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/swap.h>
18 #include <linux/falloc.h>
19 #include <linux/uio.h>
21 #include <linux/filelock.h>
23 static int fuse_send_open(struct fuse_mount *fm, u64 nodeid,
24 unsigned int open_flags, int opcode,
25 struct fuse_open_out *outargp)
27 struct fuse_open_in inarg;
30 memset(&inarg, 0, sizeof(inarg));
31 inarg.flags = open_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
32 if (!fm->fc->atomic_o_trunc)
33 inarg.flags &= ~O_TRUNC;
35 if (fm->fc->handle_killpriv_v2 &&
36 (inarg.flags & O_TRUNC) && !capable(CAP_FSETID)) {
37 inarg.open_flags |= FUSE_OPEN_KILL_SUIDGID;
43 args.in_args[0].size = sizeof(inarg);
44 args.in_args[0].value = &inarg;
46 args.out_args[0].size = sizeof(*outargp);
47 args.out_args[0].value = outargp;
49 return fuse_simple_request(fm, &args);
52 struct fuse_release_args {
53 struct fuse_args args;
54 struct fuse_release_in inarg;
58 struct fuse_file *fuse_file_alloc(struct fuse_mount *fm)
62 ff = kzalloc(sizeof(struct fuse_file), GFP_KERNEL_ACCOUNT);
67 ff->release_args = kzalloc(sizeof(*ff->release_args),
69 if (!ff->release_args) {
74 INIT_LIST_HEAD(&ff->write_entry);
75 mutex_init(&ff->readdir.lock);
76 refcount_set(&ff->count, 1);
77 RB_CLEAR_NODE(&ff->polled_node);
78 init_waitqueue_head(&ff->poll_wait);
80 ff->kh = atomic64_inc_return(&fm->fc->khctr);
85 void fuse_file_free(struct fuse_file *ff)
87 kfree(ff->release_args);
88 mutex_destroy(&ff->readdir.lock);
92 static struct fuse_file *fuse_file_get(struct fuse_file *ff)
94 refcount_inc(&ff->count);
98 static void fuse_release_end(struct fuse_mount *fm, struct fuse_args *args,
101 struct fuse_release_args *ra = container_of(args, typeof(*ra), args);
107 static void fuse_file_put(struct fuse_file *ff, bool sync, bool isdir)
109 if (refcount_dec_and_test(&ff->count)) {
110 struct fuse_args *args = &ff->release_args->args;
112 if (isdir ? ff->fm->fc->no_opendir : ff->fm->fc->no_open) {
113 /* Do nothing when client does not implement 'open' */
114 fuse_release_end(ff->fm, args, 0);
116 fuse_simple_request(ff->fm, args);
117 fuse_release_end(ff->fm, args, 0);
119 args->end = fuse_release_end;
120 if (fuse_simple_background(ff->fm, args,
121 GFP_KERNEL | __GFP_NOFAIL))
122 fuse_release_end(ff->fm, args, -ENOTCONN);
128 struct fuse_file *fuse_file_open(struct fuse_mount *fm, u64 nodeid,
129 unsigned int open_flags, bool isdir)
131 struct fuse_conn *fc = fm->fc;
132 struct fuse_file *ff;
133 int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
135 ff = fuse_file_alloc(fm);
137 return ERR_PTR(-ENOMEM);
140 /* Default for no-open */
141 ff->open_flags = FOPEN_KEEP_CACHE | (isdir ? FOPEN_CACHE_DIR : 0);
142 if (isdir ? !fc->no_opendir : !fc->no_open) {
143 struct fuse_open_out outarg;
146 err = fuse_send_open(fm, nodeid, open_flags, opcode, &outarg);
149 ff->open_flags = outarg.open_flags;
151 } else if (err != -ENOSYS) {
163 ff->open_flags &= ~FOPEN_DIRECT_IO;
170 int fuse_do_open(struct fuse_mount *fm, u64 nodeid, struct file *file,
173 struct fuse_file *ff = fuse_file_open(fm, nodeid, file->f_flags, isdir);
176 file->private_data = ff;
178 return PTR_ERR_OR_ZERO(ff);
180 EXPORT_SYMBOL_GPL(fuse_do_open);
182 static void fuse_link_write_file(struct file *file)
184 struct inode *inode = file_inode(file);
185 struct fuse_inode *fi = get_fuse_inode(inode);
186 struct fuse_file *ff = file->private_data;
188 * file may be written through mmap, so chain it onto the
189 * inodes's write_file list
191 spin_lock(&fi->lock);
192 if (list_empty(&ff->write_entry))
193 list_add(&ff->write_entry, &fi->write_files);
194 spin_unlock(&fi->lock);
197 void fuse_finish_open(struct inode *inode, struct file *file)
199 struct fuse_file *ff = file->private_data;
200 struct fuse_conn *fc = get_fuse_conn(inode);
202 if (ff->open_flags & FOPEN_STREAM)
203 stream_open(inode, file);
204 else if (ff->open_flags & FOPEN_NONSEEKABLE)
205 nonseekable_open(inode, file);
207 if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
208 struct fuse_inode *fi = get_fuse_inode(inode);
210 spin_lock(&fi->lock);
211 fi->attr_version = atomic64_inc_return(&fc->attr_version);
212 i_size_write(inode, 0);
213 spin_unlock(&fi->lock);
214 file_update_time(file);
215 fuse_invalidate_attr_mask(inode, FUSE_STATX_MODSIZE);
217 if ((file->f_mode & FMODE_WRITE) && fc->writeback_cache)
218 fuse_link_write_file(file);
221 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
223 struct fuse_mount *fm = get_fuse_mount(inode);
224 struct fuse_conn *fc = fm->fc;
226 bool is_wb_truncate = (file->f_flags & O_TRUNC) &&
227 fc->atomic_o_trunc &&
229 bool dax_truncate = (file->f_flags & O_TRUNC) &&
230 fc->atomic_o_trunc && FUSE_IS_DAX(inode);
232 if (fuse_is_bad(inode))
235 err = generic_file_open(inode, file);
239 if (is_wb_truncate || dax_truncate)
243 filemap_invalidate_lock(inode->i_mapping);
244 err = fuse_dax_break_layouts(inode, 0, 0);
246 goto out_inode_unlock;
249 if (is_wb_truncate || dax_truncate)
250 fuse_set_nowrite(inode);
252 err = fuse_do_open(fm, get_node_id(inode), file, isdir);
254 fuse_finish_open(inode, file);
256 if (is_wb_truncate || dax_truncate)
257 fuse_release_nowrite(inode);
259 struct fuse_file *ff = file->private_data;
261 if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC))
262 truncate_pagecache(inode, 0);
263 else if (!(ff->open_flags & FOPEN_KEEP_CACHE))
264 invalidate_inode_pages2(inode->i_mapping);
267 filemap_invalidate_unlock(inode->i_mapping);
269 if (is_wb_truncate || dax_truncate)
275 static void fuse_prepare_release(struct fuse_inode *fi, struct fuse_file *ff,
276 unsigned int flags, int opcode)
278 struct fuse_conn *fc = ff->fm->fc;
279 struct fuse_release_args *ra = ff->release_args;
281 /* Inode is NULL on error path of fuse_create_open() */
283 spin_lock(&fi->lock);
284 list_del(&ff->write_entry);
285 spin_unlock(&fi->lock);
287 spin_lock(&fc->lock);
288 if (!RB_EMPTY_NODE(&ff->polled_node))
289 rb_erase(&ff->polled_node, &fc->polled_files);
290 spin_unlock(&fc->lock);
292 wake_up_interruptible_all(&ff->poll_wait);
294 ra->inarg.fh = ff->fh;
295 ra->inarg.flags = flags;
296 ra->args.in_numargs = 1;
297 ra->args.in_args[0].size = sizeof(struct fuse_release_in);
298 ra->args.in_args[0].value = &ra->inarg;
299 ra->args.opcode = opcode;
300 ra->args.nodeid = ff->nodeid;
301 ra->args.force = true;
302 ra->args.nocreds = true;
305 void fuse_file_release(struct inode *inode, struct fuse_file *ff,
306 unsigned int open_flags, fl_owner_t id, bool isdir)
308 struct fuse_inode *fi = get_fuse_inode(inode);
309 struct fuse_release_args *ra = ff->release_args;
310 int opcode = isdir ? FUSE_RELEASEDIR : FUSE_RELEASE;
312 fuse_prepare_release(fi, ff, open_flags, opcode);
315 ra->inarg.release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
316 ra->inarg.lock_owner = fuse_lock_owner_id(ff->fm->fc, id);
318 /* Hold inode until release is finished */
319 ra->inode = igrab(inode);
322 * Normally this will send the RELEASE request, however if
323 * some asynchronous READ or WRITE requests are outstanding,
324 * the sending will be delayed.
326 * Make the release synchronous if this is a fuseblk mount,
327 * synchronous RELEASE is allowed (and desirable) in this case
328 * because the server can be trusted not to screw up.
330 fuse_file_put(ff, ff->fm->fc->destroy, isdir);
333 void fuse_release_common(struct file *file, bool isdir)
335 fuse_file_release(file_inode(file), file->private_data, file->f_flags,
336 (fl_owner_t) file, isdir);
339 static int fuse_open(struct inode *inode, struct file *file)
341 return fuse_open_common(inode, file, false);
344 static int fuse_release(struct inode *inode, struct file *file)
346 struct fuse_conn *fc = get_fuse_conn(inode);
349 * Dirty pages might remain despite write_inode_now() call from
350 * fuse_flush() due to writes racing with the close.
352 if (fc->writeback_cache)
353 write_inode_now(inode, 1);
355 fuse_release_common(file, false);
357 /* return value is ignored by VFS */
361 void fuse_sync_release(struct fuse_inode *fi, struct fuse_file *ff,
364 WARN_ON(refcount_read(&ff->count) > 1);
365 fuse_prepare_release(fi, ff, flags, FUSE_RELEASE);
367 * iput(NULL) is a no-op and since the refcount is 1 and everything's
368 * synchronous, we are fine with not doing igrab() here"
370 fuse_file_put(ff, true, false);
372 EXPORT_SYMBOL_GPL(fuse_sync_release);
375 * Scramble the ID space with XTEA, so that the value of the files_struct
376 * pointer is not exposed to userspace.
