?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
ocfs2: fix locking for res->tracking and dlm->tracking_list
[linux/fpc-iii.git] / fs / fuse / file.c
blob8577f3ba6dc6f34d9b680aa56954da6938d05f3a
1 /*
2 FUSE: Filesystem in Userspace
3 Copyright (C) 2001-2008 Miklos Szeredi <miklos@szeredi.hu>
4
5 This program can be distributed under the terms of the GNU GPL.
6 See the file COPYING.
7 */
8
9 #include "fuse_i.h"
10
11 #include <linux/pagemap.h>
12 #include <linux/slab.h>
13 #include <linux/kernel.h>
14 #include <linux/sched.h>
15 #include <linux/module.h>
16 #include <linux/compat.h>
17 #include <linux/swap.h>
18 #include <linux/falloc.h>
19 #include <linux/uio.h>
20
21 static const struct file_operations fuse_direct_io_file_operations;
22
23 static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
24 int opcode, struct fuse_open_out *outargp)
25 {
26 struct fuse_open_in inarg;
27 FUSE_ARGS(args);
28
29 memset(&inarg, 0, sizeof(inarg));
30 inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
31 if (!fc->atomic_o_trunc)
32 inarg.flags &= ~O_TRUNC;
33 args.in.h.opcode = opcode;
34 args.in.h.nodeid = nodeid;
35 args.in.numargs = 1;
36 args.in.args[0].size = sizeof(inarg);
37 args.in.args[0].value = &inarg;
38 args.out.numargs = 1;
39 args.out.args[0].size = sizeof(*outargp);
40 args.out.args[0].value = outargp;
41
42 return fuse_simple_request(fc, &args);
43 }
44
45 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
46 {
47 struct fuse_file *ff;
48
49 ff = kzalloc(sizeof(struct fuse_file), GFP_KERNEL);
50 if (unlikely(!ff))
51 return NULL;
52
53 ff->fc = fc;
54 ff->reserved_req = fuse_request_alloc(0);
55 if (unlikely(!ff->reserved_req)) {
56 kfree(ff);
57 return NULL;
58 }
59
60 INIT_LIST_HEAD(&ff->write_entry);
61 atomic_set(&ff->count, 0);
62 RB_CLEAR_NODE(&ff->polled_node);
63 init_waitqueue_head(&ff->poll_wait);
64
65 spin_lock(&fc->lock);
66 ff->kh = ++fc->khctr;
67 spin_unlock(&fc->lock);
68
69 return ff;
70 }
71
72 void fuse_file_free(struct fuse_file *ff)
73 {
74 fuse_request_free(ff->reserved_req);
75 kfree(ff);
76 }
77
78 struct fuse_file *fuse_file_get(struct fuse_file *ff)
79 {
80 atomic_inc(&ff->count);
81 return ff;
82 }
83
84 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
85 {
86 iput(req->misc.release.inode);
87 }
88
89 static void fuse_file_put(struct fuse_file *ff, bool sync)
90 {
91 if (atomic_dec_and_test(&ff->count)) {
92 struct fuse_req *req = ff->reserved_req;
93
94 if (ff->fc->no_open) {
95 /*
96 * Drop the release request when client does not
97 * implement 'open'
98 */
99 __clear_bit(FR_BACKGROUND, &req->flags);
100 iput(req->misc.release.inode);
101 fuse_put_request(ff->fc, req);
102 } else if (sync) {
103 __set_bit(FR_FORCE, &req->flags);
104 __clear_bit(FR_BACKGROUND, &req->flags);
105 fuse_request_send(ff->fc, req);
106 iput(req->misc.release.inode);
107 fuse_put_request(ff->fc, req);
108 } else {
109 req->end = fuse_release_end;
110 __set_bit(FR_BACKGROUND, &req->flags);
111 fuse_request_send_background(ff->fc, req);
112 }
113 kfree(ff);
114 }
115 }
116
117 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
118 bool isdir)
119 {
120 struct fuse_file *ff;
121 int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
122
123 ff = fuse_file_alloc(fc);
124 if (!ff)
125 return -ENOMEM;
126
127 ff->fh = 0;
128 ff->open_flags = FOPEN_KEEP_CACHE; /* Default for no-open */
129 if (!fc->no_open || isdir) {
130 struct fuse_open_out outarg;
131 int err;
132
133 err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
134 if (!err) {
135 ff->fh = outarg.fh;
136 ff->open_flags = outarg.open_flags;
137
138 } else if (err != -ENOSYS || isdir) {
139 fuse_file_free(ff);
140 return err;
141 } else {
142 fc->no_open = 1;
143 }
144 }
145
146 if (isdir)
147 ff->open_flags &= ~FOPEN_DIRECT_IO;
148
149 ff->nodeid = nodeid;
150 file->private_data = fuse_file_get(ff);
151
152 return 0;
153 }
154 EXPORT_SYMBOL_GPL(fuse_do_open);
155
156 static void fuse_link_write_file(struct file *file)
157 {
158 struct inode *inode = file_inode(file);
159 struct fuse_conn *fc = get_fuse_conn(inode);
160 struct fuse_inode *fi = get_fuse_inode(inode);
161 struct fuse_file *ff = file->private_data;
162 /*
163 * file may be written through mmap, so chain it onto the
164 * inodes's write_file list
165 */
166 spin_lock(&fc->lock);
167 if (list_empty(&ff->write_entry))
168 list_add(&ff->write_entry, &fi->write_files);
169 spin_unlock(&fc->lock);
170 }
171
172 void fuse_finish_open(struct inode *inode, struct file *file)
173 {
174 struct fuse_file *ff = file->private_data;
175 struct fuse_conn *fc = get_fuse_conn(inode);
176
177 if (ff->open_flags & FOPEN_DIRECT_IO)
178 file->f_op = &fuse_direct_io_file_operations;
179 if (!(ff->open_flags & FOPEN_KEEP_CACHE))
180 invalidate_inode_pages2(inode->i_mapping);
181 if (ff->open_flags & FOPEN_NONSEEKABLE)
182 nonseekable_open(inode, file);
183 if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
184 struct fuse_inode *fi = get_fuse_inode(inode);
185
186 spin_lock(&fc->lock);
187 fi->attr_version = ++fc->attr_version;
188 i_size_write(inode, 0);
189 spin_unlock(&fc->lock);
190 fuse_invalidate_attr(inode);
191 if (fc->writeback_cache)
192 file_update_time(file);
193 }
194 if ((file->f_mode & FMODE_WRITE) && fc->writeback_cache)
195 fuse_link_write_file(file);
196 }
197
198 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
199 {
200 struct fuse_conn *fc = get_fuse_conn(inode);
201 int err;
202 bool lock_inode = (file->f_flags & O_TRUNC) &&
203 fc->atomic_o_trunc &&
204 fc->writeback_cache;
205
206 err = generic_file_open(inode, file);
207 if (err)
208 return err;
209
210 if (lock_inode)
211 mutex_lock(&inode->i_mutex);
212
213 err = fuse_do_open(fc, get_node_id(inode), file, isdir);
214
215 if (!err)
216 fuse_finish_open(inode, file);
217
218 if (lock_inode)
219 mutex_unlock(&inode->i_mutex);
220
221 return err;
222 }
223
224 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
225 {
226 struct fuse_conn *fc = ff->fc;
227 struct fuse_req *req = ff->reserved_req;
228 struct fuse_release_in *inarg = &req->misc.release.in;
229
230 spin_lock(&fc->lock);
231 list_del(&ff->write_entry);
232 if (!RB_EMPTY_NODE(&ff->polled_node))
233 rb_erase(&ff->polled_node, &fc->polled_files);
234 spin_unlock(&fc->lock);
235
236 wake_up_interruptible_all(&ff->poll_wait);
237
238 inarg->fh = ff->fh;
239 inarg->flags = flags;
240 req->in.h.opcode = opcode;
241 req->in.h.nodeid = ff->nodeid;
242 req->in.numargs = 1;
243 req->in.args[0].size = sizeof(struct fuse_release_in);
244 req->in.args[0].value = inarg;
245 }
246
247 void fuse_release_common(struct file *file, int opcode)
248 {
249 struct fuse_file *ff;
250 struct fuse_req *req;
251
252 ff = file->private_data;
253 if (unlikely(!ff))
254 return;
255
256 req = ff->reserved_req;
257 fuse_prepare_release(ff, file->f_flags, opcode);
258
259 if (ff->flock) {
260 struct fuse_release_in *inarg = &req->misc.release.in;
261 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
262 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
263 (fl_owner_t) file);
264 }
265 /* Hold inode until release is finished */
266 req->misc.release.inode = igrab(file_inode(file));
267
268 /*
269 * Normally this will send the RELEASE request, however if
270 * some asynchronous READ or WRITE requests are outstanding,
271 * the sending will be delayed.
272 *
273 * Make the release synchronous if this is a fuseblk mount,
274 * synchronous RELEASE is allowed (and desirable) in this case
275 * because the server can be trusted not to screw up.
276 */
277 fuse_file_put(ff, ff->fc->destroy_req != NULL);
278 }
279
280 static int fuse_open(struct inode *inode, struct file *file)
281 {
282 return fuse_open_common(inode, file, false);
283 }
284
285 static int fuse_release(struct inode *inode, struct file *file)
286 {
287 struct fuse_conn *fc = get_fuse_conn(inode);
288
289 /* see fuse_vma_close() for !writeback_cache case */
290 if (fc->writeback_cache)
291 write_inode_now(inode, 1);
292
293 fuse_release_common(file, FUSE_RELEASE);
294
295 /* return value is ignored by VFS */
296 return 0;
297 }
298
299 void fuse_sync_release(struct fuse_file *ff, int flags)
300 {
301 WARN_ON(atomic_read(&ff->count) > 1);
302 fuse_prepare_release(ff, flags, FUSE_RELEASE);
303 __set_bit(FR_FORCE, &ff->reserved_req->flags);
304 __clear_bit(FR_BACKGROUND, &ff->reserved_req->flags);
305 fuse_request_send(ff->fc, ff->reserved_req);
306 fuse_put_request(ff->fc, ff->reserved_req);
307 kfree(ff);
308 }
309 EXPORT_SYMBOL_GPL(fuse_sync_release);
310
311 /*
312 * Scramble the ID space with XTEA, so that the value of the files_struct
313 * pointer is not exposed to userspace.
314 */
315 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
316 {
317 u32 *k = fc->scramble_key;
318 u64 v = (unsigned long) id;
319 u32 v0 = v;
320 u32 v1 = v >> 32;
321 u32 sum = 0;
322 int i;
323
324 for (i = 0; i < 32; i++) {
325 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
326 sum += 0x9E3779B9;
327 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
328 }
329
330 return (u64) v0 + ((u64) v1 << 32);
331 }
332
333 /*
334 * Check if any page in a range is under writeback
335 *
336 * This is currently done by walking the list of writepage requests
337 * for the inode, which can be pretty inefficient.
338 */
339 static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
340 pgoff_t idx_to)
341 {
342 struct fuse_conn *fc = get_fuse_conn(inode);
343 struct fuse_inode *fi = get_fuse_inode(inode);
344 struct fuse_req *req;
345 bool found = false;
346
347 spin_lock(&fc->lock);
348 list_for_each_entry(req, &fi->writepages, writepages_entry) {
349 pgoff_t curr_index;
350
351 BUG_ON(req->inode != inode);
352 curr_index = req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
353 if (idx_from < curr_index + req->num_pages &&
354 curr_index <= idx_to) {
355 found = true;
356 break;
357 }
358 }
359 spin_unlock(&fc->lock);
360
361 return found;
362 }
363
364 static inline bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
365 {
366 return fuse_range_is_writeback(inode, index, index);
367 }
368
369 /*
370 * Wait for page writeback to be completed.
371 *
372 * Since fuse doesn't rely on the VM writeback tracking, this has to
373 * use some other means.
374 */
375 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
376 {
377 struct fuse_inode *fi = get_fuse_inode(inode);
378
379 wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
380 return 0;
381 }
382
383 /*
384 * Wait for all pending writepages on the inode to finish.
