3be0301053547e92a38a1a6f775d31ac179ee469
[linux-3.10.git] / fs / fuse / file.c
1 /*
2   FUSE: Filesystem in Userspace
3   Copyright (C) 2001-2008  Miklos Szeredi <miklos@szeredi.hu>
4
5   This program can be distributed under the terms of the GNU GPL.
6   See the file COPYING.
7 */
8
9 #include "fuse_i.h"
10
11 #include <linux/pagemap.h>
12 #include <linux/slab.h>
13 #include <linux/kernel.h>
14 #include <linux/sched.h>
15
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
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 static 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                 struct inode *inode = req->misc.release.path.dentry->d_inode;
98                 struct fuse_conn *fc = get_fuse_conn(inode);
99                 req->end = fuse_release_end;
100                 fuse_request_send_background(fc, req);
101                 kfree(ff);
102         }
103 }
104
105 void fuse_finish_open(struct inode *inode, struct file *file,
106                       struct fuse_file *ff, struct fuse_open_out *outarg)
107 {
108         if (outarg->open_flags & FOPEN_DIRECT_IO)
109                 file->f_op = &fuse_direct_io_file_operations;
110         if (!(outarg->open_flags & FOPEN_KEEP_CACHE))
111                 invalidate_inode_pages2(inode->i_mapping);
112         if (outarg->open_flags & FOPEN_NONSEEKABLE)
113                 nonseekable_open(inode, file);
114         ff->fh = outarg->fh;
115         ff->nodeid = get_node_id(inode);
116         file->private_data = fuse_file_get(ff);
117 }
118
119 int fuse_open_common(struct inode *inode, struct file *file, int isdir)
120 {
121         struct fuse_conn *fc = get_fuse_conn(inode);
122         struct fuse_open_out outarg;
123         struct fuse_file *ff;
124         int err;
125
126         /* VFS checks this, but only _after_ ->open() */
127         if (file->f_flags & O_DIRECT)
128                 return -EINVAL;
129
130         err = generic_file_open(inode, file);
131         if (err)
132                 return err;
133
134         ff = fuse_file_alloc(fc);
135         if (!ff)
136                 return -ENOMEM;
137
138         err = fuse_send_open(inode, file, isdir, &outarg);
139         if (err)
140                 fuse_file_free(ff);
141         else {
142                 if (isdir)
143                         outarg.open_flags &= ~FOPEN_DIRECT_IO;
144                 fuse_finish_open(inode, file, ff, &outarg);
145         }
146
147         return err;
148 }
149
150 void fuse_release_fill(struct fuse_file *ff, u64 nodeid, int flags, int opcode)
151 {
152         struct fuse_req *req = ff->reserved_req;
153         struct fuse_release_in *inarg = &req->misc.release.in;
154
155         inarg->fh = ff->fh;
156         inarg->flags = flags;
157         req->in.h.opcode = opcode;
158         req->in.h.nodeid = nodeid;
159         req->in.numargs = 1;
160         req->in.args[0].size = sizeof(struct fuse_release_in);
161         req->in.args[0].value = inarg;
162 }
163
164 int fuse_release_common(struct inode *inode, struct file *file, int isdir)
165 {
166         struct fuse_conn *fc;
167         struct fuse_file *ff;
168         struct fuse_req *req;
169
170         ff = file->private_data;
171         if (unlikely(!ff))
172                 return 0;       /* return value is ignored by VFS */
173
174         fc = get_fuse_conn(inode);
175         req = ff->reserved_req;
176
177         fuse_release_fill(ff, get_node_id(inode), file->f_flags,
178                           isdir ? FUSE_RELEASEDIR : FUSE_RELEASE);
179
180         /* Hold vfsmount and dentry until release is finished */
181         path_get(&file->f_path);
182         req->misc.release.path = file->f_path;
183
184         spin_lock(&fc->lock);
185         list_del(&ff->write_entry);
186         if (!RB_EMPTY_NODE(&ff->polled_node))
187                 rb_erase(&ff->polled_node, &fc->polled_files);
188         spin_unlock(&fc->lock);
189
190         wake_up_interruptible_sync(&ff->poll_wait);
191         /*
192          * Normally this will send the RELEASE request, however if
193          * some asynchronous READ or WRITE requests are outstanding,
194          * the sending will be delayed.
195          */
196         fuse_file_put(ff);
197         return 0;
198 }
199
200 static int fuse_open(struct inode *inode, struct file *file)
201 {
202         return fuse_open_common(inode, file, 0);
203 }
204
205 static int fuse_release(struct inode *inode, struct file *file)
206 {
207         return fuse_release_common(inode, file, 0);
208 }
209
210 /*
211  * Scramble the ID space with XTEA, so that the value of the files_struct
212  * pointer is not exposed to userspace.
213  */
214 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
215 {
216         u32 *k = fc->scramble_key;
217         u64 v = (unsigned long) id;
218         u32 v0 = v;
219         u32 v1 = v >> 32;
220         u32 sum = 0;
221         int i;
222
223         for (i = 0; i < 32; i++) {
224                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
225                 sum += 0x9E3779B9;
226                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
227         }
228
229         return (u64) v0 + ((u64) v1 << 32);
230 }
231
232 /*
233  * Check if page is under writeback
234  *
235  * This is currently done by walking the list of writepage requests
236  * for the inode, which can be pretty inefficient.
237  */
238 static bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
239 {
240         struct fuse_conn *fc = get_fuse_conn(inode);
241         struct fuse_inode *fi = get_fuse_inode(inode);
242         struct fuse_req *req;
243         bool found = false;
244
245         spin_lock(&fc->lock);
246         list_for_each_entry(req, &fi->writepages, writepages_entry) {
247                 pgoff_t curr_index;
248
249                 BUG_ON(req->inode != inode);
250                 curr_index = req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
251                 if (curr_index == index) {
252                         found = true;
253                         break;
254                 }
255         }
256         spin_unlock(&fc->lock);
257
258         return found;
259 }
260
261 /*
262  * Wait for page writeback to be completed.
263  *
264  * Since fuse doesn't rely on the VM writeback tracking, this has to
265  * use some other means.
266  */
267 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
268 {
269         struct fuse_inode *fi = get_fuse_inode(inode);
270
271         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
272         return 0;
273 }
274
275 static int fuse_flush(struct file *file, fl_owner_t id)
276 {
277         struct inode *inode = file->f_path.dentry->d_inode;
278         struct fuse_conn *fc = get_fuse_conn(inode);
279         struct fuse_file *ff = file->private_data;
280         struct fuse_req *req;
281         struct fuse_flush_in inarg;
282         int err;
283
284         if (is_bad_inode(inode))
285                 return -EIO;
286
287         if (fc->no_flush)
288                 return 0;
289
290         req = fuse_get_req_nofail(fc, file);
291         memset(&inarg, 0, sizeof(inarg));
292         inarg.fh = ff->fh;
293         inarg.lock_owner = fuse_lock_owner_id(fc, id);
294         req->in.h.opcode = FUSE_FLUSH;
295         req->in.h.nodeid = get_node_id(inode);
296         req->in.numargs = 1;
297         req->in.args[0].size = sizeof(inarg);
298         req->in.args[0].value = &inarg;
299         req->force = 1;
300         fuse_request_send(fc, req);
301         err = req->out.h.error;
302         fuse_put_request(fc, req);
303         if (err == -ENOSYS) {
304                 fc->no_flush = 1;
305                 err = 0;
306         }
307         return err;
308 }
309
310 /*
311  * Wait for all pending writepages on the inode to finish.
312  *
313  * This is currently done by blocking further writes with FUSE_NOWRITE
314  * and waiting for all sent writes to complete.
315  *
316  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
317  * could conflict with truncation.
