Loading...
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
21static const struct file_operations fuse_direct_io_file_operations;
22
23static 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
45struct 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
72void fuse_file_free(struct fuse_file *ff)
73{
74 fuse_request_free(ff->reserved_req);
75 kfree(ff);
76}
77
78struct fuse_file *fuse_file_get(struct fuse_file *ff)
79{
80 atomic_inc(&ff->count);
81 return ff;
82}
83
84static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
85{
86 iput(req->misc.release.inode);
87}
88
89static 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
116int 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}
153EXPORT_SYMBOL_GPL(fuse_do_open);
154
155static 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
171void 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
197int 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
223static 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
246void 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
279static int fuse_open(struct inode *inode, struct file *file)
280{
281 return fuse_open_common(inode, file, false);
282}
283
284static 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
298void 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}
308EXPORT_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 */
314u64 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 */
338static 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
363static 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 */
374static 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 */
391static void fuse_sync_writes(struct inode *inode)
392{
393 fuse_set_nowrite(inode);
394 fuse_release_nowrite(inode);
395}
396
397static 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 req = fuse_get_req_nofail_nopages(fc, file);
421 memset(&inarg, 0, sizeof(inarg));
422 inarg.fh = ff->fh;
423 inarg.lock_owner = fuse_lock_owner_id(fc, id);
424 req->in.h.opcode = FUSE_FLUSH;
425 req->in.h.nodeid = get_node_id(inode);
426 req->in.numargs = 1;
427 req->in.args[0].size = sizeof(inarg);
428 req->in.args[0].value = &inarg;
429 __set_bit(FR_FORCE, &req->flags);
430 fuse_request_send(fc, req);
431 err = req->out.h.error;
432 fuse_put_request(fc, req);
433 if (err == -ENOSYS) {
434 fc->no_flush = 1;
435 err = 0;
436 }
437 return err;
438}
439
440int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
441 int datasync, int isdir)
442{
443 struct inode *inode = file->f_mapping->host;
444 struct fuse_conn *fc = get_fuse_conn(inode);
445 struct fuse_file *ff = file->private_data;
446 FUSE_ARGS(args);
447 struct fuse_fsync_in inarg;
448 int err;
449
450 if (is_bad_inode(inode))
451 return -EIO;
452
453 inode_lock(inode);
454
455 /*
456 * Start writeback against all dirty pages of the inode, then
457 * wait for all outstanding writes, before sending the FSYNC
458 * request.
459 */
460 err = filemap_write_and_wait_range(inode->i_mapping, start, end);
461 if (err)
462 goto out;
463
464 fuse_sync_writes(inode);
465 err = sync_inode_metadata(inode, 1);
466 if (err)
467 goto out;
468
469 if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
470 goto out;
471
472 memset(&inarg, 0, sizeof(inarg));
473 inarg.fh = ff->fh;
474 inarg.fsync_flags = datasync ? 1 : 0;
475 args.in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
476 args.in.h.nodeid = get_node_id(inode);
477 args.in.numargs = 1;
478 args.in.args[0].size = sizeof(inarg);
479 args.in.args[0].value = &inarg;
480 err = fuse_simple_request(fc, &args);
481 if (err == -ENOSYS) {
482 if (isdir)
483 fc->no_fsyncdir = 1;
484 else
485 fc->no_fsync = 1;
486 err = 0;
487 }
488out:
489 inode_unlock(inode);
490 return err;
491}
492
493static int fuse_fsync(struct file *file, loff_t start, loff_t end,
494 int datasync)
495{
496 return fuse_fsync_common(file, start, end, datasync, 0);
497}
498
499void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
500 size_t count, int opcode)
501{
502 struct fuse_read_in *inarg = &req->misc.read.in;
503 struct fuse_file *ff = file->private_data;
504
505 inarg->fh = ff->fh;
506 inarg->offset = pos;
507 inarg->size = count;
508 inarg->flags = file->f_flags;
509 req->in.h.opcode = opcode;
510 req->in.h.nodeid = ff->nodeid;
511 req->in.numargs = 1;
512 req->in.args[0].size = sizeof(struct fuse_read_in);
513 req->in.args[0].value = inarg;
514 req->out.argvar = 1;
515 req->out.numargs = 1;
516 req->out.args[0].size = count;
517}
518
519static void fuse_release_user_pages(struct fuse_req *req, int write)
520{
521 unsigned i;
522
523 for (i = 0; i < req->num_pages; i++) {
524 struct page *page = req->pages[i];
525 if (write)
526 set_page_dirty_lock(page);
527 put_page(page);
528 }
529}
530
531static void fuse_io_release(struct kref *kref)
532{
533 kfree(container_of(kref, struct fuse_io_priv, refcnt));
534}
535
536static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
537{
538 if (io->err)
539 return io->err;
540
541 if (io->bytes >= 0 && io->write)
542 return -EIO;
543
544 return io->bytes < 0 ? io->size : io->bytes;
545}
546
547/**
548 * In case of short read, the caller sets 'pos' to the position of
549 * actual end of fuse request in IO request. Otherwise, if bytes_requested
550 * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
551 *
552 * An example:
553 * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
554 * both submitted asynchronously. The first of them was ACKed by userspace as
555 * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
556 * second request was ACKed as short, e.g. only 1K was read, resulting in
557 * pos == 33K.
558 *
559 * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
560 * will be equal to the length of the longest contiguous fragment of
561 * transferred data starting from the beginning of IO request.
562 */
563static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
564{
565 bool is_sync = is_sync_kiocb(io->iocb);
566 int left;
567
568 spin_lock(&io->lock);
569 if (err)
570 io->err = io->err ? : err;
571 else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
572 io->bytes = pos;
573
574 left = --io->reqs;
575 if (!left && is_sync)
576 complete(io->done);
577 spin_unlock(&io->lock);
578
579 if (!left && !is_sync) {
580 ssize_t res = fuse_get_res_by_io(io);
581
582 if (res >= 0) {
583 struct inode *inode = file_inode(io->iocb->ki_filp);
584 struct fuse_conn *fc = get_fuse_conn(inode);
585 struct fuse_inode *fi = get_fuse_inode(inode);
586
587 spin_lock(&fc->lock);
588 fi->attr_version = ++fc->attr_version;
589 spin_unlock(&fc->lock);
590 }
591
592 io->iocb->ki_complete(io->iocb, res, 0);
593 }
594
595 kref_put(&io->refcnt, fuse_io_release);
596}
597
598static void fuse_aio_complete_req(struct fuse_conn *fc, struct fuse_req *req)
599{
600 struct fuse_io_priv *io = req->io;
601 ssize_t pos = -1;
602
603 fuse_release_user_pages(req, !io->write);
604
605 if (io->write) {
606 if (req->misc.write.in.size != req->misc.write.out.size)
607 pos = req->misc.write.in.offset - io->offset +
608 req->misc.write.out.size;
609 } else {
610 if (req->misc.read.in.size != req->out.args[0].size)
611 pos = req->misc.read.in.offset - io->offset +
612 req->out.args[0].size;
613 }
614
615 fuse_aio_complete(io, req->out.h.error, pos);
616}
617
618static size_t fuse_async_req_send(struct fuse_conn *fc, struct fuse_req *req,
619 size_t num_bytes, struct fuse_io_priv *io)
620{
621 spin_lock(&io->lock);
622 kref_get(&io->refcnt);
623 io->size += num_bytes;
624 io->reqs++;
625 spin_unlock(&io->lock);
626
627 req->io = io;
628 req->end = fuse_aio_complete_req;
629
630 __fuse_get_request(req);
631 fuse_request_send_background(fc, req);
632
633 return num_bytes;
634}
635
636static size_t fuse_send_read(struct fuse_req *req, struct fuse_io_priv *io,
637 loff_t pos, size_t count, fl_owner_t owner)
638{
639 struct file *file = io->file;
640 struct fuse_file *ff = file->private_data;
641 struct fuse_conn *fc = ff->fc;
642
643 fuse_read_fill(req, file, pos, count, FUSE_READ);
644 if (owner != NULL) {
645 struct fuse_read_in *inarg = &req->misc.read.in;
646
647 inarg->read_flags |= FUSE_READ_LOCKOWNER;
648 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
649 }
650
651 if (io->async)
652 return fuse_async_req_send(fc, req, count, io);
653
654 fuse_request_send(fc, req);
655 return req->out.args[0].size;
656}
657
658static void fuse_read_update_size(struct inode *inode, loff_t size,
659 u64 attr_ver)
660{
661 struct fuse_conn *fc = get_fuse_conn(inode);
662 struct fuse_inode *fi = get_fuse_inode(inode);
663
664 spin_lock(&fc->lock);
665 if (attr_ver == fi->attr_version && size < inode->i_size &&
666 !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
667 fi->attr_version = ++fc->attr_version;
668 i_size_write(inode, size);
669 }
670 spin_unlock(&fc->lock);
671}
672
673static void fuse_short_read(struct fuse_req *req, struct inode *inode,
674 u64 attr_ver)
675{
676 size_t num_read = req->out.args[0].size;
677 struct fuse_conn *fc = get_fuse_conn(inode);
678
679 if (fc->writeback_cache) {
680 /*
681 * A hole in a file. Some data after the hole are in page cache,
682 * but have not reached the client fs yet. So, the hole is not
683 * present there.
684 */
685 int i;
686 int start_idx = num_read >> PAGE_SHIFT;
687 size_t off = num_read & (PAGE_SIZE - 1);
688
689 for (i = start_idx; i < req->num_pages; i++) {
690 zero_user_segment(req->pages[i], off, PAGE_SIZE);
691 off = 0;
692 }
693 } else {
694 loff_t pos = page_offset(req->pages[0]) + num_read;
695 fuse_read_update_size(inode, pos, attr_ver);
696 }
697}
698
699static int fuse_do_readpage(struct file *file, struct page *page)
700{
701 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
702 struct inode *inode = page->mapping->host;
703 struct fuse_conn *fc = get_fuse_conn(inode);
704 struct fuse_req *req;
705 size_t num_read;
706 loff_t pos = page_offset(page);
707 size_t count = PAGE_SIZE;
708 u64 attr_ver;
709 int err;
710
711 /*
712 * Page writeback can extend beyond the lifetime of the
713 * page-cache page, so make sure we read a properly synced
714 * page.
715 */
716 fuse_wait_on_page_writeback(inode, page->index);
717
718 req = fuse_get_req(fc, 1);
719 if (IS_ERR(req))
720 return PTR_ERR(req);
721
722 attr_ver = fuse_get_attr_version(fc);
723
724 req->out.page_zeroing = 1;
725 req->out.argpages = 1;
726 req->num_pages = 1;
727 req->pages[0] = page;
728 req->page_descs[0].length = count;
729 num_read = fuse_send_read(req, &io, pos, count, NULL);
730 err = req->out.h.error;
731
732 if (!err) {
733 /*
734 * Short read means EOF. If file size is larger, truncate it
735 */
736 if (num_read < count)
737 fuse_short_read(req, inode, attr_ver);
738
739 SetPageUptodate(page);
740 }
741
742 fuse_put_request(fc, req);
743
744 return err;
745}
746
747static int fuse_readpage(struct file *file, struct page *page)
748{
749 struct inode *inode = page->mapping->host;
750 int err;
751
752 err = -EIO;
753 if (is_bad_inode(inode))
754 goto out;
755
756 err = fuse_do_readpage(file, page);
757 fuse_invalidate_atime(inode);
758 out:
759 unlock_page(page);
760 return err;
761}
762
763static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
764{
765 int i;
766 size_t count = req->misc.read.in.size;
767 size_t num_read = req->out.args[0].size;
768 struct address_space *mapping = NULL;
769
770 for (i = 0; mapping == NULL && i < req->num_pages; i++)
771 mapping = req->pages[i]->mapping;
772
773 if (mapping) {
774 struct inode *inode = mapping->host;
775
776 /*
777 * Short read means EOF. If file size is larger, truncate it
778 */
779 if (!req->out.h.error && num_read < count)
780 fuse_short_read(req, inode, req->misc.read.attr_ver);
781
782 fuse_invalidate_atime(inode);
783 }
784
785 for (i = 0; i < req->num_pages; i++) {
786 struct page *page = req->pages[i];
787 if (!req->out.h.error)
788 SetPageUptodate(page);
789 else
790 SetPageError(page);
791 unlock_page(page);
792 put_page(page);
793 }
794 if (req->ff)
795 fuse_file_put(req->ff, false);
796}
797
798static void fuse_send_readpages(struct fuse_req *req, struct file *file)
799{
800 struct fuse_file *ff = file->private_data;
801 struct fuse_conn *fc = ff->fc;
802 loff_t pos = page_offset(req->pages[0]);
803 size_t count = req->num_pages << PAGE_SHIFT;
804
805 req->out.argpages = 1;
806 req->out.page_zeroing = 1;
807 req->out.page_replace = 1;
808 fuse_read_fill(req, file, pos, count, FUSE_READ);
809 req->misc.read.attr_ver = fuse_get_attr_version(fc);
810 if (fc->async_read) {
811 req->ff = fuse_file_get(ff);
812 req->end = fuse_readpages_end;
813 fuse_request_send_background(fc, req);
814 } else {
815 fuse_request_send(fc, req);
816 fuse_readpages_end(fc, req);
817 fuse_put_request(fc, req);
818 }
819}
820
821struct fuse_fill_data {
822 struct fuse_req *req;
823 struct file *file;
824 struct inode *inode;
825 unsigned nr_pages;
826};
827
828static int fuse_readpages_fill(void *_data, struct page *page)
829{
830 struct fuse_fill_data *data = _data;
831 struct fuse_req *req = data->req;
832 struct inode *inode = data->inode;
833 struct fuse_conn *fc = get_fuse_conn(inode);
834
835 fuse_wait_on_page_writeback(inode, page->index);
836
837 if (req->num_pages &&
838 (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
839 (req->num_pages + 1) * PAGE_SIZE > fc->max_read ||
840 req->pages[req->num_pages - 1]->index + 1 != page->index)) {
841 int nr_alloc = min_t(unsigned, data->nr_pages,
842 FUSE_MAX_PAGES_PER_REQ);
843 fuse_send_readpages(req, data->file);
844 if (fc->async_read)
845 req = fuse_get_req_for_background(fc, nr_alloc);
846 else
847 req = fuse_get_req(fc, nr_alloc);
848
849 data->req = req;
850 if (IS_ERR(req)) {
851 unlock_page(page);
852 return PTR_ERR(req);
853 }
854 }
855
856 if (WARN_ON(req->num_pages >= req->max_pages)) {
857 fuse_put_request(fc, req);
858 return -EIO;
859 }
860
861 get_page(page);
862 req->pages[req->num_pages] = page;
863 req->page_descs[req->num_pages].length = PAGE_SIZE;
864 req->num_pages++;
865 data->nr_pages--;
866 return 0;
867}
868
869static int fuse_readpages(struct file *file, struct address_space *mapping,
870 struct list_head *pages, unsigned nr_pages)
871{
872 struct inode *inode = mapping->host;
873 struct fuse_conn *fc = get_fuse_conn(inode);
874 struct fuse_fill_data data;
875 int err;
876 int nr_alloc = min_t(unsigned, nr_pages, FUSE_MAX_PAGES_PER_REQ);
877
878 err = -EIO;
879 if (is_bad_inode(inode))
880 goto out;
881
882 data.file = file;
883 data.inode = inode;
884 if (fc->async_read)
885 data.req = fuse_get_req_for_background(fc, nr_alloc);
886 else
887 data.req = fuse_get_req(fc, nr_alloc);
888 data.nr_pages = nr_pages;
889 err = PTR_ERR(data.req);
890 if (IS_ERR(data.req))
891 goto out;
892
893 err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
894 if (!err) {
895 if (data.req->num_pages)
896 fuse_send_readpages(data.req, file);
897 else
898 fuse_put_request(fc, data.req);
899 }
900out:
901 return err;
902}
903
904static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
905{
906 struct inode *inode = iocb->ki_filp->f_mapping->host;
907 struct fuse_conn *fc = get_fuse_conn(inode);
908
909 /*
910 * In auto invalidate mode, always update attributes on read.
911 * Otherwise, only update if we attempt to read past EOF (to ensure
912 * i_size is up to date).
