock(&io->lock);
587
588         if (!left && !io->blocking) {
589                 ssize_t res = fuse_get_res_by_io(io);
590
591                 if (res >= 0) {
592                         struct inode *inode = file_inode(io->iocb->ki_filp);
593                         struct fuse_conn *fc = get_fuse_conn(inode);
594                         struct fuse_inode *fi = get_fuse_inode(inode);
595
596                         spin_lock(&fc->lock);
597                         fi->attr_version = ++fc->attr_version;
598                         spin_unlock(&fc->lock);
599                 }
600
601                 io->iocb->ki_complete(io->iocb, res, 0);
602         }
603
604         kref_put(&io->refcnt, fuse_io_release);
605 }
606
607 static void fuse_aio_complete_req(struct fuse_conn *fc, struct fuse_req *req)
608 {
609         struct fuse_io_priv *io = req->io;
610         ssize_t pos = -1;
611
612         fuse_release_user_pages(req, io->should_dirty);
613
614         if (io->write) {
615                 if (req->misc.write.in.size != req->misc.write.out.size)
616                         pos = req->misc.write.in.offset - io->offset +
617                                 req->misc.write.out.size;
618         } else {
619                 if (req->misc.read.in.size != req->out.args[0].size)
620                         pos = req->misc.read.in.offset - io->offset +
621                                 req->out.args[0].size;
622         }
623
624         fuse_aio_complete(io, req->out.h.error, pos);
625 }
626
627 static size_t fuse_async_req_send(struct fuse_conn *fc, struct fuse_req *req,
628                 size_t num_bytes, struct fuse_io_priv *io)
629 {
630         spin_lock(&io->lock);
631         kref_get(&io->refcnt);
632         io->size += num_bytes;
633         io->reqs++;
634         spin_unlock(&io->lock);
635
636         req->io = io;
637         req->end = fuse_aio_complete_req;
638
639         __fuse_get_request(req);
640         fuse_request_send_background(fc, req);
641
642         return num_bytes;
643 }
644
645 static size_t fuse_send_read(struct fuse_req *req, struct fuse_io_priv *io,
646                              loff_t pos, size_t count, fl_owner_t owner)
647 {
648         struct file *file = io->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         int cuse = flags & FUSE_DIO_CUSE;
1320         struct file *file = io->file;
1321         struct inode *inode = file->f_mapping->host;
1322         struct fuse_file *ff = file->private_data;
1323         struct fuse_conn *fc = ff->fc;
1324         size_t nmax = write ? fc->max_write : fc->max_read;
1325         loff_t pos = *ppos;
1326         size_t count = iov_iter_count(iter);
1327         pgoff_t idx_from = pos >> PAGE_SHIFT;
1328         pgoff_t idx_to = (pos + count - 1) >> PAGE_SHIFT;
1329         ssize_t res = 0;
1330         struct fuse_req *req;
1331         int err = 0;
1332
1333         if (io->async)
1334                 req = fuse_get_req_for_background(fc, fuse_iter_npages(iter));
1335         else
1336                 req = fuse_get_req(fc, fuse_iter_npages(iter));
1337         if (IS_ERR(req))
1338                 return PTR_ERR(req);
1339
1340         if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1341                 if (!write)
1342                         inode_lock(inode);
1343                 fuse_sync_writes(inode);
1344                 if (!write)
1345                         inode_unlock(inode);
1346         }
1347
1348         io->should_dirty = !write && iter_is_iovec(iter);
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, io->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         mapping_set_error(page->mapping, error);
1673         end_page_writeback(page);
1674         return error;
1675 }
1676
1677 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1678 {
1679         int err;
1680
1681         if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1682                 /*
1683                  * ->writepages() should be called for sync() and friends.  We
1684                  * should only get here on direct reclaim and then we are
1685                  * allowed to skip a page which is already in flight
1686                  */
1687                 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1688
1689                 redirty_page_for_writepage(wbc, page);
1690                 return 0;
1691         }
1692
1693         err = fuse_writepage_locked(page);
1694         unlock_page(page);
1695
1696         return err;
1697 }
1698
1699 struct fuse_fill_wb_data {
1700         struct fuse_req *req;
1701         struct fuse_file *ff;
1702         struct inode *inode;
1703         struct page **orig_pages;
1704 };
1705
1706 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1707 {
1708         struct fuse_req *req = data->req;
1709         struct inode *inode = data->inode;
1710         struct fuse_conn *fc = get_fuse_conn(inode);
1711         struct fuse_inode *fi = get_fuse_inode(inode);
1712         int num_pages = req->num_pages;
1713         int i;
1714
1715         req->ff = fuse_file_get(data->ff);
1716         spin_lock(&fc->lock);
1717         list_add_tail(&req->list, &fi->queued_writes);
1718         fuse_flush_writepages(inode);
1719         spin_unlock(&fc->lock);
1720
1721         for (i = 0; i < num_pages; i++)
1722                 end_page_writeback(data->orig_pages[i]);
1723 }
1724
1725 static bool fuse_writepage_in_flight(struct fuse_req *new_req,
1726                                      struct page *page)
1727 {
1728         struct fuse_conn *fc = get_fuse_conn(new_req->inode);
1729         struct fuse_inode *fi = get_fuse_inode(new_req->inode);
1730         struct fuse_req *tmp;
1731         struct fuse_req *old_req;
1732         bool found = false;
1733         pgoff_t curr_index;
1734
1735         BUG_ON(new_req->num_pages != 0);
1736
1737         spin_lock(&fc->lock);
1738         list_del(&new_req->writepages_entry);
1739         list_for_each_entry(old_req, &fi->writepages, writepages_entry) {
1740                 BUG_ON(old_req->inode != new_req->inode);
1741                 curr_index = old_req->misc.write.in.offset >> PAGE_SHIFT;
1742                 if (curr_index <= page->index &&
1743                     page->index < curr_index + old_req->num_pages) {
1744                         found = true;
1745                         break;
1746                 }
1747         }
1748         if (!found) {
1749                 list_add(&new_req->writepages_entry, &fi->writepages);
1750                 goto out_unlock;
1751         }
1752
1753         new_req->num_pages = 1;
1754         for (tmp = old_req; tmp != NULL; tmp = tmp->misc.write.next) {
1755                 BUG_ON(tmp->inode != new_req->inode);
1756                 curr_index = tmp->misc.write.in.offset >> PAGE_SHIFT;
1757                 if (tmp->num_pages == 1 &&