318  */
319 static void fuse_sync_writes(struct inode *inode)
320 {
321         fuse_set_nowrite(inode);
322         fuse_release_nowrite(inode);
323 }
324
325 int fuse_fsync_common(struct file *file, struct dentry *de, int datasync,
326                       int isdir)
327 {
328         struct inode *inode = de->d_inode;
329         struct fuse_conn *fc = get_fuse_conn(inode);
330         struct fuse_file *ff = file->private_data;
331         struct fuse_req *req;
332         struct fuse_fsync_in inarg;
333         int err;
334
335         if (is_bad_inode(inode))
336                 return -EIO;
337
338         if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
339                 return 0;
340
341         /*
342          * Start writeback against all dirty pages of the inode, then
343          * wait for all outstanding writes, before sending the FSYNC
344          * request.
345          */
346         err = write_inode_now(inode, 0);
347         if (err)
348                 return err;
349
350         fuse_sync_writes(inode);
351
352         req = fuse_get_req(fc);
353         if (IS_ERR(req))
354                 return PTR_ERR(req);
355
356         memset(&inarg, 0, sizeof(inarg));
357         inarg.fh = ff->fh;
358         inarg.fsync_flags = datasync ? 1 : 0;
359         req->in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
360         req->in.h.nodeid = get_node_id(inode);
361         req->in.numargs = 1;
362         req->in.args[0].size = sizeof(inarg);
363         req->in.args[0].value = &inarg;
364         fuse_request_send(fc, req);
365         err = req->out.h.error;
366         fuse_put_request(fc, req);
367         if (err == -ENOSYS) {
368                 if (isdir)
369                         fc->no_fsyncdir = 1;
370                 else
371                         fc->no_fsync = 1;
372                 err = 0;
373         }
374         return err;
375 }
376
377 static int fuse_fsync(struct file *file, struct dentry *de, int datasync)
378 {
379         return fuse_fsync_common(file, de, datasync, 0);
380 }
381
382 void fuse_read_fill(struct fuse_req *req, struct file *file,
383                     struct inode *inode, loff_t pos, size_t count, int opcode)
384 {
385         struct fuse_read_in *inarg = &req->misc.read.in;
386         struct fuse_file *ff = file->private_data;
387
388         inarg->fh = ff->fh;
389         inarg->offset = pos;
390         inarg->size = count;
391         inarg->flags = file->f_flags;
392         req->in.h.opcode = opcode;
393         req->in.h.nodeid = get_node_id(inode);
394         req->in.numargs = 1;
395         req->in.args[0].size = sizeof(struct fuse_read_in);
396         req->in.args[0].value = inarg;
397         req->out.argvar = 1;
398         req->out.numargs = 1;
399         req->out.args[0].size = count;
400 }
401
402 static size_t fuse_send_read(struct fuse_req *req, struct file *file,
403                              struct inode *inode, loff_t pos, size_t count,
404                              fl_owner_t owner)
405 {
406         struct fuse_conn *fc = get_fuse_conn(inode);
407
408         fuse_read_fill(req, file, inode, pos, count, FUSE_READ);
409         if (owner != NULL) {
410                 struct fuse_read_in *inarg = &req->misc.read.in;
411
412                 inarg->read_flags |= FUSE_READ_LOCKOWNER;
413                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
414         }
415         fuse_request_send(fc, req);
416         return req->out.args[0].size;
417 }
418
419 static void fuse_read_update_size(struct inode *inode, loff_t size,
420                                   u64 attr_ver)
421 {
422         struct fuse_conn *fc = get_fuse_conn(inode);
423         struct fuse_inode *fi = get_fuse_inode(inode);
424
425         spin_lock(&fc->lock);
426         if (attr_ver == fi->attr_version && size < inode->i_size) {
427                 fi->attr_version = ++fc->attr_version;
428                 i_size_write(inode, size);
429         }
430         spin_unlock(&fc->lock);
431 }
432
433 static int fuse_readpage(struct file *file, struct page *page)
434 {
435         struct inode *inode = page->mapping->host;
436         struct fuse_conn *fc = get_fuse_conn(inode);
437         struct fuse_req *req;
438         size_t num_read;
439         loff_t pos = page_offset(page);
440         size_t count = PAGE_CACHE_SIZE;
441         u64 attr_ver;
442         int err;
443
444         err = -EIO;
445         if (is_bad_inode(inode))
446                 goto out;
447
448         /*
449          * Page writeback can extend beyond the liftime of the
450          * page-cache page, so make sure we read a properly synced
451          * page.
452          */
453         fuse_wait_on_page_writeback(inode, page->index);
454
455         req = fuse_get_req(fc);
456         err = PTR_ERR(req);
457         if (IS_ERR(req))
458                 goto out;
459
460         attr_ver = fuse_get_attr_version(fc);
461
462         req->out.page_zeroing = 1;
463         req->out.argpages = 1;
464         req->num_pages = 1;
465         req->pages[0] = page;
466         num_read = fuse_send_read(req, file, inode, pos, count, NULL);
467         err = req->out.h.error;
468         fuse_put_request(fc, req);
469
470         if (!err) {
471                 /*
472                  * Short read means EOF.  If file size is larger, truncate it
473                  */
474                 if (num_read < count)
475                         fuse_read_update_size(inode, pos + num_read, attr_ver);
476
477                 SetPageUptodate(page);
478         }
479
480         fuse_invalidate_attr(inode); /* atime changed */
481  out:
482         unlock_page(page);
483         return err;
484 }
485
486 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
487 {
488         int i;
489         size_t count = req->misc.read.in.size;
490         size_t num_read = req->out.args[0].size;
491         struct inode *inode = req->pages[0]->mapping->host;
492
493         /*
494          * Short read means EOF.  If file size is larger, truncate it
495          */
496         if (!req->out.h.error && num_read < count) {
497                 loff_t pos = page_offset(req->pages[0]) + num_read;
498                 fuse_read_update_size(inode, pos, req->misc.read.attr_ver);
499         }
500
501         fuse_invalidate_attr(inode); /* atime changed */
502
503         for (i = 0; i < req->num_pages; i++) {
504                 struct page *page = req->pages[i];
505                 if (!req->out.h.error)
506                         SetPageUptodate(page);
507                 else
508                         SetPageError(page);
509                 unlock_page(page);
510         }
511         if (req->ff)
512                 fuse_file_put(req->ff);
513 }
514
515 static void fuse_send_readpages(struct fuse_req *req, struct file *file,
516                                 struct inode *inode)
517 {
518         struct fuse_conn *fc = get_fuse_conn(inode);
519         loff_t pos = page_offset(req->pages[0]);
520         size_t count = req->num_pages << PAGE_CACHE_SHIFT;
521
522         req->out.argpages = 1;
523         req->out.page_zeroing = 1;
524         fuse_read_fill(req, file, inode, pos, count, FUSE_READ);
525         req->misc.read.attr_ver = fuse_get_attr_version(fc);
526         if (fc->async_read) {
527                 struct fuse_file *ff = file->private_data;
528                 req->ff = fuse_file_get(ff);
529                 req->end = fuse_readpages_end;
530                 fuse_request_send_background(fc, req);
531         } else {
532                 fuse_request_send(fc, req);
533                 fuse_readpages_end(fc, req);
534                 fuse_put_request(fc, req);
535         }
536 }
537
538 struct fuse_fill_data {
539         struct fuse_req *req;
540         struct file *file;
541         struct inode *inode;
542 };
543
544 static int fuse_readpages_fill(void *_data, struct page *page)
545 {
546         struct fuse_fill_data *data = _data;
547         struct fuse_req *req = data->req;
548         struct inode *inode = data->inode;
549         struct fuse_conn *fc = get_fuse_conn(inode);
550
551         fuse_wait_on_page_writeback(inode, page->index);
552
553         if (req->num_pages &&
554             (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
555              (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
556              req->pages[req->num_pages - 1]->index + 1 != page->index)) {
557                 fuse_send_readpages(req, data->file, inode);
558                 data->req = req = fuse_get_req(fc);
559                 if (IS_ERR(req)) {
560                         unlock_page(page);
561                         return PTR_ERR(req);
562                 }
563         }
564         req->pages[req->num_pages] = page;
565         req->num_pages++;
566         return 0;
567 }
568
569 static int fuse_readpages(struct file *file, struct address_space *mapping,
570                           struct list_head *pages, unsigned nr_pages)
571 {
572         struct inode *inode = mapping->host;
573         struct fuse_conn *fc = get_fuse_conn(inode);
574         struct fuse_fill_data data;
575         int err;
576
577         err = -EIO;
578         if (is_bad_inode(inode))
579                 goto out;
580
581         data.file = file;
582         data.inode = inode;
583         data.req = fuse_get_req(fc);
584         err = PTR_ERR(data.req);
585         if (IS_ERR(data.req))
586                 goto out;
587
588         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
589         if (!err) {
590                 if (data.req->num_pages)
591                         fuse_send_readpages(data.req, file, inode);
592                 else
593                         fuse_put_request(fc, data.req);
594         }
595 out:
596         return err;
597 }
598
599 static ssize_t fuse_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
600                                   unsigned long nr_segs, loff_t pos)
601 {
602         struct inode *inode = iocb->ki_filp->f_mapping->host;
603
604         if (pos + iov_length(iov, nr_segs) > i_size_read(inode)) {
605                 int err;
606                 /*
607                  * If trying to read past EOF, make sure the i_size
608                  * attribute is up-to-date.
