Merge branch 'for-linus' of git://git390.marist.edu/pub/scm/linux-2.6
[linux-2.6] / 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, struct dentry *de, int datasync,
355                       int isdir)
356 {
357         struct inode *inode = de->d_inode;
358         struct fuse_conn *fc = get_fuse_conn(inode);
359         struct fuse_file *ff = file->private_data;
360         struct fuse_req *req;
361         struct fuse_fsync_in inarg;
362         int err;
363
364         if (is_bad_inode(inode))
365                 return -EIO;
366
367         if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
368                 return 0;
369
370         /*
371          * Start writeback against all dirty pages of the inode, then
372          * wait for all outstanding writes, before sending the FSYNC
373          * request.
374          */
375         err = write_inode_now(inode, 0);
376         if (err)
377                 return err;
378
379         fuse_sync_writes(inode);
380
381         req = fuse_get_req(fc);
382         if (IS_ERR(req))
383                 return PTR_ERR(req);
384
385         memset(&inarg, 0, sizeof(inarg));
386         inarg.fh = ff->fh;
387         inarg.fsync_flags = datasync ? 1 : 0;
388         req->in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
389         req->in.h.nodeid = get_node_id(inode);
390         req->in.numargs = 1;
391         req->in.args[0].size = sizeof(inarg);
392         req->in.args[0].value = &inarg;
393         fuse_request_send(fc, req);
394         err = req->out.h.error;
395         fuse_put_request(fc, req);
396         if (err == -ENOSYS) {
397                 if (isdir)
398                         fc->no_fsyncdir = 1;
399                 else
400                         fc->no_fsync = 1;
401                 err = 0;
402         }
403         return err;
404 }
405
406 static int fuse_fsync(struct file *file, struct dentry *de, int datasync)
407 {
408         return fuse_fsync_common(file, de, datasync, 0);
409 }
410
411 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
412                     size_t count, int opcode)
413 {
414         struct fuse_read_in *inarg = &req->misc.read.in;
415         struct fuse_file *ff = file->private_data;
416
417         inarg->fh = ff->fh;
418         inarg->offset = pos;
419         inarg->size = count;
420         inarg->flags = file->f_flags;
421         req->in.h.opcode = opcode;
422         req->in.h.nodeid = ff->nodeid;
423         req->in.numargs = 1;
424         req->in.args[0].size = sizeof(struct fuse_read_in);
425         req->in.args[0].value = inarg;
426         req->out.argvar = 1;
427         req->out.numargs = 1;
428         req->out.args[0].size = count;
429 }
430
431 static size_t fuse_send_read(struct fuse_req *req, struct file *file,
432                              loff_t pos, size_t count, fl_owner_t owner)
433 {
434         struct fuse_file *ff = file->private_data;
435         struct fuse_conn *fc = ff->fc;
436
437         fuse_read_fill(req, file, pos, count, FUSE_READ);
438         if (owner != NULL) {
439                 struct fuse_read_in *inarg = &req->misc.read.in;
440
441                 inarg->read_flags |= FUSE_READ_LOCKOWNER;
442                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
443         }
444         fuse_request_send(fc, req);
445         return req->out.args[0].size;
446 }
447
448 static void fuse_read_update_size(struct inode *inode, loff_t size,
449                                   u64 attr_ver)
450 {
451         struct fuse_conn *fc = get_fuse_conn(inode);
452         struct fuse_inode *fi = get_fuse_inode(inode);
453
454         spin_lock(&fc->lock);
455         if (attr_ver == fi->attr_version && size < inode->i_size) {
456                 fi->attr_version = ++fc->attr_version;
457                 i_size_write(inode, size);
458         }
459         spin_unlock(&fc->lock);
460 }
461
462 static int fuse_readpage(struct file *file, struct page *page)
463 {
464         struct inode *inode = page->mapping->host;
465         struct fuse_conn *fc = get_fuse_conn(inode);
466         struct fuse_req *req;
467         size_t num_read;
468         loff_t pos = page_offset(page);
469         size_t count = PAGE_CACHE_SIZE;
470         u64 attr_ver;
471         int err;
472
473         err = -EIO;
474         if (is_bad_inode(inode))
475                 goto out;
476
477         /*
478          * Page writeback can extend beyond the liftime of the
479          * page-cache page, so make sure we read a properly synced
480          * page.
481          */
482         fuse_wait_on_page_writeback(inode, page->index);
483
484         req = fuse_get_req(fc);
485         err = PTR_ERR(req);
486         if (IS_ERR(req))
487                 goto out;
488
489         attr_ver = fuse_get_attr_version(fc);
490
491         req->out.page_zeroing = 1;
492         req->out.argpages = 1;
493         req->num_pages = 1;
494         req->pages[0] = page;
495         num_read = fuse_send_read(req, file, pos, count, NULL);
496         err = req->out.h.error;
497         fuse_put_request(fc, req);
498
499         if (!err) {
500                 /*
501                  * Short read means EOF.  If file size is larger, truncate it
502                  */
503                 if (num_read < count)
504                         fuse_read_update_size(inode, pos + num_read, attr_ver);
505
506                 SetPageUptodate(page);
507         }
508
509         fuse_invalidate_attr(inode); /* atime changed */
510  out:
511         unlock_page(page);
512         return err;
513 }
514
515 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
516 {
517         int i;
518         size_t count = req->misc.read.in.size;
519         size_t num_read = req->out.args[0].size;
520         struct inode *inode = req->pages[0]->mapping->host;
521
522         /*
523          * Short read means EOF.  If file size is larger, truncate it
524          */
525         if (!req->out.h.error && num_read < count) {
526                 loff_t pos = page_offset(req->pages[0]) + num_read;
527                 fuse_read_update_size(inode, pos, req->misc.read.attr_ver);
528         }
529
530         fuse_invalidate_attr(inode); /* atime changed */
531
532         for (i = 0; i < req->num_pages; i++) {
533                 struct page *page = req->pages[i];
534                 if (!req->out.h.error)
535                         SetPageUptodate(page);
536                 else
537                         SetPageError(page);
538                 unlock_page(page);
539         }
540         if (req->ff)
541                 fuse_file_put(req->ff);
542 }
543
544 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
