Merge branches 'x86/asm', 'x86/cleanups', 'x86/cpudetect', 'x86/debug', 'x86/doc...
[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.argpages = 1;
390         req->out.argvar = 1;
391         req->out.numargs = 1;
392         req->out.args[0].size = count;
393 }
394
395 static size_t fuse_send_read(struct fuse_req *req, struct file *file,
396                              struct inode *inode, loff_t pos, size_t count,
397                              fl_owner_t owner)
398 {
399         struct fuse_conn *fc = get_fuse_conn(inode);
400
401         fuse_read_fill(req, file, inode, pos, count, FUSE_READ);
402         if (owner != NULL) {
403                 struct fuse_read_in *inarg = &req->misc.read.in;
404
405                 inarg->read_flags |= FUSE_READ_LOCKOWNER;
406                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
407         }
408         fuse_request_send(fc, req);
409         return req->out.args[0].size;
410 }
411
412 static void fuse_read_update_size(struct inode *inode, loff_t size,
413                                   u64 attr_ver)
414 {
415         struct fuse_conn *fc = get_fuse_conn(inode);
416         struct fuse_inode *fi = get_fuse_inode(inode);
417
418         spin_lock(&fc->lock);
419         if (attr_ver == fi->attr_version && size < inode->i_size) {
420                 fi->attr_version = ++fc->attr_version;
421                 i_size_write(inode, size);
422         }
423         spin_unlock(&fc->lock);
424 }
425
426 static int fuse_readpage(struct file *file, struct page *page)
427 {
428         struct inode *inode = page->mapping->host;
429         struct fuse_conn *fc = get_fuse_conn(inode);
430         struct fuse_req *req;
431         size_t num_read;
432         loff_t pos = page_offset(page);
433         size_t count = PAGE_CACHE_SIZE;
434         u64 attr_ver;
435         int err;
436
437         err = -EIO;
438         if (is_bad_inode(inode))
439                 goto out;
440
441         /*
442          * Page writeback can extend beyond the liftime of the
443          * page-cache page, so make sure we read a properly synced
444          * page.
445          */
446         fuse_wait_on_page_writeback(inode, page->index);
447
448         req = fuse_get_req(fc);
449         err = PTR_ERR(req);
450         if (IS_ERR(req))
451                 goto out;
452
453         attr_ver = fuse_get_attr_version(fc);
454
455         req->out.page_zeroing = 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         req->out.page_zeroing = 1;
514         fuse_read_fill(req, file, inode, pos, count, FUSE_READ);
515         req->misc.read.attr_ver = fuse_get_attr_version(fc);
516         if (fc->async_read) {
517                 struct fuse_file *ff = file->private_data;
518                 req->ff = fuse_file_get(ff);
519                 req->end = fuse_readpages_end;
520                 fuse_request_send_background(fc, req);
521         } else {
522                 fuse_request_send(fc, req);
523                 fuse_readpages_end(fc, req);
524                 fuse_put_request(fc, req);
525         }
526 }
527
528 struct fuse_fill_data {
529         struct fuse_req *req;
530         struct file *file;
531         struct inode *inode;
532 };
533
534 static int fuse_readpages_fill(void *_data, struct page *page)
535 {
536         struct fuse_fill_data *data = _data;
537         struct fuse_req *req = data->req;
538         struct inode *inode = data->inode;
539         struct fuse_conn *fc = get_fuse_conn(inode);
540
541         fuse_wait_on_page_writeback(inode, page->index);
542
543         if (req->num_pages &&
544             (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
545              (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
546              req->pages[req->num_pages - 1]->index + 1 != page->index)) {
547                 fuse_send_readpages(req, data->file, inode);
548                 data->req = req = fuse_get_req(fc);
549                 if (IS_ERR(req)) {
550                         unlock_page(page);
551                         return PTR_ERR(req);
552                 }
553         }
554         req->pages[req->num_pages] = page;
555         req->num_pages++;
556         return 0;
557 }
558
559 static int fuse_readpages(struct file *file, struct address_space *mapping,
560                           struct list_head *pages, unsigned nr_pages)
561 {
562         struct inode *inode = mapping->host;
563         struct fuse_conn *fc = get_fuse_conn(inode);
564         struct fuse_fill_data data;
565         int err;
566
567         err = -EIO;
568         if (is_bad_inode(inode))
569                 goto out;
570
571         data.file = file;
572         data.inode = inode;
573         data.req = fuse_get_req(fc);
574         err = PTR_ERR(data.req);
575         if (IS_ERR(data.req))
576                 goto out;
577
578         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
579         if (!err) {
580                 if (data.req->num_pages)
581                         fuse_send_readpages(data.req, file, inode);
582                 else
583                         fuse_put_request(fc, data.req);
584         }
585 out:
586         return err;
587 }
588
589 static ssize_t fuse_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
590                                   unsigned long nr_segs, loff_t pos)
591 {
592         struct inode *inode = iocb->ki_filp->f_mapping->host;
593
594         if (pos + iov_length(iov, nr_segs) > i_size_read(inode)) {
595                 int err;
596                 /*
597                  * If trying to read past EOF, make sure the i_size
598                  * attribute is up-to-date.