609                  */
610                 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
611                 if (err)
612                         return err;
613         }
614
615         return generic_file_aio_read(iocb, iov, nr_segs, pos);
616 }
617
618 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
619                             struct inode *inode, loff_t pos, size_t count)
620 {
621         struct fuse_conn *fc = get_fuse_conn(inode);
622         struct fuse_write_in *inarg = &req->misc.write.in;
623         struct fuse_write_out *outarg = &req->misc.write.out;
624
625         inarg->fh = ff->fh;
626         inarg->offset = pos;
627         inarg->size = count;
628         req->in.h.opcode = FUSE_WRITE;
629         req->in.h.nodeid = get_node_id(inode);
630         req->in.numargs = 2;
631         if (fc->minor < 9)
632                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
633         else
634                 req->in.args[0].size = sizeof(struct fuse_write_in);
635         req->in.args[0].value = inarg;
636         req->in.args[1].size = count;
637         req->out.numargs = 1;
638         req->out.args[0].size = sizeof(struct fuse_write_out);
639         req->out.args[0].value = outarg;
640 }
641
642 static size_t fuse_send_write(struct fuse_req *req, struct file *file,
643                               struct inode *inode, loff_t pos, size_t count,
644                               fl_owner_t owner)
645 {
646         struct fuse_conn *fc = get_fuse_conn(inode);
647         struct fuse_write_in *inarg = &req->misc.write.in;
648
649         fuse_write_fill(req, file->private_data, inode, pos, count);
650         inarg->flags = file->f_flags;
651         if (owner != NULL) {
652                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
653                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
654         }
655         fuse_request_send(fc, req);
656         return req->misc.write.out.size;
657 }
658
659 static int fuse_write_begin(struct file *file, struct address_space *mapping,
660                         loff_t pos, unsigned len, unsigned flags,
661                         struct page **pagep, void **fsdata)
662 {
663         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
664
665         *pagep = grab_cache_page_write_begin(mapping, index, flags);
666         if (!*pagep)
667                 return -ENOMEM;
668         return 0;
669 }
670
671 static void fuse_write_update_size(struct inode *inode, loff_t pos)
672 {
673         struct fuse_conn *fc = get_fuse_conn(inode);
674         struct fuse_inode *fi = get_fuse_inode(inode);
675
676         spin_lock(&fc->lock);
677         fi->attr_version = ++fc->attr_version;
678         if (pos > inode->i_size)
679                 i_size_write(inode, pos);
680         spin_unlock(&fc->lock);
681 }
682
683 static int fuse_buffered_write(struct file *file, struct inode *inode,
684                                loff_t pos, unsigned count, struct page *page)
685 {
686         int err;
687         size_t nres;
688         struct fuse_conn *fc = get_fuse_conn(inode);
689         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
690         struct fuse_req *req;
691
692         if (is_bad_inode(inode))
693                 return -EIO;
694
695         /*
696          * Make sure writepages on the same page are not mixed up with
697          * plain writes.
698          */
699         fuse_wait_on_page_writeback(inode, page->index);
700
701         req = fuse_get_req(fc);
702         if (IS_ERR(req))
703                 return PTR_ERR(req);
704
705         req->in.argpages = 1;
706         req->num_pages = 1;
707         req->pages[0] = page;
708         req->page_offset = offset;
709         nres = fuse_send_write(req, file, inode, pos, count, NULL);
710         err = req->out.h.error;
711         fuse_put_request(fc, req);
712         if (!err && !nres)
713                 err = -EIO;
714         if (!err) {
715                 pos += nres;
716                 fuse_write_update_size(inode, pos);
717                 if (count == PAGE_CACHE_SIZE)
718                         SetPageUptodate(page);
719         }
720         fuse_invalidate_attr(inode);
721         return err ? err : nres;
722 }
723
724 static int fuse_write_end(struct file *file, struct address_space *mapping,
725                         loff_t pos, unsigned len, unsigned copied,
726                         struct page *page, void *fsdata)
727 {
728         struct inode *inode = mapping->host;
729         int res = 0;
730
731         if (copied)
732                 res = fuse_buffered_write(file, inode, pos, copied, page);
733
734         unlock_page(page);
735         page_cache_release(page);
736         return res;
737 }
738
739 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
740                                     struct inode *inode, loff_t pos,
741                                     size_t count)
742 {
743         size_t res;
744         unsigned offset;
745         unsigned i;
746
747         for (i = 0; i < req->num_pages; i++)
748                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
749
750         res = fuse_send_write(req, file, inode, pos, count, NULL);
751
752         offset = req->page_offset;
753         count = res;
754         for (i = 0; i < req->num_pages; i++) {
755                 struct page *page = req->pages[i];
756
757                 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
758                         SetPageUptodate(page);
759
760                 if (count > PAGE_CACHE_SIZE - offset)
761                         count -= PAGE_CACHE_SIZE - offset;
762                 else
763                         count = 0;
764                 offset = 0;
765
766                 unlock_page(page);
767                 page_cache_release(page);
768         }
769
770         return res;
771 }
772
773 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
774                                struct address_space *mapping,
775                                struct iov_iter *ii, loff_t pos)
776 {
777         struct fuse_conn *fc = get_fuse_conn(mapping->host);
778         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
779         size_t count = 0;
780         int err;
781
782         req->in.argpages = 1;
783         req->page_offset = offset;
784
785         do {
786                 size_t tmp;
787                 struct page *page;
788                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
789                 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
790                                      iov_iter_count(ii));
791
792                 bytes = min_t(size_t, bytes, fc->max_write - count);
793
794  again:
795                 err = -EFAULT;
796                 if (iov_iter_fault_in_readable(ii, bytes))
797                         break;
798
799                 err = -ENOMEM;
800                 page = grab_cache_page_write_begin(mapping, index, 0);
801                 if (!page)
802                         break;
803
804                 pagefault_disable();
805                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
806                 pagefault_enable();
807                 flush_dcache_page(page);
808
809                 if (!tmp) {
810                         unlock_page(page);
811                         page_cache_release(page);
812                         bytes = min(bytes, iov_iter_single_seg_count(ii));
813                         goto again;
814                 }
815
816                 err = 0;
817                 req->pages[req->num_pages] = page;
818                 req->num_pages++;
819
820                 iov_iter_advance(ii, tmp);
821                 count += tmp;
822                 pos += tmp;
823                 offset += tmp;
824                 if (offset == PAGE_CACHE_SIZE)
825                         offset = 0;