913 */
914 if (fc->auto_inval_data ||
915 (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
916 int err;
917 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
918 if (err)
919 return err;
920 }
921
922 return generic_file_read_iter(iocb, to);
923}
924
925static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
926 loff_t pos, size_t count)
927{
928 struct fuse_write_in *inarg = &req->misc.write.in;
929 struct fuse_write_out *outarg = &req->misc.write.out;
930
931 inarg->fh = ff->fh;
932 inarg->offset = pos;
933 inarg->size = count;
934 req->in.h.opcode = FUSE_WRITE;
935 req->in.h.nodeid = ff->nodeid;
936 req->in.numargs = 2;
937 if (ff->fc->minor < 9)
938 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
939 else
940 req->in.args[0].size = sizeof(struct fuse_write_in);
941 req->in.args[0].value = inarg;
942 req->in.args[1].size = count;
943 req->out.numargs = 1;
944 req->out.args[0].size = sizeof(struct fuse_write_out);
945 req->out.args[0].value = outarg;
946}
947
948static size_t fuse_send_write(struct fuse_req *req, struct fuse_io_priv *io,
949 loff_t pos, size_t count, fl_owner_t owner)
950{
951 struct file *file = io->file;
952 struct fuse_file *ff = file->private_data;
953 struct fuse_conn *fc = ff->fc;
954 struct fuse_write_in *inarg = &req->misc.write.in;
955
956 fuse_write_fill(req, ff, pos, count);
957 inarg->flags = file->f_flags;
958 if (owner != NULL) {
959 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
960 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
961 }
962
963 if (io->async)
964 return fuse_async_req_send(fc, req, count, io);
965
966 fuse_request_send(fc, req);
967 return req->misc.write.out.size;
968}
969
970bool fuse_write_update_size(struct inode *inode, loff_t pos)
971{
972 struct fuse_conn *fc = get_fuse_conn(inode);
973 struct fuse_inode *fi = get_fuse_inode(inode);
974 bool ret = false;
975
976 spin_lock(&fc->lock);
977 fi->attr_version = ++fc->attr_version;
978 if (pos > inode->i_size) {
979 i_size_write(inode, pos);
980 ret = true;
981 }
982 spin_unlock(&fc->lock);
983
984 return ret;
985}
986
987static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
988 struct inode *inode, loff_t pos,
989 size_t count)
990{
991 size_t res;
992 unsigned offset;
993 unsigned i;
994 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
995
996 for (i = 0; i < req->num_pages; i++)
997 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
998
999 res = fuse_send_write(req, &io, pos, count, NULL);
1000
1001 offset = req->page_descs[0].offset;
1002 count = res;
1003 for (i = 0; i < req->num_pages; i++) {
1004 struct page *page = req->pages[i];
1005
1006 if (!req->out.h.error && !offset && count >= PAGE_SIZE)
1007 SetPageUptodate(page);
1008
1009 if (count > PAGE_SIZE - offset)
1010 count -= PAGE_SIZE - offset;
1011 else
1012 count = 0;
1013 offset = 0;
1014
1015 unlock_page(page);
1016 put_page(page);
1017 }
1018
1019 return res;
1020}
1021
1022static ssize_t fuse_fill_write_pages(struct fuse_req *req,
1023 struct address_space *mapping,
1024 struct iov_iter *ii, loff_t pos)
1025{
1026 struct fuse_conn *fc = get_fuse_conn(mapping->host);
1027 unsigned offset = pos & (PAGE_SIZE - 1);
1028 size_t count = 0;
1029 int err;
1030
1031 req->in.argpages = 1;
1032 req->page_descs[0].offset = offset;
1033
1034 do {
1035 size_t tmp;
1036 struct page *page;
1037 pgoff_t index = pos >> PAGE_SHIFT;
1038 size_t bytes = min_t(size_t, PAGE_SIZE - offset,
1039 iov_iter_count(ii));
1040
1041 bytes = min_t(size_t, bytes, fc->max_write - count);
1042
1043 again:
1044 err = -EFAULT;
1045 if (iov_iter_fault_in_readable(ii, bytes))
1046 break;
1047
1048 err = -ENOMEM;
1049 page = grab_cache_page_write_begin(mapping, index, 0);
1050 if (!page)
1051 break;
1052
1053 if (mapping_writably_mapped(mapping))
1054 flush_dcache_page(page);
1055
1056 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
1057 flush_dcache_page(page);
1058
1059 iov_iter_advance(ii, tmp);
1060 if (!tmp) {
1061 unlock_page(page);
1062 put_page(page);
1063 bytes = min(bytes, iov_iter_single_seg_count(ii));
1064 goto again;
1065 }
1066
1067 err = 0;
1068 req->pages[req->num_pages] = page;
1069 req->page_descs[req->num_pages].length = tmp;
1070 req->num_pages++;
1071
1072 count += tmp;
1073 pos += tmp;
1074 offset += tmp;
1075 if (offset == PAGE_SIZE)
1076 offset = 0;
1077
1078 if (!fc->big_writes)
1079 break;
1080 } while (iov_iter_count(ii) && count < fc->max_write &&
1081 req->num_pages < req->max_pages && offset == 0);
1082
1083 return count > 0 ? count : err;
1084}
1085
1086static inline unsigned fuse_wr_pages(loff_t pos, size_t len)
1087{
1088 return min_t(unsigned,
1089 ((pos + len - 1) >> PAGE_SHIFT) -
1090 (pos >> PAGE_SHIFT) + 1,
1091 FUSE_MAX_PAGES_PER_REQ);
1092}
1093
1094static ssize_t fuse_perform_write(struct file *file,
1095 struct address_space *mapping,
1096 struct iov_iter *ii, loff_t pos)
1097{
1098 struct inode *inode = mapping->host;
1099 struct fuse_conn *fc = get_fuse_conn(inode);
1100 struct fuse_inode *fi = get_fuse_inode(inode);
1101 int err = 0;
1102 ssize_t res = 0;
1103
1104 if (is_bad_inode(inode))
1105 return -EIO;
1106
1107 if (inode->i_size < pos + iov_iter_count(ii))
1108 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1109
1110 do {
1111 struct fuse_req *req;
1112 ssize_t count;
1113 unsigned nr_pages = fuse_wr_pages(pos, iov_iter_count(ii));
1114
1115 req = fuse_get_req(fc, nr_pages);
1116 if (IS_ERR(req)) {
1117 err = PTR_ERR(req);
1118 break;
1119 }
1120
1121 count = fuse_fill_write_pages(req, mapping, ii, pos);
1122 if (count <= 0) {
1123 err = count;
1124 } else {
1125 size_t num_written;
1126
1127 num_written = fuse_send_write_pages(req, file, inode,
1128 pos, count);
1129 err = req->out.h.error;
1130 if (!err) {
1131 res += num_written;
1132 pos += num_written;
1133
1134 /* break out of the loop on short write */
1135 if (num_written != count)
1136 err = -EIO;
1137 }
1138 }
1139 fuse_put_request(fc, req);
1140 } while (!err && iov_iter_count(ii));
1141
1142 if (res > 0)
1143 fuse_write_update_size(inode, pos);
1144
1145 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1146 fuse_invalidate_attr(inode);
1147
1148 return res > 0 ? res : err;
1149}
1150
1151static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1152{
1153 struct file *file = iocb->ki_filp;
1154 struct address_space *mapping = file->f_mapping;
1155 ssize_t written = 0;
1156 ssize_t written_buffered = 0;
1157 struct inode *inode = mapping->host;
1158 ssize_t err;
1159 loff_t endbyte = 0;
1160
1161 if (get_fuse_conn(inode)->writeback_cache) {
1162 /* Update size (EOF optimization) and mode (SUID clearing) */
1163 err = fuse_update_attributes(mapping->host, NULL, file, NULL);
1164 if (err)
1165 return err;
1166
1167 return generic_file_write_iter(iocb, from);
1168 }
1169
1170 inode_lock(inode);
1171
1172 /* We can write back this queue in page reclaim */
1173 current->backing_dev_info = inode_to_bdi(inode);
1174
1175 err = generic_write_checks(iocb, from);
1176 if (err <= 0)
1177 goto out;
1178
1179 err = file_remove_privs(file);
1180 if (err)
1181 goto out;
1182
1183 err = file_update_time(file);
1184 if (err)
1185 goto out;
1186
1187 if (iocb->ki_flags & IOCB_DIRECT) {
1188 loff_t pos = iocb->ki_pos;
1189 written = generic_file_direct_write(iocb, from, pos);
1190 if (written < 0 || !iov_iter_count(from))
1191 goto out;
1192
1193 pos += written;
1194
1195 written_buffered = fuse_perform_write(file, mapping, from, pos);
1196 if (written_buffered < 0) {
1197 err = written_buffered;
1198 goto out;
1199 }
1200 endbyte = pos + written_buffered - 1;
1201
1202 err = filemap_write_and_wait_range(file->f_mapping, pos,
1203 endbyte);
1204 if (err)
1205 goto out;
1206
1207 invalidate_mapping_pages(file->f_mapping,
1208 pos >> PAGE_SHIFT,
1209 endbyte >> PAGE_SHIFT);
1210
1211 written += written_buffered;
1212 iocb->ki_pos = pos + written_buffered;
1213 } else {
1214 written = fuse_perform_write(file, mapping, from, iocb->ki_pos);
1215 if (written >= 0)
1216 iocb->ki_pos += written;
1217 }
1218out:
1219 current->backing_dev_info = NULL;
1220 inode_unlock(inode);
1221
1222 return written ? written : err;
1223}
1224
1225static inline void fuse_page_descs_length_init(struct fuse_req *req,
1226 unsigned index, unsigned nr_pages)
1227{
1228 int i;
1229
1230 for (i = index; i < index + nr_pages; i++)
1231 req->page_descs[i].length = PAGE_SIZE -
1232 req->page_descs[i].offset;
1233}
1234
1235static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1236{
1237 return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1238}
1239
1240static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1241 size_t max_size)
1242{
1243 return min(iov_iter_single_seg_count(ii), max_size);
1244}
1245
1246static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1247 size_t *nbytesp, int write)
1248{
1249 size_t nbytes = 0; /* # bytes already packed in req */
1250 ssize_t ret = 0;
1251
1252 /* Special case for kernel I/O: can copy directly into the buffer */
1253 if (ii->type & ITER_KVEC) {
1254 unsigned long user_addr = fuse_get_user_addr(ii);
1255 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1256
1257 if (write)
1258 req->in.args[1].value = (void *) user_addr;
1259 else
1260 req->out.args[0].value = (void *) user_addr;
1261
1262 iov_iter_advance(ii, frag_size);
1263 *nbytesp = frag_size;
1264 return 0;
1265 }
1266
1267 while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1268 unsigned npages;
1269 size_t start;
1270 ret = iov_iter_get_pages(ii, &req->pages[req->num_pages],
1271 *nbytesp - nbytes,
1272 req->max_pages - req->num_pages,
1273 &start);
1274 if (ret < 0)
1275 break;
1276
1277 iov_iter_advance(ii, ret);
1278 nbytes += ret;
1279
1280 ret += start;
1281 npages = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
1282
1283 req->page_descs[req->num_pages].offset = start;
1284 fuse_page_descs_length_init(req, req->num_pages, npages);
1285
1286 req->num_pages += npages;
1287 req->page_descs[req->num_pages - 1].length -=
1288 (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1289 }
1290
1291 if (write)
1292 req->in.argpages = 1;
1293 else
1294 req->out.argpages = 1;
1295
1296 *nbytesp = nbytes;
1297
1298 return ret < 0 ? ret : 0;
1299}
1300
1301static inline int fuse_iter_npages(const struct iov_iter *ii_p)
1302{
1303 return iov_iter_npages(ii_p, FUSE_MAX_PAGES_PER_REQ);
1304}
1305
1306ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1307 loff_t *ppos, int flags)
1308{
1309 int write = flags & FUSE_DIO_WRITE;
1310 int cuse = flags & FUSE_DIO_CUSE;
1311 struct file *file = io->file;
1312 struct inode *inode = file->f_mapping->host;
1313 struct fuse_file *ff = file->private_data;
1314 struct fuse_conn *fc = ff->fc;
1315 size_t nmax = write ? fc->max_write : fc->max_read;
1316 loff_t pos = *ppos;
1317 size_t count = iov_iter_count(iter);
1318 pgoff_t idx_from = pos >> PAGE_SHIFT;
1319 pgoff_t idx_to = (pos + count - 1) >> PAGE_SHIFT;
1320 ssize_t res = 0;
1321 struct fuse_req *req;
1322 int err = 0;
1323
1324 if (io->async)
1325 req = fuse_get_req_for_background(fc, fuse_iter_npages(iter));
1326 else
1327 req = fuse_get_req(fc, fuse_iter_npages(iter));
1328 if (IS_ERR(req))
1329 return PTR_ERR(req);
1330
1331 if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1332 if (!write)
1333 inode_lock(inode);
1334 fuse_sync_writes(inode);
1335 if (!write)
1336 inode_unlock(inode);
1337 }
1338
1339 while (count) {
1340 size_t nres;
1341 fl_owner_t owner = current->files;
1342 size_t nbytes = min(count, nmax);
1343 err = fuse_get_user_pages(req, iter, &nbytes, write);
1344 if (err && !nbytes)
1345 break;
1346
1347 if (write)
1348 nres = fuse_send_write(req, io, pos, nbytes, owner);
1349 else
1350 nres = fuse_send_read(req, io, pos, nbytes, owner);
1351
1352 if (!io->async)
1353 fuse_release_user_pages(req, !write);
1354 if (req->out.h.error) {
1355 err = req->out.h.error;
1356 break;
1357 } else if (nres > nbytes) {
1358 res = 0;
1359 err = -EIO;
1360 break;
1361 }
1362 count -= nres;
1363 res += nres;
1364 pos += nres;
1365 if (nres != nbytes)
1366 break;
1367 if (count) {
1368 fuse_put_request(fc, req);
1369 if (io->async)
1370 req = fuse_get_req_for_background(fc,
1371 fuse_iter_npages(iter));
1372 else
1373 req = fuse_get_req(fc, fuse_iter_npages(iter));
1374 if (IS_ERR(req))
1375 break;
1376 }
1377 }
1378 if (!IS_ERR(req))
1379 fuse_put_request(fc, req);
1380 if (res > 0)
1381 *ppos = pos;
1382
1383 return res > 0 ? res : err;
1384}
1385EXPORT_SYMBOL_GPL(fuse_direct_io);
1386
1387static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1388 struct iov_iter *iter,
1389 loff_t *ppos)
1390{
1391 ssize_t res;
1392 struct file *file = io->file;
1393 struct inode *inode = file_inode(file);
1394
1395 if (is_bad_inode(inode))
1396 return -EIO;
1397
1398 res = fuse_direct_io(io, iter, ppos, 0);
1399
1400 fuse_invalidate_attr(inode);
1401
1402 return res;
1403}
1404
1405static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1406{
1407 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb->ki_filp);
1408 return __fuse_direct_read(&io, to, &iocb->ki_pos);
1409}
1410
1411static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1412{
1413 struct file *file = iocb->ki_filp;
1414 struct inode *inode = file_inode(file);
1415 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
1416 ssize_t res;
1417
1418 if (is_bad_inode(inode))
1419 return -EIO;
1420
1421 /* Don't allow parallel writes to the same file */
1422 inode_lock(inode);
1423 res = generic_write_checks(iocb, from);
1424 if (res > 0)
1425 res = fuse_direct_io(&io, from, &iocb->ki_pos, FUSE_DIO_WRITE);
1426 fuse_invalidate_attr(inode);
1427 if (res > 0)
1428 fuse_write_update_size(inode, iocb->ki_pos);
1429 inode_unlock(inode);
1430
1431 return res;
1432}
1433
1434static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1435{
1436 int i;
1437
1438 for (i = 0; i < req->num_pages; i++)
1439 __free_page(req->pages[i]);
1440
1441 if (req->ff)
1442 fuse_file_put(req->ff, false);
1443}
1444
1445static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1446{
1447 struct inode *inode = req->inode;
1448 struct fuse_inode *fi = get_fuse_inode(inode);
1449 struct backing_dev_info *bdi = inode_to_bdi(inode);
1450 int i;
1451
1452 list_del(&req->writepages_entry);
1453 for (i = 0; i < req->num_pages; i++) {
1454 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1455 dec_zone_page_state(req->pages[i], NR_WRITEBACK_TEMP);
1456 wb_writeout_inc(&bdi->wb);
1457 }
1458 wake_up(&fi->page_waitq);
1459}
1460
1461/* Called under fc->lock, may release and reacquire it */
1462static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req,
1463 loff_t size)
1464__releases(fc->lock)
1465__acquires(fc->lock)
1466{
1467 struct fuse_inode *fi = get_fuse_inode(req->inode);
1468 struct fuse_write_in *inarg = &req->misc.write.in;
1469 __u64 data_size = req->num_pages * PAGE_SIZE;
1470
1471 if (!fc->connected)
1472 goto out_free;
1473
1474 if (inarg->offset + data_size <= size) {
1475 inarg->size = data_size;
1476 } else if (inarg->offset < size) {
1477 inarg->size = size - inarg->offset;
1478 } else {
1479 /* Got truncated off completely */
1480 goto out_free;
1481 }
1482
1483 req->in.args[1].size = inarg->size;
1484 fi->writectr++;
1485 fuse_request_send_background_locked(fc, req);
1486 return;
1487
1488 out_free:
1489 fuse_writepage_finish(fc, req);
1490 spin_unlock(&fc->lock);
1491 fuse_writepage_free(fc, req);
1492 fuse_put_request(fc, req);
1493 spin_lock(&fc->lock);
1494}
1495
1496/*
1497 * If fi->writectr is positive (no truncate or fsync going on) send
1498 * all queued writepage requests.
1499 *
1500 * Called with fc->lock
1501 */
1502void fuse_flush_writepages(struct inode *inode)
1503__releases(fc->lock)
1504__acquires(fc->lock)
1505{
1506 struct fuse_conn *fc = get_fuse_conn(inode);
1507 struct fuse_inode *fi = get_fuse_inode(inode);
1508 size_t crop = i_size_read(inode);
1509 struct fuse_req *req;
1510
1511 while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1512 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1513 list_del_init(&req->list);
1514 fuse_send_writepage(fc, req, crop);
1515 }
1516}
1517
1518static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1519{
1520 struct inode *inode = req->inode;
1521 struct fuse_inode *fi = get_fuse_inode(inode);
1522
1523 mapping_set_error(inode->i_mapping, req->out.h.error);
1524 spin_lock(&fc->lock);
1525 while (req->misc.write.next) {
1526 struct fuse_conn *fc = get_fuse_conn(inode);
1527 struct fuse_write_in *inarg = &req->misc.write.in;
1528 struct fuse_req *next = req->misc.write.next;
1529 req->misc.write.next = next->misc.write.next;
1530 next->misc.write.next = NULL;
1531 next->ff = fuse_file_get(req->ff);
1532 list_add(&next->writepages_entry, &fi->writepages);
1533
1534 /*
1535 * Skip fuse_flush_writepages() to make it easy to crop requests
1536 * based on primary request size.
1537 *
1538 * 1st case (trivial): there are no concurrent activities using
1539 * fuse_set/release_nowrite. Then we're on safe side because
1540 * fuse_flush_writepages() would call fuse_send_writepage()
1541 * anyway.
1542 *
1543 * 2nd case: someone called fuse_set_nowrite and it is waiting
1544 * now for completion of all in-flight requests. This happens
1545 * rarely and no more than once per page, so this should be
1546 * okay.
1547 *
1548 * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1549 * of fuse_set_nowrite..fuse_release_nowrite section. The fact
1550 * that fuse_set_nowrite returned implies that all in-flight
1551 * requests were completed along with all of their secondary
1552 * requests. Further primary requests are blocked by negative
1553 * writectr. Hence there cannot be any in-flight requests and
1554 * no invocations of fuse_writepage_end() while we're in
1555 * fuse_set_nowrite..fuse_release_nowrite section.