599                  */
600                 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
601                 if (err)
602                         return err;
603         }
604
605         return generic_file_aio_read(iocb, iov, nr_segs, pos);
606 }
607
608 static void fuse_write_fill(struct fuse_req *req, struct file *file,
609                             struct fuse_file *ff, struct inode *inode,
610                             loff_t pos, size_t count, int writepage)
611 {
612         struct fuse_conn *fc = get_fuse_conn(inode);
613         struct fuse_write_in *inarg = &req->misc.write.in;
614         struct fuse_write_out *outarg = &req->misc.write.out;
615
616         memset(inarg, 0, sizeof(struct fuse_write_in));
617         inarg->fh = ff->fh;
618         inarg->offset = pos;
619         inarg->size = count;
620         inarg->write_flags = writepage ? FUSE_WRITE_CACHE : 0;
621         inarg->flags = file ? file->f_flags : 0;
622         req->in.h.opcode = FUSE_WRITE;
623         req->in.h.nodeid = get_node_id(inode);
624         req->in.argpages = 1;
625         req->in.numargs = 2;
626         if (fc->minor < 9)
627                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
628         else
629                 req->in.args[0].size = sizeof(struct fuse_write_in);
630         req->in.args[0].value = inarg;
631         req->in.args[1].size = count;
632         req->out.numargs = 1;
633         req->out.args[0].size = sizeof(struct fuse_write_out);
634         req->out.args[0].value = outarg;
635 }
636
637 static size_t fuse_send_write(struct fuse_req *req, struct file *file,
638                               struct inode *inode, loff_t pos, size_t count,
639                               fl_owner_t owner)
640 {
641         struct fuse_conn *fc = get_fuse_conn(inode);
642         fuse_write_fill(req, file, file->private_data, inode, pos, count, 0);
643         if (owner != NULL) {
644                 struct fuse_write_in *inarg = &req->misc.write.in;
645                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
646                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
647         }
648         fuse_request_send(fc, req);
649         return req->misc.write.out.size;
650 }
651
652 static int fuse_write_begin(struct file *file, struct address_space *mapping,
653                         loff_t pos, unsigned len, unsigned flags,
654                         struct page **pagep, void **fsdata)
655 {
656         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
657
658         *pagep = grab_cache_page_write_begin(mapping, index, flags);
659         if (!*pagep)
660                 return -ENOMEM;
661         return 0;
662 }
663
664 static void fuse_write_update_size(struct inode *inode, loff_t pos)
665 {
666         struct fuse_conn *fc = get_fuse_conn(inode);
667         struct fuse_inode *fi = get_fuse_inode(inode);
668
669         spin_lock(&fc->lock);
670         fi->attr_version = ++fc->attr_version;
671         if (pos > inode->i_size)
672                 i_size_write(inode, pos);
673         spin_unlock(&fc->lock);
674 }
675
676 static int fuse_buffered_write(struct file *file, struct inode *inode,
677                                loff_t pos, unsigned count, struct page *page)
678 {
679         int err;
680         size_t nres;
681         struct fuse_conn *fc = get_fuse_conn(inode);
682         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
683         struct fuse_req *req;
684
685         if (is_bad_inode(inode))
686                 return -EIO;
687
688         /*
689          * Make sure writepages on the same page are not mixed up with
690          * plain writes.
691          */
692         fuse_wait_on_page_writeback(inode, page->index);
693
694         req = fuse_get_req(fc);
695         if (IS_ERR(req))
696                 return PTR_ERR(req);
697
698         req->num_pages = 1;
699         req->pages[0] = page;
700         req->page_offset = offset;
701         nres = fuse_send_write(req, file, inode, pos, count, NULL);
702         err = req->out.h.error;
703         fuse_put_request(fc, req);
704         if (!err && !nres)
705                 err = -EIO;
706         if (!err) {
707                 pos += nres;
708                 fuse_write_update_size(inode, pos);
709                 if (count == PAGE_CACHE_SIZE)
710                         SetPageUptodate(page);
711         }
712         fuse_invalidate_attr(inode);
713         return err ? err : nres;
714 }
715
716 static int fuse_write_end(struct file *file, struct address_space *mapping,
717                         loff_t pos, unsigned len, unsigned copied,
718                         struct page *page, void *fsdata)
719 {
720         struct inode *inode = mapping->host;
721         int res = 0;
722
723         if (copied)
724                 res = fuse_buffered_write(file, inode, pos, copied, page);
725
726         unlock_page(page);
727         page_cache_release(page);
728         return res;
729 }
730
731 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
732                                     struct inode *inode, loff_t pos,
733                                     size_t count)
734 {
735         size_t res;
736         unsigned offset;
737         unsigned i;
738
739         for (i = 0; i < req->num_pages; i++)
740                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
741
742         res = fuse_send_write(req, file, inode, pos, count, NULL);
743
744         offset = req->page_offset;
745         count = res;
746         for (i = 0; i < req->num_pages; i++) {
747                 struct page *page = req->pages[i];
748
749                 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
750                         SetPageUptodate(page);
751
752                 if (count > PAGE_CACHE_SIZE - offset)
753                         count -= PAGE_CACHE_SIZE - offset;
754                 else
755                         count = 0;
756                 offset = 0;
757
758                 unlock_page(page);
759                 page_cache_release(page);
760         }
761
762         return res;
763 }
764
765 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
766                                struct address_space *mapping,
767                                struct iov_iter *ii, loff_t pos)
768 {
769         struct fuse_conn *fc = get_fuse_conn(mapping->host);
770         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
771         size_t count = 0;
772         int err;
773
774         req->page_offset = offset;
775
776         do {
777                 size_t tmp;
778                 struct page *page;
779                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
780                 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
781                                      iov_iter_count(ii));
782
783                 bytes = min_t(size_t, bytes, fc->max_write - count);
784
785  again:
786                 err = -EFAULT;
787                 if (iov_iter_fault_in_readable(ii, bytes))
788                         break;
789
790                 err = -ENOMEM;
791                 page = grab_cache_page_write_begin(mapping, index, 0);
792                 if (!page)
793                         break;
794
795                 pagefault_disable();
796                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
797                 pagefault_enable();
798                 flush_dcache_page(page);
799
800                 if (!tmp) {
