Merge branch 'for-linus' of git://git.monstr.eu/linux-2.6-microblaze
[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
16 static const struct file_operations fuse_direct_io_file_operations;
17
18 static int fuse_send_open(struct inode *inode, struct file *file, int isdir,
19                           struct fuse_open_out *outargp)
20 {
21         struct fuse_conn *fc = get_fuse_conn(inode);
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 = isdir ? FUSE_OPENDIR : FUSE_OPEN;
35         req->in.h.nodeid = get_node_id(inode);
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         ff = kmalloc(sizeof(struct fuse_file), GFP_KERNEL);
53         if (ff) {
54                 ff->reserved_req = fuse_request_alloc();
55                 if (!ff->reserved_req) {
56                         kfree(ff);
57                         return NULL;
58                 } else {
59                         INIT_LIST_HEAD(&ff->write_entry);
60                         atomic_set(&ff->count, 0);
61                         spin_lock(&fc->lock);
62                         ff->kh = ++fc->khctr;
63                         spin_unlock(&fc->lock);
64                 }
65                 RB_CLEAR_NODE(&ff->polled_node);
66                 init_waitqueue_head(&ff->poll_wait);
67         }
68         return ff;
69 }
70
71 void fuse_file_free(struct fuse_file *ff)
72 {
73         fuse_request_free(ff->reserved_req);
74         kfree(ff);
75 }
76
77 static struct fuse_file *fuse_file_get(struct fuse_file *ff)
78 {
79         atomic_inc(&ff->count);
80         return ff;
81 }
82
83 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
84 {
85         dput(req->misc.release.dentry);
86         mntput(req->misc.release.vfsmount);
87 }
88
89 static void fuse_file_put(struct fuse_file *ff)
90 {
91         if (atomic_dec_and_test(&ff->count)) {
92                 struct fuse_req *req = ff->reserved_req;
93                 struct inode *inode = req->misc.release.dentry->d_inode;
94                 struct fuse_conn *fc = get_fuse_conn(inode);
95                 req->end = fuse_release_end;
96                 fuse_request_send_background(fc, req);
97                 kfree(ff);
98         }
99 }
100
101 void fuse_finish_open(struct inode *inode, struct file *file,
102                       struct fuse_file *ff, struct fuse_open_out *outarg)
103 {
104         if (outarg->open_flags & FOPEN_DIRECT_IO)
105                 file->f_op = &fuse_direct_io_file_operations;
106         if (!(outarg->open_flags & FOPEN_KEEP_CACHE))
107                 invalidate_inode_pages2(inode->i_mapping);
108         if (outarg->open_flags & FOPEN_NONSEEKABLE)
109                 nonseekable_open(inode, file);
110         ff->fh = outarg->fh;
111         file->private_data = fuse_file_get(ff);
112 }
113
114 int fuse_open_common(struct inode *inode, struct file *file, int isdir)
115 {
116         struct fuse_conn *fc = get_fuse_conn(inode);
117         struct fuse_open_out outarg;
118         struct fuse_file *ff;
119         int err;
120
121         /* VFS checks this, but only _after_ ->open() */
122         if (file->f_flags & O_DIRECT)
123                 return -EINVAL;
124
125         err = generic_file_open(inode, file);
126         if (err)
127                 return err;
128
129         ff = fuse_file_alloc(fc);
130         if (!ff)
131                 return -ENOMEM;
132
133         err = fuse_send_open(inode, file, isdir, &outarg);
134         if (err)
135                 fuse_file_free(ff);
136         else {
137                 if (isdir)
138                         outarg.open_flags &= ~FOPEN_DIRECT_IO;
139                 fuse_finish_open(inode, file, ff, &outarg);
140         }
141
142         return err;
143 }
144
145 void fuse_release_fill(struct fuse_file *ff, u64 nodeid, int flags, int opcode)
146 {
147         struct fuse_req *req = ff->reserved_req;
148         struct fuse_release_in *inarg = &req->misc.release.in;
149
150         inarg->fh = ff->fh;
151         inarg->flags = flags;
152         req->in.h.opcode = opcode;
153         req->in.h.nodeid = nodeid;
154         req->in.numargs = 1;
155         req->in.args[0].size = sizeof(struct fuse_release_in);
156         req->in.args[0].value = inarg;
157 }
158
159 int fuse_release_common(struct inode *inode, struct file *file, int isdir)
160 {
161         struct fuse_file *ff = file->private_data;
162         if (ff) {
163                 struct fuse_conn *fc = get_fuse_conn(inode);
164                 struct fuse_req *req = ff->reserved_req;
165
166                 fuse_release_fill(ff, get_node_id(inode), file->f_flags,
167                                   isdir ? FUSE_RELEASEDIR : FUSE_RELEASE);
168
169                 /* Hold vfsmount and dentry until release is finished */
170                 req->misc.release.vfsmount = mntget(file->f_path.mnt);
171                 req->misc.release.dentry = dget(file->f_path.dentry);
172
173                 spin_lock(&fc->lock);
174                 list_del(&ff->write_entry);
175                 if (!RB_EMPTY_NODE(&ff->polled_node))
176                         rb_erase(&ff->polled_node, &fc->polled_files);
177                 spin_unlock(&fc->lock);
178
179                 wake_up_interruptible_sync(&ff->poll_wait);
180                 /*
181                  * Normally this will send the RELEASE request,
182                  * however if some asynchronous READ or WRITE requests
183                  * are outstanding, the sending will be delayed
184                  */
185                 fuse_file_put(ff);
186         }
187
188         /* Return value is ignored by VFS */
189         return 0;
190 }
191
192 static int fuse_open(struct inode *inode, struct file *file)
193 {
194         return fuse_open_common(inode, file, 0);
195 }
196
197 static int fuse_release(struct inode *inode, struct file *file)
198 {
199         return fuse_release_common(inode, file, 0);
200 }
201
202 /*
203  * Scramble the ID space with XTEA, so that the value of the files_struct
204  * pointer is not exposed to userspace.
205  */
206 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
207 {
208         u32 *k = fc->scramble_key;
209         u64 v = (unsigned long) id;
210         u32 v0 = v;
211         u32 v1 = v >> 32;
212         u32 sum = 0;
213         int i;
214
215         for (i = 0; i < 32; i++) {
216                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
217                 sum += 0x9E3779B9;
218                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
219         }
220
221         return (u64) v0 + ((u64) v1 << 32);
222 }
223
224 /*
225  * Check if page is under writeback
226  *
227  * This is currently done by walking the list of writepage requests
228  * for the inode, which can be pretty inefficient.
229  */
230 static bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
231 {
232         struct fuse_conn *fc = get_fuse_conn(inode);
233         struct fuse_inode *fi = get_fuse_inode(inode);
234         struct fuse_req *req;
235         bool found = false;
236
237         spin_lock(&fc->lock);
238         list_for_each_entry(req, &fi->writepages, writepages_entry) {
239                 pgoff_t curr_index;
240
241                 BUG_ON(req->inode != inode);
242                 curr_index = req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
243                 if (curr_index == index) {
244                         found = true;
245                         break;
246                 }
247         }
248         spin_unlock(&fc->lock);
249
250         return found;
251 }
252
253 /*
254  * Wait for page writeback to be completed.
255  *
256  * Since fuse doesn't rely on the VM writeback tracking, this has to
257  * use some other means.
