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/module.h>
16 #include <linux/compat.h>
17 #include <linux/swap.h>
18 #include <linux/falloc.h>
19 #include <linux/uio.h>
21 static const struct file_operations fuse_direct_io_file_operations;
23 static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
24 int opcode, struct fuse_open_out *outargp)
26 struct fuse_open_in inarg;
29 memset(&inarg, 0, sizeof(inarg));
30 inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
31 if (!fc->atomic_o_trunc)
32 inarg.flags &= ~O_TRUNC;
33 args.in.h.opcode = opcode;
34 args.in.h.nodeid = nodeid;
36 args.in.args[0].size = sizeof(inarg);
37 args.in.args[0].value = &inarg;
39 args.out.args[0].size = sizeof(*outargp);
40 args.out.args[0].value = outargp;
42 return fuse_simple_request(fc, &args);
45 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
49 ff = kmalloc(sizeof(struct fuse_file), GFP_KERNEL);
54 ff->reserved_req = fuse_request_alloc(0);
55 if (unlikely(!ff->reserved_req)) {
60 INIT_LIST_HEAD(&ff->write_entry);
61 atomic_set(&ff->count, 0);
62 RB_CLEAR_NODE(&ff->polled_node);
63 init_waitqueue_head(&ff->poll_wait);
67 spin_unlock(&fc->lock);
72 void fuse_file_free(struct fuse_file *ff)
74 fuse_request_free(ff->reserved_req);
78 struct fuse_file *fuse_file_get(struct fuse_file *ff)
80 atomic_inc(&ff->count);
84 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
86 iput(req->misc.release.inode);
89 static void fuse_file_put(struct fuse_file *ff, bool sync)
91 if (atomic_dec_and_test(&ff->count)) {
92 struct fuse_req *req = ff->reserved_req;
94 if (ff->fc->no_open) {
96 * Drop the release request when client does not
99 __clear_bit(FR_BACKGROUND, &req->flags);
100 iput(req->misc.release.inode);
101 fuse_put_request(ff->fc, req);
103 __clear_bit(FR_BACKGROUND, &req->flags);
104 fuse_request_send(ff->fc, req);
105 iput(req->misc.release.inode);
106 fuse_put_request(ff->fc, req);
108 req->end = fuse_release_end;
109 __set_bit(FR_BACKGROUND, &req->flags);
110 fuse_request_send_background(ff->fc, req);
116 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
119 struct fuse_file *ff;
120 int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
122 ff = fuse_file_alloc(fc);
127 ff->open_flags = FOPEN_KEEP_CACHE; /* Default for no-open */
128 if (!fc->no_open || isdir) {
129 struct fuse_open_out outarg;
132 err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
135 ff->open_flags = outarg.open_flags;
137 } else if (err != -ENOSYS || isdir) {
146 ff->open_flags &= ~FOPEN_DIRECT_IO;
149 file->private_data = fuse_file_get(ff);
153 EXPORT_SYMBOL_GPL(fuse_do_open);
155 static void fuse_link_write_file(struct file *file)
157 struct inode *inode = file_inode(file);
158 struct fuse_conn *fc = get_fuse_conn(inode);
159 struct fuse_inode *fi = get_fuse_inode(inode);
160 struct fuse_file *ff = file->private_data;
162 * file may be written through mmap, so chain it onto the
163 * inodes's write_file list
165 spin_lock(&fc->lock);
166 if (list_empty(&ff->write_entry))
167 list_add(&ff->write_entry, &fi->write_files);
168 spin_unlock(&fc->lock);
171 void fuse_finish_open(struct inode *inode, struct file *file)
173 struct fuse_file *ff = file->private_data;
174 struct fuse_conn *fc = get_fuse_conn(inode);
176 if (ff->open_flags & FOPEN_DIRECT_IO)
177 file->f_op = &fuse_direct_io_file_operations;
178 if (!(ff->open_flags & FOPEN_KEEP_CACHE))
179 invalidate_inode_pages2(inode->i_mapping);
180 if (ff->open_flags & FOPEN_NONSEEKABLE)
181 nonseekable_open(inode, file);
182 if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
183 struct fuse_inode *fi = get_fuse_inode(inode);
185 spin_lock(&fc->lock);
186 fi->attr_version = ++fc->attr_version;
187 i_size_write(inode, 0);
188 spin_unlock(&fc->lock);
189 fuse_invalidate_attr(inode);
190 if (fc->writeback_cache)
191 file_update_time(file);
193 if ((file->f_mode & FMODE_WRITE) && fc->writeback_cache)
194 fuse_link_write_file(file);
197 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
199 struct fuse_conn *fc = get_fuse_conn(inode);
201 bool lock_inode = (file->f_flags & O_TRUNC) &&
202 fc->atomic_o_trunc &&
205 err = generic_file_open(inode, file);
210 mutex_lock(&inode->i_mutex);
212 err = fuse_do_open(fc, get_node_id(inode), file, isdir);
215 fuse_finish_open(inode, file);
218 mutex_unlock(&inode->i_mutex);
223 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
225 struct fuse_conn *fc = ff->fc;
226 struct fuse_req *req = ff->reserved_req;
227 struct fuse_release_in *inarg = &req->misc.release.in;
229 spin_lock(&fc->lock);
230 list_del(&ff->write_entry);
231 if (!RB_EMPTY_NODE(&ff->polled_node))
232 rb_erase(&ff->polled_node, &fc->polled_files);
233 spin_unlock(&fc->lock);
235 wake_up_interruptible_all(&ff->poll_wait);
238 inarg->flags = flags;
239 req->in.h.opcode = opcode;
240 req->in.h.nodeid = ff->nodeid;
242 req->in.args[0].size = sizeof(struct fuse_release_in);
243 req->in.args[0].value = inarg;
246 void fuse_release_common(struct file *file, int opcode)
248 struct fuse_file *ff;
249 struct fuse_req *req;
251 ff = file->private_data;
255 req = ff->reserved_req;
256 fuse_prepare_release(ff, file->f_flags, opcode);
259 struct fuse_release_in *inarg = &req->misc.release.in;
260 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
261 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
264 /* Hold inode until release is finished */
265 req->misc.release.inode = igrab(file_inode(file));
268 * Normally this will send the RELEASE request, however if
269 * some asynchronous READ or WRITE requests are outstanding,
270 * the sending will be delayed.
272 * Make the release synchronous if this is a fuseblk mount,
273 * synchronous RELEASE is allowed (and desirable) in this case
274 * because the server can be trusted not to screw up.
276 fuse_file_put(ff, ff->fc->destroy_req != NULL);
279 static int fuse_open(struct inode *inode, struct file *file)
281 return fuse_open_common(inode, file, false);
284 static int fuse_release(struct inode *inode, struct file *file)
286 struct fuse_conn *fc = get_fuse_conn(inode);
288 /* see fuse_vma_close() for !writeback_cache case */
289 if (fc->writeback_cache)
290 write_inode_now(inode, 1);
292 fuse_release_common(file, FUSE_RELEASE);
294 /* return value is ignored by VFS */
298 void fuse_sync_release(struct fuse_file *ff, int flags)
300 WARN_ON(atomic_read(&ff->count) > 1);
301 fuse_prepare_release(ff, flags, FUSE_RELEASE);
302 __set_bit(FR_FORCE, &ff->reserved_req->flags);
303 __clear_bit(FR_BACKGROUND, &ff->reserved_req->flags);
304 fuse_request_send(ff->fc, ff->reserved_req);
305 fuse_put_request(ff->fc, ff->reserved_req);
308 EXPORT_SYMBOL_GPL(fuse_sync_release);
311 * Scramble the ID space with XTEA, so that the value of the files_struct
312 * pointer is not exposed to userspace.
314 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
316 u32 *k = fc->scramble_key;
317 u64 v = (unsigned long) id;
323 for (i = 0; i < 32; i++) {
324 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
326 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
329 return (u64) v0 + ((u64) v1 << 32);
333 * Check if any page in a range is under writeback
335 * This is currently done by walking the list of writepage requests
336 * for the inode, which can be pretty inefficient.
338 static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
341 struct fuse_conn *fc = get_fuse_conn(inode);
342 struct fuse_inode *fi = get_fuse_inode(inode);
343 struct fuse_req *req;
346 spin_lock(&fc->lock);
347 list_for_each_entry(req, &fi->writepages, writepages_entry) {
350 BUG_ON(req->inode != inode);
351 curr_index = req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
352 if (idx_from < curr_index + req->num_pages &&
353 curr_index <= idx_to) {
358 spin_unlock(&fc->lock);
363 static inline bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
365 return fuse_range_is_writeback(inode, index, index);
369 * Wait for page writeback to be completed.