1556 */
1557 fuse_send_writepage(fc, next, inarg->offset + inarg->size);
1558 }
1559 fi->writectr--;
1560 fuse_writepage_finish(fc, req);
1561 spin_unlock(&fc->lock);
1562 fuse_writepage_free(fc, req);
1563}
1564
1565static struct fuse_file *__fuse_write_file_get(struct fuse_conn *fc,
1566 struct fuse_inode *fi)
1567{
1568 struct fuse_file *ff = NULL;
1569
1570 spin_lock(&fc->lock);
1571 if (!list_empty(&fi->write_files)) {
1572 ff = list_entry(fi->write_files.next, struct fuse_file,
1573 write_entry);
1574 fuse_file_get(ff);
1575 }
1576 spin_unlock(&fc->lock);
1577
1578 return ff;
1579}
1580
1581static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1582 struct fuse_inode *fi)
1583{
1584 struct fuse_file *ff = __fuse_write_file_get(fc, fi);
1585 WARN_ON(!ff);
1586 return ff;
1587}
1588
1589int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1590{
1591 struct fuse_conn *fc = get_fuse_conn(inode);
1592 struct fuse_inode *fi = get_fuse_inode(inode);
1593 struct fuse_file *ff;
1594 int err;
1595
1596 ff = __fuse_write_file_get(fc, fi);
1597 err = fuse_flush_times(inode, ff);
1598 if (ff)
1599 fuse_file_put(ff, 0);
1600
1601 return err;
1602}
1603
1604static int fuse_writepage_locked(struct page *page)
1605{
1606 struct address_space *mapping = page->mapping;
1607 struct inode *inode = mapping->host;
1608 struct fuse_conn *fc = get_fuse_conn(inode);
1609 struct fuse_inode *fi = get_fuse_inode(inode);
1610 struct fuse_req *req;
1611 struct page *tmp_page;
1612 int error = -ENOMEM;
1613
1614 set_page_writeback(page);
1615
1616 req = fuse_request_alloc_nofs(1);
1617 if (!req)
1618 goto err;
1619
1620 /* writeback always goes to bg_queue */
1621 __set_bit(FR_BACKGROUND, &req->flags);
1622 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1623 if (!tmp_page)
1624 goto err_free;
1625
1626 error = -EIO;
1627 req->ff = fuse_write_file_get(fc, fi);
1628 if (!req->ff)
1629 goto err_nofile;
1630
1631 fuse_write_fill(req, req->ff, page_offset(page), 0);
1632
1633 copy_highpage(tmp_page, page);
1634 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1635 req->misc.write.next = NULL;
1636 req->in.argpages = 1;
1637 req->num_pages = 1;
1638 req->pages[0] = tmp_page;
1639 req->page_descs[0].offset = 0;
1640 req->page_descs[0].length = PAGE_SIZE;
1641 req->end = fuse_writepage_end;
1642 req->inode = inode;
1643
1644 inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1645 inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1646
1647 spin_lock(&fc->lock);
1648 list_add(&req->writepages_entry, &fi->writepages);
1649 list_add_tail(&req->list, &fi->queued_writes);
1650 fuse_flush_writepages(inode);
1651 spin_unlock(&fc->lock);
1652
1653 end_page_writeback(page);
1654
1655 return 0;
1656
1657err_nofile:
1658 __free_page(tmp_page);
1659err_free:
1660 fuse_request_free(req);
1661err:
1662 end_page_writeback(page);
1663 return error;
1664}
1665
1666static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1667{
1668 int err;
1669
1670 if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1671 /*
1672 * ->writepages() should be called for sync() and friends. We
1673 * should only get here on direct reclaim and then we are
1674 * allowed to skip a page which is already in flight
1675 */
1676 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1677
1678 redirty_page_for_writepage(wbc, page);
1679 return 0;
1680 }
1681
1682 err = fuse_writepage_locked(page);
1683 unlock_page(page);
1684
1685 return err;
1686}
1687
1688struct fuse_fill_wb_data {
1689 struct fuse_req *req;
1690 struct fuse_file *ff;
1691 struct inode *inode;
1692 struct page **orig_pages;
1693};
1694
1695static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1696{
1697 struct fuse_req *req = data->req;
1698 struct inode *inode = data->inode;
1699 struct fuse_conn *fc = get_fuse_conn(inode);
1700 struct fuse_inode *fi = get_fuse_inode(inode);
1701 int num_pages = req->num_pages;
1702 int i;
1703
1704 req->ff = fuse_file_get(data->ff);
1705 spin_lock(&fc->lock);
1706 list_add_tail(&req->list, &fi->queued_writes);
1707 fuse_flush_writepages(inode);
1708 spin_unlock(&fc->lock);
1709
1710 for (i = 0; i < num_pages; i++)
1711 end_page_writeback(data->orig_pages[i]);
1712}
1713
1714static bool fuse_writepage_in_flight(struct fuse_req *new_req,
1715 struct page *page)
1716{
1717 struct fuse_conn *fc = get_fuse_conn(new_req->inode);
1718 struct fuse_inode *fi = get_fuse_inode(new_req->inode);
1719 struct fuse_req *tmp;
1720 struct fuse_req *old_req;
1721 bool found = false;
1722 pgoff_t curr_index;
1723
1724 BUG_ON(new_req->num_pages != 0);
1725
1726 spin_lock(&fc->lock);
1727 list_del(&new_req->writepages_entry);
1728 list_for_each_entry(old_req, &fi->writepages, writepages_entry) {
1729 BUG_ON(old_req->inode != new_req->inode);
1730 curr_index = old_req->misc.write.in.offset >> PAGE_SHIFT;
1731 if (curr_index <= page->index &&
1732 page->index < curr_index + old_req->num_pages) {
1733 found = true;
1734 break;
1735 }
1736 }
1737 if (!found) {
1738 list_add(&new_req->writepages_entry, &fi->writepages);
1739 goto out_unlock;
1740 }
1741
1742 new_req->num_pages = 1;
1743 for (tmp = old_req; tmp != NULL; tmp = tmp->misc.write.next) {
1744 BUG_ON(tmp->inode != new_req->inode);
1745 curr_index = tmp->misc.write.in.offset >> PAGE_SHIFT;
1746 if (tmp->num_pages == 1 &&
1747 curr_index == page->index) {
1748 old_req = tmp;
1749 }
1750 }
1751
1752 if (old_req->num_pages == 1 && test_bit(FR_PENDING, &old_req->flags)) {
1753 struct backing_dev_info *bdi = inode_to_bdi(page->mapping->host);
1754
1755 copy_highpage(old_req->pages[0], page);
1756 spin_unlock(&fc->lock);
1757
1758 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1759 dec_zone_page_state(page, NR_WRITEBACK_TEMP);
1760 wb_writeout_inc(&bdi->wb);
1761 fuse_writepage_free(fc, new_req);
1762 fuse_request_free(new_req);
1763 goto out;
1764 } else {
1765 new_req->misc.write.next = old_req->misc.write.next;
1766 old_req->misc.write.next = new_req;
1767 }
1768out_unlock:
1769 spin_unlock(&fc->lock);
1770out:
1771 return found;
1772}
1773
1774static int fuse_writepages_fill(struct page *page,
1775 struct writeback_control *wbc, void *_data)
1776{
1777 struct fuse_fill_wb_data *data = _data;
1778 struct fuse_req *req = data->req;
1779 struct inode *inode = data->inode;
1780 struct fuse_conn *fc = get_fuse_conn(inode);
1781 struct page *tmp_page;
1782 bool is_writeback;
1783 int err;
1784
1785 if (!data->ff) {
1786 err = -EIO;
1787 data->ff = fuse_write_file_get(fc, get_fuse_inode(inode));
1788 if (!data->ff)
1789 goto out_unlock;
1790 }
1791
1792 /*
1793 * Being under writeback is unlikely but possible. For example direct
1794 * read to an mmaped fuse file will set the page dirty twice; once when
1795 * the pages are faulted with get_user_pages(), and then after the read
1796 * completed.
1797 */
1798 is_writeback = fuse_page_is_writeback(inode, page->index);
1799
1800 if (req && req->num_pages &&
1801 (is_writeback || req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
1802 (req->num_pages + 1) * PAGE_SIZE > fc->max_write ||
1803 data->orig_pages[req->num_pages - 1]->index + 1 != page->index)) {
1804 fuse_writepages_send(data);
1805 data->req = NULL;
1806 }
1807 err = -ENOMEM;
1808 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1809 if (!tmp_page)
1810 goto out_unlock;
1811
1812 /*
1813 * The page must not be redirtied until the writeout is completed
1814 * (i.e. userspace has sent a reply to the write request). Otherwise
1815 * there could be more than one temporary page instance for each real
1816 * page.
1817 *
1818 * This is ensured by holding the page lock in page_mkwrite() while
1819 * checking fuse_page_is_writeback(). We already hold the page lock
1820 * since clear_page_dirty_for_io() and keep it held until we add the
1821 * request to the fi->writepages list and increment req->num_pages.
1822 * After this fuse_page_is_writeback() will indicate that the page is
1823 * under writeback, so we can release the page lock.
1824 */
1825 if (data->req == NULL) {
1826 struct fuse_inode *fi = get_fuse_inode(inode);
1827
1828 err = -ENOMEM;
1829 req = fuse_request_alloc_nofs(FUSE_MAX_PAGES_PER_REQ);
1830 if (!req) {
1831 __free_page(tmp_page);
1832 goto out_unlock;
1833 }
1834
1835 fuse_write_fill(req, data->ff, page_offset(page), 0);
1836 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1837 req->misc.write.next = NULL;
1838 req->in.argpages = 1;
1839 __set_bit(FR_BACKGROUND, &req->flags);
1840 req->num_pages = 0;
1841 req->end = fuse_writepage_end;
1842 req->inode = inode;
1843
1844 spin_lock(&fc->lock);
1845 list_add(&req->writepages_entry, &fi->writepages);
1846 spin_unlock(&fc->lock);
1847
1848 data->req = req;
1849 }
1850 set_page_writeback(page);
1851
1852 copy_highpage(tmp_page, page);
1853 req->pages[req->num_pages] = tmp_page;
1854 req->page_descs[req->num_pages].offset = 0;
1855 req->page_descs[req->num_pages].length = PAGE_SIZE;
1856
1857 inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1858 inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1859
1860 err = 0;
1861 if (is_writeback && fuse_writepage_in_flight(req, page)) {
1862 end_page_writeback(page);
1863 data->req = NULL;
1864 goto out_unlock;
1865 }
1866 data->orig_pages[req->num_pages] = page;
1867
1868 /*
1869 * Protected by fc->lock against concurrent access by
1870 * fuse_page_is_writeback().
1871 */
1872 spin_lock(&fc->lock);
1873 req->num_pages++;
1874 spin_unlock(&fc->lock);
1875
1876out_unlock:
1877 unlock_page(page);
1878
1879 return err;
1880}
1881
1882static int fuse_writepages(struct address_space *mapping,
1883 struct writeback_control *wbc)
1884{
1885 struct inode *inode = mapping->host;
1886 struct fuse_fill_wb_data data;
1887 int err;
1888
1889 err = -EIO;
1890 if (is_bad_inode(inode))
1891 goto out;
1892
1893 data.inode = inode;
1894 data.req = NULL;
1895 data.ff = NULL;
1896
1897 err = -ENOMEM;
1898 data.orig_pages = kcalloc(FUSE_MAX_PAGES_PER_REQ,
1899 sizeof(struct page *),
1900 GFP_NOFS);
1901 if (!data.orig_pages)
1902 goto out;
1903
1904 err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
1905 if (data.req) {
1906 /* Ignore errors if we can write at least one page */
1907 BUG_ON(!data.req->num_pages);
1908 fuse_writepages_send(&data);
1909 err = 0;
1910 }
1911 if (data.ff)
1912 fuse_file_put(data.ff, false);
1913
1914 kfree(data.orig_pages);
1915out:
1916 return err;
1917}
1918
1919/*
1920 * It's worthy to make sure that space is reserved on disk for the write,
1921 * but how to implement it without killing performance need more thinking.
1922 */
1923static int fuse_write_begin(struct file *file, struct address_space *mapping,
1924 loff_t pos, unsigned len, unsigned flags,
1925 struct page **pagep, void **fsdata)
1926{
1927 pgoff_t index = pos >> PAGE_SHIFT;
1928 struct fuse_conn *fc = get_fuse_conn(file_inode(file));
1929 struct page *page;
1930 loff_t fsize;
1931 int err = -ENOMEM;
1932
1933 WARN_ON(!fc->writeback_cache);
1934
1935 page = grab_cache_page_write_begin(mapping, index, flags);
1936 if (!page)
1937 goto error;
1938
1939 fuse_wait_on_page_writeback(mapping->host, page->index);
1940
1941 if (PageUptodate(page) || len == PAGE_SIZE)
1942 goto success;
1943 /*
1944 * Check if the start this page comes after the end of file, in which
1945 * case the readpage can be optimized away.
1946 */
1947 fsize = i_size_read(mapping->host);
1948 if (fsize <= (pos & PAGE_MASK)) {
1949 size_t off = pos & ~PAGE_MASK;
1950 if (off)
1951 zero_user_segment(page, 0, off);
1952 goto success;
1953 }
1954 err = fuse_do_readpage(file, page);
1955 if (err)
1956 goto cleanup;
1957success:
1958 *pagep = page;
1959 return 0;
1960
1961cleanup:
1962 unlock_page(page);
1963 put_page(page);
1964error:
1965 return err;
1966}
1967
1968static int fuse_write_end(struct file *file, struct address_space *mapping,
1969 loff_t pos, unsigned len, unsigned copied,
1970 struct page *page, void *fsdata)
1971{
1972 struct inode *inode = page->mapping->host;
1973
1974 if (!PageUptodate(page)) {
1975 /* Zero any unwritten bytes at the end of the page */
1976 size_t endoff = (pos + copied) & ~PAGE_MASK;
1977 if (endoff)
1978 zero_user_segment(page, endoff, PAGE_SIZE);
1979 SetPageUptodate(page);
1980 }
1981
1982 fuse_write_update_size(inode, pos + copied);
1983 set_page_dirty(page);
1984 unlock_page(page);
1985 put_page(page);
1986
1987 return copied;
1988}
1989
1990static int fuse_launder_page(struct page *page)
1991{
1992 int err = 0;
1993 if (clear_page_dirty_for_io(page)) {
1994 struct inode *inode = page->mapping->host;
1995 err = fuse_writepage_locked(page);
1996 if (!err)
1997 fuse_wait_on_page_writeback(inode, page->index);
1998 }
1999 return err;
2000}
2001
2002/*
2003 * Write back dirty pages now, because there may not be any suitable
2004 * open files later
2005 */
2006static void fuse_vma_close(struct vm_area_struct *vma)
2007{
2008 filemap_write_and_wait(vma->vm_file->f_mapping);
2009}
2010
2011/*
2012 * Wait for writeback against this page to complete before allowing it
2013 * to be marked dirty again, and hence written back again, possibly
2014 * before the previous writepage completed.
2015 *
2016 * Block here, instead of in ->writepage(), so that the userspace fs
2017 * can only block processes actually operating on the filesystem.