378 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
380 u32 *k = fc->scramble_key;
381 u64 v = (unsigned long) id;
387 for (i = 0; i < 32; i++) {
388 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
390 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
393 return (u64) v0 + ((u64) v1 << 32);
396 struct fuse_writepage_args {
397 struct fuse_io_args ia;
398 struct rb_node writepages_entry;
399 struct list_head queue_entry;
400 struct fuse_writepage_args *next;
402 struct fuse_sync_bucket *bucket;
405 static struct fuse_writepage_args *fuse_find_writeback(struct fuse_inode *fi,
406 pgoff_t idx_from, pgoff_t idx_to)
410 n = fi->writepages.rb_node;
413 struct fuse_writepage_args *wpa;
416 wpa = rb_entry(n, struct fuse_writepage_args, writepages_entry);
417 WARN_ON(get_fuse_inode(wpa->inode) != fi);
418 curr_index = wpa->ia.write.in.offset >> PAGE_SHIFT;
419 if (idx_from >= curr_index + wpa->ia.ap.num_pages)
421 else if (idx_to < curr_index)
430 * Check if any page in a range is under writeback
432 * This is currently done by walking the list of writepage requests
433 * for the inode, which can be pretty inefficient.
435 static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
438 struct fuse_inode *fi = get_fuse_inode(inode);
441 spin_lock(&fi->lock);
442 found = fuse_find_writeback(fi, idx_from, idx_to);
443 spin_unlock(&fi->lock);
448 static inline bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
450 return fuse_range_is_writeback(inode, index, index);
454 * Wait for page writeback to be completed.
456 * Since fuse doesn't rely on the VM writeback tracking, this has to
457 * use some other means.
459 static void fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
461 struct fuse_inode *fi = get_fuse_inode(inode);
463 wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
467 * Wait for all pending writepages on the inode to finish.
469 * This is currently done by blocking further writes with FUSE_NOWRITE
470 * and waiting for all sent writes to complete.
472 * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
473 * could conflict with truncation.
475 static void fuse_sync_writes(struct inode *inode)
477 fuse_set_nowrite(inode);
478 fuse_release_nowrite(inode);
481 static int fuse_flush(struct file *file, fl_owner_t id)
483 struct inode *inode = file_inode(file);
484 struct fuse_mount *fm = get_fuse_mount(inode);
485 struct fuse_file *ff = file->private_data;
486 struct fuse_flush_in inarg;
490 if (fuse_is_bad(inode))
493 if (ff->open_flags & FOPEN_NOFLUSH && !fm->fc->writeback_cache)
496 err = write_inode_now(inode, 1);
501 fuse_sync_writes(inode);
504 err = filemap_check_errors(file->f_mapping);
509 if (fm->fc->no_flush)
512 memset(&inarg, 0, sizeof(inarg));
514 inarg.lock_owner = fuse_lock_owner_id(fm->fc, id);
515 args.opcode = FUSE_FLUSH;
516 args.nodeid = get_node_id(inode);
518 args.in_args[0].size = sizeof(inarg);
519 args.in_args[0].value = &inarg;
522 err = fuse_simple_request(fm, &args);
523 if (err == -ENOSYS) {
524 fm->fc->no_flush = 1;
530 * In memory i_blocks is not maintained by fuse, if writeback cache is
531 * enabled, i_blocks from cached attr may not be accurate.
533 if (!err && fm->fc->writeback_cache)
534 fuse_invalidate_attr_mask(inode, STATX_BLOCKS);
538 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
539 int datasync, int opcode)
541 struct inode *inode = file->f_mapping->host;
542 struct fuse_mount *fm = get_fuse_mount(inode);
543 struct fuse_file *ff = file->private_data;
545 struct fuse_fsync_in inarg;
547 memset(&inarg, 0, sizeof(inarg));
549 inarg.fsync_flags = datasync ? FUSE_FSYNC_FDATASYNC : 0;
550 args.opcode = opcode;
551 args.nodeid = get_node_id(inode);
553 args.in_args[0].size = sizeof(inarg);
554 args.in_args[0].value = &inarg;
555 return fuse_simple_request(fm, &args);
558 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
561 struct inode *inode = file->f_mapping->host;
562 struct fuse_conn *fc = get_fuse_conn(inode);
565 if (fuse_is_bad(inode))
571 * Start writeback against all dirty pages of the inode, then
572 * wait for all outstanding writes, before sending the FSYNC
575 err = file_write_and_wait_range(file, start, end);
579 fuse_sync_writes(inode);
582 * Due to implementation of fuse writeback
583 * file_write_and_wait_range() does not catch errors.
584 * We have to do this directly after fuse_sync_writes()
586 err = file_check_and_advance_wb_err(file);
590 err = sync_inode_metadata(inode, 1);
597 err = fuse_fsync_common(file, start, end, datasync, FUSE_FSYNC);
598 if (err == -ENOSYS) {
608 void fuse_read_args_fill(struct fuse_io_args *ia, struct file *file, loff_t pos,
609 size_t count, int opcode)
611 struct fuse_file *ff = file->private_data;
612 struct fuse_args *args = &ia->ap.args;
614 ia->read.in.fh = ff->fh;
615 ia->read.in.offset = pos;
616 ia->read.in.size = count;
617 ia->read.in.flags = file->f_flags;
618 args->opcode = opcode;
619 args->nodeid = ff->nodeid;
620 args->in_numargs = 1;
621 args->in_args[0].size = sizeof(ia->read.in);
622 args->in_args[0].value = &ia->read.in;
623 args->out_argvar = true;
624 args->out_numargs = 1;
625 args->out_args[0].size = count;
628 static void fuse_release_user_pages(struct fuse_args_pages *ap,
633 for (i = 0; i < ap->num_pages; i++) {
635 set_page_dirty_lock(ap->pages[i]);
636 put_page(ap->pages[i]);
640 static void fuse_io_release(struct kref *kref)
642 kfree(container_of(kref, struct fuse_io_priv, refcnt));
645 static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
650 if (io->bytes >= 0 && io->write)
653 return io->bytes < 0 ? io->size : io->bytes;
657 * In case of short read, the caller sets 'pos' to the position of
658 * actual end of fuse request in IO request. Otherwise, if bytes_requested
659 * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
662 * User requested DIO read of 64K. It was split into two 32K fuse requests,
663 * both submitted asynchronously. The first of them was ACKed by userspace as
664 * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
665 * second request was ACKed as short, e.g. only 1K was read, resulting in
668 * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
669 * will be equal to the length of the longest contiguous fragment of
670 * transferred data starting from the beginning of IO request.
672 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
676 spin_lock(&io->lock);
678 io->err = io->err ? : err;
679 else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
683 if (!left && io->blocking)
685 spin_unlock(&io->lock);
687 if (!left && !io->blocking) {
688 ssize_t res = fuse_get_res_by_io(io);
691 struct inode *inode = file_inode(io->iocb->ki_filp);
692 struct fuse_conn *fc = get_fuse_conn(inode);
693 struct fuse_inode *fi = get_fuse_inode(inode);
695 spin_lock(&fi->lock);
696 fi->attr_version = atomic64_inc_return(&fc->attr_version);
697 spin_unlock(&fi->lock);
700 io->iocb->ki_complete(io->iocb, res);
703 kref_put(&io->refcnt, fuse_io_release);
706 static struct fuse_io_args *fuse_io_alloc(struct fuse_io_priv *io,
709 struct fuse_io_args *ia;
711 ia = kzalloc(sizeof(*ia), GFP_KERNEL);
714 ia->ap.pages = fuse_pages_alloc(npages, GFP_KERNEL,
724 static void fuse_io_free(struct fuse_io_args *ia)
730 static void fuse_aio_complete_req(struct fuse_mount *fm, struct fuse_args *args,
733 struct fuse_io_args *ia = container_of(args, typeof(*ia), ap.args);
734 struct fuse_io_priv *io = ia->io;
737 fuse_release_user_pages(&ia->ap, io->should_dirty);
741 } else if (io->write) {
742 if (ia->write.out.size > ia->write.in.size) {
744 } else if (ia->write.in.size != ia->write.out.size) {
745 pos = ia->write.in.offset - io->offset +
749 u32 outsize = args->out_args[0].size;
751 if (ia->read.in.size != outsize)
752 pos = ia->read.in.offset - io->offset + outsize;
755 fuse_aio_complete(io, err, pos);
759 static ssize_t fuse_async_req_send(struct fuse_mount *fm,
760 struct fuse_io_args *ia, size_t num_bytes)
763 struct fuse_io_priv *io = ia->io;
765 spin_lock(&io->lock);
766 kref_get(&io->refcnt);
767 io->size += num_bytes;
769 spin_unlock(&io->lock);
771 ia->ap.args.end = fuse_aio_complete_req;
772 ia->ap.args.may_block = io->should_dirty;
773 err = fuse_simple_background(fm, &ia->ap.args, GFP_KERNEL);
775 fuse_aio_complete_req(fm, &ia->ap.args, err);
780 static ssize_t fuse_send_read(struct fuse_io_args *ia, loff_t pos, size_t count,
783 struct file *file = ia->io->iocb->ki_filp;
784 struct fuse_file *ff = file->private_data;
785 struct fuse_mount *fm = ff->fm;
787 fuse_read_args_fill(ia, file, pos, count, FUSE_READ);
789 ia->read.in.read_flags |= FUSE_READ_LOCKOWNER;
790 ia->read.in.lock_owner = fuse_lock_owner_id(fm->fc, owner);
794 return fuse_async_req_send(fm, ia, count);
796 return fuse_simple_request(fm, &ia->ap.args);
799 static void fuse_read_update_size(struct inode *inode, loff_t size,
802 struct fuse_conn *fc = get_fuse_conn(inode);
803 struct fuse_inode *fi = get_fuse_inode(inode);
805 spin_lock(&fi->lock);
806 if (attr_ver >= fi->attr_version && size < inode->i_size &&
807 !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
808 fi->attr_version = atomic64_inc_return(&fc->attr_version);
809 i_size_write(inode, size);
811 spin_unlock(&fi->lock);
814 static void fuse_short_read(struct inode *inode, u64 attr_ver, size_t num_read,
815 struct fuse_args_pages *ap)
817 struct fuse_conn *fc = get_fuse_conn(inode);
820 * If writeback_cache is enabled, a short read means there's a hole in
821 * the file. Some data after the hole is in page cache, but has not
822 * reached the client fs yet. So the hole is not present there.