371 * Since fuse doesn't rely on the VM writeback tracking, this has to
372 * use some other means.
374 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
376 struct fuse_inode *fi = get_fuse_inode(inode);
378 wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
383 * Wait for all pending writepages on the inode to finish.
385 * This is currently done by blocking further writes with FUSE_NOWRITE
386 * and waiting for all sent writes to complete.
388 * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
389 * could conflict with truncation.
391 static void fuse_sync_writes(struct inode *inode)
393 fuse_set_nowrite(inode);
394 fuse_release_nowrite(inode);
397 static int fuse_flush(struct file *file, fl_owner_t id)
399 struct inode *inode = file_inode(file);
400 struct fuse_conn *fc = get_fuse_conn(inode);
401 struct fuse_file *ff = file->private_data;
402 struct fuse_req *req;
403 struct fuse_flush_in inarg;
406 if (is_bad_inode(inode))
412 err = write_inode_now(inode, 1);
416 mutex_lock(&inode->i_mutex);
417 fuse_sync_writes(inode);
418 mutex_unlock(&inode->i_mutex);
420 req = fuse_get_req_nofail_nopages(fc, file);
421 memset(&inarg, 0, sizeof(inarg));
423 inarg.lock_owner = fuse_lock_owner_id(fc, id);
424 req->in.h.opcode = FUSE_FLUSH;
425 req->in.h.nodeid = get_node_id(inode);
427 req->in.args[0].size = sizeof(inarg);
428 req->in.args[0].value = &inarg;
429 __set_bit(FR_FORCE, &req->flags);
430 fuse_request_send(fc, req);
431 err = req->out.h.error;
432 fuse_put_request(fc, req);
433 if (err == -ENOSYS) {
440 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
441 int datasync, int isdir)
443 struct inode *inode = file->f_mapping->host;
444 struct fuse_conn *fc = get_fuse_conn(inode);
445 struct fuse_file *ff = file->private_data;
447 struct fuse_fsync_in inarg;
450 if (is_bad_inode(inode))
453 mutex_lock(&inode->i_mutex);
456 * Start writeback against all dirty pages of the inode, then
457 * wait for all outstanding writes, before sending the FSYNC
460 err = filemap_write_and_wait_range(inode->i_mapping, start, end);
464 fuse_sync_writes(inode);
465 err = sync_inode_metadata(inode, 1);
469 if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
472 memset(&inarg, 0, sizeof(inarg));
474 inarg.fsync_flags = datasync ? 1 : 0;
475 args.in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
476 args.in.h.nodeid = get_node_id(inode);
478 args.in.args[0].size = sizeof(inarg);
479 args.in.args[0].value = &inarg;
480 err = fuse_simple_request(fc, &args);
481 if (err == -ENOSYS) {
489 mutex_unlock(&inode->i_mutex);
493 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
496 return fuse_fsync_common(file, start, end, datasync, 0);
499 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
500 size_t count, int opcode)
502 struct fuse_read_in *inarg = &req->misc.read.in;
503 struct fuse_file *ff = file->private_data;
508 inarg->flags = file->f_flags;
509 req->in.h.opcode = opcode;
510 req->in.h.nodeid = ff->nodeid;
512 req->in.args[0].size = sizeof(struct fuse_read_in);
513 req->in.args[0].value = inarg;
515 req->out.numargs = 1;
516 req->out.args[0].size = count;
519 static void fuse_release_user_pages(struct fuse_req *req, int write)
523 for (i = 0; i < req->num_pages; i++) {
524 struct page *page = req->pages[i];
526 set_page_dirty_lock(page);
531 static void fuse_io_release(struct kref *kref)
533 kfree(container_of(kref, struct fuse_io_priv, refcnt));
536 static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
541 if (io->bytes >= 0 && io->write)
544 return io->bytes < 0 ? io->size : io->bytes;
548 * In case of short read, the caller sets 'pos' to the position of
549 * actual end of fuse request in IO request. Otherwise, if bytes_requested
550 * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
553 * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
554 * both submitted asynchronously. The first of them was ACKed by userspace as
555 * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
556 * second request was ACKed as short, e.g. only 1K was read, resulting in
559 * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
560 * will be equal to the length of the longest contiguous fragment of
561 * transferred data starting from the beginning of IO request.
563 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
565 bool is_sync = is_sync_kiocb(io->iocb);
568 spin_lock(&io->lock);
570 io->err = io->err ? : err;
571 else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
575 if (!left && is_sync)
577 spin_unlock(&io->lock);
579 if (!left && !is_sync) {
580 ssize_t res = fuse_get_res_by_io(io);
583 struct inode *inode = file_inode(io->iocb->ki_filp);
584 struct fuse_conn *fc = get_fuse_conn(inode);
585 struct fuse_inode *fi = get_fuse_inode(inode);
587 spin_lock(&fc->lock);
588 fi->attr_version = ++fc->attr_version;
589 spin_unlock(&fc->lock);
592 io->iocb->ki_complete(io->iocb, res, 0);
595 kref_put(&io->refcnt, fuse_io_release);
598 static void fuse_aio_complete_req(struct fuse_conn *fc, struct fuse_req *req)
600 struct fuse_io_priv *io = req->io;
603 fuse_release_user_pages(req, !io->write);
606 if (req->misc.write.in.size != req->misc.write.out.size)
607 pos = req->misc.write.in.offset - io->offset +
608 req->misc.write.out.size;
610 if (req->misc.read.in.size != req->out.args[0].size)
611 pos = req->misc.read.in.offset - io->offset +
612 req->out.args[0].size;
615 fuse_aio_complete(io, req->out.h.error, pos);
618 static size_t fuse_async_req_send(struct fuse_conn *fc, struct fuse_req *req,
619 size_t num_bytes, struct fuse_io_priv *io)
621 spin_lock(&io->lock);
622 kref_get(&io->refcnt);
623 io->size += num_bytes;
625 spin_unlock(&io->lock);
628 req->end = fuse_aio_complete_req;
630 __fuse_get_request(req);
631 fuse_request_send_background(fc, req);
636 static size_t fuse_send_read(struct fuse_req *req, struct fuse_io_priv *io,
637 loff_t pos, size_t count, fl_owner_t owner)
639 struct file *file = io->file;
640 struct fuse_file *ff = file->private_data;
641 struct fuse_conn *fc = ff->fc;
643 fuse_read_fill(req, file, pos, count, FUSE_READ);
645 struct fuse_read_in *inarg = &req->misc.read.in;
647 inarg->read_flags |= FUSE_READ_LOCKOWNER;
648 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
652 return fuse_async_req_send(fc, req, count, io);
654 fuse_request_send(fc, req);
655 return req->out.args[0].size;
658 static void fuse_read_update_size(struct inode *inode, loff_t size,
661 struct fuse_conn *fc = get_fuse_conn(inode);
662 struct fuse_inode *fi = get_fuse_inode(inode);
664 spin_lock(&fc->lock);
665 if (attr_ver == fi->attr_version && size < inode->i_size &&
666 !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
667 fi->attr_version = ++fc->attr_version;
668 i_size_write(inode, size);
670 spin_unlock(&fc->lock);
673 static void fuse_short_read(struct fuse_req *req, struct inode *inode,
676 size_t num_read = req->out.args[0].size;
677 struct fuse_conn *fc = get_fuse_conn(inode);
679 if (fc->writeback_cache) {
681 * A hole in a file. Some data after the hole are in page cache,
682 * but have not reached the client fs yet. So, the hole is not
686 int start_idx = num_read >> PAGE_CACHE_SHIFT;
687 size_t off = num_read & (PAGE_CACHE_SIZE - 1);
689 for (i = start_idx; i < req->num_pages; i++) {
690 zero_user_segment(req->pages[i], off, PAGE_CACHE_SIZE);
694 loff_t pos = page_offset(req->pages[0]) + num_read;
695 fuse_read_update_size(inode, pos, attr_ver);
699 static int fuse_do_readpage(struct file *file, struct page *page)
701 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
702 struct inode *inode = page->mapping->host;
703 struct fuse_conn *fc = get_fuse_conn(inode);
704 struct fuse_req *req;
706 loff_t pos = page_offset(page);
707 size_t count = PAGE_CACHE_SIZE;
712 * Page writeback can extend beyond the lifetime of the
713 * page-cache page, so make sure we read a properly synced
716 fuse_wait_on_page_writeback(inode, page->index);
718 req = fuse_get_req(fc, 1);
722 attr_ver = fuse_get_attr_version(fc);
724 req->out.page_zeroing = 1;
725 req->out.argpages = 1;