258  */
259 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
260 {
261         struct fuse_inode *fi = get_fuse_inode(inode);
262
263         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
264         return 0;
265 }
266
267 static int fuse_flush(struct file *file, fl_owner_t id)
268 {
269         struct inode *inode = file->f_path.dentry->d_inode;
270         struct fuse_conn *fc = get_fuse_conn(inode);
271         struct fuse_file *ff = file->private_data;
272         struct fuse_req *req;
273         struct fuse_flush_in inarg;
274         int err;
275
276         if (is_bad_inode(inode))
277                 return -EIO;
278
279         if (fc->no_flush)
280                 return 0;
281
282         req = fuse_get_req_nofail(fc, file);
283         memset(&inarg, 0, sizeof(inarg));
284         inarg.fh = ff->fh;
285         inarg.lock_owner = fuse_lock_owner_id(fc, id);
286         req->in.h.opcode = FUSE_FLUSH;
287         req->in.h.nodeid = get_node_id(inode);
288         req->in.numargs = 1;
289         req->in.args[0].size = sizeof(inarg);
290         req->in.args[0].value = &inarg;
291         req->force = 1;
292         fuse_request_send(fc, req);
293         err = req->out.h.error;
294         fuse_put_request(fc, req);
295         if (err == -ENOSYS) {
296                 fc->no_flush = 1;
297                 err = 0;
298         }
299         return err;
300 }
301
302 /*
303  * Wait for all pending writepages on the inode to finish.
304  *
305  * This is currently done by blocking further writes with FUSE_NOWRITE
306  * and waiting for all sent writes to complete.
307  *
308  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
309  * could conflict with truncation.
310  */
311 static void fuse_sync_writes(struct inode *inode)
312 {
313         fuse_set_nowrite(inode);
314         fuse_release_nowrite(inode);
315 }
316
317 int fuse_fsync_common(struct file *file, struct dentry *de, int datasync,
318                       int isdir)
319 {
320         struct inode *inode = de->d_inode;
321         struct fuse_conn *fc = get_fuse_conn(inode);
322         struct fuse_file *ff = file->private_data;
323         struct fuse_req *req;
324         struct fuse_fsync_in inarg;
325         int err;
326
327         if (is_bad_inode(inode))
328                 return -EIO;
329
330         if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
331                 return 0;
332
333         /*
334          * Start writeback against all dirty pages of the inode, then
335          * wait for all outstanding writes, before sending the FSYNC
336          * request.
337          */
338         err = write_inode_now(inode, 0);
339         if (err)
340                 return err;
341
342         fuse_sync_writes(inode);
343
344         req = fuse_get_req(fc);
345         if (IS_ERR(req))
346                 return PTR_ERR(req);
347
348         memset(&inarg, 0, sizeof(inarg));
349         inarg.fh = ff->fh;
350         inarg.fsync_flags = datasync ? 1 : 0;
351         req->in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
352         req->in.h.nodeid = get_node_id(inode);
353         req->in.numargs = 1;
354         req->in.args[0].size = sizeof(inarg);
355         req->in.args[0].value = &inarg;
356         fuse_request_send(fc, req);
357         err = req->out.h.error;
358         fuse_put_request(fc, req);
359         if (err == -ENOSYS) {
360                 if (isdir)
361                         fc->no_fsyncdir = 1;
362                 else
363                         fc->no_fsync = 1;
364                 err = 0;
365         }
366         return err;
367 }
368
369 static int fuse_fsync(struct file *file, struct dentry *de, int datasync)
370 {
371         return fuse_fsync_common(file, de, datasync, 0);
372 }
373
374 void fuse_read_fill(struct fuse_req *req, struct file *file,
375                     struct inode *inode, loff_t pos, size_t count, int opcode)
376 {
377         struct fuse_read_in *inarg = &req->misc.read.in;
378         struct fuse_file *ff = file->private_data;
379
380         inarg->fh = ff->fh;
381         inarg->offset = pos;
382         inarg->size = count;
383         inarg->flags = file->f_flags;
384         req->in.h.opcode = opcode;
385         req->in.h.nodeid = get_node_id(inode);
386         req->in.numargs = 1;
387         req->in.args[0].size = sizeof(struct fuse_read_in);
388         req->in.args[0].value = inarg;
389         req->out.argvar = 1;
390         req->out.numargs = 1;
391         req->out.args[0].size = count;
392 }
393
394 static size_t fuse_send_read(struct fuse_req *req, struct file *file,
395                              struct inode *inode, loff_t pos, size_t count,
396                              fl_owner_t owner)
397 {
398         struct fuse_conn *fc = get_fuse_conn(inode);
399
400         fuse_read_fill(req, file, inode, pos, count, FUSE_READ);
401         if (owner != NULL) {
402                 struct fuse_read_in *inarg = &req->misc.read.in;
403
404                 inarg->read_flags |= FUSE_READ_LOCKOWNER;
405                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
406         }
407         fuse_request_send(fc, req);
408         return req->out.args[0].size;
409 }
410
411 static void fuse_read_update_size(struct inode *inode, loff_t size,
412                                   u64 attr_ver)
413 {
414         struct fuse_conn *fc = get_fuse_conn(inode);
415         struct fuse_inode *fi = get_fuse_inode(inode);
416
417         spin_lock(&fc->lock);
418         if (attr_ver == fi->attr_version && size < inode->i_size) {
419                 fi->attr_version = ++fc->attr_version;
420                 i_size_write(inode, size);
421         }
422         spin_unlock(&fc->lock);
423 }
424
425 static int fuse_readpage(struct file *file, struct page *page)
426 {
427         struct inode *inode = page->mapping->host;
428         struct fuse_conn *fc = get_fuse_conn(inode);
429         struct fuse_req *req;
430         size_t num_read;
431         loff_t pos = page_offset(page);
432         size_t count = PAGE_CACHE_SIZE;
433         u64 attr_ver;
434         int err;
435
436         err = -EIO;
437         if (is_bad_inode(inode))
438                 goto out;
439
440         /*
441          * Page writeback can extend beyond the liftime of the
442          * page-cache page, so make sure we read a properly synced
443          * page.