2018 *
2019 * Otherwise unprivileged userspace fs would be able to block
2020 * unrelated:
2021 *
2022 * - page migration
2023 * - sync(2)
2024 * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2025 */
2026static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
2027{
2028 struct page *page = vmf->page;
2029 struct inode *inode = file_inode(vma->vm_file);
2030
2031 file_update_time(vma->vm_file);
2032 lock_page(page);
2033 if (page->mapping != inode->i_mapping) {
2034 unlock_page(page);
2035 return VM_FAULT_NOPAGE;
2036 }
2037
2038 fuse_wait_on_page_writeback(inode, page->index);
2039 return VM_FAULT_LOCKED;
2040}
2041
2042static const struct vm_operations_struct fuse_file_vm_ops = {
2043 .close = fuse_vma_close,
2044 .fault = filemap_fault,
2045 .map_pages = filemap_map_pages,
2046 .page_mkwrite = fuse_page_mkwrite,
2047};
2048
2049static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2050{
2051 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2052 fuse_link_write_file(file);
2053
2054 file_accessed(file);
2055 vma->vm_ops = &fuse_file_vm_ops;
2056 return 0;
2057}
2058
2059static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
2060{
2061 /* Can't provide the coherency needed for MAP_SHARED */
2062 if (vma->vm_flags & VM_MAYSHARE)
2063 return -ENODEV;
2064
2065 invalidate_inode_pages2(file->f_mapping);
2066
2067 return generic_file_mmap(file, vma);
2068}
2069
2070static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
2071 struct file_lock *fl)
2072{
2073 switch (ffl->type) {
2074 case F_UNLCK:
2075 break;
2076
2077 case F_RDLCK:
2078 case F_WRLCK:
2079 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2080 ffl->end < ffl->start)
2081 return -EIO;
2082
2083 fl->fl_start = ffl->start;
2084 fl->fl_end = ffl->end;
2085 fl->fl_pid = ffl->pid;
2086 break;
2087
2088 default:
2089 return -EIO;
2090 }
2091 fl->fl_type = ffl->type;
2092 return 0;
2093}
2094
2095static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2096 const struct file_lock *fl, int opcode, pid_t pid,
2097 int flock, struct fuse_lk_in *inarg)
2098{
2099 struct inode *inode = file_inode(file);
2100 struct fuse_conn *fc = get_fuse_conn(inode);
2101 struct fuse_file *ff = file->private_data;
2102
2103 memset(inarg, 0, sizeof(*inarg));
2104 inarg->fh = ff->fh;
2105 inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2106 inarg->lk.start = fl->fl_start;
2107 inarg->lk.end = fl->fl_end;
2108 inarg->lk.type = fl->fl_type;
2109 inarg->lk.pid = pid;
2110 if (flock)
2111 inarg->lk_flags |= FUSE_LK_FLOCK;
2112 args->in.h.opcode = opcode;
2113 args->in.h.nodeid = get_node_id(inode);
2114 args->in.numargs = 1;
2115 args->in.args[0].size = sizeof(*inarg);
2116 args->in.args[0].value = inarg;
2117}
2118
2119static int fuse_getlk(struct file *file, struct file_lock *fl)
2120{
2121 struct inode *inode = file_inode(file);
2122 struct fuse_conn *fc = get_fuse_conn(inode);
2123 FUSE_ARGS(args);
2124 struct fuse_lk_in inarg;
2125 struct fuse_lk_out outarg;
2126 int err;
2127
2128 fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2129 args.out.numargs = 1;
2130 args.out.args[0].size = sizeof(outarg);
2131 args.out.args[0].value = &outarg;
2132 err = fuse_simple_request(fc, &args);
2133 if (!err)
2134 err = convert_fuse_file_lock(&outarg.lk, fl);
2135
2136 return err;
2137}
2138
2139static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
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 int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2146 pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
2147 int err;
2148
2149 if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2150 /* NLM needs asynchronous locks, which we don't support yet */
2151 return -ENOLCK;
2152 }
2153
2154 /* Unlock on close is handled by the flush method */
2155 if (fl->fl_flags & FL_CLOSE)
2156 return 0;
2157
2158 fuse_lk_fill(&args, file, fl, opcode, pid, flock, &inarg);
2159 err = fuse_simple_request(fc, &args);
2160
2161 /* locking is restartable */
2162 if (err == -EINTR)
2163 err = -ERESTARTSYS;
2164
2165 return err;
2166}
2167
2168static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2169{
2170 struct inode *inode = file_inode(file);
2171 struct fuse_conn *fc = get_fuse_conn(inode);
2172 int err;
2173
2174 if (cmd == F_CANCELLK) {
2175 err = 0;
2176 } else if (cmd == F_GETLK) {
2177 if (fc->no_lock) {
2178 posix_test_lock(file, fl);
2179 err = 0;
2180 } else
2181 err = fuse_getlk(file, fl);
2182 } else {
2183 if (fc->no_lock)
2184 err = posix_lock_file(file, fl, NULL);
2185 else
2186 err = fuse_setlk(file, fl, 0);
2187 }
2188 return err;
2189}
2190
2191static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2192{
2193 struct inode *inode = file_inode(file);
2194 struct fuse_conn *fc = get_fuse_conn(inode);
2195 int err;
2196
2197 if (fc->no_flock) {
2198 err = locks_lock_file_wait(file, fl);
2199 } else {
2200 struct fuse_file *ff = file->private_data;
2201
2202 /* emulate flock with POSIX locks */
2203 ff->flock = true;
2204 err = fuse_setlk(file, fl, 1);
2205 }
2206
2207 return err;
2208}
2209
2210static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2211{
2212 struct inode *inode = mapping->host;
2213 struct fuse_conn *fc = get_fuse_conn(inode);
2214 FUSE_ARGS(args);
2215 struct fuse_bmap_in inarg;
2216 struct fuse_bmap_out outarg;
2217 int err;
2218
2219 if (!inode->i_sb->s_bdev || fc->no_bmap)
2220 return 0;
2221
2222 memset(&inarg, 0, sizeof(inarg));
2223 inarg.block = block;
2224 inarg.blocksize = inode->i_sb->s_blocksize;
2225 args.in.h.opcode = FUSE_BMAP;
2226 args.in.h.nodeid = get_node_id(inode);
2227 args.in.numargs = 1;
2228 args.in.args[0].size = sizeof(inarg);
2229 args.in.args[0].value = &inarg;
2230 args.out.numargs = 1;
2231 args.out.args[0].size = sizeof(outarg);
2232 args.out.args[0].value = &outarg;
2233 err = fuse_simple_request(fc, &args);
2234 if (err == -ENOSYS)
2235 fc->no_bmap = 1;
2236
2237 return err ? 0 : outarg.block;
2238}
2239
2240static loff_t fuse_lseek(struct file *file, loff_t offset, int whence)
2241{
2242 struct inode *inode = file->f_mapping->host;
2243 struct fuse_conn *fc = get_fuse_conn(inode);
2244 struct fuse_file *ff = file->private_data;
2245 FUSE_ARGS(args);
2246 struct fuse_lseek_in inarg = {
2247 .fh = ff->fh,
2248 .offset = offset,
2249 .whence = whence
2250 };
2251 struct fuse_lseek_out outarg;
2252 int err;
2253
2254 if (fc->no_lseek)
2255 goto fallback;
2256
2257 args.in.h.opcode = FUSE_LSEEK;
2258 args.in.h.nodeid = ff->nodeid;
2259 args.in.numargs = 1;
2260 args.in.args[0].size = sizeof(inarg);
2261 args.in.args[0].value = &inarg;
2262 args.out.numargs = 1;
2263 args.out.args[0].size = sizeof(outarg);
2264 args.out.args[0].value = &outarg;
2265 err = fuse_simple_request(fc, &args);
2266 if (err) {
2267 if (err == -ENOSYS) {
2268 fc->no_lseek = 1;
2269 goto fallback;
2270 }
2271 return err;
2272 }
2273
2274 return vfs_setpos(file, outarg.offset, inode->i_sb->s_maxbytes);
2275
2276fallback:
2277 err = fuse_update_attributes(inode, NULL, file, NULL);
2278 if (!err)
2279 return generic_file_llseek(file, offset, whence);
2280 else
2281 return err;
2282}
2283
2284static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2285{
2286 loff_t retval;
2287 struct inode *inode = file_inode(file);
2288
2289 switch (whence) {
2290 case SEEK_SET:
2291 case SEEK_CUR:
2292 /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2293 retval = generic_file_llseek(file, offset, whence);
2294 break;
2295 case SEEK_END:
2296 inode_lock(inode);
2297 retval = fuse_update_attributes(inode, NULL, file, NULL);
2298 if (!retval)
2299 retval = generic_file_llseek(file, offset, whence);
2300 inode_unlock(inode);
2301 break;
2302 case SEEK_HOLE:
2303 case SEEK_DATA:
2304 inode_lock(inode);
2305 retval = fuse_lseek(file, offset, whence);
2306 inode_unlock(inode);
2307 break;
2308 default:
2309 retval = -EINVAL;
2310 }
2311
2312 return retval;
2313}
2314
2315static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
2316 unsigned int nr_segs, size_t bytes, bool to_user)
2317{
2318 struct iov_iter ii;
2319 int page_idx = 0;
2320
2321 if (!bytes)
2322 return 0;
2323
2324 iov_iter_init(&ii, to_user ? READ : WRITE, iov, nr_segs, bytes);
2325
2326 while (iov_iter_count(&ii)) {
2327 struct page *page = pages[page_idx++];
2328 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
2329 void *kaddr;
2330
2331 kaddr = kmap(page);
2332
2333 while (todo) {
2334 char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
2335 size_t iov_len = ii.iov->iov_len - ii.iov_offset;
2336 size_t copy = min(todo, iov_len);
2337 size_t left;
2338
2339 if (!to_user)
2340 left = copy_from_user(kaddr, uaddr, copy);
2341 else
2342 left = copy_to_user(uaddr, kaddr, copy);
2343
2344 if (unlikely(left))
2345 return -EFAULT;
2346
2347 iov_iter_advance(&ii, copy);
2348 todo -= copy;
2349 kaddr += copy;
2350 }
2351
2352 kunmap(page);
2353 }
2354
2355 return 0;
2356}
2357
2358/*
2359 * CUSE servers compiled on 32bit broke on 64bit kernels because the
2360 * ABI was defined to be 'struct iovec' which is different on 32bit
2361 * and 64bit. Fortunately we can determine which structure the server
2362 * used from the size of the reply.
2363 */
2364static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2365 size_t transferred, unsigned count,
2366 bool is_compat)
2367{
2368#ifdef CONFIG_COMPAT
2369 if (count * sizeof(struct compat_iovec) == transferred) {
2370 struct compat_iovec *ciov = src;
2371 unsigned i;
2372
2373 /*
2374 * With this interface a 32bit server cannot support
2375 * non-compat (i.e. ones coming from 64bit apps) ioctl
2376 * requests
2377 */
2378 if (!is_compat)
2379 return -EINVAL;
2380
2381 for (i = 0; i < count; i++) {
2382 dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2383 dst[i].iov_len = ciov[i].iov_len;
2384 }
2385 return 0;
2386 }
2387#endif
2388
2389 if (count * sizeof(struct iovec) != transferred)
2390 return -EIO;
2391
2392 memcpy(dst, src, transferred);
2393 return 0;
2394}
2395
2396/* Make sure iov_length() won't overflow */
2397static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
2398{
2399 size_t n;
2400 u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
2401
2402 for (n = 0; n < count; n++, iov++) {
2403 if (iov->iov_len > (size_t) max)
2404 return -ENOMEM;
2405 max -= iov->iov_len;
2406 }
2407 return 0;
2408}
2409
2410static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2411 void *src, size_t transferred, unsigned count,
2412 bool is_compat)
2413{
2414 unsigned i;
2415 struct fuse_ioctl_iovec *fiov = src;
2416
2417 if (fc->minor < 16) {
2418 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2419 count, is_compat);
2420 }
2421
2422 if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2423 return -EIO;
2424
2425 for (i = 0; i < count; i++) {
2426 /* Did the server supply an inappropriate value? */
2427 if (fiov[i].base != (unsigned long) fiov[i].base ||
2428 fiov[i].len != (unsigned long) fiov[i].len)
2429 return -EIO;
2430
2431 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2432 dst[i].iov_len = (size_t) fiov[i].len;
2433
2434#ifdef CONFIG_COMPAT
2435 if (is_compat &&
2436 (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2437 (compat_size_t) dst[i].iov_len != fiov[i].len))
2438 return -EIO;
2439#endif
2440 }
2441
2442 return 0;
2443}
2444
2445
2446/*
2447 * For ioctls, there is no generic way to determine how much memory
2448 * needs to be read and/or written. Furthermore, ioctls are allowed
2449 * to dereference the passed pointer, so the parameter requires deep
2450 * copying but FUSE has no idea whatsoever about what to copy in or
2451 * out.
2452 *
2453 * This is solved by allowing FUSE server to retry ioctl with
2454 * necessary in/out iovecs. Let's assume the ioctl implementation
2455 * needs to read in the following structure.
2456 *
2457 * struct a {
2458 * char *buf;
2459 * size_t buflen;
2460 * }
2461 *
2462 * On the first callout to FUSE server, inarg->in_size and
2463 * inarg->out_size will be NULL; then, the server completes the ioctl
2464 * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2465 * the actual iov array to
2466 *
2467 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) } }
2468 *
2469 * which tells FUSE to copy in the requested area and retry the ioctl.
2470 * On the second round, the server has access to the structure and
2471 * from that it can tell what to look for next, so on the invocation,
2472 * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2473 *
2474 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) },
2475 * { .iov_base = a.buf, .iov_len = a.buflen } }
2476 *
2477 * FUSE will copy both struct a and the pointed buffer from the
2478 * process doing the ioctl and retry ioctl with both struct a and the
2479 * buffer.
2480 *
2481 * This time, FUSE server has everything it needs and completes ioctl
2482 * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2483 *
2484 * Copying data out works the same way.
2485 *
2486 * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2487 * automatically initializes in and out iovs by decoding @cmd with
2488 * _IOC_* macros and the server is not allowed to request RETRY. This
2489 * limits ioctl data transfers to well-formed ioctls and is the forced
2490 * behavior for all FUSE servers.
2491 */
2492long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2493 unsigned int flags)
2494{
2495 struct fuse_file *ff = file->private_data;
2496 struct fuse_conn *fc = ff->fc;
2497 struct fuse_ioctl_in inarg = {
2498 .fh = ff->fh,
2499 .cmd = cmd,
2500 .arg = arg,
2501 .flags = flags
2502 };
2503 struct fuse_ioctl_out outarg;
2504 struct fuse_req *req = NULL;
2505 struct page **pages = NULL;
2506 struct iovec *iov_page = NULL;
2507 struct iovec *in_iov = NULL, *out_iov = NULL;
2508 unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
2509 size_t in_size, out_size, transferred;
2510 int err;
2511
2512#if BITS_PER_LONG == 32
2513 inarg.flags |= FUSE_IOCTL_32BIT;
2514#else
2515 if (flags & FUSE_IOCTL_COMPAT)
2516 inarg.flags |= FUSE_IOCTL_32BIT;
2517#endif
2518
2519 /* assume all the iovs returned by client always fits in a page */
2520 BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2521
2522 err = -ENOMEM;
2523 pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
2524 iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2525 if (!pages || !iov_page)
2526 goto out;
2527
2528 /*
2529 * If restricted, initialize IO parameters as encoded in @cmd.
2530 * RETRY from server is not allowed.
2531 */
2532 if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2533 struct iovec *iov = iov_page;
2534
2535 iov->iov_base = (void __user *)arg;
2536 iov->iov_len = _IOC_SIZE(cmd);
2537
2538 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2539 in_iov = iov;
2540 in_iovs = 1;
2541 }
2542
2543 if (_IOC_DIR(cmd) & _IOC_READ) {
2544 out_iov = iov;
2545 out_iovs = 1;
2546 }
2547 }
2548
2549 retry:
2550 inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2551 inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2552
2553 /*
2554 * Out data can be used either for actual out data or iovs,
2555 * make sure there always is at least one page.
2556 */
2557 out_size = max_t(size_t, out_size, PAGE_SIZE);
2558 max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2559
2560 /* make sure there are enough buffer pages and init request with them */
2561 err = -ENOMEM;
2562 if (max_pages > FUSE_MAX_PAGES_PER_REQ)
2563 goto out;
2564 while (num_pages < max_pages) {
2565 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2566 if (!pages[num_pages])
2567 goto out;
2568 num_pages++;
2569 }
2570
2571 req = fuse_get_req(fc, num_pages);
2572 if (IS_ERR(req)) {
2573 err = PTR_ERR(req);
2574 req = NULL;
2575 goto out;
2576 }
2577 memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
2578 req->num_pages = num_pages;
2579 fuse_page_descs_length_init(req, 0, req->num_pages);
2580
2581 /* okay, let's send it to the client */
2582 req->in.h.opcode = FUSE_IOCTL;
2583 req->in.h.nodeid = ff->nodeid;
2584 req->in.numargs = 1;
2585 req->in.args[0].size = sizeof(inarg);
2586 req->in.args[0].value = &inarg;
2587 if (in_size) {
2588 req->in.numargs++;
2589 req->in.args[1].size = in_size;
2590 req->in.argpages = 1;
2591
2592 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
2593 false);
2594 if (err)
2595 goto out;
2596 }
2597
2598 req->out.numargs = 2;
2599 req->out.args[0].size = sizeof(outarg);
2600 req->out.args[0].value = &outarg;
2601 req->out.args[1].size = out_size;
2602 req->out.argpages = 1;
2603 req->out.argvar = 1;
2604
2605 fuse_request_send(fc, req);
2606 err = req->out.h.error;
2607 transferred = req->out.args[1].size;
2608 fuse_put_request(fc, req);
2609 req = NULL;
2610 if (err)
2611 goto out;
2612
2613 /* did it ask for retry? */
2614 if (outarg.flags & FUSE_IOCTL_RETRY) {
2615 void *vaddr;
2616
2617 /* no retry if in restricted mode */
2618 err = -EIO;
2619 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2620 goto out;
2621
2622 in_iovs = outarg.in_iovs;
2623 out_iovs = outarg.out_iovs;
2624
2625 /*
2626 * Make sure things are in boundary, separate checks
2627 * are to protect against overflow.
2628 */
2629 err = -ENOMEM;
2630 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2631 out_iovs > FUSE_IOCTL_MAX_IOV ||
2632 in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2633 goto out;
2634
2635 vaddr = kmap_atomic(pages[0]);
2636 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2637 transferred, in_iovs + out_iovs,
2638 (flags & FUSE_IOCTL_COMPAT) != 0);
2639 kunmap_atomic(vaddr);
2640 if (err)
2641 goto out;
2642
2643 in_iov = iov_page;
2644 out_iov = in_iov + in_iovs;
2645
2646 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
2647 if (err)
2648 goto out;
2649
2650 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
2651 if (err)
2652 goto out;
2653
2654 goto retry;
2655 }
2656
2657 err = -EIO;
2658 if (transferred > inarg.out_size)
2659 goto out;
2660
2661 err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
2662 out:
2663 if (req)
2664 fuse_put_request(fc, req);
2665 free_page((unsigned long) iov_page);
2666 while (num_pages)
2667 __free_page(pages[--num_pages]);
2668 kfree(pages);
2669
2670 return err ? err : outarg.result;
2671}
2672EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2673
2674long fuse_ioctl_common(struct file *file, unsigned int cmd,
2675 unsigned long arg, unsigned int flags)
2676{
2677 struct inode *inode = file_inode(file);
2678 struct fuse_conn *fc = get_fuse_conn(inode);
2679
2680 if (!fuse_allow_current_process(fc))
2681 return -EACCES;
2682
2683 if (is_bad_inode(inode))
2684 return -EIO;
2685
2686 return fuse_do_ioctl(file, cmd, arg, flags);
2687}
2688
2689static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2690 unsigned long arg)
2691{
2692 return fuse_ioctl_common(file, cmd, arg, 0);
2693}
2694
2695static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2696 unsigned long arg)
2697{
2698 return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2699}
2700
2701/*
2702 * All files which have been polled are linked to RB tree
2703 * fuse_conn->polled_files which is indexed by kh. Walk the tree and
2704 * find the matching one.
2705 */
2706static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2707 struct rb_node **parent_out)
2708{
2709 struct rb_node **link = &fc->polled_files.rb_node;
2710 struct rb_node *last = NULL;
2711
2712 while (*link) {
2713 struct fuse_file *ff;
2714
2715 last = *link;
2716 ff = rb_entry(last, struct fuse_file, polled_node);
2717
2718 if (kh < ff->kh)
2719 link = &last->rb_left;
2720 else if (kh > ff->kh)
2721 link = &last->rb_right;
2722 else
2723 return link;
2724 }
2725
2726 if (parent_out)
2727 *parent_out = last;
2728 return link;
2729}
2730
2731/*
2732 * The file is about to be polled. Make sure it's on the polled_files
2733 * RB tree. Note that files once added to the polled_files tree are
2734 * not removed before the file is released. This is because a file
2735 * polled once is likely to be polled again.
2736 */
2737static void fuse_register_polled_file(struct fuse_conn *fc,
2738 struct fuse_file *ff)
2739{
2740 spin_lock(&fc->lock);
2741 if (RB_EMPTY_NODE(&ff->polled_node)) {
2742 struct rb_node **link, *uninitialized_var(parent);
2743
2744 link = fuse_find_polled_node(fc, ff->kh, &parent);
2745 BUG_ON(*link);
2746 rb_link_node(&ff->polled_node, parent, link);
2747 rb_insert_color(&ff->polled_node, &fc->polled_files);
2748 }
2749 spin_unlock(&fc->lock);
2750}
2751
2752unsigned fuse_file_poll(struct file *file, poll_table *wait)
2753{
2754 struct fuse_file *ff = file->private_data;
2755 struct fuse_conn *fc = ff->fc;
2756 struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2757 struct fuse_poll_out outarg;
2758 FUSE_ARGS(args);
2759 int err;
2760
2761 if (fc->no_poll)
2762 return DEFAULT_POLLMASK;
2763
2764 poll_wait(file, &ff->poll_wait, wait);
2765 inarg.events = (__u32)poll_requested_events(wait);
2766
2767 /*
2768 * Ask for notification iff there's someone waiting for it.
2769 * The client may ignore the flag and always notify.
2770 */
2771 if (waitqueue_active(&ff->poll_wait)) {
2772 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2773 fuse_register_polled_file(fc, ff);
2774 }
2775
2776 args.in.h.opcode = FUSE_POLL;
2777 args.in.h.nodeid = ff->nodeid;
2778 args.in.numargs = 1;
2779 args.in.args[0].size = sizeof(inarg);
2780 args.in.args[0].value = &inarg;
2781 args.out.numargs = 1;
2782 args.out.args[0].size = sizeof(outarg);
2783 args.out.args[0].value = &outarg;
2784 err = fuse_simple_request(fc, &args);
2785
2786 if (!err)
2787 return outarg.revents;
2788 if (err == -ENOSYS) {
2789 fc->no_poll = 1;
2790 return DEFAULT_POLLMASK;
2791 }
2792 return POLLERR;
2793}
2794EXPORT_SYMBOL_GPL(fuse_file_poll);
2795
2796/*
2797 * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2798 * wakes up the poll waiters.