727 req->pages[0] = page;
728 req->page_descs[0].length = count;
729 num_read = fuse_send_read(req, &io, pos, count, NULL);
730 err = req->out.h.error;
734 * Short read means EOF. If file size is larger, truncate it
736 if (num_read < count)
737 fuse_short_read(req, inode, attr_ver);
739 SetPageUptodate(page);
742 fuse_put_request(fc, req);
747 static int fuse_readpage(struct file *file, struct page *page)
749 struct inode *inode = page->mapping->host;
753 if (is_bad_inode(inode))
756 err = fuse_do_readpage(file, page);
757 fuse_invalidate_atime(inode);
763 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
766 size_t count = req->misc.read.in.size;
767 size_t num_read = req->out.args[0].size;
768 struct address_space *mapping = NULL;
770 for (i = 0; mapping == NULL && i < req->num_pages; i++)
771 mapping = req->pages[i]->mapping;
774 struct inode *inode = mapping->host;
777 * Short read means EOF. If file size is larger, truncate it
779 if (!req->out.h.error && num_read < count)
780 fuse_short_read(req, inode, req->misc.read.attr_ver);
782 fuse_invalidate_atime(inode);
785 for (i = 0; i < req->num_pages; i++) {
786 struct page *page = req->pages[i];
787 if (!req->out.h.error)
788 SetPageUptodate(page);
792 page_cache_release(page);
795 fuse_file_put(req->ff, false);
798 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
800 struct fuse_file *ff = file->private_data;
801 struct fuse_conn *fc = ff->fc;
802 loff_t pos = page_offset(req->pages[0]);
803 size_t count = req->num_pages << PAGE_CACHE_SHIFT;
805 req->out.argpages = 1;
806 req->out.page_zeroing = 1;
807 req->out.page_replace = 1;
808 fuse_read_fill(req, file, pos, count, FUSE_READ);
809 req->misc.read.attr_ver = fuse_get_attr_version(fc);
810 if (fc->async_read) {
811 req->ff = fuse_file_get(ff);
812 req->end = fuse_readpages_end;
813 fuse_request_send_background(fc, req);
815 fuse_request_send(fc, req);
816 fuse_readpages_end(fc, req);
817 fuse_put_request(fc, req);
821 struct fuse_fill_data {
822 struct fuse_req *req;
828 static int fuse_readpages_fill(void *_data, struct page *page)
830 struct fuse_fill_data *data = _data;
831 struct fuse_req *req = data->req;
832 struct inode *inode = data->inode;
833 struct fuse_conn *fc = get_fuse_conn(inode);
835 fuse_wait_on_page_writeback(inode, page->index);
837 if (req->num_pages &&
838 (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
839 (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
840 req->pages[req->num_pages - 1]->index + 1 != page->index)) {
841 int nr_alloc = min_t(unsigned, data->nr_pages,
842 FUSE_MAX_PAGES_PER_REQ);
843 fuse_send_readpages(req, data->file);
845 req = fuse_get_req_for_background(fc, nr_alloc);
847 req = fuse_get_req(fc, nr_alloc);
856 if (WARN_ON(req->num_pages >= req->max_pages)) {
857 fuse_put_request(fc, req);
861 page_cache_get(page);
862 req->pages[req->num_pages] = page;
863 req->page_descs[req->num_pages].length = PAGE_SIZE;
869 static int fuse_readpages(struct file *file, struct address_space *mapping,
870 struct list_head *pages, unsigned nr_pages)
872 struct inode *inode = mapping->host;
873 struct fuse_conn *fc = get_fuse_conn(inode);
874 struct fuse_fill_data data;
876 int nr_alloc = min_t(unsigned, nr_pages, FUSE_MAX_PAGES_PER_REQ);
879 if (is_bad_inode(inode))
885 data.req = fuse_get_req_for_background(fc, nr_alloc);
887 data.req = fuse_get_req(fc, nr_alloc);
888 data.nr_pages = nr_pages;
889 err = PTR_ERR(data.req);
890 if (IS_ERR(data.req))
893 err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
895 if (data.req->num_pages)
896 fuse_send_readpages(data.req, file);
898 fuse_put_request(fc, data.req);
904 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
906 struct inode *inode = iocb->ki_filp->f_mapping->host;
907 struct fuse_conn *fc = get_fuse_conn(inode);
910 * In auto invalidate mode, always update attributes on read.
911 * Otherwise, only update if we attempt to read past EOF (to ensure
912 * i_size is up to date).
914 if (fc->auto_inval_data ||
915 (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
917 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
922 return generic_file_read_iter(iocb, to);
925 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
926 loff_t pos, size_t count)
928 struct fuse_write_in *inarg = &req->misc.write.in;
929 struct fuse_write_out *outarg = &req->misc.write.out;
934 req->in.h.opcode = FUSE_WRITE;
935 req->in.h.nodeid = ff->nodeid;
937 if (ff->fc->minor < 9)
938 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
940 req->in.args[0].size = sizeof(struct fuse_write_in);
941 req->in.args[0].value = inarg;
942 req->in.args[1].size = count;
943 req->out.numargs = 1;
944 req->out.args[0].size = sizeof(struct fuse_write_out);
945 req->out.args[0].value = outarg;
948 static size_t fuse_send_write(struct fuse_req *req, struct fuse_io_priv *io,
949 loff_t pos, size_t count, fl_owner_t owner)
951 struct file *file = io->file;
952 struct fuse_file *ff = file->private_data;
953 struct fuse_conn *fc = ff->fc;
954 struct fuse_write_in *inarg = &req->misc.write.in;
956 fuse_write_fill(req, ff, pos, count);
957 inarg->flags = file->f_flags;
959 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
960 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
964 return fuse_async_req_send(fc, req, count, io);
966 fuse_request_send(fc, req);
967 return req->misc.write.out.size;
970 bool fuse_write_update_size(struct inode *inode, loff_t pos)
972 struct fuse_conn *fc = get_fuse_conn(inode);
973 struct fuse_inode *fi = get_fuse_inode(inode);
976 spin_lock(&fc->lock);
977 fi->attr_version = ++fc->attr_version;
978 if (pos > inode->i_size) {
979 i_size_write(inode, pos);
982 spin_unlock(&fc->lock);
987 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
988 struct inode *inode, loff_t pos,
994 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
996 for (i = 0; i < req->num_pages; i++)
997 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
999 res = fuse_send_write(req, &io, pos, count, NULL);
1001 offset = req->page_descs[0].offset;
1003 for (i = 0; i < req->num_pages; i++) {
1004 struct page *page = req->pages[i];
1006 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
1007 SetPageUptodate(page);
1009 if (count > PAGE_CACHE_SIZE - offset)
1010 count -= PAGE_CACHE_SIZE - offset;
1016 page_cache_release(page);
1022 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
1023 struct address_space *mapping,
1024 struct iov_iter *ii, loff_t pos)
1026 struct fuse_conn *fc = get_fuse_conn(mapping->host);
1027 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
1031 req->in.argpages = 1;
1032 req->page_descs[0].offset = offset;
1037 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1038 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
1039 iov_iter_count(ii));
1041 bytes = min_t(size_t, bytes, fc->max_write - count);
1045 if (iov_iter_fault_in_readable(ii, bytes))
1049 page = grab_cache_page_write_begin(mapping, index, 0);
1053 if (mapping_writably_mapped(mapping))
1054 flush_dcache_page(page);
1056 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
1057 flush_dcache_page(page);
1059 iov_iter_advance(ii, tmp);
1062 page_cache_release(page);
1063 bytes = min(bytes, iov_iter_single_seg_count(ii));
1068 req->pages[req->num_pages] = page;
1069 req->page_descs[req->num_pages].length = tmp;
1075 if (offset == PAGE_CACHE_SIZE)
1078 if (!fc->big_writes)
1080 } while (iov_iter_count(ii) && count < fc->max_write &&
1081 req->num_pages < req->max_pages && offset == 0);
1083 return count > 0 ? count : err;
1086 static inline unsigned fuse_wr_pages(loff_t pos, size_t len)
1088 return min_t(unsigned,
1089 ((pos + len - 1) >> PAGE_CACHE_SHIFT) -
1090 (pos >> PAGE_CACHE_SHIFT) + 1,
1091 FUSE_MAX_PAGES_PER_REQ);
1094 static ssize_t fuse_perform_write(struct file *file,
1095 struct address_space *mapping,
1096 struct iov_iter *ii, loff_t pos)
1098 struct inode *inode = mapping->host;
1099 struct fuse_conn *fc = get_fuse_conn(inode);
1100 struct fuse_inode *fi = get_fuse_inode(inode);
1104 if (is_bad_inode(inode))