826
827                 if (!fc->big_writes)
828                         break;
829         } while (iov_iter_count(ii) && count < fc->max_write &&
830                  req->num_pages < FUSE_MAX_PAGES_PER_REQ && offset == 0);
831
832         return count > 0 ? count : err;
833 }
834
835 static ssize_t fuse_perform_write(struct file *file,
836                                   struct address_space *mapping,
837                                   struct iov_iter *ii, loff_t pos)
838 {
839         struct inode *inode = mapping->host;
840         struct fuse_conn *fc = get_fuse_conn(inode);
841         int err = 0;
842         ssize_t res = 0;
843
844         if (is_bad_inode(inode))
845                 return -EIO;
846
847         do {
848                 struct fuse_req *req;
849                 ssize_t count;
850
851                 req = fuse_get_req(fc);
852                 if (IS_ERR(req)) {
853                         err = PTR_ERR(req);
854                         break;
855                 }
856
857                 count = fuse_fill_write_pages(req, mapping, ii, pos);
858                 if (count <= 0) {
859                         err = count;
860                 } else {
861                         size_t num_written;
862
863                         num_written = fuse_send_write_pages(req, file, inode,
864                                                             pos, count);
865                         err = req->out.h.error;
866                         if (!err) {
867                                 res += num_written;
868                                 pos += num_written;
869
870                                 /* break out of the loop on short write */
871                                 if (num_written != count)
872                                         err = -EIO;
873                         }
874                 }
875                 fuse_put_request(fc, req);
876         } while (!err && iov_iter_count(ii));
877
878         if (res > 0)
879                 fuse_write_update_size(inode, pos);
880
881         fuse_invalidate_attr(inode);
882
883         return res > 0 ? res : err;
884 }
885
886 static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
887                                    unsigned long nr_segs, loff_t pos)
888 {
889         struct file *file = iocb->ki_filp;
890         struct address_space *mapping = file->f_mapping;
891         size_t count = 0;
892         ssize_t written = 0;
893         struct inode *inode = mapping->host;
894         ssize_t err;
895         struct iov_iter i;
896
897         WARN_ON(iocb->ki_pos != pos);
898
899         err = generic_segment_checks(iov, &nr_segs, &count, VERIFY_READ);
900         if (err)
901                 return err;
902
903         mutex_lock(&inode->i_mutex);
904         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
905
906         /* We can write back this queue in page reclaim */
907         current->backing_dev_info = mapping->backing_dev_info;
908
909         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
910         if (err)
911                 goto out;
912
913         if (count == 0)
914                 goto out;
915
916         err = file_remove_suid(file);
917         if (err)
918                 goto out;
919
920         file_update_time(file);
921
922         iov_iter_init(&i, iov, nr_segs, count, 0);
923         written = fuse_perform_write(file, mapping, &i, pos);
924         if (written >= 0)
925                 iocb->ki_pos = pos + written;
926
927 out:
928         current->backing_dev_info = NULL;
929         mutex_unlock(&inode->i_mutex);
930
931         return written ? written : err;
932 }
933
934 static void fuse_release_user_pages(struct fuse_req *req, int write)
935 {
936         unsigned i;
937
938         for (i = 0; i < req->num_pages; i++) {
939                 struct page *page = req->pages[i];
940                 if (write)
941                         set_page_dirty_lock(page);
942                 put_page(page);
943         }
944 }
945
946 static int fuse_get_user_pages(struct fuse_req *req, const char __user *buf,
947                                size_t *nbytesp, int write)
948 {
949         size_t nbytes = *nbytesp;
950         unsigned long user_addr = (unsigned long) buf;
951         unsigned offset = user_addr & ~PAGE_MASK;
952         int npages;
953
954         /* Special case for kernel I/O: can copy directly into the buffer */
955         if (segment_eq(get_fs(), KERNEL_DS)) {
956                 if (write)
957                         req->in.args[1].value = (void *) user_addr;
958                 else
959                         req->out.args[0].value = (void *) user_addr;
960
961                 return 0;
962         }
963
964         nbytes = min_t(size_t, nbytes, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
965         npages = (nbytes + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
966         npages = clamp(npages, 1, FUSE_MAX_PAGES_PER_REQ);
967         down_read(&current->mm->mmap_sem);
968         npages = get_user_pages(current, current->mm, user_addr, npages, !write,
969                                 0, req->pages, NULL);
970         up_read(&current->mm->mmap_sem);
971         if (npages < 0)
972                 return npages;
973
974         req->num_pages = npages;
975         req->page_offset = offset;
976
977         if (write)
978                 req->in.argpages = 1;
979         else
980                 req->out.argpages = 1;
981
982         nbytes = (req->num_pages << PAGE_SHIFT) - req->page_offset;
983         *nbytesp = min(*nbytesp, nbytes);
984
985         return 0;
986 }
987
988 static ssize_t fuse_direct_io(struct file *file, const char __user *buf,
989                               size_t count, loff_t *ppos, int write)
990 {
991         struct inode *inode = file->f_path.dentry->d_inode;
992         struct fuse_conn *fc = get_fuse_conn(inode);
993         size_t nmax = write ? fc->max_write : fc->max_read;
994         loff_t pos = *ppos;
995         ssize_t res = 0;
996         struct fuse_req *req;
997
998         req = fuse_get_req(fc);
999         if (IS_ERR(req))
1000                 return PTR_ERR(req);
1001
1002         while (count) {
1003                 size_t nres;
1004                 size_t nbytes = min(count, nmax);
1005                 int err = fuse_get_user_pages(req, buf, &nbytes, write);
1006                 if (err) {
1007                         res = err;
1008                         break;
1009                 }
1010
1011                 if (write)
1012                         nres = fuse_send_write(req, file, inode, pos, nbytes,
1013                                                current->files);
1014                 else
1015                         nres = fuse_send_read(req, file, inode, pos, nbytes,
1016                                               current->files);
1017                 fuse_release_user_pages(req, !write);
1018                 if (req->out.h.error) {
1019                         if (!res)
1020                                 res = req->out.h.error;
1021                         break;
1022                 } else if (nres > nbytes) {
1023                         res = -EIO;
1024                         break;
1025                 }
1026                 count -= nres;
1027                 res += nres;
1028                 pos += nres;
1029                 buf += nres;
1030                 if (nres != nbytes)
1031                         break;
1032                 if (count) {
1033                         fuse_put_request(fc, req);
1034                         req = fuse_get_req(fc);
1035                         if (IS_ERR(req))
1036                                 break;
1037                 }
1038         }
1039         fuse_put_request(fc, req);
1040         if (res > 0)
1041                 *ppos = pos;
1042
1043         return res;
1044 }
1045