545 {
546         struct fuse_file *ff = file->private_data;
547         struct fuse_conn *fc = ff->fc;
548         loff_t pos = page_offset(req->pages[0]);
549         size_t count = req->num_pages << PAGE_CACHE_SHIFT;
550
551         req->out.argpages = 1;
552         req->out.page_zeroing = 1;
553         fuse_read_fill(req, file, pos, count, FUSE_READ);
554         req->misc.read.attr_ver = fuse_get_attr_version(fc);
555         if (fc->async_read) {
556                 req->ff = fuse_file_get(ff);
557                 req->end = fuse_readpages_end;
558                 fuse_request_send_background(fc, req);
559         } else {
560                 fuse_request_send(fc, req);
561                 fuse_readpages_end(fc, req);
562                 fuse_put_request(fc, req);
563         }
564 }
565
566 struct fuse_fill_data {
567         struct fuse_req *req;
568         struct file *file;
569         struct inode *inode;
570 };
571
572 static int fuse_readpages_fill(void *_data, struct page *page)
573 {
574         struct fuse_fill_data *data = _data;
575         struct fuse_req *req = data->req;
576         struct inode *inode = data->inode;
577         struct fuse_conn *fc = get_fuse_conn(inode);
578
579         fuse_wait_on_page_writeback(inode, page->index);
580
581         if (req->num_pages &&
582             (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
583              (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
584              req->pages[req->num_pages - 1]->index + 1 != page->index)) {
585                 fuse_send_readpages(req, data->file);
586                 data->req = req = fuse_get_req(fc);
587                 if (IS_ERR(req)) {
588                         unlock_page(page);
589                         return PTR_ERR(req);
590                 }
591         }
592         req->pages[req->num_pages] = page;
593         req->num_pages++;
594         return 0;
595 }
596
597 static int fuse_readpages(struct file *file, struct address_space *mapping,
598                           struct list_head *pages, unsigned nr_pages)
599 {
600         struct inode *inode = mapping->host;
601         struct fuse_conn *fc = get_fuse_conn(inode);
602         struct fuse_fill_data data;
603         int err;
604
605         err = -EIO;
606         if (is_bad_inode(inode))
607                 goto out;
608
609         data.file = file;
610         data.inode = inode;
611         data.req = fuse_get_req(fc);
612         err = PTR_ERR(data.req);
613         if (IS_ERR(data.req))
614                 goto out;
615
616         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
617         if (!err) {
618                 if (data.req->num_pages)
619                         fuse_send_readpages(data.req, file);
620                 else
621                         fuse_put_request(fc, data.req);
622         }
623 out:
624         return err;
625 }
626
627 static ssize_t fuse_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
628                                   unsigned long nr_segs, loff_t pos)
629 {
630         struct inode *inode = iocb->ki_filp->f_mapping->host;
631
632         if (pos + iov_length(iov, nr_segs) > i_size_read(inode)) {
633                 int err;
634                 /*
635                  * If trying to read past EOF, make sure the i_size
636                  * attribute is up-to-date.
637                  */
638                 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
639                 if (err)
640                         return err;
641         }
642
643         return generic_file_aio_read(iocb, iov, nr_segs, pos);
644 }
645
646 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
647                             loff_t pos, size_t count)
648 {
649         struct fuse_write_in *inarg = &req->misc.write.in;
650         struct fuse_write_out *outarg = &req->misc.write.out;
651
652         inarg->fh = ff->fh;
653         inarg->offset = pos;
654         inarg->size = count;
655         req->in.h.opcode = FUSE_WRITE;
656         req->in.h.nodeid = ff->nodeid;
657         req->in.numargs = 2;
658         if (ff->fc->minor < 9)
659                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
660         else
661                 req->in.args[0].size = sizeof(struct fuse_write_in);
662         req->in.args[0].value = inarg;
663         req->in.args[1].size = count;
664         req->out.numargs = 1;
665         req->out.args[0].size = sizeof(struct fuse_write_out);
666         req->out.args[0].value = outarg;
667 }
668
669 static size_t fuse_send_write(struct fuse_req *req, struct file *file,
670                               loff_t pos, size_t count, fl_owner_t owner)
671 {
672         struct fuse_file *ff = file->private_data;
673         struct fuse_conn *fc = ff->fc;
674         struct fuse_write_in *inarg = &req->misc.write.in;
675
676         fuse_write_fill(req, ff, pos, count);
677         inarg->flags = file->f_flags;
678         if (owner != NULL) {
679                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
680                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
681         }
682         fuse_request_send(fc, req);
683         return req->misc.write.out.size;
684 }
685
686 static int fuse_write_begin(struct file *file, struct address_space *mapping,
687                         loff_t pos, unsigned len, unsigned flags,
688                         struct page **pagep, void **fsdata)
689 {
690         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
691
692         *pagep = grab_cache_page_write_begin(mapping, index, flags);
693         if (!*pagep)
694                 return -ENOMEM;
695         return 0;
696 }
697
698 static void fuse_write_update_size(struct inode *inode, loff_t pos)
699 {
700         struct fuse_conn *fc = get_fuse_conn(inode);
701         struct fuse_inode *fi = get_fuse_inode(inode);
702
703         spin_lock(&fc->lock);
704         fi->attr_version = ++fc->attr_version;
705         if (pos > inode->i_size)
706                 i_size_write(inode, pos);
707         spin_unlock(&fc->lock);
708 }
709
710 static int fuse_buffered_write(struct file *file, struct inode *inode,
711                                loff_t pos, unsigned count, struct page *page)
712 {
713         int err;
714         size_t nres;
715         struct fuse_conn *fc = get_fuse_conn(inode);
716         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
717         struct fuse_req *req;
718
719         if (is_bad_inode(inode))
720                 return -EIO;
721
722         /*
723          * Make sure writepages on the same page are not mixed up with
724          * plain writes.