1107 if (inode->i_size < pos + iov_iter_count(ii))
1108 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1111 struct fuse_req *req;
1113 unsigned nr_pages = fuse_wr_pages(pos, iov_iter_count(ii));
1115 req = fuse_get_req(fc, nr_pages);
1121 count = fuse_fill_write_pages(req, mapping, ii, pos);
1127 num_written = fuse_send_write_pages(req, file, inode,
1129 err = req->out.h.error;
1134 /* break out of the loop on short write */
1135 if (num_written != count)
1139 fuse_put_request(fc, req);
1140 } while (!err && iov_iter_count(ii));
1143 fuse_write_update_size(inode, pos);
1145 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1146 fuse_invalidate_attr(inode);
1148 return res > 0 ? res : err;
1151 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1153 struct file *file = iocb->ki_filp;
1154 struct address_space *mapping = file->f_mapping;
1155 ssize_t written = 0;
1156 ssize_t written_buffered = 0;
1157 struct inode *inode = mapping->host;
1161 if (get_fuse_conn(inode)->writeback_cache) {
1162 /* Update size (EOF optimization) and mode (SUID clearing) */
1163 err = fuse_update_attributes(mapping->host, NULL, file, NULL);
1167 return generic_file_write_iter(iocb, from);
1170 mutex_lock(&inode->i_mutex);
1172 /* We can write back this queue in page reclaim */
1173 current->backing_dev_info = inode_to_bdi(inode);
1175 err = generic_write_checks(iocb, from);
1179 err = file_remove_privs(file);
1183 err = file_update_time(file);
1187 if (iocb->ki_flags & IOCB_DIRECT) {
1188 loff_t pos = iocb->ki_pos;
1189 written = generic_file_direct_write(iocb, from, pos);
1190 if (written < 0 || !iov_iter_count(from))
1195 written_buffered = fuse_perform_write(file, mapping, from, pos);
1196 if (written_buffered < 0) {
1197 err = written_buffered;
1200 endbyte = pos + written_buffered - 1;
1202 err = filemap_write_and_wait_range(file->f_mapping, pos,
1207 invalidate_mapping_pages(file->f_mapping,
1208 pos >> PAGE_CACHE_SHIFT,
1209 endbyte >> PAGE_CACHE_SHIFT);
1211 written += written_buffered;
1212 iocb->ki_pos = pos + written_buffered;
1214 written = fuse_perform_write(file, mapping, from, iocb->ki_pos);
1216 iocb->ki_pos += written;
1219 current->backing_dev_info = NULL;
1220 mutex_unlock(&inode->i_mutex);
1222 return written ? written : err;
1225 static inline void fuse_page_descs_length_init(struct fuse_req *req,
1226 unsigned index, unsigned nr_pages)
1230 for (i = index; i < index + nr_pages; i++)
1231 req->page_descs[i].length = PAGE_SIZE -
1232 req->page_descs[i].offset;
1235 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1237 return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1240 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1243 return min(iov_iter_single_seg_count(ii), max_size);
1246 static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1247 size_t *nbytesp, int write)
1249 size_t nbytes = 0; /* # bytes already packed in req */
1251 /* Special case for kernel I/O: can copy directly into the buffer */
1252 if (ii->type & ITER_KVEC) {
1253 unsigned long user_addr = fuse_get_user_addr(ii);
1254 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1257 req->in.args[1].value = (void *) user_addr;
1259 req->out.args[0].value = (void *) user_addr;
1261 iov_iter_advance(ii, frag_size);
1262 *nbytesp = frag_size;
1266 while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1269 ssize_t ret = iov_iter_get_pages(ii,
1270 &req->pages[req->num_pages],
1272 req->max_pages - req->num_pages,
1277 iov_iter_advance(ii, ret);
1281 npages = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
1283 req->page_descs[req->num_pages].offset = start;
1284 fuse_page_descs_length_init(req, req->num_pages, npages);
1286 req->num_pages += npages;
1287 req->page_descs[req->num_pages - 1].length -=
1288 (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1292 req->in.argpages = 1;
1294 req->out.argpages = 1;
1301 static inline int fuse_iter_npages(const struct iov_iter *ii_p)
1303 return iov_iter_npages(ii_p, FUSE_MAX_PAGES_PER_REQ);
1306 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1307 loff_t *ppos, int flags)
1309 int write = flags & FUSE_DIO_WRITE;
1310 int cuse = flags & FUSE_DIO_CUSE;
1311 struct file *file = io->file;
1312 struct inode *inode = file->f_mapping->host;
1313 struct fuse_file *ff = file->private_data;
1314 struct fuse_conn *fc = ff->fc;
1315 size_t nmax = write ? fc->max_write : fc->max_read;
1317 size_t count = iov_iter_count(iter);
1318 pgoff_t idx_from = pos >> PAGE_CACHE_SHIFT;
1319 pgoff_t idx_to = (pos + count - 1) >> PAGE_CACHE_SHIFT;
1321 struct fuse_req *req;
1324 req = fuse_get_req_for_background(fc, fuse_iter_npages(iter));
1326 req = fuse_get_req(fc, fuse_iter_npages(iter));
1328 return PTR_ERR(req);
1330 if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1332 mutex_lock(&inode->i_mutex);
1333 fuse_sync_writes(inode);
1335 mutex_unlock(&inode->i_mutex);
1340 fl_owner_t owner = current->files;
1341 size_t nbytes = min(count, nmax);
1342 int err = fuse_get_user_pages(req, iter, &nbytes, write);
1349 nres = fuse_send_write(req, io, pos, nbytes, owner);
1351 nres = fuse_send_read(req, io, pos, nbytes, owner);
1354 fuse_release_user_pages(req, !write);
1355 if (req->out.h.error) {
1357 res = req->out.h.error;
1359 } else if (nres > nbytes) {
1369 fuse_put_request(fc, req);
1371 req = fuse_get_req_for_background(fc,
1372 fuse_iter_npages(iter));
1374 req = fuse_get_req(fc, fuse_iter_npages(iter));
1380 fuse_put_request(fc, req);
1386 EXPORT_SYMBOL_GPL(fuse_direct_io);
1388 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1389 struct iov_iter *iter,
1393 struct file *file = io->file;
1394 struct inode *inode = file_inode(file);
1396 if (is_bad_inode(inode))
1399 res = fuse_direct_io(io, iter, ppos, 0);
1401 fuse_invalidate_attr(inode);
1406 static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1408 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb->ki_filp);
1409 return __fuse_direct_read(&io, to, &iocb->ki_pos);
1412 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1414 struct file *file = iocb->ki_filp;
1415 struct inode *inode = file_inode(file);
1416 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
1419 if (is_bad_inode(inode))
1422 /* Don't allow parallel writes to the same file */
1423 mutex_lock(&inode->i_mutex);
1424 res = generic_write_checks(iocb, from);
1426 res = fuse_direct_io(&io, from, &iocb->ki_pos, FUSE_DIO_WRITE);
1427 fuse_invalidate_attr(inode);
1429 fuse_write_update_size(inode, iocb->ki_pos);
1430 mutex_unlock(&inode->i_mutex);
1435 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1439 for (i = 0; i < req->num_pages; i++)
1440 __free_page(req->pages[i]);
1443 fuse_file_put(req->ff, false);
1446 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1448 struct inode *inode = req->inode;
1449 struct fuse_inode *fi = get_fuse_inode(inode);
1450 struct backing_dev_info *bdi = inode_to_bdi(inode);
1453 list_del(&req->writepages_entry);
1454 for (i = 0; i < req->num_pages; i++) {
1455 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1456 dec_zone_page_state(req->pages[i], NR_WRITEBACK_TEMP);
1457 wb_writeout_inc(&bdi->wb);
1459 wake_up(&fi->page_waitq);
1462 /* Called under fc->lock, may release and reacquire it */
1463 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req,
1465 __releases(fc->lock)
1466 __acquires(fc->lock)
1468 struct fuse_inode *fi = get_fuse_inode(req->inode);
1469 struct fuse_write_in *inarg = &req->misc.write.in;
1470 __u64 data_size = req->num_pages * PAGE_CACHE_SIZE;
1475 if (inarg->offset + data_size <= size) {
1476 inarg->size = data_size;
1477 } else if (inarg->offset < size) {
1478 inarg->size = size - inarg->offset;
1480 /* Got truncated off completely */
1484 req->in.args[1].size = inarg->size;
1486 fuse_request_send_background_locked(fc, req);
1490 fuse_writepage_finish(fc, req);
1491 spin_unlock(&fc->lock);
1492 fuse_writepage_free(fc, req);
1493 fuse_put_request(fc, req);
1494 spin_lock(&fc->lock);
1498 * If fi->writectr is positive (no truncate or fsync going on) send
1499 * all queued writepage requests.