1046 static ssize_t fuse_direct_read(struct file *file, char __user *buf,
1047                                      size_t count, loff_t *ppos)
1048 {
1049         ssize_t res;
1050         struct inode *inode = file->f_path.dentry->d_inode;
1051
1052         if (is_bad_inode(inode))
1053                 return -EIO;
1054
1055         res = fuse_direct_io(file, buf, count, ppos, 0);
1056
1057         fuse_invalidate_attr(inode);
1058
1059         return res;
1060 }
1061
1062 static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
1063                                  size_t count, loff_t *ppos)
1064 {
1065         struct inode *inode = file->f_path.dentry->d_inode;
1066         ssize_t res;
1067
1068         if (is_bad_inode(inode))
1069                 return -EIO;
1070
1071         /* Don't allow parallel writes to the same file */
1072         mutex_lock(&inode->i_mutex);
1073         res = generic_write_checks(file, ppos, &count, 0);
1074         if (!res) {
1075                 res = fuse_direct_io(file, buf, count, ppos, 1);
1076                 if (res > 0)
1077                         fuse_write_update_size(inode, *ppos);
1078         }
1079         mutex_unlock(&inode->i_mutex);
1080
1081         fuse_invalidate_attr(inode);
1082
1083         return res;
1084 }
1085
1086 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1087 {
1088         __free_page(req->pages[0]);
1089         fuse_file_put(req->ff);
1090 }
1091
1092 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1093 {
1094         struct inode *inode = req->inode;
1095         struct fuse_inode *fi = get_fuse_inode(inode);
1096         struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
1097
1098         list_del(&req->writepages_entry);
1099         dec_bdi_stat(bdi, BDI_WRITEBACK);
1100         dec_zone_page_state(req->pages[0], NR_WRITEBACK_TEMP);
1101         bdi_writeout_inc(bdi);
1102         wake_up(&fi->page_waitq);
1103 }
1104
1105 /* Called under fc->lock, may release and reacquire it */
1106 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req)
1107 __releases(&fc->lock)
1108 __acquires(&fc->lock)
1109 {
1110         struct fuse_inode *fi = get_fuse_inode(req->inode);
1111         loff_t size = i_size_read(req->inode);
1112         struct fuse_write_in *inarg = &req->misc.write.in;
1113
1114         if (!fc->connected)
1115                 goto out_free;
1116
1117         if (inarg->offset + PAGE_CACHE_SIZE <= size) {
1118                 inarg->size = PAGE_CACHE_SIZE;
1119         } else if (inarg->offset < size) {
1120                 inarg->size = size & (PAGE_CACHE_SIZE - 1);
1121         } else {
1122                 /* Got truncated off completely */
1123                 goto out_free;
1124         }
1125
1126         req->in.args[1].size = inarg->size;
1127         fi->writectr++;
1128         fuse_request_send_background_locked(fc, req);
1129         return;
1130
1131  out_free:
1132         fuse_writepage_finish(fc, req);
1133         spin_unlock(&fc->lock);
1134         fuse_writepage_free(fc, req);
1135         fuse_put_request(fc, req);
1136         spin_lock(&fc->lock);
1137 }
1138
1139 /*
1140  * If fi->writectr is positive (no truncate or fsync going on) send
1141  * all queued writepage requests.
1142  *
1143  * Called with fc->lock
1144  */
1145 void fuse_flush_writepages(struct inode *inode)
1146 __releases(&fc->lock)
1147 __acquires(&fc->lock)
1148 {
1149         struct fuse_conn *fc = get_fuse_conn(inode);
1150         struct fuse_inode *fi = get_fuse_inode(inode);
1151         struct fuse_req *req;
1152
1153         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1154                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1155                 list_del_init(&req->list);
1156                 fuse_send_writepage(fc, req);
1157         }
1158 }
1159
1160 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1161 {
1162         struct inode *inode = req->inode;
1163         struct fuse_inode *fi = get_fuse_inode(inode);
1164
1165         mapping_set_error(inode->i_mapping, req->out.h.error);
1166         spin_lock(&fc->lock);
1167         fi->writectr--;
1168         fuse_writepage_finish(fc, req);
1169         spin_unlock(&fc->lock);
1170         fuse_writepage_free(fc, req);
1171 }
1172
1173 static int fuse_writepage_locked(struct page *page)
1174 {
1175         struct address_space *mapping = page->mapping;
1176         struct inode *inode = mapping->host;
1177         struct fuse_conn *fc = get_fuse_conn(inode);
1178         struct fuse_inode *fi = get_fuse_inode(inode);
1179         struct fuse_req *req;
1180         struct fuse_file *ff;
1181         struct page *tmp_page;
1182
1183         set_page_writeback(page);
1184
1185         req = fuse_request_alloc_nofs();
1186         if (!req)
1187                 goto err;
1188
1189         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1190         if (!tmp_page)
1191                 goto err_free;
1192
1193         spin_lock(&fc->lock);
1194         BUG_ON(list_empty(&fi->write_files));
1195         ff = list_entry(fi->write_files.next, struct fuse_file, write_entry);
1196         req->ff = fuse_file_get(ff);
1197         spin_unlock(&fc->lock);
1198
1199         fuse_write_fill(req, ff, inode, page_offset(page), 0);
1200
1201         copy_highpage(tmp_page, page);
1202         req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1203         req->in.argpages = 1;
1204         req->num_pages = 1;
1205         req->pages[0] = tmp_page;
1206         req->page_offset = 0;
1207         req->end = fuse_writepage_end;
1208         req->inode = inode;
1209
1210         inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
1211         inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1212         end_page_writeback(page);
1213
1214         spin_lock(&fc->lock);
1215         list_add(&req->writepages_entry, &fi->writepages);
1216         list_add_tail(&req->list, &fi->queued_writes);
1217         fuse_flush_writepages(inode);
1218         spin_unlock(&fc->lock);
1219
1220         return 0;
1221
1222 err_free:
1223         fuse_request_free(req);
1224 err:
1225         end_page_writeback(page);
1226         return -ENOMEM;
1227 }
1228
1229 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1230 {
1231         int err;
1232
1233         err = fuse_writepage_locked(page);
1234         unlock_page(page);
1235
1236         return err;
1237 }
1238
1239 static int fuse_launder_page(struct page *page)
1240 {
1241         int err = 0;
1242         if (clear_page_dirty_for_io(page)) {
1243                 struct inode *inode = page->mapping->host;
1244                 err = fuse_writepage_locked(page);
1245                 if (!err)
1246                         fuse_wait_on_page_writeback(inode, page->index);
1247         }
1248         return err;
1249 }
1250
1251 /*
1252  * Write back dirty pages now, because there may not be any suitable
1253  * open files later
1254  */
1255 static void fuse_vma_close(struct vm_area_struct *vma)
1256 {
1257         filemap_write_and_wait(vma->vm_file->f_mapping);
1258 }
1259
1260 /*
1261  * Wait for writeback against this page to complete before allowing it
1262  * to be marked dirty again, and hence written back again, possibly
1263  * before the previous writepage completed.
1264  *
1265  * Block here, instead of in ->writepage(), so that the userspace fs
1266  * can only block processes actually operating on the filesystem.
1267  *
1268  * Otherwise unprivileged userspace fs would be able to block
1269  * unrelated:
1270  *
1271  * - page migration
1272  * - sync(2)
1273  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
1274  */
1275 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1276 {
1277         struct page *page = vmf->page;
1278         /*
1279          * Don't use page->mapping as it may become NULL from a
1280          * concurrent truncate.
