Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mszeredi...
[linux-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
17 static const struct file_operations fuse_direct_io_file_operations;
18
19 static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
20                           int opcode, struct fuse_open_out *outargp)
21 {
22         struct fuse_open_in inarg;
23         struct fuse_req *req;
24         int err;
25
26         req = fuse_get_req(fc);
27         if (IS_ERR(req))
28                 return PTR_ERR(req);
29
30         memset(&inarg, 0, sizeof(inarg));
31         inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
32         if (!fc->atomic_o_trunc)
33                 inarg.flags &= ~O_TRUNC;
34         req->in.h.opcode = opcode;
35         req->in.h.nodeid = nodeid;
36         req->in.numargs = 1;
37         req->in.args[0].size = sizeof(inarg);
38         req->in.args[0].value = &inarg;
39         req->out.numargs = 1;
40         req->out.args[0].size = sizeof(*outargp);
41         req->out.args[0].value = outargp;
42         fuse_request_send(fc, req);
43         err = req->out.h.error;
44         fuse_put_request(fc, req);
45
46         return err;
47 }
48
49 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
50 {
51         struct fuse_file *ff;
52
53         ff = kmalloc(sizeof(struct fuse_file), GFP_KERNEL);
54         if (unlikely(!ff))
55                 return NULL;
56
57         ff->fc = fc;
58         ff->reserved_req = fuse_request_alloc();
59         if (unlikely(!ff->reserved_req)) {
60                 kfree(ff);
61                 return NULL;
62         }
63
64         INIT_LIST_HEAD(&ff->write_entry);
65         atomic_set(&ff->count, 0);
66         RB_CLEAR_NODE(&ff->polled_node);
67         init_waitqueue_head(&ff->poll_wait);
68
69         spin_lock(&fc->lock);
70         ff->kh = ++fc->khctr;
71         spin_unlock(&fc->lock);
72
73         return ff;
74 }
75
76 void fuse_file_free(struct fuse_file *ff)
77 {
78         fuse_request_free(ff->reserved_req);
79         kfree(ff);
80 }
81
82 struct fuse_file *fuse_file_get(struct fuse_file *ff)
83 {
84         atomic_inc(&ff->count);
85         return ff;
86 }
87
88 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
89 {
90         path_put(&req->misc.release.path);
91 }
92
93 static void fuse_file_put(struct fuse_file *ff)
94 {
95         if (atomic_dec_and_test(&ff->count)) {
96                 struct fuse_req *req = ff->reserved_req;
97
98                 req->end = fuse_release_end;
99                 fuse_request_send_background(ff->fc, req);
100                 kfree(ff);
101         }
102 }
103
104 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
105                  bool isdir)
106 {
107         struct fuse_open_out outarg;
108         struct fuse_file *ff;
109         int err;
110         int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
111
112         ff = fuse_file_alloc(fc);
113         if (!ff)
114                 return -ENOMEM;
115
116         err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
117         if (err) {
118                 fuse_file_free(ff);
119                 return err;
120         }
121
122         if (isdir)
123                 outarg.open_flags &= ~FOPEN_DIRECT_IO;
124
125         ff->fh = outarg.fh;
126         ff->nodeid = nodeid;
127         ff->open_flags = outarg.open_flags;
128         file->private_data = fuse_file_get(ff);
129
130         return 0;
131 }
132 EXPORT_SYMBOL_GPL(fuse_do_open);
133
134 void fuse_finish_open(struct inode *inode, struct file *file)
135 {
136         struct fuse_file *ff = file->private_data;
137
138         if (ff->open_flags & FOPEN_DIRECT_IO)
139                 file->f_op = &fuse_direct_io_file_operations;
140         if (!(ff->open_flags & FOPEN_KEEP_CACHE))
141                 invalidate_inode_pages2(inode->i_mapping);
142         if (ff->open_flags & FOPEN_NONSEEKABLE)
143                 nonseekable_open(inode, file);
144 }
145
146 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
147 {
148         struct fuse_conn *fc = get_fuse_conn(inode);
149         int err;
150
151         /* VFS checks this, but only _after_ ->open() */
152         if (file->f_flags & O_DIRECT)
153                 return -EINVAL;
154
155         err = generic_file_open(inode, file);
156         if (err)
157                 return err;
158
159         err = fuse_do_open(fc, get_node_id(inode), file, isdir);
160         if (err)
161                 return err;
162
163         fuse_finish_open(inode, file);
164
165         return 0;
166 }
167
168 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
169 {
170         struct fuse_conn *fc = ff->fc;
171         struct fuse_req *req = ff->reserved_req;
172         struct fuse_release_in *inarg = &req->misc.release.in;
173
174         spin_lock(&fc->lock);
175         list_del(&ff->write_entry);
176         if (!RB_EMPTY_NODE(&ff->polled_node))
177                 rb_erase(&ff->polled_node, &fc->polled_files);
178         spin_unlock(&fc->lock);
179
180         wake_up_interruptible_sync(&ff->poll_wait);
181
182         inarg->fh = ff->fh;
183         inarg->flags = flags;
184         req->in.h.opcode = opcode;
185         req->in.h.nodeid = ff->nodeid;
186         req->in.numargs = 1;
187         req->in.args[0].size = sizeof(struct fuse_release_in);
188         req->in.args[0].value = inarg;
189 }
190
191 void fuse_release_common(struct file *file, int opcode)
192 {
193         struct fuse_file *ff;
194         struct fuse_req *req;
195
196         ff = file->private_data;
197         if (unlikely(!ff))
198                 return;
199
200         req = ff->reserved_req;
201         fuse_prepare_release(ff, file->f_flags, opcode);
202
203         /* Hold vfsmount and dentry until release is finished */
204         path_get(&file->f_path);
205         req->misc.release.path = file->f_path;
206
207         /*
208          * Normally this will send the RELEASE request, however if
209          * some asynchronous READ or WRITE requests are outstanding,
210          * the sending will be delayed.
211          */
212         fuse_file_put(ff);
213 }
214
215 static int fuse_open(struct inode *inode, struct file *file)
216 {
217         return fuse_open_common(inode, file, false);
218 }
219
220 static int fuse_release(struct inode *inode, struct file *file)
221 {
222         fuse_release_common(file, FUSE_RELEASE);
223
224         /* return value is ignored by VFS */
225         return 0;
226 }
227
228 void fuse_sync_release(struct fuse_file *ff, int flags)
229 {
230         WARN_ON(atomic_read(&ff->count) > 1);
231         fuse_prepare_release(ff, flags, FUSE_RELEASE);
232         ff->reserved_req->force = 1;
233         fuse_request_send(ff->fc, ff->reserved_req);
234         fuse_put_request(ff->fc, ff->reserved_req);
235         kfree(ff);
236 }
237 EXPORT_SYMBOL_GPL(fuse_sync_release);
238
239 /*
240  * Scramble the ID space with XTEA, so that the value of the files_struct
241  * pointer is not exposed to userspace.
242  */
243 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
244 {
245         u32 *k = fc->scramble_key;
246         u64 v = (unsigned long) id;
247         u32 v0 = v;
248         u32 v1 = v >> 32;
249         u32 sum = 0;
250         int i;
251
252         for (i = 0; i < 32; i++) {
253                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
254                 sum += 0x9E3779B9;
255                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
256         }
257
258         return (u64) v0 + ((u64) v1 << 32);
259 }
260
261 /*
262  * Check if page is under writeback
263  *
264  * This is currently done by walking the list of writepage requests
265  * for the inode, which can be pretty inefficient.
266  */
267 static bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
268 {
269         struct fuse_conn *fc = get_fuse_conn(inode);
270         struct fuse_inode *fi = get_fuse_inode(inode);
271         struct fuse_req *req;
272         bool found = false;
273
274         spin_lock(&fc->lock);
275         list_for_each_entry(req, &fi->writepages, writepages_entry) {
276                 pgoff_t curr_index;
277
278                 BUG_ON(req->inode != inode);
279                 curr_index = req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
280                 if (curr_index == index) {
281                         found = true;
282                         break;
283                 }
284         }
285         spin_unlock(&fc->lock);
286
287         return found;
288 }
289
290 /*
291  * Wait for page writeback to be completed.