1758                     curr_index == page->index) {
1759                         old_req = tmp;
1760                 }
1761         }
1762
1763         if (old_req->num_pages == 1 && test_bit(FR_PENDING, &old_req->flags)) {
1764                 struct backing_dev_info *bdi = inode_to_bdi(page->mapping->host);
1765
1766                 copy_highpage(old_req->pages[0], page);
1767                 spin_unlock(&fc->lock);
1768
1769                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1770                 dec_node_page_state(page, NR_WRITEBACK_TEMP);
1771                 wb_writeout_inc(&bdi->wb);
1772                 fuse_writepage_free(fc, new_req);
1773                 fuse_request_free(new_req);
1774                 goto out;
1775         } else {
1776                 new_req->misc.write.next = old_req->misc.write.next;
1777                 old_req->misc.write.next = new_req;
1778         }
1779 out_unlock:
1780         spin_unlock(&fc->lock);
1781 out:
1782         return found;
1783 }
1784
1785 static int fuse_writepages_fill(struct page *page,
1786                 struct writeback_control *wbc, void *_data)
1787 {
1788         struct fuse_fill_wb_data *data = _data;
1789         struct fuse_req *req = data->req;
1790         struct inode *inode = data->inode;
1791         struct fuse_conn *fc = get_fuse_conn(inode);
1792         struct page *tmp_page;
1793         bool is_writeback;
1794         int err;
1795
1796         if (!data->ff) {
1797                 err = -EIO;
1798                 data->ff = fuse_write_file_get(fc, get_fuse_inode(inode));
1799                 if (!data->ff)
1800                         goto out_unlock;
1801         }
1802
1803         /*
1804          * Being under writeback is unlikely but possible.  For example direct
1805          * read to an mmaped fuse file will set the page dirty twice; once when
1806          * the pages are faulted with get_user_pages(), and then after the read
1807          * completed.
1808          */
1809         is_writeback = fuse_page_is_writeback(inode, page->index);
1810
1811         if (req && req->num_pages &&
1812             (is_writeback || req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
1813              (req->num_pages + 1) * PAGE_SIZE > fc->max_write ||
1814              data->orig_pages[req->num_pages - 1]->index + 1 != page->index)) {
1815                 fuse_writepages_send(data);
1816                 data->req = NULL;
1817         }
1818         err = -ENOMEM;
1819         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1820         if (!tmp_page)
1821                 goto out_unlock;
1822
1823         /*
1824          * The page must not be redirtied until the writeout is completed
1825          * (i.e. userspace has sent a reply to the write request).  Otherwise
1826          * there could be more than one temporary page instance for each real
1827          * page.
1828          *
1829          * This is ensured by holding the page lock in page_mkwrite() while
1830          * checking fuse_page_is_writeback().  We already hold the page lock
1831          * since clear_page_dirty_for_io() and keep it held until we add the
1832          * request to the fi->writepages list and increment req->num_pages.
1833          * After this fuse_page_is_writeback() will indicate that the page is
1834          * under writeback, so we can release the page lock.
1835          */
1836         if (data->req == NULL) {
1837                 struct fuse_inode *fi = get_fuse_inode(inode);
1838
1839                 err = -ENOMEM;
1840                 req = fuse_request_alloc_nofs(FUSE_MAX_PAGES_PER_REQ);
1841                 if (!req) {
1842                         __free_page(tmp_page);
1843                         goto out_unlock;
1844                 }
1845
1846                 fuse_write_fill(req, data->ff, page_offset(page), 0);
1847                 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1848                 req->misc.write.next = NULL;
1849                 req->in.argpages = 1;
1850                 __set_bit(FR_BACKGROUND, &req->flags);
1851                 req->num_pages = 0;
1852                 req->end = fuse_writepage_end;
1853                 req->inode = inode;
1854
1855                 spin_lock(&fc->lock);
1856                 list_add(&req->writepages_entry, &fi->writepages);
1857                 spin_unlock(&fc->lock);
1858
1859                 data->req = req;
1860         }
1861         set_page_writeback(page);
1862
1863         copy_highpage(tmp_page, page);
1864         req->pages[req->num_pages] = tmp_page;
1865         req->page_descs[req->num_pages].offset = 0;
1866         req->page_descs[req->num_pages].length = PAGE_SIZE;
1867
1868         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1869         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1870
1871         err = 0;
1872         if (is_writeback && fuse_writepage_in_flight(req, page)) {
1873                 end_page_writeback(page);
1874                 data->req = NULL;
1875                 goto out_unlock;
1876         }
1877         data->orig_pages[req->num_pages] = page;
1878
1879         /*
1880          * Protected by fc->lock against concurrent access by
1881          * fuse_page_is_writeback().
1882          */
1883         spin_lock(&fc->lock);
1884         req->num_pages++;
1885         spin_unlock(&fc->lock);
1886
1887 out_unlock:
1888         unlock_page(page);
1889
1890         return err;
1891 }
1892
1893 static int fuse_writepages(struct address_space *mapping,
1894                            struct writeback_control *wbc)
1895 {
1896         struct inode *inode = mapping->host;
1897         struct fuse_fill_wb_data data;
1898         int err;
1899
1900         err = -EIO;
1901         if (is_bad_inode(inode))
1902                 goto out;
1903
1904         data.inode = inode;
1905         data.req = NULL;
1906         data.ff = NULL;
1907
1908         err = -ENOMEM;
1909         data.orig_pages = kcalloc(FUSE_MAX_PAGES_PER_REQ,
1910                                   sizeof(struct page *),
1911                                   GFP_NOFS);
1912         if (!data.orig_pages)
1913                 goto out;
1914
1915         err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
1916         if (data.req) {
1917                 /* Ignore errors if we can write at least one page */
1918                 BUG_ON(!data.req->num_pages);
1919                 fuse_writepages_send(&data);
1920                 err = 0;
1921         }
1922         if (data.ff)
1923                 fuse_file_put(data.ff, false);
1924
1925         kfree(data.orig_pages);
1926 out:
1927         return err;
1928 }
1929
1930 /*
1931  * It's worthy to make sure that space is reserved on disk for the write,
1932  * but how to implement it without killing performance need more thinking.
