]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blob - fs/fuse/file.c
Merge remote-tracking branches 'asoc/topic/inntel', 'asoc/topic/input', 'asoc/topic...
[mirror_ubuntu-bionic-kernel.git] / fs / fuse / file.c
1 /*
2 FUSE: Filesystem in Userspace
3 Copyright (C) 2001-2008 Miklos Szeredi <miklos@szeredi.hu>
4
5 This program can be distributed under the terms of the GNU GPL.
6 See the file COPYING.
7 */
8
9 #include "fuse_i.h"
10
11 #include <linux/pagemap.h>
12 #include <linux/slab.h>
13 #include <linux/kernel.h>
14 #include <linux/sched.h>
15 #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 = kmalloc(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 __clear_bit(FR_BACKGROUND, &req->flags);
104 fuse_request_send(ff->fc, req);
105 iput(req->misc.release.inode);
106 fuse_put_request(ff->fc, req);
107 } else {
108 req->end = fuse_release_end;
109 __set_bit(FR_BACKGROUND, &req->flags);
110 fuse_request_send_background(ff->fc, req);
111 }
112 kfree(ff);
113 }
114 }
115
116 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
117 bool isdir)
118 {
119 struct fuse_file *ff;
120 int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
121
122 ff = fuse_file_alloc(fc);
123 if (!ff)
124 return -ENOMEM;
125
126 ff->fh = 0;
127 ff->open_flags = FOPEN_KEEP_CACHE; /* Default for no-open */
128 if (!fc->no_open || isdir) {
129 struct fuse_open_out outarg;
130 int err;
131
132 err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
133 if (!err) {
134 ff->fh = outarg.fh;
135 ff->open_flags = outarg.open_flags;
136
137 } else if (err != -ENOSYS || isdir) {
138 fuse_file_free(ff);
139 return err;
140 } else {
141 fc->no_open = 1;
142 }
143 }
144
145 if (isdir)
146 ff->open_flags &= ~FOPEN_DIRECT_IO;
147
148 ff->nodeid = nodeid;
149 file->private_data = fuse_file_get(ff);
150
151 return 0;
152 }
153 EXPORT_SYMBOL_GPL(fuse_do_open);
154
155 static void fuse_link_write_file(struct file *file)
156 {
157 struct inode *inode = file_inode(file);
158 struct fuse_conn *fc = get_fuse_conn(inode);
159 struct fuse_inode *fi = get_fuse_inode(inode);
160 struct fuse_file *ff = file->private_data;
161 /*
162 * file may be written through mmap, so chain it onto the
163 * inodes's write_file list
164 */
165 spin_lock(&fc->lock);
166 if (list_empty(&ff->write_entry))
167 list_add(&ff->write_entry, &fi->write_files);
168 spin_unlock(&fc->lock);
169 }
170
171 void fuse_finish_open(struct inode *inode, struct file *file)
172 {
173 struct fuse_file *ff = file->private_data;
174 struct fuse_conn *fc = get_fuse_conn(inode);
175
176 if (ff->open_flags & FOPEN_DIRECT_IO)
177 file->f_op = &fuse_direct_io_file_operations;
178 if (!(ff->open_flags & FOPEN_KEEP_CACHE))
179 invalidate_inode_pages2(inode->i_mapping);
180 if (ff->open_flags & FOPEN_NONSEEKABLE)
181 nonseekable_open(inode, file);
182 if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
183 struct fuse_inode *fi = get_fuse_inode(inode);
184
185 spin_lock(&fc->lock);
186 fi->attr_version = ++fc->attr_version;
187 i_size_write(inode, 0);
188 spin_unlock(&fc->lock);
189 fuse_invalidate_attr(inode);
190 if (fc->writeback_cache)
191 file_update_time(file);
192 }
193 if ((file->f_mode & FMODE_WRITE) && fc->writeback_cache)
194 fuse_link_write_file(file);
195 }
196
197 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
198 {
199 struct fuse_conn *fc = get_fuse_conn(inode);
200 int err;
201 bool lock_inode = (file->f_flags & O_TRUNC) &&
202 fc->atomic_o_trunc &&
203 fc->writeback_cache;
204
205 err = generic_file_open(inode, file);
206 if (err)
207 return err;
208
209 if (lock_inode)
210 inode_lock(inode);
211
212 err = fuse_do_open(fc, get_node_id(inode), file, isdir);
213
214 if (!err)
215 fuse_finish_open(inode, file);
216
217 if (lock_inode)
218 inode_unlock(inode);
219
220 return err;
221 }
222
223 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
224 {
225 struct fuse_conn *fc = ff->fc;
226 struct fuse_req *req = ff->reserved_req;
227 struct fuse_release_in *inarg = &req->misc.release.in;
228
229 spin_lock(&fc->lock);
230 list_del(&ff->write_entry);
231 if (!RB_EMPTY_NODE(&ff->polled_node))
232 rb_erase(&ff->polled_node, &fc->polled_files);
233 spin_unlock(&fc->lock);
234
235 wake_up_interruptible_all(&ff->poll_wait);
236
237 inarg->fh = ff->fh;
238 inarg->flags = flags;
239 req->in.h.opcode = opcode;
240 req->in.h.nodeid = ff->nodeid;
241 req->in.numargs = 1;
242 req->in.args[0].size = sizeof(struct fuse_release_in);
243 req->in.args[0].value = inarg;
244 }
245
246 void fuse_release_common(struct file *file, int opcode)
247 {
248 struct fuse_file *ff;
249 struct fuse_req *req;
250
251 ff = file->private_data;
252 if (unlikely(!ff))
253 return;
254
255 req = ff->reserved_req;
256 fuse_prepare_release(ff, file->f_flags, opcode);
257
258 if (ff->flock) {
259 struct fuse_release_in *inarg = &req->misc.release.in;
260 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
261 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
262 (fl_owner_t) file);
263 }
264 /* Hold inode until release is finished */
265 req->misc.release.inode = igrab(file_inode(file));
266
267 /*
268 * Normally this will send the RELEASE request, however if
269 * some asynchronous READ or WRITE requests are outstanding,
270 * the sending will be delayed.
271 *
272 * Make the release synchronous if this is a fuseblk mount,
273 * synchronous RELEASE is allowed (and desirable) in this case
274 * because the server can be trusted not to screw up.
275 */
276 fuse_file_put(ff, ff->fc->destroy_req != NULL);
277 }
278
279 static int fuse_open(struct inode *inode, struct file *file)
280 {
281 return fuse_open_common(inode, file, false);
282 }
283
284 static int fuse_release(struct inode *inode, struct file *file)
285 {
286 struct fuse_conn *fc = get_fuse_conn(inode);
287
288 /* see fuse_vma_close() for !writeback_cache case */
289 if (fc->writeback_cache)
290 write_inode_now(inode, 1);
291
292 fuse_release_common(file, FUSE_RELEASE);
293
294 /* return value is ignored by VFS */
295 return 0;
296 }
297
298 void fuse_sync_release(struct fuse_file *ff, int flags)
299 {
300 WARN_ON(atomic_read(&ff->count) > 1);
301 fuse_prepare_release(ff, flags, FUSE_RELEASE);
302 __set_bit(FR_FORCE, &ff->reserved_req->flags);
303 __clear_bit(FR_BACKGROUND, &ff->reserved_req->flags);
304 fuse_request_send(ff->fc, ff->reserved_req);
305 fuse_put_request(ff->fc, ff->reserved_req);
306 kfree(ff);
307 }
308 EXPORT_SYMBOL_GPL(fuse_sync_release);
309
310 /*
311 * Scramble the ID space with XTEA, so that the value of the files_struct
312 * pointer is not exposed to userspace.
313 */
314 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
315 {
316 u32 *k = fc->scramble_key;
317 u64 v = (unsigned long) id;
318 u32 v0 = v;
319 u32 v1 = v >> 32;
320 u32 sum = 0;
321 int i;
322
323 for (i = 0; i < 32; i++) {
324 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
325 sum += 0x9E3779B9;
326 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
327 }
328
329 return (u64) v0 + ((u64) v1 << 32);
330 }
331
332 /*
333 * Check if any page in a range is under writeback
334 *
335 * This is currently done by walking the list of writepage requests
336 * for the inode, which can be pretty inefficient.
337 */
338 static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
339 pgoff_t idx_to)
340 {
341 struct fuse_conn *fc = get_fuse_conn(inode);
342 struct fuse_inode *fi = get_fuse_inode(inode);
343 struct fuse_req *req;
344 bool found = false;
345
346 spin_lock(&fc->lock);
347 list_for_each_entry(req, &fi->writepages, writepages_entry) {
348 pgoff_t curr_index;
349
350 BUG_ON(req->inode != inode);
351 curr_index = req->misc.write.in.offset >> PAGE_SHIFT;
352 if (idx_from < curr_index + req->num_pages &&
353 curr_index <= idx_to) {
354 found = true;
355 break;
356 }
357 }
358 spin_unlock(&fc->lock);
359
360 return found;
361 }
362
363 static inline bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
364 {
365 return fuse_range_is_writeback(inode, index, index);
366 }
367
368 /*
369 * Wait for page writeback to be completed.
370 *
371 * Since fuse doesn't rely on the VM writeback tracking, this has to
372 * use some other means.
373 */
374 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
375 {
376 struct fuse_inode *fi = get_fuse_inode(inode);
377
378 wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
379 return 0;
380 }
381
382 /*
383 * Wait for all pending writepages on the inode to finish.
384 *
385 * This is currently done by blocking further writes with FUSE_NOWRITE
386 * and waiting for all sent writes to complete.
387 *
388 * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
389 * could conflict with truncation.
390 */
391 static void fuse_sync_writes(struct inode *inode)
392 {
393 fuse_set_nowrite(inode);
394 fuse_release_nowrite(inode);
395 }
396
397 static int fuse_flush(struct file *file, fl_owner_t id)
398 {
399 struct inode *inode = file_inode(file);
400 struct fuse_conn *fc = get_fuse_conn(inode);
401 struct fuse_file *ff = file->private_data;
402 struct fuse_req *req;
403 struct fuse_flush_in inarg;
404 int err;
405
406 if (is_bad_inode(inode))
407 return -EIO;
408
409 if (fc->no_flush)
410 return 0;
411
412 err = write_inode_now(inode, 1);
413 if (err)
414 return err;
415
416 inode_lock(inode);
417 fuse_sync_writes(inode);
418 inode_unlock(inode);
419
420 err = filemap_check_errors(file->f_mapping);
421 if (err)
422 return err;
423
424 req = fuse_get_req_nofail_nopages(fc, file);
425 memset(&inarg, 0, sizeof(inarg));
426 inarg.fh = ff->fh;
427 inarg.lock_owner = fuse_lock_owner_id(fc, id);
428 req->in.h.opcode = FUSE_FLUSH;
429 req->in.h.nodeid = get_node_id(inode);
430 req->in.numargs = 1;
431 req->in.args[0].size = sizeof(inarg);
432 req->in.args[0].value = &inarg;
433 __set_bit(FR_FORCE, &req->flags);
434 fuse_request_send(fc, req);
435 err = req->out.h.error;
436 fuse_put_request(fc, req);
437 if (err == -ENOSYS) {
438 fc->no_flush = 1;
439 err = 0;
440 }
441 return err;
442 }
443
444 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
445 int datasync, int isdir)
446 {
447 struct inode *inode = file->f_mapping->host;
448 struct fuse_conn *fc = get_fuse_conn(inode);
449 struct fuse_file *ff = file->private_data;
450 FUSE_ARGS(args);
451 struct fuse_fsync_in inarg;
452 int err;
453
454 if (is_bad_inode(inode))
455 return -EIO;
456
457 inode_lock(inode);
458
459 /*
460 * Start writeback against all dirty pages of the inode, then
461 * wait for all outstanding writes, before sending the FSYNC
462 * request.