1281          */
1282         struct inode *inode = vma->vm_file->f_mapping->host;
1283
1284         fuse_wait_on_page_writeback(inode, page->index);
1285         return 0;
1286 }
1287
1288 static struct vm_operations_struct fuse_file_vm_ops = {
1289         .close          = fuse_vma_close,
1290         .fault          = filemap_fault,
1291         .page_mkwrite   = fuse_page_mkwrite,
1292 };
1293
1294 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
1295 {
1296         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE)) {
1297                 struct inode *inode = file->f_dentry->d_inode;
1298                 struct fuse_conn *fc = get_fuse_conn(inode);
1299                 struct fuse_inode *fi = get_fuse_inode(inode);
1300                 struct fuse_file *ff = file->private_data;
1301                 /*
1302                  * file may be written through mmap, so chain it onto the
1303                  * inodes's write_file list
1304                  */
1305                 spin_lock(&fc->lock);
1306                 if (list_empty(&ff->write_entry))
1307                         list_add(&ff->write_entry, &fi->write_files);
1308                 spin_unlock(&fc->lock);
1309         }
1310         file_accessed(file);
1311         vma->vm_ops = &fuse_file_vm_ops;
1312         return 0;
1313 }
1314
1315 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
1316 {
1317         /* Can't provide the coherency needed for MAP_SHARED */
1318         if (vma->vm_flags & VM_MAYSHARE)
1319                 return -ENODEV;
1320
1321         invalidate_inode_pages2(file->f_mapping);
1322
1323         return generic_file_mmap(file, vma);
1324 }
1325
1326 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
1327                                   struct file_lock *fl)
1328 {
1329         switch (ffl->type) {
1330         case F_UNLCK:
1331                 break;
1332
1333         case F_RDLCK:
1334         case F_WRLCK:
1335                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
1336                     ffl->end < ffl->start)
1337                         return -EIO;
1338
1339                 fl->fl_start = ffl->start;
1340                 fl->fl_end = ffl->end;
1341                 fl->fl_pid = ffl->pid;
1342                 break;
1343
1344         default:
1345                 return -EIO;
1346         }
1347         fl->fl_type = ffl->type;
1348         return 0;
1349 }
1350
1351 static void fuse_lk_fill(struct fuse_req *req, struct file *file,
1352                          const struct file_lock *fl, int opcode, pid_t pid,
1353                          int flock)
1354 {
1355         struct inode *inode = file->f_path.dentry->d_inode;
1356         struct fuse_conn *fc = get_fuse_conn(inode);
1357         struct fuse_file *ff = file->private_data;
1358         struct fuse_lk_in *arg = &req->misc.lk_in;
1359
1360         arg->fh = ff->fh;
1361         arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
1362         arg->lk.start = fl->fl_start;
1363         arg->lk.end = fl->fl_end;
1364         arg->lk.type = fl->fl_type;
1365         arg->lk.pid = pid;
1366         if (flock)
1367                 arg->lk_flags |= FUSE_LK_FLOCK;
1368         req->in.h.opcode = opcode;
1369         req->in.h.nodeid = get_node_id(inode);
1370         req->in.numargs = 1;
1371         req->in.args[0].size = sizeof(*arg);
1372         req->in.args[0].value = arg;
1373 }
1374
1375 static int fuse_getlk(struct file *file, struct file_lock *fl)
1376 {
1377         struct inode *inode = file->f_path.dentry->d_inode;
1378         struct fuse_conn *fc = get_fuse_conn(inode);
1379         struct fuse_req *req;
1380         struct fuse_lk_out outarg;
1381         int err;
1382
1383         req = fuse_get_req(fc);
1384         if (IS_ERR(req))
1385                 return PTR_ERR(req);
1386
1387         fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
1388         req->out.numargs = 1;
1389         req->out.args[0].size = sizeof(outarg);
1390         req->out.args[0].value = &outarg;
1391         fuse_request_send(fc, req);
1392         err = req->out.h.error;
1393         fuse_put_request(fc, req);
1394         if (!err)
1395                 err = convert_fuse_file_lock(&outarg.lk, fl);
1396
1397         return err;
1398 }
1399
1400 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
1401 {
1402         struct inode *inode = file->f_path.dentry->d_inode;
1403         struct fuse_conn *fc = get_fuse_conn(inode);
1404         struct fuse_req *req;
1405         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
1406         pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
1407         int err;
1408
1409         if (fl->fl_lmops && fl->fl_lmops->fl_grant) {
1410                 /* NLM needs asynchronous locks, which we don't support yet */
1411                 return -ENOLCK;
1412         }
1413
1414         /* Unlock on close is handled by the flush method */
1415         if (fl->fl_flags & FL_CLOSE)
1416                 return 0;
1417
1418         req = fuse_get_req(fc);
1419         if (IS_ERR(req))
1420                 return PTR_ERR(req);
1421
1422         fuse_lk_fill(req, file, fl, opcode, pid, flock);
1423         fuse_request_send(fc, req);
1424         err = req->out.h.error;
1425         /* locking is restartable */
1426         if (err == -EINTR)
1427                 err = -ERESTARTSYS;
1428         fuse_put_request(fc, req);
1429         return err;
1430 }
1431
1432 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
1433 {
1434         struct inode *inode = file->f_path.dentry->d_inode;
1435         struct fuse_conn *fc = get_fuse_conn(inode);
1436         int err;
1437
1438         if (cmd == F_CANCELLK) {
1439                 err = 0;
1440         } else if (cmd == F_GETLK) {
1441                 if (fc->no_lock) {
1442                         posix_test_lock(file, fl);
1443                         err = 0;
1444                 } else
1445                         err = fuse_getlk(file, fl);
1446         } else {
1447                 if (fc->no_lock)
1448                         err = posix_lock_file(file, fl, NULL);
1449                 else
1450                         err = fuse_setlk(file, fl, 0);
1451         }
1452         return err;
1453 }
1454
1455 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
1456 {
1457         struct inode *inode = file->f_path.dentry->d_inode;
1458         struct fuse_conn *fc = get_fuse_conn(inode);
1459         int err;
1460
1461         if (fc->no_lock) {
1462                 err = flock_lock_file_wait(file, fl);
1463         } else {
1464                 /* emulate flock with POSIX locks */
1465                 fl->fl_owner = (fl_owner_t) file;
1466                 err = fuse_setlk(file, fl, 1);
1467         }
1468
1469         return err;
1470 }
1471
1472 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
1473 {
1474         struct inode *inode = mapping->host;
1475         struct fuse_conn *fc = get_fuse_conn(inode);
1476         struct fuse_req *req;
1477         struct fuse_bmap_in inarg;
1478         struct fuse_bmap_out outarg;
1479         int err;
1480
1481         if (!inode->i_sb->s_bdev || fc->no_bmap)
1482                 return 0;
1483
1484         req = fuse_get_req(fc);
1485         if (IS_ERR(req))
1486                 return 0;
1487
1488         memset(&inarg, 0, sizeof(inarg));
1489         inarg.block = block;
1490         inarg.blocksize = inode->i_sb->s_blocksize;
1491         req->in.h.opcode = FUSE_BMAP;
1492         req->in.h.nodeid = get_node_id(inode);
1493         req->in.numargs = 1;
1494         req->in.args[0].size = sizeof(inarg);
1495         req->in.args[0].value = &inarg;
1496         req->out.numargs = 1;
1497         req->out.args[0].size = sizeof(outarg);
1498         req->out.args[0].value = &outarg;
1499         fuse_request_send(fc, req);
1500         err = req->out.h.error;
1501         fuse_put_request(fc, req);
1502         if (err == -ENOSYS)
1503                 fc->no_bmap = 1;
1504
1505         return err ? 0 : outarg.block;
1506 }
1507
1508 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int origin)
1509 {
1510         loff_t retval;
1511         struct inode *inode = file->f_path.dentry->d_inode;
1512
1513         mutex_lock(&inode->i_mutex);
1514         switch (origin) {
1515         case SEEK_END:
1516                 retval = fuse_update_attributes(inode, NULL, file, NULL);
1517                 if (retval)
1518                         goto exit;
1519                 offset += i_size_read(inode);
1520                 break;
1521         case SEEK_CUR:
1522                 offset += file->f_pos;
1523         }
1524         retval = -EINVAL;
1525         if (offset >= 0 && offset <= inode->i_sb->s_maxbytes) {
1526                 if (offset != file->f_pos) {
1527                         file->f_pos = offset;
1528                         file->f_version = 0;
1529                 }
1530                 retval = offset;
1531         }
1532 exit:
1533         mutex_unlock(&inode->i_mutex);
1534         return retval;
1535 }
1536
1537 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
1538                         unsigned int nr_segs, size_t bytes, bool to_user)
1539 {
1540         struct iov_iter ii;
1541         int page_idx = 0;
1542
1543         if (!bytes)
1544                 return 0;
1545
1546         iov_iter_init(&ii, iov, nr_segs, bytes, 0);
1547
1548         while (iov_iter_count(&ii)) {
1549                 struct page *page = pages[page_idx++];
1550                 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
1551                 void *kaddr, *map;
1552
1553                 kaddr = map = kmap(page);
1554
1555                 while (todo) {
1556                         char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
1557                         size_t iov_len = ii.iov->iov_len - ii.iov_offset;
1558                         size_t copy = min(todo, iov_len);
1559                         size_t left;
1560
1561                         if (!to_user)
1562                                 left = copy_from_user(kaddr, uaddr, copy);
1563                         else
1564                                 left = copy_to_user(uaddr, kaddr, copy);
1565
1566                         if (unlikely(left))
1567                                 return -EFAULT;
1568
1569                         iov_iter_advance(&ii, copy);
1570                         todo -= copy;
1571                         kaddr += copy;
1572                 }
1573
1574                 kunmap(map);
1575         }
1576
1577         return 0;
1578 }
1579
1580 /*
1581  * For ioctls, there is no generic way to determine how much memory
1582  * needs to be read and/or written.  Furthermore, ioctls are allowed
1583  * to dereference the passed pointer, so the parameter requires deep
1584  * copying but FUSE has no idea whatsoever about what to copy in or
1585  * out.