444          */
445         fuse_wait_on_page_writeback(inode, page->index);
446
447         req = fuse_get_req(fc);
448         err = PTR_ERR(req);
449         if (IS_ERR(req))
450                 goto out;
451
452         attr_ver = fuse_get_attr_version(fc);
453
454         req->out.page_zeroing = 1;
455         req->out.argpages = 1;
456         req->num_pages = 1;
457         req->pages[0] = page;
458         num_read = fuse_send_read(req, file, inode, pos, count, NULL);
459         err = req->out.h.error;
460         fuse_put_request(fc, req);
461
462         if (!err) {
463                 /*
464                  * Short read means EOF.  If file size is larger, truncate it
465                  */
466                 if (num_read < count)
467                         fuse_read_update_size(inode, pos + num_read, attr_ver);
468
469                 SetPageUptodate(page);
470         }
471
472         fuse_invalidate_attr(inode); /* atime changed */
473  out:
474         unlock_page(page);
475         return err;
476 }
477
478 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
479 {
480         int i;
481         size_t count = req->misc.read.in.size;
482         size_t num_read = req->out.args[0].size;
483         struct inode *inode = req->pages[0]->mapping->host;
484
485         /*
486          * Short read means EOF.  If file size is larger, truncate it
487          */
488         if (!req->out.h.error && num_read < count) {
489                 loff_t pos = page_offset(req->pages[0]) + num_read;
490                 fuse_read_update_size(inode, pos, req->misc.read.attr_ver);
491         }
492
493         fuse_invalidate_attr(inode); /* atime changed */
494
495         for (i = 0; i < req->num_pages; i++) {
496                 struct page *page = req->pages[i];
497                 if (!req->out.h.error)
498                         SetPageUptodate(page);
499                 else
500                         SetPageError(page);
501                 unlock_page(page);
502         }
503         if (req->ff)
504                 fuse_file_put(req->ff);
505 }
506
507 static void fuse_send_readpages(struct fuse_req *req, struct file *file,
508                                 struct inode *inode)
509 {
510         struct fuse_conn *fc = get_fuse_conn(inode);
511         loff_t pos = page_offset(req->pages[0]);
512         size_t count = req->num_pages << PAGE_CACHE_SHIFT;
513
514         req->out.argpages = 1;
515         req->out.page_zeroing = 1;
516         fuse_read_fill(req, file, inode, pos, count, FUSE_READ);
517         req->misc.read.attr_ver = fuse_get_attr_version(fc);
518         if (fc->async_read) {
519                 struct fuse_file *ff = file->private_data;
520                 req->ff = fuse_file_get(ff);
521                 req->end = fuse_readpages_end;
522                 fuse_request_send_background(fc, req);
523         } else {
524                 fuse_request_send(fc, req);
525                 fuse_readpages_end(fc, req);
526                 fuse_put_request(fc, req);
527         }
528 }
529
530 struct fuse_fill_data {
531         struct fuse_req *req;
532         struct file *file;
533         struct inode *inode;
534 };
535
536 static int fuse_readpages_fill(void *_data, struct page *page)
537 {
538         struct fuse_fill_data *data = _data;
539         struct fuse_req *req = data->req;
540         struct inode *inode = data->inode;
541         struct fuse_conn *fc = get_fuse_conn(inode);
542
543         fuse_wait_on_page_writeback(inode, page->index);
544
545         if (req->num_pages &&
546             (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
547              (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
548              req->pages[req->num_pages - 1]->index + 1 != page->index)) {
549                 fuse_send_readpages(req, data->file, inode);
550                 data->req = req = fuse_get_req(fc);
551                 if (IS_ERR(req)) {
552                         unlock_page(page);
553                         return PTR_ERR(req);
554                 }
555         }
556         req->pages[req->num_pages] = page;
557         req->num_pages++;
558         return 0;
559 }
560
561 static int fuse_readpages(struct file *file, struct address_space *mapping,
562                           struct list_head *pages, unsigned nr_pages)
563 {
564         struct inode *inode = mapping->host;
565         struct fuse_conn *fc = get_fuse_conn(inode);
566         struct fuse_fill_data data;
567         int err;
568
569         err = -EIO;
570         if (is_bad_inode(inode))
571                 goto out;
572
573         data.file = file;
574         data.inode = inode;
575         data.req = fuse_get_req(fc);
576         err = PTR_ERR(data.req);
577         if (IS_ERR(data.req))
578                 goto out;
579
580         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
581         if (!err) {
582                 if (data.req->num_pages)
583                         fuse_send_readpages(data.req, file, inode);
584                 else
585                         fuse_put_request(fc, data.req);
586         }
587 out:
588         return err;
589 }
590
591 static ssize_t fuse_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
592                                   unsigned long nr_segs, loff_t pos)
593 {
594         struct inode *inode = iocb->ki_filp->f_mapping->host;
595
596         if (pos + iov_length(iov, nr_segs) > i_size_read(inode)) {
597                 int err;
598                 /*
599                  * If trying to read past EOF, make sure the i_size
600                  * attribute is up-to-date.
601                  */
602                 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
603                 if (err)
604                         return err;
605         }
606
607         return generic_file_aio_read(iocb, iov, nr_segs, pos);
608 }
609
610 static void fuse_write_fill(struct fuse_req *req, struct file *file,
611                             struct fuse_file *ff, struct inode *inode,
612                             loff_t pos, size_t count, int writepage)
613 {
614         struct fuse_conn *fc = get_fuse_conn(inode);
615         struct fuse_write_in *inarg = &req->misc.write.in;
616         struct fuse_write_out *outarg = &req->misc.write.out;
617
618         memset(inarg, 0, sizeof(struct fuse_write_in));
619         inarg->fh = ff->fh;
620         inarg->offset = pos;
621         inarg->size = count;
622         inarg->write_flags = writepage ? FUSE_WRITE_CACHE : 0;
623         inarg->flags = file ? file->f_flags : 0;
624         req->in.h.opcode = FUSE_WRITE;
625         req->in.h.nodeid = get_node_id(inode);
626         req->in.numargs = 2;
627         if (fc->minor < 9)
628                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
629         else
630                 req->in.args[0].size = sizeof(struct fuse_write_in);
631         req->in.args[0].value = inarg;
632         req->in.args[1].size = count;
633         req->out.numargs = 1;
634         req->out.args[0].size = sizeof(struct fuse_write_out);
635         req->out.args[0].value = outarg;
636 }
637
638 static size_t fuse_send_write(struct fuse_req *req, struct file *file,
639                               struct inode *inode, loff_t pos, size_t count,
640                               fl_owner_t owner)
641 {
642         struct fuse_conn *fc = get_fuse_conn(inode);
643         fuse_write_fill(req, file, file->private_data, inode, pos, count, 0);
644         if (owner != NULL) {
645                 struct fuse_write_in *inarg = &req->misc.write.in;
646                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
647                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
648         }
649         fuse_request_send(fc, req);
650         return req->misc.write.out.size;
651 }
652
653 static int fuse_write_begin(struct file *file, struct address_space *mapping,
654                         loff_t pos, unsigned len, unsigned flags,
655                         struct page **pagep, void **fsdata)
656 {
657         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
658
659         *pagep = grab_cache_page_write_begin(mapping, index, flags);
660         if (!*pagep)
661                 return -ENOMEM;
662         return 0;
663 }
664
665 static void fuse_write_update_size(struct inode *inode, loff_t pos)
666 {
667         struct fuse_conn *fc = get_fuse_conn(inode);
668         struct fuse_inode *fi = get_fuse_inode(inode);
669
670         spin_lock(&fc->lock);
671         fi->attr_version = ++fc->attr_version;
672         if (pos > inode->i_size)
673                 i_size_write(inode, pos);
674         spin_unlock(&fc->lock);
675 }
676
677 static int fuse_buffered_write(struct file *file, struct inode *inode,
678                                loff_t pos, unsigned count, struct page *page)
679 {
680         int err;
681         size_t nres;
682         struct fuse_conn *fc = get_fuse_conn(inode);
683         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
684         struct fuse_req *req;
685
686         if (is_bad_inode(inode))
687                 return -EIO;
688
689         /*
690          * Make sure writepages on the same page are not mixed up with
691          * plain writes.