801                         unlock_page(page);
802                         page_cache_release(page);
803                         bytes = min(bytes, iov_iter_single_seg_count(ii));
804                         goto again;
805                 }
806
807                 err = 0;
808                 req->pages[req->num_pages] = page;
809                 req->num_pages++;
810
811                 iov_iter_advance(ii, tmp);
812                 count += tmp;
813                 pos += tmp;
814                 offset += tmp;
815                 if (offset == PAGE_CACHE_SIZE)
816                         offset = 0;
817
818                 if (!fc->big_writes)
819                         break;
820         } while (iov_iter_count(ii) && count < fc->max_write &&
821                  req->num_pages < FUSE_MAX_PAGES_PER_REQ && offset == 0);
822
823         return count > 0 ? count : err;
824 }
825
826 static ssize_t fuse_perform_write(struct file *file,
827                                   struct address_space *mapping,
828                                   struct iov_iter *ii, loff_t pos)
829 {
830         struct inode *inode = mapping->host;
831         struct fuse_conn *fc = get_fuse_conn(inode);
832         int err = 0;
833         ssize_t res = 0;
834
835         if (is_bad_inode(inode))
836                 return -EIO;
837
838         do {
839                 struct fuse_req *req;
840                 ssize_t count;
841
842                 req = fuse_get_req(fc);
843                 if (IS_ERR(req)) {
844                         err = PTR_ERR(req);
845                         break;
846                 }
847
848                 count = fuse_fill_write_pages(req, mapping, ii, pos);
849                 if (count <= 0) {
850                         err = count;
851                 } else {
852                         size_t num_written;
853
854                         num_written = fuse_send_write_pages(req, file, inode,
855                                                             pos, count);
856                         err = req->out.h.error;
857                         if (!err) {
858                                 res += num_written;
859                                 pos += num_written;
860
861                                 /* break out of the loop on short write */
862                                 if (num_written != count)
863                                         err = -EIO;
864                         }
865                 }
866                 fuse_put_request(fc, req);
867         } while (!err && iov_iter_count(ii));
868
869         if (res > 0)
870                 fuse_write_update_size(inode, pos);
871
872         fuse_invalidate_attr(inode);
873
874         return res > 0 ? res : err;
875 }
876
877 static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
878                                    unsigned long nr_segs, loff_t pos)
879 {
880         struct file *file = iocb->ki_filp;
881         struct address_space *mapping = file->f_mapping;
882         size_t count = 0;
883         ssize_t written = 0;
884         struct inode *inode = mapping->host;
885         ssize_t err;
886         struct iov_iter i;
887
888         WARN_ON(iocb->ki_pos != pos);
889
890         err = generic_segment_checks(iov, &nr_segs, &count, VERIFY_READ);
891         if (err)
892                 return err;
893
894         mutex_lock(&inode->i_mutex);
895         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
896
897         /* We can write back this queue in page reclaim */
898         current->backing_dev_info = mapping->backing_dev_info;
899
900         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
901         if (err)
902                 goto out;
903
904         if (count == 0)
905                 goto out;
906
907         err = file_remove_suid(file);
908         if (err)
909                 goto out;
910
911         file_update_time(file);
912
913         iov_iter_init(&i, iov, nr_segs, count, 0);
914         written = fuse_perform_write(file, mapping, &i, pos);
915         if (written >= 0)
916                 iocb->ki_pos = pos + written;
917
918 out:
919         current->backing_dev_info = NULL;
920         mutex_unlock(&inode->i_mutex);
921
922         return written ? written : err;
923 }
924
925 static void fuse_release_user_pages(struct fuse_req *req, int write)
926 {
927         unsigned i;
928
929         for (i = 0; i < req->num_pages; i++) {
930                 struct page *page = req->pages[i];
931                 if (write)
932                         set_page_dirty_lock(page);
933                 put_page(page);
934         }
935 }
936
937 static int fuse_get_user_pages(struct fuse_req *req, const char __user *buf,
938                                unsigned nbytes, int write)
939 {
940         unsigned long user_addr = (unsigned long) buf;
941         unsigned offset = user_addr & ~PAGE_MASK;
942         int npages;
943
944         /* This doesn't work with nfsd */
945         if (!current->mm)
946                 return -EPERM;
947
948         nbytes = min(nbytes, (unsigned) FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
949         npages = (nbytes + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
950         npages = clamp(npages, 1, FUSE_MAX_PAGES_PER_REQ);
951         down_read(&current->mm->mmap_sem);
952         npages = get_user_pages(current, current->mm, user_addr, npages, write,
953                                 0, req->pages, NULL);
954         up_read(&current->mm->mmap_sem);
955         if (npages < 0)
956                 return npages;
957
958         req->num_pages = npages;
959         req->page_offset = offset;
960         return 0;
961 }
962
963 static ssize_t fuse_direct_io(struct file *file, const char __user *buf,
964                               size_t count, loff_t *ppos, int write)
965 {
966         struct inode *inode = file->f_path.dentry->d_inode;
967         struct fuse_conn *fc = get_fuse_conn(inode);
968         size_t nmax = write ? fc->max_write : fc->max_read;
969         loff_t pos = *ppos;
970         ssize_t res = 0;
971         struct fuse_req *req;
972
973         if (is_bad_inode(inode))
974                 return -EIO;
975
976         req = fuse_get_req(fc);
977         if (IS_ERR(req))
978                 return PTR_ERR(req);
979
980         while (count) {
981                 size_t nres;
982                 size_t nbytes_limit = min(count, nmax);
983                 size_t nbytes;
984                 int err = fuse_get_user_pages(req, buf, nbytes_limit, !write);
985                 if (err) {
986                         res = err;
987                         break;
988                 }
989                 nbytes = (req->num_pages << PAGE_SHIFT) - req->page_offset;
990                 nbytes = min(nbytes_limit, nbytes);
991                 if (write)
992                         nres = fuse_send_write(req, file, inode, pos, nbytes,
993                                                current->files);
994                 else
995                         nres = fuse_send_read(req, file, inode, pos, nbytes,
996                                               current->files);
997                 fuse_release_user_pages(req, !write);
998                 if (req->out.h.error) {
999                         if (!res)
1000                                 res = req->out.h.error;
1001                         break;
1002                 } else if (nres > nbytes) {
1003                         res = -EIO;
1004                         break;