292  *
293  * Since fuse doesn't rely on the VM writeback tracking, this has to
294  * use some other means.
295  */
296 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
297 {
298         struct fuse_inode *fi = get_fuse_inode(inode);
299
300         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
301         return 0;
302 }
303
304 static int fuse_flush(struct file *file, fl_owner_t id)
305 {
306         struct inode *inode = file->f_path.dentry->d_inode;
307         struct fuse_conn *fc = get_fuse_conn(inode);
308         struct fuse_file *ff = file->private_data;
309         struct fuse_req *req;
310         struct fuse_flush_in inarg;
311         int err;
312
313         if (is_bad_inode(inode))
314                 return -EIO;
315
316         if (fc->no_flush)
317                 return 0;
318
319         req = fuse_get_req_nofail(fc, file);
320         memset(&inarg, 0, sizeof(inarg));
321         inarg.fh = ff->fh;
322         inarg.lock_owner = fuse_lock_owner_id(fc, id);
323         req->in.h.opcode = FUSE_FLUSH;
324         req->in.h.nodeid = get_node_id(inode);
325         req->in.numargs = 1;
326         req->in.args[0].size = sizeof(inarg);
327         req->in.args[0].value = &inarg;
328         req->force = 1;
329         fuse_request_send(fc, req);
330         err = req->out.h.error;
331         fuse_put_request(fc, req);
332         if (err == -ENOSYS) {
333                 fc->no_flush = 1;
334                 err = 0;
335         }
336         return err;
337 }
338
339 /*
340  * Wait for all pending writepages on the inode to finish.
341  *
342  * This is currently done by blocking further writes with FUSE_NOWRITE
343  * and waiting for all sent writes to complete.
344  *
345  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
346  * could conflict with truncation.
347  */
348 static void fuse_sync_writes(struct inode *inode)
349 {
350         fuse_set_nowrite(inode);
351         fuse_release_nowrite(inode);
352 }
353
354 int fuse_fsync_common(struct file *file, int datasync, int isdir)
355 {
356         struct inode *inode = file->f_mapping->host;
357         struct fuse_conn *fc = get_fuse_conn(inode);
358         struct fuse_file *ff = file->private_data;
359         struct fuse_req *req;
360         struct fuse_fsync_in inarg;
361         int err;
362
363         if (is_bad_inode(inode))
364                 return -EIO;
365
366         if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
367                 return 0;
368
369         /*
370          * Start writeback against all dirty pages of the inode, then
371          * wait for all outstanding writes, before sending the FSYNC
372          * request.
373          */
374         err = write_inode_now(inode, 0);
375         if (err)
376                 return err;
377
378         fuse_sync_writes(inode);
379
380         req = fuse_get_req(fc);
381         if (IS_ERR(req))
382                 return PTR_ERR(req);
383
384         memset(&inarg, 0, sizeof(inarg));
385         inarg.fh = ff->fh;
386         inarg.fsync_flags = datasync ? 1 : 0;
387         req->in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
388         req->in.h.nodeid = get_node_id(inode);
389         req->in.numargs = 1;
390         req->in.args[0].size = sizeof(inarg);
391         req->in.args[0].value = &inarg;
392         fuse_request_send(fc, req);
393         err = req->out.h.error;
394         fuse_put_request(fc, req);
395         if (err == -ENOSYS) {
396                 if (isdir)
397                         fc->no_fsyncdir = 1;
398                 else
399                         fc->no_fsync = 1;
400                 err = 0;
401         }
402         return err;
403 }
404
405 static int fuse_fsync(struct file *file, int datasync)
406 {
407         return fuse_fsync_common(file, datasync, 0);
408 }
409
410 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
411                     size_t count, int opcode)
412 {
413         struct fuse_read_in *inarg = &req->misc.read.in;
414         struct fuse_file *ff = file->private_data;
415
416         inarg->fh = ff->fh;
417         inarg->offset = pos;
418         inarg->size = count;
419         inarg->flags = file->f_flags;
420         req->in.h.opcode = opcode;
421         req->in.h.nodeid = ff->nodeid;
422         req->in.numargs = 1;
423         req->in.args[0].size = sizeof(struct fuse_read_in);
424         req->in.args[0].value = inarg;
425         req->out.argvar = 1;
426         req->out.numargs = 1;
427         req->out.args[0].size = count;
428 }
429
430 static size_t fuse_send_read(struct fuse_req *req, struct file *file,
431                              loff_t pos, size_t count, fl_owner_t owner)
432 {
433         struct fuse_file *ff = file->private_data;
434         struct fuse_conn *fc = ff->fc;
435
436         fuse_read_fill(req, file, pos, count, FUSE_READ);
437         if (owner != NULL) {
438                 struct fuse_read_in *inarg = &req->misc.read.in;
439
440                 inarg->read_flags |= FUSE_READ_LOCKOWNER;
441                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
442         }
443         fuse_request_send(fc, req);
444         return req->out.args[0].size;
445 }
446
447 static void fuse_read_update_size(struct inode *inode, loff_t size,
448                                   u64 attr_ver)
449 {
450         struct fuse_conn *fc = get_fuse_conn(inode);
451         struct fuse_inode *fi = get_fuse_inode(inode);
452
453         spin_lock(&fc->lock);
454         if (attr_ver == fi->attr_version && size < inode->i_size) {
455                 fi->attr_version = ++fc->attr_version;
456                 i_size_write(inode, size);
457         }
458         spin_unlock(&fc->lock);
459 }
460
461 static int fuse_readpage(struct file *file, struct page *page)
462 {
463         struct inode *inode = page->mapping->host;
464         struct fuse_conn *fc = get_fuse_conn(inode);
465         struct fuse_req *req;
466         size_t num_read;
467         loff_t pos = page_offset(page);
468         size_t count = PAGE_CACHE_SIZE;
469         u64 attr_ver;
470         int err;
471
472         err = -EIO;
473         if (is_bad_inode(inode))
474                 goto out;
475
476         /*
477          * Page writeback can extend beyond the liftime of the
478          * page-cache page, so make sure we read a properly synced
479          * page.