725          */
726         fuse_wait_on_page_writeback(inode, page->index);
727
728         req = fuse_get_req(fc);
729         if (IS_ERR(req))
730                 return PTR_ERR(req);
731
732         req->in.argpages = 1;
733         req->num_pages = 1;
734         req->pages[0] = page;
735         req->page_offset = offset;
736         nres = fuse_send_write(req, file, pos, count, NULL);
737         err = req->out.h.error;
738         fuse_put_request(fc, req);
739         if (!err && !nres)
740                 err = -EIO;
741         if (!err) {
742                 pos += nres;
743                 fuse_write_update_size(inode, pos);
744                 if (count == PAGE_CACHE_SIZE)
745                         SetPageUptodate(page);
746         }
747         fuse_invalidate_attr(inode);
748         return err ? err : nres;
749 }
750
751 static int fuse_write_end(struct file *file, struct address_space *mapping,
752                         loff_t pos, unsigned len, unsigned copied,
753                         struct page *page, void *fsdata)
754 {
755         struct inode *inode = mapping->host;
756         int res = 0;
757
758         if (copied)
759                 res = fuse_buffered_write(file, inode, pos, copied, page);
760
761         unlock_page(page);
762         page_cache_release(page);
763         return res;
764 }
765
766 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
767                                     struct inode *inode, loff_t pos,
768                                     size_t count)
769 {
770         size_t res;
771         unsigned offset;
772         unsigned i;
773
774         for (i = 0; i < req->num_pages; i++)
775                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
776
777         res = fuse_send_write(req, file, pos, count, NULL);
778
779         offset = req->page_offset;
780         count = res;
781         for (i = 0; i < req->num_pages; i++) {
782                 struct page *page = req->pages[i];
783
784                 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
785                         SetPageUptodate(page);
786
787                 if (count > PAGE_CACHE_SIZE - offset)
788                         count -= PAGE_CACHE_SIZE - offset;
789                 else
790                         count = 0;
791                 offset = 0;
792
793                 unlock_page(page);
794                 page_cache_release(page);
795         }
796
797         return res;
798 }
799
800 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
801                                struct address_space *mapping,
802                                struct iov_iter *ii, loff_t pos)
803 {
804         struct fuse_conn *fc = get_fuse_conn(mapping->host);
805         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
806         size_t count = 0;
807         int err;
808
809         req->in.argpages = 1;
810         req->page_offset = offset;
811
812         do {
813                 size_t tmp;
814                 struct page *page;
815                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
816                 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
817                                      iov_iter_count(ii));
818
819                 bytes = min_t(size_t, bytes, fc->max_write - count);
820
821  again:
822                 err = -EFAULT;
823                 if (iov_iter_fault_in_readable(ii, bytes))
824                         break;
825
826                 err = -ENOMEM;
827                 page = grab_cache_page_write_begin(mapping, index, 0);
828                 if (!page)
829                         break;
830
831                 pagefault_disable();
832                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
833                 pagefault_enable();
834                 flush_dcache_page(page);
835
836                 if (!tmp) {
837                         unlock_page(page);
838                         page_cache_release(page);
839                         bytes = min(bytes, iov_iter_single_seg_count(ii));
840                         goto again;
841                 }
842
843                 err = 0;
844                 req->pages[req->num_pages] = page;
845                 req->num_pages++;
846
847                 iov_iter_advance(ii, tmp);
848                 count += tmp;
849                 pos += tmp;
850                 offset += tmp;
851                 if (offset == PAGE_CACHE_SIZE)
852                         offset = 0;
853
854                 if (!fc->big_writes)
855                         break;
856         } while (iov_iter_count(ii) && count < fc->max_write &&
857                  req->num_pages < FUSE_MAX_PAGES_PER_REQ && offset == 0);
858
859         return count > 0 ? count : err;
860 }
861
862 static ssize_t fuse_perform_write(struct file *file,
863                                   struct address_space *mapping,
864                                   struct iov_iter *ii, loff_t pos)
865 {
866         struct inode *inode = mapping->host;
867         struct fuse_conn *fc = get_fuse_conn(inode);
868         int err = 0;
869         ssize_t res = 0;
870
871         if (is_bad_inode(inode))
872                 return -EIO;
873
874         do {
875                 struct fuse_req *req;
876                 ssize_t count;
877
878                 req = fuse_get_req(fc);
879                 if (IS_ERR(req)) {
880                         err = PTR_ERR(req);
881                         break;
882                 }
883
884                 count = fuse_fill_write_pages(req, mapping, ii, pos);
885                 if (count <= 0) {
886                         err = count;
887                 } else {
888                         size_t num_written;
889
890                         num_written = fuse_send_write_pages(req, file, inode,
891                                                             pos, count);
892                         err = req->out.h.error;
893                         if (!err) {
894                                 res += num_written;
895                                 pos += num_written;
896
897                                 /* break out of the loop on short write */
898                                 if (num_written != count)
899                                         err = -EIO;
900                         }
901                 }
902                 fuse_put_request(fc, req);
903         } while (!err && iov_iter_count(ii));
904
905         if (res > 0)
906                 fuse_write_update_size(inode, pos);
907
908         fuse_invalidate_attr(inode);
909
910         return res > 0 ? res : err;
911 }
912
913 static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
914                                    unsigned long nr_segs, loff_t pos)
915 {
916         struct file *file = iocb->ki_filp;
917         struct address_space *mapping = file->f_mapping;
918         size_t count = 0;
919         ssize_t written = 0;
920         struct inode *inode = mapping->host;
921         ssize_t err;
922         struct iov_iter i;
923
924         WARN_ON(iocb->ki_pos != pos);
925
926         err = generic_segment_checks(iov, &nr_segs, &count, VERIFY_READ);
927         if (err)
928                 return err;
929
930         mutex_lock(&inode->i_mutex);
931         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
932
933         /* We can write back this queue in page reclaim */
934         current->backing_dev_info = mapping->backing_dev_info;
935
936         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
937         if (err)
938                 goto out;
939
940         if (count == 0)
941                 goto out;
942
943         err = file_remove_suid(file);
944         if (err)
945                 goto out;
946
947         file_update_time(file);
948