1933  */
1934 static int fuse_write_begin(struct file *file, struct address_space *mapping,
1935                 loff_t pos, unsigned len, unsigned flags,
1936                 struct page **pagep, void **fsdata)
1937 {
1938         pgoff_t index = pos >> PAGE_SHIFT;
1939         struct fuse_conn *fc = get_fuse_conn(file_inode(file));
1940         struct page *page;
1941         loff_t fsize;
1942         int err = -ENOMEM;
1943
1944         WARN_ON(!fc->writeback_cache);
1945
1946         page = grab_cache_page_write_begin(mapping, index, flags);
1947         if (!page)
1948                 goto error;
1949
1950         fuse_wait_on_page_writeback(mapping->host, page->index);
1951
1952         if (PageUptodate(page) || len == PAGE_SIZE)
1953                 goto success;
1954         /*
1955          * Check if the start this page comes after the end of file, in which
1956          * case the readpage can be optimized away.
1957          */
1958         fsize = i_size_read(mapping->host);
1959         if (fsize <= (pos & PAGE_MASK)) {
1960                 size_t off = pos & ~PAGE_MASK;
1961                 if (off)
1962                         zero_user_segment(page, 0, off);
1963                 goto success;
1964         }
1965         err = fuse_do_readpage(file, page);
1966         if (err)
1967                 goto cleanup;
1968 success:
1969         *pagep = page;
1970         return 0;
1971
1972 cleanup:
1973         unlock_page(page);
1974         put_page(page);
1975 error:
1976         return err;
1977 }
1978
1979 static int fuse_write_end(struct file *file, struct address_space *mapping,
1980                 loff_t pos, unsigned len, unsigned copied,
1981                 struct page *page, void *fsdata)
1982 {
1983         struct inode *inode = page->mapping->host;
1984
1985         /* Haven't copied anything?  Skip zeroing, size extending, dirtying. */
1986         if (!copied)
1987                 goto unlock;
1988
1989         if (!PageUptodate(page)) {
1990                 /* Zero any unwritten bytes at the end of the page */
1991                 size_t endoff = (pos + copied) & ~PAGE_MASK;
1992                 if (endoff)
1993                         zero_user_segment(page, endoff, PAGE_SIZE);
1994                 SetPageUptodate(page);
1995         }
1996
1997         fuse_write_update_size(inode, pos + copied);
1998         set_page_dirty(page);
1999
2000 unlock:
2001         unlock_page(page);
2002         put_page(page);
2003
2004         return copied;
2005 }
2006
2007 static int fuse_launder_page(struct page *page)
2008 {
2009         int err = 0;
2010         if (clear_page_dirty_for_io(page)) {
2011                 struct inode *inode = page->mapping->host;
2012                 err = fuse_writepage_locked(page);
2013                 if (!err)
2014                         fuse_wait_on_page_writeback(inode, page->index);
2015         }
2016         return err;
2017 }
2018
2019 /*
2020  * Write back dirty pages now, because there may not be any suitable
2021  * open files later
2022  */
2023 static void fuse_vma_close(struct vm_area_struct *vma)
2024 {
2025         filemap_write_and_wait(vma->vm_file->f_mapping);
2026 }
2027
2028 /*
2029  * Wait for writeback against this page to complete before allowing it
2030  * to be marked dirty again, and hence written back again, possibly
2031  * before the previous writepage completed.
2032  *
2033  * Block here, instead of in ->writepage(), so that the userspace fs
2034  * can only block processes actually operating on the filesystem.
2035  *
2036  * Otherwise unprivileged userspace fs would be able to block
2037  * unrelated:
2038  *
2039  * - page migration
2040  * - sync(2)
2041  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2042  */
2043 static int fuse_page_mkwrite(struct vm_fault *vmf)
2044 {
2045         struct page *page = vmf->page;
2046         struct inode *inode = file_inode(vmf->vma->vm_file);
2047
2048         file_update_time(vmf->vma->vm_file);
2049         lock_page(page);
2050         if (page->mapping != inode->i_mapping) {
2051                 unlock_page(page);
2052                 return VM_FAULT_NOPAGE;
2053         }
2054
2055         fuse_wait_on_page_writeback(inode, page->index);
2056         return VM_FAULT_LOCKED;
2057 }
2058
2059 static const struct vm_operations_struct fuse_file_vm_ops = {
2060         .close          = fuse_vma_close,
2061         .fault          = filemap_fault,
2062         .map_pages      = filemap_map_pages,
2063         .page_mkwrite   = fuse_page_mkwrite,
2064 };
2065
2066 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2067 {
2068         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2069                 fuse_link_write_file(file);
2070
2071         file_accessed(file);
2072         vma->vm_ops = &fuse_file_vm_ops;
2073         return 0;
2074 }
2075
2076 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
2077 {
2078         /* Can't provide the coherency needed for MAP_SHARED */
2079         if (vma->vm_flags & VM_MAYSHARE)
2080                 return -ENODEV;
2081
2082         invalidate_inode_pages2(file->f_mapping);
2083
2084         return generic_file_mmap(file, vma);
2085 }
2086
2087 static int convert_fuse_file_lock(struct fuse_conn *fc,
2088                                   const struct fuse_file_lock *ffl,
2089                                   struct file_lock *fl)
2090 {
2091         switch (ffl->type) {
2092         case F_UNLCK:
2093                 break;
2094
2095         case F_RDLCK:
2096         case F_WRLCK:
2097                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2098                     ffl->end < ffl->start)
2099                         return -EIO;
2100
2101                 fl->fl_start = ffl->start;
2102                 fl->fl_end = ffl->end;
2103
2104                 /*
2105                  * Convert pid into init's pid namespace.  The locks API will
2106                  * translate it into the caller's pid namespace.
