2 FUSE: Filesystem in Userspace
3 Copyright (C) 2001-2008 Miklos Szeredi <miklos@szeredi.hu>
5 This program can be distributed under the terms of the GNU GPL.
11 #include <linux/pagemap.h>
12 #include <linux/slab.h>
13 #include <linux/kernel.h>
14 #include <linux/sched.h>
15 #include <linux/sched/signal.h>
16 #include <linux/module.h>
17 #include <linux/compat.h>
18 #include <linux/swap.h>
19 #include <linux/falloc.h>
20 #include <linux/uio.h>
23 static struct page **fuse_pages_alloc(unsigned int npages, gfp_t flags,
24 struct fuse_page_desc **desc)
28 pages = kzalloc(npages * (sizeof(struct page *) +
29 sizeof(struct fuse_page_desc)), flags);
30 *desc = (void *) (pages + npages);
35 static int fuse_send_open(struct fuse_mount *fm, u64 nodeid, struct file *file,
36 int opcode, struct fuse_open_out *outargp)
38 struct fuse_open_in inarg;
41 memset(&inarg, 0, sizeof(inarg));
42 inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
43 if (!fm->fc->atomic_o_trunc)
44 inarg.flags &= ~O_TRUNC;
48 args.in_args[0].size = sizeof(inarg);
49 args.in_args[0].value = &inarg;
51 args.out_args[0].size = sizeof(*outargp);
52 args.out_args[0].value = outargp;
54 return fuse_simple_request(fm, &args);
57 struct fuse_release_args {
58 struct fuse_args args;
59 struct fuse_release_in inarg;
63 struct fuse_file *fuse_file_alloc(struct fuse_mount *fm)
67 ff = kzalloc(sizeof(struct fuse_file), GFP_KERNEL_ACCOUNT);
72 ff->release_args = kzalloc(sizeof(*ff->release_args),
74 if (!ff->release_args) {
79 INIT_LIST_HEAD(&ff->write_entry);
80 mutex_init(&ff->readdir.lock);
81 refcount_set(&ff->count, 1);
82 RB_CLEAR_NODE(&ff->polled_node);
83 init_waitqueue_head(&ff->poll_wait);
85 ff->kh = atomic64_inc_return(&fm->fc->khctr);
90 void fuse_file_free(struct fuse_file *ff)
92 kfree(ff->release_args);
93 mutex_destroy(&ff->readdir.lock);
97 static struct fuse_file *fuse_file_get(struct fuse_file *ff)
99 refcount_inc(&ff->count);
103 static void fuse_release_end(struct fuse_mount *fm, struct fuse_args *args,
106 struct fuse_release_args *ra = container_of(args, typeof(*ra), args);
112 static void fuse_file_put(struct fuse_file *ff, bool sync, bool isdir)
114 if (refcount_dec_and_test(&ff->count)) {
115 struct fuse_args *args = &ff->release_args->args;
117 if (isdir ? ff->fm->fc->no_opendir : ff->fm->fc->no_open) {
118 /* Do nothing when client does not implement 'open' */
119 fuse_release_end(ff->fm, args, 0);
121 fuse_simple_request(ff->fm, args);
122 fuse_release_end(ff->fm, args, 0);
124 args->end = fuse_release_end;
125 if (fuse_simple_background(ff->fm, args,
126 GFP_KERNEL | __GFP_NOFAIL))
127 fuse_release_end(ff->fm, args, -ENOTCONN);
133 int fuse_do_open(struct fuse_mount *fm, u64 nodeid, struct file *file,
136 struct fuse_conn *fc = fm->fc;
137 struct fuse_file *ff;
138 int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
140 ff = fuse_file_alloc(fm);
145 /* Default for no-open */
146 ff->open_flags = FOPEN_KEEP_CACHE | (isdir ? FOPEN_CACHE_DIR : 0);
147 if (isdir ? !fc->no_opendir : !fc->no_open) {
148 struct fuse_open_out outarg;
151 err = fuse_send_open(fm, nodeid, file, opcode, &outarg);
154 ff->open_flags = outarg.open_flags;
156 } else if (err != -ENOSYS) {
168 ff->open_flags &= ~FOPEN_DIRECT_IO;
171 file->private_data = ff;
175 EXPORT_SYMBOL_GPL(fuse_do_open);
177 static void fuse_link_write_file(struct file *file)
179 struct inode *inode = file_inode(file);
180 struct fuse_inode *fi = get_fuse_inode(inode);
181 struct fuse_file *ff = file->private_data;
183 * file may be written through mmap, so chain it onto the
184 * inodes's write_file list
186 spin_lock(&fi->lock);
187 if (list_empty(&ff->write_entry))
188 list_add(&ff->write_entry, &fi->write_files);
189 spin_unlock(&fi->lock);
192 void fuse_finish_open(struct inode *inode, struct file *file)
194 struct fuse_file *ff = file->private_data;
195 struct fuse_conn *fc = get_fuse_conn(inode);
197 if (ff->open_flags & FOPEN_STREAM)
198 stream_open(inode, file);
199 else if (ff->open_flags & FOPEN_NONSEEKABLE)
200 nonseekable_open(inode, file);
202 if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
203 struct fuse_inode *fi = get_fuse_inode(inode);
205 spin_lock(&fi->lock);
206 fi->attr_version = atomic64_inc_return(&fc->attr_version);
207 i_size_write(inode, 0);
208 spin_unlock(&fi->lock);
209 fuse_invalidate_attr(inode);
210 if (fc->writeback_cache)
211 file_update_time(file);
213 if ((file->f_mode & FMODE_WRITE) && fc->writeback_cache)
214 fuse_link_write_file(file);
217 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
219 struct fuse_mount *fm = get_fuse_mount(inode);
220 struct fuse_conn *fc = fm->fc;
222 bool is_wb_truncate = (file->f_flags & O_TRUNC) &&
223 fc->atomic_o_trunc &&
225 bool dax_truncate = (file->f_flags & O_TRUNC) &&
226 fc->atomic_o_trunc && FUSE_IS_DAX(inode);
228 if (fuse_is_bad(inode))
231 err = generic_file_open(inode, file);
235 if (is_wb_truncate || dax_truncate)
239 down_write(&get_fuse_inode(inode)->i_mmap_sem);
240 err = fuse_dax_break_layouts(inode, 0, 0);
242 goto out_inode_unlock;
245 if (is_wb_truncate || dax_truncate)
246 fuse_set_nowrite(inode);
248 err = fuse_do_open(fm, get_node_id(inode), file, isdir);
250 fuse_finish_open(inode, file);
252 if (is_wb_truncate || dax_truncate)
253 fuse_release_nowrite(inode);
255 struct fuse_file *ff = file->private_data;
257 if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC))
258 truncate_pagecache(inode, 0);
259 else if (!(ff->open_flags & FOPEN_KEEP_CACHE))
260 invalidate_inode_pages2(inode->i_mapping);
263 up_write(&get_fuse_inode(inode)->i_mmap_sem);
266 if (is_wb_truncate || dax_truncate)
272 static void fuse_prepare_release(struct fuse_inode *fi, struct fuse_file *ff,
273 int flags, int opcode)
275 struct fuse_conn *fc = ff->fm->fc;
276 struct fuse_release_args *ra = ff->release_args;
278 /* Inode is NULL on error path of fuse_create_open() */
280 spin_lock(&fi->lock);
281 list_del(&ff->write_entry);
282 spin_unlock(&fi->lock);
284 spin_lock(&fc->lock);
285 if (!RB_EMPTY_NODE(&ff->polled_node))
286 rb_erase(&ff->polled_node, &fc->polled_files);
287 spin_unlock(&fc->lock);
289 wake_up_interruptible_all(&ff->poll_wait);
291 ra->inarg.fh = ff->fh;
292 ra->inarg.flags = flags;
293 ra->args.in_numargs = 1;
294 ra->args.in_args[0].size = sizeof(struct fuse_release_in);
295 ra->args.in_args[0].value = &ra->inarg;
296 ra->args.opcode = opcode;
297 ra->args.nodeid = ff->nodeid;
298 ra->args.force = true;
299 ra->args.nocreds = true;
302 void fuse_release_common(struct file *file, bool isdir)
304 struct fuse_inode *fi = get_fuse_inode(file_inode(file));
305 struct fuse_file *ff = file->private_data;
306 struct fuse_release_args *ra = ff->release_args;
307 int opcode = isdir ? FUSE_RELEASEDIR : FUSE_RELEASE;
309 fuse_prepare_release(fi, ff, file->f_flags, opcode);
312 ra->inarg.release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
313 ra->inarg.lock_owner = fuse_lock_owner_id(ff->fm->fc,
316 /* Hold inode until release is finished */
317 ra->inode = igrab(file_inode(file));
320 * Normally this will send the RELEASE request, however if
321 * some asynchronous READ or WRITE requests are outstanding,
322 * the sending will be delayed.
324 * Make the release synchronous if this is a fuseblk mount,
325 * synchronous RELEASE is allowed (and desirable) in this case
326 * because the server can be trusted not to screw up.
328 fuse_file_put(ff, ff->fm->fc->destroy, isdir);
331 static int fuse_open(struct inode *inode, struct file *file)
333 return fuse_open_common(inode, file, false);
336 static int fuse_release(struct inode *inode, struct file *file)
338 struct fuse_conn *fc = get_fuse_conn(inode);
340 /* see fuse_vma_close() for !writeback_cache case */
341 if (fc->writeback_cache)
342 write_inode_now(inode, 1);
344 fuse_release_common(file, false);
346 /* return value is ignored by VFS */
350 void fuse_sync_release(struct fuse_inode *fi, struct fuse_file *ff, int flags)
352 WARN_ON(refcount_read(&ff->count) > 1);
353 fuse_prepare_release(fi, ff, flags, FUSE_RELEASE);
355 * iput(NULL) is a no-op and since the refcount is 1 and everything's
356 * synchronous, we are fine with not doing igrab() here"
358 fuse_file_put(ff, true, false);
360 EXPORT_SYMBOL_GPL(fuse_sync_release);
363 * Scramble the ID space with XTEA, so that the value of the files_struct
364 * pointer is not exposed to userspace.
366 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
368 u32 *k = fc->scramble_key;
369 u64 v = (unsigned long) id;
375 for (i = 0; i < 32; i++) {
376 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
378 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
381 return (u64) v0 + ((u64) v1 << 32);
384 struct fuse_writepage_args {
385 struct fuse_io_args ia;
386 struct rb_node writepages_entry;
387 struct list_head queue_entry;
388 struct fuse_writepage_args *next;
392 static struct fuse_writepage_args *fuse_find_writeback(struct fuse_inode *fi,
393 pgoff_t idx_from, pgoff_t idx_to)
397 n = fi->writepages.rb_node;
400 struct fuse_writepage_args *wpa;
403 wpa = rb_entry(n, struct fuse_writepage_args, writepages_entry);
404 WARN_ON(get_fuse_inode(wpa->inode) != fi);
405 curr_index = wpa->ia.write.in.offset >> PAGE_SHIFT;
406 if (idx_from >= curr_index + wpa->ia.ap.num_pages)
408 else if (idx_to < curr_index)
417 * Check if any page in a range is under writeback
419 * This is currently done by walking the list of writepage requests
420 * for the inode, which can be pretty inefficient.