463 */
464 err = filemap_write_and_wait_range(inode->i_mapping, start, end);
465 if (err)
466 goto out;
467
468 fuse_sync_writes(inode);
469
470 /*
471 * Due to implementation of fuse writeback
472 * filemap_write_and_wait_range() does not catch errors.
473 * We have to do this directly after fuse_sync_writes()
474 */
475 err = filemap_check_errors(file->f_mapping);
476 if (err)
477 goto out;
478
479 err = sync_inode_metadata(inode, 1);
480 if (err)
481 goto out;
482
483 if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
484 goto out;
485
486 memset(&inarg, 0, sizeof(inarg));
487 inarg.fh = ff->fh;
488 inarg.fsync_flags = datasync ? 1 : 0;
489 args.in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
490 args.in.h.nodeid = get_node_id(inode);
491 args.in.numargs = 1;
492 args.in.args[0].size = sizeof(inarg);
493 args.in.args[0].value = &inarg;
494 err = fuse_simple_request(fc, &args);
495 if (err == -ENOSYS) {
496 if (isdir)
497 fc->no_fsyncdir = 1;
498 else
499 fc->no_fsync = 1;
500 err = 0;
501 }
502 out:
503 inode_unlock(inode);
504 return err;
505 }
506
507 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
508 int datasync)
509 {
510 return fuse_fsync_common(file, start, end, datasync, 0);
511 }
512
513 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
514 size_t count, int opcode)
515 {
516 struct fuse_read_in *inarg = &req->misc.read.in;
517 struct fuse_file *ff = file->private_data;
518
519 inarg->fh = ff->fh;
520 inarg->offset = pos;
521 inarg->size = count;
522 inarg->flags = file->f_flags;
523 req->in.h.opcode = opcode;
524 req->in.h.nodeid = ff->nodeid;
525 req->in.numargs = 1;
526 req->in.args[0].size = sizeof(struct fuse_read_in);
527 req->in.args[0].value = inarg;
528 req->out.argvar = 1;
529 req->out.numargs = 1;
530 req->out.args[0].size = count;
531 }
532
533 static void fuse_release_user_pages(struct fuse_req *req, bool should_dirty)
534 {
535 unsigned i;
536
537 for (i = 0; i < req->num_pages; i++) {
538 struct page *page = req->pages[i];
539 if (should_dirty)
540 set_page_dirty_lock(page);
541 put_page(page);
542 }
543 }
544
545 static void fuse_io_release(struct kref *kref)
546 {
547 kfree(container_of(kref, struct fuse_io_priv, refcnt));
548 }
549
550 static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
551 {
552 if (io->err)
553 return io->err;
554
555 if (io->bytes >= 0 && io->write)
556 return -EIO;
557
558 return io->bytes < 0 ? io->size : io->bytes;
559 }
560
561 /**
562 * In case of short read, the caller sets 'pos' to the position of
563 * actual end of fuse request in IO request. Otherwise, if bytes_requested
564 * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
565 *
566 * An example:
567 * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
568 * both submitted asynchronously. The first of them was ACKed by userspace as
569 * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
570 * second request was ACKed as short, e.g. only 1K was read, resulting in
571 * pos == 33K.
572 *
573 * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
574 * will be equal to the length of the longest contiguous fragment of
575 * transferred data starting from the beginning of IO request.
576 */
577 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
578 {
579 int left;
580
581 spin_lock(&io->lock);
582 if (err)
583 io->err = io->err ? : err;
584 else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
585 io->bytes = pos;
586
587 left = --io->reqs;
588 if (!left && io->blocking)
589 complete(io->done);
590 spin_unlock(&io->lock);
591
592 if (!left && !io->blocking) {
593 ssize_t res = fuse_get_res_by_io(io);
594
595 if (res >= 0) {
596 struct inode *inode = file_inode(io->iocb->ki_filp);
597 struct fuse_conn *fc = get_fuse_conn(inode);
598 struct fuse_inode *fi = get_fuse_inode(inode);
599
600 spin_lock(&fc->lock);
601 fi->attr_version = ++fc->attr_version;
602 spin_unlock(&fc->lock);
603 }
604
605 io->iocb->ki_complete(io->iocb, res, 0);
606 }
607
608 kref_put(&io->refcnt, fuse_io_release);
609 }
610
611 static void fuse_aio_complete_req(struct fuse_conn *fc, struct fuse_req *req)
612 {
613 struct fuse_io_priv *io = req->io;
614 ssize_t pos = -1;
615
616 fuse_release_user_pages(req, !io->write);
617
618 if (io->write) {
619 if (req->misc.write.in.size != req->misc.write.out.size)
620 pos = req->misc.write.in.offset - io->offset +
621 req->misc.write.out.size;
622 } else {
623 if (req->misc.read.in.size != req->out.args[0].size)
624 pos = req->misc.read.in.offset - io->offset +
625 req->out.args[0].size;
626 }
627
628 fuse_aio_complete(io, req->out.h.error, pos);
629 }
630
631 static size_t fuse_async_req_send(struct fuse_conn *fc, struct fuse_req *req,
632 size_t num_bytes, struct fuse_io_priv *io)
633 {
634 spin_lock(&io->lock);
635 kref_get(&io->refcnt);
636 io->size += num_bytes;
637 io->reqs++;
638 spin_unlock(&io->lock);
639
640 req->io = io;
641 req->end = fuse_aio_complete_req;
642
643 __fuse_get_request(req);
644 fuse_request_send_background(fc, req);
645
646 return num_bytes;
647 }
648
649 static size_t fuse_send_read(struct fuse_req *req, struct fuse_io_priv *io,
650 loff_t pos, size_t count, fl_owner_t owner)
651 {
652 struct file *file = io->file;
653 struct fuse_file *ff = file->private_data;
654 struct fuse_conn *fc = ff->fc;
655
656 fuse_read_fill(req, file, pos, count, FUSE_READ);
657 if (owner != NULL) {
658 struct fuse_read_in *inarg = &req->misc.read.in;
659
660 inarg->read_flags |= FUSE_READ_LOCKOWNER;
661 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
662 }
663
664 if (io->async)
665 return fuse_async_req_send(fc, req, count, io);
666
667 fuse_request_send(fc, req);
668 return req->out.args[0].size;
669 }
670
671 static void fuse_read_update_size(struct inode *inode, loff_t size,
672 u64 attr_ver)
673 {
674 struct fuse_conn *fc = get_fuse_conn(inode);
675 struct fuse_inode *fi = get_fuse_inode(inode);
676
677 spin_lock(&fc->lock);
678 if (attr_ver == fi->attr_version && size < inode->i_size &&
679 !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
680 fi->attr_version = ++fc->attr_version;
681 i_size_write(inode, size);
682 }
683 spin_unlock(&fc->lock);
684 }
685
686 static void fuse_short_read(struct fuse_req *req, struct inode *inode,
687 u64 attr_ver)
688 {
689 size_t num_read = req->out.args[0].size;
690 struct fuse_conn *fc = get_fuse_conn(inode);
691
692 if (fc->writeback_cache) {
693 /*
694 * A hole in a file. Some data after the hole are in page cache,
695 * but have not reached the client fs yet. So, the hole is not
696 * present there.
697 */
698 int i;
699 int start_idx = num_read >> PAGE_SHIFT;
700 size_t off = num_read & (PAGE_SIZE - 1);
701
702 for (i = start_idx; i < req->num_pages; i++) {
703 zero_user_segment(req->pages[i], off, PAGE_SIZE);
704 off = 0;
705 }
706 } else {
707 loff_t pos = page_offset(req->pages[0]) + num_read;
708 fuse_read_update_size(inode, pos, attr_ver);
709 }
710 }
711
712 static int fuse_do_readpage(struct file *file, struct page *page)
713 {
714 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
715 struct inode *inode = page->mapping->host;
716 struct fuse_conn *fc = get_fuse_conn(inode);
717 struct fuse_req *req;
718 size_t num_read;
719 loff_t pos = page_offset(page);
720 size_t count = PAGE_SIZE;
721 u64 attr_ver;
722 int err;
723
724 /*
725 * Page writeback can extend beyond the lifetime of the
726 * page-cache page, so make sure we read a properly synced
727 * page.