2107                  */
2108                 rcu_read_lock();
2109                 fl->fl_pid = pid_nr_ns(find_pid_ns(ffl->pid, fc->pid_ns), &init_pid_ns);
2110                 rcu_read_unlock();
2111                 break;
2112
2113         default:
2114                 return -EIO;
2115         }
2116         fl->fl_type = ffl->type;
2117         return 0;
2118 }
2119
2120 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2121                          const struct file_lock *fl, int opcode, pid_t pid,
2122                          int flock, struct fuse_lk_in *inarg)
2123 {
2124         struct inode *inode = file_inode(file);
2125         struct fuse_conn *fc = get_fuse_conn(inode);
2126         struct fuse_file *ff = file->private_data;
2127
2128         memset(inarg, 0, sizeof(*inarg));
2129         inarg->fh = ff->fh;
2130         inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2131         inarg->lk.start = fl->fl_start;
2132         inarg->lk.end = fl->fl_end;
2133         inarg->lk.type = fl->fl_type;
2134         inarg->lk.pid = pid;
2135         if (flock)
2136                 inarg->lk_flags |= FUSE_LK_FLOCK;
2137         args->in.h.opcode = opcode;
2138         args->in.h.nodeid = get_node_id(inode);
2139         args->in.numargs = 1;
2140         args->in.args[0].size = sizeof(*inarg);
2141         args->in.args[0].value = inarg;
2142 }
2143
2144 static int fuse_getlk(struct file *file, struct file_lock *fl)
2145 {
2146         struct inode *inode = file_inode(file);
2147         struct fuse_conn *fc = get_fuse_conn(inode);
2148         FUSE_ARGS(args);
2149         struct fuse_lk_in inarg;
2150         struct fuse_lk_out outarg;
2151         int err;
2152
2153         fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2154         args.out.numargs = 1;
2155         args.out.args[0].size = sizeof(outarg);
2156         args.out.args[0].value = &outarg;
2157         err = fuse_simple_request(fc, &args);
2158         if (!err)
2159                 err = convert_fuse_file_lock(fc, &outarg.lk, fl);
2160
2161         return err;
2162 }
2163
2164 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2165 {
2166         struct inode *inode = file_inode(file);
2167         struct fuse_conn *fc = get_fuse_conn(inode);
2168         FUSE_ARGS(args);
2169         struct fuse_lk_in inarg;
2170         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2171         struct pid *pid = fl->fl_type != F_UNLCK ? task_tgid(current) : NULL;
2172         pid_t pid_nr = pid_nr_ns(pid, fc->pid_ns);
2173         int err;
2174
2175         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2176                 /* NLM needs asynchronous locks, which we don't support yet */
2177                 return -ENOLCK;
2178         }
2179
2180         /* Unlock on close is handled by the flush method */
2181         if ((fl->fl_flags & FL_CLOSE_POSIX) == FL_CLOSE_POSIX)
2182                 return 0;
2183
2184         if (pid && pid_nr == 0)
2185                 return -EOVERFLOW;
2186
2187         fuse_lk_fill(&args, file, fl, opcode, pid_nr, flock, &inarg);
2188         err = fuse_simple_request(fc, &args);
2189
2190         /* locking is restartable */
2191         if (err == -EINTR)
2192                 err = -ERESTARTSYS;
2193
2194         return err;
2195 }
2196
2197 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2198 {
2199         struct inode *inode = file_inode(file);
2200         struct fuse_conn *fc = get_fuse_conn(inode);
2201         int err;
2202
2203         if (cmd == F_CANCELLK) {
2204                 err = 0;
2205         } else if (cmd == F_GETLK) {
2206                 if (fc->no_lock) {
2207                         posix_test_lock(file, fl);
2208                         err = 0;
2209                 } else
2210                         err = fuse_getlk(file, fl);
2211         } else {
2212                 if (fc->no_lock)
2213                         err = posix_lock_file(file, fl, NULL);
2214                 else
2215                         err = fuse_setlk(file, fl, 0);
2216         }
2217         return err;
2218 }
2219
2220 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2221 {
2222         struct inode *inode = file_inode(file);
2223         struct fuse_conn *fc = get_fuse_conn(inode);
2224         int err;
2225
2226         if (fc->no_flock) {
2227                 err = locks_lock_file_wait(file, fl);
2228         } else {
2229                 struct fuse_file *ff = file->private_data;
2230
2231                 /* emulate flock with POSIX locks */
2232                 ff->flock = true;
2233                 err = fuse_setlk(file, fl, 1);
2234         }
2235
2236         return err;
2237 }
2238
2239 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2240 {
2241         struct inode *inode = mapping->host;
2242         struct fuse_conn *fc = get_fuse_conn(inode);
2243         FUSE_ARGS(args);
2244         struct fuse_bmap_in inarg;
2245         struct fuse_bmap_out outarg;
2246         int err;
2247
2248         if (!inode->i_sb->s_bdev || fc->no_bmap)
2249                 return 0;
2250
2251         memset(&inarg, 0, sizeof(inarg));
2252         inarg.block = block;
2253         inarg.blocksize = inode->i_sb->s_blocksize;
2254         args.in.h.opcode = FUSE_BMAP;
2255         args.in.h.nodeid = get_node_id(inode);
2256         args.in.numargs = 1;
2257         args.in.args[0].size = sizeof(inarg);
2258         args.in.args[0].value = &inarg;
2259         args.out.numargs = 1;
2260         args.out.args[0].size = sizeof(outarg);
2261         args.out.args[0].value = &outarg;
2262         err = fuse_simple_request(fc, &args);
2263         if (err == -ENOSYS)
2264                 fc->no_bmap = 1;
2265
2266         return err ? 0 : outarg.block;
2267 }
2268
2269 static loff_t fuse_lseek(struct file *file, loff_t offset, int whence)
2270 {
2271         struct inode *inode = file->f_mapping->host;
2272         struct fuse_conn *fc = get_fuse_conn(inode);
2273         struct fuse_file *ff = file->private_data;
2274         FUSE_ARGS(args);
2275         struct fuse_lseek_in inarg = {
2276                 .fh = ff->fh,
2277                 .offset = offset,
2278                 .whence = whence
2279         };
2280         struct fuse_lseek_out outarg;
2281         int err;
2282
2283         if (fc->no_lseek)
2284                 goto fallback;
2285
2286         args.in.h.opcode = FUSE_LSEEK;
2287         args.in.h.nodeid = ff->nodeid;
2288         args.in.numargs = 1;
2289         args.in.args[0].size = sizeof(inarg);
2290         args.in.args[0].value = &inarg;
2291         args.out.numargs = 1;
2292         args.out.args[0].size = sizeof(outarg);
2293         args.out.args[0].value = &outarg;
2294         err = fuse_simple_request(fc, &args);
2295         if (err) {
2296                 if (err == -ENOSYS) {
2297                         fc->no_lseek = 1;
2298                         goto fallback;
2299                 }
2300                 return err;
2301         }
2302
2303         return vfs_setpos(file, outarg.offset, inode->i_sb->s_maxbytes);
2304
2305 fallback:
2306         err = fuse_update_attributes(inode, NULL, file, NULL);
2307         if (!err)
2308                 return generic_file_llseek(file, offset, whence);
2309         else
2310                 return err;
2311 }
2312
2313 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2314 {
2315         loff_t retval;
2316         struct inode *inode = file_inode(file);
2317
2318         switch (whence) {
2319         case SEEK_SET:
2320         case SEEK_CUR:
2321                  /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2322                 retval = generic_file_llseek(file, offset, whence);
2323                 break;
2324         case SEEK_END:
2325                 inode_lock(inode);
2326                 retval = fuse_update_attributes(inode, NULL, file, NULL);
2327                 if (!retval)
2328                         retval = generic_file_llseek(file, offset, whence);
2329                 inode_unlock(inode);
2330                 break;
2331         case SEEK_HOLE:
2332         case SEEK_DATA:
2333                 inode_lock(inode);
2334                 retval = fuse_lseek(file, offset, whence);
2335                 inode_unlock(inode);
2336                 break;
2337         default:
2338                 retval = -EINVAL;
2339         }
2340
2341         return retval;
2342 }
2343
2344 /*
2345  * CUSE servers compiled on 32bit broke on 64bit kernels because the
2346  * ABI was defined to be 'struct iovec' which is different on 32bit
2347  * and 64bit.  Fortunately we can determine which structure the server
2348  * used from the size of the reply.