824 if (!fc->writeback_cache) {
825 loff_t pos = page_offset(ap->pages[0]) + num_read;
826 fuse_read_update_size(inode, pos, attr_ver);
830 static int fuse_do_readpage(struct file *file, struct page *page)
832 struct inode *inode = page->mapping->host;
833 struct fuse_mount *fm = get_fuse_mount(inode);
834 loff_t pos = page_offset(page);
835 struct fuse_page_desc desc = { .length = PAGE_SIZE };
836 struct fuse_io_args ia = {
837 .ap.args.page_zeroing = true,
838 .ap.args.out_pages = true,
847 * Page writeback can extend beyond the lifetime of the
848 * page-cache page, so make sure we read a properly synced
851 fuse_wait_on_page_writeback(inode, page->index);
853 attr_ver = fuse_get_attr_version(fm->fc);
855 /* Don't overflow end offset */
856 if (pos + (desc.length - 1) == LLONG_MAX)
859 fuse_read_args_fill(&ia, file, pos, desc.length, FUSE_READ);
860 res = fuse_simple_request(fm, &ia.ap.args);
864 * Short read means EOF. If file size is larger, truncate it
866 if (res < desc.length)
867 fuse_short_read(inode, attr_ver, res, &ia.ap);
869 SetPageUptodate(page);
874 static int fuse_read_folio(struct file *file, struct folio *folio)
876 struct page *page = &folio->page;
877 struct inode *inode = page->mapping->host;
881 if (fuse_is_bad(inode))
884 err = fuse_do_readpage(file, page);
885 fuse_invalidate_atime(inode);
891 static void fuse_readpages_end(struct fuse_mount *fm, struct fuse_args *args,
895 struct fuse_io_args *ia = container_of(args, typeof(*ia), ap.args);
896 struct fuse_args_pages *ap = &ia->ap;
897 size_t count = ia->read.in.size;
898 size_t num_read = args->out_args[0].size;
899 struct address_space *mapping = NULL;
901 for (i = 0; mapping == NULL && i < ap->num_pages; i++)
902 mapping = ap->pages[i]->mapping;
905 struct inode *inode = mapping->host;
908 * Short read means EOF. If file size is larger, truncate it
910 if (!err && num_read < count)
911 fuse_short_read(inode, ia->read.attr_ver, num_read, ap);
913 fuse_invalidate_atime(inode);
916 for (i = 0; i < ap->num_pages; i++) {
917 struct page *page = ap->pages[i];
920 SetPageUptodate(page);
927 fuse_file_put(ia->ff, false, false);
932 static void fuse_send_readpages(struct fuse_io_args *ia, struct file *file)
934 struct fuse_file *ff = file->private_data;
935 struct fuse_mount *fm = ff->fm;
936 struct fuse_args_pages *ap = &ia->ap;
937 loff_t pos = page_offset(ap->pages[0]);
938 size_t count = ap->num_pages << PAGE_SHIFT;
942 ap->args.out_pages = true;
943 ap->args.page_zeroing = true;
944 ap->args.page_replace = true;
946 /* Don't overflow end offset */
947 if (pos + (count - 1) == LLONG_MAX) {
949 ap->descs[ap->num_pages - 1].length--;
951 WARN_ON((loff_t) (pos + count) < 0);
953 fuse_read_args_fill(ia, file, pos, count, FUSE_READ);
954 ia->read.attr_ver = fuse_get_attr_version(fm->fc);
955 if (fm->fc->async_read) {
956 ia->ff = fuse_file_get(ff);
957 ap->args.end = fuse_readpages_end;
958 err = fuse_simple_background(fm, &ap->args, GFP_KERNEL);
962 res = fuse_simple_request(fm, &ap->args);
963 err = res < 0 ? res : 0;
965 fuse_readpages_end(fm, &ap->args, err);
968 static void fuse_readahead(struct readahead_control *rac)
970 struct inode *inode = rac->mapping->host;
971 struct fuse_conn *fc = get_fuse_conn(inode);
972 unsigned int i, max_pages, nr_pages = 0;
974 if (fuse_is_bad(inode))
977 max_pages = min_t(unsigned int, fc->max_pages,
978 fc->max_read / PAGE_SIZE);
981 struct fuse_io_args *ia;
982 struct fuse_args_pages *ap;
984 if (fc->num_background >= fc->congestion_threshold &&
985 rac->ra->async_size >= readahead_count(rac))
987 * Congested and only async pages left, so skip the
992 nr_pages = readahead_count(rac) - nr_pages;
993 if (nr_pages > max_pages)
994 nr_pages = max_pages;
997 ia = fuse_io_alloc(NULL, nr_pages);
1001 nr_pages = __readahead_batch(rac, ap->pages, nr_pages);
1002 for (i = 0; i < nr_pages; i++) {
1003 fuse_wait_on_page_writeback(inode,
1004 readahead_index(rac) + i);
1005 ap->descs[i].length = PAGE_SIZE;
1007 ap->num_pages = nr_pages;
1008 fuse_send_readpages(ia, rac->file);
1012 static ssize_t fuse_cache_read_iter(struct kiocb *iocb, struct iov_iter *to)
1014 struct inode *inode = iocb->ki_filp->f_mapping->host;
1015 struct fuse_conn *fc = get_fuse_conn(inode);
1018 * In auto invalidate mode, always update attributes on read.
1019 * Otherwise, only update if we attempt to read past EOF (to ensure
1020 * i_size is up to date).
1022 if (fc->auto_inval_data ||
1023 (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
1025 err = fuse_update_attributes(inode, iocb->ki_filp, STATX_SIZE);
1030 return generic_file_read_iter(iocb, to);
1033 static void fuse_write_args_fill(struct fuse_io_args *ia, struct fuse_file *ff,
1034 loff_t pos, size_t count)
1036 struct fuse_args *args = &ia->ap.args;
1038 ia->write.in.fh = ff->fh;
1039 ia->write.in.offset = pos;
1040 ia->write.in.size = count;
1041 args->opcode = FUSE_WRITE;
1042 args->nodeid = ff->nodeid;
1043 args->in_numargs = 2;
1044 if (ff->fm->fc->minor < 9)
1045 args->in_args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
1047 args->in_args[0].size = sizeof(ia->write.in);
1048 args->in_args[0].value = &ia->write.in;
1049 args->in_args[1].size = count;
1050 args->out_numargs = 1;
1051 args->out_args[0].size = sizeof(ia->write.out);
1052 args->out_args[0].value = &ia->write.out;
1055 static unsigned int fuse_write_flags(struct kiocb *iocb)
1057 unsigned int flags = iocb->ki_filp->f_flags;
1059 if (iocb_is_dsync(iocb))
1061 if (iocb->ki_flags & IOCB_SYNC)
1067 static ssize_t fuse_send_write(struct fuse_io_args *ia, loff_t pos,
1068 size_t count, fl_owner_t owner)
1070 struct kiocb *iocb = ia->io->iocb;
1071 struct file *file = iocb->ki_filp;
1072 struct fuse_file *ff = file->private_data;
1073 struct fuse_mount *fm = ff->fm;
1074 struct fuse_write_in *inarg = &ia->write.in;
1077 fuse_write_args_fill(ia, ff, pos, count);
1078 inarg->flags = fuse_write_flags(iocb);
1079 if (owner != NULL) {
1080 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
1081 inarg->lock_owner = fuse_lock_owner_id(fm->fc, owner);
1085 return fuse_async_req_send(fm, ia, count);
1087 err = fuse_simple_request(fm, &ia->ap.args);
1088 if (!err && ia->write.out.size > count)
1091 return err ?: ia->write.out.size;
1094 bool fuse_write_update_attr(struct inode *inode, loff_t pos, ssize_t written)
1096 struct fuse_conn *fc = get_fuse_conn(inode);
1097 struct fuse_inode *fi = get_fuse_inode(inode);
1100 spin_lock(&fi->lock);
1101 fi->attr_version = atomic64_inc_return(&fc->attr_version);
1102 if (written > 0 && pos > inode->i_size) {
1103 i_size_write(inode, pos);
1106 spin_unlock(&fi->lock);
1108 fuse_invalidate_attr_mask(inode, FUSE_STATX_MODSIZE);
1113 static ssize_t fuse_send_write_pages(struct fuse_io_args *ia,
1114 struct kiocb *iocb, struct inode *inode,
1115 loff_t pos, size_t count)
1117 struct fuse_args_pages *ap = &ia->ap;
1118 struct file *file = iocb->ki_filp;
1119 struct fuse_file *ff = file->private_data;
1120 struct fuse_mount *fm = ff->fm;
1121 unsigned int offset, i;
1125 for (i = 0; i < ap->num_pages; i++)
1126 fuse_wait_on_page_writeback(inode, ap->pages[i]->index);
1128 fuse_write_args_fill(ia, ff, pos, count);
1129 ia->write.in.flags = fuse_write_flags(iocb);
1130 if (fm->fc->handle_killpriv_v2 && !capable(CAP_FSETID))
1131 ia->write.in.write_flags |= FUSE_WRITE_KILL_SUIDGID;
1133 err = fuse_simple_request(fm, &ap->args);
1134 if (!err && ia->write.out.size > count)
1137 short_write = ia->write.out.size < count;
1138 offset = ap->descs[0].offset;
1139 count = ia->write.out.size;
1140 for (i = 0; i < ap->num_pages; i++) {
1141 struct page *page = ap->pages[i];
1144 ClearPageUptodate(page);
1146 if (count >= PAGE_SIZE - offset)
1147 count -= PAGE_SIZE - offset;
1150 ClearPageUptodate(page);
1155 if (ia->write.page_locked && (i == ap->num_pages - 1))
1163 static ssize_t fuse_fill_write_pages(struct fuse_io_args *ia,
1164 struct address_space *mapping,
1165 struct iov_iter *ii, loff_t pos,
1166 unsigned int max_pages)
1168 struct fuse_args_pages *ap = &ia->ap;
1169 struct fuse_conn *fc = get_fuse_conn(mapping->host);
1170 unsigned offset = pos & (PAGE_SIZE - 1);
1174 ap->args.in_pages = true;
1175 ap->descs[0].offset = offset;
1180 pgoff_t index = pos >> PAGE_SHIFT;
1181 size_t bytes = min_t(size_t, PAGE_SIZE - offset,
1182 iov_iter_count(ii));
1184 bytes = min_t(size_t, bytes, fc->max_write - count);
1188 if (fault_in_iov_iter_readable(ii, bytes))
1192 page = grab_cache_page_write_begin(mapping, index);
1196 if (mapping_writably_mapped(mapping))
1197 flush_dcache_page(page);
1199 tmp = copy_page_from_iter_atomic(page, offset, bytes, ii);
1200 flush_dcache_page(page);
1209 ap->pages[ap->num_pages] = page;
1210 ap->descs[ap->num_pages].length = tmp;
1216 if (offset == PAGE_SIZE)
1219 /* If we copied full page, mark it uptodate */
1220 if (tmp == PAGE_SIZE)
1221 SetPageUptodate(page);
1223 if (PageUptodate(page)) {
1226 ia->write.page_locked = true;
1229 if (!fc->big_writes)
1231 } while (iov_iter_count(ii) && count < fc->max_write &&
1232 ap->num_pages < max_pages && offset == 0);
1234 return count > 0 ? count : err;
1237 static inline unsigned int fuse_wr_pages(loff_t pos, size_t len,
1238 unsigned int max_pages)
1240 return min_t(unsigned int,
1241 ((pos + len - 1) >> PAGE_SHIFT) -
1242 (pos >> PAGE_SHIFT) + 1,
1246 static ssize_t fuse_perform_write(struct kiocb *iocb, struct iov_iter *ii)
1248 struct address_space *mapping = iocb->ki_filp->f_mapping;
1249 struct inode *inode = mapping->host;
1250 struct fuse_conn *fc = get_fuse_conn(inode);
1251 struct fuse_inode *fi = get_fuse_inode(inode);
1252 loff_t pos = iocb->ki_pos;
1256 if (inode->i_size < pos + iov_iter_count(ii))
1257 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1261 struct fuse_io_args ia = {};
1262 struct fuse_args_pages *ap = &ia.ap;
1263 unsigned int nr_pages = fuse_wr_pages(pos, iov_iter_count(ii),
1266 ap->pages = fuse_pages_alloc(nr_pages, GFP_KERNEL, &ap->descs);
1272 count = fuse_fill_write_pages(&ia, mapping, ii, pos, nr_pages);
1276 err = fuse_send_write_pages(&ia, iocb, inode,