692          */
693         fuse_wait_on_page_writeback(inode, page->index);
694
695         req = fuse_get_req(fc);
696         if (IS_ERR(req))
697                 return PTR_ERR(req);
698
699         req->in.argpages = 1;
700         req->num_pages = 1;
701         req->pages[0] = page;
702         req->page_offset = offset;
703         nres = fuse_send_write(req, file, inode, pos, count, NULL);
704         err = req->out.h.error;
705         fuse_put_request(fc, req);
706         if (!err && !nres)
707                 err = -EIO;
708         if (!err) {
709                 pos += nres;
710                 fuse_write_update_size(inode, pos);
711                 if (count == PAGE_CACHE_SIZE)
712                         SetPageUptodate(page);
713         }
714         fuse_invalidate_attr(inode);
715         return err ? err : nres;
716 }
717
718 static int fuse_write_end(struct file *file, struct address_space *mapping,
719                         loff_t pos, unsigned len, unsigned copied,
720                         struct page *page, void *fsdata)
721 {
722         struct inode *inode = mapping->host;
723         int res = 0;
724
725         if (copied)
726                 res = fuse_buffered_write(file, inode, pos, copied, page);
727
728         unlock_page(page);
729         page_cache_release(page);
730         return res;
731 }
732
733 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
734                                     struct inode *inode, loff_t pos,
735                                     size_t count)
736 {
737         size_t res;
738         unsigned offset;
739         unsigned i;
740
741         for (i = 0; i < req->num_pages; i++)
742                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
743
744         res = fuse_send_write(req, file, inode, pos, count, NULL);
745
746         offset = req->page_offset;
747         count = res;
748         for (i = 0; i < req->num_pages; i++) {
749                 struct page *page = req->pages[i];
750
751                 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
752                         SetPageUptodate(page);
753
754                 if (count > PAGE_CACHE_SIZE - offset)
755                         count -= PAGE_CACHE_SIZE - offset;
756                 else
757                         count = 0;
758                 offset = 0;
759
760                 unlock_page(page);
761                 page_cache_release(page);
762         }
763
764         return res;
765 }
766
767 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
768                                struct address_space *mapping,
769                                struct iov_iter *ii, loff_t pos)
770 {
771         struct fuse_conn *fc = get_fuse_conn(mapping->host);
772         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
773         size_t count = 0;
774         int err;
775
776         req->in.argpages = 1;
777         req->page_offset = offset;
778
779         do {
780                 size_t tmp;
781                 struct page *page;
782                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
783                 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
784                                      iov_iter_count(ii));
785
786                 bytes = min_t(size_t, bytes, fc->max_write - count);
787
788  again:
789                 err = -EFAULT;
790                 if (iov_iter_fault_in_readable(ii, bytes))
791                         break;
792
793                 err = -ENOMEM;
794                 page = grab_cache_page_write_begin(mapping, index, 0);
795                 if (!page)
796                         break;
797
798                 pagefault_disable();
799                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
800                 pagefault_enable();
801                 flush_dcache_page(page);
802
803                 if (!tmp) {
804                         unlock_page(page);
805                         page_cache_release(page);
806                         bytes = min(bytes, iov_iter_single_seg_count(ii));
807                         goto again;
808                 }
809
810                 err = 0;
811                 req->pages[req->num_pages] = page;
812                 req->num_pages++;
813
814                 iov_iter_advance(ii, tmp);
815                 count += tmp;
816                 pos += tmp;
817                 offset += tmp;
818                 if (offset == PAGE_CACHE_SIZE)
819                         offset = 0;
820
821                 if (!fc->big_writes)
822                         break;
823         } while (iov_iter_count(ii) && count < fc->max_write &&
824                  req->num_pages < FUSE_MAX_PAGES_PER_REQ && offset == 0);
825
826         return count > 0 ? count : err;
827 }
828
829 static ssize_t fuse_perform_write(struct file *file,
830                                   struct address_space *mapping,
831                                   struct iov_iter *ii, loff_t pos)
832 {
833         struct inode *inode = mapping->host;
834         struct fuse_conn *fc = get_fuse_conn(inode);
835         int err = 0;
836         ssize_t res = 0;
837
838         if (is_bad_inode(inode))
839                 return -EIO;
840
841         do {
842                 struct fuse_req *req;
843                 ssize_t count;
844
845                 req = fuse_get_req(fc);
846                 if (IS_ERR(req)) {
847                         err = PTR_ERR(req);
848                         break;
849                 }
850
851                 count = fuse_fill_write_pages(req, mapping, ii, pos);
852                 if (count <= 0) {
853                         err = count;
854                 } else {
855                         size_t num_written;
856
857                         num_written = fuse_send_write_pages(req, file, inode,
858                                                             pos, count);
859                         err = req->out.h.error;
860                         if (!err) {
861                                 res += num_written;
862                                 pos += num_written;
863
864                                 /* break out of the loop on short write */
865                                 if (num_written != count)
866                                         err = -EIO;
867                         }
868                 }
869                 fuse_put_request(fc, req);
870         } while (!err && iov_iter_count(ii));
871
872         if (res > 0)
873                 fuse_write_update_size(inode, pos);
874
875         fuse_invalidate_attr(inode);
876
877         return res > 0 ? res : err;
878 }
879
880 static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
881                                    unsigned long nr_segs, loff_t pos)
882 {
883         struct file *file = iocb->ki_filp;
884         struct address_space *mapping = file->f_mapping;
885         size_t count = 0;
886         ssize_t written = 0;
887         struct inode *inode = mapping->host;
888         ssize_t err;
889         struct iov_iter i;
890
891         WARN_ON(iocb->ki_pos != pos);
892
893         err = generic_segment_checks(iov, &nr_segs, &count, VERIFY_READ);
894         if (err)
895                 return err;
896
897         mutex_lock(&inode->i_mutex);
898         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
899
900         /* We can write back this queue in page reclaim */
901         current->backing_dev_info = mapping->backing_dev_info;
902
903         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
904         if (err)
905                 goto out;
906
907         if (count == 0)
908                 goto out;
909
910         err = file_remove_suid(file);
911         if (err)
912                 goto out;
913
914         file_update_time(file);
915
916         iov_iter_init(&i, iov, nr_segs, count, 0);
917         written = fuse_perform_write(file, mapping, &i, pos);
918         if (written >= 0)
919                 iocb->ki_pos = pos + written;
920
921 out:
922         current->backing_dev_info = NULL;
923         mutex_unlock(&inode->i_mutex);
924
925         return written ? written : err;
926 }
927
928 static void fuse_release_user_pages(struct fuse_req *req, int write)
929 {
930         unsigned i;
931
932         for (i = 0; i < req->num_pages; i++) {
933                 struct page *page = req->pages[i];
934                 if (write)
935                         set_page_dirty_lock(page);
936                 put_page(page);
937         }
938 }
939
940 static int fuse_get_user_pages(struct fuse_req *req, const char __user *buf,
941                                unsigned *nbytesp, int write)
942 {
943         unsigned nbytes = *nbytesp;
944         unsigned long user_addr = (unsigned long) buf;
945         unsigned offset = user_addr & ~PAGE_MASK;
946         int npages;
947
948         /* Special case for kernel I/O: can copy directly into the buffer */
949         if (segment_eq(get_fs(), KERNEL_DS)) {
950                 if (write)
951                         req->in.args[1].value = (void *) user_addr;
952                 else
953                         req->out.args[0].value = (void *) user_addr;
954
955                 return 0;
956         }
957
958         nbytes = min(nbytes, (unsigned) FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
959         npages = (nbytes + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
960         npages = clamp(npages, 1, FUSE_MAX_PAGES_PER_REQ);
961         down_read(&current->mm->mmap_sem);
962         npages = get_user_pages(current, current->mm, user_addr, npages, !write,
963                                 0, req->pages, NULL);
964         up_read(&current->mm->mmap_sem);
965         if (npages < 0)
966                 return npages;
967
968         req->num_pages = npages;
969         req->page_offset = offset;
970
971         if (write)
972                 req->in.argpages = 1;
973         else
974                 req->out.argpages = 1;