728 */
729 fuse_wait_on_page_writeback(inode, page->index);
730
731 req = fuse_get_req(fc, 1);
732 if (IS_ERR(req))
733 return PTR_ERR(req);
734
735 attr_ver = fuse_get_attr_version(fc);
736
737 req->out.page_zeroing = 1;
738 req->out.argpages = 1;
739 req->num_pages = 1;
740 req->pages[0] = page;
741 req->page_descs[0].length = count;
742 num_read = fuse_send_read(req, &io, pos, count, NULL);
743 err = req->out.h.error;
744
745 if (!err) {
746 /*
747 * Short read means EOF. If file size is larger, truncate it
748 */
749 if (num_read < count)
750 fuse_short_read(req, inode, attr_ver);
751
752 SetPageUptodate(page);
753 }
754
755 fuse_put_request(fc, req);
756
757 return err;
758 }
759
760 static int fuse_readpage(struct file *file, struct page *page)
761 {
762 struct inode *inode = page->mapping->host;
763 int err;
764
765 err = -EIO;
766 if (is_bad_inode(inode))
767 goto out;
768
769 err = fuse_do_readpage(file, page);
770 fuse_invalidate_atime(inode);
771 out:
772 unlock_page(page);
773 return err;
774 }
775
776 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
777 {
778 int i;
779 size_t count = req->misc.read.in.size;
780 size_t num_read = req->out.args[0].size;
781 struct address_space *mapping = NULL;
782
783 for (i = 0; mapping == NULL && i < req->num_pages; i++)
784 mapping = req->pages[i]->mapping;
785
786 if (mapping) {
787 struct inode *inode = mapping->host;
788
789 /*
790 * Short read means EOF. If file size is larger, truncate it
791 */
792 if (!req->out.h.error && num_read < count)
793 fuse_short_read(req, inode, req->misc.read.attr_ver);
794
795 fuse_invalidate_atime(inode);
796 }
797
798 for (i = 0; i < req->num_pages; i++) {
799 struct page *page = req->pages[i];
800 if (!req->out.h.error)
801 SetPageUptodate(page);
802 else
803 SetPageError(page);
804 unlock_page(page);
805 put_page(page);
806 }
807 if (req->ff)
808 fuse_file_put(req->ff, false);
809 }
810
811 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
812 {
813 struct fuse_file *ff = file->private_data;
814 struct fuse_conn *fc = ff->fc;
815 loff_t pos = page_offset(req->pages[0]);
816 size_t count = req->num_pages << PAGE_SHIFT;
817
818 req->out.argpages = 1;
819 req->out.page_zeroing = 1;
820 req->out.page_replace = 1;
821 fuse_read_fill(req, file, pos, count, FUSE_READ);
822 req->misc.read.attr_ver = fuse_get_attr_version(fc);
823 if (fc->async_read) {
824 req->ff = fuse_file_get(ff);
825 req->end = fuse_readpages_end;
826 fuse_request_send_background(fc, req);
827 } else {
828 fuse_request_send(fc, req);
829 fuse_readpages_end(fc, req);
830 fuse_put_request(fc, req);
831 }
832 }
833
834 struct fuse_fill_data {
835 struct fuse_req *req;
836 struct file *file;
837 struct inode *inode;
838 unsigned nr_pages;
839 };
840
841 static int fuse_readpages_fill(void *_data, struct page *page)
842 {
843 struct fuse_fill_data *data = _data;
844 struct fuse_req *req = data->req;
845 struct inode *inode = data->inode;
846 struct fuse_conn *fc = get_fuse_conn(inode);
847
848 fuse_wait_on_page_writeback(inode, page->index);
849
850 if (req->num_pages &&
851 (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
852 (req->num_pages + 1) * PAGE_SIZE > fc->max_read ||
853 req->pages[req->num_pages - 1]->index + 1 != page->index)) {
854 int nr_alloc = min_t(unsigned, data->nr_pages,
855 FUSE_MAX_PAGES_PER_REQ);
856 fuse_send_readpages(req, data->file);
857 if (fc->async_read)
858 req = fuse_get_req_for_background(fc, nr_alloc);
859 else
860 req = fuse_get_req(fc, nr_alloc);
861
862 data->req = req;
863 if (IS_ERR(req)) {
864 unlock_page(page);
865 return PTR_ERR(req);
866 }
867 }
868
869 if (WARN_ON(req->num_pages >= req->max_pages)) {
870 fuse_put_request(fc, req);
871 return -EIO;
872 }
873
874 get_page(page);
875 req->pages[req->num_pages] = page;
876 req->page_descs[req->num_pages].length = PAGE_SIZE;
877 req->num_pages++;
878 data->nr_pages--;
879 return 0;
880 }
881
882 static int fuse_readpages(struct file *file, struct address_space *mapping,
883 struct list_head *pages, unsigned nr_pages)
884 {
885 struct inode *inode = mapping->host;
886 struct fuse_conn *fc = get_fuse_conn(inode);
887 struct fuse_fill_data data;
888 int err;
889 int nr_alloc = min_t(unsigned, nr_pages, FUSE_MAX_PAGES_PER_REQ);
890
891 err = -EIO;
892 if (is_bad_inode(inode))
893 goto out;
894
895 data.file = file;
896 data.inode = inode;
897 if (fc->async_read)
898 data.req = fuse_get_req_for_background(fc, nr_alloc);
899 else
900 data.req = fuse_get_req(fc, nr_alloc);
901 data.nr_pages = nr_pages;
902 err = PTR_ERR(data.req);
903 if (IS_ERR(data.req))
904 goto out;
905
906 err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
907 if (!err) {
908 if (data.req->num_pages)
909 fuse_send_readpages(data.req, file);
910 else
911 fuse_put_request(fc, data.req);
912 }
913 out:
914 return err;
915 }
916
917 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
918 {
919 struct inode *inode = iocb->ki_filp->f_mapping->host;
920 struct fuse_conn *fc = get_fuse_conn(inode);
921
922 /*
923 * In auto invalidate mode, always update attributes on read.
924 * Otherwise, only update if we attempt to read past EOF (to ensure
925 * i_size is up to date).
926 */
927 if (fc->auto_inval_data ||
928 (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
929 int err;
930 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
931 if (err)
932 return err;
933 }
934
935 return generic_file_read_iter(iocb, to);
936 }
937
938 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
939 loff_t pos, size_t count)
940 {
941 struct fuse_write_in *inarg = &req->misc.write.in;
942 struct fuse_write_out *outarg = &req->misc.write.out;
943
944 inarg->fh = ff->fh;
945 inarg->offset = pos;
946 inarg->size = count;
947 req->in.h.opcode = FUSE_WRITE;
948 req->in.h.nodeid = ff->nodeid;
949 req->in.numargs = 2;
950 if (ff->fc->minor < 9)
951 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
952 else
953 req->in.args[0].size = sizeof(struct fuse_write_in);
954 req->in.args[0].value = inarg;
955 req->in.args[1].size = count;
956 req->out.numargs = 1;
957 req->out.args[0].size = sizeof(struct fuse_write_out);
958 req->out.args[0].value = outarg;
959 }
960
961 static size_t fuse_send_write(struct fuse_req *req, struct fuse_io_priv *io,
962 loff_t pos, size_t count, fl_owner_t owner)
963 {
964 struct file *file = io->file;
965 struct fuse_file *ff = file->private_data;
966 struct fuse_conn *fc = ff->fc;
967 struct fuse_write_in *inarg = &req->misc.write.in;
968
969 fuse_write_fill(req, ff, pos, count);
970 inarg->flags = file->f_flags;
971 if (owner != NULL) {
972 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
973 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
974 }
975
976 if (io->async)
977 return fuse_async_req_send(fc, req, count, io);
978
979 fuse_request_send(fc, req);
980 return req->misc.write.out.size;
981 }
982
983 bool fuse_write_update_size(struct inode *inode, loff_t pos)
984 {
985 struct fuse_conn *fc = get_fuse_conn(inode);
986 struct fuse_inode *fi = get_fuse_inode(inode);
987 bool ret = false;
988
989 spin_lock(&fc->lock);
990 fi->attr_version = ++fc->attr_version;
991 if (pos > inode->i_size) {
992 i_size_write(inode, pos);
993 ret = true;
994 }
995 spin_unlock(&fc->lock);
996
997 return ret;
998 }
999
1000 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
1001 struct inode *inode, loff_t pos,
1002 size_t count)
1003 {
1004 size_t res;
1005 unsigned offset;
1006 unsigned i;
1007 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
1008
1009 for (i = 0; i < req->num_pages; i++)
1010 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
1011
1012 res = fuse_send_write(req, &io, pos, count, NULL);
1013
1014 offset = req->page_descs[0].offset;
1015 count = res;
1016 for (i = 0; i < req->num_pages; i++) {
1017 struct page *page = req->pages[i];
1018
1019 if (!req->out.h.error && !offset && count >= PAGE_SIZE)
1020 SetPageUptodate(page);
1021
1022 if (count > PAGE_SIZE - offset)
1023 count -= PAGE_SIZE - offset;
1024 else
1025 count = 0;
1026 offset = 0;
1027
1028 unlock_page(page);
1029 put_page(page);
1030 }
1031
1032 return res;
1033 }
1034
1035 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
1036 struct address_space *mapping,
1037 struct iov_iter *ii, loff_t pos)
1038 {
1039 struct fuse_conn *fc = get_fuse_conn(mapping->host);
1040 unsigned offset = pos & (PAGE_SIZE - 1);
1041 size_t count = 0;
1042 int err;
1043
1044 req->in.argpages = 1;
1045 req->page_descs[0].offset = offset;
1046
1047 do {
1048 size_t tmp;
1049 struct page *page;
1050 pgoff_t index = pos >> PAGE_SHIFT;
1051 size_t bytes = min_t(size_t, PAGE_SIZE - offset,
1052 iov_iter_count(ii));
1053
1054 bytes = min_t(size_t, bytes, fc->max_write - count);
1055
1056 again:
1057 err = -EFAULT;
1058 if (iov_iter_fault_in_readable(ii, bytes))
1059 break;
1060
1061 err = -ENOMEM;
1062 page = grab_cache_page_write_begin(mapping, index, 0);
1063 if (!page)
1064 break;
1065
1066 if (mapping_writably_mapped(mapping))
1067 flush_dcache_page(page);
1068
1069 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
1070 flush_dcache_page(page);
1071
1072 iov_iter_advance(ii, tmp);
1073 if (!tmp) {
1074 unlock_page(page);
1075 put_page(page);
1076 bytes = min(bytes, iov_iter_single_seg_count(ii));
1077 goto again;
1078 }
1079
1080 err = 0;
1081 req->pages[req->num_pages] = page;
1082 req->page_descs[req->num_pages].length = tmp;
1083 req->num_pages++;
1084
1085 count += tmp;
1086 pos += tmp;
1087 offset += tmp;
1088 if (offset == PAGE_SIZE)
1089 offset = 0;
1090
1091 if (!fc->big_writes)
1092 break;
1093 } while (iov_iter_count(ii) && count < fc->max_write &&
1094 req->num_pages < req->max_pages && offset == 0);
1095
1096 return count > 0 ? count : err;
1097 }
1098
1099 static inline unsigned fuse_wr_pages(loff_t pos, size_t len)
1100 {
1101 return min_t(unsigned,
1102 ((pos + len - 1) >> PAGE_SHIFT) -
1103 (pos >> PAGE_SHIFT) + 1,
1104 FUSE_MAX_PAGES_PER_REQ);
1105 }
1106
1107 static ssize_t fuse_perform_write(struct file *file,
1108 struct address_space *mapping,
1109 struct iov_iter *ii, loff_t pos)
1110 {
1111 struct inode *inode = mapping->host;
1112 struct fuse_conn *fc = get_fuse_conn(inode);
1113 struct fuse_inode *fi = get_fuse_inode(inode);
1114 int err = 0;
1115 ssize_t res = 0;
1116
1117 if (is_bad_inode(inode))
1118 return -EIO;
1119
1120 if (inode->i_size < pos + iov_iter_count(ii))
1121 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1122