2799 */
2800int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2801 struct fuse_notify_poll_wakeup_out *outarg)
2802{
2803 u64 kh = outarg->kh;
2804 struct rb_node **link;
2805
2806 spin_lock(&fc->lock);
2807
2808 link = fuse_find_polled_node(fc, kh, NULL);
2809 if (*link) {
2810 struct fuse_file *ff;
2811
2812 ff = rb_entry(*link, struct fuse_file, polled_node);
2813 wake_up_interruptible_sync(&ff->poll_wait);
2814 }
2815
2816 spin_unlock(&fc->lock);
2817 return 0;
2818}
2819
2820static void fuse_do_truncate(struct file *file)
2821{
2822 struct inode *inode = file->f_mapping->host;
2823 struct iattr attr;
2824
2825 attr.ia_valid = ATTR_SIZE;
2826 attr.ia_size = i_size_read(inode);
2827
2828 attr.ia_file = file;
2829 attr.ia_valid |= ATTR_FILE;
2830
2831 fuse_do_setattr(inode, &attr, file);
2832}
2833
2834static inline loff_t fuse_round_up(loff_t off)
2835{
2836 return round_up(off, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
2837}
2838
2839static ssize_t
2840fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter, loff_t offset)
2841{
2842 DECLARE_COMPLETION_ONSTACK(wait);
2843 ssize_t ret = 0;
2844 struct file *file = iocb->ki_filp;
2845 struct fuse_file *ff = file->private_data;
2846 bool async_dio = ff->fc->async_dio;
2847 loff_t pos = 0;
2848 struct inode *inode;
2849 loff_t i_size;
2850 size_t count = iov_iter_count(iter);
2851 struct fuse_io_priv *io;
2852 bool is_sync = is_sync_kiocb(iocb);
2853
2854 pos = offset;
2855 inode = file->f_mapping->host;
2856 i_size = i_size_read(inode);
2857
2858 if ((iov_iter_rw(iter) == READ) && (offset > i_size))
2859 return 0;
2860
2861 /* optimization for short read */
2862 if (async_dio && iov_iter_rw(iter) != WRITE && offset + count > i_size) {
2863 if (offset >= i_size)
2864 return 0;
2865 iov_iter_truncate(iter, fuse_round_up(i_size - offset));
2866 count = iov_iter_count(iter);
2867 }
2868
2869 io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2870 if (!io)
2871 return -ENOMEM;
2872 spin_lock_init(&io->lock);
2873 kref_init(&io->refcnt);
2874 io->reqs = 1;
2875 io->bytes = -1;
2876 io->size = 0;
2877 io->offset = offset;
2878 io->write = (iov_iter_rw(iter) == WRITE);
2879 io->err = 0;
2880 io->file = file;
2881 /*
2882 * By default, we want to optimize all I/Os with async request
2883 * submission to the client filesystem if supported.
2884 */
2885 io->async = async_dio;
2886 io->iocb = iocb;
2887
2888 /*
2889 * We cannot asynchronously extend the size of a file. We have no method
2890 * to wait on real async I/O requests, so we must submit this request
2891 * synchronously.
2892 */
2893 if (!is_sync && (offset + count > i_size) &&
2894 iov_iter_rw(iter) == WRITE)
2895 io->async = false;
2896
2897 if (io->async && is_sync) {
2898 /*
2899 * Additional reference to keep io around after
2900 * calling fuse_aio_complete()
2901 */
2902 kref_get(&io->refcnt);
2903 io->done = &wait;
2904 }
2905
2906 if (iov_iter_rw(iter) == WRITE) {
2907 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
2908 fuse_invalidate_attr(inode);
2909 } else {
2910 ret = __fuse_direct_read(io, iter, &pos);
2911 }
2912
2913 if (io->async) {
2914 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2915
2916 /* we have a non-extending, async request, so return */
2917 if (!is_sync)
2918 return -EIOCBQUEUED;
2919
2920 wait_for_completion(&wait);
2921 ret = fuse_get_res_by_io(io);
2922 }
2923
2924 kref_put(&io->refcnt, fuse_io_release);
2925
2926 if (iov_iter_rw(iter) == WRITE) {
2927 if (ret > 0)
2928 fuse_write_update_size(inode, pos);
2929 else if (ret < 0 && offset + count > i_size)
2930 fuse_do_truncate(file);
2931 }
2932
2933 return ret;
2934}
2935
2936static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2937 loff_t length)
2938{
2939 struct fuse_file *ff = file->private_data;
2940 struct inode *inode = file_inode(file);
2941 struct fuse_inode *fi = get_fuse_inode(inode);
2942 struct fuse_conn *fc = ff->fc;
2943 FUSE_ARGS(args);
2944 struct fuse_fallocate_in inarg = {
2945 .fh = ff->fh,
2946 .offset = offset,
2947 .length = length,
2948 .mode = mode
2949 };
2950 int err;
2951 bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
2952 (mode & FALLOC_FL_PUNCH_HOLE);
2953
2954 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2955 return -EOPNOTSUPP;
2956
2957 if (fc->no_fallocate)
2958 return -EOPNOTSUPP;
2959
2960 if (lock_inode) {
2961 inode_lock(inode);
2962 if (mode & FALLOC_FL_PUNCH_HOLE) {
2963 loff_t endbyte = offset + length - 1;
2964 err = filemap_write_and_wait_range(inode->i_mapping,
2965 offset, endbyte);
2966 if (err)
2967 goto out;
2968
2969 fuse_sync_writes(inode);
2970 }
2971 }
2972
2973 if (!(mode & FALLOC_FL_KEEP_SIZE))
2974 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2975
2976 args.in.h.opcode = FUSE_FALLOCATE;
2977 args.in.h.nodeid = ff->nodeid;
2978 args.in.numargs = 1;
2979 args.in.args[0].size = sizeof(inarg);
2980 args.in.args[0].value = &inarg;
2981 err = fuse_simple_request(fc, &args);
2982 if (err == -ENOSYS) {
2983 fc->no_fallocate = 1;
2984 err = -EOPNOTSUPP;
2985 }
2986 if (err)
2987 goto out;
2988
2989 /* we could have extended the file */
2990 if (!(mode & FALLOC_FL_KEEP_SIZE)) {
2991 bool changed = fuse_write_update_size(inode, offset + length);
2992
2993 if (changed && fc->writeback_cache)
2994 file_update_time(file);
2995 }
2996
2997 if (mode & FALLOC_FL_PUNCH_HOLE)
2998 truncate_pagecache_range(inode, offset, offset + length - 1);
2999
3000 fuse_invalidate_attr(inode);
3001
3002out:
3003 if (!(mode & FALLOC_FL_KEEP_SIZE))
3004 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3005
3006 if (lock_inode)
3007 inode_unlock(inode);
3008
3009 return err;
3010}
3011
3012static const struct file_operations fuse_file_operations = {
3013 .llseek = fuse_file_llseek,
3014 .read_iter = fuse_file_read_iter,
3015 .write_iter = fuse_file_write_iter,
3016 .mmap = fuse_file_mmap,
3017 .open = fuse_open,
3018 .flush = fuse_flush,
3019 .release = fuse_release,
3020 .fsync = fuse_fsync,
3021 .lock = fuse_file_lock,
3022 .flock = fuse_file_flock,
3023 .splice_read = generic_file_splice_read,
3024 .unlocked_ioctl = fuse_file_ioctl,
3025 .compat_ioctl = fuse_file_compat_ioctl,
3026 .poll = fuse_file_poll,
3027 .fallocate = fuse_file_fallocate,
3028};
3029
3030static const struct file_operations fuse_direct_io_file_operations = {
3031 .llseek = fuse_file_llseek,
3032 .read_iter = fuse_direct_read_iter,
3033 .write_iter = fuse_direct_write_iter,
3034 .mmap = fuse_direct_mmap,
3035 .open = fuse_open,
3036 .flush = fuse_flush,
3037 .release = fuse_release,
3038 .fsync = fuse_fsync,
3039 .lock = fuse_file_lock,
3040 .flock = fuse_file_flock,
3041 .unlocked_ioctl = fuse_file_ioctl,
3042 .compat_ioctl = fuse_file_compat_ioctl,
3043 .poll = fuse_file_poll,
3044 .fallocate = fuse_file_fallocate,
3045 /* no splice_read */
3046};
3047
3048static const struct address_space_operations fuse_file_aops = {
3049 .readpage = fuse_readpage,
3050 .writepage = fuse_writepage,
3051 .writepages = fuse_writepages,
3052 .launder_page = fuse_launder_page,
3053 .readpages = fuse_readpages,
3054 .set_page_dirty = __set_page_dirty_nobuffers,
3055 .bmap = fuse_bmap,
3056 .direct_IO = fuse_direct_IO,
3057 .write_begin = fuse_write_begin,
3058 .write_end = fuse_write_end,
3059};
3060
3061void fuse_init_file_inode(struct inode *inode)
3062{
3063 inode->i_fop = &fuse_file_operations;
3064 inode->i_data.a_ops = &fuse_file_aops;
3065}
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
19static const struct file_operations fuse_direct_io_file_operations;
20
21static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
22 int opcode, struct fuse_open_out *outargp)
23{
24 struct fuse_open_in inarg;
25 struct fuse_req *req;
26 int err;
27
28 req = fuse_get_req(fc);
29 if (IS_ERR(req))
30 return PTR_ERR(req);
31
32 memset(&inarg, 0, sizeof(inarg));
33 inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
34 if (!fc->atomic_o_trunc)
35 inarg.flags &= ~O_TRUNC;
36 req->in.h.opcode = opcode;
37 req->in.h.nodeid = nodeid;
38 req->in.numargs = 1;
39 req->in.args[0].size = sizeof(inarg);
40 req->in.args[0].value = &inarg;
41 req->out.numargs = 1;
42 req->out.args[0].size = sizeof(*outargp);
43 req->out.args[0].value = outargp;
44 fuse_request_send(fc, req);
45 err = req->out.h.error;
46 fuse_put_request(fc, req);
47
48 return err;
49}
50
51struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
52{
53 struct fuse_file *ff;
54
55 ff = kmalloc(sizeof(struct fuse_file), GFP_KERNEL);
56 if (unlikely(!ff))
57 return NULL;
58
59 ff->fc = fc;
60 ff->reserved_req = fuse_request_alloc();
61 if (unlikely(!ff->reserved_req)) {
62 kfree(ff);
63 return NULL;
64 }
65
66 INIT_LIST_HEAD(&ff->write_entry);
67 atomic_set(&ff->count, 0);
68 RB_CLEAR_NODE(&ff->polled_node);
69 init_waitqueue_head(&ff->poll_wait);
70
71 spin_lock(&fc->lock);
72 ff->kh = ++fc->khctr;
73 spin_unlock(&fc->lock);
74
75 return ff;
76}
77
78void fuse_file_free(struct fuse_file *ff)
79{
80 fuse_request_free(ff->reserved_req);
81 kfree(ff);
82}
83
84struct fuse_file *fuse_file_get(struct fuse_file *ff)
85{
86 atomic_inc(&ff->count);
87 return ff;
88}
89
90static void fuse_release_async(struct work_struct *work)
91{
92 struct fuse_req *req;
93 struct fuse_conn *fc;
94 struct path path;
95
96 req = container_of(work, struct fuse_req, misc.release.work);
97 path = req->misc.release.path;
98 fc = get_fuse_conn(path.dentry->d_inode);
99
100 fuse_put_request(fc, req);
101 path_put(&path);
102}
103
104static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
105{
106 if (fc->destroy_req) {
107 /*
108 * If this is a fuseblk mount, then it's possible that
109 * releasing the path will result in releasing the
110 * super block and sending the DESTROY request. If
111 * the server is single threaded, this would hang.
112 * For this reason do the path_put() in a separate
113 * thread.
114 */
115 atomic_inc(&req->count);
116 INIT_WORK(&req->misc.release.work, fuse_release_async);
117 schedule_work(&req->misc.release.work);
118 } else {
119 path_put(&req->misc.release.path);
120 }
121}
122
123static void fuse_file_put(struct fuse_file *ff, bool sync)
124{
125 if (atomic_dec_and_test(&ff->count)) {
126 struct fuse_req *req = ff->reserved_req;
127
128 if (sync) {
129 fuse_request_send(ff->fc, req);
130 path_put(&req->misc.release.path);
131 fuse_put_request(ff->fc, req);
132 } else {
133 req->end = fuse_release_end;
134 fuse_request_send_background(ff->fc, req);
135 }
136 kfree(ff);
137 }
138}
139
140int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
141 bool isdir)
142{
143 struct fuse_open_out outarg;
144 struct fuse_file *ff;
145 int err;
146 int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
147
148 ff = fuse_file_alloc(fc);
149 if (!ff)
150 return -ENOMEM;
151
152 err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
153 if (err) {
154 fuse_file_free(ff);
155 return err;
156 }
157
158 if (isdir)
159 outarg.open_flags &= ~FOPEN_DIRECT_IO;
160
161 ff->fh = outarg.fh;
162 ff->nodeid = nodeid;
163 ff->open_flags = outarg.open_flags;
164 file->private_data = fuse_file_get(ff);
165
166 return 0;
167}
168EXPORT_SYMBOL_GPL(fuse_do_open);
169
170void fuse_finish_open(struct inode *inode, struct file *file)
171{
172 struct fuse_file *ff = file->private_data;
173 struct fuse_conn *fc = get_fuse_conn(inode);
174
175 if (ff->open_flags & FOPEN_DIRECT_IO)
176 file->f_op = &fuse_direct_io_file_operations;
177 if (!(ff->open_flags & FOPEN_KEEP_CACHE))
178 invalidate_inode_pages2(inode->i_mapping);
179 if (ff->open_flags & FOPEN_NONSEEKABLE)
180 nonseekable_open(inode, file);
181 if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
182 struct fuse_inode *fi = get_fuse_inode(inode);
183
184 spin_lock(&fc->lock);
185 fi->attr_version = ++fc->attr_version;
186 i_size_write(inode, 0);
187 spin_unlock(&fc->lock);
188 fuse_invalidate_attr(inode);
189 }
190}
191
192int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
193{
194 struct fuse_conn *fc = get_fuse_conn(inode);
195 int err;
196
197 /* VFS checks this, but only _after_ ->open() */
198 if (file->f_flags & O_DIRECT)
199 return -EINVAL;
200
201 err = generic_file_open(inode, file);
202 if (err)
203 return err;
204
205 err = fuse_do_open(fc, get_node_id(inode), file, isdir);
206 if (err)
207 return err;
208
209 fuse_finish_open(inode, file);
210
211 return 0;
212}
213
214static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
215{
216 struct fuse_conn *fc = ff->fc;
217 struct fuse_req *req = ff->reserved_req;
218 struct fuse_release_in *inarg = &req->misc.release.in;
219
220 spin_lock(&fc->lock);
221 list_del(&ff->write_entry);
222 if (!RB_EMPTY_NODE(&ff->polled_node))
223 rb_erase(&ff->polled_node, &fc->polled_files);
224 spin_unlock(&fc->lock);
225
226 wake_up_interruptible_all(&ff->poll_wait);
227
228 inarg->fh = ff->fh;
229 inarg->flags = flags;
230 req->in.h.opcode = opcode;
231 req->in.h.nodeid = ff->nodeid;
232 req->in.numargs = 1;
233 req->in.args[0].size = sizeof(struct fuse_release_in);
234 req->in.args[0].value = inarg;
235}
236
237void fuse_release_common(struct file *file, int opcode)
238{
239 struct fuse_file *ff;
240 struct fuse_req *req;
241
242 ff = file->private_data;
243 if (unlikely(!ff))
244 return;
245
246 req = ff->reserved_req;
247 fuse_prepare_release(ff, file->f_flags, opcode);
248
249 if (ff->flock) {
250 struct fuse_release_in *inarg = &req->misc.release.in;
251 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
252 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
253 (fl_owner_t) file);
254 }
255 /* Hold vfsmount and dentry until release is finished */
256 path_get(&file->f_path);
257 req->misc.release.path = file->f_path;
258
259 /*
260 * Normally this will send the RELEASE request, however if
261 * some asynchronous READ or WRITE requests are outstanding,
262 * the sending will be delayed.
263 *
264 * Make the release synchronous if this is a fuseblk mount,
265 * synchronous RELEASE is allowed (and desirable) in this case
266 * because the server can be trusted not to screw up.
267 */
268 fuse_file_put(ff, ff->fc->destroy_req != NULL);
269}
270
271static int fuse_open(struct inode *inode, struct file *file)
272{
273 return fuse_open_common(inode, file, false);
274}
275
276static int fuse_release(struct inode *inode, struct file *file)
277{
278 fuse_release_common(file, FUSE_RELEASE);
279
280 /* return value is ignored by VFS */
281 return 0;
282}
283
284void fuse_sync_release(struct fuse_file *ff, int flags)
285{
286 WARN_ON(atomic_read(&ff->count) > 1);
287 fuse_prepare_release(ff, flags, FUSE_RELEASE);
288 ff->reserved_req->force = 1;
289 fuse_request_send(ff->fc, ff->reserved_req);
290 fuse_put_request(ff->fc, ff->reserved_req);
291 kfree(ff);
292}
293EXPORT_SYMBOL_GPL(fuse_sync_release);
294
295/*
296 * Scramble the ID space with XTEA, so that the value of the files_struct
297 * pointer is not exposed to userspace.
298 */
299u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
300{
301 u32 *k = fc->scramble_key;
302 u64 v = (unsigned long) id;
303 u32 v0 = v;
304 u32 v1 = v >> 32;
305 u32 sum = 0;
306 int i;
307
308 for (i = 0; i < 32; i++) {
309 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
310 sum += 0x9E3779B9;
311 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
312 }
313
314 return (u64) v0 + ((u64) v1 << 32);
315}
316
317/*
318 * Check if page is under writeback
319 *
320 * This is currently done by walking the list of writepage requests
321 * for the inode, which can be pretty inefficient.