480          */
481         fuse_wait_on_page_writeback(inode, page->index);
482
483         req = fuse_get_req(fc);
484         err = PTR_ERR(req);
485         if (IS_ERR(req))
486                 goto out;
487
488         attr_ver = fuse_get_attr_version(fc);
489
490         req->out.page_zeroing = 1;
491         req->out.argpages = 1;
492         req->num_pages = 1;
493         req->pages[0] = page;
494         num_read = fuse_send_read(req, file, pos, count, NULL);
495         err = req->out.h.error;
496         fuse_put_request(fc, req);
497
498         if (!err) {
499                 /*
500                  * Short read means EOF.  If file size is larger, truncate it
501                  */
502                 if (num_read < count)
503                         fuse_read_update_size(inode, pos + num_read, attr_ver);
504
505                 SetPageUptodate(page);
506         }
507
508         fuse_invalidate_attr(inode); /* atime changed */
509  out:
510         unlock_page(page);
511         return err;
512 }
513
514 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
515 {
516         int i;
517         size_t count = req->misc.read.in.size;
518         size_t num_read = req->out.args[0].size;
519         struct address_space *mapping = NULL;
520
521         for (i = 0; mapping == NULL && i < req->num_pages; i++)
522                 mapping = req->pages[i]->mapping;
523
524         if (mapping) {
525                 struct inode *inode = mapping->host;
526
527                 /*
528                  * Short read means EOF. If file size is larger, truncate it
529                  */
530                 if (!req->out.h.error && num_read < count) {
531                         loff_t pos;
532
533                         pos = page_offset(req->pages[0]) + num_read;
534                         fuse_read_update_size(inode, pos,
535                                               req->misc.read.attr_ver);
536                 }
537                 fuse_invalidate_attr(inode); /* atime changed */
538         }
539
540         for (i = 0; i < req->num_pages; i++) {
541                 struct page *page = req->pages[i];
542                 if (!req->out.h.error)
543                         SetPageUptodate(page);
544                 else
545                         SetPageError(page);
546                 unlock_page(page);
547                 page_cache_release(page);
548         }
549         if (req->ff)
550                 fuse_file_put(req->ff);
551 }
552
553 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
554 {
555         struct fuse_file *ff = file->private_data;
556         struct fuse_conn *fc = ff->fc;
557         loff_t pos = page_offset(req->pages[0]);
558         size_t count = req->num_pages << PAGE_CACHE_SHIFT;
559
560         req->out.argpages = 1;
561         req->out.page_zeroing = 1;
562         req->out.page_replace = 1;
563         fuse_read_fill(req, file, pos, count, FUSE_READ);
564         req->misc.read.attr_ver = fuse_get_attr_version(fc);
565         if (fc->async_read) {
566                 req->ff = fuse_file_get(ff);
567                 req->end = fuse_readpages_end;
568                 fuse_request_send_background(fc, req);
569         } else {
570                 fuse_request_send(fc, req);
571                 fuse_readpages_end(fc, req);
572                 fuse_put_request(fc, req);
573         }
574 }
575
576 struct fuse_fill_data {
577         struct fuse_req *req;
578         struct file *file;
579         struct inode *inode;
580 };
581
582 static int fuse_readpages_fill(void *_data, struct page *page)
583 {
584         struct fuse_fill_data *data = _data;
585         struct fuse_req *req = data->req;
586         struct inode *inode = data->inode;
587         struct fuse_conn *fc = get_fuse_conn(inode);
588
589         fuse_wait_on_page_writeback(inode, page->index);
590
591         if (req->num_pages &&
592             (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
593              (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
594              req->pages[req->num_pages - 1]->index + 1 != page->index)) {
595                 fuse_send_readpages(req, data->file);
596                 data->req = req = fuse_get_req(fc);
597                 if (IS_ERR(req)) {
598                         unlock_page(page);
599                         return PTR_ERR(req);
600                 }
601         }
602         page_cache_get(page);
603         req->pages[req->num_pages] = page;
604         req->num_pages++;
605         return 0;
606 }
607
608 static int fuse_readpages(struct file *file, struct address_space *mapping,
609                           struct list_head *pages, unsigned nr_pages)
610 {
611         struct inode *inode = mapping->host;
612         struct fuse_conn *fc = get_fuse_conn(inode);
613         struct fuse_fill_data data;
614         int err;
615
616         err = -EIO;
617         if (is_bad_inode(inode))
618                 goto out;
619
620         data.file = file;
621         data.inode = inode;
622         data.req = fuse_get_req(fc);
623         err = PTR_ERR(data.req);
624         if (IS_ERR(data.req))
625                 goto out;
626
627         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
628         if (!err) {
629                 if (data.req->num_pages)
630                         fuse_send_readpages(data.req, file);
631                 else
632                         fuse_put_request(fc, data.req);
633         }
634 out:
635         return err;
636 }
637
638 static ssize_t fuse_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
639                                   unsigned long nr_segs, loff_t pos)
640 {
641         struct inode *inode = iocb->ki_filp->f_mapping->host;
642
643         if (pos + iov_length(iov, nr_segs) > i_size_read(inode)) {
644                 int err;
645                 /*
646                  * If trying to read past EOF, make sure the i_size
647                  * attribute is up-to-date.
648                  */
649                 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
650                 if (err)
651                         return err;
652         }
653
654         return generic_file_aio_read(iocb, iov, nr_segs, pos);
655 }
656
657 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
658                             loff_t pos, size_t count)
659 {
660         struct fuse_write_in *inarg = &req->misc.write.in;
661         struct fuse_write_out *outarg = &req->misc.write.out;
662
663         inarg->fh = ff->fh;
664         inarg->offset = pos;
665         inarg->size = count;
666         req->in.h.opcode = FUSE_WRITE;
667         req->in.h.nodeid = ff->nodeid;
668         req->in.numargs = 2;
669         if (ff->fc->minor < 9)
670                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
671         else
672                 req->in.args[0].size = sizeof(struct fuse_write_in);
673         req->in.args[0].value = inarg;
674         req->in.args[1].size = count;
675         req->out.numargs = 1;
676         req->out.args[0].size = sizeof(struct fuse_write_out);
677         req->out.args[0].value = outarg;
678 }
679
680 static size_t fuse_send_write(struct fuse_req *req, struct file *file,
681                               loff_t pos, size_t count, fl_owner_t owner)
682 {
683         struct fuse_file *ff = file->private_data;
684         struct fuse_conn *fc = ff->fc;
685         struct fuse_write_in *inarg = &req->misc.write.in;
686
687         fuse_write_fill(req, ff, pos, count);
688         inarg->flags = file->f_flags;
689         if (owner != NULL) {
690                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
691                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
692         }
693         fuse_request_send(fc, req);
694         return req->misc.write.out.size;
695 }
696
697 static int fuse_write_begin(struct file *file, struct address_space *mapping,
698                         loff_t pos, unsigned len, unsigned flags,
699                         struct page **pagep, void **fsdata)
700 {
701         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
702
703         *pagep = grab_cache_page_write_begin(mapping, index, flags);
704         if (!*pagep)
705                 return -ENOMEM;
706         return 0;
707 }
708
709 static void fuse_write_update_size(struct inode *inode, loff_t pos)
710 {
711         struct fuse_conn *fc = get_fuse_conn(inode);
712         struct fuse_inode *fi = get_fuse_inode(inode);
713
714         spin_lock(&fc->lock);
715         fi->attr_version = ++fc->attr_version;
716         if (pos > inode->i_size)
717                 i_size_write(inode, pos);
718         spin_unlock(&fc->lock);
719 }
720
721 static int fuse_buffered_write(struct file *file, struct inode *inode,
722                                loff_t pos, unsigned count, struct page *page)
723 {
724         int err;
725         size_t nres;
726         struct fuse_conn *fc = get_fuse_conn(inode);
727         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
728         struct fuse_req *req;
729
730         if (is_bad_inode(inode))
731                 return -EIO;
732
733         /*
734          * Make sure writepages on the same page are not mixed up with
735          * plain writes.