1123 do {
1124 struct fuse_req *req;
1125 ssize_t count;
1126 unsigned nr_pages = fuse_wr_pages(pos, iov_iter_count(ii));
1127
1128 req = fuse_get_req(fc, nr_pages);
1129 if (IS_ERR(req)) {
1130 err = PTR_ERR(req);
1131 break;
1132 }
1133
1134 count = fuse_fill_write_pages(req, mapping, ii, pos);
1135 if (count <= 0) {
1136 err = count;
1137 } else {
1138 size_t num_written;
1139
1140 num_written = fuse_send_write_pages(req, file, inode,
1141 pos, count);
1142 err = req->out.h.error;
1143 if (!err) {
1144 res += num_written;
1145 pos += num_written;
1146
1147 /* break out of the loop on short write */
1148 if (num_written != count)
1149 err = -EIO;
1150 }
1151 }
1152 fuse_put_request(fc, req);
1153 } while (!err && iov_iter_count(ii));
1154
1155 if (res > 0)
1156 fuse_write_update_size(inode, pos);
1157
1158 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1159 fuse_invalidate_attr(inode);
1160
1161 return res > 0 ? res : err;
1162 }
1163
1164 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1165 {
1166 struct file *file = iocb->ki_filp;
1167 struct address_space *mapping = file->f_mapping;
1168 ssize_t written = 0;
1169 ssize_t written_buffered = 0;
1170 struct inode *inode = mapping->host;
1171 ssize_t err;
1172 loff_t endbyte = 0;
1173
1174 if (get_fuse_conn(inode)->writeback_cache) {
1175 /* Update size (EOF optimization) and mode (SUID clearing) */
1176 err = fuse_update_attributes(mapping->host, NULL, file, NULL);
1177 if (err)
1178 return err;
1179
1180 return generic_file_write_iter(iocb, from);
1181 }
1182
1183 inode_lock(inode);
1184
1185 /* We can write back this queue in page reclaim */
1186 current->backing_dev_info = inode_to_bdi(inode);
1187
1188 err = generic_write_checks(iocb, from);
1189 if (err <= 0)
1190 goto out;
1191
1192 err = file_remove_privs(file);
1193 if (err)
1194 goto out;
1195
1196 err = file_update_time(file);
1197 if (err)
1198 goto out;
1199
1200 if (iocb->ki_flags & IOCB_DIRECT) {
1201 loff_t pos = iocb->ki_pos;
1202 written = generic_file_direct_write(iocb, from);
1203 if (written < 0 || !iov_iter_count(from))
1204 goto out;
1205
1206 pos += written;
1207
1208 written_buffered = fuse_perform_write(file, mapping, from, pos);
1209 if (written_buffered < 0) {
1210 err = written_buffered;
1211 goto out;
1212 }
1213 endbyte = pos + written_buffered - 1;
1214
1215 err = filemap_write_and_wait_range(file->f_mapping, pos,
1216 endbyte);
1217 if (err)
1218 goto out;
1219
1220 invalidate_mapping_pages(file->f_mapping,
1221 pos >> PAGE_SHIFT,
1222 endbyte >> PAGE_SHIFT);
1223
1224 written += written_buffered;
1225 iocb->ki_pos = pos + written_buffered;
1226 } else {
1227 written = fuse_perform_write(file, mapping, from, iocb->ki_pos);
1228 if (written >= 0)
1229 iocb->ki_pos += written;
1230 }
1231 out:
1232 current->backing_dev_info = NULL;
1233 inode_unlock(inode);
1234
1235 return written ? written : err;
1236 }
1237
1238 static inline void fuse_page_descs_length_init(struct fuse_req *req,
1239 unsigned index, unsigned nr_pages)
1240 {
1241 int i;
1242
1243 for (i = index; i < index + nr_pages; i++)
1244 req->page_descs[i].length = PAGE_SIZE -
1245 req->page_descs[i].offset;
1246 }
1247
1248 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1249 {
1250 return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1251 }
1252
1253 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1254 size_t max_size)
1255 {
1256 return min(iov_iter_single_seg_count(ii), max_size);
1257 }
1258
1259 static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1260 size_t *nbytesp, int write)
1261 {
1262 size_t nbytes = 0; /* # bytes already packed in req */
1263 ssize_t ret = 0;
1264
1265 /* Special case for kernel I/O: can copy directly into the buffer */
1266 if (ii->type & ITER_KVEC) {
1267 unsigned long user_addr = fuse_get_user_addr(ii);
1268 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1269
1270 if (write)
1271 req->in.args[1].value = (void *) user_addr;
1272 else
1273 req->out.args[0].value = (void *) user_addr;
1274
1275 iov_iter_advance(ii, frag_size);
1276 *nbytesp = frag_size;
1277 return 0;
1278 }
1279
1280 while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1281 unsigned npages;
1282 size_t start;
1283 ret = iov_iter_get_pages(ii, &req->pages[req->num_pages],
1284 *nbytesp - nbytes,
1285 req->max_pages - req->num_pages,
1286 &start);
1287 if (ret < 0)
1288 break;
1289
1290 iov_iter_advance(ii, ret);
1291 nbytes += ret;
1292
1293 ret += start;
1294 npages = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
1295
1296 req->page_descs[req->num_pages].offset = start;
1297 fuse_page_descs_length_init(req, req->num_pages, npages);
1298
1299 req->num_pages += npages;
1300 req->page_descs[req->num_pages - 1].length -=
1301 (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1302 }
1303
1304 if (write)
1305 req->in.argpages = 1;
1306 else
1307 req->out.argpages = 1;
1308
1309 *nbytesp = nbytes;
1310
1311 return ret < 0 ? ret : 0;
1312 }
1313
1314 static inline int fuse_iter_npages(const struct iov_iter *ii_p)
1315 {
1316 return iov_iter_npages(ii_p, FUSE_MAX_PAGES_PER_REQ);
1317 }
1318
1319 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1320 loff_t *ppos, int flags)
1321 {
1322 int write = flags & FUSE_DIO_WRITE;
1323 bool should_dirty = !write && iter_is_iovec(iter);
1324 int cuse = flags & FUSE_DIO_CUSE;
1325 struct file *file = io->file;
1326 struct inode *inode = file->f_mapping->host;
1327 struct fuse_file *ff = file->private_data;
1328 struct fuse_conn *fc = ff->fc;
1329 size_t nmax = write ? fc->max_write : fc->max_read;
1330 loff_t pos = *ppos;
1331 size_t count = iov_iter_count(iter);
1332 pgoff_t idx_from = pos >> PAGE_SHIFT;
1333 pgoff_t idx_to = (pos + count - 1) >> PAGE_SHIFT;
1334 ssize_t res = 0;
1335 struct fuse_req *req;
1336 int err = 0;
1337
1338 if (io->async)
1339 req = fuse_get_req_for_background(fc, fuse_iter_npages(iter));
1340 else
1341 req = fuse_get_req(fc, fuse_iter_npages(iter));
1342 if (IS_ERR(req))
1343 return PTR_ERR(req);
1344
1345 if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1346 if (!write)
1347 inode_lock(inode);
1348 fuse_sync_writes(inode);
1349 if (!write)
1350 inode_unlock(inode);
1351 }
1352
1353 while (count) {
1354 size_t nres;
1355 fl_owner_t owner = current->files;
1356 size_t nbytes = min(count, nmax);
1357 err = fuse_get_user_pages(req, iter, &nbytes, write);
1358 if (err && !nbytes)
1359 break;
1360
1361 if (write)
1362 nres = fuse_send_write(req, io, pos, nbytes, owner);
1363 else
1364 nres = fuse_send_read(req, io, pos, nbytes, owner);
1365
1366 if (!io->async)
1367 fuse_release_user_pages(req, should_dirty);
1368 if (req->out.h.error) {
1369 err = req->out.h.error;
1370 break;
1371 } else if (nres > nbytes) {
1372 res = 0;
1373 err = -EIO;
1374 break;
1375 }
1376 count -= nres;
1377 res += nres;
1378 pos += nres;
1379 if (nres != nbytes)
1380 break;
1381 if (count) {
1382 fuse_put_request(fc, req);
1383 if (io->async)
1384 req = fuse_get_req_for_background(fc,
1385 fuse_iter_npages(iter));
1386 else
1387 req = fuse_get_req(fc, fuse_iter_npages(iter));
1388 if (IS_ERR(req))
1389 break;
1390 }
1391 }
1392 if (!IS_ERR(req))
1393 fuse_put_request(fc, req);
1394 if (res > 0)
1395 *ppos = pos;
1396
1397 return res > 0 ? res : err;
1398 }
1399 EXPORT_SYMBOL_GPL(fuse_direct_io);
1400
1401 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1402 struct iov_iter *iter,
1403 loff_t *ppos)
1404 {
1405 ssize_t res;
1406 struct file *file = io->file;
1407 struct inode *inode = file_inode(file);
1408
1409 if (is_bad_inode(inode))
1410 return -EIO;
1411
1412 res = fuse_direct_io(io, iter, ppos, 0);
1413
1414 fuse_invalidate_attr(inode);
1415
1416 return res;
1417 }
1418
1419 static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1420 {
1421 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb->ki_filp);
1422 return __fuse_direct_read(&io, to, &iocb->ki_pos);
1423 }
1424
1425 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1426 {
1427 struct file *file = iocb->ki_filp;
1428 struct inode *inode = file_inode(file);
1429 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
1430 ssize_t res;
1431
1432 if (is_bad_inode(inode))
1433 return -EIO;
1434
1435 /* Don't allow parallel writes to the same file */
1436 inode_lock(inode);
1437 res = generic_write_checks(iocb, from);
1438 if (res > 0)
1439 res = fuse_direct_io(&io, from, &iocb->ki_pos, FUSE_DIO_WRITE);
1440 fuse_invalidate_attr(inode);
1441 if (res > 0)
1442 fuse_write_update_size(inode, iocb->ki_pos);
1443 inode_unlock(inode);
1444
1445 return res;
1446 }
1447
1448 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1449 {
1450 int i;
1451
1452 for (i = 0; i < req->num_pages; i++)
1453 __free_page(req->pages[i]);
1454
1455 if (req->ff)
1456 fuse_file_put(req->ff, false);
1457 }
1458
1459 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1460 {
1461 struct inode *inode = req->inode;
1462 struct fuse_inode *fi = get_fuse_inode(inode);
1463 struct backing_dev_info *bdi = inode_to_bdi(inode);
1464 int i;
1465
1466 list_del(&req->writepages_entry);
1467 for (i = 0; i < req->num_pages; i++) {
1468 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1469 dec_node_page_state(req->pages[i], NR_WRITEBACK_TEMP);
1470 wb_writeout_inc(&bdi->wb);
1471 }
1472 wake_up(&fi->page_waitq);
1473 }
1474
1475 /* Called under fc->lock, may release and reacquire it */
1476 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req,
1477 loff_t size)
1478 __releases(fc->lock)
1479 __acquires(fc->lock)
1480 {
1481 struct fuse_inode *fi = get_fuse_inode(req->inode);
1482 struct fuse_write_in *inarg = &req->misc.write.in;
1483 __u64 data_size = req->num_pages * PAGE_SIZE;
1484
1485 if (!fc->connected)
1486 goto out_free;
1487
1488 if (inarg->offset + data_size <= size) {
1489 inarg->size = data_size;
1490 } else if (inarg->offset < size) {
1491 inarg->size = size - inarg->offset;
1492 } else {
1493 /* Got truncated off completely */
1494 goto out_free;
1495 }
1496
1497 req->in.args[1].size = inarg->size;
1498 fi->writectr++;
1499 fuse_request_send_background_locked(fc, req);
1500 return;
1501
1502 out_free:
1503 fuse_writepage_finish(fc, req);
1504 spin_unlock(&fc->lock);
1505 fuse_writepage_free(fc, req);
1506 fuse_put_request(fc, req);
1507 spin_lock(&fc->lock);
1508 }
1509
1510 /*
1511 * If fi->writectr is positive (no truncate or fsync going on) send
1512 * all queued writepage requests.