975
976         nbytes = (req->num_pages << PAGE_SHIFT) - req->page_offset;
977         *nbytesp = min(*nbytesp, nbytes);
978
979         return 0;
980 }
981
982 static ssize_t fuse_direct_io(struct file *file, const char __user *buf,
983                               size_t count, loff_t *ppos, int write)
984 {
985         struct inode *inode = file->f_path.dentry->d_inode;
986         struct fuse_conn *fc = get_fuse_conn(inode);
987         size_t nmax = write ? fc->max_write : fc->max_read;
988         loff_t pos = *ppos;
989         ssize_t res = 0;
990         struct fuse_req *req;
991
992         if (is_bad_inode(inode))
993                 return -EIO;
994
995         req = fuse_get_req(fc);
996         if (IS_ERR(req))
997                 return PTR_ERR(req);
998
999         while (count) {
1000                 size_t nres;
1001                 size_t nbytes = min(count, nmax);
1002                 int err = fuse_get_user_pages(req, buf, &nbytes, write);
1003                 if (err) {
1004                         res = err;
1005                         break;
1006                 }
1007
1008                 if (write)
1009                         nres = fuse_send_write(req, file, inode, pos, nbytes,
1010                                                current->files);
1011                 else
1012                         nres = fuse_send_read(req, file, inode, pos, nbytes,
1013                                               current->files);
1014                 fuse_release_user_pages(req, !write);
1015                 if (req->out.h.error) {
1016                         if (!res)
1017                                 res = req->out.h.error;
1018                         break;
1019                 } else if (nres > nbytes) {
1020                         res = -EIO;
1021                         break;
1022                 }
1023                 count -= nres;
1024                 res += nres;
1025                 pos += nres;
1026                 buf += nres;
1027                 if (nres != nbytes)
1028                         break;
1029                 if (count) {
1030                         fuse_put_request(fc, req);
1031                         req = fuse_get_req(fc);
1032                         if (IS_ERR(req))
1033                                 break;
1034                 }
1035         }
1036         fuse_put_request(fc, req);
1037         if (res > 0) {
1038                 if (write)
1039                         fuse_write_update_size(inode, pos);
1040                 *ppos = pos;
1041         }
1042         fuse_invalidate_attr(inode);
1043
1044         return res;
1045 }
1046
1047 static ssize_t fuse_direct_read(struct file *file, char __user *buf,
1048                                      size_t count, loff_t *ppos)
1049 {
1050         return fuse_direct_io(file, buf, count, ppos, 0);
1051 }
1052
1053 static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
1054                                  size_t count, loff_t *ppos)
1055 {
1056         struct inode *inode = file->f_path.dentry->d_inode;
1057         ssize_t res;
1058         /* Don't allow parallel writes to the same file */
1059         mutex_lock(&inode->i_mutex);
1060         res = generic_write_checks(file, ppos, &count, 0);
1061         if (!res)
1062                 res = fuse_direct_io(file, buf, count, ppos, 1);
1063         mutex_unlock(&inode->i_mutex);
1064         return res;
1065 }
1066
1067 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1068 {
1069         __free_page(req->pages[0]);
1070         fuse_file_put(req->ff);
1071 }
1072
1073 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1074 {
1075         struct inode *inode = req->inode;
1076         struct fuse_inode *fi = get_fuse_inode(inode);
1077         struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
1078
1079         list_del(&req->writepages_entry);
1080         dec_bdi_stat(bdi, BDI_WRITEBACK);
1081         dec_zone_page_state(req->pages[0], NR_WRITEBACK_TEMP);
1082         bdi_writeout_inc(bdi);
1083         wake_up(&fi->page_waitq);
1084 }
1085
1086 /* Called under fc->lock, may release and reacquire it */
1087 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req)
1088 __releases(&fc->lock)
1089 __acquires(&fc->lock)
1090 {
1091         struct fuse_inode *fi = get_fuse_inode(req->inode);
1092         loff_t size = i_size_read(req->inode);
1093         struct fuse_write_in *inarg = &req->misc.write.in;
1094
1095         if (!fc->connected)
1096                 goto out_free;
1097
1098         if (inarg->offset + PAGE_CACHE_SIZE <= size) {
1099                 inarg->size = PAGE_CACHE_SIZE;
1100         } else if (inarg->offset < size) {
1101                 inarg->size = size & (PAGE_CACHE_SIZE - 1);
1102         } else {
1103                 /* Got truncated off completely */
1104                 goto out_free;
1105         }
1106
1107         req->in.args[1].size = inarg->size;
1108         fi->writectr++;
1109         fuse_request_send_background_locked(fc, req);
1110         return;
1111
1112  out_free:
1113         fuse_writepage_finish(fc, req);
1114         spin_unlock(&fc->lock);
1115         fuse_writepage_free(fc, req);
1116         fuse_put_request(fc, req);
1117         spin_lock(&fc->lock);
1118 }
1119
1120 /*
1121  * If fi->writectr is positive (no truncate or fsync going on) send
1122  * all queued writepage requests.
1123  *
1124  * Called with fc->lock
1125  */
1126 void fuse_flush_writepages(struct inode *inode)
1127 __releases(&fc->lock)
1128 __acquires(&fc->lock)
1129 {
1130         struct fuse_conn *fc = get_fuse_conn(inode);
1131         struct fuse_inode *fi = get_fuse_inode(inode);
1132         struct fuse_req *req;
1133
1134         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1135                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1136                 list_del_init(&req->list);
1137                 fuse_send_writepage(fc, req);
1138         }
1139 }
1140
1141 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1142 {
1143         struct inode *inode = req->inode;
1144         struct fuse_inode *fi = get_fuse_inode(inode);
1145
1146         mapping_set_error(inode->i_mapping, req->out.h.error);
1147         spin_lock(&fc->lock);
1148         fi->writectr--;
1149         fuse_writepage_finish(fc, req);
1150         spin_unlock(&fc->lock);
1151         fuse_writepage_free(fc, req);
1152 }
1153
1154 static int fuse_writepage_locked(struct page *page)
1155 {
1156         struct address_space *mapping = page->mapping;
1157         struct inode *inode = mapping->host;
1158         struct fuse_conn *fc = get_fuse_conn(inode);
1159         struct fuse_inode *fi = get_fuse_inode(inode);
1160         struct fuse_req *req;
1161         struct fuse_file *ff;
1162         struct page *tmp_page;
1163
1164         set_page_writeback(page);
1165
1166         req = fuse_request_alloc_nofs();
1167         if (!req)
1168                 goto err;
1169
1170         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1171         if (!tmp_page)
1172                 goto err_free;
1173
1174         spin_lock(&fc->lock);
1175         BUG_ON(list_empty(&fi->write_files));
1176         ff = list_entry(fi->write_files.next, struct fuse_file, write_entry);
1177         req->ff = fuse_file_get(ff);
1178         spin_unlock(&fc->lock);
1179
1180         fuse_write_fill(req, NULL, ff, inode, page_offset(page), 0, 1);
1181
1182         copy_highpage(tmp_page, page);
1183         req->in.argpages = 1;
1184         req->num_pages = 1;
1185         req->pages[0] = tmp_page;
1186         req->page_offset = 0;
1187         req->end = fuse_writepage_end;
1188         req->inode = inode;
1189
1190         inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
1191         inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1192         end_page_writeback(page);
1193
1194         spin_lock(&fc->lock);
1195         list_add(&req->writepages_entry, &fi->writepages);
1196         list_add_tail(&req->list, &fi->queued_writes);
1197         fuse_flush_writepages(inode);
1198         spin_unlock(&fc->lock);
1199
1200         return 0;
1201
1202 err_free:
1203         fuse_request_free(req);
1204 err:
1205         end_page_writeback(page);
1206         return -ENOMEM;
1207 }
1208
1209 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1210 {
1211         int err;
1212
1213         err = fuse_writepage_locked(page);
1214         unlock_page(page);
1215
1216         return err;
1217 }
1218
1219 static int fuse_launder_page(struct page *page)
1220 {
1221         int err = 0;
1222         if (clear_page_dirty_for_io(page)) {
1223                 struct inode *inode = page->mapping->host;
1224                 err = fuse_writepage_locked(page);
1225                 if (!err)
1226                         fuse_wait_on_page_writeback(inode, page->index);
1227         }
1228         return err;
1229 }
1230
1231 /*
1232  * Write back dirty pages now, because there may not be any suitable
1233  * open files later
1234  */
1235 static void fuse_vma_close(struct vm_area_struct *vma)
1236 {
1237         filemap_write_and_wait(vma->vm_file->f_mapping);
1238 }
1239
1240 /*
1241  * Wait for writeback against this page to complete before allowing it
1242  * to be marked dirty again, and hence written back again, possibly
1243  * before the previous writepage completed.
1244  *
1245  * Block here, instead of in ->writepage(), so that the userspace fs
1246  * can only block processes actually operating on the filesystem.
1247  *
1248  * Otherwise unprivileged userspace fs would be able to block
1249  * unrelated:
1250  *
1251  * - page migration
1252  * - sync(2)
1253  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
1254  */
1255 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1256 {
1257         struct page *page = vmf->page;
1258         /*
1259          * Don't use page->mapping as it may become NULL from a
1260          * concurrent truncate.