736          */
737         fuse_wait_on_page_writeback(inode, page->index);
738
739         req = fuse_get_req(fc);
740         if (IS_ERR(req))
741                 return PTR_ERR(req);
742
743         req->in.argpages = 1;
744         req->num_pages = 1;
745         req->pages[0] = page;
746         req->page_offset = offset;
747         nres = fuse_send_write(req, file, pos, count, NULL);
748         err = req->out.h.error;
749         fuse_put_request(fc, req);
750         if (!err && !nres)
751                 err = -EIO;
752         if (!err) {
753                 pos += nres;
754                 fuse_write_update_size(inode, pos);
755                 if (count == PAGE_CACHE_SIZE)
756                         SetPageUptodate(page);
757         }
758         fuse_invalidate_attr(inode);
759         return err ? err : nres;
760 }
761
762 static int fuse_write_end(struct file *file, struct address_space *mapping,
763                         loff_t pos, unsigned len, unsigned copied,
764                         struct page *page, void *fsdata)
765 {
766         struct inode *inode = mapping->host;
767         int res = 0;
768
769         if (copied)
770                 res = fuse_buffered_write(file, inode, pos, copied, page);
771
772         unlock_page(page);
773         page_cache_release(page);
774         return res;
775 }
776
777 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
778                                     struct inode *inode, loff_t pos,
779                                     size_t count)
780 {
781         size_t res;
782         unsigned offset;
783         unsigned i;
784
785         for (i = 0; i < req->num_pages; i++)
786                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
787
788         res = fuse_send_write(req, file, pos, count, NULL);
789
790         offset = req->page_offset;
791         count = res;
792         for (i = 0; i < req->num_pages; i++) {
793                 struct page *page = req->pages[i];
794
795                 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
796                         SetPageUptodate(page);
797
798                 if (count > PAGE_CACHE_SIZE - offset)
799                         count -= PAGE_CACHE_SIZE - offset;
800                 else
801                         count = 0;
802                 offset = 0;
803
804                 unlock_page(page);
805                 page_cache_release(page);
806         }
807
808         return res;
809 }
810
811 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
812                                struct address_space *mapping,
813                                struct iov_iter *ii, loff_t pos)
814 {
815         struct fuse_conn *fc = get_fuse_conn(mapping->host);
816         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
817         size_t count = 0;
818         int err;
819
820         req->in.argpages = 1;
821         req->page_offset = offset;
822
823         do {
824                 size_t tmp;
825                 struct page *page;
826                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
827                 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
828                                      iov_iter_count(ii));
829
830                 bytes = min_t(size_t, bytes, fc->max_write - count);
831
832  again:
833                 err = -EFAULT;
834                 if (iov_iter_fault_in_readable(ii, bytes))
835                         break;
836
837                 err = -ENOMEM;
838                 page = grab_cache_page_write_begin(mapping, index, 0);
839                 if (!page)
840                         break;
841
842                 if (mapping_writably_mapped(mapping))
843                         flush_dcache_page(page);
844
845                 pagefault_disable();
846                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
847                 pagefault_enable();
848                 flush_dcache_page(page);
849
850                 if (!tmp) {
851                         unlock_page(page);
852                         page_cache_release(page);
853                         bytes = min(bytes, iov_iter_single_seg_count(ii));
854                         goto again;
855                 }
856
857                 err = 0;
858                 req->pages[req->num_pages] = page;
859                 req->num_pages++;
860
861                 iov_iter_advance(ii, tmp);
862                 count += tmp;
863                 pos += tmp;
864                 offset += tmp;
865                 if (offset == PAGE_CACHE_SIZE)
866                         offset = 0;
867
868                 if (!fc->big_writes)
869                         break;
870         } while (iov_iter_count(ii) && count < fc->max_write &&
871                  req->num_pages < FUSE_MAX_PAGES_PER_REQ && offset == 0);
872
873         return count > 0 ? count : err;
874 }
875
876 static ssize_t fuse_perform_write(struct file *file,
877                                   struct address_space *mapping,
878                                   struct iov_iter *ii, loff_t pos)
879 {
880         struct inode *inode = mapping->host;
881         struct fuse_conn *fc = get_fuse_conn(inode);
882         int err = 0;
883         ssize_t res = 0;
884
885         if (is_bad_inode(inode))
886                 return -EIO;
887
888         do {
889                 struct fuse_req *req;
890                 ssize_t count;
891
892                 req = fuse_get_req(fc);
893                 if (IS_ERR(req)) {
894                         err = PTR_ERR(req);
895                         break;
896                 }
897
898                 count = fuse_fill_write_pages(req, mapping, ii, pos);
899                 if (count <= 0) {
900                         err = count;
901                 } else {
902                         size_t num_written;
903
904                         num_written = fuse_send_write_pages(req, file, inode,
905                                                             pos, count);
906                         err = req->out.h.error;
907                         if (!err) {
908                                 res += num_written;
909                                 pos += num_written;
910
911                                 /* break out of the loop on short write */
912                                 if (num_written != count)
913                                         err = -EIO;
914                         }
915                 }
916                 fuse_put_request(fc, req);
917         } while (!err && iov_iter_count(ii));
918
919         if (res > 0)
920                 fuse_write_update_size(inode, pos);
921
922         fuse_invalidate_attr(inode);
923
924         return res > 0 ? res : err;
925 }
926
927 static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
928                                    unsigned long nr_segs, loff_t pos)
929 {
930         struct file *file = iocb->ki_filp;
931         struct address_space *mapping = file->f_mapping;
932         size_t count = 0;
933         ssize_t written = 0;
934         struct inode *inode = mapping->host;
935         ssize_t err;
936         struct iov_iter i;
937
938         WARN_ON(iocb->ki_pos != pos);
939
940         err = generic_segment_checks(iov, &nr_segs, &count, VERIFY_READ);
941         if (err)
942                 return err;
943
944         mutex_lock(&inode->i_mutex);
945         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
946
947         /* We can write back this queue in page reclaim */
948         current->backing_dev_info = mapping->backing_dev_info;
949
950         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
951         if (err)
952                 goto out;
953
954         if (count == 0)
955                 goto out;
956
957         err = file_remove_suid(file);
958         if (err)
959                 goto out;
960
961         file_update_time(file);
962
963         iov_iter_init(&i, iov, nr_segs, count, 0);
964         written = fuse_perform_write(file, mapping, &i, pos);
965         if (written >= 0)
966                 iocb->ki_pos = pos + written;
967
968 out:
969         current->backing_dev_info = NULL;
970         mutex_unlock(&inode->i_mutex);
971
972         return written ? written : err;
973 }
974
975 static void fuse_release_user_pages(struct fuse_req *req, int write)
976 {
977         unsigned i;
978
979         for (i = 0; i < req->num_pages; i++) {
980                 struct page *page = req->pages[i];
981                 if (write)
982                         set_page_dirty_lock(page);
983                 put_page(page);
984         }
985 }
986
987 static int fuse_get_user_pages(struct fuse_req *req, const char __user *buf,
988                                size_t *nbytesp, int write)
989 {
990         size_t nbytes = *nbytesp;
991         unsigned long user_addr = (unsigned long) buf;
992         unsigned offset = user_addr & ~PAGE_MASK;
993         int npages;
994
995         /* Special case for kernel I/O: can copy directly into the buffer */
996         if (segment_eq(get_fs(), KERNEL_DS)) {
997                 if (write)
998                         req->in.args[1].value = (void *) user_addr;
999                 else
1000                         req->out.args[0].value = (void *) user_addr;
1001
1002                 return 0;
1003         }
1004
1005         nbytes = min_t(size_t, nbytes, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
1006         npages = (nbytes + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1007         npages = clamp(npages, 1, FUSE_MAX_PAGES_PER_REQ);
1008         npages = get_user_pages_fast(user_addr, npages, !write, req->pages);
1009         if (npages < 0)
1010                 return npages;
1011
1012         req->num_pages = npages;
1013         req->page_offset = offset;
1014
1015         if (write)
1016                 req->in.argpages = 1;
1017         else
1018                 req->out.argpages = 1;
1019
1020         nbytes = (req->num_pages << PAGE_SHIFT) - req->page_offset;
1021         *nbytesp = min(*nbytesp, nbytes);
1022
1023         return 0;
1024 }
1025