422 static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
425 struct fuse_inode *fi = get_fuse_inode(inode);
428 spin_lock(&fi->lock);
429 found = fuse_find_writeback(fi, idx_from, idx_to);
430 spin_unlock(&fi->lock);
435 static inline bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
437 return fuse_range_is_writeback(inode, index, index);
441 * Wait for page writeback to be completed.
443 * Since fuse doesn't rely on the VM writeback tracking, this has to
444 * use some other means.
446 static void fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
448 struct fuse_inode *fi = get_fuse_inode(inode);
450 wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
454 * Wait for all pending writepages on the inode to finish.
456 * This is currently done by blocking further writes with FUSE_NOWRITE
457 * and waiting for all sent writes to complete.
459 * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
460 * could conflict with truncation.
462 static void fuse_sync_writes(struct inode *inode)
464 fuse_set_nowrite(inode);
465 fuse_release_nowrite(inode);
468 static int fuse_flush(struct file *file, fl_owner_t id)
470 struct inode *inode = file_inode(file);
471 struct fuse_mount *fm = get_fuse_mount(inode);
472 struct fuse_file *ff = file->private_data;
473 struct fuse_flush_in inarg;
477 if (fuse_is_bad(inode))
480 err = write_inode_now(inode, 1);
485 fuse_sync_writes(inode);
488 err = filemap_check_errors(file->f_mapping);
493 if (fm->fc->no_flush)
496 memset(&inarg, 0, sizeof(inarg));
498 inarg.lock_owner = fuse_lock_owner_id(fm->fc, id);
499 args.opcode = FUSE_FLUSH;
500 args.nodeid = get_node_id(inode);
502 args.in_args[0].size = sizeof(inarg);
503 args.in_args[0].value = &inarg;
506 err = fuse_simple_request(fm, &args);
507 if (err == -ENOSYS) {
508 fm->fc->no_flush = 1;
514 * In memory i_blocks is not maintained by fuse, if writeback cache is
515 * enabled, i_blocks from cached attr may not be accurate.
517 if (!err && fm->fc->writeback_cache)
518 fuse_invalidate_attr(inode);
522 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
523 int datasync, int opcode)
525 struct inode *inode = file->f_mapping->host;
526 struct fuse_mount *fm = get_fuse_mount(inode);
527 struct fuse_file *ff = file->private_data;
529 struct fuse_fsync_in inarg;
531 memset(&inarg, 0, sizeof(inarg));
533 inarg.fsync_flags = datasync ? FUSE_FSYNC_FDATASYNC : 0;
534 args.opcode = opcode;
535 args.nodeid = get_node_id(inode);
537 args.in_args[0].size = sizeof(inarg);
538 args.in_args[0].value = &inarg;
539 return fuse_simple_request(fm, &args);
542 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
545 struct inode *inode = file->f_mapping->host;
546 struct fuse_conn *fc = get_fuse_conn(inode);
549 if (fuse_is_bad(inode))
555 * Start writeback against all dirty pages of the inode, then
556 * wait for all outstanding writes, before sending the FSYNC
559 err = file_write_and_wait_range(file, start, end);
563 fuse_sync_writes(inode);
566 * Due to implementation of fuse writeback
567 * file_write_and_wait_range() does not catch errors.
568 * We have to do this directly after fuse_sync_writes()
570 err = file_check_and_advance_wb_err(file);
574 err = sync_inode_metadata(inode, 1);
581 err = fuse_fsync_common(file, start, end, datasync, FUSE_FSYNC);
582 if (err == -ENOSYS) {
592 void fuse_read_args_fill(struct fuse_io_args *ia, struct file *file, loff_t pos,
593 size_t count, int opcode)
595 struct fuse_file *ff = file->private_data;
596 struct fuse_args *args = &ia->ap.args;
598 ia->read.in.fh = ff->fh;
599 ia->read.in.offset = pos;
600 ia->read.in.size = count;
601 ia->read.in.flags = file->f_flags;
602 args->opcode = opcode;
603 args->nodeid = ff->nodeid;
604 args->in_numargs = 1;
605 args->in_args[0].size = sizeof(ia->read.in);
606 args->in_args[0].value = &ia->read.in;
607 args->out_argvar = true;
608 args->out_numargs = 1;
609 args->out_args[0].size = count;
612 static void fuse_release_user_pages(struct fuse_args_pages *ap,
617 for (i = 0; i < ap->num_pages; i++) {
619 set_page_dirty_lock(ap->pages[i]);
620 put_page(ap->pages[i]);
624 static void fuse_io_release(struct kref *kref)
626 kfree(container_of(kref, struct fuse_io_priv, refcnt));
629 static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
634 if (io->bytes >= 0 && io->write)
637 return io->bytes < 0 ? io->size : io->bytes;
641 * In case of short read, the caller sets 'pos' to the position of
642 * actual end of fuse request in IO request. Otherwise, if bytes_requested
643 * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
646 * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
647 * both submitted asynchronously. The first of them was ACKed by userspace as
648 * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
649 * second request was ACKed as short, e.g. only 1K was read, resulting in
652 * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
653 * will be equal to the length of the longest contiguous fragment of
654 * transferred data starting from the beginning of IO request.
656 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
660 spin_lock(&io->lock);
662 io->err = io->err ? : err;
663 else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
667 if (!left && io->blocking)
669 spin_unlock(&io->lock);
671 if (!left && !io->blocking) {
672 ssize_t res = fuse_get_res_by_io(io);
675 struct inode *inode = file_inode(io->iocb->ki_filp);
676 struct fuse_conn *fc = get_fuse_conn(inode);
677 struct fuse_inode *fi = get_fuse_inode(inode);
679 spin_lock(&fi->lock);
680 fi->attr_version = atomic64_inc_return(&fc->attr_version);
681 spin_unlock(&fi->lock);
684 io->iocb->ki_complete(io->iocb, res, 0);
687 kref_put(&io->refcnt, fuse_io_release);
690 static struct fuse_io_args *fuse_io_alloc(struct fuse_io_priv *io,
693 struct fuse_io_args *ia;
695 ia = kzalloc(sizeof(*ia), GFP_KERNEL);
698 ia->ap.pages = fuse_pages_alloc(npages, GFP_KERNEL,
708 static void fuse_io_free(struct fuse_io_args *ia)
714 static void fuse_aio_complete_req(struct fuse_mount *fm, struct fuse_args *args,
717 struct fuse_io_args *ia = container_of(args, typeof(*ia), ap.args);
718 struct fuse_io_priv *io = ia->io;
721 fuse_release_user_pages(&ia->ap, io->should_dirty);
725 } else if (io->write) {
726 if (ia->write.out.size > ia->write.in.size) {
728 } else if (ia->write.in.size != ia->write.out.size) {
729 pos = ia->write.in.offset - io->offset +
733 u32 outsize = args->out_args[0].size;
735 if (ia->read.in.size != outsize)
736 pos = ia->read.in.offset - io->offset + outsize;
739 fuse_aio_complete(io, err, pos);
743 static ssize_t fuse_async_req_send(struct fuse_mount *fm,
744 struct fuse_io_args *ia, size_t num_bytes)
747 struct fuse_io_priv *io = ia->io;
749 spin_lock(&io->lock);
750 kref_get(&io->refcnt);
751 io->size += num_bytes;
753 spin_unlock(&io->lock);
755 ia->ap.args.end = fuse_aio_complete_req;
756 ia->ap.args.may_block = io->should_dirty;
757 err = fuse_simple_background(fm, &ia->ap.args, GFP_KERNEL);
759 fuse_aio_complete_req(fm, &ia->ap.args, err);
764 static ssize_t fuse_send_read(struct fuse_io_args *ia, loff_t pos, size_t count,
767 struct file *file = ia->io->iocb->ki_filp;
768 struct fuse_file *ff = file->private_data;
769 struct fuse_mount *fm = ff->fm;
771 fuse_read_args_fill(ia, file, pos, count, FUSE_READ);
773 ia->read.in.read_flags |= FUSE_READ_LOCKOWNER;
774 ia->read.in.lock_owner = fuse_lock_owner_id(fm->fc, owner);
778 return fuse_async_req_send(fm, ia, count);
780 return fuse_simple_request(fm, &ia->ap.args);
783 static void fuse_read_update_size(struct inode *inode, loff_t size,
786 struct fuse_conn *fc = get_fuse_conn(inode);
787 struct fuse_inode *fi = get_fuse_inode(inode);
789 spin_lock(&fi->lock);
790 if (attr_ver >= fi->attr_version && size < inode->i_size &&
791 !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
792 fi->attr_version = atomic64_inc_return(&fc->attr_version);
793 i_size_write(inode, size);
795 spin_unlock(&fi->lock);
798 static void fuse_short_read(struct inode *inode, u64 attr_ver, size_t num_read,
799 struct fuse_args_pages *ap)
801 struct fuse_conn *fc = get_fuse_conn(inode);
803 if (fc->writeback_cache) {
805 * A hole in a file. Some data after the hole are in page cache,
806 * but have not reached the client fs yet. So, the hole is not
810 int start_idx = num_read >> PAGE_SHIFT;
811 size_t off = num_read & (PAGE_SIZE - 1);
813 for (i = start_idx; i < ap->num_pages; i++) {
814 zero_user_segment(ap->pages[i], off, PAGE_SIZE);
818 loff_t pos = page_offset(ap->pages[0]) + num_read;
819 fuse_read_update_size(inode, pos, attr_ver);
823 static int fuse_do_readpage(struct file *file, struct page *page)
825 struct inode *inode = page->mapping->host;
826 struct fuse_mount *fm = get_fuse_mount(inode);
827 loff_t pos = page_offset(page);
828 struct fuse_page_desc desc = { .length = PAGE_SIZE };
829 struct fuse_io_args ia = {
830 .ap.args.page_zeroing = true,
831 .ap.args.out_pages = true,
840 * Page writeback can extend beyond the lifetime of the
841 * page-cache page, so make sure we read a properly synced
844 fuse_wait_on_page_writeback(inode, page->index);
846 attr_ver = fuse_get_attr_version(fm->fc);
848 /* Don't overflow end offset */
849 if (pos + (desc.length - 1) == LLONG_MAX)
852 fuse_read_args_fill(&ia, file, pos, desc.length, FUSE_READ);
853 res = fuse_simple_request(fm, &ia.ap.args);
857 * Short read means EOF. If file size is larger, truncate it
859 if (res < desc.length)
860 fuse_short_read(inode, attr_ver, res, &ia.ap);
862 SetPageUptodate(page);
867 static int fuse_readpage(struct file *file, struct page *page)
869 struct inode *inode = page->mapping->host;
873 if (fuse_is_bad(inode))
876 err = fuse_do_readpage(file, page);
877 fuse_invalidate_atime(inode);
883 static void fuse_readpages_end(struct fuse_mount *fm, struct fuse_args *args,
887 struct fuse_io_args *ia = container_of(args, typeof(*ia), ap.args);
888 struct fuse_args_pages *ap = &ia->ap;
889 size_t count = ia->read.in.size;
890 size_t num_read = args->out_args[0].size;
891 struct address_space *mapping = NULL;
893 for (i = 0; mapping == NULL && i < ap->num_pages; i++)
894 mapping = ap->pages[i]->mapping;
897 struct inode *inode = mapping->host;
900 * Short read means EOF. If file size is larger, truncate it
902 if (!err && num_read < count)
903 fuse_short_read(inode, ia->read.attr_ver, num_read, ap);
905 fuse_invalidate_atime(inode);
908 for (i = 0; i < ap->num_pages; i++) {
909 struct page *page = ap->pages[i];
912 SetPageUptodate(page);
919 fuse_file_put(ia->ff, false, false);
924 static void fuse_send_readpages(struct fuse_io_args *ia, struct file *file)
926 struct fuse_file *ff = file->private_data;
927 struct fuse_mount *fm = ff->fm;
928 struct fuse_args_pages *ap = &ia->ap;
929 loff_t pos = page_offset(ap->pages[0]);
930 size_t count = ap->num_pages << PAGE_SHIFT;
934 ap->args.out_pages = true;
935 ap->args.page_zeroing = true;
936 ap->args.page_replace = true;
938 /* Don't overflow end offset */
939 if (pos + (count - 1) == LLONG_MAX) {
941 ap->descs[ap->num_pages - 1].length--;
943 WARN_ON((loff_t) (pos + count) < 0);
945 fuse_read_args_fill(ia, file, pos, count, FUSE_READ);
946 ia->read.attr_ver = fuse_get_attr_version(fm->fc);
947 if (fm->fc->async_read) {
948 ia->ff = fuse_file_get(ff);
949 ap->args.end = fuse_readpages_end;
950 err = fuse_simple_background(fm, &ap->args, GFP_KERNEL);
954 res = fuse_simple_request(fm, &ap->args);
955 err = res < 0 ? res : 0;
957 fuse_readpages_end(fm, &ap->args, err);
960 static void fuse_readahead(struct readahead_control *rac)
962 struct inode *inode = rac->mapping->host;
963 struct fuse_conn *fc = get_fuse_conn(inode);
964 unsigned int i, max_pages, nr_pages = 0;
966 if (fuse_is_bad(inode))
969 max_pages = min_t(unsigned int, fc->max_pages,
970 fc->max_read / PAGE_SIZE);
973 struct fuse_io_args *ia;
974 struct fuse_args_pages *ap;
976 nr_pages = readahead_count(rac) - nr_pages;
977 if (nr_pages > max_pages)
978 nr_pages = max_pages;
981 ia = fuse_io_alloc(NULL, nr_pages);
985 nr_pages = __readahead_batch(rac, ap->pages, nr_pages);
986 for (i = 0; i < nr_pages; i++) {
987 fuse_wait_on_page_writeback(inode,
988 readahead_index(rac) + i);
989 ap->descs[i].length = PAGE_SIZE;
991 ap->num_pages = nr_pages;
992 fuse_send_readpages(ia, rac->file);
996 static ssize_t fuse_cache_read_iter(struct kiocb *iocb, struct iov_iter *to)
998 struct inode *inode = iocb->ki_filp->f_mapping->host;
999 struct fuse_conn *fc = get_fuse_conn(inode);
1002 * In auto invalidate mode, always update attributes on read.