1261          */
1262         struct inode *inode = vma->vm_file->f_mapping->host;
1263
1264         fuse_wait_on_page_writeback(inode, page->index);
1265         return 0;
1266 }
1267
1268 static struct vm_operations_struct fuse_file_vm_ops = {
1269         .close          = fuse_vma_close,
1270         .fault          = filemap_fault,
1271         .page_mkwrite   = fuse_page_mkwrite,
1272 };
1273
1274 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
1275 {
1276         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE)) {
1277                 struct inode *inode = file->f_dentry->d_inode;
1278                 struct fuse_conn *fc = get_fuse_conn(inode);
1279                 struct fuse_inode *fi = get_fuse_inode(inode);
1280                 struct fuse_file *ff = file->private_data;
1281                 /*
1282                  * file may be written through mmap, so chain it onto the
1283                  * inodes's write_file list
1284                  */
1285                 spin_lock(&fc->lock);
1286                 if (list_empty(&ff->write_entry))
1287                         list_add(&ff->write_entry, &fi->write_files);
1288                 spin_unlock(&fc->lock);
1289         }
1290         file_accessed(file);
1291         vma->vm_ops = &fuse_file_vm_ops;
1292         return 0;
1293 }
1294
1295 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
1296 {
1297         /* Can't provide the coherency needed for MAP_SHARED */
1298         if (vma->vm_flags & VM_MAYSHARE)
1299                 return -ENODEV;
1300
1301         return generic_file_mmap(file, vma);
1302 }
1303
1304 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
1305                                   struct file_lock *fl)
1306 {
1307         switch (ffl->type) {
1308         case F_UNLCK:
1309                 break;
1310
1311         case F_RDLCK:
1312         case F_WRLCK:
1313                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
1314                     ffl->end < ffl->start)
1315                         return -EIO;
1316
1317                 fl->fl_start = ffl->start;
1318                 fl->fl_end = ffl->end;
1319                 fl->fl_pid = ffl->pid;
1320                 break;
1321
1322         default:
1323                 return -EIO;
1324         }
1325         fl->fl_type = ffl->type;
1326         return 0;
1327 }
1328
1329 static void fuse_lk_fill(struct fuse_req *req, struct file *file,
1330                          const struct file_lock *fl, int opcode, pid_t pid,
1331                          int flock)
1332 {
1333         struct inode *inode = file->f_path.dentry->d_inode;
1334         struct fuse_conn *fc = get_fuse_conn(inode);
1335         struct fuse_file *ff = file->private_data;
1336         struct fuse_lk_in *arg = &req->misc.lk_in;
1337
1338         arg->fh = ff->fh;
1339         arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
1340         arg->lk.start = fl->fl_start;
1341         arg->lk.end = fl->fl_end;
1342         arg->lk.type = fl->fl_type;
1343         arg->lk.pid = pid;
1344         if (flock)
1345                 arg->lk_flags |= FUSE_LK_FLOCK;
1346         req->in.h.opcode = opcode;
1347         req->in.h.nodeid = get_node_id(inode);
1348         req->in.numargs = 1;
1349         req->in.args[0].size = sizeof(*arg);
1350         req->in.args[0].value = arg;
1351 }
1352
1353 static int fuse_getlk(struct file *file, struct file_lock *fl)
1354 {
1355         struct inode *inode = file->f_path.dentry->d_inode;
1356         struct fuse_conn *fc = get_fuse_conn(inode);
1357         struct fuse_req *req;
1358         struct fuse_lk_out outarg;
1359         int err;
1360
1361         req = fuse_get_req(fc);
1362         if (IS_ERR(req))
1363                 return PTR_ERR(req);
1364
1365         fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
1366         req->out.numargs = 1;
1367         req->out.args[0].size = sizeof(outarg);
1368         req->out.args[0].value = &outarg;
1369         fuse_request_send(fc, req);
1370         err = req->out.h.error;
1371         fuse_put_request(fc, req);
1372         if (!err)
1373                 err = convert_fuse_file_lock(&outarg.lk, fl);
1374
1375         return err;
1376 }
1377
1378 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
1379 {
1380         struct inode *inode = file->f_path.dentry->d_inode;
1381         struct fuse_conn *fc = get_fuse_conn(inode);
1382         struct fuse_req *req;
1383         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
1384         pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
1385         int err;
1386
1387         if (fl->fl_lmops && fl->fl_lmops->fl_grant) {
1388                 /* NLM needs asynchronous locks, which we don't support yet */
1389                 return -ENOLCK;
1390         }
1391
1392         /* Unlock on close is handled by the flush method */
1393         if (fl->fl_flags & FL_CLOSE)
1394                 return 0;
1395
1396         req = fuse_get_req(fc);
1397         if (IS_ERR(req))
1398                 return PTR_ERR(req);
1399
1400         fuse_lk_fill(req, file, fl, opcode, pid, flock);
1401         fuse_request_send(fc, req);
1402         err = req->out.h.error;
1403         /* locking is restartable */
1404         if (err == -EINTR)
1405                 err = -ERESTARTSYS;
1406         fuse_put_request(fc, req);
1407         return err;
1408 }
1409
1410 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
1411 {
1412         struct inode *inode = file->f_path.dentry->d_inode;
1413         struct fuse_conn *fc = get_fuse_conn(inode);
1414         int err;
1415
1416         if (cmd == F_CANCELLK) {
1417                 err = 0;
1418         } else if (cmd == F_GETLK) {
1419                 if (fc->no_lock) {
1420                         posix_test_lock(file, fl);
1421                         err = 0;
1422                 } else
1423                         err = fuse_getlk(file, fl);
1424         } else {
1425                 if (fc->no_lock)
1426                         err = posix_lock_file(file, fl, NULL);
1427                 else
1428                         err = fuse_setlk(file, fl, 0);
1429         }
1430         return err;
1431 }
1432
1433 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
1434 {
1435         struct inode *inode = file->f_path.dentry->d_inode;
1436         struct fuse_conn *fc = get_fuse_conn(inode);
1437         int err;
1438
1439         if (fc->no_lock) {
1440                 err = flock_lock_file_wait(file, fl);
1441         } else {
1442                 /* emulate flock with POSIX locks */
1443                 fl->fl_owner = (fl_owner_t) file;
1444                 err = fuse_setlk(file, fl, 1);
1445         }
1446
1447         return err;
1448 }
1449
1450 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
1451 {
1452         struct inode *inode = mapping->host;
1453         struct fuse_conn *fc = get_fuse_conn(inode);
1454         struct fuse_req *req;
1455         struct fuse_bmap_in inarg;
1456         struct fuse_bmap_out outarg;
1457         int err;
1458
1459         if (!inode->i_sb->s_bdev || fc->no_bmap)
1460                 return 0;
1461
1462         req = fuse_get_req(fc);
1463         if (IS_ERR(req))
1464                 return 0;
1465
1466         memset(&inarg, 0, sizeof(inarg));
1467         inarg.block = block;
1468         inarg.blocksize = inode->i_sb->s_blocksize;
1469         req->in.h.opcode = FUSE_BMAP;
1470         req->in.h.nodeid = get_node_id(inode);
1471         req->in.numargs = 1;
1472         req->in.args[0].size = sizeof(inarg);
1473         req->in.args[0].value = &inarg;
1474         req->out.numargs = 1;
1475         req->out.args[0].size = sizeof(outarg);
1476         req->out.args[0].value = &outarg;
1477         fuse_request_send(fc, req);
1478         err = req->out.h.error;
1479         fuse_put_request(fc, req);
1480         if (err == -ENOSYS)
1481                 fc->no_bmap = 1;
1482
1483         return err ? 0 : outarg.block;
1484 }
1485
1486 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int origin)
1487 {
1488         loff_t retval;
1489         struct inode *inode = file->f_path.dentry->d_inode;
1490
1491         mutex_lock(&inode->i_mutex);
1492         switch (origin) {
1493         case SEEK_END:
1494                 retval = fuse_update_attributes(inode, NULL, file, NULL);
1495                 if (retval)
1496                         goto exit;
1497                 offset += i_size_read(inode);
1498                 break;
1499         case SEEK_CUR:
1500                 offset += file->f_pos;
1501         }
1502         retval = -EINVAL;
1503         if (offset >= 0 && offset <= inode->i_sb->s_maxbytes) {
1504                 if (offset != file->f_pos) {
1505                         file->f_pos = offset;
1506                         file->f_version = 0;
1507                 }
1508                 retval = offset;
1509         }
1510 exit:
1511         mutex_unlock(&inode->i_mutex);
1512         return retval;
1513 }
1514
1515 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
1516                         unsigned int nr_segs, size_t bytes, bool to_user)
1517 {
1518         struct iov_iter ii;
1519         int page_idx = 0;
1520
1521         if (!bytes)
1522                 return 0;
1523
1524         iov_iter_init(&ii, iov, nr_segs, bytes, 0);
1525
1526         while (iov_iter_count(&ii)) {
1527                 struct page *page = pages[page_idx++];
1528                 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
1529                 void *kaddr, *map;
1530
1531                 kaddr = map = kmap(page);
1532
1533                 while (todo) {
1534                         char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
1535                         size_t iov_len = ii.iov->iov_len - ii.iov_offset;
1536                         size_t copy = min(todo, iov_len);
1537                         size_t left;
1538
1539                         if (!to_user)
1540                                 left = copy_from_user(kaddr, uaddr, copy);
1541                         else
1542                                 left = copy_to_user(uaddr, kaddr, copy);
1543
1544                         if (unlikely(left))
1545                                 return -EFAULT;
1546
1547                         iov_iter_advance(&ii, copy);
1548                         todo -= copy;
1549                         kaddr += copy;
1550                 }
1551
1552                 kunmap(map);
1553         }
1554
1555         return 0;
1556 }
1557
1558 /*
1559  * For ioctls, there is no generic way to determine how much memory
1560  * needs to be read and/or written.  Furthermore, ioctls are allowed
1561  * to dereference the passed pointer, so the parameter requires deep
1562  * copying but FUSE has no idea whatsoever about what to copy in or
1563  * out.