322 */
323static bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
324{
325 struct fuse_conn *fc = get_fuse_conn(inode);
326 struct fuse_inode *fi = get_fuse_inode(inode);
327 struct fuse_req *req;
328 bool found = false;
329
330 spin_lock(&fc->lock);
331 list_for_each_entry(req, &fi->writepages, writepages_entry) {
332 pgoff_t curr_index;
333
334 BUG_ON(req->inode != inode);
335 curr_index = req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
336 if (curr_index == index) {
337 found = true;
338 break;
339 }
340 }
341 spin_unlock(&fc->lock);
342
343 return found;
344}
345
346/*
347 * Wait for page writeback to be completed.
348 *
349 * Since fuse doesn't rely on the VM writeback tracking, this has to
350 * use some other means.
351 */
352static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
353{
354 struct fuse_inode *fi = get_fuse_inode(inode);
355
356 wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
357 return 0;
358}
359
360static int fuse_flush(struct file *file, fl_owner_t id)
361{
362 struct inode *inode = file->f_path.dentry->d_inode;
363 struct fuse_conn *fc = get_fuse_conn(inode);
364 struct fuse_file *ff = file->private_data;
365 struct fuse_req *req;
366 struct fuse_flush_in inarg;
367 int err;
368
369 if (is_bad_inode(inode))
370 return -EIO;
371
372 if (fc->no_flush)
373 return 0;
374
375 req = fuse_get_req_nofail(fc, file);
376 memset(&inarg, 0, sizeof(inarg));
377 inarg.fh = ff->fh;
378 inarg.lock_owner = fuse_lock_owner_id(fc, id);
379 req->in.h.opcode = FUSE_FLUSH;
380 req->in.h.nodeid = get_node_id(inode);
381 req->in.numargs = 1;
382 req->in.args[0].size = sizeof(inarg);
383 req->in.args[0].value = &inarg;
384 req->force = 1;
385 fuse_request_send(fc, req);
386 err = req->out.h.error;
387 fuse_put_request(fc, req);
388 if (err == -ENOSYS) {
389 fc->no_flush = 1;
390 err = 0;
391 }
392 return err;
393}
394
395/*
396 * Wait for all pending writepages on the inode to finish.
397 *
398 * This is currently done by blocking further writes with FUSE_NOWRITE
399 * and waiting for all sent writes to complete.
400 *
401 * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
402 * could conflict with truncation.
403 */
404static void fuse_sync_writes(struct inode *inode)
405{
406 fuse_set_nowrite(inode);
407 fuse_release_nowrite(inode);
408}
409
410int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
411 int datasync, int isdir)
412{
413 struct inode *inode = file->f_mapping->host;
414 struct fuse_conn *fc = get_fuse_conn(inode);
415 struct fuse_file *ff = file->private_data;
416 struct fuse_req *req;
417 struct fuse_fsync_in inarg;
418 int err;
419
420 if (is_bad_inode(inode))
421 return -EIO;
422
423 err = filemap_write_and_wait_range(inode->i_mapping, start, end);
424 if (err)
425 return err;
426
427 if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
428 return 0;
429
430 mutex_lock(&inode->i_mutex);
431
432 /*
433 * Start writeback against all dirty pages of the inode, then
434 * wait for all outstanding writes, before sending the FSYNC
435 * request.
436 */
437 err = write_inode_now(inode, 0);
438 if (err)
439 goto out;
440
441 fuse_sync_writes(inode);
442
443 req = fuse_get_req(fc);
444 if (IS_ERR(req)) {
445 err = PTR_ERR(req);
446 goto out;
447 }
448
449 memset(&inarg, 0, sizeof(inarg));
450 inarg.fh = ff->fh;
451 inarg.fsync_flags = datasync ? 1 : 0;
452 req->in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
453 req->in.h.nodeid = get_node_id(inode);
454 req->in.numargs = 1;
455 req->in.args[0].size = sizeof(inarg);
456 req->in.args[0].value = &inarg;
457 fuse_request_send(fc, req);
458 err = req->out.h.error;
459 fuse_put_request(fc, req);
460 if (err == -ENOSYS) {
461 if (isdir)
462 fc->no_fsyncdir = 1;
463 else
464 fc->no_fsync = 1;
465 err = 0;
466 }
467out:
468 mutex_unlock(&inode->i_mutex);
469 return err;
470}
471
472static int fuse_fsync(struct file *file, loff_t start, loff_t end,
473 int datasync)
474{
475 return fuse_fsync_common(file, start, end, datasync, 0);
476}
477
478void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
479 size_t count, int opcode)
480{
481 struct fuse_read_in *inarg = &req->misc.read.in;
482 struct fuse_file *ff = file->private_data;
483
484 inarg->fh = ff->fh;
485 inarg->offset = pos;
486 inarg->size = count;
487 inarg->flags = file->f_flags;
488 req->in.h.opcode = opcode;
489 req->in.h.nodeid = ff->nodeid;
490 req->in.numargs = 1;
491 req->in.args[0].size = sizeof(struct fuse_read_in);
492 req->in.args[0].value = inarg;
493 req->out.argvar = 1;
494 req->out.numargs = 1;
495 req->out.args[0].size = count;
496}
497
498static size_t fuse_send_read(struct fuse_req *req, struct file *file,
499 loff_t pos, size_t count, fl_owner_t owner)
500{
501 struct fuse_file *ff = file->private_data;
502 struct fuse_conn *fc = ff->fc;
503
504 fuse_read_fill(req, file, pos, count, FUSE_READ);
505 if (owner != NULL) {
506 struct fuse_read_in *inarg = &req->misc.read.in;
507
508 inarg->read_flags |= FUSE_READ_LOCKOWNER;
509 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
510 }
511 fuse_request_send(fc, req);
512 return req->out.args[0].size;
513}
514
515static void fuse_read_update_size(struct inode *inode, loff_t size,
516 u64 attr_ver)
517{
518 struct fuse_conn *fc = get_fuse_conn(inode);
519 struct fuse_inode *fi = get_fuse_inode(inode);
520
521 spin_lock(&fc->lock);
522 if (attr_ver == fi->attr_version && size < inode->i_size) {
523 fi->attr_version = ++fc->attr_version;
524 i_size_write(inode, size);
525 }
526 spin_unlock(&fc->lock);
527}
528
529static int fuse_readpage(struct file *file, struct page *page)
530{
531 struct inode *inode = page->mapping->host;
532 struct fuse_conn *fc = get_fuse_conn(inode);
533 struct fuse_req *req;
534 size_t num_read;
535 loff_t pos = page_offset(page);
536 size_t count = PAGE_CACHE_SIZE;
537 u64 attr_ver;
538 int err;
539
540 err = -EIO;
541 if (is_bad_inode(inode))
542 goto out;
543
544 /*
545 * Page writeback can extend beyond the lifetime of the
546 * page-cache page, so make sure we read a properly synced
547 * page.
548 */
549 fuse_wait_on_page_writeback(inode, page->index);
550
551 req = fuse_get_req(fc);
552 err = PTR_ERR(req);
553 if (IS_ERR(req))
554 goto out;
555
556 attr_ver = fuse_get_attr_version(fc);
557
558 req->out.page_zeroing = 1;
559 req->out.argpages = 1;
560 req->num_pages = 1;
561 req->pages[0] = page;
562 num_read = fuse_send_read(req, file, pos, count, NULL);
563 err = req->out.h.error;
564 fuse_put_request(fc, req);
565
566 if (!err) {
567 /*
568 * Short read means EOF. If file size is larger, truncate it
569 */
570 if (num_read < count)
571 fuse_read_update_size(inode, pos + num_read, attr_ver);
572
573 SetPageUptodate(page);
574 }
575
576 fuse_invalidate_attr(inode); /* atime changed */
577 out:
578 unlock_page(page);
579 return err;
580}
581
582static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
583{
584 int i;
585 size_t count = req->misc.read.in.size;
586 size_t num_read = req->out.args[0].size;
587 struct address_space *mapping = NULL;
588
589 for (i = 0; mapping == NULL && i < req->num_pages; i++)
590 mapping = req->pages[i]->mapping;
591
592 if (mapping) {
593 struct inode *inode = mapping->host;
594
595 /*
596 * Short read means EOF. If file size is larger, truncate it
597 */
598 if (!req->out.h.error && num_read < count) {
599 loff_t pos;
600
601 pos = page_offset(req->pages[0]) + num_read;
602 fuse_read_update_size(inode, pos,
603 req->misc.read.attr_ver);
604 }
605 fuse_invalidate_attr(inode); /* atime changed */
606 }
607
608 for (i = 0; i < req->num_pages; i++) {
609 struct page *page = req->pages[i];
610 if (!req->out.h.error)
611 SetPageUptodate(page);
612 else
613 SetPageError(page);
614 unlock_page(page);
615 page_cache_release(page);
616 }
617 if (req->ff)
618 fuse_file_put(req->ff, false);
619}
620
621static void fuse_send_readpages(struct fuse_req *req, struct file *file)
622{
623 struct fuse_file *ff = file->private_data;
624 struct fuse_conn *fc = ff->fc;
625 loff_t pos = page_offset(req->pages[0]);
626 size_t count = req->num_pages << PAGE_CACHE_SHIFT;
627
628 req->out.argpages = 1;
629 req->out.page_zeroing = 1;
630 req->out.page_replace = 1;
631 fuse_read_fill(req, file, pos, count, FUSE_READ);
632 req->misc.read.attr_ver = fuse_get_attr_version(fc);
633 if (fc->async_read) {
634 req->ff = fuse_file_get(ff);
635 req->end = fuse_readpages_end;
636 fuse_request_send_background(fc, req);
637 } else {
638 fuse_request_send(fc, req);
639 fuse_readpages_end(fc, req);
640 fuse_put_request(fc, req);
641 }
642}
643
644struct fuse_fill_data {
645 struct fuse_req *req;
646 struct file *file;
647 struct inode *inode;
648};
649
650static int fuse_readpages_fill(void *_data, struct page *page)
651{
652 struct fuse_fill_data *data = _data;
653 struct fuse_req *req = data->req;
654 struct inode *inode = data->inode;
655 struct fuse_conn *fc = get_fuse_conn(inode);
656
657 fuse_wait_on_page_writeback(inode, page->index);
658
659 if (req->num_pages &&
660 (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
661 (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
662 req->pages[req->num_pages - 1]->index + 1 != page->index)) {
663 fuse_send_readpages(req, data->file);
664 data->req = req = fuse_get_req(fc);
665 if (IS_ERR(req)) {
666 unlock_page(page);
667 return PTR_ERR(req);
668 }
669 }
670 page_cache_get(page);
671 req->pages[req->num_pages] = page;
672 req->num_pages++;
673 return 0;
674}
675
676static int fuse_readpages(struct file *file, struct address_space *mapping,
677 struct list_head *pages, unsigned nr_pages)
678{
679 struct inode *inode = mapping->host;
680 struct fuse_conn *fc = get_fuse_conn(inode);
681 struct fuse_fill_data data;
682 int err;
683
684 err = -EIO;
685 if (is_bad_inode(inode))
686 goto out;
687
688 data.file = file;
689 data.inode = inode;
690 data.req = fuse_get_req(fc);
691 err = PTR_ERR(data.req);
692 if (IS_ERR(data.req))
693 goto out;
694
695 err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
696 if (!err) {
697 if (data.req->num_pages)
698 fuse_send_readpages(data.req, file);
699 else
700 fuse_put_request(fc, data.req);
701 }
702out:
703 return err;
704}
705
706static ssize_t fuse_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
707 unsigned long nr_segs, loff_t pos)
708{
709 struct inode *inode = iocb->ki_filp->f_mapping->host;
710
711 if (pos + iov_length(iov, nr_segs) > i_size_read(inode)) {
712 int err;
713 /*
714 * If trying to read past EOF, make sure the i_size
715 * attribute is up-to-date.
716 */
717 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
718 if (err)
719 return err;
720 }
721
722 return generic_file_aio_read(iocb, iov, nr_segs, pos);
723}
724
725static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
726 loff_t pos, size_t count)
727{
728 struct fuse_write_in *inarg = &req->misc.write.in;
729 struct fuse_write_out *outarg = &req->misc.write.out;
730
731 inarg->fh = ff->fh;
732 inarg->offset = pos;
733 inarg->size = count;
734 req->in.h.opcode = FUSE_WRITE;
735 req->in.h.nodeid = ff->nodeid;
736 req->in.numargs = 2;
737 if (ff->fc->minor < 9)
738 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
739 else
740 req->in.args[0].size = sizeof(struct fuse_write_in);
741 req->in.args[0].value = inarg;
742 req->in.args[1].size = count;
743 req->out.numargs = 1;
744 req->out.args[0].size = sizeof(struct fuse_write_out);
745 req->out.args[0].value = outarg;
746}
747
748static size_t fuse_send_write(struct fuse_req *req, struct file *file,
749 loff_t pos, size_t count, fl_owner_t owner)
750{
751 struct fuse_file *ff = file->private_data;
752 struct fuse_conn *fc = ff->fc;
753 struct fuse_write_in *inarg = &req->misc.write.in;
754
755 fuse_write_fill(req, ff, pos, count);
756 inarg->flags = file->f_flags;
757 if (owner != NULL) {
758 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
759 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
760 }
761 fuse_request_send(fc, req);
762 return req->misc.write.out.size;
763}
764
765void fuse_write_update_size(struct inode *inode, loff_t pos)
766{
767 struct fuse_conn *fc = get_fuse_conn(inode);
768 struct fuse_inode *fi = get_fuse_inode(inode);
769
770 spin_lock(&fc->lock);
771 fi->attr_version = ++fc->attr_version;
772 if (pos > inode->i_size)
773 i_size_write(inode, pos);
774 spin_unlock(&fc->lock);
775}
776
777static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
778 struct inode *inode, loff_t pos,
779 size_t count)
780{
781 size_t res;
782 unsigned offset;
783 unsigned i;
784
785 for (i = 0; i < req->num_pages; i++)
786 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
787
788 res = fuse_send_write(req, file, pos, count, NULL);
789
790 offset = req->page_offset;
791 count = res;
792 for (i = 0; i < req->num_pages; i++) {
793 struct page *page = req->pages[i];
794
795 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
796 SetPageUptodate(page);
797
798 if (count > PAGE_CACHE_SIZE - offset)
799 count -= PAGE_CACHE_SIZE - offset;
800 else
801 count = 0;
802 offset = 0;
803
804 unlock_page(page);
805 page_cache_release(page);
806 }
807
808 return res;
809}
810
811static ssize_t fuse_fill_write_pages(struct fuse_req *req,
812 struct address_space *mapping,
813 struct iov_iter *ii, loff_t pos)
814{
815 struct fuse_conn *fc = get_fuse_conn(mapping->host);
816 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
817 size_t count = 0;
818 int err;
819
820 req->in.argpages = 1;
821 req->page_offset = offset;
822
823 do {
824 size_t tmp;
825 struct page *page;
826 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
827 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
828 iov_iter_count(ii));
829
830 bytes = min_t(size_t, bytes, fc->max_write - count);
831
832 again:
833 err = -EFAULT;
834 if (iov_iter_fault_in_readable(ii, bytes))
835 break;
836
837 err = -ENOMEM;
838 page = grab_cache_page_write_begin(mapping, index, 0);
839 if (!page)
840 break;
841
842 if (mapping_writably_mapped(mapping))
843 flush_dcache_page(page);
844
845 pagefault_disable();
846 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
847 pagefault_enable();
848 flush_dcache_page(page);
849
850 mark_page_accessed(page);
851
852 if (!tmp) {
853 unlock_page(page);
854 page_cache_release(page);
855 bytes = min(bytes, iov_iter_single_seg_count(ii));
856 goto again;
857 }
858
859 err = 0;
860 req->pages[req->num_pages] = page;
861 req->num_pages++;
862
863 iov_iter_advance(ii, tmp);
864 count += tmp;
865 pos += tmp;
866 offset += tmp;
867 if (offset == PAGE_CACHE_SIZE)
868 offset = 0;
869
870 if (!fc->big_writes)
871 break;
872 } while (iov_iter_count(ii) && count < fc->max_write &&
873 req->num_pages < FUSE_MAX_PAGES_PER_REQ && offset == 0);
874
875 return count > 0 ? count : err;
876}
877
878static ssize_t fuse_perform_write(struct file *file,
879 struct address_space *mapping,
880 struct iov_iter *ii, loff_t pos)
881{
882 struct inode *inode = mapping->host;
883 struct fuse_conn *fc = get_fuse_conn(inode);
884 int err = 0;
885 ssize_t res = 0;
886
887 if (is_bad_inode(inode))
888 return -EIO;
889
890 do {
891 struct fuse_req *req;
892 ssize_t count;
893
894 req = fuse_get_req(fc);
895 if (IS_ERR(req)) {
896 err = PTR_ERR(req);
897 break;
898 }
899
900 count = fuse_fill_write_pages(req, mapping, ii, pos);
901 if (count <= 0) {
902 err = count;
903 } else {
904 size_t num_written;
905
906 num_written = fuse_send_write_pages(req, file, inode,
907 pos, count);
908 err = req->out.h.error;
909 if (!err) {
910 res += num_written;
911 pos += num_written;
912
913 /* break out of the loop on short write */
914 if (num_written != count)
915 err = -EIO;
916 }
917 }
918 fuse_put_request(fc, req);
919 } while (!err && iov_iter_count(ii));
920
921 if (res > 0)
922 fuse_write_update_size(inode, pos);
923
924 fuse_invalidate_attr(inode);
925
926 return res > 0 ? res : err;
927}
928
929static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
930 unsigned long nr_segs, loff_t pos)
931{
932 struct file *file = iocb->ki_filp;
933 struct address_space *mapping = file->f_mapping;
934 size_t count = 0;
935 ssize_t written = 0;
936 struct inode *inode = mapping->host;
937 ssize_t err;
938 struct iov_iter i;
939
940 WARN_ON(iocb->ki_pos != pos);
941
942 err = generic_segment_checks(iov, &nr_segs, &count, VERIFY_READ);
943 if (err)
944 return err;
945
946 mutex_lock(&inode->i_mutex);
947 vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
948
949 /* We can write back this queue in page reclaim */
950 current->backing_dev_info = mapping->backing_dev_info;
951
952 err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
953 if (err)
954 goto out;
955
956 if (count == 0)
957 goto out;
958
959 err = file_remove_suid(file);
960 if (err)
961 goto out;
962
963 file_update_time(file);
964
965 iov_iter_init(&i, iov, nr_segs, count, 0);
966 written = fuse_perform_write(file, mapping, &i, pos);
967 if (written >= 0)
968 iocb->ki_pos = pos + written;
969
970out:
971 current->backing_dev_info = NULL;
972 mutex_unlock(&inode->i_mutex);
973
974 return written ? written : err;
975}
976
977static void fuse_release_user_pages(struct fuse_req *req, int write)
978{
979 unsigned i;
980
981 for (i = 0; i < req->num_pages; i++) {
982 struct page *page = req->pages[i];
983 if (write)
984 set_page_dirty_lock(page);
985 put_page(page);
986 }
987}
988
989static int fuse_get_user_pages(struct fuse_req *req, const char __user *buf,
990 size_t *nbytesp, int write)
991{
992 size_t nbytes = *nbytesp;
993 unsigned long user_addr = (unsigned long) buf;
994 unsigned offset = user_addr & ~PAGE_MASK;
995 int npages;
996
997 /* Special case for kernel I/O: can copy directly into the buffer */