1586  *
1587  * This is solved by allowing FUSE server to retry ioctl with
1588  * necessary in/out iovecs.  Let's assume the ioctl implementation
1589  * needs to read in the following structure.
1590  *
1591  * struct a {
1592  *      char    *buf;
1593  *      size_t  buflen;
1594  * }
1595  *
1596  * On the first callout to FUSE server, inarg->in_size and
1597  * inarg->out_size will be NULL; then, the server completes the ioctl
1598  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
1599  * the actual iov array to
1600  *
1601  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
1602  *
1603  * which tells FUSE to copy in the requested area and retry the ioctl.
1604  * On the second round, the server has access to the structure and
1605  * from that it can tell what to look for next, so on the invocation,
1606  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
1607  *
1608  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
1609  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
1610  *
1611  * FUSE will copy both struct a and the pointed buffer from the
1612  * process doing the ioctl and retry ioctl with both struct a and the
1613  * buffer.
1614  *
1615  * This time, FUSE server has everything it needs and completes ioctl
1616  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
1617  *
1618  * Copying data out works the same way.
1619  *
1620  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
1621  * automatically initializes in and out iovs by decoding @cmd with
1622  * _IOC_* macros and the server is not allowed to request RETRY.  This
1623  * limits ioctl data transfers to well-formed ioctls and is the forced
1624  * behavior for all FUSE servers.
1625  */
1626 static long fuse_file_do_ioctl(struct file *file, unsigned int cmd,
1627                                unsigned long arg, unsigned int flags)
1628 {
1629         struct inode *inode = file->f_dentry->d_inode;
1630         struct fuse_file *ff = file->private_data;
1631         struct fuse_conn *fc = get_fuse_conn(inode);
1632         struct fuse_ioctl_in inarg = {
1633                 .fh = ff->fh,
1634                 .cmd = cmd,
1635                 .arg = arg,
1636                 .flags = flags
1637         };
1638         struct fuse_ioctl_out outarg;
1639         struct fuse_req *req = NULL;
1640         struct page **pages = NULL;
1641         struct page *iov_page = NULL;
1642         struct iovec *in_iov = NULL, *out_iov = NULL;
1643         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
1644         size_t in_size, out_size, transferred;
1645         int err;
1646
1647         /* assume all the iovs returned by client always fits in a page */
1648         BUILD_BUG_ON(sizeof(struct iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
1649
1650         if (!fuse_allow_task(fc, current))
1651                 return -EACCES;
1652
1653         err = -EIO;
1654         if (is_bad_inode(inode))
1655                 goto out;
1656
1657         err = -ENOMEM;
1658         pages = kzalloc(sizeof(pages[0]) * FUSE_MAX_PAGES_PER_REQ, GFP_KERNEL);
1659         iov_page = alloc_page(GFP_KERNEL);
1660         if (!pages || !iov_page)
1661                 goto out;
1662
1663         /*
1664          * If restricted, initialize IO parameters as encoded in @cmd.
1665          * RETRY from server is not allowed.
1666          */
1667         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
1668                 struct iovec *iov = page_address(iov_page);
1669
1670                 iov->iov_base = (void __user *)arg;
1671                 iov->iov_len = _IOC_SIZE(cmd);
1672
1673                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
1674                         in_iov = iov;
1675                         in_iovs = 1;
1676                 }
1677
1678                 if (_IOC_DIR(cmd) & _IOC_READ) {
1679                         out_iov = iov;
1680                         out_iovs = 1;
1681                 }
1682         }
1683
1684  retry:
1685         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
1686         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
1687
1688         /*
1689          * Out data can be used either for actual out data or iovs,
1690          * make sure there always is at least one page.
1691          */
1692         out_size = max_t(size_t, out_size, PAGE_SIZE);
1693         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
1694
1695         /* make sure there are enough buffer pages and init request with them */
1696         err = -ENOMEM;
1697         if (max_pages > FUSE_MAX_PAGES_PER_REQ)
1698                 goto out;
1699         while (num_pages < max_pages) {
1700                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
1701                 if (!pages[num_pages])
1702                         goto out;
1703                 num_pages++;
1704         }
1705
1706         req = fuse_get_req(fc);
1707         if (IS_ERR(req)) {
1708                 err = PTR_ERR(req);
1709                 req = NULL;
1710                 goto out;
1711         }
1712         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
1713         req->num_pages = num_pages;
1714
1715         /* okay, let's send it to the client */
1716         req->in.h.opcode = FUSE_IOCTL;
1717         req->in.h.nodeid = get_node_id(inode);
1718         req->in.numargs = 1;
1719         req->in.args[0].size = sizeof(inarg);
1720         req->in.args[0].value = &inarg;
1721         if (in_size) {
1722                 req->in.numargs++;
1723                 req->in.args[1].size = in_size;
1724                 req->in.argpages = 1;
1725
1726                 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
1727                                            false);
1728                 if (err)
1729                         goto out;
1730         }
1731
1732         req->out.numargs = 2;
1733         req->out.args[0].size = sizeof(outarg);
1734         req->out.args[0].value = &outarg;
1735         req->out.args[1].size = out_size;
1736         req->out.argpages = 1;
1737         req->out.argvar = 1;
1738
1739         fuse_request_send(fc, req);
1740         err = req->out.h.error;
1741         transferred = req->out.args[1].size;
1742         fuse_put_request(fc, req);
1743         req = NULL;
1744         if (err)
1745                 goto out;
1746
1747         /* did it ask for retry? */
1748         if (outarg.flags & FUSE_IOCTL_RETRY) {
1749                 char *vaddr;
1750
1751                 /* no retry if in restricted mode */
1752                 err = -EIO;
1753                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
1754                         goto out;
1755
1756                 in_iovs = outarg.in_iovs;
1757                 out_iovs = outarg.out_iovs;
1758
1759                 /*
1760                  * Make sure things are in boundary, separate checks
1761                  * are to protect against overflow.