1005                 }
1006                 count -= nres;
1007                 res += nres;
1008                 pos += nres;
1009                 buf += nres;
1010                 if (nres != nbytes)
1011                         break;
1012                 if (count) {
1013                         fuse_put_request(fc, req);
1014                         req = fuse_get_req(fc);
1015                         if (IS_ERR(req))
1016                                 break;
1017                 }
1018         }
1019         fuse_put_request(fc, req);
1020         if (res > 0) {
1021                 if (write)
1022                         fuse_write_update_size(inode, pos);
1023                 *ppos = pos;
1024         }
1025         fuse_invalidate_attr(inode);
1026
1027         return res;
1028 }
1029
1030 static ssize_t fuse_direct_read(struct file *file, char __user *buf,
1031                                      size_t count, loff_t *ppos)
1032 {
1033         return fuse_direct_io(file, buf, count, ppos, 0);
1034 }
1035
1036 static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
1037                                  size_t count, loff_t *ppos)
1038 {
1039         struct inode *inode = file->f_path.dentry->d_inode;
1040         ssize_t res;
1041         /* Don't allow parallel writes to the same file */
1042         mutex_lock(&inode->i_mutex);
1043         res = generic_write_checks(file, ppos, &count, 0);
1044         if (!res)
1045                 res = fuse_direct_io(file, buf, count, ppos, 1);
1046         mutex_unlock(&inode->i_mutex);
1047         return res;
1048 }
1049
1050 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1051 {
1052         __free_page(req->pages[0]);
1053         fuse_file_put(req->ff);
1054 }
1055
1056 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1057 {
1058         struct inode *inode = req->inode;
1059         struct fuse_inode *fi = get_fuse_inode(inode);
1060         struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
1061
1062         list_del(&req->writepages_entry);
1063         dec_bdi_stat(bdi, BDI_WRITEBACK);
1064         dec_zone_page_state(req->pages[0], NR_WRITEBACK_TEMP);
1065         bdi_writeout_inc(bdi);
1066         wake_up(&fi->page_waitq);
1067 }
1068
1069 /* Called under fc->lock, may release and reacquire it */
1070 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req)
1071 __releases(&fc->lock)
1072 __acquires(&fc->lock)
1073 {
1074         struct fuse_inode *fi = get_fuse_inode(req->inode);
1075         loff_t size = i_size_read(req->inode);
1076         struct fuse_write_in *inarg = &req->misc.write.in;
1077
1078         if (!fc->connected)
1079                 goto out_free;
1080
1081         if (inarg->offset + PAGE_CACHE_SIZE <= size) {
1082                 inarg->size = PAGE_CACHE_SIZE;
1083         } else if (inarg->offset < size) {
1084                 inarg->size = size & (PAGE_CACHE_SIZE - 1);
1085         } else {
1086                 /* Got truncated off completely */
1087                 goto out_free;
1088         }
1089
1090         req->in.args[1].size = inarg->size;
1091         fi->writectr++;
1092         fuse_request_send_background_locked(fc, req);
1093         return;
1094
1095  out_free:
1096         fuse_writepage_finish(fc, req);
1097         spin_unlock(&fc->lock);
1098         fuse_writepage_free(fc, req);
1099         fuse_put_request(fc, req);
1100         spin_lock(&fc->lock);
1101 }
1102
1103 /*
1104  * If fi->writectr is positive (no truncate or fsync going on) send
1105  * all queued writepage requests.
1106  *
1107  * Called with fc->lock
1108  */
1109 void fuse_flush_writepages(struct inode *inode)
1110 __releases(&fc->lock)
1111 __acquires(&fc->lock)
1112 {
1113         struct fuse_conn *fc = get_fuse_conn(inode);
1114         struct fuse_inode *fi = get_fuse_inode(inode);
1115         struct fuse_req *req;
1116
1117         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1118                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1119                 list_del_init(&req->list);
1120                 fuse_send_writepage(fc, req);
1121         }
1122 }
1123
1124 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1125 {
1126         struct inode *inode = req->inode;
1127         struct fuse_inode *fi = get_fuse_inode(inode);
1128
1129         mapping_set_error(inode->i_mapping, req->out.h.error);
1130         spin_lock(&fc->lock);
1131         fi->writectr--;
1132         fuse_writepage_finish(fc, req);
1133         spin_unlock(&fc->lock);
1134         fuse_writepage_free(fc, req);
1135 }
1136
1137 static int fuse_writepage_locked(struct page *page)
1138 {
1139         struct address_space *mapping = page->mapping;
1140         struct inode *inode = mapping->host;
1141         struct fuse_conn *fc = get_fuse_conn(inode);
1142         struct fuse_inode *fi = get_fuse_inode(inode);
1143         struct fuse_req *req;
1144         struct fuse_file *ff;
1145         struct page *tmp_page;
1146
1147         set_page_writeback(page);
1148
1149         req = fuse_request_alloc_nofs();
1150         if (!req)
1151                 goto err;
1152
1153         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1154         if (!tmp_page)
1155                 goto err_free;
1156
1157         spin_lock(&fc->lock);
1158         BUG_ON(list_empty(&fi->write_files));
1159         ff = list_entry(fi->write_files.next, struct fuse_file, write_entry);
1160         req->ff = fuse_file_get(ff);
1161         spin_unlock(&fc->lock);
1162
1163         fuse_write_fill(req, NULL, ff, inode, page_offset(page), 0, 1);
1164
1165         copy_highpage(tmp_page, page);
1166         req->num_pages = 1;
1167         req->pages[0] = tmp_page;
1168         req->page_offset = 0;
1169         req->end = fuse_writepage_end;
1170         req->inode = inode;
1171
1172         inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
1173         inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1174         end_page_writeback(page);
1175
1176         spin_lock(&fc->lock);
1177         list_add(&req->writepages_entry, &fi->writepages);
1178         list_add_tail(&req->list, &fi->queued_writes);
1179         fuse_flush_writepages(inode);
1180         spin_unlock(&fc->lock);
1181
1182         return 0;
1183
1184 err_free:
1185         fuse_request_free(req);
1186 err:
1187         end_page_writeback(page);
1188         return -ENOMEM;
1189 }
1190
1191 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1192 {
1193         int err;
1194
1195         err = fuse_writepage_locked(page);
1196         unlock_page(page);
1197
1198         return err;
1199 }
1200
1201 static int fuse_launder_page(struct page *page)
1202 {
1203         int err = 0;
1204         if (clear_page_dirty_for_io(page)) {
1205                 struct inode *inode = page->mapping->host;
1206                 err = fuse_writepage_locked(page);
1207                 if (!err)
1208                         fuse_wait_on_page_writeback(inode, page->index);
1209         }
1210         return err;
1211 }
1212
1213 /*
1214  * Write back dirty pages now, because there may not be any suitable
1215  * open files later
1216  */
1217 static void fuse_vma_close(struct vm_area_struct *vma)
1218 {
1219         filemap_write_and_wait(vma->vm_file->f_mapping);
1220 }
1221
1222 /*
1223  * Wait for writeback against this page to complete before allowing it
1224  * to be marked dirty again, and hence written back again, possibly
1225  * before the previous writepage completed.