1003 * Otherwise, only update if we attempt to read past EOF (to ensure
1004 * i_size is up to date).
1006 if (fc->auto_inval_data ||
1007 (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
1009 err = fuse_update_attributes(inode, iocb->ki_filp);
1014 return generic_file_read_iter(iocb, to);
1017 static void fuse_write_args_fill(struct fuse_io_args *ia, struct fuse_file *ff,
1018 loff_t pos, size_t count)
1020 struct fuse_args *args = &ia->ap.args;
1022 ia->write.in.fh = ff->fh;
1023 ia->write.in.offset = pos;
1024 ia->write.in.size = count;
1025 args->opcode = FUSE_WRITE;
1026 args->nodeid = ff->nodeid;
1027 args->in_numargs = 2;
1028 if (ff->fm->fc->minor < 9)
1029 args->in_args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
1031 args->in_args[0].size = sizeof(ia->write.in);
1032 args->in_args[0].value = &ia->write.in;
1033 args->in_args[1].size = count;
1034 args->out_numargs = 1;
1035 args->out_args[0].size = sizeof(ia->write.out);
1036 args->out_args[0].value = &ia->write.out;
1039 static unsigned int fuse_write_flags(struct kiocb *iocb)
1041 unsigned int flags = iocb->ki_filp->f_flags;
1043 if (iocb->ki_flags & IOCB_DSYNC)
1045 if (iocb->ki_flags & IOCB_SYNC)
1051 static ssize_t fuse_send_write(struct fuse_io_args *ia, loff_t pos,
1052 size_t count, fl_owner_t owner)
1054 struct kiocb *iocb = ia->io->iocb;
1055 struct file *file = iocb->ki_filp;
1056 struct fuse_file *ff = file->private_data;
1057 struct fuse_mount *fm = ff->fm;
1058 struct fuse_write_in *inarg = &ia->write.in;
1061 fuse_write_args_fill(ia, ff, pos, count);
1062 inarg->flags = fuse_write_flags(iocb);
1063 if (owner != NULL) {
1064 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
1065 inarg->lock_owner = fuse_lock_owner_id(fm->fc, owner);
1069 return fuse_async_req_send(fm, ia, count);
1071 err = fuse_simple_request(fm, &ia->ap.args);
1072 if (!err && ia->write.out.size > count)
1075 return err ?: ia->write.out.size;
1078 bool fuse_write_update_size(struct inode *inode, loff_t pos)
1080 struct fuse_conn *fc = get_fuse_conn(inode);
1081 struct fuse_inode *fi = get_fuse_inode(inode);
1084 spin_lock(&fi->lock);
1085 fi->attr_version = atomic64_inc_return(&fc->attr_version);
1086 if (pos > inode->i_size) {
1087 i_size_write(inode, pos);
1090 spin_unlock(&fi->lock);
1095 static ssize_t fuse_send_write_pages(struct fuse_io_args *ia,
1096 struct kiocb *iocb, struct inode *inode,
1097 loff_t pos, size_t count)
1099 struct fuse_args_pages *ap = &ia->ap;
1100 struct file *file = iocb->ki_filp;
1101 struct fuse_file *ff = file->private_data;
1102 struct fuse_mount *fm = ff->fm;
1103 unsigned int offset, i;
1107 for (i = 0; i < ap->num_pages; i++)
1108 fuse_wait_on_page_writeback(inode, ap->pages[i]->index);
1110 fuse_write_args_fill(ia, ff, pos, count);
1111 ia->write.in.flags = fuse_write_flags(iocb);
1113 err = fuse_simple_request(fm, &ap->args);
1114 if (!err && ia->write.out.size > count)
1117 short_write = ia->write.out.size < count;
1118 offset = ap->descs[0].offset;
1119 count = ia->write.out.size;
1120 for (i = 0; i < ap->num_pages; i++) {
1121 struct page *page = ap->pages[i];
1124 ClearPageUptodate(page);
1126 if (count >= PAGE_SIZE - offset)
1127 count -= PAGE_SIZE - offset;
1130 ClearPageUptodate(page);
1135 if (ia->write.page_locked && (i == ap->num_pages - 1))
1143 static ssize_t fuse_fill_write_pages(struct fuse_io_args *ia,
1144 struct address_space *mapping,
1145 struct iov_iter *ii, loff_t pos,
1146 unsigned int max_pages)
1148 struct fuse_args_pages *ap = &ia->ap;
1149 struct fuse_conn *fc = get_fuse_conn(mapping->host);
1150 unsigned offset = pos & (PAGE_SIZE - 1);
1154 ap->args.in_pages = true;
1155 ap->descs[0].offset = offset;
1160 pgoff_t index = pos >> PAGE_SHIFT;
1161 size_t bytes = min_t(size_t, PAGE_SIZE - offset,
1162 iov_iter_count(ii));
1164 bytes = min_t(size_t, bytes, fc->max_write - count);
1168 if (iov_iter_fault_in_readable(ii, bytes))
1172 page = grab_cache_page_write_begin(mapping, index, 0);
1176 if (mapping_writably_mapped(mapping))
1177 flush_dcache_page(page);
1179 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
1180 flush_dcache_page(page);
1182 iov_iter_advance(ii, tmp);
1186 bytes = min(bytes, iov_iter_single_seg_count(ii));
1191 ap->pages[ap->num_pages] = page;
1192 ap->descs[ap->num_pages].length = tmp;
1198 if (offset == PAGE_SIZE)
1201 /* If we copied full page, mark it uptodate */
1202 if (tmp == PAGE_SIZE)
1203 SetPageUptodate(page);
1205 if (PageUptodate(page)) {
1208 ia->write.page_locked = true;
1211 if (!fc->big_writes)
1213 } while (iov_iter_count(ii) && count < fc->max_write &&
1214 ap->num_pages < max_pages && offset == 0);
1216 return count > 0 ? count : err;
1219 static inline unsigned int fuse_wr_pages(loff_t pos, size_t len,
1220 unsigned int max_pages)
1222 return min_t(unsigned int,
1223 ((pos + len - 1) >> PAGE_SHIFT) -
1224 (pos >> PAGE_SHIFT) + 1,
1228 static ssize_t fuse_perform_write(struct kiocb *iocb,
1229 struct address_space *mapping,
1230 struct iov_iter *ii, loff_t pos)
1232 struct inode *inode = mapping->host;
1233 struct fuse_conn *fc = get_fuse_conn(inode);
1234 struct fuse_inode *fi = get_fuse_inode(inode);
1238 if (inode->i_size < pos + iov_iter_count(ii))
1239 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1243 struct fuse_io_args ia = {};
1244 struct fuse_args_pages *ap = &ia.ap;
1245 unsigned int nr_pages = fuse_wr_pages(pos, iov_iter_count(ii),
1248 ap->pages = fuse_pages_alloc(nr_pages, GFP_KERNEL, &ap->descs);
1254 count = fuse_fill_write_pages(&ia, mapping, ii, pos, nr_pages);
1258 err = fuse_send_write_pages(&ia, iocb, inode,
1261 size_t num_written = ia.write.out.size;
1266 /* break out of the loop on short write */
1267 if (num_written != count)
1272 } while (!err && iov_iter_count(ii));
1275 fuse_write_update_size(inode, pos);
1277 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1278 fuse_invalidate_attr(inode);
1280 return res > 0 ? res : err;
1283 static ssize_t fuse_cache_write_iter(struct kiocb *iocb, struct iov_iter *from)
1285 struct file *file = iocb->ki_filp;
1286 struct address_space *mapping = file->f_mapping;
1287 ssize_t written = 0;
1288 ssize_t written_buffered = 0;
1289 struct inode *inode = mapping->host;
1293 if (get_fuse_conn(inode)->writeback_cache) {
1294 /* Update size (EOF optimization) and mode (SUID clearing) */
1295 err = fuse_update_attributes(mapping->host, file);
1299 return generic_file_write_iter(iocb, from);
1304 /* We can write back this queue in page reclaim */
1305 current->backing_dev_info = inode_to_bdi(inode);
1307 err = generic_write_checks(iocb, from);
1311 err = file_remove_privs(file);
1315 err = file_update_time(file);
1319 if (iocb->ki_flags & IOCB_DIRECT) {
1320 loff_t pos = iocb->ki_pos;
1321 written = generic_file_direct_write(iocb, from);
1322 if (written < 0 || !iov_iter_count(from))
1327 written_buffered = fuse_perform_write(iocb, mapping, from, pos);
1328 if (written_buffered < 0) {
1329 err = written_buffered;
1332 endbyte = pos + written_buffered - 1;
1334 err = filemap_write_and_wait_range(file->f_mapping, pos,
1339 invalidate_mapping_pages(file->f_mapping,
1341 endbyte >> PAGE_SHIFT);
1343 written += written_buffered;
1344 iocb->ki_pos = pos + written_buffered;
1346 written = fuse_perform_write(iocb, mapping, from, iocb->ki_pos);
1348 iocb->ki_pos += written;
1351 current->backing_dev_info = NULL;
1352 inode_unlock(inode);
1354 written = generic_write_sync(iocb, written);
1356 return written ? written : err;
1359 static inline void fuse_page_descs_length_init(struct fuse_page_desc *descs,
1361 unsigned int nr_pages)
1365 for (i = index; i < index + nr_pages; i++)
1366 descs[i].length = PAGE_SIZE - descs[i].offset;
1369 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1371 return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1374 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1377 return min(iov_iter_single_seg_count(ii), max_size);
1380 static int fuse_get_user_pages(struct fuse_args_pages *ap, struct iov_iter *ii,
1381 size_t *nbytesp, int write,
1382 unsigned int max_pages)
1384 size_t nbytes = 0; /* # bytes already packed in req */
1387 /* Special case for kernel I/O: can copy directly into the buffer */
1388 if (iov_iter_is_kvec(ii)) {
1389 unsigned long user_addr = fuse_get_user_addr(ii);
1390 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1393 ap->args.in_args[1].value = (void *) user_addr;
1395 ap->args.out_args[0].value = (void *) user_addr;