1279 size_t num_written = ia.write.out.size;
1284 /* break out of the loop on short write */
1285 if (num_written != count)
1290 } while (!err && iov_iter_count(ii));
1292 fuse_write_update_attr(inode, pos, res);
1293 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1297 iocb->ki_pos += res;
1301 static ssize_t fuse_cache_write_iter(struct kiocb *iocb, struct iov_iter *from)
1303 struct file *file = iocb->ki_filp;
1304 struct address_space *mapping = file->f_mapping;
1305 ssize_t written = 0;
1306 struct inode *inode = mapping->host;
1308 struct fuse_conn *fc = get_fuse_conn(inode);
1310 if (fc->writeback_cache) {
1311 /* Update size (EOF optimization) and mode (SUID clearing) */
1312 err = fuse_update_attributes(mapping->host, file,
1313 STATX_SIZE | STATX_MODE);
1317 if (fc->handle_killpriv_v2 &&
1318 setattr_should_drop_suidgid(&nop_mnt_idmap,
1319 file_inode(file))) {
1323 return generic_file_write_iter(iocb, from);
1329 err = generic_write_checks(iocb, from);
1333 err = file_remove_privs(file);
1337 err = file_update_time(file);
1341 if (iocb->ki_flags & IOCB_DIRECT) {
1342 written = generic_file_direct_write(iocb, from);
1343 if (written < 0 || !iov_iter_count(from))
1345 written = direct_write_fallback(iocb, from, written,
1346 fuse_perform_write(iocb, from));
1348 written = fuse_perform_write(iocb, from);
1351 inode_unlock(inode);
1353 written = generic_write_sync(iocb, written);
1355 return written ? written : err;
1358 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1360 return (unsigned long)iter_iov(ii)->iov_base + ii->iov_offset;
1363 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1366 return min(iov_iter_single_seg_count(ii), max_size);
1369 static int fuse_get_user_pages(struct fuse_args_pages *ap, struct iov_iter *ii,
1370 size_t *nbytesp, int write,
1371 unsigned int max_pages)
1373 size_t nbytes = 0; /* # bytes already packed in req */
1376 /* Special case for kernel I/O: can copy directly into the buffer */
1377 if (iov_iter_is_kvec(ii)) {
1378 unsigned long user_addr = fuse_get_user_addr(ii);
1379 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1382 ap->args.in_args[1].value = (void *) user_addr;
1384 ap->args.out_args[0].value = (void *) user_addr;
1386 iov_iter_advance(ii, frag_size);
1387 *nbytesp = frag_size;
1391 while (nbytes < *nbytesp && ap->num_pages < max_pages) {
1394 ret = iov_iter_get_pages2(ii, &ap->pages[ap->num_pages],
1396 max_pages - ap->num_pages,
1404 npages = DIV_ROUND_UP(ret, PAGE_SIZE);
1406 ap->descs[ap->num_pages].offset = start;
1407 fuse_page_descs_length_init(ap->descs, ap->num_pages, npages);
1409 ap->num_pages += npages;
1410 ap->descs[ap->num_pages - 1].length -=
1411 (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1414 ap->args.user_pages = true;
1416 ap->args.in_pages = true;
1418 ap->args.out_pages = true;
1422 return ret < 0 ? ret : 0;
1425 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1426 loff_t *ppos, int flags)
1428 int write = flags & FUSE_DIO_WRITE;
1429 int cuse = flags & FUSE_DIO_CUSE;
1430 struct file *file = io->iocb->ki_filp;
1431 struct address_space *mapping = file->f_mapping;
1432 struct inode *inode = mapping->host;
1433 struct fuse_file *ff = file->private_data;
1434 struct fuse_conn *fc = ff->fm->fc;
1435 size_t nmax = write ? fc->max_write : fc->max_read;
1437 size_t count = iov_iter_count(iter);
1438 pgoff_t idx_from = pos >> PAGE_SHIFT;
1439 pgoff_t idx_to = (pos + count - 1) >> PAGE_SHIFT;
1442 struct fuse_io_args *ia;
1443 unsigned int max_pages;
1444 bool fopen_direct_io = ff->open_flags & FOPEN_DIRECT_IO;
1446 max_pages = iov_iter_npages(iter, fc->max_pages);
1447 ia = fuse_io_alloc(io, max_pages);
1451 if (fopen_direct_io && fc->direct_io_allow_mmap) {
1452 res = filemap_write_and_wait_range(mapping, pos, pos + count - 1);
1458 if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1461 fuse_sync_writes(inode);
1463 inode_unlock(inode);
1466 if (fopen_direct_io && write) {
1467 res = invalidate_inode_pages2_range(mapping, idx_from, idx_to);
1474 io->should_dirty = !write && user_backed_iter(iter);
1477 fl_owner_t owner = current->files;
1478 size_t nbytes = min(count, nmax);
1480 err = fuse_get_user_pages(&ia->ap, iter, &nbytes, write,
1486 if (!capable(CAP_FSETID))
1487 ia->write.in.write_flags |= FUSE_WRITE_KILL_SUIDGID;
1489 nres = fuse_send_write(ia, pos, nbytes, owner);
1491 nres = fuse_send_read(ia, pos, nbytes, owner);
1494 if (!io->async || nres < 0) {
1495 fuse_release_user_pages(&ia->ap, io->should_dirty);
1500 iov_iter_revert(iter, nbytes);
1504 WARN_ON(nres > nbytes);
1509 if (nres != nbytes) {
1510 iov_iter_revert(iter, nbytes - nres);
1514 max_pages = iov_iter_npages(iter, fc->max_pages);
1515 ia = fuse_io_alloc(io, max_pages);
1525 return res > 0 ? res : err;
1527 EXPORT_SYMBOL_GPL(fuse_direct_io);
1529 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1530 struct iov_iter *iter,
1534 struct inode *inode = file_inode(io->iocb->ki_filp);
1536 res = fuse_direct_io(io, iter, ppos, 0);
1538 fuse_invalidate_atime(inode);
1543 static ssize_t fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter);
1545 static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1549 if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1550 res = fuse_direct_IO(iocb, to);
1552 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1554 res = __fuse_direct_read(&io, to, &iocb->ki_pos);
1560 static bool fuse_direct_write_extending_i_size(struct kiocb *iocb,
1561 struct iov_iter *iter)
1563 struct inode *inode = file_inode(iocb->ki_filp);
1565 return iocb->ki_pos + iov_iter_count(iter) > i_size_read(inode);
1568 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1570 struct inode *inode = file_inode(iocb->ki_filp);
1571 struct file *file = iocb->ki_filp;
1572 struct fuse_file *ff = file->private_data;
1573 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1575 bool exclusive_lock =
1576 !(ff->open_flags & FOPEN_PARALLEL_DIRECT_WRITES) ||
1577 get_fuse_conn(inode)->direct_io_allow_mmap ||
1578 iocb->ki_flags & IOCB_APPEND ||
1579 fuse_direct_write_extending_i_size(iocb, from);
1582 * Take exclusive lock if
1583 * - Parallel direct writes are disabled - a user space decision
1584 * - Parallel direct writes are enabled and i_size is being extended.
1585 * - Shared mmap on direct_io file is supported (FUSE_DIRECT_IO_ALLOW_MMAP).
1586 * This might not be needed at all, but needs further investigation.
1591 inode_lock_shared(inode);
1593 /* A race with truncate might have come up as the decision for
1594 * the lock type was done without holding the lock, check again.
1596 if (fuse_direct_write_extending_i_size(iocb, from)) {
1597 inode_unlock_shared(inode);
1599 exclusive_lock = true;
1603 res = generic_write_checks(iocb, from);
1605 if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1606 res = fuse_direct_IO(iocb, from);
1608 res = fuse_direct_io(&io, from, &iocb->ki_pos,
1610 fuse_write_update_attr(inode, iocb->ki_pos, res);
1614 inode_unlock(inode);
1616 inode_unlock_shared(inode);
1621 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
1623 struct file *file = iocb->ki_filp;
1624 struct fuse_file *ff = file->private_data;
1625 struct inode *inode = file_inode(file);
1627 if (fuse_is_bad(inode))
1630 if (FUSE_IS_DAX(inode))
1631 return fuse_dax_read_iter(iocb, to);
1633 if (!(ff->open_flags & FOPEN_DIRECT_IO))
1634 return fuse_cache_read_iter(iocb, to);
1636 return fuse_direct_read_iter(iocb, to);
1639 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1641 struct file *file = iocb->ki_filp;
1642 struct fuse_file *ff = file->private_data;
1643 struct inode *inode = file_inode(file);
1645 if (fuse_is_bad(inode))
1648 if (FUSE_IS_DAX(inode))
1649 return fuse_dax_write_iter(iocb, from);
1651 if (!(ff->open_flags & FOPEN_DIRECT_IO))
1652 return fuse_cache_write_iter(iocb, from);
1654 return fuse_direct_write_iter(iocb, from);
1657 static void fuse_writepage_free(struct fuse_writepage_args *wpa)
1659 struct fuse_args_pages *ap = &wpa->ia.ap;
1663 fuse_sync_bucket_dec(wpa->bucket);
1665 for (i = 0; i < ap->num_pages; i++)
1666 __free_page(ap->pages[i]);
1669 fuse_file_put(wpa->ia.ff, false, false);
1675 static void fuse_writepage_finish(struct fuse_mount *fm,
1676 struct fuse_writepage_args *wpa)
1678 struct fuse_args_pages *ap = &wpa->ia.ap;
1679 struct inode *inode = wpa->inode;
1680 struct fuse_inode *fi = get_fuse_inode(inode);
1681 struct backing_dev_info *bdi = inode_to_bdi(inode);
1684 for (i = 0; i < ap->num_pages; i++) {
1685 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1686 dec_node_page_state(ap->pages[i], NR_WRITEBACK_TEMP);
1687 wb_writeout_inc(&bdi->wb);
1689 wake_up(&fi->page_waitq);
1692 /* Called under fi->lock, may release and reacquire it */
1693 static void fuse_send_writepage(struct fuse_mount *fm,
1694 struct fuse_writepage_args *wpa, loff_t size)
1695 __releases(fi->lock)
1696 __acquires(fi->lock)
1698 struct fuse_writepage_args *aux, *next;
1699 struct fuse_inode *fi = get_fuse_inode(wpa->inode);
1700 struct fuse_write_in *inarg = &wpa->ia.write.in;
1701 struct fuse_args *args = &wpa->ia.ap.args;
1702 __u64 data_size = wpa->ia.ap.num_pages * PAGE_SIZE;
1706 if (inarg->offset + data_size <= size) {
1707 inarg->size = data_size;
1708 } else if (inarg->offset < size) {
1709 inarg->size = size - inarg->offset;
1711 /* Got truncated off completely */
1715 args->in_args[1].size = inarg->size;
1717 args->nocreds = true;
1719 err = fuse_simple_background(fm, args, GFP_ATOMIC);
1720 if (err == -ENOMEM) {
1721 spin_unlock(&fi->lock);
1722 err = fuse_simple_background(fm, args, GFP_NOFS | __GFP_NOFAIL);
1723 spin_lock(&fi->lock);
1726 /* Fails on broken connection only */
1734 rb_erase(&wpa->writepages_entry, &fi->writepages);
1735 fuse_writepage_finish(fm, wpa);
1736 spin_unlock(&fi->lock);
1738 /* After fuse_writepage_finish() aux request list is private */
1739 for (aux = wpa->next; aux; aux = next) {
1742 fuse_writepage_free(aux);
1745 fuse_writepage_free(wpa);
1746 spin_lock(&fi->lock);
1750 * If fi->writectr is positive (no truncate or fsync going on) send
1751 * all queued writepage requests.