385 *
386 * This is currently done by blocking further writes with FUSE_NOWRITE
387 * and waiting for all sent writes to complete.
388 *
389 * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
390 * could conflict with truncation.
391 */
392 static void fuse_sync_writes(struct inode *inode)
393 {
394 fuse_set_nowrite(inode);
395 fuse_release_nowrite(inode);
396 }
397
398 static int fuse_flush(struct file *file, fl_owner_t id)
399 {
400 struct inode *inode = file_inode(file);
401 struct fuse_conn *fc = get_fuse_conn(inode);
402 struct fuse_file *ff = file->private_data;
403 struct fuse_req *req;
404 struct fuse_flush_in inarg;
405 int err;
406
407 if (is_bad_inode(inode))
408 return -EIO;
409
410 if (fc->no_flush)
411 return 0;
412
413 err = write_inode_now(inode, 1);
414 if (err)
415 return err;
416
417 mutex_lock(&inode->i_mutex);
418 fuse_sync_writes(inode);
419 mutex_unlock(&inode->i_mutex);
420
421 if (test_bit(AS_ENOSPC, &file->f_mapping->flags) &&
422 test_and_clear_bit(AS_ENOSPC, &file->f_mapping->flags))
423 err = -ENOSPC;
424 if (test_bit(AS_EIO, &file->f_mapping->flags) &&
425 test_and_clear_bit(AS_EIO, &file->f_mapping->flags))
426 err = -EIO;
427 if (err)
428 return err;
429
430 req = fuse_get_req_nofail_nopages(fc, file);
431 memset(&inarg, 0, sizeof(inarg));
432 inarg.fh = ff->fh;
433 inarg.lock_owner = fuse_lock_owner_id(fc, id);
434 req->in.h.opcode = FUSE_FLUSH;
435 req->in.h.nodeid = get_node_id(inode);
436 req->in.numargs = 1;
437 req->in.args[0].size = sizeof(inarg);
438 req->in.args[0].value = &inarg;
439 __set_bit(FR_FORCE, &req->flags);
440 fuse_request_send(fc, req);
441 err = req->out.h.error;
442 fuse_put_request(fc, req);
443 if (err == -ENOSYS) {
444 fc->no_flush = 1;
445 err = 0;
446 }
447 return err;
448 }
449
450 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
451 int datasync, int isdir)
452 {
453 struct inode *inode = file->f_mapping->host;
454 struct fuse_conn *fc = get_fuse_conn(inode);
455 struct fuse_file *ff = file->private_data;
456 FUSE_ARGS(args);
457 struct fuse_fsync_in inarg;
458 int err;
459
460 if (is_bad_inode(inode))
461 return -EIO;
462
463 mutex_lock(&inode->i_mutex);
464
465 /*
466 * Start writeback against all dirty pages of the inode, then
467 * wait for all outstanding writes, before sending the FSYNC
468 * request.
469 */
470 err = filemap_write_and_wait_range(inode->i_mapping, start, end);
471 if (err)
472 goto out;
473
474 fuse_sync_writes(inode);
475
476 /*
477 * Due to implementation of fuse writeback
478 * filemap_write_and_wait_range() does not catch errors.
479 * We have to do this directly after fuse_sync_writes()
480 */
481 if (test_bit(AS_ENOSPC, &file->f_mapping->flags) &&
482 test_and_clear_bit(AS_ENOSPC, &file->f_mapping->flags))
483 err = -ENOSPC;
484 if (test_bit(AS_EIO, &file->f_mapping->flags) &&
485 test_and_clear_bit(AS_EIO, &file->f_mapping->flags))
486 err = -EIO;
487 if (err)
488 goto out;
489
490 err = sync_inode_metadata(inode, 1);
491 if (err)
492 goto out;
493
494 if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
495 goto out;
496
497 memset(&inarg, 0, sizeof(inarg));
498 inarg.fh = ff->fh;
499 inarg.fsync_flags = datasync ? 1 : 0;
500 args.in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
501 args.in.h.nodeid = get_node_id(inode);
502 args.in.numargs = 1;
503 args.in.args[0].size = sizeof(inarg);
504 args.in.args[0].value = &inarg;
505 err = fuse_simple_request(fc, &args);
506 if (err == -ENOSYS) {
507 if (isdir)
508 fc->no_fsyncdir = 1;
509 else
510 fc->no_fsync = 1;
511 err = 0;
512 }
513 out:
514 mutex_unlock(&inode->i_mutex);
515 return err;
516 }
517
518 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
519 int datasync)
520 {
521 return fuse_fsync_common(file, start, end, datasync, 0);
522 }
523
524 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
525 size_t count, int opcode)
526 {
527 struct fuse_read_in *inarg = &req->misc.read.in;
528 struct fuse_file *ff = file->private_data;
529
530 inarg->fh = ff->fh;
531 inarg->offset = pos;
532 inarg->size = count;
533 inarg->flags = file->f_flags;
534 req->in.h.opcode = opcode;
535 req->in.h.nodeid = ff->nodeid;
536 req->in.numargs = 1;
537 req->in.args[0].size = sizeof(struct fuse_read_in);
538 req->in.args[0].value = inarg;
539 req->out.argvar = 1;
540 req->out.numargs = 1;
541 req->out.args[0].size = count;
542 }
543
544 static void fuse_release_user_pages(struct fuse_req *req, bool should_dirty)
545 {
546 unsigned i;
547
548 for (i = 0; i < req->num_pages; i++) {
549 struct page *page = req->pages[i];
550 if (should_dirty)
551 set_page_dirty_lock(page);
552 put_page(page);
553 }
554 }
555
556 static void fuse_io_release(struct kref *kref)
557 {
558 kfree(container_of(kref, struct fuse_io_priv, refcnt));
559 }
560
561 static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
562 {
563 if (io->err)
564 return io->err;
565
566 if (io->bytes >= 0 && io->write)
567 return -EIO;
568
569 return io->bytes < 0 ? io->size : io->bytes;
570 }
571
572 /**
573 * In case of short read, the caller sets 'pos' to the position of
574 * actual end of fuse request in IO request. Otherwise, if bytes_requested
575 * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
576 *
577 * An example:
578 * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
579 * both submitted asynchronously. The first of them was ACKed by userspace as
580 * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
581 * second request was ACKed as short, e.g. only 1K was read, resulting in
582 * pos == 33K.
583 *
584 * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
585 * will be equal to the length of the longest contiguous fragment of
586 * transferred data starting from the beginning of IO request.
587 */
588 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
589 {
590 bool is_sync = is_sync_kiocb(io->iocb);
591 int left;
592
593 spin_lock(&io->lock);
594 if (err)
595 io->err = io->err ? : err;
596 else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
597 io->bytes = pos;
598
599 left = --io->reqs;
600 if (!left && is_sync)
601 complete(io->done);
602 spin_unlock(&io->lock);
603
604 if (!left && !is_sync) {
605 ssize_t res = fuse_get_res_by_io(io);
606
607 if (res >= 0) {
608 struct inode *inode = file_inode(io->iocb->ki_filp);
609 struct fuse_conn *fc = get_fuse_conn(inode);
610 struct fuse_inode *fi = get_fuse_inode(inode);
611
612 spin_lock(&fc->lock);
613 fi->attr_version = ++fc->attr_version;
614 spin_unlock(&fc->lock);
615 }
616
617 io->iocb->ki_complete(io->iocb, res, 0);
618 }
619
620 kref_put(&io->refcnt, fuse_io_release);
621 }
622
623 static void fuse_aio_complete_req(struct fuse_conn *fc, struct fuse_req *req)
624 {
625 struct fuse_io_priv *io = req->io;
626 ssize_t pos = -1;
627
628 fuse_release_user_pages(req, !io->write);
629
630 if (io->write) {
631 if (req->misc.write.in.size != req->misc.write.out.size)
632 pos = req->misc.write.in.offset - io->offset +
633 req->misc.write.out.size;
634 } else {
635 if (req->misc.read.in.size != req->out.args[0].size)
636 pos = req->misc.read.in.offset - io->offset +
637 req->out.args[0].size;
638 }
639
640 fuse_aio_complete(io, req->out.h.error, pos);
641 }
642
643 static size_t fuse_async_req_send(struct fuse_conn *fc, struct fuse_req *req,
644 size_t num_bytes, struct fuse_io_priv *io)
645 {
646 spin_lock(&io->lock);
647 kref_get(&io->refcnt);
648 io->size += num_bytes;
649 io->reqs++;
650 spin_unlock(&io->lock);
651
652 req->io = io;
653 req->end = fuse_aio_complete_req;
654
655 __fuse_get_request(req);
656 fuse_request_send_background(fc, req);
657
658 return num_bytes;
659 }
660
661 static size_t fuse_send_read(struct fuse_req *req, struct fuse_io_priv *io,
662 loff_t pos, size_t count, fl_owner_t owner)
663 {
664 struct file *file = io->file;
665 struct fuse_file *ff = file->private_data;
666 struct fuse_conn *fc = ff->fc;
667
668 fuse_read_fill(req, file, pos, count, FUSE_READ);
669 if (owner != NULL) {
670 struct fuse_read_in *inarg = &req->misc.read.in;
671
672 inarg->read_flags |= FUSE_READ_LOCKOWNER;
673 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
674 }
675
676 if (io->async)
677 return fuse_async_req_send(fc, req, count, io);
678
679 fuse_request_send(fc, req);
680 return req->out.args[0].size;
681 }
682
683 static void fuse_read_update_size(struct inode *inode, loff_t size,
684 u64 attr_ver)
685 {
686 struct fuse_conn *fc = get_fuse_conn(inode);
687 struct fuse_inode *fi = get_fuse_inode(inode);
688
689 spin_lock(&fc->lock);
690 if (attr_ver == fi->attr_version && size < inode->i_size &&
691 !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
692 fi->attr_version = ++fc->attr_version;
693 i_size_write(inode, size);
694 }
695 spin_unlock(&fc->lock);
696 }
697
698 static void fuse_short_read(struct fuse_req *req, struct inode *inode,
699 u64 attr_ver)
700 {
701 size_t num_read = req->out.args[0].size;
702 struct fuse_conn *fc = get_fuse_conn(inode);
703
704 if (fc->writeback_cache) {
705 /*
706 * A hole in a file. Some data after the hole are in page cache,
707 * but have not reached the client fs yet. So, the hole is not
708 * present there.
709 */
710 int i;
711 int start_idx = num_read >> PAGE_CACHE_SHIFT;
712 size_t off = num_read & (PAGE_CACHE_SIZE - 1);
713
714 for (i = start_idx; i < req->num_pages; i++) {
715 zero_user_segment(req->pages[i], off, PAGE_CACHE_SIZE);
716 off = 0;
717 }
718 } else {
719 loff_t pos = page_offset(req->pages[0]) + num_read;
720 fuse_read_update_size(inode, pos, attr_ver);
721 }
722 }
723
724 static int fuse_do_readpage(struct file *file, struct page *page)
725 {
726 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
727 struct inode *inode = page->mapping->host;
728 struct fuse_conn *fc = get_fuse_conn(inode);
729 struct fuse_req *req;
730 size_t num_read;
731 loff_t pos = page_offset(page);
732 size_t count = PAGE_CACHE_SIZE;
733 u64 attr_ver;
734 int err;
735
736 /*
737 * Page writeback can extend beyond the lifetime of the
738 * page-cache page, so make sure we read a properly synced
739 * page.