2349  */
2350 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2351                                      size_t transferred, unsigned count,
2352                                      bool is_compat)
2353 {
2354 #ifdef CONFIG_COMPAT
2355         if (count * sizeof(struct compat_iovec) == transferred) {
2356                 struct compat_iovec *ciov = src;
2357                 unsigned i;
2358
2359                 /*
2360                  * With this interface a 32bit server cannot support
2361                  * non-compat (i.e. ones coming from 64bit apps) ioctl
2362                  * requests
2363                  */
2364                 if (!is_compat)
2365                         return -EINVAL;
2366
2367                 for (i = 0; i < count; i++) {
2368                         dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2369                         dst[i].iov_len = ciov[i].iov_len;
2370                 }
2371                 return 0;
2372         }
2373 #endif
2374
2375         if (count * sizeof(struct iovec) != transferred)
2376                 return -EIO;
2377
2378         memcpy(dst, src, transferred);
2379         return 0;
2380 }
2381
2382 /* Make sure iov_length() won't overflow */
2383 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
2384 {
2385         size_t n;
2386         u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
2387
2388         for (n = 0; n < count; n++, iov++) {
2389                 if (iov->iov_len > (size_t) max)
2390                         return -ENOMEM;
2391                 max -= iov->iov_len;
2392         }
2393         return 0;
2394 }
2395
2396 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2397                                  void *src, size_t transferred, unsigned count,
2398                                  bool is_compat)
2399 {
2400         unsigned i;
2401         struct fuse_ioctl_iovec *fiov = src;
2402
2403         if (fc->minor < 16) {
2404                 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2405                                                  count, is_compat);
2406         }
2407
2408         if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2409                 return -EIO;
2410
2411         for (i = 0; i < count; i++) {
2412                 /* Did the server supply an inappropriate value? */
2413                 if (fiov[i].base != (unsigned long) fiov[i].base ||
2414                     fiov[i].len != (unsigned long) fiov[i].len)
2415                         return -EIO;
2416
2417                 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2418                 dst[i].iov_len = (size_t) fiov[i].len;
2419
2420 #ifdef CONFIG_COMPAT
2421                 if (is_compat &&
2422                     (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2423                      (compat_size_t) dst[i].iov_len != fiov[i].len))
2424                         return -EIO;
2425 #endif
2426         }
2427
2428         return 0;
2429 }
2430
2431
2432 /*
2433  * For ioctls, there is no generic way to determine how much memory
2434  * needs to be read and/or written.  Furthermore, ioctls are allowed
2435  * to dereference the passed pointer, so the parameter requires deep
2436  * copying but FUSE has no idea whatsoever about what to copy in or
2437  * out.
2438  *
2439  * This is solved by allowing FUSE server to retry ioctl with
2440  * necessary in/out iovecs.  Let's assume the ioctl implementation
2441  * needs to read in the following structure.
2442  *
2443  * struct a {
2444  *      char    *buf;
2445  *      size_t  buflen;
2446  * }
2447  *
2448  * On the first callout to FUSE server, inarg->in_size and
2449  * inarg->out_size will be NULL; then, the server completes the ioctl
2450  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2451  * the actual iov array to
2452  *
2453  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
2454  *
2455  * which tells FUSE to copy in the requested area and retry the ioctl.
2456  * On the second round, the server has access to the structure and
2457  * from that it can tell what to look for next, so on the invocation,
2458  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2459  *
2460  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
2461  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
2462  *
2463  * FUSE will copy both struct a and the pointed buffer from the
2464  * process doing the ioctl and retry ioctl with both struct a and the
2465  * buffer.
2466  *
2467  * This time, FUSE server has everything it needs and completes ioctl
2468  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2469  *
2470  * Copying data out works the same way.
2471  *
2472  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2473  * automatically initializes in and out iovs by decoding @cmd with
2474  * _IOC_* macros and the server is not allowed to request RETRY.  This
2475  * limits ioctl data transfers to well-formed ioctls and is the forced
2476  * behavior for all FUSE servers.