1753 * Called with fi->lock
1755 void fuse_flush_writepages(struct inode *inode)
1756 __releases(fi->lock)
1757 __acquires(fi->lock)
1759 struct fuse_mount *fm = get_fuse_mount(inode);
1760 struct fuse_inode *fi = get_fuse_inode(inode);
1761 loff_t crop = i_size_read(inode);
1762 struct fuse_writepage_args *wpa;
1764 while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1765 wpa = list_entry(fi->queued_writes.next,
1766 struct fuse_writepage_args, queue_entry);
1767 list_del_init(&wpa->queue_entry);
1768 fuse_send_writepage(fm, wpa, crop);
1772 static struct fuse_writepage_args *fuse_insert_writeback(struct rb_root *root,
1773 struct fuse_writepage_args *wpa)
1775 pgoff_t idx_from = wpa->ia.write.in.offset >> PAGE_SHIFT;
1776 pgoff_t idx_to = idx_from + wpa->ia.ap.num_pages - 1;
1777 struct rb_node **p = &root->rb_node;
1778 struct rb_node *parent = NULL;
1780 WARN_ON(!wpa->ia.ap.num_pages);
1782 struct fuse_writepage_args *curr;
1786 curr = rb_entry(parent, struct fuse_writepage_args,
1788 WARN_ON(curr->inode != wpa->inode);
1789 curr_index = curr->ia.write.in.offset >> PAGE_SHIFT;
1791 if (idx_from >= curr_index + curr->ia.ap.num_pages)
1792 p = &(*p)->rb_right;
1793 else if (idx_to < curr_index)
1799 rb_link_node(&wpa->writepages_entry, parent, p);
1800 rb_insert_color(&wpa->writepages_entry, root);
1804 static void tree_insert(struct rb_root *root, struct fuse_writepage_args *wpa)
1806 WARN_ON(fuse_insert_writeback(root, wpa));
1809 static void fuse_writepage_end(struct fuse_mount *fm, struct fuse_args *args,
1812 struct fuse_writepage_args *wpa =
1813 container_of(args, typeof(*wpa), ia.ap.args);
1814 struct inode *inode = wpa->inode;
1815 struct fuse_inode *fi = get_fuse_inode(inode);
1816 struct fuse_conn *fc = get_fuse_conn(inode);
1818 mapping_set_error(inode->i_mapping, error);
1820 * A writeback finished and this might have updated mtime/ctime on
1821 * server making local mtime/ctime stale. Hence invalidate attrs.
1822 * Do this only if writeback_cache is not enabled. If writeback_cache
1823 * is enabled, we trust local ctime/mtime.
1825 if (!fc->writeback_cache)
1826 fuse_invalidate_attr_mask(inode, FUSE_STATX_MODIFY);
1827 spin_lock(&fi->lock);
1828 rb_erase(&wpa->writepages_entry, &fi->writepages);
1830 struct fuse_mount *fm = get_fuse_mount(inode);
1831 struct fuse_write_in *inarg = &wpa->ia.write.in;
1832 struct fuse_writepage_args *next = wpa->next;
1834 wpa->next = next->next;
1836 next->ia.ff = fuse_file_get(wpa->ia.ff);
1837 tree_insert(&fi->writepages, next);
1840 * Skip fuse_flush_writepages() to make it easy to crop requests
1841 * based on primary request size.
1843 * 1st case (trivial): there are no concurrent activities using
1844 * fuse_set/release_nowrite. Then we're on safe side because
1845 * fuse_flush_writepages() would call fuse_send_writepage()
1848 * 2nd case: someone called fuse_set_nowrite and it is waiting
1849 * now for completion of all in-flight requests. This happens
1850 * rarely and no more than once per page, so this should be
1853 * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1854 * of fuse_set_nowrite..fuse_release_nowrite section. The fact
1855 * that fuse_set_nowrite returned implies that all in-flight
1856 * requests were completed along with all of their secondary
1857 * requests. Further primary requests are blocked by negative
1858 * writectr. Hence there cannot be any in-flight requests and
1859 * no invocations of fuse_writepage_end() while we're in
1860 * fuse_set_nowrite..fuse_release_nowrite section.
1862 fuse_send_writepage(fm, next, inarg->offset + inarg->size);
1865 fuse_writepage_finish(fm, wpa);
1866 spin_unlock(&fi->lock);
1867 fuse_writepage_free(wpa);
1870 static struct fuse_file *__fuse_write_file_get(struct fuse_inode *fi)
1872 struct fuse_file *ff;
1874 spin_lock(&fi->lock);
1875 ff = list_first_entry_or_null(&fi->write_files, struct fuse_file,
1879 spin_unlock(&fi->lock);
1884 static struct fuse_file *fuse_write_file_get(struct fuse_inode *fi)
1886 struct fuse_file *ff = __fuse_write_file_get(fi);
1891 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1893 struct fuse_inode *fi = get_fuse_inode(inode);
1894 struct fuse_file *ff;
1898 * Inode is always written before the last reference is dropped and
1899 * hence this should not be reached from reclaim.
1901 * Writing back the inode from reclaim can deadlock if the request
1902 * processing itself needs an allocation. Allocations triggering
1903 * reclaim while serving a request can't be prevented, because it can
1904 * involve any number of unrelated userspace processes.
1906 WARN_ON(wbc->for_reclaim);
1908 ff = __fuse_write_file_get(fi);
1909 err = fuse_flush_times(inode, ff);
1911 fuse_file_put(ff, false, false);
1916 static struct fuse_writepage_args *fuse_writepage_args_alloc(void)
1918 struct fuse_writepage_args *wpa;
1919 struct fuse_args_pages *ap;
1921 wpa = kzalloc(sizeof(*wpa), GFP_NOFS);
1925 ap->pages = fuse_pages_alloc(1, GFP_NOFS, &ap->descs);
1935 static void fuse_writepage_add_to_bucket(struct fuse_conn *fc,
1936 struct fuse_writepage_args *wpa)
1942 /* Prevent resurrection of dead bucket in unlikely race with syncfs */
1944 wpa->bucket = rcu_dereference(fc->curr_bucket);
1945 } while (unlikely(!atomic_inc_not_zero(&wpa->bucket->count)));
1949 static int fuse_writepage_locked(struct page *page)
1951 struct address_space *mapping = page->mapping;
1952 struct inode *inode = mapping->host;
1953 struct fuse_conn *fc = get_fuse_conn(inode);
1954 struct fuse_inode *fi = get_fuse_inode(inode);
1955 struct fuse_writepage_args *wpa;
1956 struct fuse_args_pages *ap;
1957 struct page *tmp_page;
1958 int error = -ENOMEM;
1960 set_page_writeback(page);
1962 wpa = fuse_writepage_args_alloc();
1967 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1972 wpa->ia.ff = fuse_write_file_get(fi);
1976 fuse_writepage_add_to_bucket(fc, wpa);
1977 fuse_write_args_fill(&wpa->ia, wpa->ia.ff, page_offset(page), 0);
1979 copy_highpage(tmp_page, page);
1980 wpa->ia.write.in.write_flags |= FUSE_WRITE_CACHE;
1982 ap->args.in_pages = true;
1984 ap->pages[0] = tmp_page;
1985 ap->descs[0].offset = 0;
1986 ap->descs[0].length = PAGE_SIZE;
1987 ap->args.end = fuse_writepage_end;
1990 inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1991 inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1993 spin_lock(&fi->lock);
1994 tree_insert(&fi->writepages, wpa);
1995 list_add_tail(&wpa->queue_entry, &fi->queued_writes);
1996 fuse_flush_writepages(inode);
1997 spin_unlock(&fi->lock);
1999 end_page_writeback(page);
2004 __free_page(tmp_page);
2008 mapping_set_error(page->mapping, error);
2009 end_page_writeback(page);
2013 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
2015 struct fuse_conn *fc = get_fuse_conn(page->mapping->host);
2018 if (fuse_page_is_writeback(page->mapping->host, page->index)) {
2020 * ->writepages() should be called for sync() and friends. We
2021 * should only get here on direct reclaim and then we are
2022 * allowed to skip a page which is already in flight
2024 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
2026 redirty_page_for_writepage(wbc, page);
2031 if (wbc->sync_mode == WB_SYNC_NONE &&
2032 fc->num_background >= fc->congestion_threshold)
2033 return AOP_WRITEPAGE_ACTIVATE;
2035 err = fuse_writepage_locked(page);
2041 struct fuse_fill_wb_data {
2042 struct fuse_writepage_args *wpa;
2043 struct fuse_file *ff;
2044 struct inode *inode;
2045 struct page **orig_pages;
2046 unsigned int max_pages;
2049 static bool fuse_pages_realloc(struct fuse_fill_wb_data *data)
2051 struct fuse_args_pages *ap = &data->wpa->ia.ap;
2052 struct fuse_conn *fc = get_fuse_conn(data->inode);
2053 struct page **pages;
2054 struct fuse_page_desc *descs;
2055 unsigned int npages = min_t(unsigned int,
2056 max_t(unsigned int, data->max_pages * 2,
2057 FUSE_DEFAULT_MAX_PAGES_PER_REQ),
2059 WARN_ON(npages <= data->max_pages);
2061 pages = fuse_pages_alloc(npages, GFP_NOFS, &descs);
2065 memcpy(pages, ap->pages, sizeof(struct page *) * ap->num_pages);
2066 memcpy(descs, ap->descs, sizeof(struct fuse_page_desc) * ap->num_pages);
2070 data->max_pages = npages;
2075 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
2077 struct fuse_writepage_args *wpa = data->wpa;
2078 struct inode *inode = data->inode;
2079 struct fuse_inode *fi = get_fuse_inode(inode);
2080 int num_pages = wpa->ia.ap.num_pages;
2083 wpa->ia.ff = fuse_file_get(data->ff);
2084 spin_lock(&fi->lock);
2085 list_add_tail(&wpa->queue_entry, &fi->queued_writes);
2086 fuse_flush_writepages(inode);
2087 spin_unlock(&fi->lock);
2089 for (i = 0; i < num_pages; i++)
2090 end_page_writeback(data->orig_pages[i]);
2094 * Check under fi->lock if the page is under writeback, and insert it onto the
2095 * rb_tree if not. Otherwise iterate auxiliary write requests, to see if there's
2096 * one already added for a page at this offset. If there's none, then insert
2097 * this new request onto the auxiliary list, otherwise reuse the existing one by
2098 * swapping the new temp page with the old one.