740 */
741 fuse_wait_on_page_writeback(inode, page->index);
742
743 req = fuse_get_req(fc, 1);
744 if (IS_ERR(req))
745 return PTR_ERR(req);
746
747 attr_ver = fuse_get_attr_version(fc);
748
749 req->out.page_zeroing = 1;
750 req->out.argpages = 1;
751 req->num_pages = 1;
752 req->pages[0] = page;
753 req->page_descs[0].length = count;
754 num_read = fuse_send_read(req, &io, pos, count, NULL);
755 err = req->out.h.error;
756
757 if (!err) {
758 /*
759 * Short read means EOF. If file size is larger, truncate it
760 */
761 if (num_read < count)
762 fuse_short_read(req, inode, attr_ver);
763
764 SetPageUptodate(page);
765 }
766
767 fuse_put_request(fc, req);
768
769 return err;
770 }
771
772 static int fuse_readpage(struct file *file, struct page *page)
773 {
774 struct inode *inode = page->mapping->host;
775 int err;
776
777 err = -EIO;
778 if (is_bad_inode(inode))
779 goto out;
780
781 err = fuse_do_readpage(file, page);
782 fuse_invalidate_atime(inode);
783 out:
784 unlock_page(page);
785 return err;
786 }
787
788 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
789 {
790 int i;
791 size_t count = req->misc.read.in.size;
792 size_t num_read = req->out.args[0].size;
793 struct address_space *mapping = NULL;
794
795 for (i = 0; mapping == NULL && i < req->num_pages; i++)
796 mapping = req->pages[i]->mapping;
797
798 if (mapping) {
799 struct inode *inode = mapping->host;
800
801 /*
802 * Short read means EOF. If file size is larger, truncate it
803 */
804 if (!req->out.h.error && num_read < count)
805 fuse_short_read(req, inode, req->misc.read.attr_ver);
806
807 fuse_invalidate_atime(inode);
808 }
809
810 for (i = 0; i < req->num_pages; i++) {
811 struct page *page = req->pages[i];
812 if (!req->out.h.error)
813 SetPageUptodate(page);
814 else
815 SetPageError(page);
816 unlock_page(page);
817 page_cache_release(page);
818 }
819 if (req->ff)
820 fuse_file_put(req->ff, false);
821 }
822
823 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
824 {
825 struct fuse_file *ff = file->private_data;
826 struct fuse_conn *fc = ff->fc;
827 loff_t pos = page_offset(req->pages[0]);
828 size_t count = req->num_pages << PAGE_CACHE_SHIFT;
829
830 req->out.argpages = 1;
831 req->out.page_zeroing = 1;
832 req->out.page_replace = 1;
833 fuse_read_fill(req, file, pos, count, FUSE_READ);
834 req->misc.read.attr_ver = fuse_get_attr_version(fc);
835 if (fc->async_read) {
836 req->ff = fuse_file_get(ff);
837 req->end = fuse_readpages_end;
838 fuse_request_send_background(fc, req);
839 } else {
840 fuse_request_send(fc, req);
841 fuse_readpages_end(fc, req);
842 fuse_put_request(fc, req);
843 }
844 }
845
846 struct fuse_fill_data {
847 struct fuse_req *req;
848 struct file *file;
849 struct inode *inode;
850 unsigned nr_pages;
851 };
852
853 static int fuse_readpages_fill(void *_data, struct page *page)
854 {
855 struct fuse_fill_data *data = _data;
856 struct fuse_req *req = data->req;
857 struct inode *inode = data->inode;
858 struct fuse_conn *fc = get_fuse_conn(inode);
859
860 fuse_wait_on_page_writeback(inode, page->index);
861
862 if (req->num_pages &&
863 (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
864 (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
865 req->pages[req->num_pages - 1]->index + 1 != page->index)) {
866 int nr_alloc = min_t(unsigned, data->nr_pages,
867 FUSE_MAX_PAGES_PER_REQ);
868 fuse_send_readpages(req, data->file);
869 if (fc->async_read)
870 req = fuse_get_req_for_background(fc, nr_alloc);
871 else
872 req = fuse_get_req(fc, nr_alloc);
873
874 data->req = req;
875 if (IS_ERR(req)) {
876 unlock_page(page);
877 return PTR_ERR(req);
878 }
879 }
880
881 if (WARN_ON(req->num_pages >= req->max_pages)) {
882 unlock_page(page);
883 fuse_put_request(fc, req);
884 return -EIO;
885 }
886
887 page_cache_get(page);
888 req->pages[req->num_pages] = page;
889 req->page_descs[req->num_pages].length = PAGE_SIZE;
890 req->num_pages++;
891 data->nr_pages--;
892 return 0;
893 }
894
895 static int fuse_readpages(struct file *file, struct address_space *mapping,
896 struct list_head *pages, unsigned nr_pages)
897 {
898 struct inode *inode = mapping->host;
899 struct fuse_conn *fc = get_fuse_conn(inode);
900 struct fuse_fill_data data;
901 int err;
902 int nr_alloc = min_t(unsigned, nr_pages, FUSE_MAX_PAGES_PER_REQ);
903
904 err = -EIO;
905 if (is_bad_inode(inode))
906 goto out;
907
908 data.file = file;
909 data.inode = inode;
910 if (fc->async_read)
911 data.req = fuse_get_req_for_background(fc, nr_alloc);
912 else
913 data.req = fuse_get_req(fc, nr_alloc);
914 data.nr_pages = nr_pages;
915 err = PTR_ERR(data.req);
916 if (IS_ERR(data.req))
917 goto out;
918
919 err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
920 if (!err) {
921 if (data.req->num_pages)
922 fuse_send_readpages(data.req, file);
923 else
924 fuse_put_request(fc, data.req);
925 }
926 out:
927 return err;
928 }
929
930 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
931 {
932 struct inode *inode = iocb->ki_filp->f_mapping->host;
933 struct fuse_conn *fc = get_fuse_conn(inode);
934
935 /*
936 * In auto invalidate mode, always update attributes on read.
937 * Otherwise, only update if we attempt to read past EOF (to ensure
938 * i_size is up to date).
939 */
940 if (fc->auto_inval_data ||
941 (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
942 int err;
943 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
944 if (err)
945 return err;
946 }
947
948 return generic_file_read_iter(iocb, to);
949 }
950
951 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
952 loff_t pos, size_t count)
953 {
954 struct fuse_write_in *inarg = &req->misc.write.in;
955 struct fuse_write_out *outarg = &req->misc.write.out;
956
957 inarg->fh = ff->fh;
958 inarg->offset = pos;
959 inarg->size = count;
960 req->in.h.opcode = FUSE_WRITE;
961 req->in.h.nodeid = ff->nodeid;
962 req->in.numargs = 2;
963 if (ff->fc->minor < 9)
964 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
965 else
966 req->in.args[0].size = sizeof(struct fuse_write_in);
967 req->in.args[0].value = inarg;
968 req->in.args[1].size = count;
969 req->out.numargs = 1;
970 req->out.args[0].size = sizeof(struct fuse_write_out);
971 req->out.args[0].value = outarg;
972 }
973
974 static size_t fuse_send_write(struct fuse_req *req, struct fuse_io_priv *io,
975 loff_t pos, size_t count, fl_owner_t owner)
976 {
977 struct file *file = io->file;
978 struct fuse_file *ff = file->private_data;
979 struct fuse_conn *fc = ff->fc;
980 struct fuse_write_in *inarg = &req->misc.write.in;
981
982 fuse_write_fill(req, ff, pos, count);
983 inarg->flags = file->f_flags;
984 if (owner != NULL) {
985 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
986 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
987 }
988
989 if (io->async)
990 return fuse_async_req_send(fc, req, count, io);
991
992 fuse_request_send(fc, req);
993 return req->misc.write.out.size;
994 }
995
996 bool fuse_write_update_size(struct inode *inode, loff_t pos)
997 {
998 struct fuse_conn *fc = get_fuse_conn(inode);
999 struct fuse_inode *fi = get_fuse_inode(inode);
1000 bool ret = false;
1001
1002 spin_lock(&fc->lock);
1003 fi->attr_version = ++fc->attr_version;
1004 if (pos > inode->i_size) {
1005 i_size_write(inode, pos);
1006 ret = true;
1007 }
1008 spin_unlock(&fc->lock);
1009
1010 return ret;
1011 }
1012
1013 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
1014 struct inode *inode, loff_t pos,
1015 size_t count)
1016 {
1017 size_t res;
1018 unsigned offset;
1019 unsigned i;
1020 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
1021
1022 for (i = 0; i < req->num_pages; i++)
1023 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
1024
1025 res = fuse_send_write(req, &io, pos, count, NULL);
1026
1027 offset = req->page_descs[0].offset;
1028 count = res;
1029 for (i = 0; i < req->num_pages; i++) {
1030 struct page *page = req->pages[i];
1031
1032 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
1033 SetPageUptodate(page);
1034
1035 if (count > PAGE_CACHE_SIZE - offset)
1036 count -= PAGE_CACHE_SIZE - offset;
1037 else
1038 count = 0;
1039 offset = 0;
1040
1041 unlock_page(page);
1042 page_cache_release(page);
1043 }
1044
1045 return res;
1046 }
1047
1048 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
1049 struct address_space *mapping,
1050 struct iov_iter *ii, loff_t pos)
1051 {
1052 struct fuse_conn *fc = get_fuse_conn(mapping->host);
1053 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
1054 size_t count = 0;
1055 int err;
1056
1057 req->in.argpages = 1;
1058 req->page_descs[0].offset = offset;
1059
1060 do {
1061 size_t tmp;
1062 struct page *page;
1063 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1064 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
1065 iov_iter_count(ii));
1066
1067 bytes = min_t(size_t, bytes, fc->max_write - count);
1068
1069 again:
1070 err = -EFAULT;
1071 if (iov_iter_fault_in_readable(ii, bytes))
1072 break;
1073
1074 err = -ENOMEM;
1075 page = grab_cache_page_write_begin(mapping, index, 0);
1076 if (!page)
1077 break;
1078
1079 if (mapping_writably_mapped(mapping))
1080 flush_dcache_page(page);
1081
1082 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
1083 flush_dcache_page(page);
1084
1085 iov_iter_advance(ii, tmp);
1086 if (!tmp) {
1087 unlock_page(page);
1088 page_cache_release(page);
1089 bytes = min(bytes, iov_iter_single_seg_count(ii));
1090 goto again;
1091 }
1092
1093 err = 0;
1094 req->pages[req->num_pages] = page;
1095 req->page_descs[req->num_pages].length = tmp;
1096 req->num_pages++;
1097
1098 count += tmp;
1099 pos += tmp;
1100 offset += tmp;
1101 if (offset == PAGE_CACHE_SIZE)
1102 offset = 0;
1103
1104 if (!fc->big_writes)
1105 break;
1106 } while (iov_iter_count(ii) && count < fc->max_write &&
1107 req->num_pages < req->max_pages && offset == 0);
1108
1109 return count > 0 ? count : err;
1110 }
1111
1112 static inline unsigned fuse_wr_pages(loff_t pos, size_t len)
1113 {
1114 return min_t(unsigned,
1115 ((pos + len - 1) >> PAGE_CACHE_SHIFT) -
1116 (pos >> PAGE_CACHE_SHIFT) + 1,
1117 FUSE_MAX_PAGES_PER_REQ);
1118 }
1119
1120 static ssize_t fuse_perform_write(struct file *file,
1121 struct address_space *mapping,
1122 struct iov_iter *ii, loff_t pos)
1123 {
1124 struct inode *inode = mapping->host;
1125 struct fuse_conn *fc = get_fuse_conn(inode);
1126 struct fuse_inode *fi = get_fuse_inode(inode);
1127 int err = 0;
1128 ssize_t res = 0;
1129
1130 if (is_bad_inode(inode))
1131 return -EIO;
1132
1133 if (inode->i_size < pos + iov_iter_count(ii))