1397 iov_iter_advance(ii, frag_size);
1398 *nbytesp = frag_size;
1402 while (nbytes < *nbytesp && ap->num_pages < max_pages) {
1405 ret = iov_iter_get_pages(ii, &ap->pages[ap->num_pages],
1407 max_pages - ap->num_pages,
1412 iov_iter_advance(ii, ret);
1416 npages = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
1418 ap->descs[ap->num_pages].offset = start;
1419 fuse_page_descs_length_init(ap->descs, ap->num_pages, npages);
1421 ap->num_pages += npages;
1422 ap->descs[ap->num_pages - 1].length -=
1423 (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1426 ap->args.user_pages = true;
1428 ap->args.in_pages = true;
1430 ap->args.out_pages = true;
1434 return ret < 0 ? ret : 0;
1437 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1438 loff_t *ppos, int flags)
1440 int write = flags & FUSE_DIO_WRITE;
1441 int cuse = flags & FUSE_DIO_CUSE;
1442 struct file *file = io->iocb->ki_filp;
1443 struct inode *inode = file->f_mapping->host;
1444 struct fuse_file *ff = file->private_data;
1445 struct fuse_conn *fc = ff->fm->fc;
1446 size_t nmax = write ? fc->max_write : fc->max_read;
1448 size_t count = iov_iter_count(iter);
1449 pgoff_t idx_from = pos >> PAGE_SHIFT;
1450 pgoff_t idx_to = (pos + count - 1) >> PAGE_SHIFT;
1453 struct fuse_io_args *ia;
1454 unsigned int max_pages;
1456 max_pages = iov_iter_npages(iter, fc->max_pages);
1457 ia = fuse_io_alloc(io, max_pages);
1462 if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1465 fuse_sync_writes(inode);
1467 inode_unlock(inode);
1470 io->should_dirty = !write && iter_is_iovec(iter);
1473 fl_owner_t owner = current->files;
1474 size_t nbytes = min(count, nmax);
1476 err = fuse_get_user_pages(&ia->ap, iter, &nbytes, write,
1482 if (!capable(CAP_FSETID))
1483 ia->write.in.write_flags |= FUSE_WRITE_KILL_PRIV;
1485 nres = fuse_send_write(ia, pos, nbytes, owner);
1487 nres = fuse_send_read(ia, pos, nbytes, owner);
1490 if (!io->async || nres < 0) {
1491 fuse_release_user_pages(&ia->ap, io->should_dirty);
1496 iov_iter_revert(iter, nbytes);
1500 WARN_ON(nres > nbytes);
1505 if (nres != nbytes) {
1506 iov_iter_revert(iter, nbytes - nres);
1510 max_pages = iov_iter_npages(iter, fc->max_pages);
1511 ia = fuse_io_alloc(io, max_pages);
1521 return res > 0 ? res : err;
1523 EXPORT_SYMBOL_GPL(fuse_direct_io);
1525 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1526 struct iov_iter *iter,
1530 struct inode *inode = file_inode(io->iocb->ki_filp);
1532 res = fuse_direct_io(io, iter, ppos, 0);
1534 fuse_invalidate_atime(inode);
1539 static ssize_t fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter);
1541 static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1545 if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1546 res = fuse_direct_IO(iocb, to);
1548 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1550 res = __fuse_direct_read(&io, to, &iocb->ki_pos);
1556 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1558 struct inode *inode = file_inode(iocb->ki_filp);
1559 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1562 /* Don't allow parallel writes to the same file */
1564 res = generic_write_checks(iocb, from);
1566 if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1567 res = fuse_direct_IO(iocb, from);
1569 res = fuse_direct_io(&io, from, &iocb->ki_pos,
1573 fuse_invalidate_attr(inode);
1575 fuse_write_update_size(inode, iocb->ki_pos);
1576 inode_unlock(inode);
1581 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
1583 struct file *file = iocb->ki_filp;
1584 struct fuse_file *ff = file->private_data;
1585 struct inode *inode = file_inode(file);
1587 if (fuse_is_bad(inode))
1590 if (FUSE_IS_DAX(inode))
1591 return fuse_dax_read_iter(iocb, to);
1593 if (!(ff->open_flags & FOPEN_DIRECT_IO))
1594 return fuse_cache_read_iter(iocb, to);
1596 return fuse_direct_read_iter(iocb, to);
1599 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1601 struct file *file = iocb->ki_filp;
1602 struct fuse_file *ff = file->private_data;
1603 struct inode *inode = file_inode(file);
1605 if (fuse_is_bad(inode))
1608 if (FUSE_IS_DAX(inode))
1609 return fuse_dax_write_iter(iocb, from);
1611 if (!(ff->open_flags & FOPEN_DIRECT_IO))
1612 return fuse_cache_write_iter(iocb, from);
1614 return fuse_direct_write_iter(iocb, from);
1617 static void fuse_writepage_free(struct fuse_writepage_args *wpa)
1619 struct fuse_args_pages *ap = &wpa->ia.ap;
1622 for (i = 0; i < ap->num_pages; i++)
1623 __free_page(ap->pages[i]);
1626 fuse_file_put(wpa->ia.ff, false, false);
1632 static void fuse_writepage_finish(struct fuse_mount *fm,
1633 struct fuse_writepage_args *wpa)
1635 struct fuse_args_pages *ap = &wpa->ia.ap;
1636 struct inode *inode = wpa->inode;
1637 struct fuse_inode *fi = get_fuse_inode(inode);
1638 struct backing_dev_info *bdi = inode_to_bdi(inode);
1641 for (i = 0; i < ap->num_pages; i++) {
1642 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1643 dec_node_page_state(ap->pages[i], NR_WRITEBACK_TEMP);
1644 wb_writeout_inc(&bdi->wb);
1646 wake_up(&fi->page_waitq);
1649 /* Called under fi->lock, may release and reacquire it */
1650 static void fuse_send_writepage(struct fuse_mount *fm,
1651 struct fuse_writepage_args *wpa, loff_t size)
1652 __releases(fi->lock)
1653 __acquires(fi->lock)
1655 struct fuse_writepage_args *aux, *next;
1656 struct fuse_inode *fi = get_fuse_inode(wpa->inode);
1657 struct fuse_write_in *inarg = &wpa->ia.write.in;
1658 struct fuse_args *args = &wpa->ia.ap.args;
1659 __u64 data_size = wpa->ia.ap.num_pages * PAGE_SIZE;
1663 if (inarg->offset + data_size <= size) {
1664 inarg->size = data_size;
1665 } else if (inarg->offset < size) {
1666 inarg->size = size - inarg->offset;
1668 /* Got truncated off completely */
1672 args->in_args[1].size = inarg->size;
1674 args->nocreds = true;
1676 err = fuse_simple_background(fm, args, GFP_ATOMIC);
1677 if (err == -ENOMEM) {
1678 spin_unlock(&fi->lock);
1679 err = fuse_simple_background(fm, args, GFP_NOFS | __GFP_NOFAIL);
1680 spin_lock(&fi->lock);
1683 /* Fails on broken connection only */
1691 rb_erase(&wpa->writepages_entry, &fi->writepages);
1692 fuse_writepage_finish(fm, wpa);
1693 spin_unlock(&fi->lock);
1695 /* After fuse_writepage_finish() aux request list is private */
1696 for (aux = wpa->next; aux; aux = next) {
1699 fuse_writepage_free(aux);
1702 fuse_writepage_free(wpa);
1703 spin_lock(&fi->lock);
1707 * If fi->writectr is positive (no truncate or fsync going on) send
1708 * all queued writepage requests.
1710 * Called with fi->lock
1712 void fuse_flush_writepages(struct inode *inode)
1713 __releases(fi->lock)
1714 __acquires(fi->lock)
1716 struct fuse_mount *fm = get_fuse_mount(inode);
1717 struct fuse_inode *fi = get_fuse_inode(inode);
1718 loff_t crop = i_size_read(inode);
1719 struct fuse_writepage_args *wpa;
1721 while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1722 wpa = list_entry(fi->queued_writes.next,
1723 struct fuse_writepage_args, queue_entry);
1724 list_del_init(&wpa->queue_entry);
1725 fuse_send_writepage(fm, wpa, crop);
1729 static struct fuse_writepage_args *fuse_insert_writeback(struct rb_root *root,
1730 struct fuse_writepage_args *wpa)
1732 pgoff_t idx_from = wpa->ia.write.in.offset >> PAGE_SHIFT;
1733 pgoff_t idx_to = idx_from + wpa->ia.ap.num_pages - 1;
1734 struct rb_node **p = &root->rb_node;
1735 struct rb_node *parent = NULL;
1737 WARN_ON(!wpa->ia.ap.num_pages);
1739 struct fuse_writepage_args *curr;
1743 curr = rb_entry(parent, struct fuse_writepage_args,
1745 WARN_ON(curr->inode != wpa->inode);
1746 curr_index = curr->ia.write.in.offset >> PAGE_SHIFT;
1748 if (idx_from >= curr_index + curr->ia.ap.num_pages)
1749 p = &(*p)->rb_right;
1750 else if (idx_to < curr_index)
1756 rb_link_node(&wpa->writepages_entry, parent, p);
1757 rb_insert_color(&wpa->writepages_entry, root);
1761 static void tree_insert(struct rb_root *root, struct fuse_writepage_args *wpa)
1763 WARN_ON(fuse_insert_writeback(root, wpa));
1766 static void fuse_writepage_end(struct fuse_mount *fm, struct fuse_args *args,
1769 struct fuse_writepage_args *wpa =
1770 container_of(args, typeof(*wpa), ia.ap.args);
1771 struct inode *inode = wpa->inode;
1772 struct fuse_inode *fi = get_fuse_inode(inode);
1773 struct fuse_conn *fc = get_fuse_conn(inode);
1775 mapping_set_error(inode->i_mapping, error);
1777 * A writeback finished and this might have updated mtime/ctime on
1778 * server making local mtime/ctime stale. Hence invalidate attrs.