2100 static bool fuse_writepage_add(struct fuse_writepage_args *new_wpa,
2103 struct fuse_inode *fi = get_fuse_inode(new_wpa->inode);
2104 struct fuse_writepage_args *tmp;
2105 struct fuse_writepage_args *old_wpa;
2106 struct fuse_args_pages *new_ap = &new_wpa->ia.ap;
2108 WARN_ON(new_ap->num_pages != 0);
2109 new_ap->num_pages = 1;
2111 spin_lock(&fi->lock);
2112 old_wpa = fuse_insert_writeback(&fi->writepages, new_wpa);
2114 spin_unlock(&fi->lock);
2118 for (tmp = old_wpa->next; tmp; tmp = tmp->next) {
2121 WARN_ON(tmp->inode != new_wpa->inode);
2122 curr_index = tmp->ia.write.in.offset >> PAGE_SHIFT;
2123 if (curr_index == page->index) {
2124 WARN_ON(tmp->ia.ap.num_pages != 1);
2125 swap(tmp->ia.ap.pages[0], new_ap->pages[0]);
2131 new_wpa->next = old_wpa->next;
2132 old_wpa->next = new_wpa;
2135 spin_unlock(&fi->lock);
2138 struct backing_dev_info *bdi = inode_to_bdi(new_wpa->inode);
2140 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
2141 dec_node_page_state(new_ap->pages[0], NR_WRITEBACK_TEMP);
2142 wb_writeout_inc(&bdi->wb);
2143 fuse_writepage_free(new_wpa);
2149 static bool fuse_writepage_need_send(struct fuse_conn *fc, struct page *page,
2150 struct fuse_args_pages *ap,
2151 struct fuse_fill_wb_data *data)
2153 WARN_ON(!ap->num_pages);
2156 * Being under writeback is unlikely but possible. For example direct
2157 * read to an mmaped fuse file will set the page dirty twice; once when
2158 * the pages are faulted with get_user_pages(), and then after the read
2161 if (fuse_page_is_writeback(data->inode, page->index))
2164 /* Reached max pages */
2165 if (ap->num_pages == fc->max_pages)
2168 /* Reached max write bytes */
2169 if ((ap->num_pages + 1) * PAGE_SIZE > fc->max_write)
2173 if (data->orig_pages[ap->num_pages - 1]->index + 1 != page->index)
2176 /* Need to grow the pages array? If so, did the expansion fail? */
2177 if (ap->num_pages == data->max_pages && !fuse_pages_realloc(data))
2183 static int fuse_writepages_fill(struct folio *folio,
2184 struct writeback_control *wbc, void *_data)
2186 struct fuse_fill_wb_data *data = _data;
2187 struct fuse_writepage_args *wpa = data->wpa;
2188 struct fuse_args_pages *ap = &wpa->ia.ap;
2189 struct inode *inode = data->inode;
2190 struct fuse_inode *fi = get_fuse_inode(inode);
2191 struct fuse_conn *fc = get_fuse_conn(inode);
2192 struct page *tmp_page;
2197 data->ff = fuse_write_file_get(fi);
2202 if (wpa && fuse_writepage_need_send(fc, &folio->page, ap, data)) {
2203 fuse_writepages_send(data);
2208 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2213 * The page must not be redirtied until the writeout is completed
2214 * (i.e. userspace has sent a reply to the write request). Otherwise
2215 * there could be more than one temporary page instance for each real
2218 * This is ensured by holding the page lock in page_mkwrite() while
2219 * checking fuse_page_is_writeback(). We already hold the page lock
2220 * since clear_page_dirty_for_io() and keep it held until we add the
2221 * request to the fi->writepages list and increment ap->num_pages.
2222 * After this fuse_page_is_writeback() will indicate that the page is
2223 * under writeback, so we can release the page lock.
2225 if (data->wpa == NULL) {
2227 wpa = fuse_writepage_args_alloc();
2229 __free_page(tmp_page);
2232 fuse_writepage_add_to_bucket(fc, wpa);
2234 data->max_pages = 1;
2237 fuse_write_args_fill(&wpa->ia, data->ff, folio_pos(folio), 0);
2238 wpa->ia.write.in.write_flags |= FUSE_WRITE_CACHE;
2240 ap->args.in_pages = true;
2241 ap->args.end = fuse_writepage_end;
2245 folio_start_writeback(folio);
2247 copy_highpage(tmp_page, &folio->page);
2248 ap->pages[ap->num_pages] = tmp_page;
2249 ap->descs[ap->num_pages].offset = 0;
2250 ap->descs[ap->num_pages].length = PAGE_SIZE;
2251 data->orig_pages[ap->num_pages] = &folio->page;
2253 inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
2254 inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
2259 * Protected by fi->lock against concurrent access by
2260 * fuse_page_is_writeback().
2262 spin_lock(&fi->lock);
2264 spin_unlock(&fi->lock);
2265 } else if (fuse_writepage_add(wpa, &folio->page)) {
2268 folio_end_writeback(folio);
2271 folio_unlock(folio);
2276 static int fuse_writepages(struct address_space *mapping,
2277 struct writeback_control *wbc)
2279 struct inode *inode = mapping->host;
2280 struct fuse_conn *fc = get_fuse_conn(inode);
2281 struct fuse_fill_wb_data data;
2285 if (fuse_is_bad(inode))
2288 if (wbc->sync_mode == WB_SYNC_NONE &&
2289 fc->num_background >= fc->congestion_threshold)
2297 data.orig_pages = kcalloc(fc->max_pages,
2298 sizeof(struct page *),
2300 if (!data.orig_pages)
2303 err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
2305 WARN_ON(!data.wpa->ia.ap.num_pages);
2306 fuse_writepages_send(&data);
2309 fuse_file_put(data.ff, false, false);
2311 kfree(data.orig_pages);
2317 * It's worthy to make sure that space is reserved on disk for the write,
2318 * but how to implement it without killing performance need more thinking.
2320 static int fuse_write_begin(struct file *file, struct address_space *mapping,
2321 loff_t pos, unsigned len, struct page **pagep, void **fsdata)
2323 pgoff_t index = pos >> PAGE_SHIFT;
2324 struct fuse_conn *fc = get_fuse_conn(file_inode(file));
2329 WARN_ON(!fc->writeback_cache);
2331 page = grab_cache_page_write_begin(mapping, index);
2335 fuse_wait_on_page_writeback(mapping->host, page->index);
2337 if (PageUptodate(page) || len == PAGE_SIZE)
2340 * Check if the start this page comes after the end of file, in which
2341 * case the readpage can be optimized away.
2343 fsize = i_size_read(mapping->host);
2344 if (fsize <= (pos & PAGE_MASK)) {
2345 size_t off = pos & ~PAGE_MASK;
2347 zero_user_segment(page, 0, off);
2350 err = fuse_do_readpage(file, page);
2364 static int fuse_write_end(struct file *file, struct address_space *mapping,
2365 loff_t pos, unsigned len, unsigned copied,
2366 struct page *page, void *fsdata)
2368 struct inode *inode = page->mapping->host;
2370 /* Haven't copied anything? Skip zeroing, size extending, dirtying. */
2375 if (!PageUptodate(page)) {
2376 /* Zero any unwritten bytes at the end of the page */
2377 size_t endoff = pos & ~PAGE_MASK;
2379 zero_user_segment(page, endoff, PAGE_SIZE);
2380 SetPageUptodate(page);
2383 if (pos > inode->i_size)
2384 i_size_write(inode, pos);
2386 set_page_dirty(page);
2395 static int fuse_launder_folio(struct folio *folio)
2398 if (folio_clear_dirty_for_io(folio)) {
2399 struct inode *inode = folio->mapping->host;
2401 /* Serialize with pending writeback for the same page */
2402 fuse_wait_on_page_writeback(inode, folio->index);
2403 err = fuse_writepage_locked(&folio->page);
2405 fuse_wait_on_page_writeback(inode, folio->index);
2411 * Write back dirty data/metadata now (there may not be any suitable
2412 * open files later for data)
2414 static void fuse_vma_close(struct vm_area_struct *vma)
2418 err = write_inode_now(vma->vm_file->f_mapping->host, 1);
2419 mapping_set_error(vma->vm_file->f_mapping, err);
2423 * Wait for writeback against this page to complete before allowing it
2424 * to be marked dirty again, and hence written back again, possibly
2425 * before the previous writepage completed.
2427 * Block here, instead of in ->writepage(), so that the userspace fs
2428 * can only block processes actually operating on the filesystem.