1134 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1135
1136 do {
1137 struct fuse_req *req;
1138 ssize_t count;
1139 unsigned nr_pages = fuse_wr_pages(pos, iov_iter_count(ii));
1140
1141 req = fuse_get_req(fc, nr_pages);
1142 if (IS_ERR(req)) {
1143 err = PTR_ERR(req);
1144 break;
1145 }
1146
1147 count = fuse_fill_write_pages(req, mapping, ii, pos);
1148 if (count <= 0) {
1149 err = count;
1150 } else {
1151 size_t num_written;
1152
1153 num_written = fuse_send_write_pages(req, file, inode,
1154 pos, count);
1155 err = req->out.h.error;
1156 if (!err) {
1157 res += num_written;
1158 pos += num_written;
1159
1160 /* break out of the loop on short write */
1161 if (num_written != count)
1162 err = -EIO;
1163 }
1164 }
1165 fuse_put_request(fc, req);
1166 } while (!err && iov_iter_count(ii));
1167
1168 if (res > 0)
1169 fuse_write_update_size(inode, pos);
1170
1171 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1172 fuse_invalidate_attr(inode);
1173
1174 return res > 0 ? res : err;
1175 }
1176
1177 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1178 {
1179 struct file *file = iocb->ki_filp;
1180 struct address_space *mapping = file->f_mapping;
1181 ssize_t written = 0;
1182 ssize_t written_buffered = 0;
1183 struct inode *inode = mapping->host;
1184 ssize_t err;
1185 loff_t endbyte = 0;
1186
1187 if (get_fuse_conn(inode)->writeback_cache) {
1188 /* Update size (EOF optimization) and mode (SUID clearing) */
1189 err = fuse_update_attributes(mapping->host, NULL, file, NULL);
1190 if (err)
1191 return err;
1192
1193 return generic_file_write_iter(iocb, from);
1194 }
1195
1196 mutex_lock(&inode->i_mutex);
1197
1198 /* We can write back this queue in page reclaim */
1199 current->backing_dev_info = inode_to_bdi(inode);
1200
1201 err = generic_write_checks(iocb, from);
1202 if (err <= 0)
1203 goto out;
1204
1205 err = file_remove_privs(file);
1206 if (err)
1207 goto out;
1208
1209 err = file_update_time(file);
1210 if (err)
1211 goto out;
1212
1213 if (iocb->ki_flags & IOCB_DIRECT) {
1214 loff_t pos = iocb->ki_pos;
1215 written = generic_file_direct_write(iocb, from, pos);
1216 if (written < 0 || !iov_iter_count(from))
1217 goto out;
1218
1219 pos += written;
1220
1221 written_buffered = fuse_perform_write(file, mapping, from, pos);
1222 if (written_buffered < 0) {
1223 err = written_buffered;
1224 goto out;
1225 }
1226 endbyte = pos + written_buffered - 1;
1227
1228 err = filemap_write_and_wait_range(file->f_mapping, pos,
1229 endbyte);
1230 if (err)
1231 goto out;
1232
1233 invalidate_mapping_pages(file->f_mapping,
1234 pos >> PAGE_CACHE_SHIFT,
1235 endbyte >> PAGE_CACHE_SHIFT);
1236
1237 written += written_buffered;
1238 iocb->ki_pos = pos + written_buffered;
1239 } else {
1240 written = fuse_perform_write(file, mapping, from, iocb->ki_pos);
1241 if (written >= 0)
1242 iocb->ki_pos += written;
1243 }
1244 out:
1245 current->backing_dev_info = NULL;
1246 mutex_unlock(&inode->i_mutex);
1247
1248 return written ? written : err;
1249 }
1250
1251 static inline void fuse_page_descs_length_init(struct fuse_req *req,
1252 unsigned index, unsigned nr_pages)
1253 {
1254 int i;
1255
1256 for (i = index; i < index + nr_pages; i++)
1257 req->page_descs[i].length = PAGE_SIZE -
1258 req->page_descs[i].offset;
1259 }
1260
1261 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1262 {
1263 return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1264 }
1265
1266 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1267 size_t max_size)
1268 {
1269 return min(iov_iter_single_seg_count(ii), max_size);
1270 }
1271
1272 static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1273 size_t *nbytesp, int write)
1274 {
1275 size_t nbytes = 0; /* # bytes already packed in req */
1276
1277 /* Special case for kernel I/O: can copy directly into the buffer */
1278 if (ii->type & ITER_KVEC) {
1279 unsigned long user_addr = fuse_get_user_addr(ii);
1280 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1281
1282 if (write)
1283 req->in.args[1].value = (void *) user_addr;
1284 else
1285 req->out.args[0].value = (void *) user_addr;
1286
1287 iov_iter_advance(ii, frag_size);
1288 *nbytesp = frag_size;
1289 return 0;
1290 }
1291
1292 while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1293 unsigned npages;
1294 size_t start;
1295 ssize_t ret = iov_iter_get_pages(ii,
1296 &req->pages[req->num_pages],
1297 *nbytesp - nbytes,
1298 req->max_pages - req->num_pages,
1299 &start);
1300 if (ret < 0)
1301 return ret;
1302
1303 iov_iter_advance(ii, ret);
1304 nbytes += ret;
1305
1306 ret += start;
1307 npages = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
1308
1309 req->page_descs[req->num_pages].offset = start;
1310 fuse_page_descs_length_init(req, req->num_pages, npages);
1311
1312 req->num_pages += npages;
1313 req->page_descs[req->num_pages - 1].length -=
1314 (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1315 }
1316
1317 if (write)
1318 req->in.argpages = 1;
1319 else
1320 req->out.argpages = 1;
1321
1322 *nbytesp = nbytes;
1323
1324 return 0;
1325 }
1326
1327 static inline int fuse_iter_npages(const struct iov_iter *ii_p)
1328 {
1329 return iov_iter_npages(ii_p, FUSE_MAX_PAGES_PER_REQ);
1330 }
1331
1332 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1333 loff_t *ppos, int flags)
1334 {
1335 int write = flags & FUSE_DIO_WRITE;
1336 bool should_dirty = !write && iter_is_iovec(iter);
1337 int cuse = flags & FUSE_DIO_CUSE;
1338 struct file *file = io->file;
1339 struct inode *inode = file->f_mapping->host;
1340 struct fuse_file *ff = file->private_data;
1341 struct fuse_conn *fc = ff->fc;
1342 size_t nmax = write ? fc->max_write : fc->max_read;
1343 loff_t pos = *ppos;
1344 size_t count = iov_iter_count(iter);
1345 pgoff_t idx_from = pos >> PAGE_CACHE_SHIFT;
1346 pgoff_t idx_to = (pos + count - 1) >> PAGE_CACHE_SHIFT;
1347 ssize_t res = 0;
1348 struct fuse_req *req;
1349
1350 if (io->async)
1351 req = fuse_get_req_for_background(fc, fuse_iter_npages(iter));
1352 else
1353 req = fuse_get_req(fc, fuse_iter_npages(iter));
1354 if (IS_ERR(req))
1355 return PTR_ERR(req);
1356
1357 if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1358 if (!write)
1359 mutex_lock(&inode->i_mutex);
1360 fuse_sync_writes(inode);
1361 if (!write)
1362 mutex_unlock(&inode->i_mutex);
1363 }
1364
1365 while (count) {
1366 size_t nres;
1367 fl_owner_t owner = current->files;
1368 size_t nbytes = min(count, nmax);
1369 int err = fuse_get_user_pages(req, iter, &nbytes, write);
1370 if (err) {
1371 res = err;
1372 break;
1373 }
1374
1375 if (write)
1376 nres = fuse_send_write(req, io, pos, nbytes, owner);
1377 else
1378 nres = fuse_send_read(req, io, pos, nbytes, owner);
1379
1380 if (!io->async)
1381 fuse_release_user_pages(req, should_dirty);
1382 if (req->out.h.error) {
1383 if (!res)
1384 res = req->out.h.error;
1385 break;
1386 } else if (nres > nbytes) {
1387 res = -EIO;
1388 break;
1389 }
1390 count -= nres;
1391 res += nres;
1392 pos += nres;
1393 if (nres != nbytes)
1394 break;
1395 if (count) {
1396 fuse_put_request(fc, req);
1397 if (io->async)
1398 req = fuse_get_req_for_background(fc,
1399 fuse_iter_npages(iter));
1400 else
1401 req = fuse_get_req(fc, fuse_iter_npages(iter));
1402 if (IS_ERR(req))
1403 break;
1404 }
1405 }
1406 if (!IS_ERR(req))
1407 fuse_put_request(fc, req);
1408 if (res > 0)
1409 *ppos = pos;
1410
1411 return res;
1412 }
1413 EXPORT_SYMBOL_GPL(fuse_direct_io);
1414
1415 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1416 struct iov_iter *iter,
1417 loff_t *ppos)
1418 {
1419 ssize_t res;
1420 struct file *file = io->file;
1421 struct inode *inode = file_inode(file);
1422
1423 if (is_bad_inode(inode))
1424 return -EIO;
1425
1426 res = fuse_direct_io(io, iter, ppos, 0);
1427
1428 fuse_invalidate_attr(inode);
1429
1430 return res;
1431 }
1432
1433 static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1434 {
1435 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb->ki_filp);
1436 return __fuse_direct_read(&io, to, &iocb->ki_pos);
1437 }
1438
1439 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1440 {
1441 struct file *file = iocb->ki_filp;
1442 struct inode *inode = file_inode(file);
1443 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
1444 ssize_t res;
1445
1446 if (is_bad_inode(inode))
1447 return -EIO;
1448
1449 /* Don't allow parallel writes to the same file */
1450 mutex_lock(&inode->i_mutex);
1451 res = generic_write_checks(iocb, from);
1452 if (res > 0)
1453 res = fuse_direct_io(&io, from, &iocb->ki_pos, FUSE_DIO_WRITE);
1454 fuse_invalidate_attr(inode);
1455 if (res > 0)
1456 fuse_write_update_size(inode, iocb->ki_pos);
1457 mutex_unlock(&inode->i_mutex);
1458
1459 return res;
1460 }
1461
1462 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1463 {
1464 int i;
1465
1466 for (i = 0; i < req->num_pages; i++)
1467 __free_page(req->pages[i]);
1468
1469 if (req->ff)
1470 fuse_file_put(req->ff, false);
1471 }
1472
1473 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1474 {
1475 struct inode *inode = req->inode;
1476 struct fuse_inode *fi = get_fuse_inode(inode);
1477 struct backing_dev_info *bdi = inode_to_bdi(inode);
1478 int i;
1479
1480 list_del(&req->writepages_entry);
1481 for (i = 0; i < req->num_pages; i++) {
1482 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1483 dec_zone_page_state(req->pages[i], NR_WRITEBACK_TEMP);
1484 wb_writeout_inc(&bdi->wb);
1485 }
1486 wake_up(&fi->page_waitq);
1487 }
1488
1489 /* Called under fc->lock, may release and reacquire it */
1490 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req,
1491 loff_t size)
1492 __releases(fc->lock)
1493 __acquires(fc->lock)
1494 {
1495 struct fuse_inode *fi = get_fuse_inode(req->inode);
1496 struct fuse_write_in *inarg = &req->misc.write.in;
1497 __u64 data_size = req->num_pages * PAGE_CACHE_SIZE;
1498
1499 if (!fc->connected)
1500 goto out_free;
1501
1502 if (inarg->offset + data_size <= size) {
1503 inarg->size = data_size;
1504 } else if (inarg->offset < size) {
1505 inarg->size = size - inarg->offset;
1506 } else {
1507 /* Got truncated off completely */
1508 goto out_free;
1509 }
1510
1511 req->in.args[1].size = inarg->size;
1512 fi->writectr++;
1513 fuse_request_send_background_locked(fc, req);
1514 return;
1515
1516 out_free:
1517 fuse_writepage_finish(fc, req);
1518 spin_unlock(&fc->lock);
1519 fuse_writepage_free(fc, req);
1520 fuse_put_request(fc, req);
1521 spin_lock(&fc->lock);
1522 }
1523
1524 /*
1525 * If fi->writectr is positive (no truncate or fsync going on) send
1526 * all queued writepage requests.