1513 *
1514 * Called with fc->lock
1515 */
1516 void fuse_flush_writepages(struct inode *inode)
1517 __releases(fc->lock)
1518 __acquires(fc->lock)
1519 {
1520 struct fuse_conn *fc = get_fuse_conn(inode);
1521 struct fuse_inode *fi = get_fuse_inode(inode);
1522 size_t crop = i_size_read(inode);
1523 struct fuse_req *req;
1524
1525 while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1526 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1527 list_del_init(&req->list);
1528 fuse_send_writepage(fc, req, crop);
1529 }
1530 }
1531
1532 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1533 {
1534 struct inode *inode = req->inode;
1535 struct fuse_inode *fi = get_fuse_inode(inode);
1536
1537 mapping_set_error(inode->i_mapping, req->out.h.error);
1538 spin_lock(&fc->lock);
1539 while (req->misc.write.next) {
1540 struct fuse_conn *fc = get_fuse_conn(inode);
1541 struct fuse_write_in *inarg = &req->misc.write.in;
1542 struct fuse_req *next = req->misc.write.next;
1543 req->misc.write.next = next->misc.write.next;
1544 next->misc.write.next = NULL;
1545 next->ff = fuse_file_get(req->ff);
1546 list_add(&next->writepages_entry, &fi->writepages);
1547
1548 /*
1549 * Skip fuse_flush_writepages() to make it easy to crop requests
1550 * based on primary request size.
1551 *
1552 * 1st case (trivial): there are no concurrent activities using
1553 * fuse_set/release_nowrite. Then we're on safe side because
1554 * fuse_flush_writepages() would call fuse_send_writepage()
1555 * anyway.
1556 *
1557 * 2nd case: someone called fuse_set_nowrite and it is waiting
1558 * now for completion of all in-flight requests. This happens
1559 * rarely and no more than once per page, so this should be
1560 * okay.
1561 *
1562 * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1563 * of fuse_set_nowrite..fuse_release_nowrite section. The fact
1564 * that fuse_set_nowrite returned implies that all in-flight
1565 * requests were completed along with all of their secondary
1566 * requests. Further primary requests are blocked by negative
1567 * writectr. Hence there cannot be any in-flight requests and
1568 * no invocations of fuse_writepage_end() while we're in
1569 * fuse_set_nowrite..fuse_release_nowrite section.
1570 */
1571 fuse_send_writepage(fc, next, inarg->offset + inarg->size);
1572 }
1573 fi->writectr--;
1574 fuse_writepage_finish(fc, req);
1575 spin_unlock(&fc->lock);
1576 fuse_writepage_free(fc, req);
1577 }
1578
1579 static struct fuse_file *__fuse_write_file_get(struct fuse_conn *fc,
1580 struct fuse_inode *fi)
1581 {
1582 struct fuse_file *ff = NULL;
1583
1584 spin_lock(&fc->lock);
1585 if (!list_empty(&fi->write_files)) {
1586 ff = list_entry(fi->write_files.next, struct fuse_file,
1587 write_entry);
1588 fuse_file_get(ff);
1589 }
1590 spin_unlock(&fc->lock);
1591
1592 return ff;
1593 }
1594
1595 static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1596 struct fuse_inode *fi)
1597 {
1598 struct fuse_file *ff = __fuse_write_file_get(fc, fi);
1599 WARN_ON(!ff);
1600 return ff;
1601 }
1602
1603 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1604 {
1605 struct fuse_conn *fc = get_fuse_conn(inode);
1606 struct fuse_inode *fi = get_fuse_inode(inode);
1607 struct fuse_file *ff;
1608 int err;
1609
1610 ff = __fuse_write_file_get(fc, fi);
1611 err = fuse_flush_times(inode, ff);
1612 if (ff)
1613 fuse_file_put(ff, 0);
1614
1615 return err;
1616 }
1617
1618 static int fuse_writepage_locked(struct page *page)
1619 {
1620 struct address_space *mapping = page->mapping;
1621 struct inode *inode = mapping->host;
1622 struct fuse_conn *fc = get_fuse_conn(inode);
1623 struct fuse_inode *fi = get_fuse_inode(inode);
1624 struct fuse_req *req;
1625 struct page *tmp_page;
1626 int error = -ENOMEM;
1627
1628 set_page_writeback(page);
1629
1630 req = fuse_request_alloc_nofs(1);
1631 if (!req)
1632 goto err;
1633
1634 /* writeback always goes to bg_queue */
1635 __set_bit(FR_BACKGROUND, &req->flags);
1636 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1637 if (!tmp_page)
1638 goto err_free;
1639
1640 error = -EIO;
1641 req->ff = fuse_write_file_get(fc, fi);
1642 if (!req->ff)
1643 goto err_nofile;
1644
1645 fuse_write_fill(req, req->ff, page_offset(page), 0);
1646
1647 copy_highpage(tmp_page, page);
1648 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1649 req->misc.write.next = NULL;
1650 req->in.argpages = 1;
1651 req->num_pages = 1;
1652 req->pages[0] = tmp_page;
1653 req->page_descs[0].offset = 0;
1654 req->page_descs[0].length = PAGE_SIZE;
1655 req->end = fuse_writepage_end;
1656 req->inode = inode;
1657
1658 inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1659 inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1660
1661 spin_lock(&fc->lock);
1662 list_add(&req->writepages_entry, &fi->writepages);
1663 list_add_tail(&req->list, &fi->queued_writes);
1664 fuse_flush_writepages(inode);
1665 spin_unlock(&fc->lock);
1666
1667 end_page_writeback(page);
1668
1669 return 0;
1670
1671 err_nofile:
1672 __free_page(tmp_page);
1673 err_free:
1674 fuse_request_free(req);
1675 err:
1676 end_page_writeback(page);
1677 return error;
1678 }
1679
1680 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1681 {
1682 int err;
1683
1684 if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1685 /*
1686 * ->writepages() should be called for sync() and friends. We
1687 * should only get here on direct reclaim and then we are
1688 * allowed to skip a page which is already in flight
1689 */
1690 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1691
1692 redirty_page_for_writepage(wbc, page);
1693 return 0;
1694 }
1695
1696 err = fuse_writepage_locked(page);
1697 unlock_page(page);
1698
1699 return err;
1700 }
1701
1702 struct fuse_fill_wb_data {
1703 struct fuse_req *req;
1704 struct fuse_file *ff;
1705 struct inode *inode;
1706 struct page **orig_pages;
1707 };
1708
1709 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1710 {
1711 struct fuse_req *req = data->req;
1712 struct inode *inode = data->inode;
1713 struct fuse_conn *fc = get_fuse_conn(inode);
1714 struct fuse_inode *fi = get_fuse_inode(inode);
1715 int num_pages = req->num_pages;
1716 int i;
1717
1718 req->ff = fuse_file_get(data->ff);
1719 spin_lock(&fc->lock);
1720 list_add_tail(&req->list, &fi->queued_writes);
1721 fuse_flush_writepages(inode);
1722 spin_unlock(&fc->lock);
1723
1724 for (i = 0; i < num_pages; i++)
1725 end_page_writeback(data->orig_pages[i]);
1726 }
1727
1728 static bool fuse_writepage_in_flight(struct fuse_req *new_req,
1729 struct page *page)
1730 {
1731 struct fuse_conn *fc = get_fuse_conn(new_req->inode);
1732 struct fuse_inode *fi = get_fuse_inode(new_req->inode);
1733 struct fuse_req *tmp;
1734 struct fuse_req *old_req;
1735 bool found = false;
1736 pgoff_t curr_index;
1737
1738 BUG_ON(new_req->num_pages != 0);
1739
1740 spin_lock(&fc->lock);
1741 list_del(&new_req->writepages_entry);
1742 list_for_each_entry(old_req, &fi->writepages, writepages_entry) {
1743 BUG_ON(old_req->inode != new_req->inode);
1744 curr_index = old_req->misc.write.in.offset >> PAGE_SHIFT;
1745 if (curr_index <= page->index &&
1746 page->index < curr_index + old_req->num_pages) {
1747 found = true;
1748 break;
1749 }
1750 }
1751 if (!found) {
1752 list_add(&new_req->writepages_entry, &fi->writepages);
1753 goto out_unlock;
1754 }
1755
1756 new_req->num_pages = 1;
1757 for (tmp = old_req; tmp != NULL; tmp = tmp->misc.write.next) {
1758 BUG_ON(tmp->inode != new_req->inode);
1759 curr_index = tmp->misc.write.in.offset >> PAGE_SHIFT;
1760 if (tmp->num_pages == 1 &&
1761 curr_index == page->index) {
1762 old_req = tmp;
1763 }
1764 }
1765
1766 if (old_req->num_pages == 1 && test_bit(FR_PENDING, &old_req->flags)) {
1767 struct backing_dev_info *bdi = inode_to_bdi(page->mapping->host);
1768
1769 copy_highpage(old_req->pages[0], page);
1770 spin_unlock(&fc->lock);
1771
1772 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1773 dec_node_page_state(page, NR_WRITEBACK_TEMP);
1774 wb_writeout_inc(&bdi->wb);
1775 fuse_writepage_free(fc, new_req);
1776 fuse_request_free(new_req);
1777 goto out;
1778 } else {
1779 new_req->misc.write.next = old_req->misc.write.next;
1780 old_req->misc.write.next = new_req;
1781 }
1782 out_unlock:
1783 spin_unlock(&fc->lock);
1784 out:
1785 return found;
1786 }
1787
1788 static int fuse_writepages_fill(struct page *page,
1789 struct writeback_control *wbc, void *_data)
1790 {
1791 struct fuse_fill_wb_data *data = _data;
1792 struct fuse_req *req = data->req;
1793 struct inode *inode = data->inode;
1794 struct fuse_conn *fc = get_fuse_conn(inode);
1795 struct page *tmp_page;
1796 bool is_writeback;
1797 int err;
1798
1799 if (!data->ff) {
1800 err = -EIO;
1801 data->ff = fuse_write_file_get(fc, get_fuse_inode(inode));
1802 if (!data->ff)
1803 goto out_unlock;
1804 }
1805
1806 /*
1807 * Being under writeback is unlikely but possible. For example direct
1808 * read to an mmaped fuse file will set the page dirty twice; once when
1809 * the pages are faulted with get_user_pages(), and then after the read
1810 * completed.
1811 */
1812 is_writeback = fuse_page_is_writeback(inode, page->index);
1813
1814 if (req && req->num_pages &&
1815 (is_writeback || req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
1816 (req->num_pages + 1) * PAGE_SIZE > fc->max_write ||
1817 data->orig_pages[req->num_pages - 1]->index + 1 != page->index)) {
1818 fuse_writepages_send(data);
1819 data->req = NULL;
1820 }
1821 err = -ENOMEM;
1822 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1823 if (!tmp_page)
1824 goto out_unlock;
1825
1826 /*
1827 * The page must not be redirtied until the writeout is completed
1828 * (i.e. userspace has sent a reply to the write request). Otherwise
1829 * there could be more than one temporary page instance for each real
1830 * page.
1831 *
1832 * This is ensured by holding the page lock in page_mkwrite() while
1833 * checking fuse_page_is_writeback(). We already hold the page lock
1834 * since clear_page_dirty_for_io() and keep it held until we add the
1835 * request to the fi->writepages list and increment req->num_pages.
1836 * After this fuse_page_is_writeback() will indicate that the page is
1837 * under writeback, so we can release the page lock.