2430 * Otherwise unprivileged userspace fs would be able to block
2435 * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2437 static vm_fault_t fuse_page_mkwrite(struct vm_fault *vmf)
2439 struct page *page = vmf->page;
2440 struct inode *inode = file_inode(vmf->vma->vm_file);
2442 file_update_time(vmf->vma->vm_file);
2444 if (page->mapping != inode->i_mapping) {
2446 return VM_FAULT_NOPAGE;
2449 fuse_wait_on_page_writeback(inode, page->index);
2450 return VM_FAULT_LOCKED;
2453 static const struct vm_operations_struct fuse_file_vm_ops = {
2454 .close = fuse_vma_close,
2455 .fault = filemap_fault,
2456 .map_pages = filemap_map_pages,
2457 .page_mkwrite = fuse_page_mkwrite,
2460 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2462 struct fuse_file *ff = file->private_data;
2463 struct fuse_conn *fc = ff->fm->fc;
2465 /* DAX mmap is superior to direct_io mmap */
2466 if (FUSE_IS_DAX(file_inode(file)))
2467 return fuse_dax_mmap(file, vma);
2469 if (ff->open_flags & FOPEN_DIRECT_IO) {
2470 /* Can't provide the coherency needed for MAP_SHARED
2471 * if FUSE_DIRECT_IO_ALLOW_MMAP isn't set.
2473 if ((vma->vm_flags & VM_MAYSHARE) && !fc->direct_io_allow_mmap)
2476 invalidate_inode_pages2(file->f_mapping);
2478 return generic_file_mmap(file, vma);
2481 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2482 fuse_link_write_file(file);
2484 file_accessed(file);
2485 vma->vm_ops = &fuse_file_vm_ops;
2489 static int convert_fuse_file_lock(struct fuse_conn *fc,
2490 const struct fuse_file_lock *ffl,
2491 struct file_lock *fl)
2493 switch (ffl->type) {
2499 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2500 ffl->end < ffl->start)
2503 fl->fl_start = ffl->start;
2504 fl->fl_end = ffl->end;
2507 * Convert pid into init's pid namespace. The locks API will
2508 * translate it into the caller's pid namespace.
2511 fl->fl_pid = pid_nr_ns(find_pid_ns(ffl->pid, fc->pid_ns), &init_pid_ns);
2518 fl->fl_type = ffl->type;
2522 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2523 const struct file_lock *fl, int opcode, pid_t pid,
2524 int flock, struct fuse_lk_in *inarg)
2526 struct inode *inode = file_inode(file);
2527 struct fuse_conn *fc = get_fuse_conn(inode);
2528 struct fuse_file *ff = file->private_data;
2530 memset(inarg, 0, sizeof(*inarg));
2532 inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2533 inarg->lk.start = fl->fl_start;
2534 inarg->lk.end = fl->fl_end;
2535 inarg->lk.type = fl->fl_type;
2536 inarg->lk.pid = pid;
2538 inarg->lk_flags |= FUSE_LK_FLOCK;
2539 args->opcode = opcode;
2540 args->nodeid = get_node_id(inode);
2541 args->in_numargs = 1;
2542 args->in_args[0].size = sizeof(*inarg);
2543 args->in_args[0].value = inarg;
2546 static int fuse_getlk(struct file *file, struct file_lock *fl)
2548 struct inode *inode = file_inode(file);
2549 struct fuse_mount *fm = get_fuse_mount(inode);
2551 struct fuse_lk_in inarg;
2552 struct fuse_lk_out outarg;
2555 fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2556 args.out_numargs = 1;
2557 args.out_args[0].size = sizeof(outarg);
2558 args.out_args[0].value = &outarg;
2559 err = fuse_simple_request(fm, &args);
2561 err = convert_fuse_file_lock(fm->fc, &outarg.lk, fl);
2566 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2568 struct inode *inode = file_inode(file);
2569 struct fuse_mount *fm = get_fuse_mount(inode);
2571 struct fuse_lk_in inarg;
2572 int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2573 struct pid *pid = fl->fl_type != F_UNLCK ? task_tgid(current) : NULL;
2574 pid_t pid_nr = pid_nr_ns(pid, fm->fc->pid_ns);
2577 if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2578 /* NLM needs asynchronous locks, which we don't support yet */
2582 /* Unlock on close is handled by the flush method */
2583 if ((fl->fl_flags & FL_CLOSE_POSIX) == FL_CLOSE_POSIX)
2586 fuse_lk_fill(&args, file, fl, opcode, pid_nr, flock, &inarg);
2587 err = fuse_simple_request(fm, &args);
2589 /* locking is restartable */
2596 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2598 struct inode *inode = file_inode(file);
2599 struct fuse_conn *fc = get_fuse_conn(inode);
2602 if (cmd == F_CANCELLK) {
2604 } else if (cmd == F_GETLK) {
2606 posix_test_lock(file, fl);
2609 err = fuse_getlk(file, fl);
2612 err = posix_lock_file(file, fl, NULL);
2614 err = fuse_setlk(file, fl, 0);
2619 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2621 struct inode *inode = file_inode(file);
2622 struct fuse_conn *fc = get_fuse_conn(inode);
2626 err = locks_lock_file_wait(file, fl);
2628 struct fuse_file *ff = file->private_data;
2630 /* emulate flock with POSIX locks */
2632 err = fuse_setlk(file, fl, 1);
2638 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2640 struct inode *inode = mapping->host;
2641 struct fuse_mount *fm = get_fuse_mount(inode);
2643 struct fuse_bmap_in inarg;
2644 struct fuse_bmap_out outarg;
2647 if (!inode->i_sb->s_bdev || fm->fc->no_bmap)
2650 memset(&inarg, 0, sizeof(inarg));
2651 inarg.block = block;
2652 inarg.blocksize = inode->i_sb->s_blocksize;
2653 args.opcode = FUSE_BMAP;
2654 args.nodeid = get_node_id(inode);
2655 args.in_numargs = 1;
2656 args.in_args[0].size = sizeof(inarg);
2657 args.in_args[0].value = &inarg;
2658 args.out_numargs = 1;
2659 args.out_args[0].size = sizeof(outarg);
2660 args.out_args[0].value = &outarg;
2661 err = fuse_simple_request(fm, &args);
2663 fm->fc->no_bmap = 1;
2665 return err ? 0 : outarg.block;
2668 static loff_t fuse_lseek(struct file *file, loff_t offset, int whence)
2670 struct inode *inode = file->f_mapping->host;
2671 struct fuse_mount *fm = get_fuse_mount(inode);
2672 struct fuse_file *ff = file->private_data;
2674 struct fuse_lseek_in inarg = {
2679 struct fuse_lseek_out outarg;
2682 if (fm->fc->no_lseek)
2685 args.opcode = FUSE_LSEEK;
2686 args.nodeid = ff->nodeid;
2687 args.in_numargs = 1;
2688 args.in_args[0].size = sizeof(inarg);
2689 args.in_args[0].value = &inarg;
2690 args.out_numargs = 1;
2691 args.out_args[0].size = sizeof(outarg);
2692 args.out_args[0].value = &outarg;
2693 err = fuse_simple_request(fm, &args);
2695 if (err == -ENOSYS) {
2696 fm->fc->no_lseek = 1;
2702 return vfs_setpos(file, outarg.offset, inode->i_sb->s_maxbytes);
2705 err = fuse_update_attributes(inode, file, STATX_SIZE);
2707 return generic_file_llseek(file, offset, whence);
2712 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2715 struct inode *inode = file_inode(file);
2720 /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2721 retval = generic_file_llseek(file, offset, whence);
2725 retval = fuse_update_attributes(inode, file, STATX_SIZE);
2727 retval = generic_file_llseek(file, offset, whence);
2728 inode_unlock(inode);
2733 retval = fuse_lseek(file, offset, whence);
2734 inode_unlock(inode);
2744 * All files which have been polled are linked to RB tree
2745 * fuse_conn->polled_files which is indexed by kh. Walk the tree and
2746 * find the matching one.
2748 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2749 struct rb_node **parent_out)
2751 struct rb_node **link = &fc->polled_files.rb_node;
2752 struct rb_node *last = NULL;
2755 struct fuse_file *ff;
2758 ff = rb_entry(last, struct fuse_file, polled_node);
2761 link = &last->rb_left;
2762 else if (kh > ff->kh)
2763 link = &last->rb_right;
2774 * The file is about to be polled. Make sure it's on the polled_files
2775 * RB tree. Note that files once added to the polled_files tree are
2776 * not removed before the file is released. This is because a file
2777 * polled once is likely to be polled again.
2779 static void fuse_register_polled_file(struct fuse_conn *fc,
2780 struct fuse_file *ff)
2782 spin_lock(&fc->lock);
2783 if (RB_EMPTY_NODE(&ff->polled_node)) {
2784 struct rb_node **link, *parent;
2786 link = fuse_find_polled_node(fc, ff->kh, &parent);
2788 rb_link_node(&ff->polled_node, parent, link);
2789 rb_insert_color(&ff->polled_node, &fc->polled_files);
2791 spin_unlock(&fc->lock);
2794 __poll_t fuse_file_poll(struct file *file, poll_table *wait)
2796 struct fuse_file *ff = file->private_data;
2797 struct fuse_mount *fm = ff->fm;
2798 struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2799 struct fuse_poll_out outarg;
2803 if (fm->fc->no_poll)
2804 return DEFAULT_POLLMASK;
2806 poll_wait(file, &ff->poll_wait, wait);
2807 inarg.events = mangle_poll(poll_requested_events(wait));
2810 * Ask for notification iff there's someone waiting for it.
2811 * The client may ignore the flag and always notify.
2813 if (waitqueue_active(&ff->poll_wait)) {
2814 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2815 fuse_register_polled_file(fm->fc, ff);
2818 args.opcode = FUSE_POLL;
2819 args.nodeid = ff->nodeid;
2820 args.in_numargs = 1;
2821 args.in_args[0].size = sizeof(inarg);
2822 args.in_args[0].value = &inarg;
2823 args.out_numargs = 1;
2824 args.out_args[0].size = sizeof(outarg);
2825 args.out_args[0].value = &outarg;
2826 err = fuse_simple_request(fm, &args);
2829 return demangle_poll(outarg.revents);
2830 if (err == -ENOSYS) {
2831 fm->fc->no_poll = 1;
2832 return DEFAULT_POLLMASK;
2836 EXPORT_SYMBOL_GPL(fuse_file_poll);
2839 * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2840 * wakes up the poll waiters.