1527 *
1528 * Called with fc->lock
1529 */
1530 void fuse_flush_writepages(struct inode *inode)
1531 __releases(fc->lock)
1532 __acquires(fc->lock)
1533 {
1534 struct fuse_conn *fc = get_fuse_conn(inode);
1535 struct fuse_inode *fi = get_fuse_inode(inode);
1536 size_t crop = i_size_read(inode);
1537 struct fuse_req *req;
1538
1539 while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1540 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1541 list_del_init(&req->list);
1542 fuse_send_writepage(fc, req, crop);
1543 }
1544 }
1545
1546 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1547 {
1548 struct inode *inode = req->inode;
1549 struct fuse_inode *fi = get_fuse_inode(inode);
1550
1551 mapping_set_error(inode->i_mapping, req->out.h.error);
1552 spin_lock(&fc->lock);
1553 while (req->misc.write.next) {
1554 struct fuse_conn *fc = get_fuse_conn(inode);
1555 struct fuse_write_in *inarg = &req->misc.write.in;
1556 struct fuse_req *next = req->misc.write.next;
1557 req->misc.write.next = next->misc.write.next;
1558 next->misc.write.next = NULL;
1559 next->ff = fuse_file_get(req->ff);
1560 list_add(&next->writepages_entry, &fi->writepages);
1561
1562 /*
1563 * Skip fuse_flush_writepages() to make it easy to crop requests
1564 * based on primary request size.
1565 *
1566 * 1st case (trivial): there are no concurrent activities using
1567 * fuse_set/release_nowrite. Then we're on safe side because
1568 * fuse_flush_writepages() would call fuse_send_writepage()
1569 * anyway.
1570 *
1571 * 2nd case: someone called fuse_set_nowrite and it is waiting
1572 * now for completion of all in-flight requests. This happens
1573 * rarely and no more than once per page, so this should be
1574 * okay.
1575 *
1576 * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1577 * of fuse_set_nowrite..fuse_release_nowrite section. The fact
1578 * that fuse_set_nowrite returned implies that all in-flight
1579 * requests were completed along with all of their secondary
1580 * requests. Further primary requests are blocked by negative
1581 * writectr. Hence there cannot be any in-flight requests and
1582 * no invocations of fuse_writepage_end() while we're in
1583 * fuse_set_nowrite..fuse_release_nowrite section.
1584 */
1585 fuse_send_writepage(fc, next, inarg->offset + inarg->size);
1586 }
1587 fi->writectr--;
1588 fuse_writepage_finish(fc, req);
1589 spin_unlock(&fc->lock);
1590 fuse_writepage_free(fc, req);
1591 }
1592
1593 static struct fuse_file *__fuse_write_file_get(struct fuse_conn *fc,
1594 struct fuse_inode *fi)
1595 {
1596 struct fuse_file *ff = NULL;
1597
1598 spin_lock(&fc->lock);
1599 if (!list_empty(&fi->write_files)) {
1600 ff = list_entry(fi->write_files.next, struct fuse_file,
1601 write_entry);
1602 fuse_file_get(ff);
1603 }
1604 spin_unlock(&fc->lock);
1605
1606 return ff;
1607 }
1608
1609 static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1610 struct fuse_inode *fi)
1611 {
1612 struct fuse_file *ff = __fuse_write_file_get(fc, fi);
1613 WARN_ON(!ff);
1614 return ff;
1615 }
1616
1617 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1618 {
1619 struct fuse_conn *fc = get_fuse_conn(inode);
1620 struct fuse_inode *fi = get_fuse_inode(inode);
1621 struct fuse_file *ff;
1622 int err;
1623
1624 ff = __fuse_write_file_get(fc, fi);
1625 err = fuse_flush_times(inode, ff);
1626 if (ff)
1627 fuse_file_put(ff, 0);
1628
1629 return err;
1630 }
1631
1632 static int fuse_writepage_locked(struct page *page)
1633 {
1634 struct address_space *mapping = page->mapping;
1635 struct inode *inode = mapping->host;
1636 struct fuse_conn *fc = get_fuse_conn(inode);
1637 struct fuse_inode *fi = get_fuse_inode(inode);
1638 struct fuse_req *req;
1639 struct page *tmp_page;
1640 int error = -ENOMEM;
1641
1642 set_page_writeback(page);
1643
1644 req = fuse_request_alloc_nofs(1);
1645 if (!req)
1646 goto err;
1647
1648 /* writeback always goes to bg_queue */
1649 __set_bit(FR_BACKGROUND, &req->flags);
1650 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1651 if (!tmp_page)
1652 goto err_free;
1653
1654 error = -EIO;
1655 req->ff = fuse_write_file_get(fc, fi);
1656 if (!req->ff)
1657 goto err_nofile;
1658
1659 fuse_write_fill(req, req->ff, page_offset(page), 0);
1660
1661 copy_highpage(tmp_page, page);
1662 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1663 req->misc.write.next = NULL;
1664 req->in.argpages = 1;
1665 req->num_pages = 1;
1666 req->pages[0] = tmp_page;
1667 req->page_descs[0].offset = 0;
1668 req->page_descs[0].length = PAGE_SIZE;
1669 req->end = fuse_writepage_end;
1670 req->inode = inode;
1671
1672 inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1673 inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1674
1675 spin_lock(&fc->lock);
1676 list_add(&req->writepages_entry, &fi->writepages);
1677 list_add_tail(&req->list, &fi->queued_writes);
1678 fuse_flush_writepages(inode);
1679 spin_unlock(&fc->lock);
1680
1681 end_page_writeback(page);
1682
1683 return 0;
1684
1685 err_nofile:
1686 __free_page(tmp_page);
1687 err_free:
1688 fuse_request_free(req);
1689 err:
1690 end_page_writeback(page);
1691 return error;
1692 }
1693
1694 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1695 {
1696 int err;
1697
1698 if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1699 /*
1700 * ->writepages() should be called for sync() and friends. We
1701 * should only get here on direct reclaim and then we are
1702 * allowed to skip a page which is already in flight
1703 */
1704 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1705
1706 redirty_page_for_writepage(wbc, page);
1707 return 0;
1708 }
1709
1710 err = fuse_writepage_locked(page);
1711 unlock_page(page);
1712
1713 return err;
1714 }
1715
1716 struct fuse_fill_wb_data {
1717 struct fuse_req *req;
1718 struct fuse_file *ff;
1719 struct inode *inode;
1720 struct page **orig_pages;
1721 };
1722
1723 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1724 {
1725 struct fuse_req *req = data->req;
1726 struct inode *inode = data->inode;
1727 struct fuse_conn *fc = get_fuse_conn(inode);
1728 struct fuse_inode *fi = get_fuse_inode(inode);
1729 int num_pages = req->num_pages;
1730 int i;
1731
1732 req->ff = fuse_file_get(data->ff);
1733 spin_lock(&fc->lock);
1734 list_add_tail(&req->list, &fi->queued_writes);
1735 fuse_flush_writepages(inode);
1736 spin_unlock(&fc->lock);
1737
1738 for (i = 0; i < num_pages; i++)
1739 end_page_writeback(data->orig_pages[i]);
1740 }
1741
1742 static bool fuse_writepage_in_flight(struct fuse_req *new_req,
1743 struct page *page)
1744 {
1745 struct fuse_conn *fc = get_fuse_conn(new_req->inode);
1746 struct fuse_inode *fi = get_fuse_inode(new_req->inode);
1747 struct fuse_req *tmp;
1748 struct fuse_req *old_req;
1749 bool found = false;
1750 pgoff_t curr_index;
1751
1752 BUG_ON(new_req->num_pages != 0);
1753
1754 spin_lock(&fc->lock);
1755 list_del(&new_req->writepages_entry);
1756 list_for_each_entry(old_req, &fi->writepages, writepages_entry) {
1757 BUG_ON(old_req->inode != new_req->inode);
1758 curr_index = old_req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
1759 if (curr_index <= page->index &&
1760 page->index < curr_index + old_req->num_pages) {
1761 found = true;
1762 break;
1763 }
1764 }
1765 if (!found) {
1766 list_add(&new_req->writepages_entry, &fi->writepages);
1767 goto out_unlock;
1768 }
1769
1770 new_req->num_pages = 1;
1771 for (tmp = old_req; tmp != NULL; tmp = tmp->misc.write.next) {
1772 BUG_ON(tmp->inode != new_req->inode);
1773 curr_index = tmp->misc.write.in.offset >> PAGE_CACHE_SHIFT;
1774 if (tmp->num_pages == 1 &&
1775 curr_index == page->index) {
1776 old_req = tmp;
1777 }
1778 }
1779
1780 if (old_req->num_pages == 1 && test_bit(FR_PENDING, &old_req->flags)) {
1781 struct backing_dev_info *bdi = inode_to_bdi(page->mapping->host);
1782
1783 copy_highpage(old_req->pages[0], page);
1784 spin_unlock(&fc->lock);
1785
1786 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1787 dec_zone_page_state(page, NR_WRITEBACK_TEMP);
1788 wb_writeout_inc(&bdi->wb);
1789 fuse_writepage_free(fc, new_req);
1790 fuse_request_free(new_req);
1791 goto out;
1792 } else {
1793 new_req->misc.write.next = old_req->misc.write.next;
1794 old_req->misc.write.next = new_req;
1795 }
1796 out_unlock:
1797 spin_unlock(&fc->lock);
1798 out:
1799 return found;
1800 }
1801
1802 static int fuse_writepages_fill(struct page *page,
1803 struct writeback_control *wbc, void *_data)
1804 {
1805 struct fuse_fill_wb_data *data = _data;
1806 struct fuse_req *req = data->req;
1807 struct inode *inode = data->inode;
1808 struct fuse_conn *fc = get_fuse_conn(inode);
1809 struct page *tmp_page;
1810 bool is_writeback;
1811 int err;
1812
1813 if (!data->ff) {
1814 err = -EIO;
1815 data->ff = fuse_write_file_get(fc, get_fuse_inode(inode));
1816 if (!data->ff)
1817 goto out_unlock;
1818 }
1819
1820 /*
1821 * Being under writeback is unlikely but possible. For example direct
1822 * read to an mmaped fuse file will set the page dirty twice; once when
1823 * the pages are faulted with get_user_pages(), and then after the read
1824 * completed.
1825 */
1826 is_writeback = fuse_page_is_writeback(inode, page->index);
1827
1828 if (req && req->num_pages &&
1829 (is_writeback || req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
1830 (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_write ||
1831 data->orig_pages[req->num_pages - 1]->index + 1 != page->index)) {
1832 fuse_writepages_send(data);
1833 data->req = NULL;
1834 }
1835 err = -ENOMEM;
1836 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1837 if (!tmp_page)
1838 goto out_unlock;
1839
1840 /*
1841 * The page must not be redirtied until the writeout is completed
1842 * (i.e. userspace has sent a reply to the write request). Otherwise
1843 * there could be more than one temporary page instance for each real
1844 * page.
1845 *
1846 * This is ensured by holding the page lock in page_mkwrite() while
1847 * checking fuse_page_is_writeback(). We already hold the page lock
1848 * since clear_page_dirty_for_io() and keep it held until we add the
1849 * request to the fi->writepages list and increment req->num_pages.
1850 * After this fuse_page_is_writeback() will indicate that the page is
1851 * under writeback, so we can release the page lock.
1852 */
1853 if (data->req == NULL) {
1854 struct fuse_inode *fi = get_fuse_inode(inode);
1855
1856 err = -ENOMEM;
1857 req = fuse_request_alloc_nofs(FUSE_MAX_PAGES_PER_REQ);
1858 if (!req) {
1859 __free_page(tmp_page);
1860 goto out_unlock;
1861 }
1862
1863 fuse_write_fill(req, data->ff, page_offset(page), 0);
1864 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1865 req->misc.write.next = NULL;
1866 req->in.argpages = 1;
1867 __set_bit(FR_BACKGROUND, &req->flags);
1868 req->num_pages = 0;
1869 req->end = fuse_writepage_end;
1870 req->inode = inode;
1871
1872 spin_lock(&fc->lock);
1873 list_add(&req->writepages_entry, &fi->writepages);
1874 spin_unlock(&fc->lock);
1875
1876 data->req = req;
1877 }
1878 set_page_writeback(page);
1879
1880 copy_highpage(tmp_page, page);
1881 req->pages[req->num_pages] = tmp_page;
1882 req->page_descs[req->num_pages].offset = 0;
1883 req->page_descs[req->num_pages].length = PAGE_SIZE;
1884
1885 inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1886 inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1887
1888 err = 0;
1889 if (is_writeback && fuse_writepage_in_flight(req, page)) {
1890 end_page_writeback(page);
1891 data->req = NULL;
1892 goto out_unlock;
1893 }
1894 data->orig_pages[req->num_pages] = page;
1895
1896 /*
1897 * Protected by fc->lock against concurrent access by
1898 * fuse_page_is_writeback().