1226  *
1227  * Block here, instead of in ->writepage(), so that the userspace fs
1228  * can only block processes actually operating on the filesystem.
1229  *
1230  * Otherwise unprivileged userspace fs would be able to block
1231  * unrelated:
1232  *
1233  * - page migration
1234  * - sync(2)
1235  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
1236  */
1237 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct page *page)
1238 {
1239         /*
1240          * Don't use page->mapping as it may become NULL from a
1241          * concurrent truncate.
1242          */
1243         struct inode *inode = vma->vm_file->f_mapping->host;
1244
1245         fuse_wait_on_page_writeback(inode, page->index);
1246         return 0;
1247 }
1248
1249 static struct vm_operations_struct fuse_file_vm_ops = {
1250         .close          = fuse_vma_close,
1251         .fault          = filemap_fault,
1252         .page_mkwrite   = fuse_page_mkwrite,
1253 };
1254
1255 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
1256 {
1257         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE)) {
1258                 struct inode *inode = file->f_dentry->d_inode;
1259                 struct fuse_conn *fc = get_fuse_conn(inode);
1260                 struct fuse_inode *fi = get_fuse_inode(inode);
1261                 struct fuse_file *ff = file->private_data;
1262                 /*
1263                  * file may be written through mmap, so chain it onto the
1264                  * inodes's write_file list
1265                  */
1266                 spin_lock(&fc->lock);
1267                 if (list_empty(&ff->write_entry))
1268                         list_add(&ff->write_entry, &fi->write_files);
1269                 spin_unlock(&fc->lock);
1270         }
1271         file_accessed(file);
1272         vma->vm_ops = &fuse_file_vm_ops;
1273         return 0;
1274 }
1275
1276 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
1277                                   struct file_lock *fl)
1278 {
1279         switch (ffl->type) {
1280         case F_UNLCK:
1281                 break;
1282
1283         case F_RDLCK:
1284         case F_WRLCK:
1285                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
1286                     ffl->end < ffl->start)
1287                         return -EIO;
1288
1289                 fl->fl_start = ffl->start;
1290                 fl->fl_end = ffl->end;
1291                 fl->fl_pid = ffl->pid;
1292                 break;
1293
1294         default:
1295                 return -EIO;
1296         }
1297         fl->fl_type = ffl->type;
1298         return 0;
1299 }
1300
1301 static void fuse_lk_fill(struct fuse_req *req, struct file *file,
1302                          const struct file_lock *fl, int opcode, pid_t pid,
1303                          int flock)
1304 {
1305         struct inode *inode = file->f_path.dentry->d_inode;
1306         struct fuse_conn *fc = get_fuse_conn(inode);
1307         struct fuse_file *ff = file->private_data;
1308         struct fuse_lk_in *arg = &req->misc.lk_in;
1309
1310         arg->fh = ff->fh;
1311         arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
1312         arg->lk.start = fl->fl_start;
1313         arg->lk.end = fl->fl_end;
1314         arg->lk.type = fl->fl_type;
1315         arg->lk.pid = pid;
1316         if (flock)
1317                 arg->lk_flags |= FUSE_LK_FLOCK;
1318         req->in.h.opcode = opcode;
1319         req->in.h.nodeid = get_node_id(inode);
1320         req->in.numargs = 1;
1321         req->in.args[0].size = sizeof(*arg);
1322         req->in.args[0].value = arg;
1323 }
1324
1325 static int fuse_getlk(struct file *file, struct file_lock *fl)
1326 {
1327         struct inode *inode = file->f_path.dentry->d_inode;
1328         struct fuse_conn *fc = get_fuse_conn(inode);
1329         struct fuse_req *req;
1330         struct fuse_lk_out outarg;
1331         int err;
1332
1333         req = fuse_get_req(fc);
1334         if (IS_ERR(req))
1335                 return PTR_ERR(req);
1336
1337         fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
1338         req->out.numargs = 1;
1339         req->out.args[0].size = sizeof(outarg);
1340         req->out.args[0].value = &outarg;
1341         fuse_request_send(fc, req);
1342         err = req->out.h.error;
1343         fuse_put_request(fc, req);
1344         if (!err)
1345                 err = convert_fuse_file_lock(&outarg.lk, fl);
1346
1347         return err;
1348 }
1349
1350 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
1351 {
1352         struct inode *inode = file->f_path.dentry->d_inode;
1353         struct fuse_conn *fc = get_fuse_conn(inode);
1354         struct fuse_req *req;
1355         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
1356         pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
1357         int err;
1358
1359         if (fl->fl_lmops && fl->fl_lmops->fl_grant) {
1360                 /* NLM needs asynchronous locks, which we don't support yet */
1361                 return -ENOLCK;
1362         }
1363
1364         /* Unlock on close is handled by the flush method */
1365         if (fl->fl_flags & FL_CLOSE)
1366                 return 0;
1367
1368         req = fuse_get_req(fc);
1369         if (IS_ERR(req))
1370                 return PTR_ERR(req);
1371
1372         fuse_lk_fill(req, file, fl, opcode, pid, flock);
1373         fuse_request_send(fc, req);
1374         err = req->out.h.error;
1375         /* locking is restartable */
1376         if (err == -EINTR)
1377                 err = -ERESTARTSYS;
1378         fuse_put_request(fc, req);
1379         return err;
1380 }
1381
1382 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
1383 {
1384         struct inode *inode = file->f_path.dentry->d_inode;
1385         struct fuse_conn *fc = get_fuse_conn(inode);
1386         int err;
1387
1388         if (cmd == F_CANCELLK) {
1389                 err = 0;
1390         } else if (cmd == F_GETLK) {
1391                 if (fc->no_lock) {
1392                         posix_test_lock(file, fl);
1393                         err = 0;
1394                 } else
1395                         err = fuse_getlk(file, fl);
1396         } else {
1397                 if (fc->no_lock)
1398                         err = posix_lock_file(file, fl, NULL);
1399                 else
1400                         err = fuse_setlk(file, fl, 0);