1564  *
1565  * This is solved by allowing FUSE server to retry ioctl with
1566  * necessary in/out iovecs.  Let's assume the ioctl implementation
1567  * needs to read in the following structure.
1568  *
1569  * struct a {
1570  *      char    *buf;
1571  *      size_t  buflen;
1572  * }
1573  *
1574  * On the first callout to FUSE server, inarg->in_size and
1575  * inarg->out_size will be NULL; then, the server completes the ioctl
1576  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
1577  * the actual iov array to
1578  *
1579  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
1580  *
1581  * which tells FUSE to copy in the requested area and retry the ioctl.
1582  * On the second round, the server has access to the structure and
1583  * from that it can tell what to look for next, so on the invocation,
1584  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
1585  *
1586  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
1587  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
1588  *
1589  * FUSE will copy both struct a and the pointed buffer from the
1590  * process doing the ioctl and retry ioctl with both struct a and the
1591  * buffer.
1592  *
1593  * This time, FUSE server has everything it needs and completes ioctl
1594  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
1595  *
1596  * Copying data out works the same way.
1597  *
1598  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
1599  * automatically initializes in and out iovs by decoding @cmd with
1600  * _IOC_* macros and the server is not allowed to request RETRY.  This
1601  * limits ioctl data transfers to well-formed ioctls and is the forced
1602  * behavior for all FUSE servers.
1603  */
1604 static long fuse_file_do_ioctl(struct file *file, unsigned int cmd,
1605                                unsigned long arg, unsigned int flags)
1606 {
1607         struct inode *inode = file->f_dentry->d_inode;
1608         struct fuse_file *ff = file->private_data;
1609         struct fuse_conn *fc = get_fuse_conn(inode);
1610         struct fuse_ioctl_in inarg = {
1611                 .fh = ff->fh,
1612                 .cmd = cmd,
1613                 .arg = arg,
1614                 .flags = flags
1615         };
1616         struct fuse_ioctl_out outarg;
1617         struct fuse_req *req = NULL;
1618         struct page **pages = NULL;
1619         struct page *iov_page = NULL;
1620         struct iovec *in_iov = NULL, *out_iov = NULL;
1621         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
1622         size_t in_size, out_size, transferred;
1623         int err;
1624
1625         /* assume all the iovs returned by client always fits in a page */
1626         BUILD_BUG_ON(sizeof(struct iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
1627
1628         if (!fuse_allow_task(fc, current))
1629                 return -EACCES;
1630
1631         err = -EIO;
1632         if (is_bad_inode(inode))
1633                 goto out;
1634
1635         err = -ENOMEM;
1636         pages = kzalloc(sizeof(pages[0]) * FUSE_MAX_PAGES_PER_REQ, GFP_KERNEL);
1637         iov_page = alloc_page(GFP_KERNEL);
1638         if (!pages || !iov_page)
1639                 goto out;
1640
1641         /*
1642          * If restricted, initialize IO parameters as encoded in @cmd.
1643          * RETRY from server is not allowed.
1644          */
1645         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
1646                 struct iovec *iov = page_address(iov_page);
1647
1648                 iov->iov_base = (void __user *)arg;
1649                 iov->iov_len = _IOC_SIZE(cmd);
1650
1651                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
1652                         in_iov = iov;
1653                         in_iovs = 1;
1654                 }
1655
1656                 if (_IOC_DIR(cmd) & _IOC_READ) {
1657                         out_iov = iov;
1658                         out_iovs = 1;
1659                 }
1660         }
1661
1662  retry:
1663         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
1664         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
1665
1666         /*
1667          * Out data can be used either for actual out data or iovs,
1668          * make sure there always is at least one page.
1669          */
1670         out_size = max_t(size_t, out_size, PAGE_SIZE);
1671         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
1672
1673         /* make sure there are enough buffer pages and init request with them */
1674         err = -ENOMEM;
1675         if (max_pages > FUSE_MAX_PAGES_PER_REQ)
1676                 goto out;
1677         while (num_pages < max_pages) {
1678                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
1679                 if (!pages[num_pages])
1680                         goto out;
1681                 num_pages++;
1682         }
1683
1684         req = fuse_get_req(fc);
1685         if (IS_ERR(req)) {
1686                 err = PTR_ERR(req);
1687                 req = NULL;
1688                 goto out;
1689         }
1690         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
1691         req->num_pages = num_pages;
1692
1693         /* okay, let's send it to the client */
1694         req->in.h.opcode = FUSE_IOCTL;
1695         req->in.h.nodeid = get_node_id(inode);
1696         req->in.numargs = 1;
1697         req->in.args[0].size = sizeof(inarg);
1698         req->in.args[0].value = &inarg;
1699         if (in_size) {
1700                 req->in.numargs++;
1701                 req->in.args[1].size = in_size;
1702                 req->in.argpages = 1;
1703
1704                 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
1705                                            false);
1706                 if (err)
1707                         goto out;
1708         }
1709
1710         req->out.numargs = 2;
1711         req->out.args[0].size = sizeof(outarg);
1712         req->out.args[0].value = &outarg;
1713         req->out.args[1].size = out_size;
1714         req->out.argpages = 1;
1715         req->out.argvar = 1;
1716
1717         fuse_request_send(fc, req);
1718         err = req->out.h.error;
1719         transferred = req->out.args[1].size;
1720         fuse_put_request(fc, req);
1721         req = NULL;
1722         if (err)
1723                 goto out;
1724
1725         /* did it ask for retry? */
1726         if (outarg.flags & FUSE_IOCTL_RETRY) {
1727                 char *vaddr;
1728
1729                 /* no retry if in restricted mode */
1730                 err = -EIO;
1731                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
1732                         goto out;
1733
1734                 in_iovs = outarg.in_iovs;
1735                 out_iovs = outarg.out_iovs;
1736
1737                 /*
1738                  * Make sure things are in boundary, separate checks
1739                  * are to protect against overflow.