949         iov_iter_init(&i, iov, nr_segs, count, 0);
950         written = fuse_perform_write(file, mapping, &i, pos);
951         if (written >= 0)
952                 iocb->ki_pos = pos + written;
953
954 out:
955         current->backing_dev_info = NULL;
956         mutex_unlock(&inode->i_mutex);
957
958         return written ? written : err;
959 }
960
961 static void fuse_release_user_pages(struct fuse_req *req, int write)
962 {
963         unsigned i;
964
965         for (i = 0; i < req->num_pages; i++) {
966                 struct page *page = req->pages[i];
967                 if (write)
968                         set_page_dirty_lock(page);
969                 put_page(page);
970         }
971 }
972
973 static int fuse_get_user_pages(struct fuse_req *req, const char __user *buf,
974                                size_t *nbytesp, int write)
975 {
976         size_t nbytes = *nbytesp;
977         unsigned long user_addr = (unsigned long) buf;
978         unsigned offset = user_addr & ~PAGE_MASK;
979         int npages;
980
981         /* Special case for kernel I/O: can copy directly into the buffer */
982         if (segment_eq(get_fs(), KERNEL_DS)) {
983                 if (write)
984                         req->in.args[1].value = (void *) user_addr;
985                 else
986                         req->out.args[0].value = (void *) user_addr;
987
988                 return 0;
989         }
990
991         nbytes = min_t(size_t, nbytes, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
992         npages = (nbytes + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
993         npages = clamp(npages, 1, FUSE_MAX_PAGES_PER_REQ);
994         down_read(&current->mm->mmap_sem);
995         npages = get_user_pages(current, current->mm, user_addr, npages, !write,
996                                 0, req->pages, NULL);
997         up_read(&current->mm->mmap_sem);
998         if (npages < 0)
999                 return npages;
1000
1001         req->num_pages = npages;
1002         req->page_offset = offset;
1003
1004         if (write)
1005                 req->in.argpages = 1;
1006         else
1007                 req->out.argpages = 1;
1008
1009         nbytes = (req->num_pages << PAGE_SHIFT) - req->page_offset;
1010         *nbytesp = min(*nbytesp, nbytes);
1011
1012         return 0;
1013 }
1014
1015 ssize_t fuse_direct_io(struct file *file, const char __user *buf,
1016                        size_t count, loff_t *ppos, int write)
1017 {
1018         struct fuse_file *ff = file->private_data;
1019         struct fuse_conn *fc = ff->fc;
1020         size_t nmax = write ? fc->max_write : fc->max_read;
1021         loff_t pos = *ppos;
1022         ssize_t res = 0;
1023         struct fuse_req *req;
1024
1025         req = fuse_get_req(fc);
1026         if (IS_ERR(req))
1027                 return PTR_ERR(req);
1028
1029         while (count) {
1030                 size_t nres;
1031                 fl_owner_t owner = current->files;
1032                 size_t nbytes = min(count, nmax);
1033                 int err = fuse_get_user_pages(req, buf, &nbytes, write);
1034                 if (err) {
1035                         res = err;
1036                         break;
1037                 }
1038
1039                 if (write)
1040                         nres = fuse_send_write(req, file, pos, nbytes, owner);
1041                 else
1042                         nres = fuse_send_read(req, file, pos, nbytes, owner);
1043
1044                 fuse_release_user_pages(req, !write);
1045                 if (req->out.h.error) {
1046                         if (!res)
1047                                 res = req->out.h.error;
1048                         break;
1049                 } else if (nres > nbytes) {
1050                         res = -EIO;
1051                         break;
1052                 }
1053                 count -= nres;
1054                 res += nres;
1055                 pos += nres;
1056                 buf += nres;
1057                 if (nres != nbytes)
1058                         break;
1059                 if (count) {
1060                         fuse_put_request(fc, req);
1061                         req = fuse_get_req(fc);
1062                         if (IS_ERR(req))
1063                                 break;
1064                 }
1065         }
1066         fuse_put_request(fc, req);
1067         if (res > 0)
1068                 *ppos = pos;
1069
1070         return res;
1071 }
1072 EXPORT_SYMBOL_GPL(fuse_direct_io);
1073
1074 static ssize_t fuse_direct_read(struct file *file, char __user *buf,
1075                                      size_t count, loff_t *ppos)
1076 {
1077         ssize_t res;
1078         struct inode *inode = file->f_path.dentry->d_inode;
1079
1080         if (is_bad_inode(inode))
1081                 return -EIO;
1082
1083         res = fuse_direct_io(file, buf, count, ppos, 0);
1084
1085         fuse_invalidate_attr(inode);
1086
1087         return res;
1088 }
1089
1090 static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
1091                                  size_t count, loff_t *ppos)
1092 {
1093         struct inode *inode = file->f_path.dentry->d_inode;
1094         ssize_t res;
1095
1096         if (is_bad_inode(inode))
1097                 return -EIO;
1098
1099         /* Don't allow parallel writes to the same file */
1100         mutex_lock(&inode->i_mutex);
1101         res = generic_write_checks(file, ppos, &count, 0);
1102         if (!res) {
1103                 res = fuse_direct_io(file, buf, count, ppos, 1);
1104                 if (res > 0)
1105                         fuse_write_update_size(inode, *ppos);
1106         }
1107         mutex_unlock(&inode->i_mutex);
1108
1109         fuse_invalidate_attr(inode);
1110
1111         return res;
1112 }
1113
1114 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1115 {
1116         __free_page(req->pages[0]);
1117         fuse_file_put(req->ff);
1118 }
1119
1120 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1121 {
1122         struct inode *inode = req->inode;
1123         struct fuse_inode *fi = get_fuse_inode(inode);
1124         struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
1125
1126         list_del(&req->writepages_entry);
1127         dec_bdi_stat(bdi, BDI_WRITEBACK);
1128         dec_zone_page_state(req->pages[0], NR_WRITEBACK_TEMP);
1129         bdi_writeout_inc(bdi);
1130         wake_up(&fi->page_waitq);
1131 }
1132
1133 /* Called under fc->lock, may release and reacquire it */
1134 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req)
1135 __releases(&fc->lock)
1136 __acquires(&fc->lock)
1137 {
1138         struct fuse_inode *fi = get_fuse_inode(req->inode);
1139         loff_t size = i_size_read(req->inode);
1140         struct fuse_write_in *inarg = &req->misc.write.in;
1141
1142         if (!fc->connected)
1143                 goto out_free;
1144
1145         if (inarg->offset + PAGE_CACHE_SIZE <= size) {
1146                 inarg->size = PAGE_CACHE_SIZE;
1147         } else if (inarg->offset < size) {
1148                 inarg->size = size & (PAGE_CACHE_SIZE - 1);
1149         } else {
1150                 /* Got truncated off completely */
1151                 goto out_free;
1152         }
1153
1154         req->in.args[1].size = inarg->size;
1155         fi->writectr++;
1156         fuse_request_send_background_locked(fc, req);
1157         return;
1158
1159  out_free:
1160         fuse_writepage_finish(fc, req);
1161         spin_unlock(&fc->lock);
1162         fuse_writepage_free(fc, req);
1163         fuse_put_request(fc, req);
1164         spin_lock(&fc->lock);
1165 }
1166
1167 /*
1168  * If fi->writectr is positive (no truncate or fsync going on) send
1169  * all queued writepage requests.