1899 */
1900 spin_lock(&fc->lock);
1901 req->num_pages++;
1902 spin_unlock(&fc->lock);
1903
1904 out_unlock:
1905 unlock_page(page);
1906
1907 return err;
1908 }
1909
1910 static int fuse_writepages(struct address_space *mapping,
1911 struct writeback_control *wbc)
1912 {
1913 struct inode *inode = mapping->host;
1914 struct fuse_fill_wb_data data;
1915 int err;
1916
1917 err = -EIO;
1918 if (is_bad_inode(inode))
1919 goto out;
1920
1921 data.inode = inode;
1922 data.req = NULL;
1923 data.ff = NULL;
1924
1925 err = -ENOMEM;
1926 data.orig_pages = kcalloc(FUSE_MAX_PAGES_PER_REQ,
1927 sizeof(struct page *),
1928 GFP_NOFS);
1929 if (!data.orig_pages)
1930 goto out;
1931
1932 err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
1933 if (data.req) {
1934 /* Ignore errors if we can write at least one page */
1935 BUG_ON(!data.req->num_pages);
1936 fuse_writepages_send(&data);
1937 err = 0;
1938 }
1939 if (data.ff)
1940 fuse_file_put(data.ff, false);
1941
1942 kfree(data.orig_pages);
1943 out:
1944 return err;
1945 }
1946
1947 /*
1948 * It's worthy to make sure that space is reserved on disk for the write,
1949 * but how to implement it without killing performance need more thinking.
1950 */
1951 static int fuse_write_begin(struct file *file, struct address_space *mapping,
1952 loff_t pos, unsigned len, unsigned flags,
1953 struct page **pagep, void **fsdata)
1954 {
1955 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1956 struct fuse_conn *fc = get_fuse_conn(file_inode(file));
1957 struct page *page;
1958 loff_t fsize;
1959 int err = -ENOMEM;
1960
1961 WARN_ON(!fc->writeback_cache);
1962
1963 page = grab_cache_page_write_begin(mapping, index, flags);
1964 if (!page)
1965 goto error;
1966
1967 fuse_wait_on_page_writeback(mapping->host, page->index);
1968
1969 if (PageUptodate(page) || len == PAGE_CACHE_SIZE)
1970 goto success;
1971 /*
1972 * Check if the start this page comes after the end of file, in which
1973 * case the readpage can be optimized away.
1974 */
1975 fsize = i_size_read(mapping->host);
1976 if (fsize <= (pos & PAGE_CACHE_MASK)) {
1977 size_t off = pos & ~PAGE_CACHE_MASK;
1978 if (off)
1979 zero_user_segment(page, 0, off);
1980 goto success;
1981 }
1982 err = fuse_do_readpage(file, page);
1983 if (err)
1984 goto cleanup;
1985 success:
1986 *pagep = page;
1987 return 0;
1988
1989 cleanup:
1990 unlock_page(page);
1991 page_cache_release(page);
1992 error:
1993 return err;
1994 }
1995
1996 static int fuse_write_end(struct file *file, struct address_space *mapping,
1997 loff_t pos, unsigned len, unsigned copied,
1998 struct page *page, void *fsdata)
1999 {
2000 struct inode *inode = page->mapping->host;
2001
2002 /* Haven't copied anything? Skip zeroing, size extending, dirtying. */
2003 if (!copied)
2004 goto unlock;
2005
2006 if (!PageUptodate(page)) {
2007 /* Zero any unwritten bytes at the end of the page */
2008 size_t endoff = (pos + copied) & ~PAGE_CACHE_MASK;
2009 if (endoff)
2010 zero_user_segment(page, endoff, PAGE_CACHE_SIZE);
2011 SetPageUptodate(page);
2012 }
2013
2014 fuse_write_update_size(inode, pos + copied);
2015 set_page_dirty(page);
2016
2017 unlock:
2018 unlock_page(page);
2019 page_cache_release(page);
2020
2021 return copied;
2022 }
2023
2024 static int fuse_launder_page(struct page *page)
2025 {
2026 int err = 0;
2027 if (clear_page_dirty_for_io(page)) {
2028 struct inode *inode = page->mapping->host;
2029 err = fuse_writepage_locked(page);
2030 if (!err)
2031 fuse_wait_on_page_writeback(inode, page->index);
2032 }
2033 return err;
2034 }
2035
2036 /*
2037 * Write back dirty pages now, because there may not be any suitable
2038 * open files later
2039 */
2040 static void fuse_vma_close(struct vm_area_struct *vma)
2041 {
2042 filemap_write_and_wait(vma->vm_file->f_mapping);
2043 }
2044
2045 /*
2046 * Wait for writeback against this page to complete before allowing it
2047 * to be marked dirty again, and hence written back again, possibly
2048 * before the previous writepage completed.
2049 *
2050 * Block here, instead of in ->writepage(), so that the userspace fs
2051 * can only block processes actually operating on the filesystem.
2052 *
2053 * Otherwise unprivileged userspace fs would be able to block
2054 * unrelated:
2055 *
2056 * - page migration
2057 * - sync(2)
2058 * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2059 */
2060 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
2061 {
2062 struct page *page = vmf->page;
2063 struct inode *inode = file_inode(vma->vm_file);
2064
2065 file_update_time(vma->vm_file);
2066 lock_page(page);
2067 if (page->mapping != inode->i_mapping) {
2068 unlock_page(page);
2069 return VM_FAULT_NOPAGE;
2070 }
2071
2072 fuse_wait_on_page_writeback(inode, page->index);
2073 return VM_FAULT_LOCKED;
2074 }
2075
2076 static const struct vm_operations_struct fuse_file_vm_ops = {
2077 .close = fuse_vma_close,
2078 .fault = filemap_fault,
2079 .map_pages = filemap_map_pages,
2080 .page_mkwrite = fuse_page_mkwrite,
2081 };
2082
2083 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2084 {
2085 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2086 fuse_link_write_file(file);
2087
2088 file_accessed(file);
2089 vma->vm_ops = &fuse_file_vm_ops;
2090 return 0;
2091 }
2092
2093 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
2094 {
2095 /* Can't provide the coherency needed for MAP_SHARED */
2096 if (vma->vm_flags & VM_MAYSHARE)
2097 return -ENODEV;
2098
2099 invalidate_inode_pages2(file->f_mapping);
2100
2101 return generic_file_mmap(file, vma);
2102 }
2103
2104 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
2105 struct file_lock *fl)
2106 {
2107 switch (ffl->type) {
2108 case F_UNLCK:
2109 break;
2110
2111 case F_RDLCK:
2112 case F_WRLCK:
2113 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2114 ffl->end < ffl->start)
2115 return -EIO;
2116
2117 fl->fl_start = ffl->start;
2118 fl->fl_end = ffl->end;
2119 fl->fl_pid = ffl->pid;
2120 break;
2121
2122 default:
2123 return -EIO;
2124 }
2125 fl->fl_type = ffl->type;
2126 return 0;
2127 }
2128
2129 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2130 const struct file_lock *fl, int opcode, pid_t pid,
2131 int flock, struct fuse_lk_in *inarg)
2132 {
2133 struct inode *inode = file_inode(file);
2134 struct fuse_conn *fc = get_fuse_conn(inode);
2135 struct fuse_file *ff = file->private_data;
2136
2137 memset(inarg, 0, sizeof(*inarg));
2138 inarg->fh = ff->fh;
2139 inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2140 inarg->lk.start = fl->fl_start;
2141 inarg->lk.end = fl->fl_end;
2142 inarg->lk.type = fl->fl_type;
2143 inarg->lk.pid = pid;
2144 if (flock)
2145 inarg->lk_flags |= FUSE_LK_FLOCK;
2146 args->in.h.opcode = opcode;
2147 args->in.h.nodeid = get_node_id(inode);
2148 args->in.numargs = 1;
2149 args->in.args[0].size = sizeof(*inarg);
2150 args->in.args[0].value = inarg;
2151 }
2152
2153 static int fuse_getlk(struct file *file, struct file_lock *fl)
2154 {
2155 struct inode *inode = file_inode(file);
2156 struct fuse_conn *fc = get_fuse_conn(inode);
2157 FUSE_ARGS(args);
2158 struct fuse_lk_in inarg;
2159 struct fuse_lk_out outarg;
2160 int err;
2161
2162 fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2163 args.out.numargs = 1;
2164 args.out.args[0].size = sizeof(outarg);
2165 args.out.args[0].value = &outarg;
2166 err = fuse_simple_request(fc, &args);
2167 if (!err)
2168 err = convert_fuse_file_lock(&outarg.lk, fl);
2169
2170 return err;
2171 }
2172
2173 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2174 {
2175 struct inode *inode = file_inode(file);
2176 struct fuse_conn *fc = get_fuse_conn(inode);
2177 FUSE_ARGS(args);
2178 struct fuse_lk_in inarg;
2179 int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2180 pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
2181 int err;
2182
2183 if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2184 /* NLM needs asynchronous locks, which we don't support yet */
2185 return -ENOLCK;
2186 }
2187
2188 /* Unlock on close is handled by the flush method */
2189 if (fl->fl_flags & FL_CLOSE)
2190 return 0;
2191
2192 fuse_lk_fill(&args, file, fl, opcode, pid, flock, &inarg);
2193 err = fuse_simple_request(fc, &args);
2194
2195 /* locking is restartable */
2196 if (err == -EINTR)
2197 err = -ERESTARTSYS;
2198
2199 return err;
2200 }
2201
2202 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2203 {
2204 struct inode *inode = file_inode(file);
2205 struct fuse_conn *fc = get_fuse_conn(inode);
2206 int err;
2207
2208 if (cmd == F_CANCELLK) {
2209 err = 0;
2210 } else if (cmd == F_GETLK) {
2211 if (fc->no_lock) {
2212 posix_test_lock(file, fl);
2213 err = 0;
2214 } else
2215 err = fuse_getlk(file, fl);
2216 } else {
2217 if (fc->no_lock)
2218 err = posix_lock_file(file, fl, NULL);
2219 else
2220 err = fuse_setlk(file, fl, 0);
2221 }
2222 return err;
2223 }
2224
2225 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2226 {
2227 struct inode *inode = file_inode(file);
2228 struct fuse_conn *fc = get_fuse_conn(inode);
2229 int err;
2230
2231 if (fc->no_flock) {
2232 err = locks_lock_file_wait(file, fl);
2233 } else {
2234 struct fuse_file *ff = file->private_data;
2235
2236 /* emulate flock with POSIX locks */
2237 ff->flock = true;
2238 err = fuse_setlk(file, fl, 1);
2239 }
2240
2241 return err;
2242 }
2243
2244 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2245 {
2246 struct inode *inode = mapping->host;
2247 struct fuse_conn *fc = get_fuse_conn(inode);
2248 FUSE_ARGS(args);
2249 struct fuse_bmap_in inarg;
2250 struct fuse_bmap_out outarg;
2251 int err;
2252
2253 if (!inode->i_sb->s_bdev || fc->no_bmap)
2254 return 0;
2255
2256 memset(&inarg, 0, sizeof(inarg));
2257 inarg.block = block;
2258 inarg.blocksize = inode->i_sb->s_blocksize;
2259 args.in.h.opcode = FUSE_BMAP;
2260 args.in.h.nodeid = get_node_id(inode);
2261 args.in.numargs = 1;
2262 args.in.args[0].size = sizeof(inarg);
2263 args.in.args[0].value = &inarg;
2264 args.out.numargs = 1;
2265 args.out.args[0].size = sizeof(outarg);
2266 args.out.args[0].value = &outarg;
2267 err = fuse_simple_request(fc, &args);
2268 if (err == -ENOSYS)
2269 fc->no_bmap = 1;
2270
2271 return err ? 0 : outarg.block;
2272 }
2273
2274 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2275 {
2276 loff_t retval;
2277 struct inode *inode = file_inode(file);
2278
2279 /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2280 if (whence == SEEK_CUR || whence == SEEK_SET)
2281 return generic_file_llseek(file, offset, whence);
2282
2283 mutex_lock(&inode->i_mutex);
2284 retval = fuse_update_attributes(inode, NULL, file, NULL);
2285 if (!retval)
2286 retval = generic_file_llseek(file, offset, whence);
2287 mutex_unlock(&inode->i_mutex);
2288
2289 return retval;
2290 }
2291
2292 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
2293 unsigned int nr_segs, size_t bytes, bool to_user)
2294 {
2295 struct iov_iter ii;
2296 int page_idx = 0;
2297
2298 if (!bytes)
2299 return 0;
2300
2301 iov_iter_init(&ii, to_user ? READ : WRITE, iov, nr_segs, bytes);
2302
2303 while (iov_iter_count(&ii)) {
2304 struct page *page = pages[page_idx++];
2305 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
2306 void *kaddr;
2307
2308 kaddr = kmap(page);
2309
2310 while (todo) {
2311 char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
2312 size_t iov_len = ii.iov->iov_len - ii.iov_offset;
2313 size_t copy = min(todo, iov_len);
2314 size_t left;
2315
2316 if (!to_user)
2317 left = copy_from_user(kaddr, uaddr, copy);
2318 else
2319 left = copy_to_user(uaddr, kaddr, copy);
2320
2321 if (unlikely(left))
2322 return -EFAULT;
2323
2324 iov_iter_advance(&ii, copy);
2325 todo -= copy;
2326 kaddr += copy;
2327 }
2328
2329 kunmap(page);
2330 }
2331
2332 return 0;
2333 }
2334
2335 /*
2336 * CUSE servers compiled on 32bit broke on 64bit kernels because the
2337 * ABI was defined to be 'struct iovec' which is different on 32bit
2338 * and 64bit. Fortunately we can determine which structure the server
2339 * used from the size of the reply.