2842 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2843 struct fuse_notify_poll_wakeup_out *outarg)
2845 u64 kh = outarg->kh;
2846 struct rb_node **link;
2848 spin_lock(&fc->lock);
2850 link = fuse_find_polled_node(fc, kh, NULL);
2852 struct fuse_file *ff;
2854 ff = rb_entry(*link, struct fuse_file, polled_node);
2855 wake_up_interruptible_sync(&ff->poll_wait);
2858 spin_unlock(&fc->lock);
2862 static void fuse_do_truncate(struct file *file)
2864 struct inode *inode = file->f_mapping->host;
2867 attr.ia_valid = ATTR_SIZE;
2868 attr.ia_size = i_size_read(inode);
2870 attr.ia_file = file;
2871 attr.ia_valid |= ATTR_FILE;
2873 fuse_do_setattr(file_dentry(file), &attr, file);
2876 static inline loff_t fuse_round_up(struct fuse_conn *fc, loff_t off)
2878 return round_up(off, fc->max_pages << PAGE_SHIFT);
2882 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
2884 DECLARE_COMPLETION_ONSTACK(wait);
2886 struct file *file = iocb->ki_filp;
2887 struct fuse_file *ff = file->private_data;
2889 struct inode *inode;
2891 size_t count = iov_iter_count(iter), shortened = 0;
2892 loff_t offset = iocb->ki_pos;
2893 struct fuse_io_priv *io;
2896 inode = file->f_mapping->host;
2897 i_size = i_size_read(inode);
2899 if ((iov_iter_rw(iter) == READ) && (offset >= i_size))
2902 io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2905 spin_lock_init(&io->lock);
2906 kref_init(&io->refcnt);
2910 io->offset = offset;
2911 io->write = (iov_iter_rw(iter) == WRITE);
2914 * By default, we want to optimize all I/Os with async request
2915 * submission to the client filesystem if supported.
2917 io->async = ff->fm->fc->async_dio;
2919 io->blocking = is_sync_kiocb(iocb);
2921 /* optimization for short read */
2922 if (io->async && !io->write && offset + count > i_size) {
2923 iov_iter_truncate(iter, fuse_round_up(ff->fm->fc, i_size - offset));
2924 shortened = count - iov_iter_count(iter);
2929 * We cannot asynchronously extend the size of a file.
2930 * In such case the aio will behave exactly like sync io.
2932 if ((offset + count > i_size) && io->write)
2933 io->blocking = true;
2935 if (io->async && io->blocking) {
2937 * Additional reference to keep io around after
2938 * calling fuse_aio_complete()
2940 kref_get(&io->refcnt);
2944 if (iov_iter_rw(iter) == WRITE) {
2945 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
2946 fuse_invalidate_attr_mask(inode, FUSE_STATX_MODSIZE);
2948 ret = __fuse_direct_read(io, iter, &pos);
2950 iov_iter_reexpand(iter, iov_iter_count(iter) + shortened);
2953 bool blocking = io->blocking;
2955 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2957 /* we have a non-extending, async request, so return */
2959 return -EIOCBQUEUED;
2961 wait_for_completion(&wait);
2962 ret = fuse_get_res_by_io(io);
2965 kref_put(&io->refcnt, fuse_io_release);
2967 if (iov_iter_rw(iter) == WRITE) {
2968 fuse_write_update_attr(inode, pos, ret);
2969 /* For extending writes we already hold exclusive lock */
2970 if (ret < 0 && offset + count > i_size)
2971 fuse_do_truncate(file);
2977 static int fuse_writeback_range(struct inode *inode, loff_t start, loff_t end)
2979 int err = filemap_write_and_wait_range(inode->i_mapping, start, LLONG_MAX);
2982 fuse_sync_writes(inode);
2987 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2990 struct fuse_file *ff = file->private_data;
2991 struct inode *inode = file_inode(file);
2992 struct fuse_inode *fi = get_fuse_inode(inode);
2993 struct fuse_mount *fm = ff->fm;
2995 struct fuse_fallocate_in inarg = {
3002 bool block_faults = FUSE_IS_DAX(inode) &&
3003 (!(mode & FALLOC_FL_KEEP_SIZE) ||
3004 (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE)));
3006 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3007 FALLOC_FL_ZERO_RANGE))
3010 if (fm->fc->no_fallocate)
3015 filemap_invalidate_lock(inode->i_mapping);
3016 err = fuse_dax_break_layouts(inode, 0, 0);
3021 if (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE)) {
3022 loff_t endbyte = offset + length - 1;
3024 err = fuse_writeback_range(inode, offset, endbyte);
3029 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
3030 offset + length > i_size_read(inode)) {
3031 err = inode_newsize_ok(inode, offset + length);
3036 err = file_modified(file);
3040 if (!(mode & FALLOC_FL_KEEP_SIZE))
3041 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3043 args.opcode = FUSE_FALLOCATE;
3044 args.nodeid = ff->nodeid;
3045 args.in_numargs = 1;
3046 args.in_args[0].size = sizeof(inarg);
3047 args.in_args[0].value = &inarg;
3048 err = fuse_simple_request(fm, &args);
3049 if (err == -ENOSYS) {
3050 fm->fc->no_fallocate = 1;
3056 /* we could have extended the file */
3057 if (!(mode & FALLOC_FL_KEEP_SIZE)) {
3058 if (fuse_write_update_attr(inode, offset + length, length))
3059 file_update_time(file);
3062 if (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE))
3063 truncate_pagecache_range(inode, offset, offset + length - 1);
3065 fuse_invalidate_attr_mask(inode, FUSE_STATX_MODSIZE);
3068 if (!(mode & FALLOC_FL_KEEP_SIZE))
3069 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3072 filemap_invalidate_unlock(inode->i_mapping);
3074 inode_unlock(inode);
3076 fuse_flush_time_update(inode);
3081 static ssize_t __fuse_copy_file_range(struct file *file_in, loff_t pos_in,
3082 struct file *file_out, loff_t pos_out,
3083 size_t len, unsigned int flags)
3085 struct fuse_file *ff_in = file_in->private_data;
3086 struct fuse_file *ff_out = file_out->private_data;
3087 struct inode *inode_in = file_inode(file_in);
3088 struct inode *inode_out = file_inode(file_out);
3089 struct fuse_inode *fi_out = get_fuse_inode(inode_out);
3090 struct fuse_mount *fm = ff_in->fm;
3091 struct fuse_conn *fc = fm->fc;
3093 struct fuse_copy_file_range_in inarg = {
3096 .nodeid_out = ff_out->nodeid,
3097 .fh_out = ff_out->fh,
3102 struct fuse_write_out outarg;
3104 /* mark unstable when write-back is not used, and file_out gets
3106 bool is_unstable = (!fc->writeback_cache) &&
3107 ((pos_out + len) > inode_out->i_size);
3109 if (fc->no_copy_file_range)
3112 if (file_inode(file_in)->i_sb != file_inode(file_out)->i_sb)
3115 inode_lock(inode_in);
3116 err = fuse_writeback_range(inode_in, pos_in, pos_in + len - 1);
3117 inode_unlock(inode_in);
3121 inode_lock(inode_out);
3123 err = file_modified(file_out);
3128 * Write out dirty pages in the destination file before sending the COPY
3129 * request to userspace. After the request is completed, truncate off
3130 * pages (including partial ones) from the cache that have been copied,
3131 * since these contain stale data at that point.
3133 * This should be mostly correct, but if the COPY writes to partial
3134 * pages (at the start or end) and the parts not covered by the COPY are
3135 * written through a memory map after calling fuse_writeback_range(),
3136 * then these partial page modifications will be lost on truncation.
3138 * It is unlikely that someone would rely on such mixed style
3139 * modifications. Yet this does give less guarantees than if the
3140 * copying was performed with write(2).
3142 * To fix this a mapping->invalidate_lock could be used to prevent new
3143 * faults while the copy is ongoing.
3145 err = fuse_writeback_range(inode_out, pos_out, pos_out + len - 1);
3150 set_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3152 args.opcode = FUSE_COPY_FILE_RANGE;
3153 args.nodeid = ff_in->nodeid;
3154 args.in_numargs = 1;
3155 args.in_args[0].size = sizeof(inarg);
3156 args.in_args[0].value = &inarg;
3157 args.out_numargs = 1;
3158 args.out_args[0].size = sizeof(outarg);
3159 args.out_args[0].value = &outarg;
3160 err = fuse_simple_request(fm, &args);
3161 if (err == -ENOSYS) {
3162 fc->no_copy_file_range = 1;
3168 truncate_inode_pages_range(inode_out->i_mapping,
3169 ALIGN_DOWN(pos_out, PAGE_SIZE),
3170 ALIGN(pos_out + outarg.size, PAGE_SIZE) - 1);
3172 file_update_time(file_out);
3173 fuse_write_update_attr(inode_out, pos_out + outarg.size, outarg.size);
3178 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3180 inode_unlock(inode_out);
3181 file_accessed(file_in);
3183 fuse_flush_time_update(inode_out);
3188 static ssize_t fuse_copy_file_range(struct file *src_file, loff_t src_off,
3189 struct file *dst_file, loff_t dst_off,
3190 size_t len, unsigned int flags)
3194 ret = __fuse_copy_file_range(src_file, src_off, dst_file, dst_off,
3197 if (ret == -EOPNOTSUPP || ret == -EXDEV)
3198 ret = generic_copy_file_range(src_file, src_off, dst_file,
3199 dst_off, len, flags);
3203 static const struct file_operations fuse_file_operations = {
3204 .llseek = fuse_file_llseek,
3205 .read_iter = fuse_file_read_iter,
3206 .write_iter = fuse_file_write_iter,
3207 .mmap = fuse_file_mmap,
3209 .flush = fuse_flush,
3210 .release = fuse_release,
3211 .fsync = fuse_fsync,
3212 .lock = fuse_file_lock,
3213 .get_unmapped_area = thp_get_unmapped_area,
3214 .flock = fuse_file_flock,
3215 .splice_read = filemap_splice_read,
3216 .splice_write = iter_file_splice_write,
3217 .unlocked_ioctl = fuse_file_ioctl,
3218 .compat_ioctl = fuse_file_compat_ioctl,
3219 .poll = fuse_file_poll,
3220 .fallocate = fuse_file_fallocate,
3221 .copy_file_range = fuse_copy_file_range,
3224 static const struct address_space_operations fuse_file_aops = {
3225 .read_folio = fuse_read_folio,
3226 .readahead = fuse_readahead,
3227 .writepage = fuse_writepage,
3228 .writepages = fuse_writepages,
3229 .launder_folio = fuse_launder_folio,
3230 .dirty_folio = filemap_dirty_folio,
3232 .direct_IO = fuse_direct_IO,
3233 .write_begin = fuse_write_begin,
3234 .write_end = fuse_write_end,
3237 void fuse_init_file_inode(struct inode *inode, unsigned int flags)
3239 struct fuse_inode *fi = get_fuse_inode(inode);
3241 inode->i_fop = &fuse_file_operations;
3242 inode->i_data.a_ops = &fuse_file_aops;
3244 INIT_LIST_HEAD(&fi->write_files);
3245 INIT_LIST_HEAD(&fi->queued_writes);
3247 init_waitqueue_head(&fi->page_waitq);
3248 fi->writepages = RB_ROOT;
3250 if (IS_ENABLED(CONFIG_FUSE_DAX))
3251 fuse_dax_inode_init(inode, flags);