1170  *
1171  * Called with fc->lock
1172  */
1173 void fuse_flush_writepages(struct inode *inode)
1174 __releases(&fc->lock)
1175 __acquires(&fc->lock)
1176 {
1177         struct fuse_conn *fc = get_fuse_conn(inode);
1178         struct fuse_inode *fi = get_fuse_inode(inode);
1179         struct fuse_req *req;
1180
1181         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1182                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1183                 list_del_init(&req->list);
1184                 fuse_send_writepage(fc, req);
1185         }
1186 }
1187
1188 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1189 {
1190         struct inode *inode = req->inode;
1191         struct fuse_inode *fi = get_fuse_inode(inode);
1192
1193         mapping_set_error(inode->i_mapping, req->out.h.error);
1194         spin_lock(&fc->lock);
1195         fi->writectr--;
1196         fuse_writepage_finish(fc, req);
1197         spin_unlock(&fc->lock);
1198         fuse_writepage_free(fc, req);
1199 }
1200
1201 static int fuse_writepage_locked(struct page *page)
1202 {
1203         struct address_space *mapping = page->mapping;
1204         struct inode *inode = mapping->host;
1205         struct fuse_conn *fc = get_fuse_conn(inode);
1206         struct fuse_inode *fi = get_fuse_inode(inode);
1207         struct fuse_req *req;
1208         struct fuse_file *ff;
1209         struct page *tmp_page;
1210
1211         set_page_writeback(page);
1212
1213         req = fuse_request_alloc_nofs();
1214         if (!req)
1215                 goto err;
1216
1217         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1218         if (!tmp_page)
1219                 goto err_free;
1220
1221         spin_lock(&fc->lock);
1222         BUG_ON(list_empty(&fi->write_files));
1223         ff = list_entry(fi->write_files.next, struct fuse_file, write_entry);
1224         req->ff = fuse_file_get(ff);
1225         spin_unlock(&fc->lock);
1226
1227         fuse_write_fill(req, ff, page_offset(page), 0);
1228
1229         copy_highpage(tmp_page, page);
1230         req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1231         req->in.argpages = 1;
1232         req->num_pages = 1;
1233         req->pages[0] = tmp_page;
1234         req->page_offset = 0;
1235         req->end = fuse_writepage_end;
1236         req->inode = inode;
1237
1238         inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
1239         inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1240         end_page_writeback(page);
1241
1242         spin_lock(&fc->lock);
1243         list_add(&req->writepages_entry, &fi->writepages);
1244         list_add_tail(&req->list, &fi->queued_writes);
1245         fuse_flush_writepages(inode);
1246         spin_unlock(&fc->lock);
1247
1248         return 0;
1249
1250 err_free:
1251         fuse_request_free(req);
1252 err:
1253         end_page_writeback(page);
1254         return -ENOMEM;
1255 }
1256
1257 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1258 {
1259         int err;
1260
1261         err = fuse_writepage_locked(page);
1262         unlock_page(page);
1263
1264         return err;
1265 }
1266
1267 static int fuse_launder_page(struct page *page)
1268 {
1269         int err = 0;
1270         if (clear_page_dirty_for_io(page)) {
1271                 struct inode *inode = page->mapping->host;
1272                 err = fuse_writepage_locked(page);
1273                 if (!err)
1274                         fuse_wait_on_page_writeback(inode, page->index);
1275         }
1276         return err;
1277 }
1278
1279 /*
1280  * Write back dirty pages now, because there may not be any suitable
1281  * open files later
1282  */
1283 static void fuse_vma_close(struct vm_area_struct *vma)
1284 {
1285         filemap_write_and_wait(vma->vm_file->f_mapping);
1286 }
1287
1288 /*
1289  * Wait for writeback against this page to complete before allowing it
1290  * to be marked dirty again, and hence written back again, possibly
1291  * before the previous writepage completed.
1292  *
1293  * Block here, instead of in ->writepage(), so that the userspace fs
1294  * can only block processes actually operating on the filesystem.
1295  *
1296  * Otherwise unprivileged userspace fs would be able to block
1297  * unrelated:
1298  *
1299  * - page migration
1300  * - sync(2)
1301  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
1302  */
1303 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1304 {
1305         struct page *page = vmf->page;
1306         /*
1307          * Don't use page->mapping as it may become NULL from a
1308          * concurrent truncate.
1309          */
1310         struct inode *inode = vma->vm_file->f_mapping->host;
1311
1312         fuse_wait_on_page_writeback(inode, page->index);
1313         return 0;
1314 }
1315
1316 static struct vm_operations_struct fuse_file_vm_ops = {
1317         .close          = fuse_vma_close,
1318         .fault          = filemap_fault,
1319         .page_mkwrite   = fuse_page_mkwrite,
1320 };
1321
1322 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
1323 {
1324         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE)) {
1325                 struct inode *inode = file->f_dentry->d_inode;
1326                 struct fuse_conn *fc = get_fuse_conn(inode);
1327                 struct fuse_inode *fi = get_fuse_inode(inode);
1328                 struct fuse_file *ff = file->private_data;
1329                 /*
1330                  * file may be written through mmap, so chain it onto the
1331                  * inodes's write_file list
1332                  */
1333                 spin_lock(&fc->lock);
1334                 if (list_empty(&ff->write_entry))
1335                         list_add(&ff->write_entry, &fi->write_files);
1336                 spin_unlock(&fc->lock);
1337         }
1338         file_accessed(file);
1339         vma->vm_ops = &fuse_file_vm_ops;
1340         return 0;
1341 }
1342
1343 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
1344 {
1345         /* Can't provide the coherency needed for MAP_SHARED */
1346         if (vma->vm_flags & VM_MAYSHARE)
1347                 return -ENODEV;
1348
1349         invalidate_inode_pages2(file->f_mapping);
1350
1351         return generic_file_mmap(file, vma);
1352 }
1353
1354 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
1355                                   struct file_lock *fl)
1356 {
1357         switch (ffl->type) {
1358         case F_UNLCK:
1359                 break;
1360
1361         case F_RDLCK:
1362         case F_WRLCK:
1363                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
1364                     ffl->end < ffl->start)
1365                         return -EIO;
1366
1367                 fl->fl_start = ffl->start;
1368                 fl->fl_end = ffl->end;
1369                 fl->fl_pid = ffl->pid;
1370                 break;
1371
1372         default:
1373                 return -EIO;
1374         }
1375         fl->fl_type = ffl->type;
1376         return 0;
1377 }
1378
1379 static void fuse_lk_fill(struct fuse_req *req, struct file *file,
1380                          const struct file_lock *fl, int opcode, pid_t pid,
1381                          int flock)
1382 {
1383         struct inode *inode = file->f_path.dentry->d_inode;
1384         struct fuse_conn *fc = get_fuse_conn(inode);
1385         struct fuse_file *ff = file->private_data;
1386         struct fuse_lk_in *arg = &req->misc.lk_in;
1387
1388         arg->fh = ff->fh;
1389         arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
1390         arg->lk.start = fl->fl_start;
1391         arg->lk.end = fl->fl_end;
1392         arg->lk.type = fl->fl_type;
1393         arg->lk.pid = pid;
1394         if (flock)
1395                 arg->lk_flags |= FUSE_LK_FLOCK;
1396         req->in.h.opcode = opcode;
1397         req->in.h.nodeid = get_node_id(inode);
1398         req->in.numargs = 1;
1399         req->in.args[0].size = sizeof(*arg);
1400         req->in.args[0].value = arg;
1401 }
1402
1403 static int fuse_getlk(struct file *file, struct file_lock *fl)