2340 */
2341 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2342 size_t transferred, unsigned count,
2343 bool is_compat)
2344 {
2345 #ifdef CONFIG_COMPAT
2346 if (count * sizeof(struct compat_iovec) == transferred) {
2347 struct compat_iovec *ciov = src;
2348 unsigned i;
2349
2350 /*
2351 * With this interface a 32bit server cannot support
2352 * non-compat (i.e. ones coming from 64bit apps) ioctl
2353 * requests
2354 */
2355 if (!is_compat)
2356 return -EINVAL;
2357
2358 for (i = 0; i < count; i++) {
2359 dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2360 dst[i].iov_len = ciov[i].iov_len;
2361 }
2362 return 0;
2363 }
2364 #endif
2365
2366 if (count * sizeof(struct iovec) != transferred)
2367 return -EIO;
2368
2369 memcpy(dst, src, transferred);
2370 return 0;
2371 }
2372
2373 /* Make sure iov_length() won't overflow */
2374 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
2375 {
2376 size_t n;
2377 u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
2378
2379 for (n = 0; n < count; n++, iov++) {
2380 if (iov->iov_len > (size_t) max)
2381 return -ENOMEM;
2382 max -= iov->iov_len;
2383 }
2384 return 0;
2385 }
2386
2387 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2388 void *src, size_t transferred, unsigned count,
2389 bool is_compat)
2390 {
2391 unsigned i;
2392 struct fuse_ioctl_iovec *fiov = src;
2393
2394 if (fc->minor < 16) {
2395 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2396 count, is_compat);
2397 }
2398
2399 if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2400 return -EIO;
2401
2402 for (i = 0; i < count; i++) {
2403 /* Did the server supply an inappropriate value? */
2404 if (fiov[i].base != (unsigned long) fiov[i].base ||
2405 fiov[i].len != (unsigned long) fiov[i].len)
2406 return -EIO;
2407
2408 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2409 dst[i].iov_len = (size_t) fiov[i].len;
2410
2411 #ifdef CONFIG_COMPAT
2412 if (is_compat &&
2413 (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2414 (compat_size_t) dst[i].iov_len != fiov[i].len))
2415 return -EIO;
2416 #endif
2417 }
2418
2419 return 0;
2420 }
2421
2422
2423 /*
2424 * For ioctls, there is no generic way to determine how much memory
2425 * needs to be read and/or written. Furthermore, ioctls are allowed
2426 * to dereference the passed pointer, so the parameter requires deep
2427 * copying but FUSE has no idea whatsoever about what to copy in or
2428 * out.
2429 *
2430 * This is solved by allowing FUSE server to retry ioctl with
2431 * necessary in/out iovecs. Let's assume the ioctl implementation
2432 * needs to read in the following structure.
2433 *
2434 * struct a {
2435 * char *buf;
2436 * size_t buflen;
2437 * }
2438 *
2439 * On the first callout to FUSE server, inarg->in_size and
2440 * inarg->out_size will be NULL; then, the server completes the ioctl
2441 * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2442 * the actual iov array to
2443 *
2444 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) } }
2445 *
2446 * which tells FUSE to copy in the requested area and retry the ioctl.
2447 * On the second round, the server has access to the structure and
2448 * from that it can tell what to look for next, so on the invocation,
2449 * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2450 *
2451 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) },
2452 * { .iov_base = a.buf, .iov_len = a.buflen } }
2453 *
2454 * FUSE will copy both struct a and the pointed buffer from the
2455 * process doing the ioctl and retry ioctl with both struct a and the
2456 * buffer.
2457 *
2458 * This time, FUSE server has everything it needs and completes ioctl
2459 * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2460 *
2461 * Copying data out works the same way.
2462 *
2463 * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2464 * automatically initializes in and out iovs by decoding @cmd with
2465 * _IOC_* macros and the server is not allowed to request RETRY. This
2466 * limits ioctl data transfers to well-formed ioctls and is the forced
2467 * behavior for all FUSE servers.
2468 */
2469 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2470 unsigned int flags)
2471 {
2472 struct fuse_file *ff = file->private_data;
2473 struct fuse_conn *fc = ff->fc;
2474 struct fuse_ioctl_in inarg = {
2475 .fh = ff->fh,
2476 .cmd = cmd,
2477 .arg = arg,
2478 .flags = flags
2479 };
2480 struct fuse_ioctl_out outarg;
2481 struct fuse_req *req = NULL;
2482 struct page **pages = NULL;
2483 struct iovec *iov_page = NULL;
2484 struct iovec *in_iov = NULL, *out_iov = NULL;
2485 unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
2486 size_t in_size, out_size, transferred;
2487 int err;
2488
2489 #if BITS_PER_LONG == 32
2490 inarg.flags |= FUSE_IOCTL_32BIT;
2491 #else
2492 if (flags & FUSE_IOCTL_COMPAT)
2493 inarg.flags |= FUSE_IOCTL_32BIT;
2494 #endif
2495
2496 /* assume all the iovs returned by client always fits in a page */
2497 BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2498
2499 err = -ENOMEM;
2500 pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
2501 iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2502 if (!pages || !iov_page)
2503 goto out;
2504
2505 /*
2506 * If restricted, initialize IO parameters as encoded in @cmd.
2507 * RETRY from server is not allowed.
2508 */
2509 if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2510 struct iovec *iov = iov_page;
2511
2512 iov->iov_base = (void __user *)arg;
2513 iov->iov_len = _IOC_SIZE(cmd);
2514
2515 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2516 in_iov = iov;
2517 in_iovs = 1;
2518 }
2519
2520 if (_IOC_DIR(cmd) & _IOC_READ) {
2521 out_iov = iov;
2522 out_iovs = 1;
2523 }
2524 }
2525
2526 retry:
2527 inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2528 inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2529
2530 /*
2531 * Out data can be used either for actual out data or iovs,
2532 * make sure there always is at least one page.
2533 */
2534 out_size = max_t(size_t, out_size, PAGE_SIZE);
2535 max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2536
2537 /* make sure there are enough buffer pages and init request with them */
2538 err = -ENOMEM;
2539 if (max_pages > FUSE_MAX_PAGES_PER_REQ)
2540 goto out;
2541 while (num_pages < max_pages) {
2542 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2543 if (!pages[num_pages])
2544 goto out;
2545 num_pages++;
2546 }
2547
2548 req = fuse_get_req(fc, num_pages);
2549 if (IS_ERR(req)) {
2550 err = PTR_ERR(req);
2551 req = NULL;
2552 goto out;
2553 }
2554 memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
2555 req->num_pages = num_pages;
2556 fuse_page_descs_length_init(req, 0, req->num_pages);
2557
2558 /* okay, let's send it to the client */
2559 req->in.h.opcode = FUSE_IOCTL;
2560 req->in.h.nodeid = ff->nodeid;
2561 req->in.numargs = 1;
2562 req->in.args[0].size = sizeof(inarg);
2563 req->in.args[0].value = &inarg;
2564 if (in_size) {
2565 req->in.numargs++;
2566 req->in.args[1].size = in_size;
2567 req->in.argpages = 1;
2568
2569 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
2570 false);
2571 if (err)
2572 goto out;
2573 }
2574
2575 req->out.numargs = 2;
2576 req->out.args[0].size = sizeof(outarg);
2577 req->out.args[0].value = &outarg;
2578 req->out.args[1].size = out_size;
2579 req->out.argpages = 1;
2580 req->out.argvar = 1;
2581
2582 fuse_request_send(fc, req);
2583 err = req->out.h.error;
2584 transferred = req->out.args[1].size;
2585 fuse_put_request(fc, req);
2586 req = NULL;
2587 if (err)
2588 goto out;
2589
2590 /* did it ask for retry? */
2591 if (outarg.flags & FUSE_IOCTL_RETRY) {
2592 void *vaddr;
2593
2594 /* no retry if in restricted mode */
2595 err = -EIO;
2596 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2597 goto out;
2598
2599 in_iovs = outarg.in_iovs;
2600 out_iovs = outarg.out_iovs;
2601
2602 /*
2603 * Make sure things are in boundary, separate checks
2604 * are to protect against overflow.
2605 */
2606 err = -ENOMEM;
2607 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2608 out_iovs > FUSE_IOCTL_MAX_IOV ||
2609 in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2610 goto out;
2611
2612 vaddr = kmap_atomic(pages[0]);
2613 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2614 transferred, in_iovs + out_iovs,
2615 (flags & FUSE_IOCTL_COMPAT) != 0);
2616 kunmap_atomic(vaddr);
2617 if (err)
2618 goto out;
2619
2620 in_iov = iov_page;
2621 out_iov = in_iov + in_iovs;
2622
2623 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
2624 if (err)
2625 goto out;
2626
2627 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
2628 if (err)
2629 goto out;
2630
2631 goto retry;
2632 }
2633
2634 err = -EIO;
2635 if (transferred > inarg.out_size)
2636 goto out;
2637
2638 err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
2639 out:
2640 if (req)
2641 fuse_put_request(fc, req);
2642 free_page((unsigned long) iov_page);
2643 while (num_pages)
2644 __free_page(pages[--num_pages]);
2645 kfree(pages);
2646
2647 return err ? err : outarg.result;
2648 }
2649 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2650
2651 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2652 unsigned long arg, unsigned int flags)
2653 {
2654 struct inode *inode = file_inode(file);
2655 struct fuse_conn *fc = get_fuse_conn(inode);
2656
2657 if (!fuse_allow_current_process(fc))
2658 return -EACCES;
2659
2660 if (is_bad_inode(inode))
2661 return -EIO;
2662
2663 return fuse_do_ioctl(file, cmd, arg, flags);
2664 }
2665
2666 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2667 unsigned long arg)
2668 {
2669 return fuse_ioctl_common(file, cmd, arg, 0);
2670 }
2671
2672 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2673 unsigned long arg)
2674 {
2675 return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2676 }
2677
2678 /*
2679 * All files which have been polled are linked to RB tree
2680 * fuse_conn->polled_files which is indexed by kh. Walk the tree and
2681 * find the matching one.