1401         }
1402         return err;
1403 }
1404
1405 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
1406 {
1407         struct inode *inode = file->f_path.dentry->d_inode;
1408         struct fuse_conn *fc = get_fuse_conn(inode);
1409         int err;
1410
1411         if (fc->no_lock) {
1412                 err = flock_lock_file_wait(file, fl);
1413         } else {
1414                 /* emulate flock with POSIX locks */
1415                 fl->fl_owner = (fl_owner_t) file;
1416                 err = fuse_setlk(file, fl, 1);
1417         }
1418
1419         return err;
1420 }
1421
1422 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
1423 {
1424         struct inode *inode = mapping->host;
1425         struct fuse_conn *fc = get_fuse_conn(inode);
1426         struct fuse_req *req;
1427         struct fuse_bmap_in inarg;
1428         struct fuse_bmap_out outarg;
1429         int err;
1430
1431         if (!inode->i_sb->s_bdev || fc->no_bmap)
1432                 return 0;
1433
1434         req = fuse_get_req(fc);
1435         if (IS_ERR(req))
1436                 return 0;
1437
1438         memset(&inarg, 0, sizeof(inarg));
1439         inarg.block = block;
1440         inarg.blocksize = inode->i_sb->s_blocksize;
1441         req->in.h.opcode = FUSE_BMAP;
1442         req->in.h.nodeid = get_node_id(inode);
1443         req->in.numargs = 1;
1444         req->in.args[0].size = sizeof(inarg);
1445         req->in.args[0].value = &inarg;
1446         req->out.numargs = 1;
1447         req->out.args[0].size = sizeof(outarg);
1448         req->out.args[0].value = &outarg;
1449         fuse_request_send(fc, req);
1450         err = req->out.h.error;
1451         fuse_put_request(fc, req);
1452         if (err == -ENOSYS)
1453                 fc->no_bmap = 1;
1454
1455         return err ? 0 : outarg.block;
1456 }
1457
1458 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int origin)
1459 {
1460         loff_t retval;
1461         struct inode *inode = file->f_path.dentry->d_inode;
1462
1463         mutex_lock(&inode->i_mutex);
1464         switch (origin) {
1465         case SEEK_END:
1466                 retval = fuse_update_attributes(inode, NULL, file, NULL);
1467                 if (retval)
1468                         return retval;
1469                 offset += i_size_read(inode);
1470                 break;
1471         case SEEK_CUR:
1472                 offset += file->f_pos;
1473         }
1474         retval = -EINVAL;
1475         if (offset >= 0 && offset <= inode->i_sb->s_maxbytes) {
1476                 if (offset != file->f_pos) {
1477                         file->f_pos = offset;
1478                         file->f_version = 0;
1479                 }
1480                 retval = offset;
1481         }
1482         mutex_unlock(&inode->i_mutex);
1483         return retval;
1484 }
1485
1486 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
1487                         unsigned int nr_segs, size_t bytes, bool to_user)
1488 {
1489         struct iov_iter ii;
1490         int page_idx = 0;
1491
1492         if (!bytes)
1493                 return 0;
1494
1495         iov_iter_init(&ii, iov, nr_segs, bytes, 0);
1496
1497         while (iov_iter_count(&ii)) {
1498                 struct page *page = pages[page_idx++];
1499                 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
1500                 void *kaddr, *map;
1501
1502                 kaddr = map = kmap(page);
1503
1504                 while (todo) {
1505                         char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
1506                         size_t iov_len = ii.iov->iov_len - ii.iov_offset;
1507                         size_t copy = min(todo, iov_len);
1508                         size_t left;
1509
1510                         if (!to_user)
1511                                 left = copy_from_user(kaddr, uaddr, copy);
1512                         else
1513                                 left = copy_to_user(uaddr, kaddr, copy);
1514
1515                         if (unlikely(left))
1516                                 return -EFAULT;
1517
1518                         iov_iter_advance(&ii, copy);
1519                         todo -= copy;
1520                         kaddr += copy;
1521                 }
1522
1523                 kunmap(map);
1524         }
1525
1526         return 0;
1527 }
1528
1529 /*
1530  * For ioctls, there is no generic way to determine how much memory
1531  * needs to be read and/or written.  Furthermore, ioctls are allowed
1532  * to dereference the passed pointer, so the parameter requires deep
1533  * copying but FUSE has no idea whatsoever about what to copy in or
1534  * out.
1535  *
1536  * This is solved by allowing FUSE server to retry ioctl with
1537  * necessary in/out iovecs.  Let's assume the ioctl implementation
1538  * needs to read in the following structure.
1539  *
1540  * struct a {
1541  *      char    *buf;
1542  *      size_t  buflen;
1543  * }
1544  *
1545  * On the first callout to FUSE server, inarg->in_size and
1546  * inarg->out_size will be NULL; then, the server completes the ioctl
1547  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
1548  * the actual iov array to
1549  *
1550  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
1551  *
1552  * which tells FUSE to copy in the requested area and retry the ioctl.
1553  * On the second round, the server has access to the structure and
1554  * from that it can tell what to look for next, so on the invocation,
1555  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
1556  *
1557  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
1558  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
1559  *
1560  * FUSE will copy both struct a and the pointed buffer from the
1561  * process doing the ioctl and retry ioctl with both struct a and the
1562  * buffer.
1563  *
1564  * This time, FUSE server has everything it needs and completes ioctl
1565  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
1566  *
1567  * Copying data out works the same way.
1568  *
1569  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
1570  * automatically initializes in and out iovs by decoding @cmd with
1571  * _IOC_* macros and the server is not allowed to request RETRY.  This
1572  * limits ioctl data transfers to well-formed ioctls and is the forced
1573  * behavior for all FUSE servers.