1501 * Called with fc->lock
1503 void fuse_flush_writepages(struct inode *inode)
1504 __releases(fc->lock)
1505 __acquires(fc->lock)
1507 struct fuse_conn *fc = get_fuse_conn(inode);
1508 struct fuse_inode *fi = get_fuse_inode(inode);
1509 size_t crop = i_size_read(inode);
1510 struct fuse_req *req;
1512 while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1513 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1514 list_del_init(&req->list);
1515 fuse_send_writepage(fc, req, crop);
1519 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1521 struct inode *inode = req->inode;
1522 struct fuse_inode *fi = get_fuse_inode(inode);
1524 mapping_set_error(inode->i_mapping, req->out.h.error);
1525 spin_lock(&fc->lock);
1526 while (req->misc.write.next) {
1527 struct fuse_conn *fc = get_fuse_conn(inode);
1528 struct fuse_write_in *inarg = &req->misc.write.in;
1529 struct fuse_req *next = req->misc.write.next;
1530 req->misc.write.next = next->misc.write.next;
1531 next->misc.write.next = NULL;
1532 next->ff = fuse_file_get(req->ff);
1533 list_add(&next->writepages_entry, &fi->writepages);
1536 * Skip fuse_flush_writepages() to make it easy to crop requests
1537 * based on primary request size.
1539 * 1st case (trivial): there are no concurrent activities using
1540 * fuse_set/release_nowrite. Then we're on safe side because
1541 * fuse_flush_writepages() would call fuse_send_writepage()
1544 * 2nd case: someone called fuse_set_nowrite and it is waiting
1545 * now for completion of all in-flight requests. This happens
1546 * rarely and no more than once per page, so this should be
1549 * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1550 * of fuse_set_nowrite..fuse_release_nowrite section. The fact
1551 * that fuse_set_nowrite returned implies that all in-flight
1552 * requests were completed along with all of their secondary
1553 * requests. Further primary requests are blocked by negative
1554 * writectr. Hence there cannot be any in-flight requests and
1555 * no invocations of fuse_writepage_end() while we're in
1556 * fuse_set_nowrite..fuse_release_nowrite section.
1558 fuse_send_writepage(fc, next, inarg->offset + inarg->size);
1561 fuse_writepage_finish(fc, req);
1562 spin_unlock(&fc->lock);
1563 fuse_writepage_free(fc, req);
1566 static struct fuse_file *__fuse_write_file_get(struct fuse_conn *fc,
1567 struct fuse_inode *fi)
1569 struct fuse_file *ff = NULL;
1571 spin_lock(&fc->lock);
1572 if (!list_empty(&fi->write_files)) {
1573 ff = list_entry(fi->write_files.next, struct fuse_file,
1577 spin_unlock(&fc->lock);
1582 static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1583 struct fuse_inode *fi)
1585 struct fuse_file *ff = __fuse_write_file_get(fc, fi);
1590 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1592 struct fuse_conn *fc = get_fuse_conn(inode);
1593 struct fuse_inode *fi = get_fuse_inode(inode);
1594 struct fuse_file *ff;
1597 ff = __fuse_write_file_get(fc, fi);
1598 err = fuse_flush_times(inode, ff);
1600 fuse_file_put(ff, 0);
1605 static int fuse_writepage_locked(struct page *page)
1607 struct address_space *mapping = page->mapping;
1608 struct inode *inode = mapping->host;
1609 struct fuse_conn *fc = get_fuse_conn(inode);
1610 struct fuse_inode *fi = get_fuse_inode(inode);
1611 struct fuse_req *req;
1612 struct page *tmp_page;
1613 int error = -ENOMEM;
1615 set_page_writeback(page);
1617 req = fuse_request_alloc_nofs(1);
1621 /* writeback always goes to bg_queue */
1622 __set_bit(FR_BACKGROUND, &req->flags);
1623 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1628 req->ff = fuse_write_file_get(fc, fi);
1632 fuse_write_fill(req, req->ff, page_offset(page), 0);
1634 copy_highpage(tmp_page, page);
1635 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1636 req->misc.write.next = NULL;
1637 req->in.argpages = 1;
1639 req->pages[0] = tmp_page;
1640 req->page_descs[0].offset = 0;
1641 req->page_descs[0].length = PAGE_SIZE;
1642 req->end = fuse_writepage_end;
1645 inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1646 inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1648 spin_lock(&fc->lock);
1649 list_add(&req->writepages_entry, &fi->writepages);
1650 list_add_tail(&req->list, &fi->queued_writes);
1651 fuse_flush_writepages(inode);
1652 spin_unlock(&fc->lock);
1654 end_page_writeback(page);
1659 __free_page(tmp_page);
1661 fuse_request_free(req);
1663 end_page_writeback(page);
1667 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1671 if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1673 * ->writepages() should be called for sync() and friends. We
1674 * should only get here on direct reclaim and then we are
1675 * allowed to skip a page which is already in flight
1677 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1679 redirty_page_for_writepage(wbc, page);
1683 err = fuse_writepage_locked(page);
1689 struct fuse_fill_wb_data {
1690 struct fuse_req *req;
1691 struct fuse_file *ff;
1692 struct inode *inode;
1693 struct page **orig_pages;
1696 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1698 struct fuse_req *req = data->req;
1699 struct inode *inode = data->inode;
1700 struct fuse_conn *fc = get_fuse_conn(inode);
1701 struct fuse_inode *fi = get_fuse_inode(inode);
1702 int num_pages = req->num_pages;
1705 req->ff = fuse_file_get(data->ff);
1706 spin_lock(&fc->lock);
1707 list_add_tail(&req->list, &fi->queued_writes);
1708 fuse_flush_writepages(inode);
1709 spin_unlock(&fc->lock);
1711 for (i = 0; i < num_pages; i++)
1712 end_page_writeback(data->orig_pages[i]);
1715 static bool fuse_writepage_in_flight(struct fuse_req *new_req,
1718 struct fuse_conn *fc = get_fuse_conn(new_req->inode);
1719 struct fuse_inode *fi = get_fuse_inode(new_req->inode);
1720 struct fuse_req *tmp;
1721 struct fuse_req *old_req;
1725 BUG_ON(new_req->num_pages != 0);
1727 spin_lock(&fc->lock);
1728 list_del(&new_req->writepages_entry);
1729 list_for_each_entry(old_req, &fi->writepages, writepages_entry) {
1730 BUG_ON(old_req->inode != new_req->inode);
1731 curr_index = old_req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
1732 if (curr_index <= page->index &&
1733 page->index < curr_index + old_req->num_pages) {
1739 list_add(&new_req->writepages_entry, &fi->writepages);
1743 new_req->num_pages = 1;
1744 for (tmp = old_req; tmp != NULL; tmp = tmp->misc.write.next) {
1745 BUG_ON(tmp->inode != new_req->inode);
1746 curr_index = tmp->misc.write.in.offset >> PAGE_CACHE_SHIFT;
1747 if (tmp->num_pages == 1 &&
1748 curr_index == page->index) {
1753 if (old_req->num_pages == 1 && test_bit(FR_PENDING, &old_req->flags)) {
1754 struct backing_dev_info *bdi = inode_to_bdi(page->mapping->host);
1756 copy_highpage(old_req->pages[0], page);
1757 spin_unlock(&fc->lock);
1759 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1760 dec_zone_page_state(page, NR_WRITEBACK_TEMP);
1761 wb_writeout_inc(&bdi->wb);
1762 fuse_writepage_free(fc, new_req);
1763 fuse_request_free(new_req);
1766 new_req->misc.write.next = old_req->misc.write.next;
1767 old_req->misc.write.next = new_req;
1770 spin_unlock(&fc->lock);
1775 static int fuse_writepages_fill(struct page *page,
1776 struct writeback_control *wbc, void *_data)
1778 struct fuse_fill_wb_data *data = _data;
1779 struct fuse_req *req = data->req;
1780 struct inode *inode = data->inode;
1781 struct fuse_conn *fc = get_fuse_conn(inode);
1782 struct page *tmp_page;
1788 data->ff = fuse_write_file_get(fc, get_fuse_inode(inode));
1794 * Being under writeback is unlikely but possible. For example direct
1795 * read to an mmaped fuse file will set the page dirty twice; once when
1796 * the pages are faulted with get_user_pages(), and then after the read
1799 is_writeback = fuse_page_is_writeback(inode, page->index);
1801 if (req && req->num_pages &&
1802 (is_writeback || req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
1803 (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_write ||
1804 data->orig_pages[req->num_pages - 1]->index + 1 != page->index)) {
1805 fuse_writepages_send(data);
1809 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1814 * The page must not be redirtied until the writeout is completed
1815 * (i.e. userspace has sent a reply to the write request). Otherwise
1816 * there could be more than one temporary page instance for each real
1819 * This is ensured by holding the page lock in page_mkwrite() while
1820 * checking fuse_page_is_writeback(). We already hold the page lock
1821 * since clear_page_dirty_for_io() and keep it held until we add the
1822 * request to the fi->writepages list and increment req->num_pages.