1740                  */
1741                 err = -ENOMEM;
1742                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
1743                     out_iovs > FUSE_IOCTL_MAX_IOV ||
1744                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
1745                         goto out;
1746
1747                 err = -EIO;
1748                 if ((in_iovs + out_iovs) * sizeof(struct iovec) != transferred)
1749                         goto out;
1750
1751                 /* okay, copy in iovs and retry */
1752                 vaddr = kmap_atomic(pages[0], KM_USER0);
1753                 memcpy(page_address(iov_page), vaddr, transferred);
1754                 kunmap_atomic(vaddr, KM_USER0);
1755
1756                 in_iov = page_address(iov_page);
1757                 out_iov = in_iov + in_iovs;
1758
1759                 goto retry;
1760         }
1761
1762         err = -EIO;
1763         if (transferred > inarg.out_size)
1764                 goto out;
1765
1766         err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
1767  out:
1768         if (req)
1769                 fuse_put_request(fc, req);
1770         if (iov_page)
1771                 __free_page(iov_page);
1772         while (num_pages)
1773                 __free_page(pages[--num_pages]);
1774         kfree(pages);
1775
1776         return err ? err : outarg.result;
1777 }
1778
1779 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
1780                             unsigned long arg)
1781 {
1782         return fuse_file_do_ioctl(file, cmd, arg, 0);
1783 }
1784
1785 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
1786                                    unsigned long arg)
1787 {
1788         return fuse_file_do_ioctl(file, cmd, arg, FUSE_IOCTL_COMPAT);
1789 }
1790
1791 /*
1792  * All files which have been polled are linked to RB tree
1793  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
1794  * find the matching one.
1795  */
1796 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
1797                                               struct rb_node **parent_out)
1798 {
1799         struct rb_node **link = &fc->polled_files.rb_node;
1800         struct rb_node *last = NULL;
1801
1802         while (*link) {
1803                 struct fuse_file *ff;
1804
1805                 last = *link;
1806                 ff = rb_entry(last, struct fuse_file, polled_node);
1807
1808                 if (kh < ff->kh)
1809                         link = &last->rb_left;
1810                 else if (kh > ff->kh)
1811                         link = &last->rb_right;
1812                 else
1813                         return link;
1814         }
1815
1816         if (parent_out)
1817                 *parent_out = last;
1818         return link;
1819 }
1820
1821 /*
1822  * The file is about to be polled.  Make sure it's on the polled_files
1823  * RB tree.  Note that files once added to the polled_files tree are
1824  * not removed before the file is released.  This is because a file
1825  * polled once is likely to be polled again.
1826  */
1827 static void fuse_register_polled_file(struct fuse_conn *fc,
1828                                       struct fuse_file *ff)
1829 {
1830         spin_lock(&fc->lock);
1831         if (RB_EMPTY_NODE(&ff->polled_node)) {
1832                 struct rb_node **link, *parent;
1833
1834                 link = fuse_find_polled_node(fc, ff->kh, &parent);
1835                 BUG_ON(*link);
1836                 rb_link_node(&ff->polled_node, parent, link);
1837                 rb_insert_color(&ff->polled_node, &fc->polled_files);
1838         }
1839         spin_unlock(&fc->lock);
1840 }
1841
1842 static unsigned fuse_file_poll(struct file *file, poll_table *wait)
1843 {
1844         struct inode *inode = file->f_dentry->d_inode;
1845         struct fuse_file *ff = file->private_data;
1846         struct fuse_conn *fc = get_fuse_conn(inode);
1847         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
1848         struct fuse_poll_out outarg;
1849         struct fuse_req *req;
1850         int err;
1851
1852         if (fc->no_poll)
1853                 return DEFAULT_POLLMASK;
1854
1855         poll_wait(file, &ff->poll_wait, wait);
1856
1857         /*
1858          * Ask for notification iff there's someone waiting for it.
1859          * The client may ignore the flag and always notify.
1860          */
1861         if (waitqueue_active(&ff->poll_wait)) {
1862                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
1863                 fuse_register_polled_file(fc, ff);
1864         }
1865
1866         req = fuse_get_req(fc);
1867         if (IS_ERR(req))
1868                 return PTR_ERR(req);
1869
1870         req->in.h.opcode = FUSE_POLL;
1871         req->in.h.nodeid = get_node_id(inode);
1872         req->in.numargs = 1;
1873         req->in.args[0].size = sizeof(inarg);
1874         req->in.args[0].value = &inarg;
1875         req->out.numargs = 1;
1876         req->out.args[0].size = sizeof(outarg);
1877         req->out.args[0].value = &outarg;
1878         fuse_request_send(fc, req);
1879         err = req->out.h.error;
1880         fuse_put_request(fc, req);
1881
1882         if (!err)
1883                 return outarg.revents;
1884         if (err == -ENOSYS) {
1885                 fc->no_poll = 1;
1886                 return DEFAULT_POLLMASK;
1887         }
1888         return POLLERR;
1889 }
1890
1891 /*
1892  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
1893  * wakes up the poll waiters.
1894  */
1895 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
1896                             struct fuse_notify_poll_wakeup_out *outarg)
1897 {
1898         u64 kh = outarg->kh;
1899         struct rb_node **link;
1900
1901         spin_lock(&fc->lock);
1902
1903         link = fuse_find_polled_node(fc, kh, NULL);
1904         if (*link) {
1905                 struct fuse_file *ff;
1906
1907                 ff = rb_entry(*link, struct fuse_file, polled_node);
1908                 wake_up_interruptible_sync(&ff->poll_wait);
1909         }
1910
1911         spin_unlock(&fc->lock);
1912         return 0;
1913 }
1914
1915 static const struct file_operations fuse_file_operations = {
1916         .llseek         = fuse_file_llseek,
1917         .read           = do_sync_read,
1918         .aio_read       = fuse_file_aio_read,
1919         .write          = do_sync_write,
1920         .aio_write      = fuse_file_aio_write,
1921         .mmap           = fuse_file_mmap,
1922         .open           = fuse_open,
1923         .flush          = fuse_flush,
1924         .release        = fuse_release,
1925         .fsync          = fuse_fsync,
1926         .lock           = fuse_file_lock,
1927         .flock          = fuse_file_flock,
1928         .splice_read    = generic_file_splice_read,
1929         .unlocked_ioctl = fuse_file_ioctl,
1930         .compat_ioctl   = fuse_file_compat_ioctl,
1931         .poll           = fuse_file_poll,
1932 };
1933
1934 static const struct file_operations fuse_direct_io_file_operations = {
1935         .llseek         = fuse_file_llseek,
1936         .read           = fuse_direct_read,
1937         .write          = fuse_direct_write,
1938         .mmap           = fuse_direct_mmap,
1939         .open           = fuse_open,
1940         .flush          = fuse_flush,
1941         .release        = fuse_release,
1942         .fsync          = fuse_fsync,
1943         .lock           = fuse_file_lock,
1944         .flock          = fuse_file_flock,
1945         .unlocked_ioctl = fuse_file_ioctl,
1946         .compat_ioctl   = fuse_file_compat_ioctl,
1947         .poll           = fuse_file_poll,
1948         /* no splice_read */
1949 };
1950
1951 static const struct address_space_operations fuse_file_aops  = {
1952         .readpage       = fuse_readpage,
1953         .writepage      = fuse_writepage,
1954         .launder_page   = fuse_launder_page,
1955         .write_begin    = fuse_write_begin,
1956         .write_end      = fuse_write_end,
1957         .readpages      = fuse_readpages,
1958         .set_page_dirty = __set_page_dirty_nobuffers,
1959         .bmap           = fuse_bmap,
1960 };
1961
1962 void fuse_init_file_inode(struct inode *inode)
1963 {
1964         inode->i_fop = &fuse_file_operations;
1965         inode->i_data.a_ops = &fuse_file_aops;
1966 }