1404 {
1405         struct inode *inode = file->f_path.dentry->d_inode;
1406         struct fuse_conn *fc = get_fuse_conn(inode);
1407         struct fuse_req *req;
1408         struct fuse_lk_out outarg;
1409         int err;
1410
1411         req = fuse_get_req(fc);
1412         if (IS_ERR(req))
1413                 return PTR_ERR(req);
1414
1415         fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
1416         req->out.numargs = 1;
1417         req->out.args[0].size = sizeof(outarg);
1418         req->out.args[0].value = &outarg;
1419         fuse_request_send(fc, req);
1420         err = req->out.h.error;
1421         fuse_put_request(fc, req);
1422         if (!err)
1423                 err = convert_fuse_file_lock(&outarg.lk, fl);
1424
1425         return err;
1426 }
1427
1428 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
1429 {
1430         struct inode *inode = file->f_path.dentry->d_inode;
1431         struct fuse_conn *fc = get_fuse_conn(inode);
1432         struct fuse_req *req;
1433         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
1434         pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
1435         int err;
1436
1437         if (fl->fl_lmops && fl->fl_lmops->fl_grant) {
1438                 /* NLM needs asynchronous locks, which we don't support yet */
1439                 return -ENOLCK;
1440         }
1441
1442         /* Unlock on close is handled by the flush method */
1443         if (fl->fl_flags & FL_CLOSE)
1444                 return 0;
1445
1446         req = fuse_get_req(fc);
1447         if (IS_ERR(req))
1448                 return PTR_ERR(req);
1449
1450         fuse_lk_fill(req, file, fl, opcode, pid, flock);
1451         fuse_request_send(fc, req);
1452         err = req->out.h.error;
1453         /* locking is restartable */
1454         if (err == -EINTR)
1455                 err = -ERESTARTSYS;
1456         fuse_put_request(fc, req);
1457         return err;
1458 }
1459
1460 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
1461 {
1462         struct inode *inode = file->f_path.dentry->d_inode;
1463         struct fuse_conn *fc = get_fuse_conn(inode);
1464         int err;
1465
1466         if (cmd == F_CANCELLK) {
1467                 err = 0;
1468         } else if (cmd == F_GETLK) {
1469                 if (fc->no_lock) {
1470                         posix_test_lock(file, fl);
1471                         err = 0;
1472                 } else
1473                         err = fuse_getlk(file, fl);
1474         } else {
1475                 if (fc->no_lock)
1476                         err = posix_lock_file(file, fl, NULL);
1477                 else
1478                         err = fuse_setlk(file, fl, 0);
1479         }
1480         return err;
1481 }
1482
1483 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
1484 {
1485         struct inode *inode = file->f_path.dentry->d_inode;
1486         struct fuse_conn *fc = get_fuse_conn(inode);
1487         int err;
1488
1489         if (fc->no_lock) {
1490                 err = flock_lock_file_wait(file, fl);
1491         } else {
1492                 /* emulate flock with POSIX locks */
1493                 fl->fl_owner = (fl_owner_t) file;
1494                 err = fuse_setlk(file, fl, 1);
1495         }
1496
1497         return err;
1498 }
1499
1500 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
1501 {
1502         struct inode *inode = mapping->host;
1503         struct fuse_conn *fc = get_fuse_conn(inode);
1504         struct fuse_req *req;
1505         struct fuse_bmap_in inarg;
1506         struct fuse_bmap_out outarg;
1507         int err;
1508
1509         if (!inode->i_sb->s_bdev || fc->no_bmap)
1510                 return 0;
1511
1512         req = fuse_get_req(fc);
1513         if (IS_ERR(req))
1514                 return 0;
1515
1516         memset(&inarg, 0, sizeof(inarg));
1517         inarg.block = block;
1518         inarg.blocksize = inode->i_sb->s_blocksize;
1519         req->in.h.opcode = FUSE_BMAP;
1520         req->in.h.nodeid = get_node_id(inode);
1521         req->in.numargs = 1;
1522         req->in.args[0].size = sizeof(inarg);
1523         req->in.args[0].value = &inarg;
1524         req->out.numargs = 1;
1525         req->out.args[0].size = sizeof(outarg);
1526         req->out.args[0].value = &outarg;
1527         fuse_request_send(fc, req);
1528         err = req->out.h.error;
1529         fuse_put_request(fc, req);
1530         if (err == -ENOSYS)
1531                 fc->no_bmap = 1;
1532
1533         return err ? 0 : outarg.block;
1534 }
1535
1536 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int origin)
1537 {
1538         loff_t retval;
1539         struct inode *inode = file->f_path.dentry->d_inode;
1540
1541         mutex_lock(&inode->i_mutex);
1542         switch (origin) {
1543         case SEEK_END:
1544                 retval = fuse_update_attributes(inode, NULL, file, NULL);
1545                 if (retval)
1546                         goto exit;
1547                 offset += i_size_read(inode);
1548                 break;
1549         case SEEK_CUR:
1550                 offset += file->f_pos;
1551         }
1552         retval = -EINVAL;
1553         if (offset >= 0 && offset <= inode->i_sb->s_maxbytes) {
1554                 if (offset != file->f_pos) {
1555                         file->f_pos = offset;
1556                         file->f_version = 0;
1557                 }
1558                 retval = offset;
1559         }
1560 exit:
1561         mutex_unlock(&inode->i_mutex);
1562         return retval;
1563 }
1564
1565 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
1566                         unsigned int nr_segs, size_t bytes, bool to_user)
1567 {
1568         struct iov_iter ii;
1569         int page_idx = 0;
1570
1571         if (!bytes)
1572                 return 0;
1573
1574         iov_iter_init(&ii, iov, nr_segs, bytes, 0);
1575
1576         while (iov_iter_count(&ii)) {
1577                 struct page *page = pages[page_idx++];
1578                 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
1579                 void *kaddr, *map;
1580
1581                 kaddr = map = kmap(page);
1582
1583                 while (todo) {
1584                         char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
1585                         size_t iov_len = ii.iov->iov_len - ii.iov_offset;
1586                         size_t copy = min(todo, iov_len);
1587                         size_t left;
1588
1589                         if (!to_user)
1590                                 left = copy_from_user(kaddr, uaddr, copy);
1591                         else
1592                                 left = copy_to_user(uaddr, kaddr, copy);
1593
1594                         if (unlikely(left))
1595                                 return -EFAULT;
1596
1597                         iov_iter_advance(&ii, copy);
1598                         todo -= copy;
1599                         kaddr += copy;
1600                 }
1601
1602                 kunmap(map);
1603         }
1604
1605         return 0;
1606 }
1607
1608 /*
1609  * For ioctls, there is no generic way to determine how much memory
1610  * needs to be read and/or written.  Furthermore, ioctls are allowed
1611  * to dereference the passed pointer, so the parameter requires deep
1612  * copying but FUSE has no idea whatsoever about what to copy in or
1613  * out.
1614  *
1615  * This is solved by allowing FUSE server to retry ioctl with
1616  * necessary in/out iovecs.  Let's assume the ioctl implementation
1617  * needs to read in the following structure.
1618  *
1619  * struct a {
1620  *      char    *buf;
1621  *      size_t  buflen;
1622  * }
1623  *
1624  * On the first callout to FUSE server, inarg->in_size and
1625  * inarg->out_size will be NULL; then, the server completes the ioctl
1626  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
1627  * the actual iov array to
1628  *
1629  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
1630  *
1631  * which tells FUSE to copy in the requested area and retry the ioctl.