1823 * After this fuse_page_is_writeback() will indicate that the page is
1824 * under writeback, so we can release the page lock.
1826 if (data->req == NULL) {
1827 struct fuse_inode *fi = get_fuse_inode(inode);
1830 req = fuse_request_alloc_nofs(FUSE_MAX_PAGES_PER_REQ);
1832 __free_page(tmp_page);
1836 fuse_write_fill(req, data->ff, page_offset(page), 0);
1837 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1838 req->misc.write.next = NULL;
1839 req->in.argpages = 1;
1840 __set_bit(FR_BACKGROUND, &req->flags);
1842 req->end = fuse_writepage_end;
1845 spin_lock(&fc->lock);
1846 list_add(&req->writepages_entry, &fi->writepages);
1847 spin_unlock(&fc->lock);
1851 set_page_writeback(page);
1853 copy_highpage(tmp_page, page);
1854 req->pages[req->num_pages] = tmp_page;
1855 req->page_descs[req->num_pages].offset = 0;
1856 req->page_descs[req->num_pages].length = PAGE_SIZE;
1858 inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1859 inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1862 if (is_writeback && fuse_writepage_in_flight(req, page)) {
1863 end_page_writeback(page);
1867 data->orig_pages[req->num_pages] = page;
1870 * Protected by fc->lock against concurrent access by
1871 * fuse_page_is_writeback().
1873 spin_lock(&fc->lock);
1875 spin_unlock(&fc->lock);
1883 static int fuse_writepages(struct address_space *mapping,
1884 struct writeback_control *wbc)
1886 struct inode *inode = mapping->host;
1887 struct fuse_fill_wb_data data;
1891 if (is_bad_inode(inode))
1899 data.orig_pages = kcalloc(FUSE_MAX_PAGES_PER_REQ,
1900 sizeof(struct page *),
1902 if (!data.orig_pages)
1905 err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
1907 /* Ignore errors if we can write at least one page */
1908 BUG_ON(!data.req->num_pages);
1909 fuse_writepages_send(&data);
1913 fuse_file_put(data.ff, false);
1915 kfree(data.orig_pages);
1921 * It's worthy to make sure that space is reserved on disk for the write,
1922 * but how to implement it without killing performance need more thinking.
1924 static int fuse_write_begin(struct file *file, struct address_space *mapping,
1925 loff_t pos, unsigned len, unsigned flags,
1926 struct page **pagep, void **fsdata)
1928 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1929 struct fuse_conn *fc = get_fuse_conn(file_inode(file));
1934 WARN_ON(!fc->writeback_cache);
1936 page = grab_cache_page_write_begin(mapping, index, flags);
1940 fuse_wait_on_page_writeback(mapping->host, page->index);
1942 if (PageUptodate(page) || len == PAGE_CACHE_SIZE)
1945 * Check if the start this page comes after the end of file, in which
1946 * case the readpage can be optimized away.
1948 fsize = i_size_read(mapping->host);
1949 if (fsize <= (pos & PAGE_CACHE_MASK)) {
1950 size_t off = pos & ~PAGE_CACHE_MASK;
1952 zero_user_segment(page, 0, off);
1955 err = fuse_do_readpage(file, page);
1964 page_cache_release(page);
1969 static int fuse_write_end(struct file *file, struct address_space *mapping,
1970 loff_t pos, unsigned len, unsigned copied,
1971 struct page *page, void *fsdata)
1973 struct inode *inode = page->mapping->host;
1975 if (!PageUptodate(page)) {
1976 /* Zero any unwritten bytes at the end of the page */
1977 size_t endoff = (pos + copied) & ~PAGE_CACHE_MASK;
1979 zero_user_segment(page, endoff, PAGE_CACHE_SIZE);
1980 SetPageUptodate(page);
1983 fuse_write_update_size(inode, pos + copied);
1984 set_page_dirty(page);
1986 page_cache_release(page);
1991 static int fuse_launder_page(struct page *page)
1994 if (clear_page_dirty_for_io(page)) {
1995 struct inode *inode = page->mapping->host;
1996 err = fuse_writepage_locked(page);
1998 fuse_wait_on_page_writeback(inode, page->index);
2004 * Write back dirty pages now, because there may not be any suitable
2007 static void fuse_vma_close(struct vm_area_struct *vma)
2009 filemap_write_and_wait(vma->vm_file->f_mapping);
2013 * Wait for writeback against this page to complete before allowing it
2014 * to be marked dirty again, and hence written back again, possibly
2015 * before the previous writepage completed.
2017 * Block here, instead of in ->writepage(), so that the userspace fs
2018 * can only block processes actually operating on the filesystem.
2020 * Otherwise unprivileged userspace fs would be able to block
2025 * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2027 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
2029 struct page *page = vmf->page;
2030 struct inode *inode = file_inode(vma->vm_file);
2032 file_update_time(vma->vm_file);
2034 if (page->mapping != inode->i_mapping) {
2036 return VM_FAULT_NOPAGE;
2039 fuse_wait_on_page_writeback(inode, page->index);
2040 return VM_FAULT_LOCKED;
2043 static const struct vm_operations_struct fuse_file_vm_ops = {
2044 .close = fuse_vma_close,
2045 .fault = filemap_fault,
2046 .map_pages = filemap_map_pages,
2047 .page_mkwrite = fuse_page_mkwrite,
2050 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2052 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2053 fuse_link_write_file(file);
2055 file_accessed(file);
2056 vma->vm_ops = &fuse_file_vm_ops;
2060 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
2062 /* Can't provide the coherency needed for MAP_SHARED */
2063 if (vma->vm_flags & VM_MAYSHARE)
2066 invalidate_inode_pages2(file->f_mapping);
2068 return generic_file_mmap(file, vma);
2071 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
2072 struct file_lock *fl)
2074 switch (ffl->type) {
2080 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2081 ffl->end < ffl->start)
2084 fl->fl_start = ffl->start;
2085 fl->fl_end = ffl->end;
2086 fl->fl_pid = ffl->pid;
2092 fl->fl_type = ffl->type;
2096 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2097 const struct file_lock *fl, int opcode, pid_t pid,
2098 int flock, struct fuse_lk_in *inarg)
2100 struct inode *inode = file_inode(file);
2101 struct fuse_conn *fc = get_fuse_conn(inode);
2102 struct fuse_file *ff = file->private_data;
2104 memset(inarg, 0, sizeof(*inarg));
2106 inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2107 inarg->lk.start = fl->fl_start;
2108 inarg->lk.end = fl->fl_end;
2109 inarg->lk.type = fl->fl_type;
2110 inarg->lk.pid = pid;
2112 inarg->lk_flags |= FUSE_LK_FLOCK;
2113 args->in.h.opcode = opcode;
2114 args->in.h.nodeid = get_node_id(inode);
2115 args->in.numargs = 1;
2116 args->in.args[0].size = sizeof(*inarg);
2117 args->in.args[0].value = inarg;
2120 static int fuse_getlk(struct file *file, struct file_lock *fl)
2122 struct inode *inode = file_inode(file);
2123 struct fuse_conn *fc = get_fuse_conn(inode);
2125 struct fuse_lk_in inarg;
2126 struct fuse_lk_out outarg;
2129 fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2130 args.out.numargs = 1;
2131 args.out.args[0].size = sizeof(outarg);
2132 args.out.args[0].value = &outarg;
2133 err = fuse_simple_request(fc, &args);
2135 err = convert_fuse_file_lock(&outarg.lk, fl);
2140 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2142 struct inode *inode = file_inode(file);
2143 struct fuse_conn *fc = get_fuse_conn(inode);
2145 struct fuse_lk_in inarg;
2146 int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2147 pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
2150 if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2151 /* NLM needs asynchronous locks, which we don't support yet */
2155 /* Unlock on close is handled by the flush method */
2156 if (fl->fl_flags & FL_CLOSE)
2159 fuse_lk_fill(&args, file, fl, opcode, pid, flock, &inarg);
2160 err = fuse_simple_request(fc, &args);
2162 /* locking is restartable */
2169 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2171 struct inode *inode = file_inode(file);
2172 struct fuse_conn *fc = get_fuse_conn(inode);
2175 if (cmd == F_CANCELLK) {
2177 } else if (cmd == F_GETLK) {
2179 posix_test_lock(file, fl);
2182 err = fuse_getlk(file, fl);
2185 err = posix_lock_file(file, fl, NULL);
2187 err = fuse_setlk(file, fl, 0);
2192 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2194 struct inode *inode = file_inode(file);
2195 struct fuse_conn *fc = get_fuse_conn(inode);
2199 err = locks_lock_file_wait(file, fl);
2201 struct fuse_file *ff = file->private_data;
2203 /* emulate flock with POSIX locks */
2205 err = fuse_setlk(file, fl, 1);
2211 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2213 struct inode *inode = mapping->host;
2214 struct fuse_conn *fc = get_fuse_conn(inode);