1838 */
1839 if (data->req == NULL) {
1840 struct fuse_inode *fi = get_fuse_inode(inode);
1841
1842 err = -ENOMEM;
1843 req = fuse_request_alloc_nofs(FUSE_MAX_PAGES_PER_REQ);
1844 if (!req) {
1845 __free_page(tmp_page);
1846 goto out_unlock;
1847 }
1848
1849 fuse_write_fill(req, data->ff, page_offset(page), 0);
1850 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1851 req->misc.write.next = NULL;
1852 req->in.argpages = 1;
1853 __set_bit(FR_BACKGROUND, &req->flags);
1854 req->num_pages = 0;
1855 req->end = fuse_writepage_end;
1856 req->inode = inode;
1857
1858 spin_lock(&fc->lock);
1859 list_add(&req->writepages_entry, &fi->writepages);
1860 spin_unlock(&fc->lock);
1861
1862 data->req = req;
1863 }
1864 set_page_writeback(page);
1865
1866 copy_highpage(tmp_page, page);
1867 req->pages[req->num_pages] = tmp_page;
1868 req->page_descs[req->num_pages].offset = 0;
1869 req->page_descs[req->num_pages].length = PAGE_SIZE;
1870
1871 inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1872 inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1873
1874 err = 0;
1875 if (is_writeback && fuse_writepage_in_flight(req, page)) {
1876 end_page_writeback(page);
1877 data->req = NULL;
1878 goto out_unlock;
1879 }
1880 data->orig_pages[req->num_pages] = page;
1881
1882 /*
1883 * Protected by fc->lock against concurrent access by
1884 * fuse_page_is_writeback().
1885 */
1886 spin_lock(&fc->lock);
1887 req->num_pages++;
1888 spin_unlock(&fc->lock);
1889
1890 out_unlock:
1891 unlock_page(page);
1892
1893 return err;
1894 }
1895
1896 static int fuse_writepages(struct address_space *mapping,
1897 struct writeback_control *wbc)
1898 {
1899 struct inode *inode = mapping->host;
1900 struct fuse_fill_wb_data data;
1901 int err;
1902
1903 err = -EIO;
1904 if (is_bad_inode(inode))
1905 goto out;
1906
1907 data.inode = inode;
1908 data.req = NULL;
1909 data.ff = NULL;
1910
1911 err = -ENOMEM;
1912 data.orig_pages = kcalloc(FUSE_MAX_PAGES_PER_REQ,
1913 sizeof(struct page *),
1914 GFP_NOFS);
1915 if (!data.orig_pages)
1916 goto out;
1917
1918 err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
1919 if (data.req) {
1920 /* Ignore errors if we can write at least one page */
1921 BUG_ON(!data.req->num_pages);
1922 fuse_writepages_send(&data);
1923 err = 0;
1924 }
1925 if (data.ff)
1926 fuse_file_put(data.ff, false);
1927
1928 kfree(data.orig_pages);
1929 out:
1930 return err;
1931 }
1932
1933 /*
1934 * It's worthy to make sure that space is reserved on disk for the write,
1935 * but how to implement it without killing performance need more thinking.
1936 */
1937 static int fuse_write_begin(struct file *file, struct address_space *mapping,
1938 loff_t pos, unsigned len, unsigned flags,
1939 struct page **pagep, void **fsdata)
1940 {
1941 pgoff_t index = pos >> PAGE_SHIFT;
1942 struct fuse_conn *fc = get_fuse_conn(file_inode(file));
1943 struct page *page;
1944 loff_t fsize;
1945 int err = -ENOMEM;
1946
1947 WARN_ON(!fc->writeback_cache);
1948
1949 page = grab_cache_page_write_begin(mapping, index, flags);
1950 if (!page)
1951 goto error;
1952
1953 fuse_wait_on_page_writeback(mapping->host, page->index);
1954
1955 if (PageUptodate(page) || len == PAGE_SIZE)
1956 goto success;
1957 /*
1958 * Check if the start this page comes after the end of file, in which
1959 * case the readpage can be optimized away.
1960 */
1961 fsize = i_size_read(mapping->host);
1962 if (fsize <= (pos & PAGE_MASK)) {
1963 size_t off = pos & ~PAGE_MASK;
1964 if (off)
1965 zero_user_segment(page, 0, off);
1966 goto success;
1967 }
1968 err = fuse_do_readpage(file, page);
1969 if (err)
1970 goto cleanup;
1971 success:
1972 *pagep = page;
1973 return 0;
1974
1975 cleanup:
1976 unlock_page(page);
1977 put_page(page);
1978 error:
1979 return err;
1980 }
1981
1982 static int fuse_write_end(struct file *file, struct address_space *mapping,
1983 loff_t pos, unsigned len, unsigned copied,
1984 struct page *page, void *fsdata)
1985 {
1986 struct inode *inode = page->mapping->host;
1987
1988 /* Haven't copied anything? Skip zeroing, size extending, dirtying. */
1989 if (!copied)
1990 goto unlock;
1991
1992 if (!PageUptodate(page)) {
1993 /* Zero any unwritten bytes at the end of the page */
1994 size_t endoff = (pos + copied) & ~PAGE_MASK;
1995 if (endoff)
1996 zero_user_segment(page, endoff, PAGE_SIZE);
1997 SetPageUptodate(page);
1998 }
1999
2000 fuse_write_update_size(inode, pos + copied);
2001 set_page_dirty(page);
2002
2003 unlock:
2004 unlock_page(page);
2005 put_page(page);
2006
2007 return copied;
2008 }
2009
2010 static int fuse_launder_page(struct page *page)
2011 {
2012 int err = 0;
2013 if (clear_page_dirty_for_io(page)) {
2014 struct inode *inode = page->mapping->host;
2015 err = fuse_writepage_locked(page);
2016 if (!err)
2017 fuse_wait_on_page_writeback(inode, page->index);
2018 }
2019 return err;
2020 }
2021
2022 /*
2023 * Write back dirty pages now, because there may not be any suitable
2024 * open files later
2025 */
2026 static void fuse_vma_close(struct vm_area_struct *vma)
2027 {
2028 filemap_write_and_wait(vma->vm_file->f_mapping);
2029 }
2030
2031 /*
2032 * Wait for writeback against this page to complete before allowing it
2033 * to be marked dirty again, and hence written back again, possibly
2034 * before the previous writepage completed.
2035 *
2036 * Block here, instead of in ->writepage(), so that the userspace fs
2037 * can only block processes actually operating on the filesystem.
2038 *
2039 * Otherwise unprivileged userspace fs would be able to block
2040 * unrelated:
2041 *
2042 * - page migration
2043 * - sync(2)
2044 * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2045 */
2046 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
2047 {
2048 struct page *page = vmf->page;
2049 struct inode *inode = file_inode(vma->vm_file);
2050
2051 file_update_time(vma->vm_file);
2052 lock_page(page);
2053 if (page->mapping != inode->i_mapping) {
2054 unlock_page(page);
2055 return VM_FAULT_NOPAGE;
2056 }
2057
2058 fuse_wait_on_page_writeback(inode, page->index);
2059 return VM_FAULT_LOCKED;
2060 }
2061
2062 static const struct vm_operations_struct fuse_file_vm_ops = {
2063 .close = fuse_vma_close,
2064 .fault = filemap_fault,
2065 .map_pages = filemap_map_pages,
2066 .page_mkwrite = fuse_page_mkwrite,
2067 };
2068
2069 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2070 {
2071 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2072 fuse_link_write_file(file);
2073
2074 file_accessed(file);
2075 vma->vm_ops = &fuse_file_vm_ops;
2076 return 0;
2077 }
2078
2079 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
2080 {
2081 /* Can't provide the coherency needed for MAP_SHARED */
2082 if (vma->vm_flags & VM_MAYSHARE)
2083 return -ENODEV;
2084
2085 invalidate_inode_pages2(file->f_mapping);
2086
2087 return generic_file_mmap(file, vma);
2088 }
2089
2090 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
2091 struct file_lock *fl)
2092 {
2093 switch (ffl->type) {
2094 case F_UNLCK:
2095 break;
2096
2097 case F_RDLCK:
2098 case F_WRLCK:
2099 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2100 ffl->end < ffl->start)
2101 return -EIO;
2102
2103 fl->fl_start = ffl->start;
2104 fl->fl_end = ffl->end;
2105 fl->fl_pid = ffl->pid;
2106 break;
2107
2108 default:
2109 return -EIO;
2110 }
2111 fl->fl_type = ffl->type;
2112 return 0;
2113 }
2114
2115 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2116 const struct file_lock *fl, int opcode, pid_t pid,
2117 int flock, struct fuse_lk_in *inarg)
2118 {
2119 struct inode *inode = file_inode(file);
2120 struct fuse_conn *fc = get_fuse_conn(inode);
2121 struct fuse_file *ff = file->private_data;
2122
2123 memset(inarg, 0, sizeof(*inarg));
2124 inarg->fh = ff->fh;
2125 inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2126 inarg->lk.start = fl->fl_start;
2127 inarg->lk.end = fl->fl_end;
2128 inarg->lk.type = fl->fl_type;
2129 inarg->lk.pid = pid;
2130 if (flock)
2131 inarg->lk_flags |= FUSE_LK_FLOCK;
2132 args->in.h.opcode = opcode;
2133 args->in.h.nodeid = get_node_id(inode);
2134 args->in.numargs = 1;
2135 args->in.args[0].size = sizeof(*inarg);
2136 args->in.args[0].value = inarg;
2137 }
2138
2139 static int fuse_getlk(struct file *file, struct file_lock *fl)
2140 {
2141 struct inode *inode = file_inode(file);
2142 struct fuse_conn *fc = get_fuse_conn(inode);
2143 FUSE_ARGS(args);
2144 struct fuse_lk_in inarg;
2145 struct fuse_lk_out outarg;
2146 int err;
2147
2148 fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2149 args.out.numargs = 1;
2150 args.out.args[0].size = sizeof(outarg);
2151 args.out.args[0].value = &outarg;
2152 err = fuse_simple_request(fc, &args);
2153 if (!err)
2154 err = convert_fuse_file_lock(&outarg.lk, fl);
2155
2156 return err;
2157 }
2158
2159 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2160 {
2161 struct inode *inode = file_inode(file);
2162 struct fuse_conn *fc = get_fuse_conn(inode);
2163 FUSE_ARGS(args);
2164 struct fuse_lk_in inarg;
2165 int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2166 pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
2167 int err;
2168
2169 if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2170 /* NLM needs asynchronous locks, which we don't support yet */
2171 return -ENOLCK;
2172 }
2173
2174 /* Unlock on close is handled by the flush method */
2175 if (fl->fl_flags & FL_CLOSE)
2176 return 0;
2177
2178 fuse_lk_fill(&args, file, fl, opcode, pid, flock, &inarg);
2179 err = fuse_simple_request(fc, &args);
2180
2181 /* locking is restartable */
2182 if (err == -EINTR)
2183 err = -ERESTARTSYS;
2184
2185 return err;
2186 }
2187
2188 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2189 {
2190 struct inode *inode = file_inode(file);
2191 struct fuse_conn *fc = get_fuse_conn(inode);
2192 int err;
2193
2194 if (cmd == F_CANCELLK) {
2195 err = 0;
2196 } else if (cmd == F_GETLK) {
2197 if (fc->no_lock) {
2198 posix_test_lock(file, fl);
2199 err = 0;
2200 } else
2201 err = fuse_getlk(file, fl);
2202 } else {
2203 if (fc->no_lock)