1762                  */
1763                 err = -ENOMEM;
1764                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
1765                     out_iovs > FUSE_IOCTL_MAX_IOV ||
1766                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
1767                         goto out;
1768
1769                 err = -EIO;
1770                 if ((in_iovs + out_iovs) * sizeof(struct iovec) != transferred)
1771                         goto out;
1772
1773                 /* okay, copy in iovs and retry */
1774                 vaddr = kmap_atomic(pages[0], KM_USER0);
1775                 memcpy(page_address(iov_page), vaddr, transferred);
1776                 kunmap_atomic(vaddr, KM_USER0);
1777
1778                 in_iov = page_address(iov_page);
1779                 out_iov = in_iov + in_iovs;
1780
1781                 goto retry;
1782         }
1783
1784         err = -EIO;
1785         if (transferred > inarg.out_size)
1786                 goto out;
1787
1788         err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
1789  out:
1790         if (req)
1791                 fuse_put_request(fc, req);
1792         if (iov_page)
1793                 __free_page(iov_page);
1794         while (num_pages)
1795                 __free_page(pages[--num_pages]);
1796         kfree(pages);
1797
1798         return err ? err : outarg.result;
1799 }
1800
1801 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
1802                             unsigned long arg)
1803 {
1804         return fuse_file_do_ioctl(file, cmd, arg, 0);
1805 }
1806
1807 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
1808                                    unsigned long arg)
1809 {
1810         return fuse_file_do_ioctl(file, cmd, arg, FUSE_IOCTL_COMPAT);
1811 }
1812
1813 /*
1814  * All files which have been polled are linked to RB tree
1815  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
1816  * find the matching one.
1817  */
1818 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
1819                                               struct rb_node **parent_out)
1820 {
1821         struct rb_node **link = &fc->polled_files.rb_node;
1822         struct rb_node *last = NULL;
1823
1824         while (*link) {
1825                 struct fuse_file *ff;
1826
1827                 last = *link;
1828                 ff = rb_entry(last, struct fuse_file, polled_node);
1829
1830                 if (kh < ff->kh)
1831                         link = &last->rb_left;
1832                 else if (kh > ff->kh)
1833                         link = &last->rb_right;
1834                 else
1835                         return link;
1836         }
1837
1838         if (parent_out)
1839                 *parent_out = last;
1840         return link;
1841 }
1842
1843 /*
1844  * The file is about to be polled.  Make sure it's on the polled_files
1845  * RB tree.  Note that files once added to the polled_files tree are
1846  * not removed before the file is released.  This is because a file
1847  * polled once is likely to be polled again.
1848  */
1849 static void fuse_register_polled_file(struct fuse_conn *fc,
1850                                       struct fuse_file *ff)
1851 {
1852         spin_lock(&fc->lock);
1853         if (RB_EMPTY_NODE(&ff->polled_node)) {
1854                 struct rb_node **link, *parent;
1855
1856                 link = fuse_find_polled_node(fc, ff->kh, &parent);
1857                 BUG_ON(*link);
1858                 rb_link_node(&ff->polled_node, parent, link);
1859                 rb_insert_color(&ff->polled_node, &fc->polled_files);
1860         }
1861         spin_unlock(&fc->lock);
1862 }
1863
1864 static unsigned fuse_file_poll(struct file *file, poll_table *wait)
1865 {
1866         struct inode *inode = file->f_dentry->d_inode;
1867         struct fuse_file *ff = file->private_data;
1868         struct fuse_conn *fc = get_fuse_conn(inode);
1869         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
1870         struct fuse_poll_out outarg;
1871         struct fuse_req *req;
1872         int err;
1873
1874         if (fc->no_poll)
1875                 return DEFAULT_POLLMASK;
1876
1877         poll_wait(file, &ff->poll_wait, wait);
1878
1879         /*
1880          * Ask for notification iff there's someone waiting for it.
1881          * The client may ignore the flag and always notify.
1882          */
1883         if (waitqueue_active(&ff->poll_wait)) {
1884                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
1885                 fuse_register_polled_file(fc, ff);
1886         }
1887
1888         req = fuse_get_req(fc);
1889         if (IS_ERR(req))
1890                 return PTR_ERR(req);
1891
1892         req->in.h.opcode = FUSE_POLL;
1893         req->in.h.nodeid = get_node_id(inode);
1894         req->in.numargs = 1;
1895         req->in.args[0].size = sizeof(inarg);
1896         req->in.args[0].value = &inarg;
1897         req->out.numargs = 1;
1898         req->out.args[0].size = sizeof(outarg);
1899         req->out.args[0].value = &outarg;
1900         fuse_request_send(fc, req);
1901         err = req->out.h.error;
1902         fuse_put_request(fc, req);
1903
1904         if (!err)
1905                 return outarg.revents;
1906         if (err == -ENOSYS) {
1907                 fc->no_poll = 1;
1908                 return DEFAULT_POLLMASK;
1909         }
1910         return POLLERR;
1911 }
1912
1913 /*
1914  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
1915  * wakes up the poll waiters.
1916  */
1917 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
1918                             struct fuse_notify_poll_wakeup_out *outarg)
1919 {
1920         u64 kh = outarg->kh;
1921         struct rb_node **link;
1922
1923         spin_lock(&fc->lock);
1924
1925         link = fuse_find_polled_node(fc, kh, NULL);
1926         if (*link) {
1927                 struct fuse_file *ff;
1928
1929                 ff = rb_entry(*link, struct fuse_file, polled_node);
1930                 wake_up_interruptible_sync(&ff->poll_wait);
1931         }
1932
1933         spin_unlock(&fc->lock);
1934         return 0;
1935 }
1936
1937 static const struct file_operations fuse_file_operations = {
1938         .llseek         = fuse_file_llseek,
1939         .read           = do_sync_read,
1940         .aio_read       = fuse_file_aio_read,
1941         .write          = do_sync_write,
1942         .aio_write      = fuse_file_aio_write,
1943         .mmap           = fuse_file_mmap,
1944         .open           = fuse_open,
1945         .flush          = fuse_flush,
1946         .release        = fuse_release,
1947         .fsync          = fuse_fsync,
1948         .lock           = fuse_file_lock,
1949         .flock          = fuse_file_flock,
1950         .splice_read    = generic_file_splice_read,
1951         .unlocked_ioctl = fuse_file_ioctl,
1952         .compat_ioctl   = fuse_file_compat_ioctl,
1953         .poll           = fuse_file_poll,
1954 };
1955
1956 static const struct file_operations fuse_direct_io_file_operations = {
1957         .llseek         = fuse_file_llseek,
1958         .read           = fuse_direct_read,
1959         .write          = fuse_direct_write,
1960         .mmap           = fuse_direct_mmap,
1961         .open           = fuse_open,
1962         .flush          = fuse_flush,
1963         .release        = fuse_release,
1964         .fsync          = fuse_fsync,
1965         .lock           = fuse_file_lock,
1966         .flock          = fuse_file_flock,
1967         .unlocked_ioctl = fuse_file_ioctl,
1968         .compat_ioctl   = fuse_file_compat_ioctl,
1969         .poll           = fuse_file_poll,
1970         /* no splice_read */
1971 };
1972
1973 static const struct address_space_operations fuse_file_aops  = {
1974         .readpage       = fuse_readpage,
1975         .writepage      = fuse_writepage,
1976         .launder_page   = fuse_launder_page,
1977         .write_begin    = fuse_write_begin,
1978         .write_end      = fuse_write_end,
1979         .readpages      = fuse_readpages,
1980         .set_page_dirty = __set_page_dirty_nobuffers,
1981         .bmap           = fuse_bmap,
1982 };
1983
1984 void fuse_init_file_inode(struct inode *inode)
1985 {
1986         inode->i_fop = &fuse_file_operations;
1987         inode->i_data.a_ops = &fuse_file_aops;
1988 }