1026 ssize_t fuse_direct_io(struct file *file, const char __user *buf,
1027                        size_t count, loff_t *ppos, int write)
1028 {
1029         struct fuse_file *ff = file->private_data;
1030         struct fuse_conn *fc = ff->fc;
1031         size_t nmax = write ? fc->max_write : fc->max_read;
1032         loff_t pos = *ppos;
1033         ssize_t res = 0;
1034         struct fuse_req *req;
1035
1036         req = fuse_get_req(fc);
1037         if (IS_ERR(req))
1038                 return PTR_ERR(req);
1039
1040         while (count) {
1041                 size_t nres;
1042                 fl_owner_t owner = current->files;
1043                 size_t nbytes = min(count, nmax);
1044                 int err = fuse_get_user_pages(req, buf, &nbytes, write);
1045                 if (err) {
1046                         res = err;
1047                         break;
1048                 }
1049
1050                 if (write)
1051                         nres = fuse_send_write(req, file, pos, nbytes, owner);
1052                 else
1053                         nres = fuse_send_read(req, file, pos, nbytes, owner);
1054
1055                 fuse_release_user_pages(req, !write);
1056                 if (req->out.h.error) {
1057                         if (!res)
1058                                 res = req->out.h.error;
1059                         break;
1060                 } else if (nres > nbytes) {
1061                         res = -EIO;
1062                         break;
1063                 }
1064                 count -= nres;
1065                 res += nres;
1066                 pos += nres;
1067                 buf += nres;
1068                 if (nres != nbytes)
1069                         break;
1070                 if (count) {
1071                         fuse_put_request(fc, req);
1072                         req = fuse_get_req(fc);
1073                         if (IS_ERR(req))
1074                                 break;
1075                 }
1076         }
1077         if (!IS_ERR(req))
1078                 fuse_put_request(fc, req);
1079         if (res > 0)
1080                 *ppos = pos;
1081
1082         return res;
1083 }
1084 EXPORT_SYMBOL_GPL(fuse_direct_io);
1085
1086 static ssize_t fuse_direct_read(struct file *file, char __user *buf,
1087                                      size_t count, loff_t *ppos)
1088 {
1089         ssize_t res;
1090         struct inode *inode = file->f_path.dentry->d_inode;
1091
1092         if (is_bad_inode(inode))
1093                 return -EIO;
1094
1095         res = fuse_direct_io(file, buf, count, ppos, 0);
1096
1097         fuse_invalidate_attr(inode);
1098
1099         return res;
1100 }
1101
1102 static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
1103                                  size_t count, loff_t *ppos)
1104 {
1105         struct inode *inode = file->f_path.dentry->d_inode;
1106         ssize_t res;
1107
1108         if (is_bad_inode(inode))
1109                 return -EIO;
1110
1111         /* Don't allow parallel writes to the same file */
1112         mutex_lock(&inode->i_mutex);
1113         res = generic_write_checks(file, ppos, &count, 0);
1114         if (!res) {
1115                 res = fuse_direct_io(file, buf, count, ppos, 1);
1116                 if (res > 0)
1117                         fuse_write_update_size(inode, *ppos);
1118         }
1119         mutex_unlock(&inode->i_mutex);
1120
1121         fuse_invalidate_attr(inode);
1122
1123         return res;
1124 }
1125
1126 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1127 {
1128         __free_page(req->pages[0]);
1129         fuse_file_put(req->ff);
1130 }
1131
1132 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1133 {
1134         struct inode *inode = req->inode;
1135         struct fuse_inode *fi = get_fuse_inode(inode);
1136         struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
1137
1138         list_del(&req->writepages_entry);
1139         dec_bdi_stat(bdi, BDI_WRITEBACK);
1140         dec_zone_page_state(req->pages[0], NR_WRITEBACK_TEMP);
1141         bdi_writeout_inc(bdi);
1142         wake_up(&fi->page_waitq);
1143 }
1144
1145 /* Called under fc->lock, may release and reacquire it */
1146 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req)
1147 __releases(&fc->lock)
1148 __acquires(&fc->lock)
1149 {
1150         struct fuse_inode *fi = get_fuse_inode(req->inode);
1151         loff_t size = i_size_read(req->inode);
1152         struct fuse_write_in *inarg = &req->misc.write.in;
1153
1154         if (!fc->connected)
1155                 goto out_free;
1156
1157         if (inarg->offset + PAGE_CACHE_SIZE <= size) {
1158                 inarg->size = PAGE_CACHE_SIZE;
1159         } else if (inarg->offset < size) {
1160                 inarg->size = size & (PAGE_CACHE_SIZE - 1);
1161         } else {
1162                 /* Got truncated off completely */
1163                 goto out_free;
1164         }
1165
1166         req->in.args[1].size = inarg->size;
1167         fi->writectr++;
1168         fuse_request_send_background_locked(fc, req);
1169         return;
1170
1171  out_free:
1172         fuse_writepage_finish(fc, req);
1173         spin_unlock(&fc->lock);
1174         fuse_writepage_free(fc, req);
1175         fuse_put_request(fc, req);
1176         spin_lock(&fc->lock);
1177 }
1178
1179 /*
1180  * If fi->writectr is positive (no truncate or fsync going on) send
1181  * all queued writepage requests.
1182  *
1183  * Called with fc->lock
1184  */
1185 void fuse_flush_writepages(struct inode *inode)
1186 __releases(&fc->lock)
1187 __acquires(&fc->lock)
1188 {
1189         struct fuse_conn *fc = get_fuse_conn(inode);
1190         struct fuse_inode *fi = get_fuse_inode(inode);
1191         struct fuse_req *req;
1192
1193         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1194                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1195                 list_del_init(&req->list);
1196                 fuse_send_writepage(fc, req);
1197         }
1198 }
1199
1200 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1201 {
1202         struct inode *inode = req->inode;
1203         struct fuse_inode *fi = get_fuse_inode(inode);
1204
1205         mapping_set_error(inode->i_mapping, req->out.h.error);
1206         spin_lock(&fc->lock);
1207         fi->writectr--;
1208         fuse_writepage_finish(fc, req);
1209         spin_unlock(&fc->lock);
1210         fuse_writepage_free(fc, req);
1211 }
1212
1213 static int fuse_writepage_locked(struct page *page)
1214 {
1215         struct address_space *mapping = page->mapping;
1216         struct inode *inode = mapping->host;
1217         struct fuse_conn *fc = get_fuse_conn(inode);
1218         struct fuse_inode *fi = get_fuse_inode(inode);
1219         struct fuse_req *req;
1220         struct fuse_file *ff;
1221         struct page *tmp_page;
1222
1223         set_page_writeback(page);
1224
1225         req = fuse_request_alloc_nofs();
1226         if (!req)
1227                 goto err;
1228
1229         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1230         if (!tmp_page)
1231                 goto err_free;
1232
1233         spin_lock(&fc->lock);
1234         BUG_ON(list_empty(&fi->write_files));
1235         ff = list_entry(fi->write_files.next, struct fuse_file, write_entry);
1236         req->ff = fuse_file_get(ff);
1237         spin_unlock(&fc->lock);
1238
1239         fuse_write_fill(req, ff, page_offset(page), 0);
1240
1241         copy_highpage(tmp_page, page);
1242         req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1243         req->in.argpages = 1;
1244         req->num_pages = 1;
1245         req->pages[0] = tmp_page;
1246         req->page_offset = 0;
1247         req->end = fuse_writepage_end;
1248         req->inode = inode;
1249
1250         inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
1251         inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1252         end_page_writeback(page);
1253
1254         spin_lock(&fc->lock);
1255         list_add(&req->writepages_entry, &fi->writepages);
1256         list_add_tail(&req->list, &fi->queued_writes);
1257         fuse_flush_writepages(inode);
1258         spin_unlock(&fc->lock);
1259
1260         return 0;
1261
1262 err_free:
1263         fuse_request_free(req);
1264 err:
1265         end_page_writeback(page);
1266         return -ENOMEM;
1267 }
1268
1269 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1270 {
1271         int err;
1272
1273         err = fuse_writepage_locked(page);
1274         unlock_page(page);
1275
1276         return err;
1277 }
1278
1279 static int fuse_launder_page(struct page *page)
1280 {
1281         int err = 0;
1282         if (clear_page_dirty_for_io(page)) {
1283                 struct inode *inode = page->mapping->host;
1284                 err = fuse_writepage_locked(page);
1285                 if (!err)
1286                         fuse_wait_on_page_writeback(inode, page->index);
1287         }
1288         return err;
1289 }
1290
1291 /*
1292  * Write back dirty pages now, because there may not be any suitable
1293  * open files later
1294  */
1295 static void fuse_vma_close(struct vm_area_struct *vma)
1296 {
1297         filemap_write_and_wait(vma->vm_file->f_mapping);
1298 }
1299
1300 /*
1301  * Wait for writeback against this page to complete before allowing it
1302  * to be marked dirty again, and hence written back again, possibly
1303  * before the previous writepage completed.
1304  *
1305  * Block here, instead of in ->writepage(), so that the userspace fs
1306  * can only block processes actually operating on the filesystem.
1307  *
1308  * Otherwise unprivileged userspace fs would be able to block
1309  * unrelated:
1310  *
1311  * - page migration
1312  * - sync(2)
1313  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
1314  */
1315 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1316 {
1317         struct page *page = vmf->page;
1318         /*
1319          * Don't use page->mapping as it may become NULL from a
1320          * concurrent truncate.