2204 err = posix_lock_file(file, fl, NULL);
2205 else
2206 err = fuse_setlk(file, fl, 0);
2207 }
2208 return err;
2209 }
2210
2211 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2212 {
2213 struct inode *inode = file_inode(file);
2214 struct fuse_conn *fc = get_fuse_conn(inode);
2215 int err;
2216
2217 if (fc->no_flock) {
2218 err = locks_lock_file_wait(file, fl);
2219 } else {
2220 struct fuse_file *ff = file->private_data;
2221
2222 /* emulate flock with POSIX locks */
2223 ff->flock = true;
2224 err = fuse_setlk(file, fl, 1);
2225 }
2226
2227 return err;
2228 }
2229
2230 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2231 {
2232 struct inode *inode = mapping->host;
2233 struct fuse_conn *fc = get_fuse_conn(inode);
2234 FUSE_ARGS(args);
2235 struct fuse_bmap_in inarg;
2236 struct fuse_bmap_out outarg;
2237 int err;
2238
2239 if (!inode->i_sb->s_bdev || fc->no_bmap)
2240 return 0;
2241
2242 memset(&inarg, 0, sizeof(inarg));
2243 inarg.block = block;
2244 inarg.blocksize = inode->i_sb->s_blocksize;
2245 args.in.h.opcode = FUSE_BMAP;
2246 args.in.h.nodeid = get_node_id(inode);
2247 args.in.numargs = 1;
2248 args.in.args[0].size = sizeof(inarg);
2249 args.in.args[0].value = &inarg;
2250 args.out.numargs = 1;
2251 args.out.args[0].size = sizeof(outarg);
2252 args.out.args[0].value = &outarg;
2253 err = fuse_simple_request(fc, &args);
2254 if (err == -ENOSYS)
2255 fc->no_bmap = 1;
2256
2257 return err ? 0 : outarg.block;
2258 }
2259
2260 static loff_t fuse_lseek(struct file *file, loff_t offset, int whence)
2261 {
2262 struct inode *inode = file->f_mapping->host;
2263 struct fuse_conn *fc = get_fuse_conn(inode);
2264 struct fuse_file *ff = file->private_data;
2265 FUSE_ARGS(args);
2266 struct fuse_lseek_in inarg = {
2267 .fh = ff->fh,
2268 .offset = offset,
2269 .whence = whence
2270 };
2271 struct fuse_lseek_out outarg;
2272 int err;
2273
2274 if (fc->no_lseek)
2275 goto fallback;
2276
2277 args.in.h.opcode = FUSE_LSEEK;
2278 args.in.h.nodeid = ff->nodeid;
2279 args.in.numargs = 1;
2280 args.in.args[0].size = sizeof(inarg);
2281 args.in.args[0].value = &inarg;
2282 args.out.numargs = 1;
2283 args.out.args[0].size = sizeof(outarg);
2284 args.out.args[0].value = &outarg;
2285 err = fuse_simple_request(fc, &args);
2286 if (err) {
2287 if (err == -ENOSYS) {
2288 fc->no_lseek = 1;
2289 goto fallback;
2290 }
2291 return err;
2292 }
2293
2294 return vfs_setpos(file, outarg.offset, inode->i_sb->s_maxbytes);
2295
2296 fallback:
2297 err = fuse_update_attributes(inode, NULL, file, NULL);
2298 if (!err)
2299 return generic_file_llseek(file, offset, whence);
2300 else
2301 return err;
2302 }
2303
2304 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2305 {
2306 loff_t retval;
2307 struct inode *inode = file_inode(file);
2308
2309 switch (whence) {
2310 case SEEK_SET:
2311 case SEEK_CUR:
2312 /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2313 retval = generic_file_llseek(file, offset, whence);
2314 break;
2315 case SEEK_END:
2316 inode_lock(inode);
2317 retval = fuse_update_attributes(inode, NULL, file, NULL);
2318 if (!retval)
2319 retval = generic_file_llseek(file, offset, whence);
2320 inode_unlock(inode);
2321 break;
2322 case SEEK_HOLE:
2323 case SEEK_DATA:
2324 inode_lock(inode);
2325 retval = fuse_lseek(file, offset, whence);
2326 inode_unlock(inode);
2327 break;
2328 default:
2329 retval = -EINVAL;
2330 }
2331
2332 return retval;
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, c;
2487 int err, i;
2488 struct iov_iter ii;
2489
2490 #if BITS_PER_LONG == 32
2491 inarg.flags |= FUSE_IOCTL_32BIT;
2492 #else
2493 if (flags & FUSE_IOCTL_COMPAT)
2494 inarg.flags |= FUSE_IOCTL_32BIT;
2495 #endif
2496
2497 /* assume all the iovs returned by client always fits in a page */
2498 BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2499
2500 err = -ENOMEM;
2501 pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
2502 iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2503 if (!pages || !iov_page)
2504 goto out;
2505
2506 /*
2507 * If restricted, initialize IO parameters as encoded in @cmd.
2508 * RETRY from server is not allowed.
2509 */
2510 if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2511 struct iovec *iov = iov_page;
2512
2513 iov->iov_base = (void __user *)arg;
2514 iov->iov_len = _IOC_SIZE(cmd);
2515
2516 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2517 in_iov = iov;
2518 in_iovs = 1;
2519 }
2520
2521 if (_IOC_DIR(cmd) & _IOC_READ) {
2522 out_iov = iov;
2523 out_iovs = 1;
2524 }
2525 }
2526
2527 retry:
2528 inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2529 inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2530
2531 /*
2532 * Out data can be used either for actual out data or iovs,
2533 * make sure there always is at least one page.
2534 */
2535 out_size = max_t(size_t, out_size, PAGE_SIZE);
2536 max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2537
2538 /* make sure there are enough buffer pages and init request with them */
2539 err = -ENOMEM;
2540 if (max_pages > FUSE_MAX_PAGES_PER_REQ)
2541 goto out;
2542 while (num_pages < max_pages) {
2543 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2544 if (!pages[num_pages])
2545 goto out;
2546 num_pages++;
2547 }
2548
2549 req = fuse_get_req(fc, num_pages);
2550 if (IS_ERR(req)) {
2551 err = PTR_ERR(req);
2552 req = NULL;
2553 goto out;
2554 }
2555 memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
2556 req->num_pages = num_pages;
2557 fuse_page_descs_length_init(req, 0, req->num_pages);
2558
2559 /* okay, let's send it to the client */
2560 req->in.h.opcode = FUSE_IOCTL;
2561 req->in.h.nodeid = ff->nodeid;
2562 req->in.numargs = 1;
2563 req->in.args[0].size = sizeof(inarg);
2564 req->in.args[0].value = &inarg;
2565 if (in_size) {
2566 req->in.numargs++;
2567 req->in.args[1].size = in_size;
2568 req->in.argpages = 1;
2569
2570 err = -EFAULT;
2571 iov_iter_init(&ii, WRITE, in_iov, in_iovs, in_size);
2572 for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= num_pages); i++) {
2573 c = copy_page_from_iter(pages[i], 0, PAGE_SIZE, &ii);
2574 if (c != PAGE_SIZE && iov_iter_count(&ii))
2575 goto out;
2576 }
2577 }
2578
2579 req->out.numargs = 2;
2580 req->out.args[0].size = sizeof(outarg);
2581 req->out.args[0].value = &outarg;
2582 req->out.args[1].size = out_size;
2583 req->out.argpages = 1;
2584 req->out.argvar = 1;
2585
2586 fuse_request_send(fc, req);
2587 err = req->out.h.error;
2588 transferred = req->out.args[1].size;
2589 fuse_put_request(fc, req);
2590 req = NULL;
2591 if (err)
2592 goto out;
2593
2594 /* did it ask for retry? */
2595 if (outarg.flags & FUSE_IOCTL_RETRY) {
2596 void *vaddr;
2597
2598 /* no retry if in restricted mode */
2599 err = -EIO;
2600 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2601 goto out;
2602
2603 in_iovs = outarg.in_iovs;
2604 out_iovs = outarg.out_iovs;
2605
2606 /*
2607 * Make sure things are in boundary, separate checks
2608 * are to protect against overflow.
2609 */
2610 err = -ENOMEM;
2611 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2612 out_iovs > FUSE_IOCTL_MAX_IOV ||
2613 in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2614 goto out;
2615
2616 vaddr = kmap_atomic(pages[0]);
2617 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2618 transferred, in_iovs + out_iovs,
2619 (flags & FUSE_IOCTL_COMPAT) != 0);
2620 kunmap_atomic(vaddr);
2621 if (err)
2622 goto out;
2623
2624 in_iov = iov_page;
2625 out_iov = in_iov + in_iovs;
2626
2627 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
2628 if (err)
2629 goto out;
2630
2631 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
2632 if (err)
2633 goto out;
2634
2635 goto retry;
2636 }
2637
2638 err = -EIO;
2639 if (transferred > inarg.out_size)
2640 goto out;
2641
2642 err = -EFAULT;
2643 iov_iter_init(&ii, READ, out_iov, out_iovs, transferred);
2644 for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= num_pages); i++) {
2645 c = copy_page_to_iter(pages[i], 0, PAGE_SIZE, &ii);
2646 if (c != PAGE_SIZE && iov_iter_count(&ii))
2647 goto out;
2648 }
2649 err = 0;
2650 out:
2651 if (req)
2652 fuse_put_request(fc, req);
2653 free_page((unsigned long) iov_page);
2654 while (num_pages)
2655 __free_page(pages[--num_pages]);
2656 kfree(pages);
2657
2658 return err ? err : outarg.result;
2659 }
2660 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2661
2662 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2663 unsigned long arg, unsigned int flags)
2664 {
2665 struct inode *inode = file_inode(file);
2666 struct fuse_conn *fc = get_fuse_conn(inode);
2667
2668 if (!fuse_allow_current_process(fc))
2669 return -EACCES;
2670
2671 if (is_bad_inode(inode))
2672 return -EIO;
2673
2674 return fuse_do_ioctl(file, cmd, arg, flags);
2675 }
2676
2677 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2678 unsigned long arg)
2679 {
2680 return fuse_ioctl_common(file, cmd, arg, 0);
2681 }
2682
2683 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2684 unsigned long arg)
2685 {
2686 return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2687 }
2688
2689 /*
2690 * All files which have been polled are linked to RB tree
2691 * fuse_conn->polled_files which is indexed by kh. Walk the tree and
2692 * find the matching one.
2693 */
2694 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2695 struct rb_node **parent_out)
2696 {
2697 struct rb_node **link = &fc->polled_files.rb_node;
2698 struct rb_node *last = NULL;
2699
2700 while (*link) {
2701 struct fuse_file *ff;
2702
2703 last = *link;
2704 ff = rb_entry(last, struct fuse_file, polled_node);
2705
2706 if (kh < ff->kh)
2707 link = &last->rb_left;
2708 else if (kh > ff->kh)
2709 link = &last->rb_right;
2710 else
2711 return link;
2712 }
2713
2714 if (parent_out)
2715 *parent_out = last;
2716 return link;
2717 }
2718
2719 /*
2720 * The file is about to be polled. Make sure it's on the polled_files
2721 * RB tree. Note that files once added to the polled_files tree are
2722 * not removed before the file is released. This is because a file
2723 * polled once is likely to be polled again.