2216 struct fuse_bmap_in inarg;
2217 struct fuse_bmap_out outarg;
2220 if (!inode->i_sb->s_bdev || fc->no_bmap)
2223 memset(&inarg, 0, sizeof(inarg));
2224 inarg.block = block;
2225 inarg.blocksize = inode->i_sb->s_blocksize;
2226 args.in.h.opcode = FUSE_BMAP;
2227 args.in.h.nodeid = get_node_id(inode);
2228 args.in.numargs = 1;
2229 args.in.args[0].size = sizeof(inarg);
2230 args.in.args[0].value = &inarg;
2231 args.out.numargs = 1;
2232 args.out.args[0].size = sizeof(outarg);
2233 args.out.args[0].value = &outarg;
2234 err = fuse_simple_request(fc, &args);
2238 return err ? 0 : outarg.block;
2241 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2244 struct inode *inode = file_inode(file);
2246 /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2247 if (whence == SEEK_CUR || whence == SEEK_SET)
2248 return generic_file_llseek(file, offset, whence);
2250 mutex_lock(&inode->i_mutex);
2251 retval = fuse_update_attributes(inode, NULL, file, NULL);
2253 retval = generic_file_llseek(file, offset, whence);
2254 mutex_unlock(&inode->i_mutex);
2259 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
2260 unsigned int nr_segs, size_t bytes, bool to_user)
2268 iov_iter_init(&ii, to_user ? READ : WRITE, iov, nr_segs, bytes);
2270 while (iov_iter_count(&ii)) {
2271 struct page *page = pages[page_idx++];
2272 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
2278 char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
2279 size_t iov_len = ii.iov->iov_len - ii.iov_offset;
2280 size_t copy = min(todo, iov_len);
2284 left = copy_from_user(kaddr, uaddr, copy);
2286 left = copy_to_user(uaddr, kaddr, copy);
2291 iov_iter_advance(&ii, copy);
2303 * CUSE servers compiled on 32bit broke on 64bit kernels because the
2304 * ABI was defined to be 'struct iovec' which is different on 32bit
2305 * and 64bit. Fortunately we can determine which structure the server
2306 * used from the size of the reply.
2308 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2309 size_t transferred, unsigned count,
2312 #ifdef CONFIG_COMPAT
2313 if (count * sizeof(struct compat_iovec) == transferred) {
2314 struct compat_iovec *ciov = src;
2318 * With this interface a 32bit server cannot support
2319 * non-compat (i.e. ones coming from 64bit apps) ioctl
2325 for (i = 0; i < count; i++) {
2326 dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2327 dst[i].iov_len = ciov[i].iov_len;
2333 if (count * sizeof(struct iovec) != transferred)
2336 memcpy(dst, src, transferred);
2340 /* Make sure iov_length() won't overflow */
2341 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
2344 u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
2346 for (n = 0; n < count; n++, iov++) {
2347 if (iov->iov_len > (size_t) max)
2349 max -= iov->iov_len;
2354 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2355 void *src, size_t transferred, unsigned count,
2359 struct fuse_ioctl_iovec *fiov = src;
2361 if (fc->minor < 16) {
2362 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2366 if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2369 for (i = 0; i < count; i++) {
2370 /* Did the server supply an inappropriate value? */
2371 if (fiov[i].base != (unsigned long) fiov[i].base ||
2372 fiov[i].len != (unsigned long) fiov[i].len)
2375 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2376 dst[i].iov_len = (size_t) fiov[i].len;
2378 #ifdef CONFIG_COMPAT
2380 (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2381 (compat_size_t) dst[i].iov_len != fiov[i].len))
2391 * For ioctls, there is no generic way to determine how much memory
2392 * needs to be read and/or written. Furthermore, ioctls are allowed
2393 * to dereference the passed pointer, so the parameter requires deep
2394 * copying but FUSE has no idea whatsoever about what to copy in or
2397 * This is solved by allowing FUSE server to retry ioctl with
2398 * necessary in/out iovecs. Let's assume the ioctl implementation
2399 * needs to read in the following structure.
2406 * On the first callout to FUSE server, inarg->in_size and
2407 * inarg->out_size will be NULL; then, the server completes the ioctl
2408 * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2409 * the actual iov array to
2411 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) } }
2413 * which tells FUSE to copy in the requested area and retry the ioctl.
2414 * On the second round, the server has access to the structure and
2415 * from that it can tell what to look for next, so on the invocation,
2416 * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2418 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) },
2419 * { .iov_base = a.buf, .iov_len = a.buflen } }
2421 * FUSE will copy both struct a and the pointed buffer from the
2422 * process doing the ioctl and retry ioctl with both struct a and the
2425 * This time, FUSE server has everything it needs and completes ioctl
2426 * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2428 * Copying data out works the same way.
2430 * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2431 * automatically initializes in and out iovs by decoding @cmd with
2432 * _IOC_* macros and the server is not allowed to request RETRY. This
2433 * limits ioctl data transfers to well-formed ioctls and is the forced
2434 * behavior for all FUSE servers.
2436 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2439 struct fuse_file *ff = file->private_data;
2440 struct fuse_conn *fc = ff->fc;
2441 struct fuse_ioctl_in inarg = {
2447 struct fuse_ioctl_out outarg;
2448 struct fuse_req *req = NULL;
2449 struct page **pages = NULL;
2450 struct iovec *iov_page = NULL;
2451 struct iovec *in_iov = NULL, *out_iov = NULL;
2452 unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
2453 size_t in_size, out_size, transferred;
2456 #if BITS_PER_LONG == 32
2457 inarg.flags |= FUSE_IOCTL_32BIT;
2459 if (flags & FUSE_IOCTL_COMPAT)
2460 inarg.flags |= FUSE_IOCTL_32BIT;
2463 /* assume all the iovs returned by client always fits in a page */
2464 BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2467 pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
2468 iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2469 if (!pages || !iov_page)
2473 * If restricted, initialize IO parameters as encoded in @cmd.
2474 * RETRY from server is not allowed.
2476 if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2477 struct iovec *iov = iov_page;
2479 iov->iov_base = (void __user *)arg;
2480 iov->iov_len = _IOC_SIZE(cmd);
2482 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2487 if (_IOC_DIR(cmd) & _IOC_READ) {
2494 inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2495 inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2498 * Out data can be used either for actual out data or iovs,
2499 * make sure there always is at least one page.
2501 out_size = max_t(size_t, out_size, PAGE_SIZE);
2502 max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2504 /* make sure there are enough buffer pages and init request with them */
2506 if (max_pages > FUSE_MAX_PAGES_PER_REQ)
2508 while (num_pages < max_pages) {
2509 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2510 if (!pages[num_pages])
2515 req = fuse_get_req(fc, num_pages);
2521 memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
2522 req->num_pages = num_pages;
2523 fuse_page_descs_length_init(req, 0, req->num_pages);
2525 /* okay, let's send it to the client */
2526 req->in.h.opcode = FUSE_IOCTL;
2527 req->in.h.nodeid = ff->nodeid;
2528 req->in.numargs = 1;
2529 req->in.args[0].size = sizeof(inarg);
2530 req->in.args[0].value = &inarg;
2533 req->in.args[1].size = in_size;
2534 req->in.argpages = 1;
2536 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
2542 req->out.numargs = 2;
2543 req->out.args[0].size = sizeof(outarg);
2544 req->out.args[0].value = &outarg;
2545 req->out.args[1].size = out_size;
2546 req->out.argpages = 1;
2547 req->out.argvar = 1;
2549 fuse_request_send(fc, req);
2550 err = req->out.h.error;
2551 transferred = req->out.args[1].size;
2552 fuse_put_request(fc, req);
2557 /* did it ask for retry? */
2558 if (outarg.flags & FUSE_IOCTL_RETRY) {
2561 /* no retry if in restricted mode */
2563 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2566 in_iovs = outarg.in_iovs;
2567 out_iovs = outarg.out_iovs;
2570 * Make sure things are in boundary, separate checks
2571 * are to protect against overflow.