2477  */
2478 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2479                    unsigned int flags)
2480 {
2481         struct fuse_file *ff = file->private_data;
2482         struct fuse_conn *fc = ff->fc;
2483         struct fuse_ioctl_in inarg = {
2484                 .fh = ff->fh,
2485                 .cmd = cmd,
2486                 .arg = arg,
2487                 .flags = flags
2488         };
2489         struct fuse_ioctl_out outarg;
2490         struct fuse_req *req = NULL;
2491         struct page **pages = NULL;
2492         struct iovec *iov_page = NULL;
2493         struct iovec *in_iov = NULL, *out_iov = NULL;
2494         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
2495         size_t in_size, out_size, transferred, c;
2496         int err, i;
2497         struct iov_iter ii;
2498
2499 #if BITS_PER_LONG == 32
2500         inarg.flags |= FUSE_IOCTL_32BIT;
2501 #else
2502         if (flags & FUSE_IOCTL_COMPAT)
2503                 inarg.flags |= FUSE_IOCTL_32BIT;
2504 #endif
2505
2506         /* assume all the iovs returned by client always fits in a page */
2507         BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2508
2509         err = -ENOMEM;
2510         pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
2511         iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2512         if (!pages || !iov_page)
2513                 goto out;
2514
2515         /*
2516          * If restricted, initialize IO parameters as encoded in @cmd.
2517          * RETRY from server is not allowed.
2518          */
2519         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2520                 struct iovec *iov = iov_page;
2521
2522                 iov->iov_base = (void __user *)arg;
2523                 iov->iov_len = _IOC_SIZE(cmd);
2524
2525                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2526                         in_iov = iov;
2527                         in_iovs = 1;
2528                 }
2529
2530                 if (_IOC_DIR(cmd) & _IOC_READ) {
2531                         out_iov = iov;
2532                         out_iovs = 1;
2533                 }
2534         }
2535
2536  retry:
2537         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2538         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2539
2540         /*
2541          * Out data can be used either for actual out data or iovs,
2542          * make sure there always is at least one page.
2543          */
2544         out_size = max_t(size_t, out_size, PAGE_SIZE);
2545         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2546
2547         /* make sure there are enough buffer pages and init request with them */
2548         err = -ENOMEM;
2549         if (max_pages > FUSE_MAX_PAGES_PER_REQ)
2550                 goto out;
2551         while (num_pages < max_pages) {
2552                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2553                 if (!pages[num_pages])
2554                         goto out;
2555                 num_pages++;
2556         }
2557
2558         req = fuse_get_req(fc, num_pages);
2559         if (IS_ERR(req)) {
2560                 err = PTR_ERR(req);
2561                 req = NULL;
2562                 goto out;
2563         }
2564         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
2565         req->num_pages = num_pages;
2566         fuse_page_descs_length_init(req, 0, req->num_pages);
2567
2568         /* okay, let's send it to the client */
2569         req->in.h.opcode = FUSE_IOCTL;
2570         req->in.h.nodeid = ff->nodeid;
2571         req->in.numargs = 1;
2572         req->in.args[0].size = sizeof(inarg);
2573         req->in.args[0].value = &inarg;
2574         if (in_size) {
2575                 req->in.numargs++;
2576                 req->in.args[1].size = in_size;
2577                 req->in.argpages = 1;
2578
2579                 err = -EFAULT;
2580                 iov_iter_init(&ii, WRITE, in_iov, in_iovs, in_size);
2581                 for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= num_pages); i++) {
2582                         c = copy_page_from_iter(pages[i], 0, PAGE_SIZE, &ii);
2583                         if (c != PAGE_SIZE && iov_iter_count(&ii))
2584                                 goto out;
2585                 }
2586         }
2587
2588         req->out.numargs = 2;
2589         req->out.args[0].size = sizeof(outarg);
2590         req->out.args[0].value = &outarg;
2591         req->out.args[1].size = out_size;
2592         req->out.argpages = 1;
2593         req->out.argvar = 1;
2594
2595         fuse_request_send(fc, req);
2596         err = req->out.h.error;
2597         transferred = req->out.args[1].size;
2598         fuse_put_request(fc, req);
2599         req = NULL;
2600         if (err)
2601                 goto out;
2602
2603         /* did it ask for retry? */
2604         if (outarg.flags & FUSE_IOCTL_RETRY) {
2605                 void *vaddr;
2606
2607                 /* no retry if in restricted mode */
2608                 err = -EIO;
2609                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2610                         goto out;
2611
2612                 in_iovs = outarg.in_iovs;
2613                 out_iovs = outarg.out_iovs;
2614
2615                 /*
2616                  * Make sure things are in boundary, separate checks
2617                  * are to protect against overflow.
2618                  */
2619                 err = -ENOMEM;
2620                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2621                     out_iovs > FUSE_IOCTL_MAX_IOV ||
2622                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2623                         goto out;
2624
2625                 vaddr = kmap_atomic(pages[0]);
2626                 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2627                                             transferred, in_iovs + out_iovs,
2628                                             (flags & FUSE_IOCTL_COMPAT) != 0);
2629                 kunmap_atomic(vaddr);
2630                 if (err)
2631                         goto out;
2632
2633                 in_iov = iov_page;
2634                 out_iov = in_iov + in_iovs;
2635
2636                 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
2637                 if (err)
2638                         goto out;
2639
2640                 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
2641                 if (err)
2642                         goto out;
2643
2644                 goto retry;
2645         }
2646
2647         err = -EIO;
2648         if (transferred > inarg.out_size)
2649                 goto out;
2650
2651         err = -EFAULT;
2652         iov_iter_init(&ii, READ, out_iov, out_iovs, transferred);
2653         for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= num_pages); i++) {
2654                 c = copy_page_to_iter(pages[i], 0, PAGE_SIZE, &ii);
2655                 if (c != PAGE_SIZE && iov_iter_count(&ii))
2656                         goto out;
2657         }
2658         err = 0;
2659  out:
2660         if (req)
2661                 fuse_put_request(fc, req);
2662         free_page((unsigned long) iov_page);
2663         while (num_pages)
2664                 __free_page(pages[--num_pages]);
2665         kfree(pages);
2666
2667         return err ? err : outarg.result;
2668 }
2669 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2670
2671 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2672                        unsigned long arg, unsigned int flags)
2673 {
2674         struct inode *inode = file_inode(file);
2675         struct fuse_conn *fc = get_fuse_conn(inode);
2676
2677         if (!fuse_allow_current_process(fc))
2678                 return -EACCES;
2679
2680         if (is_bad_inode(inode))
2681                 return -EIO;
2682
2683         return fuse_do_ioctl(file, cmd, arg, flags);
2684 }
2685
2686 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2687                             unsigned long arg)
2688 {
2689         return fuse_ioctl_common(file, cmd, arg, 0);
2690 }
2691
2692 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2693                                    unsigned long arg)
2694 {
2695         return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2696 }
2697
2698 /*
2699  * All files which have been polled are linked to RB tree
2700  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2701  * find the matching one.