1779 * Do this only if writeback_cache is not enabled. If writeback_cache
1780 * is enabled, we trust local ctime/mtime.
1782 if (!fc->writeback_cache)
1783 fuse_invalidate_attr(inode);
1784 spin_lock(&fi->lock);
1785 rb_erase(&wpa->writepages_entry, &fi->writepages);
1787 struct fuse_mount *fm = get_fuse_mount(inode);
1788 struct fuse_write_in *inarg = &wpa->ia.write.in;
1789 struct fuse_writepage_args *next = wpa->next;
1791 wpa->next = next->next;
1793 next->ia.ff = fuse_file_get(wpa->ia.ff);
1794 tree_insert(&fi->writepages, next);
1797 * Skip fuse_flush_writepages() to make it easy to crop requests
1798 * based on primary request size.
1800 * 1st case (trivial): there are no concurrent activities using
1801 * fuse_set/release_nowrite. Then we're on safe side because
1802 * fuse_flush_writepages() would call fuse_send_writepage()
1805 * 2nd case: someone called fuse_set_nowrite and it is waiting
1806 * now for completion of all in-flight requests. This happens
1807 * rarely and no more than once per page, so this should be
1810 * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1811 * of fuse_set_nowrite..fuse_release_nowrite section. The fact
1812 * that fuse_set_nowrite returned implies that all in-flight
1813 * requests were completed along with all of their secondary
1814 * requests. Further primary requests are blocked by negative
1815 * writectr. Hence there cannot be any in-flight requests and
1816 * no invocations of fuse_writepage_end() while we're in
1817 * fuse_set_nowrite..fuse_release_nowrite section.
1819 fuse_send_writepage(fm, next, inarg->offset + inarg->size);
1822 fuse_writepage_finish(fm, wpa);
1823 spin_unlock(&fi->lock);
1824 fuse_writepage_free(wpa);
1827 static struct fuse_file *__fuse_write_file_get(struct fuse_conn *fc,
1828 struct fuse_inode *fi)
1830 struct fuse_file *ff = NULL;
1832 spin_lock(&fi->lock);
1833 if (!list_empty(&fi->write_files)) {
1834 ff = list_entry(fi->write_files.next, struct fuse_file,
1838 spin_unlock(&fi->lock);
1843 static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1844 struct fuse_inode *fi)
1846 struct fuse_file *ff = __fuse_write_file_get(fc, fi);
1851 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1853 struct fuse_conn *fc = get_fuse_conn(inode);
1854 struct fuse_inode *fi = get_fuse_inode(inode);
1855 struct fuse_file *ff;
1859 * Inode is always written before the last reference is dropped and
1860 * hence this should not be reached from reclaim.
1862 * Writing back the inode from reclaim can deadlock if the request
1863 * processing itself needs an allocation. Allocations triggering
1864 * reclaim while serving a request can't be prevented, because it can
1865 * involve any number of unrelated userspace processes.
1867 WARN_ON(wbc->for_reclaim);
1869 ff = __fuse_write_file_get(fc, fi);
1870 err = fuse_flush_times(inode, ff);
1872 fuse_file_put(ff, false, false);
1877 static struct fuse_writepage_args *fuse_writepage_args_alloc(void)
1879 struct fuse_writepage_args *wpa;
1880 struct fuse_args_pages *ap;
1882 wpa = kzalloc(sizeof(*wpa), GFP_NOFS);
1886 ap->pages = fuse_pages_alloc(1, GFP_NOFS, &ap->descs);
1896 static int fuse_writepage_locked(struct page *page)
1898 struct address_space *mapping = page->mapping;
1899 struct inode *inode = mapping->host;
1900 struct fuse_conn *fc = get_fuse_conn(inode);
1901 struct fuse_inode *fi = get_fuse_inode(inode);
1902 struct fuse_writepage_args *wpa;
1903 struct fuse_args_pages *ap;
1904 struct page *tmp_page;
1905 int error = -ENOMEM;
1907 set_page_writeback(page);
1909 wpa = fuse_writepage_args_alloc();
1914 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1919 wpa->ia.ff = fuse_write_file_get(fc, fi);
1923 fuse_write_args_fill(&wpa->ia, wpa->ia.ff, page_offset(page), 0);
1925 copy_highpage(tmp_page, page);
1926 wpa->ia.write.in.write_flags |= FUSE_WRITE_CACHE;
1928 ap->args.in_pages = true;
1930 ap->pages[0] = tmp_page;
1931 ap->descs[0].offset = 0;
1932 ap->descs[0].length = PAGE_SIZE;
1933 ap->args.end = fuse_writepage_end;
1936 inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1937 inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1939 spin_lock(&fi->lock);
1940 tree_insert(&fi->writepages, wpa);
1941 list_add_tail(&wpa->queue_entry, &fi->queued_writes);
1942 fuse_flush_writepages(inode);
1943 spin_unlock(&fi->lock);
1945 end_page_writeback(page);
1950 __free_page(tmp_page);
1954 mapping_set_error(page->mapping, error);
1955 end_page_writeback(page);
1959 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1963 if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1965 * ->writepages() should be called for sync() and friends. We
1966 * should only get here on direct reclaim and then we are
1967 * allowed to skip a page which is already in flight
1969 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1971 redirty_page_for_writepage(wbc, page);
1976 err = fuse_writepage_locked(page);
1982 struct fuse_fill_wb_data {
1983 struct fuse_writepage_args *wpa;
1984 struct fuse_file *ff;
1985 struct inode *inode;
1986 struct page **orig_pages;
1987 unsigned int max_pages;
1990 static bool fuse_pages_realloc(struct fuse_fill_wb_data *data)
1992 struct fuse_args_pages *ap = &data->wpa->ia.ap;
1993 struct fuse_conn *fc = get_fuse_conn(data->inode);
1994 struct page **pages;
1995 struct fuse_page_desc *descs;
1996 unsigned int npages = min_t(unsigned int,
1997 max_t(unsigned int, data->max_pages * 2,
1998 FUSE_DEFAULT_MAX_PAGES_PER_REQ),
2000 WARN_ON(npages <= data->max_pages);
2002 pages = fuse_pages_alloc(npages, GFP_NOFS, &descs);
2006 memcpy(pages, ap->pages, sizeof(struct page *) * ap->num_pages);
2007 memcpy(descs, ap->descs, sizeof(struct fuse_page_desc) * ap->num_pages);
2011 data->max_pages = npages;
2016 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
2018 struct fuse_writepage_args *wpa = data->wpa;
2019 struct inode *inode = data->inode;
2020 struct fuse_inode *fi = get_fuse_inode(inode);
2021 int num_pages = wpa->ia.ap.num_pages;
2024 wpa->ia.ff = fuse_file_get(data->ff);
2025 spin_lock(&fi->lock);
2026 list_add_tail(&wpa->queue_entry, &fi->queued_writes);
2027 fuse_flush_writepages(inode);
2028 spin_unlock(&fi->lock);
2030 for (i = 0; i < num_pages; i++)
2031 end_page_writeback(data->orig_pages[i]);
2035 * Check under fi->lock if the page is under writeback, and insert it onto the
2036 * rb_tree if not. Otherwise iterate auxiliary write requests, to see if there's
2037 * one already added for a page at this offset. If there's none, then insert
2038 * this new request onto the auxiliary list, otherwise reuse the existing one by
2039 * swapping the new temp page with the old one.
2041 static bool fuse_writepage_add(struct fuse_writepage_args *new_wpa,
2044 struct fuse_inode *fi = get_fuse_inode(new_wpa->inode);
2045 struct fuse_writepage_args *tmp;
2046 struct fuse_writepage_args *old_wpa;
2047 struct fuse_args_pages *new_ap = &new_wpa->ia.ap;
2049 WARN_ON(new_ap->num_pages != 0);
2050 new_ap->num_pages = 1;
2052 spin_lock(&fi->lock);
2053 old_wpa = fuse_insert_writeback(&fi->writepages, new_wpa);
2055 spin_unlock(&fi->lock);
2059 for (tmp = old_wpa->next; tmp; tmp = tmp->next) {
2062 WARN_ON(tmp->inode != new_wpa->inode);
2063 curr_index = tmp->ia.write.in.offset >> PAGE_SHIFT;
2064 if (curr_index == page->index) {
2065 WARN_ON(tmp->ia.ap.num_pages != 1);
2066 swap(tmp->ia.ap.pages[0], new_ap->pages[0]);
2072 new_wpa->next = old_wpa->next;
2073 old_wpa->next = new_wpa;
2076 spin_unlock(&fi->lock);
2079 struct backing_dev_info *bdi = inode_to_bdi(new_wpa->inode);
2081 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
2082 dec_node_page_state(new_ap->pages[0], NR_WRITEBACK_TEMP);
2083 wb_writeout_inc(&bdi->wb);
2084 fuse_writepage_free(new_wpa);
2090 static bool fuse_writepage_need_send(struct fuse_conn *fc, struct page *page,
2091 struct fuse_args_pages *ap,
2092 struct fuse_fill_wb_data *data)
2094 WARN_ON(!ap->num_pages);
2097 * Being under writeback is unlikely but possible. For example direct
2098 * read to an mmaped fuse file will set the page dirty twice; once when
2099 * the pages are faulted with get_user_pages(), and then after the read
2102 if (fuse_page_is_writeback(data->inode, page->index))
2105 /* Reached max pages */
2106 if (ap->num_pages == fc->max_pages)
2109 /* Reached max write bytes */
2110 if ((ap->num_pages + 1) * PAGE_SIZE > fc->max_write)
2114 if (data->orig_pages[ap->num_pages - 1]->index + 1 != page->index)
2117 /* Need to grow the pages array? If so, did the expansion fail? */
2118 if (ap->num_pages == data->max_pages && !fuse_pages_realloc(data))
2124 static int fuse_writepages_fill(struct page *page,
2125 struct writeback_control *wbc, void *_data)
2127 struct fuse_fill_wb_data *data = _data;
2128 struct fuse_writepage_args *wpa = data->wpa;
2129 struct fuse_args_pages *ap = &wpa->ia.ap;
2130 struct inode *inode = data->inode;
2131 struct fuse_inode *fi = get_fuse_inode(inode);
2132 struct fuse_conn *fc = get_fuse_conn(inode);
2133 struct page *tmp_page;
2138 data->ff = fuse_write_file_get(fc, fi);
2143 if (wpa && fuse_writepage_need_send(fc, page, ap, data)) {
2144 fuse_writepages_send(data);
2149 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2154 * The page must not be redirtied until the writeout is completed
2155 * (i.e. userspace has sent a reply to the write request). Otherwise
2156 * there could be more than one temporary page instance for each real
2159 * This is ensured by holding the page lock in page_mkwrite() while
2160 * checking fuse_page_is_writeback(). We already hold the page lock
2161 * since clear_page_dirty_for_io() and keep it held until we add the
2162 * request to the fi->writepages list and increment ap->num_pages.
2163 * After this fuse_page_is_writeback() will indicate that the page is
2164 * under writeback, so we can release the page lock.