998 if (segment_eq(get_fs(), KERNEL_DS)) {
999 if (write)
1000 req->in.args[1].value = (void *) user_addr;
1001 else
1002 req->out.args[0].value = (void *) user_addr;
1003
1004 return 0;
1005 }
1006
1007 nbytes = min_t(size_t, nbytes, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
1008 npages = (nbytes + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1009 npages = clamp(npages, 1, FUSE_MAX_PAGES_PER_REQ);
1010 npages = get_user_pages_fast(user_addr, npages, !write, req->pages);
1011 if (npages < 0)
1012 return npages;
1013
1014 req->num_pages = npages;
1015 req->page_offset = offset;
1016
1017 if (write)
1018 req->in.argpages = 1;
1019 else
1020 req->out.argpages = 1;
1021
1022 nbytes = (req->num_pages << PAGE_SHIFT) - req->page_offset;
1023 *nbytesp = min(*nbytesp, nbytes);
1024
1025 return 0;
1026}
1027
1028ssize_t fuse_direct_io(struct file *file, const char __user *buf,
1029 size_t count, loff_t *ppos, int write)
1030{
1031 struct fuse_file *ff = file->private_data;
1032 struct fuse_conn *fc = ff->fc;
1033 size_t nmax = write ? fc->max_write : fc->max_read;
1034 loff_t pos = *ppos;
1035 ssize_t res = 0;
1036 struct fuse_req *req;
1037
1038 req = fuse_get_req(fc);
1039 if (IS_ERR(req))
1040 return PTR_ERR(req);
1041
1042 while (count) {
1043 size_t nres;
1044 fl_owner_t owner = current->files;
1045 size_t nbytes = min(count, nmax);
1046 int err = fuse_get_user_pages(req, buf, &nbytes, write);
1047 if (err) {
1048 res = err;
1049 break;
1050 }
1051
1052 if (write)
1053 nres = fuse_send_write(req, file, pos, nbytes, owner);
1054 else
1055 nres = fuse_send_read(req, file, pos, nbytes, owner);
1056
1057 fuse_release_user_pages(req, !write);
1058 if (req->out.h.error) {
1059 if (!res)
1060 res = req->out.h.error;
1061 break;
1062 } else if (nres > nbytes) {
1063 res = -EIO;
1064 break;
1065 }
1066 count -= nres;
1067 res += nres;
1068 pos += nres;
1069 buf += nres;
1070 if (nres != nbytes)
1071 break;
1072 if (count) {
1073 fuse_put_request(fc, req);
1074 req = fuse_get_req(fc);
1075 if (IS_ERR(req))
1076 break;
1077 }
1078 }
1079 if (!IS_ERR(req))
1080 fuse_put_request(fc, req);
1081 if (res > 0)
1082 *ppos = pos;
1083
1084 return res;
1085}
1086EXPORT_SYMBOL_GPL(fuse_direct_io);
1087
1088static ssize_t fuse_direct_read(struct file *file, char __user *buf,
1089 size_t count, loff_t *ppos)
1090{
1091 ssize_t res;
1092 struct inode *inode = file->f_path.dentry->d_inode;
1093
1094 if (is_bad_inode(inode))
1095 return -EIO;
1096
1097 res = fuse_direct_io(file, buf, count, ppos, 0);
1098
1099 fuse_invalidate_attr(inode);
1100
1101 return res;
1102}
1103
1104static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
1105 size_t count, loff_t *ppos)
1106{
1107 struct inode *inode = file->f_path.dentry->d_inode;
1108 ssize_t res;
1109
1110 if (is_bad_inode(inode))
1111 return -EIO;
1112
1113 /* Don't allow parallel writes to the same file */
1114 mutex_lock(&inode->i_mutex);
1115 res = generic_write_checks(file, ppos, &count, 0);
1116 if (!res) {
1117 res = fuse_direct_io(file, buf, count, ppos, 1);
1118 if (res > 0)
1119 fuse_write_update_size(inode, *ppos);
1120 }
1121 mutex_unlock(&inode->i_mutex);
1122
1123 fuse_invalidate_attr(inode);
1124
1125 return res;
1126}
1127
1128static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1129{
1130 __free_page(req->pages[0]);
1131 fuse_file_put(req->ff, false);
1132}
1133
1134static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1135{
1136 struct inode *inode = req->inode;
1137 struct fuse_inode *fi = get_fuse_inode(inode);
1138 struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
1139
1140 list_del(&req->writepages_entry);
1141 dec_bdi_stat(bdi, BDI_WRITEBACK);
1142 dec_zone_page_state(req->pages[0], NR_WRITEBACK_TEMP);
1143 bdi_writeout_inc(bdi);
1144 wake_up(&fi->page_waitq);
1145}
1146
1147/* Called under fc->lock, may release and reacquire it */
1148static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req)
1149__releases(fc->lock)
1150__acquires(fc->lock)
1151{
1152 struct fuse_inode *fi = get_fuse_inode(req->inode);
1153 loff_t size = i_size_read(req->inode);
1154 struct fuse_write_in *inarg = &req->misc.write.in;
1155
1156 if (!fc->connected)
1157 goto out_free;
1158
1159 if (inarg->offset + PAGE_CACHE_SIZE <= size) {
1160 inarg->size = PAGE_CACHE_SIZE;
1161 } else if (inarg->offset < size) {
1162 inarg->size = size & (PAGE_CACHE_SIZE - 1);
1163 } else {
1164 /* Got truncated off completely */
1165 goto out_free;
1166 }
1167
1168 req->in.args[1].size = inarg->size;
1169 fi->writectr++;
1170 fuse_request_send_background_locked(fc, req);
1171 return;
1172
1173 out_free:
1174 fuse_writepage_finish(fc, req);
1175 spin_unlock(&fc->lock);
1176 fuse_writepage_free(fc, req);
1177 fuse_put_request(fc, req);
1178 spin_lock(&fc->lock);
1179}
1180
1181/*
1182 * If fi->writectr is positive (no truncate or fsync going on) send
1183 * all queued writepage requests.
1184 *
1185 * Called with fc->lock
1186 */
1187void fuse_flush_writepages(struct inode *inode)
1188__releases(fc->lock)
1189__acquires(fc->lock)
1190{
1191 struct fuse_conn *fc = get_fuse_conn(inode);
1192 struct fuse_inode *fi = get_fuse_inode(inode);
1193 struct fuse_req *req;
1194
1195 while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1196 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1197 list_del_init(&req->list);
1198 fuse_send_writepage(fc, req);
1199 }
1200}
1201
1202static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1203{
1204 struct inode *inode = req->inode;
1205 struct fuse_inode *fi = get_fuse_inode(inode);
1206
1207 mapping_set_error(inode->i_mapping, req->out.h.error);
1208 spin_lock(&fc->lock);
1209 fi->writectr--;
1210 fuse_writepage_finish(fc, req);
1211 spin_unlock(&fc->lock);
1212 fuse_writepage_free(fc, req);
1213}
1214
1215static int fuse_writepage_locked(struct page *page)
1216{
1217 struct address_space *mapping = page->mapping;
1218 struct inode *inode = mapping->host;
1219 struct fuse_conn *fc = get_fuse_conn(inode);
1220 struct fuse_inode *fi = get_fuse_inode(inode);
1221 struct fuse_req *req;
1222 struct fuse_file *ff;
1223 struct page *tmp_page;
1224
1225 set_page_writeback(page);
1226
1227 req = fuse_request_alloc_nofs();
1228 if (!req)
1229 goto err;
1230
1231 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1232 if (!tmp_page)
1233 goto err_free;
1234
1235 spin_lock(&fc->lock);
1236 BUG_ON(list_empty(&fi->write_files));
1237 ff = list_entry(fi->write_files.next, struct fuse_file, write_entry);
1238 req->ff = fuse_file_get(ff);
1239 spin_unlock(&fc->lock);
1240
1241 fuse_write_fill(req, ff, page_offset(page), 0);
1242
1243 copy_highpage(tmp_page, page);
1244 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1245 req->in.argpages = 1;
1246 req->num_pages = 1;
1247 req->pages[0] = tmp_page;
1248 req->page_offset = 0;
1249 req->end = fuse_writepage_end;
1250 req->inode = inode;
1251
1252 inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
1253 inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1254 end_page_writeback(page);
1255
1256 spin_lock(&fc->lock);
1257 list_add(&req->writepages_entry, &fi->writepages);
1258 list_add_tail(&req->list, &fi->queued_writes);
1259 fuse_flush_writepages(inode);
1260 spin_unlock(&fc->lock);
1261
1262 return 0;
1263
1264err_free:
1265 fuse_request_free(req);
1266err:
1267 end_page_writeback(page);
1268 return -ENOMEM;
1269}
1270
1271static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1272{
1273 int err;
1274
1275 err = fuse_writepage_locked(page);
1276 unlock_page(page);
1277
1278 return err;
1279}
1280
1281static int fuse_launder_page(struct page *page)
1282{
1283 int err = 0;
1284 if (clear_page_dirty_for_io(page)) {
1285 struct inode *inode = page->mapping->host;
1286 err = fuse_writepage_locked(page);
1287 if (!err)
1288 fuse_wait_on_page_writeback(inode, page->index);
1289 }
1290 return err;
1291}
1292
1293/*
1294 * Write back dirty pages now, because there may not be any suitable
1295 * open files later
1296 */
1297static void fuse_vma_close(struct vm_area_struct *vma)
1298{
1299 filemap_write_and_wait(vma->vm_file->f_mapping);
1300}
1301
1302/*
1303 * Wait for writeback against this page to complete before allowing it
1304 * to be marked dirty again, and hence written back again, possibly
1305 * before the previous writepage completed.
1306 *
1307 * Block here, instead of in ->writepage(), so that the userspace fs
1308 * can only block processes actually operating on the filesystem.
1309 *
1310 * Otherwise unprivileged userspace fs would be able to block
1311 * unrelated:
1312 *
1313 * - page migration
1314 * - sync(2)
1315 * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
1316 */
1317static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1318{
1319 struct page *page = vmf->page;
1320 /*
1321 * Don't use page->mapping as it may become NULL from a
1322 * concurrent truncate.
1323 */
1324 struct inode *inode = vma->vm_file->f_mapping->host;
1325
1326 fuse_wait_on_page_writeback(inode, page->index);
1327 return 0;
1328}
1329
1330static const struct vm_operations_struct fuse_file_vm_ops = {
1331 .close = fuse_vma_close,
1332 .fault = filemap_fault,
1333 .page_mkwrite = fuse_page_mkwrite,
1334};
1335
1336static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
1337{
1338 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE)) {
1339 struct inode *inode = file->f_dentry->d_inode;
1340 struct fuse_conn *fc = get_fuse_conn(inode);
1341 struct fuse_inode *fi = get_fuse_inode(inode);
1342 struct fuse_file *ff = file->private_data;
1343 /*
1344 * file may be written through mmap, so chain it onto the
1345 * inodes's write_file list
1346 */
1347 spin_lock(&fc->lock);
1348 if (list_empty(&ff->write_entry))
1349 list_add(&ff->write_entry, &fi->write_files);
1350 spin_unlock(&fc->lock);
1351 }
1352 file_accessed(file);
1353 vma->vm_ops = &fuse_file_vm_ops;
1354 return 0;
1355}
1356
1357static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
1358{
1359 /* Can't provide the coherency needed for MAP_SHARED */
1360 if (vma->vm_flags & VM_MAYSHARE)
1361 return -ENODEV;
1362
1363 invalidate_inode_pages2(file->f_mapping);
1364
1365 return generic_file_mmap(file, vma);
1366}
1367
1368static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
1369 struct file_lock *fl)
1370{
1371 switch (ffl->type) {
1372 case F_UNLCK:
1373 break;
1374
1375 case F_RDLCK:
1376 case F_WRLCK:
1377 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
1378 ffl->end < ffl->start)
1379 return -EIO;
1380
1381 fl->fl_start = ffl->start;
1382 fl->fl_end = ffl->end;
1383 fl->fl_pid = ffl->pid;
1384 break;
1385
1386 default:
1387 return -EIO;
1388 }
1389 fl->fl_type = ffl->type;
1390 return 0;
1391}
1392
1393static void fuse_lk_fill(struct fuse_req *req, struct file *file,
1394 const struct file_lock *fl, int opcode, pid_t pid,
1395 int flock)
1396{
1397 struct inode *inode = file->f_path.dentry->d_inode;
1398 struct fuse_conn *fc = get_fuse_conn(inode);
1399 struct fuse_file *ff = file->private_data;
1400 struct fuse_lk_in *arg = &req->misc.lk_in;
1401
1402 arg->fh = ff->fh;
1403 arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
1404 arg->lk.start = fl->fl_start;
1405 arg->lk.end = fl->fl_end;
1406 arg->lk.type = fl->fl_type;
1407 arg->lk.pid = pid;
1408 if (flock)
1409 arg->lk_flags |= FUSE_LK_FLOCK;
1410 req->in.h.opcode = opcode;
1411 req->in.h.nodeid = get_node_id(inode);
1412 req->in.numargs = 1;
1413 req->in.args[0].size = sizeof(*arg);
1414 req->in.args[0].value = arg;
1415}
1416
1417static int fuse_getlk(struct file *file, struct file_lock *fl)
1418{
1419 struct inode *inode = file->f_path.dentry->d_inode;
1420 struct fuse_conn *fc = get_fuse_conn(inode);
1421 struct fuse_req *req;
1422 struct fuse_lk_out outarg;
1423 int err;
1424
1425 req = fuse_get_req(fc);
1426 if (IS_ERR(req))
1427 return PTR_ERR(req);
1428
1429 fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
1430 req->out.numargs = 1;
1431 req->out.args[0].size = sizeof(outarg);
1432 req->out.args[0].value = &outarg;
1433 fuse_request_send(fc, req);
1434 err = req->out.h.error;
1435 fuse_put_request(fc, req);
1436 if (!err)
1437 err = convert_fuse_file_lock(&outarg.lk, fl);
1438
1439 return err;
1440}
1441
1442static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
1443{
1444 struct inode *inode = file->f_path.dentry->d_inode;
1445 struct fuse_conn *fc = get_fuse_conn(inode);
1446 struct fuse_req *req;
1447 int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
1448 pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
1449 int err;
1450
1451 if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
1452 /* NLM needs asynchronous locks, which we don't support yet */
1453 return -ENOLCK;
1454 }
1455
1456 /* Unlock on close is handled by the flush method */
1457 if (fl->fl_flags & FL_CLOSE)
1458 return 0;
1459
1460 req = fuse_get_req(fc);
1461 if (IS_ERR(req))
1462 return PTR_ERR(req);
1463
1464 fuse_lk_fill(req, file, fl, opcode, pid, flock);
1465 fuse_request_send(fc, req);
1466 err = req->out.h.error;
1467 /* locking is restartable */
1468 if (err == -EINTR)
1469 err = -ERESTARTSYS;
1470 fuse_put_request(fc, req);
1471 return err;
1472}
1473
1474static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
1475{
1476 struct inode *inode = file->f_path.dentry->d_inode;
1477 struct fuse_conn *fc = get_fuse_conn(inode);
1478 int err;
1479
1480 if (cmd == F_CANCELLK) {
1481 err = 0;
1482 } else if (cmd == F_GETLK) {
1483 if (fc->no_lock) {
1484 posix_test_lock(file, fl);
1485 err = 0;
1486 } else
1487 err = fuse_getlk(file, fl);
1488 } else {
1489 if (fc->no_lock)
1490 err = posix_lock_file(file, fl, NULL);
1491 else
1492 err = fuse_setlk(file, fl, 0);
1493 }
1494 return err;
1495}
1496
1497static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
1498{
1499 struct inode *inode = file->f_path.dentry->d_inode;
1500 struct fuse_conn *fc = get_fuse_conn(inode);
1501 int err;
1502
1503 if (fc->no_flock) {
1504 err = flock_lock_file_wait(file, fl);
1505 } else {
1506 struct fuse_file *ff = file->private_data;
1507
1508 /* emulate flock with POSIX locks */
1509 fl->fl_owner = (fl_owner_t) file;
1510 ff->flock = true;
1511 err = fuse_setlk(file, fl, 1);
1512 }
1513
1514 return err;
1515}
1516
1517static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
1518{
1519 struct inode *inode = mapping->host;
1520 struct fuse_conn *fc = get_fuse_conn(inode);
1521 struct fuse_req *req;
1522 struct fuse_bmap_in inarg;
1523 struct fuse_bmap_out outarg;
1524 int err;
1525
1526 if (!inode->i_sb->s_bdev || fc->no_bmap)
1527 return 0;
1528
1529 req = fuse_get_req(fc);
1530 if (IS_ERR(req))
1531 return 0;
1532
1533 memset(&inarg, 0, sizeof(inarg));
1534 inarg.block = block;
1535 inarg.blocksize = inode->i_sb->s_blocksize;
1536 req->in.h.opcode = FUSE_BMAP;
1537 req->in.h.nodeid = get_node_id(inode);
1538 req->in.numargs = 1;
1539 req->in.args[0].size = sizeof(inarg);
1540 req->in.args[0].value = &inarg;
1541 req->out.numargs = 1;
1542 req->out.args[0].size = sizeof(outarg);
1543 req->out.args[0].value = &outarg;
1544 fuse_request_send(fc, req);
1545 err = req->out.h.error;
1546 fuse_put_request(fc, req);
1547 if (err == -ENOSYS)
1548 fc->no_bmap = 1;
1549
1550 return err ? 0 : outarg.block;
1551}
1552
1553static loff_t fuse_file_llseek(struct file *file, loff_t offset, int origin)
1554{
1555 loff_t retval;
1556 struct inode *inode = file->f_path.dentry->d_inode;
1557
1558 mutex_lock(&inode->i_mutex);
1559 if (origin != SEEK_CUR || origin != SEEK_SET) {
1560 retval = fuse_update_attributes(inode, NULL, file, NULL);
1561 if (retval)
1562 goto exit;
1563 }
1564
1565 switch (origin) {
1566 case SEEK_END:
1567 offset += i_size_read(inode);
1568 break;
1569 case SEEK_CUR:
1570 offset += file->f_pos;
1571 break;
1572 case SEEK_DATA:
1573 if (offset >= i_size_read(inode)) {
1574 retval = -ENXIO;
1575 goto exit;
1576 }
1577 break;
1578 case SEEK_HOLE:
1579 if (offset >= i_size_read(inode)) {
1580 retval = -ENXIO;
1581 goto exit;
1582 }
1583 offset = i_size_read(inode);
1584 break;
1585 }
1586 retval = -EINVAL;
1587 if (offset >= 0 && offset <= inode->i_sb->s_maxbytes) {
1588 if (offset != file->f_pos) {
1589 file->f_pos = offset;
1590 file->f_version = 0;
1591 }
1592 retval = offset;
1593 }
1594exit:
1595 mutex_unlock(&inode->i_mutex);
1596 return retval;
1597}
1598
1599static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
1600 unsigned int nr_segs, size_t bytes, bool to_user)
1601{
1602 struct iov_iter ii;
1603 int page_idx = 0;
1604
1605 if (!bytes)
1606 return 0;
1607
1608 iov_iter_init(&ii, iov, nr_segs, bytes, 0);
1609
1610 while (iov_iter_count(&ii)) {
1611 struct page *page = pages[page_idx++];
1612 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
1613 void *kaddr;
1614
1615 kaddr = kmap(page);
1616
1617 while (todo) {
1618 char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
1619 size_t iov_len = ii.iov->iov_len - ii.iov_offset;
1620 size_t copy = min(todo, iov_len);
1621 size_t left;
1622
1623 if (!to_user)
1624 left = copy_from_user(kaddr, uaddr, copy);
1625 else
1626 left = copy_to_user(uaddr, kaddr, copy);
1627
1628 if (unlikely(left))
1629 return -EFAULT;
1630
1631 iov_iter_advance(&ii, copy);
1632 todo -= copy;
1633 kaddr += copy;
1634 }
1635
1636 kunmap(page);
1637 }
1638
1639 return 0;
1640}
1641
1642/*
1643 * CUSE servers compiled on 32bit broke on 64bit kernels because the
1644 * ABI was defined to be 'struct iovec' which is different on 32bit
1645 * and 64bit. Fortunately we can determine which structure the server
1646 * used from the size of the reply.