2702  */
2703 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2704                                               struct rb_node **parent_out)
2705 {
2706         struct rb_node **link = &fc->polled_files.rb_node;
2707         struct rb_node *last = NULL;
2708
2709         while (*link) {
2710                 struct fuse_file *ff;
2711
2712                 last = *link;
2713                 ff = rb_entry(last, struct fuse_file, polled_node);
2714
2715                 if (kh < ff->kh)
2716                         link = &last->rb_left;
2717                 else if (kh > ff->kh)
2718                         link = &last->rb_right;
2719                 else
2720                         return link;
2721         }
2722
2723         if (parent_out)
2724                 *parent_out = last;
2725         return link;
2726 }
2727
2728 /*
2729  * The file is about to be polled.  Make sure it's on the polled_files
2730  * RB tree.  Note that files once added to the polled_files tree are
2731  * not removed before the file is released.  This is because a file
2732  * polled once is likely to be polled again.
2733  */
2734 static void fuse_register_polled_file(struct fuse_conn *fc,
2735                                       struct fuse_file *ff)
2736 {
2737         spin_lock(&fc->lock);
2738         if (RB_EMPTY_NODE(&ff->polled_node)) {
2739                 struct rb_node **link, *uninitialized_var(parent);
2740
2741                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2742                 BUG_ON(*link);
2743                 rb_link_node(&ff->polled_node, parent, link);
2744                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2745         }
2746         spin_unlock(&fc->lock);
2747 }
2748
2749 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2750 {
2751         struct fuse_file *ff = file->private_data;
2752         struct fuse_conn *fc = ff->fc;
2753         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2754         struct fuse_poll_out outarg;
2755         FUSE_ARGS(args);
2756         int err;
2757
2758         if (fc->no_poll)
2759                 return DEFAULT_POLLMASK;
2760
2761         poll_wait(file, &ff->poll_wait, wait);
2762         inarg.events = (__u32)poll_requested_events(wait);
2763
2764         /*
2765          * Ask for notification iff there's someone waiting for it.
2766          * The client may ignore the flag and always notify.
2767          */
2768         if (waitqueue_active(&ff->poll_wait)) {
2769                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2770                 fuse_register_polled_file(fc, ff);
2771         }
2772
2773         args.in.h.opcode = FUSE_POLL;
2774         args.in.h.nodeid = ff->nodeid;
2775         args.in.numargs = 1;
2776         args.in.args[0].size = sizeof(inarg);
2777         args.in.args[0].value = &inarg;
2778         args.out.numargs = 1;
2779         args.out.args[0].size = sizeof(outarg);
2780         args.out.args[0].value = &outarg;
2781         err = fuse_simple_request(fc, &args);
2782
2783         if (!err)
2784                 return outarg.revents;
2785         if (err == -ENOSYS) {
2786                 fc->no_poll = 1;
2787                 return DEFAULT_POLLMASK;
2788         }
2789         return POLLERR;
2790 }
2791 EXPORT_SYMBOL_GPL(fuse_file_poll);
2792
2793 /*
2794  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2795  * wakes up the poll waiters.
2796  */
2797 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2798                             struct fuse_notify_poll_wakeup_out *outarg)
2799 {
2800         u64 kh = outarg->kh;
2801         struct rb_node **link;
2802
2803         spin_lock(&fc->lock);
2804
2805         link = fuse_find_polled_node(fc, kh, NULL);
2806         if (*link) {
2807                 struct fuse_file *ff;
2808
2809                 ff = rb_entry(*link, struct fuse_file, polled_node);
2810                 wake_up_interruptible_sync(&ff->poll_wait);
2811         }
2812
2813         spin_unlock(&fc->lock);
2814         return 0;
2815 }
2816
2817 static void fuse_do_truncate(struct file *file)
2818 {
2819         struct inode *inode = file->f_mapping->host;
2820         struct iattr attr;
2821
2822         attr.ia_valid = ATTR_SIZE;
2823         attr.ia_size = i_size_read(inode);
2824
2825         attr.ia_file = file;
2826         attr.ia_valid |= ATTR_FILE;
2827
2828         fuse_do_setattr(file_dentry(file), &attr, file);
2829 }
2830
2831 static inline loff_t fuse_round_up(loff_t off)
2832 {
2833         return round_up(off, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
2834 }
2835
2836 static ssize_t
2837 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
2838 {
2839         DECLARE_COMPLETION_ONSTACK(wait);
2840         ssize_t ret = 0;
2841         struct file *file = iocb->ki_filp;
2842         struct fuse_file *ff = file->private_data;
2843         bool async_dio = ff->fc->async_dio;
2844         loff_t pos = 0;
2845         struct inode *inode;
2846         loff_t i_size;
2847         size_t count = iov_iter_count(iter);
2848         loff_t offset = iocb->ki_pos;
2849         struct fuse_io_priv *io;
2850
2851         pos = offset;
2852         inode = file->f_mapping->host;
2853         i_size = i_size_read(inode);
2854
2855         if ((iov_iter_rw(iter) == READ) && (offset > i_size))
2856                 return 0;
2857
2858         /* optimization for short read */
2859         if (async_dio && iov_iter_rw(iter) != WRITE && offset + count > i_size) {
2860                 if (offset >= i_size)
2861                         return 0;
2862                 iov_iter_truncate(iter, fuse_round_up(i_size - offset));
2863                 count = iov_iter_count(iter);
2864         }
2865
2866         io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2867         if (!io)
2868                 return -ENOMEM;
2869         spin_lock_init(&io->lock);
2870         kref_init(&io->refcnt);
2871         io->reqs = 1;
2872         io->bytes = -1;
2873         io->size = 0;
2874         io->offset = offset;
2875         io->write = (iov_iter_rw(iter) == WRITE);
2876         io->err = 0;
2877         io->file = file;
2878         /*
2879          * By default, we want to optimize all I/Os with async request
2880          * submission to the client filesystem if supported.
2881          */
2882         io->async = async_dio;
2883         io->iocb = iocb;
2884         io->blocking = is_sync_kiocb(iocb);
2885
2886         /*
2887          * We cannot asynchronously extend the size of a file.