2166 if (data->wpa == NULL) {
2168 wpa = fuse_writepage_args_alloc();
2170 __free_page(tmp_page);
2173 data->max_pages = 1;
2176 fuse_write_args_fill(&wpa->ia, data->ff, page_offset(page), 0);
2177 wpa->ia.write.in.write_flags |= FUSE_WRITE_CACHE;
2179 ap->args.in_pages = true;
2180 ap->args.end = fuse_writepage_end;
2184 set_page_writeback(page);
2186 copy_highpage(tmp_page, page);
2187 ap->pages[ap->num_pages] = tmp_page;
2188 ap->descs[ap->num_pages].offset = 0;
2189 ap->descs[ap->num_pages].length = PAGE_SIZE;
2190 data->orig_pages[ap->num_pages] = page;
2192 inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
2193 inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
2198 * Protected by fi->lock against concurrent access by
2199 * fuse_page_is_writeback().
2201 spin_lock(&fi->lock);
2203 spin_unlock(&fi->lock);
2204 } else if (fuse_writepage_add(wpa, page)) {
2207 end_page_writeback(page);
2215 static int fuse_writepages(struct address_space *mapping,
2216 struct writeback_control *wbc)
2218 struct inode *inode = mapping->host;
2219 struct fuse_conn *fc = get_fuse_conn(inode);
2220 struct fuse_fill_wb_data data;
2224 if (fuse_is_bad(inode))
2232 data.orig_pages = kcalloc(fc->max_pages,
2233 sizeof(struct page *),
2235 if (!data.orig_pages)
2238 err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
2240 WARN_ON(!data.wpa->ia.ap.num_pages);
2241 fuse_writepages_send(&data);
2244 fuse_file_put(data.ff, false, false);
2246 kfree(data.orig_pages);
2252 * It's worthy to make sure that space is reserved on disk for the write,
2253 * but how to implement it without killing performance need more thinking.
2255 static int fuse_write_begin(struct file *file, struct address_space *mapping,
2256 loff_t pos, unsigned len, unsigned flags,
2257 struct page **pagep, void **fsdata)
2259 pgoff_t index = pos >> PAGE_SHIFT;
2260 struct fuse_conn *fc = get_fuse_conn(file_inode(file));
2265 WARN_ON(!fc->writeback_cache);
2267 page = grab_cache_page_write_begin(mapping, index, flags);
2271 fuse_wait_on_page_writeback(mapping->host, page->index);
2273 if (PageUptodate(page) || len == PAGE_SIZE)
2276 * Check if the start this page comes after the end of file, in which
2277 * case the readpage can be optimized away.
2279 fsize = i_size_read(mapping->host);
2280 if (fsize <= (pos & PAGE_MASK)) {
2281 size_t off = pos & ~PAGE_MASK;
2283 zero_user_segment(page, 0, off);
2286 err = fuse_do_readpage(file, page);
2300 static int fuse_write_end(struct file *file, struct address_space *mapping,
2301 loff_t pos, unsigned len, unsigned copied,
2302 struct page *page, void *fsdata)
2304 struct inode *inode = page->mapping->host;
2306 /* Haven't copied anything? Skip zeroing, size extending, dirtying. */
2310 if (!PageUptodate(page)) {
2311 /* Zero any unwritten bytes at the end of the page */
2312 size_t endoff = (pos + copied) & ~PAGE_MASK;
2314 zero_user_segment(page, endoff, PAGE_SIZE);
2315 SetPageUptodate(page);
2318 fuse_write_update_size(inode, pos + copied);
2319 set_page_dirty(page);
2328 static int fuse_launder_page(struct page *page)
2331 if (clear_page_dirty_for_io(page)) {
2332 struct inode *inode = page->mapping->host;
2333 err = fuse_writepage_locked(page);
2335 fuse_wait_on_page_writeback(inode, page->index);
2341 * Write back dirty pages now, because there may not be any suitable
2344 static void fuse_vma_close(struct vm_area_struct *vma)
2346 filemap_write_and_wait(vma->vm_file->f_mapping);
2350 * Wait for writeback against this page to complete before allowing it
2351 * to be marked dirty again, and hence written back again, possibly
2352 * before the previous writepage completed.
2354 * Block here, instead of in ->writepage(), so that the userspace fs
2355 * can only block processes actually operating on the filesystem.
2357 * Otherwise unprivileged userspace fs would be able to block
2362 * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2364 static vm_fault_t fuse_page_mkwrite(struct vm_fault *vmf)
2366 struct page *page = vmf->page;
2367 struct inode *inode = file_inode(vmf->vma->vm_file);
2369 file_update_time(vmf->vma->vm_file);
2371 if (page->mapping != inode->i_mapping) {
2373 return VM_FAULT_NOPAGE;
2376 fuse_wait_on_page_writeback(inode, page->index);
2377 return VM_FAULT_LOCKED;
2380 static const struct vm_operations_struct fuse_file_vm_ops = {
2381 .close = fuse_vma_close,
2382 .fault = filemap_fault,
2383 .map_pages = filemap_map_pages,
2384 .page_mkwrite = fuse_page_mkwrite,
2387 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2389 struct fuse_file *ff = file->private_data;
2391 /* DAX mmap is superior to direct_io mmap */
2392 if (FUSE_IS_DAX(file_inode(file)))
2393 return fuse_dax_mmap(file, vma);
2395 if (ff->open_flags & FOPEN_DIRECT_IO) {
2396 /* Can't provide the coherency needed for MAP_SHARED */
2397 if (vma->vm_flags & VM_MAYSHARE)
2400 invalidate_inode_pages2(file->f_mapping);
2402 return generic_file_mmap(file, vma);
2405 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2406 fuse_link_write_file(file);
2408 file_accessed(file);
2409 vma->vm_ops = &fuse_file_vm_ops;
2413 static int convert_fuse_file_lock(struct fuse_conn *fc,
2414 const struct fuse_file_lock *ffl,
2415 struct file_lock *fl)
2417 switch (ffl->type) {
2423 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2424 ffl->end < ffl->start)
2427 fl->fl_start = ffl->start;
2428 fl->fl_end = ffl->end;
2431 * Convert pid into init's pid namespace. The locks API will
2432 * translate it into the caller's pid namespace.
2435 fl->fl_pid = pid_nr_ns(find_pid_ns(ffl->pid, fc->pid_ns), &init_pid_ns);
2442 fl->fl_type = ffl->type;
2446 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2447 const struct file_lock *fl, int opcode, pid_t pid,
2448 int flock, struct fuse_lk_in *inarg)
2450 struct inode *inode = file_inode(file);
2451 struct fuse_conn *fc = get_fuse_conn(inode);
2452 struct fuse_file *ff = file->private_data;
2454 memset(inarg, 0, sizeof(*inarg));
2456 inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2457 inarg->lk.start = fl->fl_start;
2458 inarg->lk.end = fl->fl_end;
2459 inarg->lk.type = fl->fl_type;
2460 inarg->lk.pid = pid;
2462 inarg->lk_flags |= FUSE_LK_FLOCK;
2463 args->opcode = opcode;
2464 args->nodeid = get_node_id(inode);
2465 args->in_numargs = 1;
2466 args->in_args[0].size = sizeof(*inarg);
2467 args->in_args[0].value = inarg;
2470 static int fuse_getlk(struct file *file, struct file_lock *fl)
2472 struct inode *inode = file_inode(file);
2473 struct fuse_mount *fm = get_fuse_mount(inode);
2475 struct fuse_lk_in inarg;
2476 struct fuse_lk_out outarg;
2479 fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2480 args.out_numargs = 1;
2481 args.out_args[0].size = sizeof(outarg);
2482 args.out_args[0].value = &outarg;
2483 err = fuse_simple_request(fm, &args);
2485 err = convert_fuse_file_lock(fm->fc, &outarg.lk, fl);
2490 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2492 struct inode *inode = file_inode(file);
2493 struct fuse_mount *fm = get_fuse_mount(inode);
2495 struct fuse_lk_in inarg;
2496 int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2497 struct pid *pid = fl->fl_type != F_UNLCK ? task_tgid(current) : NULL;
2498 pid_t pid_nr = pid_nr_ns(pid, fm->fc->pid_ns);
2501 if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2502 /* NLM needs asynchronous locks, which we don't support yet */
2506 /* Unlock on close is handled by the flush method */
2507 if ((fl->fl_flags & FL_CLOSE_POSIX) == FL_CLOSE_POSIX)
2510 fuse_lk_fill(&args, file, fl, opcode, pid_nr, flock, &inarg);
2511 err = fuse_simple_request(fm, &args);
2513 /* locking is restartable */
2520 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2522 struct inode *inode = file_inode(file);
2523 struct fuse_conn *fc = get_fuse_conn(inode);
2526 if (cmd == F_CANCELLK) {
2528 } else if (cmd == F_GETLK) {
2530 posix_test_lock(file, fl);
2533 err = fuse_getlk(file, fl);
2536 err = posix_lock_file(file, fl, NULL);
2538 err = fuse_setlk(file, fl, 0);
2543 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2545 struct inode *inode = file_inode(file);
2546 struct fuse_conn *fc = get_fuse_conn(inode);
2550 err = locks_lock_file_wait(file, fl);
2552 struct fuse_file *ff = file->private_data;
2554 /* emulate flock with POSIX locks */
2556 err = fuse_setlk(file, fl, 1);
2562 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2564 struct inode *inode = mapping->host;
2565 struct fuse_mount *fm = get_fuse_mount(inode);
2567 struct fuse_bmap_in inarg;
2568 struct fuse_bmap_out outarg;
2571 if (!inode->i_sb->s_bdev || fm->fc->no_bmap)
2574 memset(&inarg, 0, sizeof(inarg));
2575 inarg.block = block;
2576 inarg.blocksize = inode->i_sb->s_blocksize;
2577 args.opcode = FUSE_BMAP;
2578 args.nodeid = get_node_id(inode);
2579 args.in_numargs = 1;
2580 args.in_args[0].size = sizeof(inarg);
2581 args.in_args[0].value = &inarg;
2582 args.out_numargs = 1;
2583 args.out_args[0].size = sizeof(outarg);
2584 args.out_args[0].value = &outarg;
2585 err = fuse_simple_request(fm, &args);
2587 fm->fc->no_bmap = 1;
2589 return err ? 0 : outarg.block;
2592 static loff_t fuse_lseek(struct file *file, loff_t offset, int whence)
2594 struct inode *inode = file->f_mapping->host;
2595 struct fuse_mount *fm = get_fuse_mount(inode);
2596 struct fuse_file *ff = file->private_data;
2598 struct fuse_lseek_in inarg = {
2603 struct fuse_lseek_out outarg;
2606 if (fm->fc->no_lseek)
2609 args.opcode = FUSE_LSEEK;
2610 args.nodeid = ff->nodeid;
2611 args.in_numargs = 1;
2612 args.in_args[0].size = sizeof(inarg);
2613 args.in_args[0].value = &inarg;
2614 args.out_numargs = 1;
2615 args.out_args[0].size = sizeof(outarg);
2616 args.out_args[0].value = &outarg;
2617 err = fuse_simple_request(fm, &args);
2619 if (err == -ENOSYS) {
2620 fm->fc->no_lseek = 1;
2626 return vfs_setpos(file, outarg.offset, inode->i_sb->s_maxbytes);
2629 err = fuse_update_attributes(inode, file);
2631 return generic_file_llseek(file, offset, whence);
2636 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2639 struct inode *inode = file_inode(file);
2644 /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2645 retval = generic_file_llseek(file, offset, whence);
2649 retval = fuse_update_attributes(inode, file);
2651 retval = generic_file_llseek(file, offset, whence);
2652 inode_unlock(inode);
2657 retval = fuse_lseek(file, offset, whence);
2658 inode_unlock(inode);
2668 * CUSE servers compiled on 32bit broke on 64bit kernels because the
2669 * ABI was defined to be 'struct iovec' which is different on 32bit
2670 * and 64bit. Fortunately we can determine which structure the server
2671 * used from the size of the reply.