1632  * On the second round, the server has access to the structure and
1633  * from that it can tell what to look for next, so on the invocation,
1634  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
1635  *
1636  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
1637  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
1638  *
1639  * FUSE will copy both struct a and the pointed buffer from the
1640  * process doing the ioctl and retry ioctl with both struct a and the
1641  * buffer.
1642  *
1643  * This time, FUSE server has everything it needs and completes ioctl
1644  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
1645  *
1646  * Copying data out works the same way.
1647  *
1648  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
1649  * automatically initializes in and out iovs by decoding @cmd with
1650  * _IOC_* macros and the server is not allowed to request RETRY.  This
1651  * limits ioctl data transfers to well-formed ioctls and is the forced
1652  * behavior for all FUSE servers.
1653  */
1654 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
1655                    unsigned int flags)
1656 {
1657         struct fuse_file *ff = file->private_data;
1658         struct fuse_conn *fc = ff->fc;
1659         struct fuse_ioctl_in inarg = {
1660                 .fh = ff->fh,
1661                 .cmd = cmd,
1662                 .arg = arg,
1663                 .flags = flags
1664         };
1665         struct fuse_ioctl_out outarg;
1666         struct fuse_req *req = NULL;
1667         struct page **pages = NULL;
1668         struct page *iov_page = NULL;
1669         struct iovec *in_iov = NULL, *out_iov = NULL;
1670         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
1671         size_t in_size, out_size, transferred;
1672         int err;
1673
1674         /* assume all the iovs returned by client always fits in a page */
1675         BUILD_BUG_ON(sizeof(struct iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
1676
1677         err = -ENOMEM;
1678         pages = kzalloc(sizeof(pages[0]) * FUSE_MAX_PAGES_PER_REQ, GFP_KERNEL);
1679         iov_page = alloc_page(GFP_KERNEL);
1680         if (!pages || !iov_page)
1681                 goto out;
1682
1683         /*
1684          * If restricted, initialize IO parameters as encoded in @cmd.
1685          * RETRY from server is not allowed.
1686          */
1687         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
1688                 struct iovec *iov = page_address(iov_page);
1689
1690                 iov->iov_base = (void __user *)arg;
1691                 iov->iov_len = _IOC_SIZE(cmd);
1692
1693                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
1694                         in_iov = iov;
1695                         in_iovs = 1;
1696                 }
1697
1698                 if (_IOC_DIR(cmd) & _IOC_READ) {
1699                         out_iov = iov;
1700                         out_iovs = 1;
1701                 }
1702         }
1703
1704  retry:
1705         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
1706         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
1707
1708         /*
1709          * Out data can be used either for actual out data or iovs,
1710          * make sure there always is at least one page.
1711          */
1712         out_size = max_t(size_t, out_size, PAGE_SIZE);
1713         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
1714
1715         /* make sure there are enough buffer pages and init request with them */
1716         err = -ENOMEM;
1717         if (max_pages > FUSE_MAX_PAGES_PER_REQ)
1718                 goto out;
1719         while (num_pages < max_pages) {
1720                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
1721                 if (!pages[num_pages])
1722                         goto out;
1723                 num_pages++;
1724         }
1725
1726         req = fuse_get_req(fc);
1727         if (IS_ERR(req)) {
1728                 err = PTR_ERR(req);
1729                 req = NULL;
1730                 goto out;
1731         }
1732         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
1733         req->num_pages = num_pages;
1734
1735         /* okay, let's send it to the client */
1736         req->in.h.opcode = FUSE_IOCTL;
1737         req->in.h.nodeid = ff->nodeid;
1738         req->in.numargs = 1;
1739         req->in.args[0].size = sizeof(inarg);
1740         req->in.args[0].value = &inarg;
1741         if (in_size) {
1742                 req->in.numargs++;
1743                 req->in.args[1].size = in_size;
1744                 req->in.argpages = 1;
1745
1746                 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
1747                                            false);
1748                 if (err)
1749                         goto out;
1750         }
1751
1752         req->out.numargs = 2;
1753         req->out.args[0].size = sizeof(outarg);
1754         req->out.args[0].value = &outarg;
1755         req->out.args[1].size = out_size;
1756         req->out.argpages = 1;
1757         req->out.argvar = 1;
1758
1759         fuse_request_send(fc, req);
1760         err = req->out.h.error;
1761         transferred = req->out.args[1].size;
1762         fuse_put_request(fc, req);
1763         req = NULL;
1764         if (err)
1765                 goto out;
1766
1767         /* did it ask for retry? */
1768         if (outarg.flags & FUSE_IOCTL_RETRY) {
1769                 char *vaddr;
1770
1771                 /* no retry if in restricted mode */
1772                 err = -EIO;
1773                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
1774                         goto out;
1775
1776                 in_iovs = outarg.in_iovs;
1777                 out_iovs = outarg.out_iovs;
1778
1779                 /*
1780                  * Make sure things are in boundary, separate checks
1781                  * are to protect against overflow.
1782                  */
1783                 err = -ENOMEM;
1784                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
1785                     out_iovs > FUSE_IOCTL_MAX_IOV ||
1786                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
1787                         goto out;
1788
1789                 err = -EIO;
1790                 if ((in_iovs + out_iovs) * sizeof(struct iovec) != transferred)
1791                         goto out;
1792
1793                 /* okay, copy in iovs and retry */
1794                 vaddr = kmap_atomic(pages[0], KM_USER0);
1795                 memcpy(page_address(iov_page), vaddr, transferred);
1796                 kunmap_atomic(vaddr, KM_USER0);
1797
1798                 in_iov = page_address(iov_page);
1799                 out_iov = in_iov + in_iovs;
1800
1801                 goto retry;
1802         }
1803
1804         err = -EIO;
1805         if (transferred > inarg.out_size)
1806                 goto out;
1807
1808         err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
1809  out:
1810         if (req)
1811                 fuse_put_request(fc, req);
1812         if (iov_page)
1813                 __free_page(iov_page);
1814         while (num_pages)
1815                 __free_page(pages[--num_pages]);
1816         kfree(pages);
1817
1818         return err ? err : outarg.result;
1819 }
1820 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
1821
1822 static long fuse_file_ioctl_common(struct file *file, unsigned int cmd,
1823                                    unsigned long arg, unsigned int flags)
1824 {
1825         struct inode *inode = file->f_dentry->d_inode;
1826         struct fuse_conn *fc = get_fuse_conn(inode);
1827
1828         if (!fuse_allow_task(fc, current))
1829                 return -EACCES;
1830
1831         if (is_bad_inode(inode))
1832                 return -EIO;
1833
1834         return fuse_do_ioctl(file, cmd, arg, flags);
1835 }
1836
1837 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
1838                             unsigned long arg)
1839 {
1840         return fuse_file_ioctl_common(file, cmd, arg, 0);
1841 }
1842
1843 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
1844                                    unsigned long arg)
1845 {
1846         return fuse_file_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
1847 }
1848
1849 /*
1850  * All files which have been polled are linked to RB tree
1851  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
1852  * find the matching one.