2888          * In such case the aio will behave exactly like sync io.
2889          */
2890         if ((offset + count > i_size) && iov_iter_rw(iter) == WRITE)
2891                 io->blocking = true;
2892
2893         if (io->async && io->blocking) {
2894                 /*
2895                  * Additional reference to keep io around after
2896                  * calling fuse_aio_complete()
2897                  */
2898                 kref_get(&io->refcnt);
2899                 io->done = &wait;
2900         }
2901
2902         if (iov_iter_rw(iter) == WRITE) {
2903                 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
2904                 fuse_invalidate_attr(inode);
2905         } else {
2906                 ret = __fuse_direct_read(io, iter, &pos);
2907         }
2908
2909         if (io->async) {
2910                 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2911
2912                 /* we have a non-extending, async request, so return */
2913                 if (!io->blocking)
2914                         return -EIOCBQUEUED;
2915
2916                 wait_for_completion(&wait);
2917                 ret = fuse_get_res_by_io(io);
2918         }
2919
2920         kref_put(&io->refcnt, fuse_io_release);
2921
2922         if (iov_iter_rw(iter) == WRITE) {
2923                 if (ret > 0)
2924                         fuse_write_update_size(inode, pos);
2925                 else if (ret < 0 && offset + count > i_size)
2926                         fuse_do_truncate(file);
2927         }
2928
2929         return ret;
2930 }
2931
2932 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2933                                 loff_t length)
2934 {
2935         struct fuse_file *ff = file->private_data;
2936         struct inode *inode = file_inode(file);
2937         struct fuse_inode *fi = get_fuse_inode(inode);
2938         struct fuse_conn *fc = ff->fc;
2939         FUSE_ARGS(args);
2940         struct fuse_fallocate_in inarg = {
2941                 .fh = ff->fh,
2942                 .offset = offset,
2943                 .length = length,
2944                 .mode = mode
2945         };
2946         int err;
2947         bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
2948                            (mode & FALLOC_FL_PUNCH_HOLE);
2949
2950         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2951                 return -EOPNOTSUPP;
2952
2953         if (fc->no_fallocate)
2954                 return -EOPNOTSUPP;
2955
2956         if (lock_inode) {
2957                 inode_lock(inode);
2958                 if (mode & FALLOC_FL_PUNCH_HOLE) {
2959                         loff_t endbyte = offset + length - 1;
2960                         err = filemap_write_and_wait_range(inode->i_mapping,
2961                                                            offset, endbyte);
2962                         if (err)
2963                                 goto out;
2964
2965                         fuse_sync_writes(inode);
2966                 }
2967         }
2968
2969         if (!(mode & FALLOC_FL_KEEP_SIZE))
2970                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2971
2972         args.in.h.opcode = FUSE_FALLOCATE;
2973         args.in.h.nodeid = ff->nodeid;
2974         args.in.numargs = 1;
2975         args.in.args[0].size = sizeof(inarg);
2976         args.in.args[0].value = &inarg;
2977         err = fuse_simple_request(fc, &args);
2978         if (err == -ENOSYS) {
2979                 fc->no_fallocate = 1;
2980                 err = -EOPNOTSUPP;
2981         }
2982         if (err)
2983                 goto out;
2984
2985         /* we could have extended the file */
2986         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
2987                 bool changed = fuse_write_update_size(inode, offset + length);
2988
2989                 if (changed && fc->writeback_cache)
2990                         file_update_time(file);
2991         }
2992
2993         if (mode & FALLOC_FL_PUNCH_HOLE)
2994                 truncate_pagecache_range(inode, offset, offset + length - 1);
2995
2996         fuse_invalidate_attr(inode);
2997
2998 out:
2999         if (!(mode & FALLOC_FL_KEEP_SIZE))
3000                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3001
3002         if (lock_inode)
3003                 inode_unlock(inode);
3004
3005         return err;
3006 }
3007
3008 static const struct file_operations fuse_file_operations = {
3009         .llseek         = fuse_file_llseek,
3010         .read_iter      = fuse_file_read_iter,
3011         .write_iter     = fuse_file_write_iter,
3012         .mmap           = fuse_file_mmap,
3013         .open           = fuse_open,
3014         .flush          = fuse_flush,
3015         .release        = fuse_release,
3016         .fsync          = fuse_fsync,
3017         .lock           = fuse_file_lock,
3018         .flock          = fuse_file_flock,
3019         .splice_read    = generic_file_splice_read,
3020         .unlocked_ioctl = fuse_file_ioctl,
3021         .compat_ioctl   = fuse_file_compat_ioctl,
3022         .poll           = fuse_file_poll,
3023         .fallocate      = fuse_file_fallocate,
3024 };
3025
3026 static const struct file_operations fuse_direct_io_file_operations = {
3027         .llseek         = fuse_file_llseek,
3028         .read_iter      = fuse_direct_read_iter,
3029         .write_iter     = fuse_direct_write_iter,
3030         .mmap           = fuse_direct_mmap,
3031         .open           = fuse_open,
3032         .flush          = fuse_flush,
3033         .release        = fuse_release,
3034         .fsync          = fuse_fsync,
3035         .lock           = fuse_file_lock,
3036         .flock          = fuse_file_flock,
3037         .unlocked_ioctl = fuse_file_ioctl,
3038         .compat_ioctl   = fuse_file_compat_ioctl,
3039         .poll           = fuse_file_poll,
3040         .fallocate      = fuse_file_fallocate,
3041         /* no splice_read */
3042 };
3043
3044 static const struct address_space_operations fuse_file_aops  = {
3045         .readpage       = fuse_readpage,
3046         .writepage      = fuse_writepage,
3047         .writepages     = fuse_writepages,
3048         .launder_page   = fuse_launder_page,
3049         .readpages      = fuse_readpages,
3050         .set_page_dirty = __set_page_dirty_nobuffers,
3051         .bmap           = fuse_bmap,
3052         .direct_IO      = fuse_direct_IO,
3053         .write_begin    = fuse_write_begin,
3054         .write_end      = fuse_write_end,
3055 };
3056
3057 void fuse_init_file_inode(struct inode *inode)
3058 {
3059         inode->i_fop = &fuse_file_operations;
3060         inode->i_data.a_ops = &fuse_file_aops;
3061 }