2673 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2674 size_t transferred, unsigned count,
2677 #ifdef CONFIG_COMPAT
2678 if (count * sizeof(struct compat_iovec) == transferred) {
2679 struct compat_iovec *ciov = src;
2683 * With this interface a 32bit server cannot support
2684 * non-compat (i.e. ones coming from 64bit apps) ioctl
2690 for (i = 0; i < count; i++) {
2691 dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2692 dst[i].iov_len = ciov[i].iov_len;
2698 if (count * sizeof(struct iovec) != transferred)
2701 memcpy(dst, src, transferred);
2705 /* Make sure iov_length() won't overflow */
2706 static int fuse_verify_ioctl_iov(struct fuse_conn *fc, struct iovec *iov,
2710 u32 max = fc->max_pages << PAGE_SHIFT;
2712 for (n = 0; n < count; n++, iov++) {
2713 if (iov->iov_len > (size_t) max)
2715 max -= iov->iov_len;
2720 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2721 void *src, size_t transferred, unsigned count,
2725 struct fuse_ioctl_iovec *fiov = src;
2727 if (fc->minor < 16) {
2728 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2732 if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2735 for (i = 0; i < count; i++) {
2736 /* Did the server supply an inappropriate value? */
2737 if (fiov[i].base != (unsigned long) fiov[i].base ||
2738 fiov[i].len != (unsigned long) fiov[i].len)
2741 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2742 dst[i].iov_len = (size_t) fiov[i].len;
2744 #ifdef CONFIG_COMPAT
2746 (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2747 (compat_size_t) dst[i].iov_len != fiov[i].len))
2757 * For ioctls, there is no generic way to determine how much memory
2758 * needs to be read and/or written. Furthermore, ioctls are allowed
2759 * to dereference the passed pointer, so the parameter requires deep
2760 * copying but FUSE has no idea whatsoever about what to copy in or
2763 * This is solved by allowing FUSE server to retry ioctl with
2764 * necessary in/out iovecs. Let's assume the ioctl implementation
2765 * needs to read in the following structure.
2772 * On the first callout to FUSE server, inarg->in_size and
2773 * inarg->out_size will be NULL; then, the server completes the ioctl
2774 * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2775 * the actual iov array to
2777 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) } }
2779 * which tells FUSE to copy in the requested area and retry the ioctl.
2780 * On the second round, the server has access to the structure and
2781 * from that it can tell what to look for next, so on the invocation,
2782 * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2784 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) },
2785 * { .iov_base = a.buf, .iov_len = a.buflen } }
2787 * FUSE will copy both struct a and the pointed buffer from the
2788 * process doing the ioctl and retry ioctl with both struct a and the
2791 * This time, FUSE server has everything it needs and completes ioctl
2792 * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2794 * Copying data out works the same way.
2796 * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2797 * automatically initializes in and out iovs by decoding @cmd with
2798 * _IOC_* macros and the server is not allowed to request RETRY. This
2799 * limits ioctl data transfers to well-formed ioctls and is the forced
2800 * behavior for all FUSE servers.
2802 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2805 struct fuse_file *ff = file->private_data;
2806 struct fuse_mount *fm = ff->fm;
2807 struct fuse_ioctl_in inarg = {
2813 struct fuse_ioctl_out outarg;
2814 struct iovec *iov_page = NULL;
2815 struct iovec *in_iov = NULL, *out_iov = NULL;
2816 unsigned int in_iovs = 0, out_iovs = 0, max_pages;
2817 size_t in_size, out_size, c;
2818 ssize_t transferred;
2821 struct fuse_args_pages ap = {};
2823 #if BITS_PER_LONG == 32
2824 inarg.flags |= FUSE_IOCTL_32BIT;
2826 if (flags & FUSE_IOCTL_COMPAT) {
2827 inarg.flags |= FUSE_IOCTL_32BIT;
2828 #ifdef CONFIG_X86_X32
2829 if (in_x32_syscall())
2830 inarg.flags |= FUSE_IOCTL_COMPAT_X32;
2835 /* assume all the iovs returned by client always fits in a page */
2836 BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2839 ap.pages = fuse_pages_alloc(fm->fc->max_pages, GFP_KERNEL, &ap.descs);
2840 iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2841 if (!ap.pages || !iov_page)
2844 fuse_page_descs_length_init(ap.descs, 0, fm->fc->max_pages);
2847 * If restricted, initialize IO parameters as encoded in @cmd.
2848 * RETRY from server is not allowed.
2850 if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2851 struct iovec *iov = iov_page;
2853 iov->iov_base = (void __user *)arg;
2856 case FS_IOC_GETFLAGS:
2857 case FS_IOC_SETFLAGS:
2858 iov->iov_len = sizeof(int);
2861 iov->iov_len = _IOC_SIZE(cmd);
2865 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2870 if (_IOC_DIR(cmd) & _IOC_READ) {
2877 inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2878 inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2881 * Out data can be used either for actual out data or iovs,
2882 * make sure there always is at least one page.
2884 out_size = max_t(size_t, out_size, PAGE_SIZE);
2885 max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2887 /* make sure there are enough buffer pages and init request with them */
2889 if (max_pages > fm->fc->max_pages)
2891 while (ap.num_pages < max_pages) {
2892 ap.pages[ap.num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2893 if (!ap.pages[ap.num_pages])
2899 /* okay, let's send it to the client */
2900 ap.args.opcode = FUSE_IOCTL;
2901 ap.args.nodeid = ff->nodeid;
2902 ap.args.in_numargs = 1;
2903 ap.args.in_args[0].size = sizeof(inarg);
2904 ap.args.in_args[0].value = &inarg;
2906 ap.args.in_numargs++;
2907 ap.args.in_args[1].size = in_size;
2908 ap.args.in_pages = true;
2911 iov_iter_init(&ii, WRITE, in_iov, in_iovs, in_size);
2912 for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= ap.num_pages); i++) {
2913 c = copy_page_from_iter(ap.pages[i], 0, PAGE_SIZE, &ii);
2914 if (c != PAGE_SIZE && iov_iter_count(&ii))
2919 ap.args.out_numargs = 2;
2920 ap.args.out_args[0].size = sizeof(outarg);
2921 ap.args.out_args[0].value = &outarg;
2922 ap.args.out_args[1].size = out_size;
2923 ap.args.out_pages = true;
2924 ap.args.out_argvar = true;
2926 transferred = fuse_simple_request(fm, &ap.args);
2928 if (transferred < 0)
2931 /* did it ask for retry? */
2932 if (outarg.flags & FUSE_IOCTL_RETRY) {
2935 /* no retry if in restricted mode */
2937 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2940 in_iovs = outarg.in_iovs;
2941 out_iovs = outarg.out_iovs;
2944 * Make sure things are in boundary, separate checks
2945 * are to protect against overflow.
2948 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2949 out_iovs > FUSE_IOCTL_MAX_IOV ||
2950 in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2953 vaddr = kmap_atomic(ap.pages[0]);
2954 err = fuse_copy_ioctl_iovec(fm->fc, iov_page, vaddr,
2955 transferred, in_iovs + out_iovs,
2956 (flags & FUSE_IOCTL_COMPAT) != 0);
2957 kunmap_atomic(vaddr);
2962 out_iov = in_iov + in_iovs;
2964 err = fuse_verify_ioctl_iov(fm->fc, in_iov, in_iovs);
2968 err = fuse_verify_ioctl_iov(fm->fc, out_iov, out_iovs);
2976 if (transferred > inarg.out_size)
2980 iov_iter_init(&ii, READ, out_iov, out_iovs, transferred);
2981 for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= ap.num_pages); i++) {
2982 c = copy_page_to_iter(ap.pages[i], 0, PAGE_SIZE, &ii);
2983 if (c != PAGE_SIZE && iov_iter_count(&ii))
2988 free_page((unsigned long) iov_page);
2989 while (ap.num_pages)
2990 __free_page(ap.pages[--ap.num_pages]);
2993 return err ? err : outarg.result;
2995 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2997 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2998 unsigned long arg, unsigned int flags)
3000 struct inode *inode = file_inode(file);
3001 struct fuse_conn *fc = get_fuse_conn(inode);
3003 if (!fuse_allow_current_process(fc))
3006 if (fuse_is_bad(inode))
3009 return fuse_do_ioctl(file, cmd, arg, flags);
3012 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
3015 return fuse_ioctl_common(file, cmd, arg, 0);
3018 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
3021 return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
3025 * All files which have been polled are linked to RB tree
3026 * fuse_conn->polled_files which is indexed by kh. Walk the tree and
3027 * find the matching one.
3029 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
3030 struct rb_node **parent_out)
3032 struct rb_node **link = &fc->polled_files.rb_node;
3033 struct rb_node *last = NULL;
3036 struct fuse_file *ff;
3039 ff = rb_entry(last, struct fuse_file, polled_node);
3042 link = &last->rb_left;
3043 else if (kh > ff->kh)
3044 link = &last->rb_right;
3055 * The file is about to be polled. Make sure it's on the polled_files
3056 * RB tree. Note that files once added to the polled_files tree are
3057 * not removed before the file is released. This is because a file
3058 * polled once is likely to be polled again.
3060 static void fuse_register_polled_file(struct fuse_conn *fc,
3061 struct fuse_file *ff)
3063 spin_lock(&fc->lock);
3064 if (RB_EMPTY_NODE(&ff->polled_node)) {
3065 struct rb_node **link, *parent;
3067 link = fuse_find_polled_node(fc, ff->kh, &parent);
3069 rb_link_node(&ff->polled_node, parent, link);
3070 rb_insert_color(&ff->polled_node, &fc->polled_files);
3072 spin_unlock(&fc->lock);
3075 __poll_t fuse_file_poll(struct file *file, poll_table *wait)
3077 struct fuse_file *ff = file->private_data;
3078 struct fuse_mount *fm = ff->fm;
3079 struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
3080 struct fuse_poll_out outarg;
3084 if (fm->fc->no_poll)
3085 return DEFAULT_POLLMASK;
3087 poll_wait(file, &ff->poll_wait, wait);
3088 inarg.events = mangle_poll(poll_requested_events(wait));
3091 * Ask for notification iff there's someone waiting for it.
3092 * The client may ignore the flag and always notify.
3094 if (waitqueue_active(&ff->poll_wait)) {
3095 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
3096 fuse_register_polled_file(fm->fc, ff);
3099 args.opcode = FUSE_POLL;
3100 args.nodeid = ff->nodeid;
3101 args.in_numargs = 1;
3102 args.in_args[0].size = sizeof(inarg);
3103 args.in_args[0].value = &inarg;
3104 args.out_numargs = 1;
3105 args.out_args[0].size = sizeof(outarg);
3106 args.out_args[0].value = &outarg;
3107 err = fuse_simple_request(fm, &args);
3110 return demangle_poll(outarg.revents);
3111 if (err == -ENOSYS) {
3112 fm->fc->no_poll = 1;
3113 return DEFAULT_POLLMASK;
3117 EXPORT_SYMBOL_GPL(fuse_file_poll);
3120 * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
3121 * wakes up the poll waiters.