1321          */
1322         struct inode *inode = vma->vm_file->f_mapping->host;
1323
1324         fuse_wait_on_page_writeback(inode, page->index);
1325         return 0;
1326 }
1327
1328 static const struct vm_operations_struct fuse_file_vm_ops = {
1329         .close          = fuse_vma_close,
1330         .fault          = filemap_fault,
1331         .page_mkwrite   = fuse_page_mkwrite,
1332 };
1333
1334 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
1335 {
1336         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE)) {
1337                 struct inode *inode = file->f_dentry->d_inode;
1338                 struct fuse_conn *fc = get_fuse_conn(inode);
1339                 struct fuse_inode *fi = get_fuse_inode(inode);
1340                 struct fuse_file *ff = file->private_data;
1341                 /*
1342                  * file may be written through mmap, so chain it onto the
1343                  * inodes's write_file list
1344                  */
1345                 spin_lock(&fc->lock);
1346                 if (list_empty(&ff->write_entry))
1347                         list_add(&ff->write_entry, &fi->write_files);
1348                 spin_unlock(&fc->lock);
1349         }
1350         file_accessed(file);
1351         vma->vm_ops = &fuse_file_vm_ops;
1352         return 0;
1353 }
1354
1355 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
1356 {
1357         /* Can't provide the coherency needed for MAP_SHARED */
1358         if (vma->vm_flags & VM_MAYSHARE)
1359                 return -ENODEV;
1360
1361         invalidate_inode_pages2(file->f_mapping);
1362
1363         return generic_file_mmap(file, vma);
1364 }
1365
1366 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
1367                                   struct file_lock *fl)
1368 {
1369         switch (ffl->type) {
1370         case F_UNLCK:
1371                 break;
1372
1373         case F_RDLCK:
1374         case F_WRLCK:
1375                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
1376                     ffl->end < ffl->start)
1377                         return -EIO;
1378
1379                 fl->fl_start = ffl->start;
1380                 fl->fl_end = ffl->end;
1381                 fl->fl_pid = ffl->pid;
1382                 break;
1383
1384         default:
1385                 return -EIO;
1386         }
1387         fl->fl_type = ffl->type;
1388         return 0;
1389 }
1390
1391 static void fuse_lk_fill(struct fuse_req *req, struct file *file,
1392                          const struct file_lock *fl, int opcode, pid_t pid,
1393                          int flock)
1394 {
1395         struct inode *inode = file->f_path.dentry->d_inode;
1396         struct fuse_conn *fc = get_fuse_conn(inode);
1397         struct fuse_file *ff = file->private_data;
1398         struct fuse_lk_in *arg = &req->misc.lk_in;
1399
1400         arg->fh = ff->fh;
1401         arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
1402         arg->lk.start = fl->fl_start;
1403         arg->lk.end = fl->fl_end;
1404         arg->lk.type = fl->fl_type;
1405         arg->lk.pid = pid;
1406         if (flock)
1407                 arg->lk_flags |= FUSE_LK_FLOCK;
1408         req->in.h.opcode = opcode;
1409         req->in.h.nodeid = get_node_id(inode);
1410         req->in.numargs = 1;
1411         req->in.args[0].size = sizeof(*arg);
1412         req->in.args[0].value = arg;
1413 }
1414
1415 static int fuse_getlk(struct file *file, struct file_lock *fl)
1416 {
1417         struct inode *inode = file->f_path.dentry->d_inode;
1418         struct fuse_conn *fc = get_fuse_conn(inode);
1419         struct fuse_req *req;
1420         struct fuse_lk_out outarg;
1421         int err;
1422
1423         req = fuse_get_req(fc);
1424         if (IS_ERR(req))
1425                 return PTR_ERR(req);
1426
1427         fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
1428         req->out.numargs = 1;
1429         req->out.args[0].size = sizeof(outarg);
1430         req->out.args[0].value = &outarg;
1431         fuse_request_send(fc, req);
1432         err = req->out.h.error;
1433         fuse_put_request(fc, req);
1434         if (!err)
1435                 err = convert_fuse_file_lock(&outarg.lk, fl);
1436
1437         return err;
1438 }
1439
1440 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
1441 {
1442         struct inode *inode = file->f_path.dentry->d_inode;
1443         struct fuse_conn *fc = get_fuse_conn(inode);
1444         struct fuse_req *req;
1445         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
1446         pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
1447         int err;
1448
1449         if (fl->fl_lmops && fl->fl_lmops->fl_grant) {
1450                 /* NLM needs asynchronous locks, which we don't support yet */
1451                 return -ENOLCK;
1452         }
1453
1454         /* Unlock on close is handled by the flush method */
1455         if (fl->fl_flags & FL_CLOSE)
1456                 return 0;
1457
1458         req = fuse_get_req(fc);
1459         if (IS_ERR(req))
1460                 return PTR_ERR(req);
1461
1462         fuse_lk_fill(req, file, fl, opcode, pid, flock);
1463         fuse_request_send(fc, req);
1464         err = req->out.h.error;
1465         /* locking is restartable */
1466         if (err == -EINTR)
1467                 err = -ERESTARTSYS;
1468         fuse_put_request(fc, req);
1469         return err;
1470 }
1471
1472 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
1473 {
1474         struct inode *inode = file->f_path.dentry->d_inode;
1475         struct fuse_conn *fc = get_fuse_conn(inode);
1476         int err;
1477
1478         if (cmd == F_CANCELLK) {
1479                 err = 0;
1480         } else if (cmd == F_GETLK) {
1481                 if (fc->no_lock) {
1482                         posix_test_lock(file, fl);
1483                         err = 0;
1484                 } else
1485                         err = fuse_getlk(file, fl);
1486         } else {
1487                 if (fc->no_lock)
1488                         err = posix_lock_file(file, fl, NULL);
1489                 else
1490                         err = fuse_setlk(file, fl, 0);
1491         }
1492         return err;
1493 }
1494
1495 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
1496 {
1497         struct inode *inode = file->f_path.dentry->d_inode;
1498         struct fuse_conn *fc = get_fuse_conn(inode);
1499         int err;
1500
1501         if (fc->no_lock) {
1502                 err = flock_lock_file_wait(file, fl);
1503         } else {
1504                 /* emulate flock with POSIX locks */
1505                 fl->fl_owner = (fl_owner_t) file;
1506                 err = fuse_setlk(file, fl, 1);
1507         }
1508
1509         return err;
1510 }
1511
1512 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
1513 {
1514         struct inode *inode = mapping->host;
1515         struct fuse_conn *fc = get_fuse_conn(inode);
1516         struct fuse_req *req;
1517         struct fuse_bmap_in inarg;
1518         struct fuse_bmap_out outarg;
1519         int err;
1520
1521         if (!inode->i_sb->s_bdev || fc->no_bmap)
1522                 return 0;
1523
1524         req = fuse_get_req(fc);
1525         if (IS_ERR(req))
1526                 return 0;
1527
1528         memset(&inarg, 0, sizeof(inarg));
1529         inarg.block = block;
1530         inarg.blocksize = inode->i_sb->s_blocksize;
1531         req->in.h.opcode = FUSE_BMAP;
1532         req->in.h.nodeid = get_node_id(inode);
1533         req->in.numargs = 1;
1534         req->in.args[0].size = sizeof(inarg);
1535         req->in.args[0].value = &inarg;
1536         req->out.numargs = 1;
1537         req->out.args[0].size = sizeof(outarg);
1538         req->out.args[0].value = &outarg;
1539         fuse_request_send(fc, req);
1540         err = req->out.h.error;
1541         fuse_put_request(fc, req);
1542         if (err == -ENOSYS)
1543                 fc->no_bmap = 1;
1544
1545         return err ? 0 : outarg.block;
1546 }
1547
1548 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int origin)
1549 {
1550         loff_t retval;
1551         struct inode *inode = file->f_path.dentry->d_inode;
1552
1553         mutex_lock(&inode->i_mutex);
1554         switch (origin) {
1555         case SEEK_END:
1556                 retval = fuse_update_attributes(inode, NULL, file, NULL);
1557                 if (retval)
1558                         goto exit;
1559                 offset += i_size_read(inode);
1560                 break;
1561         case SEEK_CUR:
1562                 offset += file->f_pos;
1563         }
1564         retval = -EINVAL;
1565         if (offset >= 0 && offset <= inode->i_sb->s_maxbytes) {
1566                 if (offset != file->f_pos) {
1567                         file->f_pos = offset;
1568                         file->f_version = 0;
1569                 }
1570                 retval = offset;
1571         }
1572 exit:
1573         mutex_unlock(&inode->i_mutex);
1574         return retval;
1575 }
1576
1577 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
1578                         unsigned int nr_segs, size_t bytes, bool to_user)
1579 {
1580         struct iov_iter ii;
1581         int page_idx = 0;
1582
1583         if (!bytes)
1584                 return 0;
1585
1586         iov_iter_init(&ii, iov, nr_segs, bytes, 0);
1587
1588         while (iov_iter_count(&ii)) {
1589                 struct page *page = pages[page_idx++];
1590                 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
1591                 void *kaddr;
1592
1593                 kaddr = kmap(page);
1594
1595                 while (todo) {
1596                         char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
1597                         size_t iov_len = ii.iov->iov_len - ii.iov_offset;
1598                         size_t copy = min(todo, iov_len);
1599                         size_t left;
1600
1601                         if (!to_user)
1602                                 left = copy_from_user(kaddr, uaddr, copy);
1603                         else
1604                                 left = copy_to_user(uaddr, kaddr, copy);
1605
1606                         if (unlikely(left))
1607                                 return -EFAULT;
1608
1609                         iov_iter_advance(&ii, copy);
1610                         todo -= copy;
1611                         kaddr += copy;
1612                 }
1613
1614                 kunmap(page);
1615         }
1616
1617         return 0;
1618 }
1619
1620 /*
1621  * For ioctls, there is no generic way to determine how much memory
1622  * needs to be read and/or written.  Furthermore, ioctls are allowed
1623  * to dereference the passed pointer, so the parameter requires deep
1624  * copying but FUSE has no idea whatsoever about what to copy in or
1625  * out.