1574  */
1575 static long fuse_file_do_ioctl(struct file *file, unsigned int cmd,
1576                                unsigned long arg, unsigned int flags)
1577 {
1578         struct inode *inode = file->f_dentry->d_inode;
1579         struct fuse_file *ff = file->private_data;
1580         struct fuse_conn *fc = get_fuse_conn(inode);
1581         struct fuse_ioctl_in inarg = {
1582                 .fh = ff->fh,
1583                 .cmd = cmd,
1584                 .arg = arg,
1585                 .flags = flags
1586         };
1587         struct fuse_ioctl_out outarg;
1588         struct fuse_req *req = NULL;
1589         struct page **pages = NULL;
1590         struct page *iov_page = NULL;
1591         struct iovec *in_iov = NULL, *out_iov = NULL;
1592         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
1593         size_t in_size, out_size, transferred;
1594         int err;
1595
1596         /* assume all the iovs returned by client always fits in a page */
1597         BUILD_BUG_ON(sizeof(struct iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
1598
1599         if (!fuse_allow_task(fc, current))
1600                 return -EACCES;
1601
1602         err = -EIO;
1603         if (is_bad_inode(inode))
1604                 goto out;
1605
1606         err = -ENOMEM;
1607         pages = kzalloc(sizeof(pages[0]) * FUSE_MAX_PAGES_PER_REQ, GFP_KERNEL);
1608         iov_page = alloc_page(GFP_KERNEL);
1609         if (!pages || !iov_page)
1610                 goto out;
1611
1612         /*
1613          * If restricted, initialize IO parameters as encoded in @cmd.
1614          * RETRY from server is not allowed.
1615          */
1616         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
1617                 struct iovec *iov = page_address(iov_page);
1618
1619                 iov->iov_base = (void __user *)arg;
1620                 iov->iov_len = _IOC_SIZE(cmd);
1621
1622                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
1623                         in_iov = iov;
1624                         in_iovs = 1;
1625                 }
1626
1627                 if (_IOC_DIR(cmd) & _IOC_READ) {
1628                         out_iov = iov;
1629                         out_iovs = 1;
1630                 }
1631         }
1632
1633  retry:
1634         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
1635         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
1636
1637         /*
1638          * Out data can be used either for actual out data or iovs,
1639          * make sure there always is at least one page.
1640          */
1641         out_size = max_t(size_t, out_size, PAGE_SIZE);
1642         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
1643
1644         /* make sure there are enough buffer pages and init request with them */
1645         err = -ENOMEM;
1646         if (max_pages > FUSE_MAX_PAGES_PER_REQ)
1647                 goto out;
1648         while (num_pages < max_pages) {
1649                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
1650                 if (!pages[num_pages])
1651                         goto out;
1652                 num_pages++;
1653         }
1654
1655         req = fuse_get_req(fc);
1656         if (IS_ERR(req)) {
1657                 err = PTR_ERR(req);
1658                 req = NULL;
1659                 goto out;
1660         }
1661         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
1662         req->num_pages = num_pages;
1663
1664         /* okay, let's send it to the client */
1665         req->in.h.opcode = FUSE_IOCTL;
1666         req->in.h.nodeid = get_node_id(inode);
1667         req->in.numargs = 1;
1668         req->in.args[0].size = sizeof(inarg);
1669         req->in.args[0].value = &inarg;
1670         if (in_size) {
1671                 req->in.numargs++;
1672                 req->in.args[1].size = in_size;
1673                 req->in.argpages = 1;
1674
1675                 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
1676                                            false);
1677                 if (err)
1678                         goto out;
1679         }
1680
1681         req->out.numargs = 2;
1682         req->out.args[0].size = sizeof(outarg);
1683         req->out.args[0].value = &outarg;
1684         req->out.args[1].size = out_size;
1685         req->out.argpages = 1;
1686         req->out.argvar = 1;
1687
1688         fuse_request_send(fc, req);
1689         err = req->out.h.error;
1690         transferred = req->out.args[1].size;
1691         fuse_put_request(fc, req);
1692         req = NULL;
1693         if (err)
1694                 goto out;
1695
1696         /* did it ask for retry? */
1697         if (outarg.flags & FUSE_IOCTL_RETRY) {
1698                 char *vaddr;
1699
1700                 /* no retry if in restricted mode */
1701                 err = -EIO;
1702                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
1703                         goto out;
1704
1705                 in_iovs = outarg.in_iovs;
1706                 out_iovs = outarg.out_iovs;
1707
1708                 /*
1709                  * Make sure things are in boundary, separate checks
1710                  * are to protect against overflow.
1711                  */
1712                 err = -ENOMEM;
1713                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
1714                     out_iovs > FUSE_IOCTL_MAX_IOV ||
1715                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
1716                         goto out;
1717
1718                 err = -EIO;
1719                 if ((in_iovs + out_iovs) * sizeof(struct iovec) != transferred)
1720                         goto out;
1721
1722                 /* okay, copy in iovs and retry */
1723                 vaddr = kmap_atomic(pages[0], KM_USER0);
1724                 memcpy(page_address(iov_page), vaddr, transferred);
1725                 kunmap_atomic(vaddr, KM_USER0);
1726
1727                 in_iov = page_address(iov_page);
1728                 out_iov = in_iov + in_iovs;
1729
1730                 goto retry;
1731         }
1732
1733         err = -EIO;
1734         if (transferred > inarg.out_size)
1735                 goto out;
1736
1737         err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
1738  out:
1739         if (req)
1740                 fuse_put_request(fc, req);
1741         if (iov_page)
1742                 __free_page(iov_page);
1743         while (num_pages)
1744                 __free_page(pages[--num_pages]);
1745         kfree(pages);
1746
1747         return err ? err : outarg.result;
1748 }
1749
1750 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
1751                             unsigned long arg)
1752 {
1753         return fuse_file_do_ioctl(file, cmd, arg, 0);
1754 }
1755
1756 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
1757                                    unsigned long arg)
1758 {
1759         return fuse_file_do_ioctl(file, cmd, arg, FUSE_IOCTL_COMPAT);
1760 }
1761
1762 /*
1763  * All files which have been polled are linked to RB tree
1764  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
1765  * find the matching one.