2682 */
2683 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2684 struct rb_node **parent_out)
2685 {
2686 struct rb_node **link = &fc->polled_files.rb_node;
2687 struct rb_node *last = NULL;
2688
2689 while (*link) {
2690 struct fuse_file *ff;
2691
2692 last = *link;
2693 ff = rb_entry(last, struct fuse_file, polled_node);
2694
2695 if (kh < ff->kh)
2696 link = &last->rb_left;
2697 else if (kh > ff->kh)
2698 link = &last->rb_right;
2699 else
2700 return link;
2701 }
2702
2703 if (parent_out)
2704 *parent_out = last;
2705 return link;
2706 }
2707
2708 /*
2709 * The file is about to be polled. Make sure it's on the polled_files
2710 * RB tree. Note that files once added to the polled_files tree are
2711 * not removed before the file is released. This is because a file
2712 * polled once is likely to be polled again.
2713 */
2714 static void fuse_register_polled_file(struct fuse_conn *fc,
2715 struct fuse_file *ff)
2716 {
2717 spin_lock(&fc->lock);
2718 if (RB_EMPTY_NODE(&ff->polled_node)) {
2719 struct rb_node **link, *uninitialized_var(parent);
2720
2721 link = fuse_find_polled_node(fc, ff->kh, &parent);
2722 BUG_ON(*link);
2723 rb_link_node(&ff->polled_node, parent, link);
2724 rb_insert_color(&ff->polled_node, &fc->polled_files);
2725 }
2726 spin_unlock(&fc->lock);
2727 }
2728
2729 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2730 {
2731 struct fuse_file *ff = file->private_data;
2732 struct fuse_conn *fc = ff->fc;
2733 struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2734 struct fuse_poll_out outarg;
2735 FUSE_ARGS(args);
2736 int err;
2737
2738 if (fc->no_poll)
2739 return DEFAULT_POLLMASK;
2740
2741 poll_wait(file, &ff->poll_wait, wait);
2742 inarg.events = (__u32)poll_requested_events(wait);
2743
2744 /*
2745 * Ask for notification iff there's someone waiting for it.
2746 * The client may ignore the flag and always notify.
2747 */
2748 if (waitqueue_active(&ff->poll_wait)) {
2749 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2750 fuse_register_polled_file(fc, ff);
2751 }
2752
2753 args.in.h.opcode = FUSE_POLL;
2754 args.in.h.nodeid = ff->nodeid;
2755 args.in.numargs = 1;
2756 args.in.args[0].size = sizeof(inarg);
2757 args.in.args[0].value = &inarg;
2758 args.out.numargs = 1;
2759 args.out.args[0].size = sizeof(outarg);
2760 args.out.args[0].value = &outarg;
2761 err = fuse_simple_request(fc, &args);
2762
2763 if (!err)
2764 return outarg.revents;
2765 if (err == -ENOSYS) {
2766 fc->no_poll = 1;
2767 return DEFAULT_POLLMASK;
2768 }
2769 return POLLERR;
2770 }
2771 EXPORT_SYMBOL_GPL(fuse_file_poll);
2772
2773 /*
2774 * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2775 * wakes up the poll waiters.
2776 */
2777 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2778 struct fuse_notify_poll_wakeup_out *outarg)
2779 {
2780 u64 kh = outarg->kh;
2781 struct rb_node **link;
2782
2783 spin_lock(&fc->lock);
2784
2785 link = fuse_find_polled_node(fc, kh, NULL);
2786 if (*link) {
2787 struct fuse_file *ff;
2788
2789 ff = rb_entry(*link, struct fuse_file, polled_node);
2790 wake_up_interruptible_sync(&ff->poll_wait);
2791 }
2792
2793 spin_unlock(&fc->lock);
2794 return 0;
2795 }
2796
2797 static void fuse_do_truncate(struct file *file)
2798 {
2799 struct inode *inode = file->f_mapping->host;
2800 struct iattr attr;
2801
2802 attr.ia_valid = ATTR_SIZE;
2803 attr.ia_size = i_size_read(inode);
2804
2805 attr.ia_file = file;
2806 attr.ia_valid |= ATTR_FILE;
2807
2808 fuse_do_setattr(inode, &attr, file);
2809 }
2810
2811 static inline loff_t fuse_round_up(loff_t off)
2812 {
2813 return round_up(off, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
2814 }
2815
2816 static ssize_t
2817 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter, loff_t offset)
2818 {
2819 DECLARE_COMPLETION_ONSTACK(wait);
2820 ssize_t ret = 0;
2821 struct file *file = iocb->ki_filp;
2822 struct fuse_file *ff = file->private_data;
2823 bool async_dio = ff->fc->async_dio;
2824 loff_t pos = 0;
2825 struct inode *inode;
2826 loff_t i_size;
2827 size_t count = iov_iter_count(iter);
2828 struct fuse_io_priv *io;
2829 bool is_sync = is_sync_kiocb(iocb);
2830
2831 pos = offset;
2832 inode = file->f_mapping->host;
2833 i_size = i_size_read(inode);
2834
2835 if ((iov_iter_rw(iter) == READ) && (offset > i_size))
2836 return 0;
2837
2838 /* optimization for short read */
2839 if (async_dio && iov_iter_rw(iter) != WRITE && offset + count > i_size) {
2840 if (offset >= i_size)
2841 return 0;
2842 iov_iter_truncate(iter, fuse_round_up(i_size - offset));
2843 count = iov_iter_count(iter);
2844 }
2845
2846 io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2847 if (!io)
2848 return -ENOMEM;
2849 spin_lock_init(&io->lock);
2850 kref_init(&io->refcnt);
2851 io->reqs = 1;
2852 io->bytes = -1;
2853 io->size = 0;
2854 io->offset = offset;
2855 io->write = (iov_iter_rw(iter) == WRITE);
2856 io->err = 0;
2857 io->file = file;
2858 /*
2859 * By default, we want to optimize all I/Os with async request
2860 * submission to the client filesystem if supported.
2861 */
2862 io->async = async_dio;
2863 io->iocb = iocb;
2864
2865 /*
2866 * We cannot asynchronously extend the size of a file. We have no method
2867 * to wait on real async I/O requests, so we must submit this request
2868 * synchronously.
2869 */
2870 if (!is_sync && (offset + count > i_size) &&
2871 iov_iter_rw(iter) == WRITE)
2872 io->async = false;
2873
2874 if (io->async && is_sync) {
2875 /*
2876 * Additional reference to keep io around after
2877 * calling fuse_aio_complete()
2878 */
2879 kref_get(&io->refcnt);
2880 io->done = &wait;
2881 }
2882
2883 if (iov_iter_rw(iter) == WRITE) {
2884 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
2885 fuse_invalidate_attr(inode);
2886 } else {
2887 ret = __fuse_direct_read(io, iter, &pos);
2888 }
2889
2890 if (io->async) {
2891 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2892
2893 /* we have a non-extending, async request, so return */
2894 if (!is_sync)
2895 return -EIOCBQUEUED;
2896
2897 wait_for_completion(&wait);
2898 ret = fuse_get_res_by_io(io);
2899 }
2900
2901 kref_put(&io->refcnt, fuse_io_release);
2902
2903 if (iov_iter_rw(iter) == WRITE) {
2904 if (ret > 0)
2905 fuse_write_update_size(inode, pos);
2906 else if (ret < 0 && offset + count > i_size)
2907 fuse_do_truncate(file);
2908 }
2909
2910 return ret;
2911 }
2912
2913 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2914 loff_t length)
2915 {
2916 struct fuse_file *ff = file->private_data;
2917 struct inode *inode = file_inode(file);
2918 struct fuse_inode *fi = get_fuse_inode(inode);
2919 struct fuse_conn *fc = ff->fc;
2920 FUSE_ARGS(args);
2921 struct fuse_fallocate_in inarg = {
2922 .fh = ff->fh,
2923 .offset = offset,
2924 .length = length,
2925 .mode = mode
2926 };
2927 int err;
2928 bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
2929 (mode & FALLOC_FL_PUNCH_HOLE);
2930
2931 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2932 return -EOPNOTSUPP;
2933
2934 if (fc->no_fallocate)
2935 return -EOPNOTSUPP;
2936
2937 if (lock_inode) {
2938 mutex_lock(&inode->i_mutex);
2939 if (mode & FALLOC_FL_PUNCH_HOLE) {
2940 loff_t endbyte = offset + length - 1;
2941 err = filemap_write_and_wait_range(inode->i_mapping,
2942 offset, endbyte);
2943 if (err)
2944 goto out;
2945
2946 fuse_sync_writes(inode);
2947 }
2948 }
2949
2950 if (!(mode & FALLOC_FL_KEEP_SIZE))
2951 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2952
2953 args.in.h.opcode = FUSE_FALLOCATE;
2954 args.in.h.nodeid = ff->nodeid;
2955 args.in.numargs = 1;
2956 args.in.args[0].size = sizeof(inarg);
2957 args.in.args[0].value = &inarg;
2958 err = fuse_simple_request(fc, &args);
2959 if (err == -ENOSYS) {
2960 fc->no_fallocate = 1;
2961 err = -EOPNOTSUPP;
2962 }
2963 if (err)
2964 goto out;
2965
2966 /* we could have extended the file */
2967 if (!(mode & FALLOC_FL_KEEP_SIZE)) {
2968 bool changed = fuse_write_update_size(inode, offset + length);
2969
2970 if (changed && fc->writeback_cache)
2971 file_update_time(file);
2972 }
2973
2974 if (mode & FALLOC_FL_PUNCH_HOLE)
2975 truncate_pagecache_range(inode, offset, offset + length - 1);
2976
2977 fuse_invalidate_attr(inode);
2978
2979 out:
2980 if (!(mode & FALLOC_FL_KEEP_SIZE))
2981 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2982
2983 if (lock_inode)
2984 mutex_unlock(&inode->i_mutex);
2985
2986 return err;
2987 }
2988
2989 static const struct file_operations fuse_file_operations = {
2990 .llseek = fuse_file_llseek,
2991 .read_iter = fuse_file_read_iter,
2992 .write_iter = fuse_file_write_iter,
2993 .mmap = fuse_file_mmap,
2994 .open = fuse_open,
2995 .flush = fuse_flush,
2996 .release = fuse_release,
2997 .fsync = fuse_fsync,
2998 .lock = fuse_file_lock,
2999 .flock = fuse_file_flock,
3000 .splice_read = generic_file_splice_read,
3001 .unlocked_ioctl = fuse_file_ioctl,
3002 .compat_ioctl = fuse_file_compat_ioctl,
3003 .poll = fuse_file_poll,
3004 .fallocate = fuse_file_fallocate,
3005 };
3006
3007 static const struct file_operations fuse_direct_io_file_operations = {
3008 .llseek = fuse_file_llseek,
3009 .read_iter = fuse_direct_read_iter,
3010 .write_iter = fuse_direct_write_iter,
3011 .mmap = fuse_direct_mmap,
3012 .open = fuse_open,
3013 .flush = fuse_flush,
3014 .release = fuse_release,
3015 .fsync = fuse_fsync,
3016 .lock = fuse_file_lock,
3017 .flock = fuse_file_flock,
3018 .unlocked_ioctl = fuse_file_ioctl,
3019 .compat_ioctl = fuse_file_compat_ioctl,
3020 .poll = fuse_file_poll,
3021 .fallocate = fuse_file_fallocate,
3022 /* no splice_read */
3023 };
3024
3025 static const struct address_space_operations fuse_file_aops = {
3026 .readpage = fuse_readpage,
3027 .writepage = fuse_writepage,
3028 .writepages = fuse_writepages,
3029 .launder_page = fuse_launder_page,
3030 .readpages = fuse_readpages,
3031 .set_page_dirty = __set_page_dirty_nobuffers,
3032 .bmap = fuse_bmap,
3033 .direct_IO = fuse_direct_IO,
3034 .write_begin = fuse_write_begin,
3035 .write_end = fuse_write_end,
3036 };
3037
3038 void fuse_init_file_inode(struct inode *inode)
3039 {
3040 inode->i_fop = &fuse_file_operations;
3041 inode->i_data.a_ops = &fuse_file_aops;
3042 }
Linux with added FPC-III support