1853  */
1854 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
1855                                               struct rb_node **parent_out)
1856 {
1857         struct rb_node **link = &fc->polled_files.rb_node;
1858         struct rb_node *last = NULL;
1859
1860         while (*link) {
1861                 struct fuse_file *ff;
1862
1863                 last = *link;
1864                 ff = rb_entry(last, struct fuse_file, polled_node);
1865
1866                 if (kh < ff->kh)
1867                         link = &last->rb_left;
1868                 else if (kh > ff->kh)
1869                         link = &last->rb_right;
1870                 else
1871                         return link;
1872         }
1873
1874         if (parent_out)
1875                 *parent_out = last;
1876         return link;
1877 }
1878
1879 /*
1880  * The file is about to be polled.  Make sure it's on the polled_files
1881  * RB tree.  Note that files once added to the polled_files tree are
1882  * not removed before the file is released.  This is because a file
1883  * polled once is likely to be polled again.
1884  */
1885 static void fuse_register_polled_file(struct fuse_conn *fc,
1886                                       struct fuse_file *ff)
1887 {
1888         spin_lock(&fc->lock);
1889         if (RB_EMPTY_NODE(&ff->polled_node)) {
1890                 struct rb_node **link, *parent;
1891
1892                 link = fuse_find_polled_node(fc, ff->kh, &parent);
1893                 BUG_ON(*link);
1894                 rb_link_node(&ff->polled_node, parent, link);
1895                 rb_insert_color(&ff->polled_node, &fc->polled_files);
1896         }
1897         spin_unlock(&fc->lock);
1898 }
1899
1900 unsigned fuse_file_poll(struct file *file, poll_table *wait)
1901 {
1902         struct fuse_file *ff = file->private_data;
1903         struct fuse_conn *fc = ff->fc;
1904         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
1905         struct fuse_poll_out outarg;
1906         struct fuse_req *req;
1907         int err;
1908
1909         if (fc->no_poll)
1910                 return DEFAULT_POLLMASK;
1911
1912         poll_wait(file, &ff->poll_wait, wait);
1913
1914         /*
1915          * Ask for notification iff there's someone waiting for it.
1916          * The client may ignore the flag and always notify.
1917          */
1918         if (waitqueue_active(&ff->poll_wait)) {
1919                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
1920                 fuse_register_polled_file(fc, ff);
1921         }
1922
1923         req = fuse_get_req(fc);
1924         if (IS_ERR(req))
1925                 return PTR_ERR(req);
1926
1927         req->in.h.opcode = FUSE_POLL;
1928         req->in.h.nodeid = ff->nodeid;
1929         req->in.numargs = 1;
1930         req->in.args[0].size = sizeof(inarg);
1931         req->in.args[0].value = &inarg;
1932         req->out.numargs = 1;
1933         req->out.args[0].size = sizeof(outarg);
1934         req->out.args[0].value = &outarg;
1935         fuse_request_send(fc, req);
1936         err = req->out.h.error;
1937         fuse_put_request(fc, req);
1938
1939         if (!err)
1940                 return outarg.revents;
1941         if (err == -ENOSYS) {
1942                 fc->no_poll = 1;
1943                 return DEFAULT_POLLMASK;
1944         }
1945         return POLLERR;
1946 }
1947 EXPORT_SYMBOL_GPL(fuse_file_poll);
1948
1949 /*
1950  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
1951  * wakes up the poll waiters.
1952  */
1953 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
1954                             struct fuse_notify_poll_wakeup_out *outarg)
1955 {
1956         u64 kh = outarg->kh;
1957         struct rb_node **link;
1958
1959         spin_lock(&fc->lock);
1960
1961         link = fuse_find_polled_node(fc, kh, NULL);
1962         if (*link) {
1963                 struct fuse_file *ff;
1964
1965                 ff = rb_entry(*link, struct fuse_file, polled_node);
1966                 wake_up_interruptible_sync(&ff->poll_wait);
1967         }
1968
1969         spin_unlock(&fc->lock);
1970         return 0;
1971 }
1972
1973 static const struct file_operations fuse_file_operations = {
1974         .llseek         = fuse_file_llseek,
1975         .read           = do_sync_read,
1976         .aio_read       = fuse_file_aio_read,
1977         .write          = do_sync_write,
1978         .aio_write      = fuse_file_aio_write,
1979         .mmap           = fuse_file_mmap,
1980         .open           = fuse_open,
1981         .flush          = fuse_flush,
1982         .release        = fuse_release,
1983         .fsync          = fuse_fsync,
1984         .lock           = fuse_file_lock,
1985         .flock          = fuse_file_flock,
1986         .splice_read    = generic_file_splice_read,
1987         .unlocked_ioctl = fuse_file_ioctl,
1988         .compat_ioctl   = fuse_file_compat_ioctl,
1989         .poll           = fuse_file_poll,
1990 };
1991
1992 static const struct file_operations fuse_direct_io_file_operations = {
1993         .llseek         = fuse_file_llseek,
1994         .read           = fuse_direct_read,
1995         .write          = fuse_direct_write,
1996         .mmap           = fuse_direct_mmap,
1997         .open           = fuse_open,
1998         .flush          = fuse_flush,
1999         .release        = fuse_release,
2000         .fsync          = fuse_fsync,
2001         .lock           = fuse_file_lock,
2002         .flock          = fuse_file_flock,
2003         .unlocked_ioctl = fuse_file_ioctl,
2004         .compat_ioctl   = fuse_file_compat_ioctl,
2005         .poll           = fuse_file_poll,
2006         /* no splice_read */
2007 };
2008
2009 static const struct address_space_operations fuse_file_aops  = {
2010         .readpage       = fuse_readpage,
2011         .writepage      = fuse_writepage,
2012         .launder_page   = fuse_launder_page,
2013         .write_begin    = fuse_write_begin,
2014         .write_end      = fuse_write_end,
2015         .readpages      = fuse_readpages,
2016         .set_page_dirty = __set_page_dirty_nobuffers,
2017         .bmap           = fuse_bmap,
2018 };
2019
2020 void fuse_init_file_inode(struct inode *inode)
2021 {
2022         inode->i_fop = &fuse_file_operations;
2023         inode->i_data.a_ops = &fuse_file_aops;
2024 }