1766  */
1767 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
1768                                               struct rb_node **parent_out)
1769 {
1770         struct rb_node **link = &fc->polled_files.rb_node;
1771         struct rb_node *last = NULL;
1772
1773         while (*link) {
1774                 struct fuse_file *ff;
1775
1776                 last = *link;
1777                 ff = rb_entry(last, struct fuse_file, polled_node);
1778
1779                 if (kh < ff->kh)
1780                         link = &last->rb_left;
1781                 else if (kh > ff->kh)
1782                         link = &last->rb_right;
1783                 else
1784                         return link;
1785         }
1786
1787         if (parent_out)
1788                 *parent_out = last;
1789         return link;
1790 }
1791
1792 /*
1793  * The file is about to be polled.  Make sure it's on the polled_files
1794  * RB tree.  Note that files once added to the polled_files tree are
1795  * not removed before the file is released.  This is because a file
1796  * polled once is likely to be polled again.
1797  */
1798 static void fuse_register_polled_file(struct fuse_conn *fc,
1799                                       struct fuse_file *ff)
1800 {
1801         spin_lock(&fc->lock);
1802         if (RB_EMPTY_NODE(&ff->polled_node)) {
1803                 struct rb_node **link, *parent;
1804
1805                 link = fuse_find_polled_node(fc, ff->kh, &parent);
1806                 BUG_ON(*link);
1807                 rb_link_node(&ff->polled_node, parent, link);
1808                 rb_insert_color(&ff->polled_node, &fc->polled_files);
1809         }
1810         spin_unlock(&fc->lock);
1811 }
1812
1813 static unsigned fuse_file_poll(struct file *file, poll_table *wait)
1814 {
1815         struct inode *inode = file->f_dentry->d_inode;
1816         struct fuse_file *ff = file->private_data;
1817         struct fuse_conn *fc = get_fuse_conn(inode);
1818         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
1819         struct fuse_poll_out outarg;
1820         struct fuse_req *req;
1821         int err;
1822
1823         if (fc->no_poll)
1824                 return DEFAULT_POLLMASK;
1825
1826         poll_wait(file, &ff->poll_wait, wait);
1827
1828         /*
1829          * Ask for notification iff there's someone waiting for it.
1830          * The client may ignore the flag and always notify.
1831          */
1832         if (waitqueue_active(&ff->poll_wait)) {
1833                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
1834                 fuse_register_polled_file(fc, ff);
1835         }
1836
1837         req = fuse_get_req(fc);
1838         if (IS_ERR(req))
1839                 return PTR_ERR(req);
1840
1841         req->in.h.opcode = FUSE_POLL;
1842         req->in.h.nodeid = get_node_id(inode);
1843         req->in.numargs = 1;
1844         req->in.args[0].size = sizeof(inarg);
1845         req->in.args[0].value = &inarg;
1846         req->out.numargs = 1;
1847         req->out.args[0].size = sizeof(outarg);
1848         req->out.args[0].value = &outarg;
1849         fuse_request_send(fc, req);
1850         err = req->out.h.error;
1851         fuse_put_request(fc, req);
1852
1853         if (!err)
1854                 return outarg.revents;
1855         if (err == -ENOSYS) {
1856                 fc->no_poll = 1;
1857                 return DEFAULT_POLLMASK;
1858         }
1859         return POLLERR;
1860 }
1861
1862 /*
1863  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
1864  * wakes up the poll waiters.
1865  */
1866 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
1867                             struct fuse_notify_poll_wakeup_out *outarg)
1868 {
1869         u64 kh = outarg->kh;
1870         struct rb_node **link;
1871
1872         spin_lock(&fc->lock);
1873
1874         link = fuse_find_polled_node(fc, kh, NULL);
1875         if (*link) {
1876                 struct fuse_file *ff;
1877
1878                 ff = rb_entry(*link, struct fuse_file, polled_node);
1879                 wake_up_interruptible_sync(&ff->poll_wait);
1880         }
1881
1882         spin_unlock(&fc->lock);
1883         return 0;
1884 }
1885
1886 static const struct file_operations fuse_file_operations = {
1887         .llseek         = fuse_file_llseek,
1888         .read           = do_sync_read,
1889         .aio_read       = fuse_file_aio_read,
1890         .write          = do_sync_write,
1891         .aio_write      = fuse_file_aio_write,
1892         .mmap           = fuse_file_mmap,
1893         .open           = fuse_open,
1894         .flush          = fuse_flush,
1895         .release        = fuse_release,
1896         .fsync          = fuse_fsync,
1897         .lock           = fuse_file_lock,
1898         .flock          = fuse_file_flock,
1899         .splice_read    = generic_file_splice_read,
1900         .unlocked_ioctl = fuse_file_ioctl,
1901         .compat_ioctl   = fuse_file_compat_ioctl,
1902         .poll           = fuse_file_poll,
1903 };
1904
1905 static const struct file_operations fuse_direct_io_file_operations = {
1906         .llseek         = fuse_file_llseek,
1907         .read           = fuse_direct_read,
1908         .write          = fuse_direct_write,
1909         .open           = fuse_open,
1910         .flush          = fuse_flush,
1911         .release        = fuse_release,
1912         .fsync          = fuse_fsync,
1913         .lock           = fuse_file_lock,
1914         .flock          = fuse_file_flock,
1915         .unlocked_ioctl = fuse_file_ioctl,
1916         .compat_ioctl   = fuse_file_compat_ioctl,
1917         .poll           = fuse_file_poll,
1918         /* no mmap and splice_read */
1919 };
1920
1921 static const struct address_space_operations fuse_file_aops  = {
1922         .readpage       = fuse_readpage,
1923         .writepage      = fuse_writepage,
1924         .launder_page   = fuse_launder_page,
1925         .write_begin    = fuse_write_begin,
1926         .write_end      = fuse_write_end,
1927         .readpages      = fuse_readpages,
1928         .set_page_dirty = __set_page_dirty_nobuffers,
1929         .bmap           = fuse_bmap,
1930 };
1931
1932 void fuse_init_file_inode(struct inode *inode)
1933 {
1934         inode->i_fop = &fuse_file_operations;
1935         inode->i_data.a_ops = &fuse_file_aops;
1936 }