1626  *
1627  * This is solved by allowing FUSE server to retry ioctl with
1628  * necessary in/out iovecs.  Let's assume the ioctl implementation
1629  * needs to read in the following structure.
1630  *
1631  * struct a {
1632  *      char    *buf;
1633  *      size_t  buflen;
1634  * }
1635  *
1636  * On the first callout to FUSE server, inarg->in_size and
1637  * inarg->out_size will be NULL; then, the server completes the ioctl
1638  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
1639  * the actual iov array to
1640  *
1641  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
1642  *
1643  * which tells FUSE to copy in the requested area and retry the ioctl.
1644  * On the second round, the server has access to the structure and
1645  * from that it can tell what to look for next, so on the invocation,
1646  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
1647  *
1648  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
1649  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
1650  *
1651  * FUSE will copy both struct a and the pointed buffer from the
1652  * process doing the ioctl and retry ioctl with both struct a and the
1653  * buffer.
1654  *
1655  * This time, FUSE server has everything it needs and completes ioctl
1656  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
1657  *
1658  * Copying data out works the same way.
1659  *
1660  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
1661  * automatically initializes in and out iovs by decoding @cmd with
1662  * _IOC_* macros and the server is not allowed to request RETRY.  This
1663  * limits ioctl data transfers to well-formed ioctls and is the forced
1664  * behavior for all FUSE servers.
1665  */
1666 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
1667                    unsigned int flags)
1668 {
1669         struct fuse_file *ff = file->private_data;
1670         struct fuse_conn *fc = ff->fc;
1671         struct fuse_ioctl_in inarg = {
1672                 .fh = ff->fh,
1673                 .cmd = cmd,
1674                 .arg = arg,
1675                 .flags = flags
1676         };
1677         struct fuse_ioctl_out outarg;
1678         struct fuse_req *req = NULL;
1679         struct page **pages = NULL;
1680         struct page *iov_page = NULL;
1681         struct iovec *in_iov = NULL, *out_iov = NULL;
1682         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
1683         size_t in_size, out_size, transferred;
1684         int err;
1685
1686         /* assume all the iovs returned by client always fits in a page */
1687         BUILD_BUG_ON(sizeof(struct iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
1688
1689         err = -ENOMEM;
1690         pages = kzalloc(sizeof(pages[0]) * FUSE_MAX_PAGES_PER_REQ, GFP_KERNEL);
1691         iov_page = alloc_page(GFP_KERNEL);
1692         if (!pages || !iov_page)
1693                 goto out;
1694
1695         /*
1696          * If restricted, initialize IO parameters as encoded in @cmd.
1697          * RETRY from server is not allowed.
1698          */
1699         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
1700                 struct iovec *iov = page_address(iov_page);
1701
1702                 iov->iov_base = (void __user *)arg;
1703                 iov->iov_len = _IOC_SIZE(cmd);
1704
1705                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
1706                         in_iov = iov;
1707                         in_iovs = 1;
1708                 }
1709
1710                 if (_IOC_DIR(cmd) & _IOC_READ) {
1711                         out_iov = iov;
1712                         out_iovs = 1;
1713                 }
1714         }
1715
1716  retry:
1717         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
1718         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
1719
1720         /*
1721          * Out data can be used either for actual out data or iovs,
1722          * make sure there always is at least one page.
1723          */
1724         out_size = max_t(size_t, out_size, PAGE_SIZE);
1725         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
1726
1727         /* make sure there are enough buffer pages and init request with them */
1728         err = -ENOMEM;
1729         if (max_pages > FUSE_MAX_PAGES_PER_REQ)
1730                 goto out;
1731         while (num_pages < max_pages) {
1732                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
1733                 if (!pages[num_pages])
1734                         goto out;
1735                 num_pages++;
1736         }
1737
1738         req = fuse_get_req(fc);
1739         if (IS_ERR(req)) {
1740                 err = PTR_ERR(req);
1741                 req = NULL;
1742                 goto out;
1743         }
1744         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
1745         req->num_pages = num_pages;
1746
1747         /* okay, let's send it to the client */
1748         req->in.h.opcode = FUSE_IOCTL;
1749         req->in.h.nodeid = ff->nodeid;
1750         req->in.numargs = 1;
1751         req->in.args[0].size = sizeof(inarg);
1752         req->in.args[0].value = &inarg;
1753         if (in_size) {
1754                 req->in.numargs++;
1755                 req->in.args[1].size = in_size;
1756                 req->in.argpages = 1;
1757
1758                 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
1759                                            false);
1760                 if (err)
1761                         goto out;
1762         }
1763
1764         req->out.numargs = 2;
1765         req->out.args[0].size = sizeof(outarg);
1766         req->out.args[0].value = &outarg;
1767         req->out.args[1].size = out_size;
1768         req->out.argpages = 1;
1769         req->out.argvar = 1;
1770
1771         fuse_request_send(fc, req);
1772         err = req->out.h.error;
1773         transferred = req->out.args[1].size;
1774         fuse_put_request(fc, req);
1775         req = NULL;
1776         if (err)
1777                 goto out;
1778
1779         /* did it ask for retry? */
1780         if (outarg.flags & FUSE_IOCTL_RETRY) {
1781                 char *vaddr;
1782
1783                 /* no retry if in restricted mode */
1784                 err = -EIO;
1785                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
1786                         goto out;
1787
1788                 in_iovs = outarg.in_iovs;
1789                 out_iovs = outarg.out_iovs;
1790
1791                 /*
1792                  * Make sure things are in boundary, separate checks
1793                  * are to protect against overflow.