3123 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
3124 struct fuse_notify_poll_wakeup_out *outarg)
3126 u64 kh = outarg->kh;
3127 struct rb_node **link;
3129 spin_lock(&fc->lock);
3131 link = fuse_find_polled_node(fc, kh, NULL);
3133 struct fuse_file *ff;
3135 ff = rb_entry(*link, struct fuse_file, polled_node);
3136 wake_up_interruptible_sync(&ff->poll_wait);
3139 spin_unlock(&fc->lock);
3143 static void fuse_do_truncate(struct file *file)
3145 struct inode *inode = file->f_mapping->host;
3148 attr.ia_valid = ATTR_SIZE;
3149 attr.ia_size = i_size_read(inode);
3151 attr.ia_file = file;
3152 attr.ia_valid |= ATTR_FILE;
3154 fuse_do_setattr(file_dentry(file), &attr, file);
3157 static inline loff_t fuse_round_up(struct fuse_conn *fc, loff_t off)
3159 return round_up(off, fc->max_pages << PAGE_SHIFT);
3163 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
3165 DECLARE_COMPLETION_ONSTACK(wait);
3167 struct file *file = iocb->ki_filp;
3168 struct fuse_file *ff = file->private_data;
3170 struct inode *inode;
3172 size_t count = iov_iter_count(iter), shortened = 0;
3173 loff_t offset = iocb->ki_pos;
3174 struct fuse_io_priv *io;
3177 inode = file->f_mapping->host;
3178 i_size = i_size_read(inode);
3180 if ((iov_iter_rw(iter) == READ) && (offset >= i_size))
3183 io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
3186 spin_lock_init(&io->lock);
3187 kref_init(&io->refcnt);
3191 io->offset = offset;
3192 io->write = (iov_iter_rw(iter) == WRITE);
3195 * By default, we want to optimize all I/Os with async request
3196 * submission to the client filesystem if supported.
3198 io->async = ff->fm->fc->async_dio;
3200 io->blocking = is_sync_kiocb(iocb);
3202 /* optimization for short read */
3203 if (io->async && !io->write && offset + count > i_size) {
3204 iov_iter_truncate(iter, fuse_round_up(ff->fm->fc, i_size - offset));
3205 shortened = count - iov_iter_count(iter);
3210 * We cannot asynchronously extend the size of a file.
3211 * In such case the aio will behave exactly like sync io.
3213 if ((offset + count > i_size) && io->write)
3214 io->blocking = true;
3216 if (io->async && io->blocking) {
3218 * Additional reference to keep io around after
3219 * calling fuse_aio_complete()
3221 kref_get(&io->refcnt);
3225 if (iov_iter_rw(iter) == WRITE) {
3226 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
3227 fuse_invalidate_attr(inode);
3229 ret = __fuse_direct_read(io, iter, &pos);
3231 iov_iter_reexpand(iter, iov_iter_count(iter) + shortened);
3234 bool blocking = io->blocking;
3236 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
3238 /* we have a non-extending, async request, so return */
3240 return -EIOCBQUEUED;
3242 wait_for_completion(&wait);
3243 ret = fuse_get_res_by_io(io);
3246 kref_put(&io->refcnt, fuse_io_release);
3248 if (iov_iter_rw(iter) == WRITE) {
3250 fuse_write_update_size(inode, pos);
3251 else if (ret < 0 && offset + count > i_size)
3252 fuse_do_truncate(file);
3258 static int fuse_writeback_range(struct inode *inode, loff_t start, loff_t end)
3260 int err = filemap_write_and_wait_range(inode->i_mapping, start, LLONG_MAX);
3263 fuse_sync_writes(inode);
3268 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
3271 struct fuse_file *ff = file->private_data;
3272 struct inode *inode = file_inode(file);
3273 struct fuse_inode *fi = get_fuse_inode(inode);
3274 struct fuse_mount *fm = ff->fm;
3276 struct fuse_fallocate_in inarg = {
3283 bool block_faults = FUSE_IS_DAX(inode) &&
3284 (!(mode & FALLOC_FL_KEEP_SIZE) ||
3285 (mode & FALLOC_FL_PUNCH_HOLE));
3287 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
3290 if (fm->fc->no_fallocate)
3295 down_write(&fi->i_mmap_sem);
3296 err = fuse_dax_break_layouts(inode, 0, 0);
3301 if (mode & FALLOC_FL_PUNCH_HOLE) {
3302 loff_t endbyte = offset + length - 1;
3304 err = fuse_writeback_range(inode, offset, endbyte);
3309 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
3310 offset + length > i_size_read(inode)) {
3311 err = inode_newsize_ok(inode, offset + length);
3316 err = file_modified(file);
3320 if (!(mode & FALLOC_FL_KEEP_SIZE))
3321 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3323 args.opcode = FUSE_FALLOCATE;
3324 args.nodeid = ff->nodeid;
3325 args.in_numargs = 1;
3326 args.in_args[0].size = sizeof(inarg);
3327 args.in_args[0].value = &inarg;
3328 err = fuse_simple_request(fm, &args);
3329 if (err == -ENOSYS) {
3330 fm->fc->no_fallocate = 1;
3336 /* we could have extended the file */
3337 if (!(mode & FALLOC_FL_KEEP_SIZE)) {
3338 bool changed = fuse_write_update_size(inode, offset + length);
3340 if (changed && fm->fc->writeback_cache)
3341 file_update_time(file);
3344 if (mode & FALLOC_FL_PUNCH_HOLE)
3345 truncate_pagecache_range(inode, offset, offset + length - 1);
3347 fuse_invalidate_attr(inode);
3350 if (!(mode & FALLOC_FL_KEEP_SIZE))
3351 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3354 up_write(&fi->i_mmap_sem);
3356 inode_unlock(inode);
3358 fuse_flush_time_update(inode);
3363 static ssize_t __fuse_copy_file_range(struct file *file_in, loff_t pos_in,
3364 struct file *file_out, loff_t pos_out,
3365 size_t len, unsigned int flags)
3367 struct fuse_file *ff_in = file_in->private_data;
3368 struct fuse_file *ff_out = file_out->private_data;
3369 struct inode *inode_in = file_inode(file_in);
3370 struct inode *inode_out = file_inode(file_out);
3371 struct fuse_inode *fi_out = get_fuse_inode(inode_out);
3372 struct fuse_mount *fm = ff_in->fm;
3373 struct fuse_conn *fc = fm->fc;
3375 struct fuse_copy_file_range_in inarg = {
3378 .nodeid_out = ff_out->nodeid,
3379 .fh_out = ff_out->fh,
3384 struct fuse_write_out outarg;
3386 /* mark unstable when write-back is not used, and file_out gets
3388 bool is_unstable = (!fc->writeback_cache) &&
3389 ((pos_out + len) > inode_out->i_size);
3391 if (fc->no_copy_file_range)
3394 if (file_inode(file_in)->i_sb != file_inode(file_out)->i_sb)
3397 inode_lock(inode_in);
3398 err = fuse_writeback_range(inode_in, pos_in, pos_in + len - 1);
3399 inode_unlock(inode_in);
3403 inode_lock(inode_out);
3405 err = file_modified(file_out);
3410 * Write out dirty pages in the destination file before sending the COPY
3411 * request to userspace. After the request is completed, truncate off
3412 * pages (including partial ones) from the cache that have been copied,
3413 * since these contain stale data at that point.
3415 * This should be mostly correct, but if the COPY writes to partial
3416 * pages (at the start or end) and the parts not covered by the COPY are
3417 * written through a memory map after calling fuse_writeback_range(),
3418 * then these partial page modifications will be lost on truncation.
3420 * It is unlikely that someone would rely on such mixed style
3421 * modifications. Yet this does give less guarantees than if the
3422 * copying was performed with write(2).
3424 * To fix this a i_mmap_sem style lock could be used to prevent new
3425 * faults while the copy is ongoing.
3427 err = fuse_writeback_range(inode_out, pos_out, pos_out + len - 1);
3432 set_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3434 args.opcode = FUSE_COPY_FILE_RANGE;
3435 args.nodeid = ff_in->nodeid;
3436 args.in_numargs = 1;
3437 args.in_args[0].size = sizeof(inarg);
3438 args.in_args[0].value = &inarg;
3439 args.out_numargs = 1;
3440 args.out_args[0].size = sizeof(outarg);
3441 args.out_args[0].value = &outarg;
3442 err = fuse_simple_request(fm, &args);
3443 if (err == -ENOSYS) {
3444 fc->no_copy_file_range = 1;
3450 truncate_inode_pages_range(inode_out->i_mapping,
3451 ALIGN_DOWN(pos_out, PAGE_SIZE),
3452 ALIGN(pos_out + outarg.size, PAGE_SIZE) - 1);
3454 if (fc->writeback_cache) {
3455 fuse_write_update_size(inode_out, pos_out + outarg.size);
3456 file_update_time(file_out);
3459 fuse_invalidate_attr(inode_out);
3464 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3466 inode_unlock(inode_out);
3467 file_accessed(file_in);
3469 fuse_flush_time_update(inode_out);
3474 static ssize_t fuse_copy_file_range(struct file *src_file, loff_t src_off,
3475 struct file *dst_file, loff_t dst_off,
3476 size_t len, unsigned int flags)
3480 ret = __fuse_copy_file_range(src_file, src_off, dst_file, dst_off,
3483 if (ret == -EOPNOTSUPP || ret == -EXDEV)
3484 ret = generic_copy_file_range(src_file, src_off, dst_file,
3485 dst_off, len, flags);
3489 static const struct file_operations fuse_file_operations = {
3490 .llseek = fuse_file_llseek,
3491 .read_iter = fuse_file_read_iter,
3492 .write_iter = fuse_file_write_iter,
3493 .mmap = fuse_file_mmap,
3495 .flush = fuse_flush,
3496 .release = fuse_release,
3497 .fsync = fuse_fsync,
3498 .lock = fuse_file_lock,
3499 .get_unmapped_area = thp_get_unmapped_area,
3500 .flock = fuse_file_flock,
3501 .splice_read = generic_file_splice_read,
3502 .splice_write = iter_file_splice_write,
3503 .unlocked_ioctl = fuse_file_ioctl,
3504 .compat_ioctl = fuse_file_compat_ioctl,
3505 .poll = fuse_file_poll,
3506 .fallocate = fuse_file_fallocate,
3507 .copy_file_range = fuse_copy_file_range,
3510 static const struct address_space_operations fuse_file_aops = {
3511 .readpage = fuse_readpage,
3512 .readahead = fuse_readahead,
3513 .writepage = fuse_writepage,
3514 .writepages = fuse_writepages,
3515 .launder_page = fuse_launder_page,
3516 .set_page_dirty = __set_page_dirty_nobuffers,
3518 .direct_IO = fuse_direct_IO,
3519 .write_begin = fuse_write_begin,
3520 .write_end = fuse_write_end,
3523 void fuse_init_file_inode(struct inode *inode)
3525 struct fuse_inode *fi = get_fuse_inode(inode);
3527 inode->i_fop = &fuse_file_operations;
3528 inode->i_data.a_ops = &fuse_file_aops;
3530 INIT_LIST_HEAD(&fi->write_files);
3531 INIT_LIST_HEAD(&fi->queued_writes);
3533 init_waitqueue_head(&fi->page_waitq);
3534 fi->writepages = RB_ROOT;
3536 if (IS_ENABLED(CONFIG_FUSE_DAX))
3537 fuse_dax_inode_init(inode);