1647 */
1648static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
1649 size_t transferred, unsigned count,
1650 bool is_compat)
1651{
1652#ifdef CONFIG_COMPAT
1653 if (count * sizeof(struct compat_iovec) == transferred) {
1654 struct compat_iovec *ciov = src;
1655 unsigned i;
1656
1657 /*
1658 * With this interface a 32bit server cannot support
1659 * non-compat (i.e. ones coming from 64bit apps) ioctl
1660 * requests
1661 */
1662 if (!is_compat)
1663 return -EINVAL;
1664
1665 for (i = 0; i < count; i++) {
1666 dst[i].iov_base = compat_ptr(ciov[i].iov_base);
1667 dst[i].iov_len = ciov[i].iov_len;
1668 }
1669 return 0;
1670 }
1671#endif
1672
1673 if (count * sizeof(struct iovec) != transferred)
1674 return -EIO;
1675
1676 memcpy(dst, src, transferred);
1677 return 0;
1678}
1679
1680/* Make sure iov_length() won't overflow */
1681static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
1682{
1683 size_t n;
1684 u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
1685
1686 for (n = 0; n < count; n++) {
1687 if (iov->iov_len > (size_t) max)
1688 return -ENOMEM;
1689 max -= iov->iov_len;
1690 }
1691 return 0;
1692}
1693
1694static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
1695 void *src, size_t transferred, unsigned count,
1696 bool is_compat)
1697{
1698 unsigned i;
1699 struct fuse_ioctl_iovec *fiov = src;
1700
1701 if (fc->minor < 16) {
1702 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
1703 count, is_compat);
1704 }
1705
1706 if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
1707 return -EIO;
1708
1709 for (i = 0; i < count; i++) {
1710 /* Did the server supply an inappropriate value? */
1711 if (fiov[i].base != (unsigned long) fiov[i].base ||
1712 fiov[i].len != (unsigned long) fiov[i].len)
1713 return -EIO;
1714
1715 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
1716 dst[i].iov_len = (size_t) fiov[i].len;
1717
1718#ifdef CONFIG_COMPAT
1719 if (is_compat &&
1720 (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
1721 (compat_size_t) dst[i].iov_len != fiov[i].len))
1722 return -EIO;
1723#endif
1724 }
1725
1726 return 0;
1727}
1728
1729
1730/*
1731 * For ioctls, there is no generic way to determine how much memory
1732 * needs to be read and/or written. Furthermore, ioctls are allowed
1733 * to dereference the passed pointer, so the parameter requires deep
1734 * copying but FUSE has no idea whatsoever about what to copy in or
1735 * out.
1736 *
1737 * This is solved by allowing FUSE server to retry ioctl with
1738 * necessary in/out iovecs. Let's assume the ioctl implementation
1739 * needs to read in the following structure.
1740 *
1741 * struct a {
1742 * char *buf;
1743 * size_t buflen;
1744 * }
1745 *
1746 * On the first callout to FUSE server, inarg->in_size and
1747 * inarg->out_size will be NULL; then, the server completes the ioctl
1748 * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
1749 * the actual iov array to
1750 *
1751 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) } }
1752 *
1753 * which tells FUSE to copy in the requested area and retry the ioctl.
1754 * On the second round, the server has access to the structure and
1755 * from that it can tell what to look for next, so on the invocation,
1756 * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
1757 *
1758 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) },
1759 * { .iov_base = a.buf, .iov_len = a.buflen } }
1760 *
1761 * FUSE will copy both struct a and the pointed buffer from the
1762 * process doing the ioctl and retry ioctl with both struct a and the
1763 * buffer.
1764 *
1765 * This time, FUSE server has everything it needs and completes ioctl
1766 * without FUSE_IOCTL_RETRY which finishes the ioctl call.
1767 *
1768 * Copying data out works the same way.
1769 *
1770 * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
1771 * automatically initializes in and out iovs by decoding @cmd with
1772 * _IOC_* macros and the server is not allowed to request RETRY. This
1773 * limits ioctl data transfers to well-formed ioctls and is the forced
1774 * behavior for all FUSE servers.
1775 */
1776long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
1777 unsigned int flags)
1778{
1779 struct fuse_file *ff = file->private_data;
1780 struct fuse_conn *fc = ff->fc;
1781 struct fuse_ioctl_in inarg = {
1782 .fh = ff->fh,
1783 .cmd = cmd,
1784 .arg = arg,
1785 .flags = flags
1786 };
1787 struct fuse_ioctl_out outarg;
1788 struct fuse_req *req = NULL;
1789 struct page **pages = NULL;
1790 struct iovec *iov_page = NULL;
1791 struct iovec *in_iov = NULL, *out_iov = NULL;
1792 unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
1793 size_t in_size, out_size, transferred;
1794 int err;
1795
1796#if BITS_PER_LONG == 32
1797 inarg.flags |= FUSE_IOCTL_32BIT;
1798#else
1799 if (flags & FUSE_IOCTL_COMPAT)
1800 inarg.flags |= FUSE_IOCTL_32BIT;
1801#endif
1802
1803 /* assume all the iovs returned by client always fits in a page */
1804 BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
1805
1806 err = -ENOMEM;
1807 pages = kzalloc(sizeof(pages[0]) * FUSE_MAX_PAGES_PER_REQ, GFP_KERNEL);
1808 iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
1809 if (!pages || !iov_page)
1810 goto out;
1811
1812 /*
1813 * If restricted, initialize IO parameters as encoded in @cmd.
1814 * RETRY from server is not allowed.
1815 */
1816 if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
1817 struct iovec *iov = iov_page;
1818
1819 iov->iov_base = (void __user *)arg;
1820 iov->iov_len = _IOC_SIZE(cmd);
1821
1822 if (_IOC_DIR(cmd) & _IOC_WRITE) {
1823 in_iov = iov;
1824 in_iovs = 1;
1825 }
1826
1827 if (_IOC_DIR(cmd) & _IOC_READ) {
1828 out_iov = iov;
1829 out_iovs = 1;
1830 }
1831 }
1832
1833 retry:
1834 inarg.in_size = in_size = iov_length(in_iov, in_iovs);
1835 inarg.out_size = out_size = iov_length(out_iov, out_iovs);
1836
1837 /*
1838 * Out data can be used either for actual out data or iovs,
1839 * make sure there always is at least one page.
1840 */
1841 out_size = max_t(size_t, out_size, PAGE_SIZE);
1842 max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
1843
1844 /* make sure there are enough buffer pages and init request with them */
1845 err = -ENOMEM;
1846 if (max_pages > FUSE_MAX_PAGES_PER_REQ)
1847 goto out;
1848 while (num_pages < max_pages) {
1849 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
1850 if (!pages[num_pages])
1851 goto out;
1852 num_pages++;
1853 }
1854
1855 req = fuse_get_req(fc);
1856 if (IS_ERR(req)) {
1857 err = PTR_ERR(req);
1858 req = NULL;
1859 goto out;
1860 }
1861 memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
1862 req->num_pages = num_pages;
1863
1864 /* okay, let's send it to the client */
1865 req->in.h.opcode = FUSE_IOCTL;
1866 req->in.h.nodeid = ff->nodeid;
1867 req->in.numargs = 1;
1868 req->in.args[0].size = sizeof(inarg);
1869 req->in.args[0].value = &inarg;
1870 if (in_size) {
1871 req->in.numargs++;
1872 req->in.args[1].size = in_size;
1873 req->in.argpages = 1;
1874
1875 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
1876 false);
1877 if (err)
1878 goto out;
1879 }
1880
1881 req->out.numargs = 2;
1882 req->out.args[0].size = sizeof(outarg);
1883 req->out.args[0].value = &outarg;
1884 req->out.args[1].size = out_size;
1885 req->out.argpages = 1;
1886 req->out.argvar = 1;
1887
1888 fuse_request_send(fc, req);
1889 err = req->out.h.error;
1890 transferred = req->out.args[1].size;
1891 fuse_put_request(fc, req);
1892 req = NULL;
1893 if (err)
1894 goto out;
1895
1896 /* did it ask for retry? */
1897 if (outarg.flags & FUSE_IOCTL_RETRY) {
1898 void *vaddr;
1899
1900 /* no retry if in restricted mode */
1901 err = -EIO;
1902 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
1903 goto out;
1904
1905 in_iovs = outarg.in_iovs;
1906 out_iovs = outarg.out_iovs;
1907
1908 /*
1909 * Make sure things are in boundary, separate checks
1910 * are to protect against overflow.
1911 */
1912 err = -ENOMEM;
1913 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
1914 out_iovs > FUSE_IOCTL_MAX_IOV ||
1915 in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
1916 goto out;
1917
1918 vaddr = kmap_atomic(pages[0], KM_USER0);
1919 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
1920 transferred, in_iovs + out_iovs,
1921 (flags & FUSE_IOCTL_COMPAT) != 0);
1922 kunmap_atomic(vaddr, KM_USER0);
1923 if (err)
1924 goto out;
1925
1926 in_iov = iov_page;
1927 out_iov = in_iov + in_iovs;
1928
1929 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
1930 if (err)
1931 goto out;
1932
1933 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
1934 if (err)
1935 goto out;
1936
1937 goto retry;
1938 }
1939
1940 err = -EIO;
1941 if (transferred > inarg.out_size)
1942 goto out;
1943
1944 err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
1945 out:
1946 if (req)
1947 fuse_put_request(fc, req);
1948 free_page((unsigned long) iov_page);
1949 while (num_pages)
1950 __free_page(pages[--num_pages]);
1951 kfree(pages);
1952
1953 return err ? err : outarg.result;
1954}
1955EXPORT_SYMBOL_GPL(fuse_do_ioctl);
1956
1957static long fuse_file_ioctl_common(struct file *file, unsigned int cmd,
1958 unsigned long arg, unsigned int flags)
1959{
1960 struct inode *inode = file->f_dentry->d_inode;
1961 struct fuse_conn *fc = get_fuse_conn(inode);
1962
1963 if (!fuse_allow_task(fc, current))
1964 return -EACCES;
1965
1966 if (is_bad_inode(inode))
1967 return -EIO;
1968
1969 return fuse_do_ioctl(file, cmd, arg, flags);
1970}
1971
1972static long fuse_file_ioctl(struct file *file, unsigned int cmd,
1973 unsigned long arg)
1974{
1975 return fuse_file_ioctl_common(file, cmd, arg, 0);
1976}
1977
1978static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
1979 unsigned long arg)
1980{
1981 return fuse_file_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
1982}
1983
1984/*
1985 * All files which have been polled are linked to RB tree
1986 * fuse_conn->polled_files which is indexed by kh. Walk the tree and
1987 * find the matching one.
1988 */
1989static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
1990 struct rb_node **parent_out)
1991{
1992 struct rb_node **link = &fc->polled_files.rb_node;
1993 struct rb_node *last = NULL;
1994
1995 while (*link) {
1996 struct fuse_file *ff;
1997
1998 last = *link;
1999 ff = rb_entry(last, struct fuse_file, polled_node);
2000
2001 if (kh < ff->kh)
2002 link = &last->rb_left;
2003 else if (kh > ff->kh)
2004 link = &last->rb_right;
2005 else
2006 return link;
2007 }
2008
2009 if (parent_out)
2010 *parent_out = last;
2011 return link;
2012}
2013
2014/*
2015 * The file is about to be polled. Make sure it's on the polled_files
2016 * RB tree. Note that files once added to the polled_files tree are
2017 * not removed before the file is released. This is because a file
2018 * polled once is likely to be polled again.
2019 */
2020static void fuse_register_polled_file(struct fuse_conn *fc,
2021 struct fuse_file *ff)
2022{
2023 spin_lock(&fc->lock);
2024 if (RB_EMPTY_NODE(&ff->polled_node)) {
2025 struct rb_node **link, *parent;
2026
2027 link = fuse_find_polled_node(fc, ff->kh, &parent);
2028 BUG_ON(*link);
2029 rb_link_node(&ff->polled_node, parent, link);
2030 rb_insert_color(&ff->polled_node, &fc->polled_files);
2031 }
2032 spin_unlock(&fc->lock);
2033}
2034
2035unsigned fuse_file_poll(struct file *file, poll_table *wait)
2036{
2037 struct fuse_file *ff = file->private_data;
2038 struct fuse_conn *fc = ff->fc;
2039 struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2040 struct fuse_poll_out outarg;
2041 struct fuse_req *req;
2042 int err;
2043
2044 if (fc->no_poll)
2045 return DEFAULT_POLLMASK;
2046
2047 poll_wait(file, &ff->poll_wait, wait);
2048
2049 /*
2050 * Ask for notification iff there's someone waiting for it.
2051 * The client may ignore the flag and always notify.
2052 */
2053 if (waitqueue_active(&ff->poll_wait)) {
2054 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2055 fuse_register_polled_file(fc, ff);
2056 }
2057
2058 req = fuse_get_req(fc);
2059 if (IS_ERR(req))
2060 return POLLERR;
2061
2062 req->in.h.opcode = FUSE_POLL;
2063 req->in.h.nodeid = ff->nodeid;
2064 req->in.numargs = 1;
2065 req->in.args[0].size = sizeof(inarg);
2066 req->in.args[0].value = &inarg;
2067 req->out.numargs = 1;
2068 req->out.args[0].size = sizeof(outarg);
2069 req->out.args[0].value = &outarg;
2070 fuse_request_send(fc, req);
2071 err = req->out.h.error;
2072 fuse_put_request(fc, req);
2073
2074 if (!err)
2075 return outarg.revents;
2076 if (err == -ENOSYS) {
2077 fc->no_poll = 1;
2078 return DEFAULT_POLLMASK;
2079 }
2080 return POLLERR;
2081}
2082EXPORT_SYMBOL_GPL(fuse_file_poll);
2083
2084/*
2085 * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2086 * wakes up the poll waiters.
2087 */
2088int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2089 struct fuse_notify_poll_wakeup_out *outarg)
2090{
2091 u64 kh = outarg->kh;
2092 struct rb_node **link;
2093
2094 spin_lock(&fc->lock);
2095
2096 link = fuse_find_polled_node(fc, kh, NULL);
2097 if (*link) {
2098 struct fuse_file *ff;
2099
2100 ff = rb_entry(*link, struct fuse_file, polled_node);
2101 wake_up_interruptible_sync(&ff->poll_wait);
2102 }
2103
2104 spin_unlock(&fc->lock);
2105 return 0;
2106}
2107
2108static const struct file_operations fuse_file_operations = {
2109 .llseek = fuse_file_llseek,
2110 .read = do_sync_read,
2111 .aio_read = fuse_file_aio_read,
2112 .write = do_sync_write,
2113 .aio_write = fuse_file_aio_write,
2114 .mmap = fuse_file_mmap,
2115 .open = fuse_open,
2116 .flush = fuse_flush,
2117 .release = fuse_release,
2118 .fsync = fuse_fsync,
2119 .lock = fuse_file_lock,
2120 .flock = fuse_file_flock,
2121 .splice_read = generic_file_splice_read,
2122 .unlocked_ioctl = fuse_file_ioctl,
2123 .compat_ioctl = fuse_file_compat_ioctl,
2124 .poll = fuse_file_poll,
2125};
2126
2127static const struct file_operations fuse_direct_io_file_operations = {
2128 .llseek = fuse_file_llseek,
2129 .read = fuse_direct_read,
2130 .write = fuse_direct_write,
2131 .mmap = fuse_direct_mmap,
2132 .open = fuse_open,
2133 .flush = fuse_flush,
2134 .release = fuse_release,
2135 .fsync = fuse_fsync,
2136 .lock = fuse_file_lock,
2137 .flock = fuse_file_flock,
2138 .unlocked_ioctl = fuse_file_ioctl,
2139 .compat_ioctl = fuse_file_compat_ioctl,
2140 .poll = fuse_file_poll,
2141 /* no splice_read */
2142};
2143
2144static const struct address_space_operations fuse_file_aops = {
2145 .readpage = fuse_readpage,
2146 .writepage = fuse_writepage,
2147 .launder_page = fuse_launder_page,
2148 .readpages = fuse_readpages,
2149 .set_page_dirty = __set_page_dirty_nobuffers,
2150 .bmap = fuse_bmap,
2151};
2152
2153void fuse_init_file_inode(struct inode *inode)
2154{
2155 inode->i_fop = &fuse_file_operations;
2156 inode->i_data.a_ops = &fuse_file_aops;
2157}