1794                  */
1795                 err = -ENOMEM;
1796                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
1797                     out_iovs > FUSE_IOCTL_MAX_IOV ||
1798                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
1799                         goto out;
1800
1801                 err = -EIO;
1802                 if ((in_iovs + out_iovs) * sizeof(struct iovec) != transferred)
1803                         goto out;
1804
1805                 /* okay, copy in iovs and retry */
1806                 vaddr = kmap_atomic(pages[0], KM_USER0);
1807                 memcpy(page_address(iov_page), vaddr, transferred);
1808                 kunmap_atomic(vaddr, KM_USER0);
1809
1810                 in_iov = page_address(iov_page);
1811                 out_iov = in_iov + in_iovs;
1812
1813                 goto retry;
1814         }
1815
1816         err = -EIO;
1817         if (transferred > inarg.out_size)
1818                 goto out;
1819
1820         err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
1821  out:
1822         if (req)
1823                 fuse_put_request(fc, req);
1824         if (iov_page)
1825                 __free_page(iov_page);
1826         while (num_pages)
1827                 __free_page(pages[--num_pages]);
1828         kfree(pages);
1829
1830         return err ? err : outarg.result;
1831 }
1832 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
1833
1834 static long fuse_file_ioctl_common(struct file *file, unsigned int cmd,
1835                                    unsigned long arg, unsigned int flags)
1836 {
1837         struct inode *inode = file->f_dentry->d_inode;
1838         struct fuse_conn *fc = get_fuse_conn(inode);
1839
1840         if (!fuse_allow_task(fc, current))
1841                 return -EACCES;
1842
1843         if (is_bad_inode(inode))
1844                 return -EIO;
1845
1846         return fuse_do_ioctl(file, cmd, arg, flags);
1847 }
1848
1849 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
1850                             unsigned long arg)
1851 {
1852         return fuse_file_ioctl_common(file, cmd, arg, 0);
1853 }
1854
1855 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
1856                                    unsigned long arg)
1857 {
1858         return fuse_file_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
1859 }
1860
1861 /*
1862  * All files which have been polled are linked to RB tree
1863  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
1864  * find the matching one.
1865  */
1866 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
1867                                               struct rb_node **parent_out)
1868 {
1869         struct rb_node **link = &fc->polled_files.rb_node;
1870         struct rb_node *last = NULL;
1871
1872         while (*link) {
1873                 struct fuse_file *ff;
1874
1875                 last = *link;
1876                 ff = rb_entry(last, struct fuse_file, polled_node);
1877
1878                 if (kh < ff->kh)
1879                         link = &last->rb_left;
1880                 else if (kh > ff->kh)
1881                         link = &last->rb_right;
1882                 else
1883                         return link;
1884         }
1885
1886         if (parent_out)
1887                 *parent_out = last;
1888         return link;
1889 }
1890
1891 /*
1892  * The file is about to be polled.  Make sure it's on the polled_files
1893  * RB tree.  Note that files once added to the polled_files tree are
1894  * not removed before the file is released.  This is because a file
1895  * polled once is likely to be polled again.
1896  */
1897 static void fuse_register_polled_file(struct fuse_conn *fc,
1898                                       struct fuse_file *ff)
1899 {
1900         spin_lock(&fc->lock);
1901         if (RB_EMPTY_NODE(&ff->polled_node)) {
1902                 struct rb_node **link, *parent;
1903
1904                 link = fuse_find_polled_node(fc, ff->kh, &parent);
1905                 BUG_ON(*link);
1906                 rb_link_node(&ff->polled_node, parent, link);
1907                 rb_insert_color(&ff->polled_node, &fc->polled_files);
1908         }
1909         spin_unlock(&fc->lock);
1910 }
1911
1912 unsigned fuse_file_poll(struct file *file, poll_table *wait)
1913 {
1914         struct fuse_file *ff = file->private_data;
1915         struct fuse_conn *fc = ff->fc;
1916         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
1917         struct fuse_poll_out outarg;
1918         struct fuse_req *req;
1919         int err;
1920
1921         if (fc->no_poll)
1922                 return DEFAULT_POLLMASK;
1923
1924         poll_wait(file, &ff->poll_wait, wait);
1925
1926         /*
1927          * Ask for notification iff there's someone waiting for it.
1928          * The client may ignore the flag and always notify.
1929          */
1930         if (waitqueue_active(&ff->poll_wait)) {
1931                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
1932                 fuse_register_polled_file(fc, ff);
1933         }
1934
1935         req = fuse_get_req(fc);
1936         if (IS_ERR(req))
1937                 return POLLERR;
1938
1939         req->in.h.opcode = FUSE_POLL;
1940         req->in.h.nodeid = ff->nodeid;
1941         req->in.numargs = 1;
1942         req->in.args[0].size = sizeof(inarg);
1943         req->in.args[0].value = &inarg;
1944         req->out.numargs = 1;
1945         req->out.args[0].size = sizeof(outarg);
1946         req->out.args[0].value = &outarg;
1947         fuse_request_send(fc, req);
1948         err = req->out.h.error;
1949         fuse_put_request(fc, req);
1950
1951         if (!err)
1952                 return outarg.revents;
1953         if (err == -ENOSYS) {
1954                 fc->no_poll = 1;
1955                 return DEFAULT_POLLMASK;
1956         }
1957         return POLLERR;
1958 }
1959 EXPORT_SYMBOL_GPL(fuse_file_poll);
1960
1961 /*
1962  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
1963  * wakes up the poll waiters.
1964  */
1965 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
1966                             struct fuse_notify_poll_wakeup_out *outarg)
1967 {
1968         u64 kh = outarg->kh;
1969         struct rb_node **link;
1970
1971         spin_lock(&fc->lock);
1972
1973         link = fuse_find_polled_node(fc, kh, NULL);
1974         if (*link) {
1975                 struct fuse_file *ff;
1976
1977                 ff = rb_entry(*link, struct fuse_file, polled_node);
1978                 wake_up_interruptible_sync(&ff->poll_wait);
1979         }
1980
1981         spin_unlock(&fc->lock);
1982         return 0;
1983 }
1984
1985 static const struct file_operations fuse_file_operations = {
1986         .llseek         = fuse_file_llseek,
1987         .read           = do_sync_read,
1988         .aio_read       = fuse_file_aio_read,
1989         .write          = do_sync_write,
1990         .aio_write      = fuse_file_aio_write,
1991         .mmap           = fuse_file_mmap,
1992         .open           = fuse_open,
1993         .flush          = fuse_flush,
1994         .release        = fuse_release,
1995         .fsync          = fuse_fsync,
1996         .lock           = fuse_file_lock,
1997         .flock          = fuse_file_flock,
1998         .splice_read    = generic_file_splice_read,
1999         .unlocked_ioctl = fuse_file_ioctl,
2000         .compat_ioctl   = fuse_file_compat_ioctl,
2001         .poll           = fuse_file_poll,
2002 };
2003
2004 static const struct file_operations fuse_direct_io_file_operations = {
2005         .llseek         = fuse_file_llseek,
2006         .read           = fuse_direct_read,
2007         .write          = fuse_direct_write,
2008         .mmap           = fuse_direct_mmap,
2009         .open           = fuse_open,
2010         .flush          = fuse_flush,
2011         .release        = fuse_release,
2012         .fsync          = fuse_fsync,
2013         .lock           = fuse_file_lock,
2014         .flock          = fuse_file_flock,
2015         .unlocked_ioctl = fuse_file_ioctl,
2016         .compat_ioctl   = fuse_file_compat_ioctl,
2017         .poll           = fuse_file_poll,
2018         /* no splice_read */
2019 };
2020
2021 static const struct address_space_operations fuse_file_aops  = {
2022         .readpage       = fuse_readpage,
2023         .writepage      = fuse_writepage,
2024         .launder_page   = fuse_launder_page,
2025         .write_begin    = fuse_write_begin,
2026         .write_end      = fuse_write_end,
2027         .readpages      = fuse_readpages,
2028         .set_page_dirty = __set_page_dirty_nobuffers,
2029         .bmap           = fuse_bmap,
2030 };
2031
2032 void fuse_init_file_inode(struct inode *inode)
2033 {
2034         inode->i_fop = &fuse_file_operations;
2035         inode->i_data.a_ops = &fuse_file_aops;
2036 }