2724 */
2725 static void fuse_register_polled_file(struct fuse_conn *fc,
2726 struct fuse_file *ff)
2727 {
2728 spin_lock(&fc->lock);
2729 if (RB_EMPTY_NODE(&ff->polled_node)) {
2730 struct rb_node **link, *uninitialized_var(parent);
2731
2732 link = fuse_find_polled_node(fc, ff->kh, &parent);
2733 BUG_ON(*link);
2734 rb_link_node(&ff->polled_node, parent, link);
2735 rb_insert_color(&ff->polled_node, &fc->polled_files);
2736 }
2737 spin_unlock(&fc->lock);
2738 }
2739
2740 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2741 {
2742 struct fuse_file *ff = file->private_data;
2743 struct fuse_conn *fc = ff->fc;
2744 struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2745 struct fuse_poll_out outarg;
2746 FUSE_ARGS(args);
2747 int err;
2748
2749 if (fc->no_poll)
2750 return DEFAULT_POLLMASK;
2751
2752 poll_wait(file, &ff->poll_wait, wait);
2753 inarg.events = (__u32)poll_requested_events(wait);
2754
2755 /*
2756 * Ask for notification iff there's someone waiting for it.
2757 * The client may ignore the flag and always notify.
2758 */
2759 if (waitqueue_active(&ff->poll_wait)) {
2760 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2761 fuse_register_polled_file(fc, ff);
2762 }
2763
2764 args.in.h.opcode = FUSE_POLL;
2765 args.in.h.nodeid = ff->nodeid;
2766 args.in.numargs = 1;
2767 args.in.args[0].size = sizeof(inarg);
2768 args.in.args[0].value = &inarg;
2769 args.out.numargs = 1;
2770 args.out.args[0].size = sizeof(outarg);
2771 args.out.args[0].value = &outarg;
2772 err = fuse_simple_request(fc, &args);
2773
2774 if (!err)
2775 return outarg.revents;
2776 if (err == -ENOSYS) {
2777 fc->no_poll = 1;
2778 return DEFAULT_POLLMASK;
2779 }
2780 return POLLERR;
2781 }
2782 EXPORT_SYMBOL_GPL(fuse_file_poll);
2783
2784 /*
2785 * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2786 * wakes up the poll waiters.
2787 */
2788 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2789 struct fuse_notify_poll_wakeup_out *outarg)
2790 {
2791 u64 kh = outarg->kh;
2792 struct rb_node **link;
2793
2794 spin_lock(&fc->lock);
2795
2796 link = fuse_find_polled_node(fc, kh, NULL);
2797 if (*link) {
2798 struct fuse_file *ff;
2799
2800 ff = rb_entry(*link, struct fuse_file, polled_node);
2801 wake_up_interruptible_sync(&ff->poll_wait);
2802 }
2803
2804 spin_unlock(&fc->lock);
2805 return 0;
2806 }
2807
2808 static void fuse_do_truncate(struct file *file)
2809 {
2810 struct inode *inode = file->f_mapping->host;
2811 struct iattr attr;
2812
2813 attr.ia_valid = ATTR_SIZE;
2814 attr.ia_size = i_size_read(inode);
2815
2816 attr.ia_file = file;
2817 attr.ia_valid |= ATTR_FILE;
2818
2819 fuse_do_setattr(file_dentry(file), &attr, file);
2820 }
2821
2822 static inline loff_t fuse_round_up(loff_t off)
2823 {
2824 return round_up(off, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
2825 }
2826
2827 static ssize_t
2828 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
2829 {
2830 DECLARE_COMPLETION_ONSTACK(wait);
2831 ssize_t ret = 0;
2832 struct file *file = iocb->ki_filp;
2833 struct fuse_file *ff = file->private_data;
2834 bool async_dio = ff->fc->async_dio;
2835 loff_t pos = 0;
2836 struct inode *inode;
2837 loff_t i_size;
2838 size_t count = iov_iter_count(iter);
2839 loff_t offset = iocb->ki_pos;
2840 struct fuse_io_priv *io;
2841
2842 pos = offset;
2843 inode = file->f_mapping->host;
2844 i_size = i_size_read(inode);
2845
2846 if ((iov_iter_rw(iter) == READ) && (offset > i_size))
2847 return 0;
2848
2849 /* optimization for short read */
2850 if (async_dio && iov_iter_rw(iter) != WRITE && offset + count > i_size) {
2851 if (offset >= i_size)
2852 return 0;
2853 iov_iter_truncate(iter, fuse_round_up(i_size - offset));
2854 count = iov_iter_count(iter);
2855 }
2856
2857 io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2858 if (!io)
2859 return -ENOMEM;
2860 spin_lock_init(&io->lock);
2861 kref_init(&io->refcnt);
2862 io->reqs = 1;
2863 io->bytes = -1;
2864 io->size = 0;
2865 io->offset = offset;
2866 io->write = (iov_iter_rw(iter) == WRITE);
2867 io->err = 0;
2868 io->file = file;
2869 /*
2870 * By default, we want to optimize all I/Os with async request
2871 * submission to the client filesystem if supported.
2872 */
2873 io->async = async_dio;
2874 io->iocb = iocb;
2875 io->blocking = is_sync_kiocb(iocb);
2876
2877 /*
2878 * We cannot asynchronously extend the size of a file.
2879 * In such case the aio will behave exactly like sync io.
2880 */
2881 if ((offset + count > i_size) && iov_iter_rw(iter) == WRITE)
2882 io->blocking = true;
2883
2884 if (io->async && io->blocking) {
2885 /*
2886 * Additional reference to keep io around after
2887 * calling fuse_aio_complete()
2888 */
2889 kref_get(&io->refcnt);
2890 io->done = &wait;
2891 }
2892
2893 if (iov_iter_rw(iter) == WRITE) {
2894 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
2895 fuse_invalidate_attr(inode);
2896 } else {
2897 ret = __fuse_direct_read(io, iter, &pos);
2898 }
2899
2900 if (io->async) {
2901 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2902
2903 /* we have a non-extending, async request, so return */
2904 if (!io->blocking)
2905 return -EIOCBQUEUED;
2906
2907 wait_for_completion(&wait);
2908 ret = fuse_get_res_by_io(io);
2909 }
2910
2911 kref_put(&io->refcnt, fuse_io_release);
2912
2913 if (iov_iter_rw(iter) == WRITE) {
2914 if (ret > 0)
2915 fuse_write_update_size(inode, pos);
2916 else if (ret < 0 && offset + count > i_size)
2917 fuse_do_truncate(file);
2918 }
2919
2920 return ret;
2921 }
2922
2923 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2924 loff_t length)
2925 {
2926 struct fuse_file *ff = file->private_data;
2927 struct inode *inode = file_inode(file);
2928 struct fuse_inode *fi = get_fuse_inode(inode);
2929 struct fuse_conn *fc = ff->fc;
2930 FUSE_ARGS(args);
2931 struct fuse_fallocate_in inarg = {
2932 .fh = ff->fh,
2933 .offset = offset,
2934 .length = length,
2935 .mode = mode
2936 };
2937 int err;
2938 bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
2939 (mode & FALLOC_FL_PUNCH_HOLE);
2940
2941 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2942 return -EOPNOTSUPP;
2943
2944 if (fc->no_fallocate)
2945 return -EOPNOTSUPP;
2946
2947 if (lock_inode) {
2948 inode_lock(inode);
2949 if (mode & FALLOC_FL_PUNCH_HOLE) {
2950 loff_t endbyte = offset + length - 1;
2951 err = filemap_write_and_wait_range(inode->i_mapping,
2952 offset, endbyte);
2953 if (err)
2954 goto out;
2955
2956 fuse_sync_writes(inode);
2957 }
2958 }
2959
2960 if (!(mode & FALLOC_FL_KEEP_SIZE))
2961 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2962
2963 args.in.h.opcode = FUSE_FALLOCATE;
2964 args.in.h.nodeid = ff->nodeid;
2965 args.in.numargs = 1;
2966 args.in.args[0].size = sizeof(inarg);
2967 args.in.args[0].value = &inarg;
2968 err = fuse_simple_request(fc, &args);
2969 if (err == -ENOSYS) {
2970 fc->no_fallocate = 1;
2971 err = -EOPNOTSUPP;
2972 }
2973 if (err)
2974 goto out;
2975
2976 /* we could have extended the file */
2977 if (!(mode & FALLOC_FL_KEEP_SIZE)) {
2978 bool changed = fuse_write_update_size(inode, offset + length);
2979
2980 if (changed && fc->writeback_cache)
2981 file_update_time(file);
2982 }
2983
2984 if (mode & FALLOC_FL_PUNCH_HOLE)
2985 truncate_pagecache_range(inode, offset, offset + length - 1);
2986
2987 fuse_invalidate_attr(inode);
2988
2989 out:
2990 if (!(mode & FALLOC_FL_KEEP_SIZE))
2991 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2992
2993 if (lock_inode)
2994 inode_unlock(inode);
2995
2996 return err;
2997 }
2998
2999 static const struct file_operations fuse_file_operations = {
3000 .llseek = fuse_file_llseek,
3001 .read_iter = fuse_file_read_iter,
3002 .write_iter = fuse_file_write_iter,
3003 .mmap = fuse_file_mmap,
3004 .open = fuse_open,
3005 .flush = fuse_flush,
3006 .release = fuse_release,
3007 .fsync = fuse_fsync,
3008 .lock = fuse_file_lock,
3009 .flock = fuse_file_flock,
3010 .splice_read = generic_file_splice_read,
3011 .unlocked_ioctl = fuse_file_ioctl,
3012 .compat_ioctl = fuse_file_compat_ioctl,
3013 .poll = fuse_file_poll,
3014 .fallocate = fuse_file_fallocate,
3015 };
3016
3017 static const struct file_operations fuse_direct_io_file_operations = {
3018 .llseek = fuse_file_llseek,
3019 .read_iter = fuse_direct_read_iter,
3020 .write_iter = fuse_direct_write_iter,
3021 .mmap = fuse_direct_mmap,
3022 .open = fuse_open,
3023 .flush = fuse_flush,
3024 .release = fuse_release,
3025 .fsync = fuse_fsync,
3026 .lock = fuse_file_lock,
3027 .flock = fuse_file_flock,
3028 .unlocked_ioctl = fuse_file_ioctl,
3029 .compat_ioctl = fuse_file_compat_ioctl,
3030 .poll = fuse_file_poll,
3031 .fallocate = fuse_file_fallocate,
3032 /* no splice_read */
3033 };
3034
3035 static const struct address_space_operations fuse_file_aops = {
3036 .readpage = fuse_readpage,
3037 .writepage = fuse_writepage,
3038 .writepages = fuse_writepages,
3039 .launder_page = fuse_launder_page,
3040 .readpages = fuse_readpages,
3041 .set_page_dirty = __set_page_dirty_nobuffers,
3042 .bmap = fuse_bmap,
3043 .direct_IO = fuse_direct_IO,
3044 .write_begin = fuse_write_begin,
3045 .write_end = fuse_write_end,
3046 };
3047
3048 void fuse_init_file_inode(struct inode *inode)
3049 {
3050 inode->i_fop = &fuse_file_operations;
3051 inode->i_data.a_ops = &fuse_file_aops;
3052 }