2574 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2575 out_iovs > FUSE_IOCTL_MAX_IOV ||
2576 in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2579 vaddr = kmap_atomic(pages[0]);
2580 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2581 transferred, in_iovs + out_iovs,
2582 (flags & FUSE_IOCTL_COMPAT) != 0);
2583 kunmap_atomic(vaddr);
2588 out_iov = in_iov + in_iovs;
2590 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
2594 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
2602 if (transferred > inarg.out_size)
2605 err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
2608 fuse_put_request(fc, req);
2609 free_page((unsigned long) iov_page);
2611 __free_page(pages[--num_pages]);
2614 return err ? err : outarg.result;
2616 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2618 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2619 unsigned long arg, unsigned int flags)
2621 struct inode *inode = file_inode(file);
2622 struct fuse_conn *fc = get_fuse_conn(inode);
2624 if (!fuse_allow_current_process(fc))
2627 if (is_bad_inode(inode))
2630 return fuse_do_ioctl(file, cmd, arg, flags);
2633 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2636 return fuse_ioctl_common(file, cmd, arg, 0);
2639 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2642 return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2646 * All files which have been polled are linked to RB tree
2647 * fuse_conn->polled_files which is indexed by kh. Walk the tree and
2648 * find the matching one.
2650 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2651 struct rb_node **parent_out)
2653 struct rb_node **link = &fc->polled_files.rb_node;
2654 struct rb_node *last = NULL;
2657 struct fuse_file *ff;
2660 ff = rb_entry(last, struct fuse_file, polled_node);
2663 link = &last->rb_left;
2664 else if (kh > ff->kh)
2665 link = &last->rb_right;
2676 * The file is about to be polled. Make sure it's on the polled_files
2677 * RB tree. Note that files once added to the polled_files tree are
2678 * not removed before the file is released. This is because a file
2679 * polled once is likely to be polled again.
2681 static void fuse_register_polled_file(struct fuse_conn *fc,
2682 struct fuse_file *ff)
2684 spin_lock(&fc->lock);
2685 if (RB_EMPTY_NODE(&ff->polled_node)) {
2686 struct rb_node **link, *uninitialized_var(parent);
2688 link = fuse_find_polled_node(fc, ff->kh, &parent);
2690 rb_link_node(&ff->polled_node, parent, link);
2691 rb_insert_color(&ff->polled_node, &fc->polled_files);
2693 spin_unlock(&fc->lock);
2696 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2698 struct fuse_file *ff = file->private_data;
2699 struct fuse_conn *fc = ff->fc;
2700 struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2701 struct fuse_poll_out outarg;
2706 return DEFAULT_POLLMASK;
2708 poll_wait(file, &ff->poll_wait, wait);
2709 inarg.events = (__u32)poll_requested_events(wait);
2712 * Ask for notification iff there's someone waiting for it.
2713 * The client may ignore the flag and always notify.
2715 if (waitqueue_active(&ff->poll_wait)) {
2716 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2717 fuse_register_polled_file(fc, ff);
2720 args.in.h.opcode = FUSE_POLL;
2721 args.in.h.nodeid = ff->nodeid;
2722 args.in.numargs = 1;
2723 args.in.args[0].size = sizeof(inarg);
2724 args.in.args[0].value = &inarg;
2725 args.out.numargs = 1;
2726 args.out.args[0].size = sizeof(outarg);
2727 args.out.args[0].value = &outarg;
2728 err = fuse_simple_request(fc, &args);
2731 return outarg.revents;
2732 if (err == -ENOSYS) {
2734 return DEFAULT_POLLMASK;
2738 EXPORT_SYMBOL_GPL(fuse_file_poll);
2741 * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2742 * wakes up the poll waiters.
2744 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2745 struct fuse_notify_poll_wakeup_out *outarg)
2747 u64 kh = outarg->kh;
2748 struct rb_node **link;
2750 spin_lock(&fc->lock);
2752 link = fuse_find_polled_node(fc, kh, NULL);
2754 struct fuse_file *ff;
2756 ff = rb_entry(*link, struct fuse_file, polled_node);
2757 wake_up_interruptible_sync(&ff->poll_wait);
2760 spin_unlock(&fc->lock);
2764 static void fuse_do_truncate(struct file *file)
2766 struct inode *inode = file->f_mapping->host;
2769 attr.ia_valid = ATTR_SIZE;
2770 attr.ia_size = i_size_read(inode);
2772 attr.ia_file = file;
2773 attr.ia_valid |= ATTR_FILE;
2775 fuse_do_setattr(inode, &attr, file);
2778 static inline loff_t fuse_round_up(loff_t off)
2780 return round_up(off, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
2784 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter, loff_t offset)
2786 DECLARE_COMPLETION_ONSTACK(wait);
2788 struct file *file = iocb->ki_filp;
2789 struct fuse_file *ff = file->private_data;
2790 bool async_dio = ff->fc->async_dio;
2792 struct inode *inode;
2794 size_t count = iov_iter_count(iter);
2795 struct fuse_io_priv *io;
2796 bool is_sync = is_sync_kiocb(iocb);
2799 inode = file->f_mapping->host;
2800 i_size = i_size_read(inode);
2802 if ((iov_iter_rw(iter) == READ) && (offset > i_size))
2805 /* optimization for short read */
2806 if (async_dio && iov_iter_rw(iter) != WRITE && offset + count > i_size) {
2807 if (offset >= i_size)
2809 iov_iter_truncate(iter, fuse_round_up(i_size - offset));
2810 count = iov_iter_count(iter);
2813 io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2816 spin_lock_init(&io->lock);
2817 kref_init(&io->refcnt);
2821 io->offset = offset;
2822 io->write = (iov_iter_rw(iter) == WRITE);
2826 * By default, we want to optimize all I/Os with async request
2827 * submission to the client filesystem if supported.
2829 io->async = async_dio;
2833 * We cannot asynchronously extend the size of a file. We have no method
2834 * to wait on real async I/O requests, so we must submit this request
2837 if (!is_sync && (offset + count > i_size) &&
2838 iov_iter_rw(iter) == WRITE)
2841 if (io->async && is_sync) {
2843 * Additional reference to keep io around after
2844 * calling fuse_aio_complete()
2846 kref_get(&io->refcnt);
2850 if (iov_iter_rw(iter) == WRITE) {
2851 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
2852 fuse_invalidate_attr(inode);
2854 ret = __fuse_direct_read(io, iter, &pos);
2858 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2860 /* we have a non-extending, async request, so return */
2862 return -EIOCBQUEUED;
2864 wait_for_completion(&wait);
2865 ret = fuse_get_res_by_io(io);
2868 kref_put(&io->refcnt, fuse_io_release);
2870 if (iov_iter_rw(iter) == WRITE) {
2872 fuse_write_update_size(inode, pos);
2873 else if (ret < 0 && offset + count > i_size)
2874 fuse_do_truncate(file);
2880 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2883 struct fuse_file *ff = file->private_data;
2884 struct inode *inode = file_inode(file);
2885 struct fuse_inode *fi = get_fuse_inode(inode);
2886 struct fuse_conn *fc = ff->fc;
2888 struct fuse_fallocate_in inarg = {
2895 bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
2896 (mode & FALLOC_FL_PUNCH_HOLE);
2898 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2901 if (fc->no_fallocate)
2905 mutex_lock(&inode->i_mutex);
2906 if (mode & FALLOC_FL_PUNCH_HOLE) {
2907 loff_t endbyte = offset + length - 1;
2908 err = filemap_write_and_wait_range(inode->i_mapping,
2913 fuse_sync_writes(inode);
2917 if (!(mode & FALLOC_FL_KEEP_SIZE))
2918 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2920 args.in.h.opcode = FUSE_FALLOCATE;
2921 args.in.h.nodeid = ff->nodeid;
2922 args.in.numargs = 1;
2923 args.in.args[0].size = sizeof(inarg);
2924 args.in.args[0].value = &inarg;
2925 err = fuse_simple_request(fc, &args);
2926 if (err == -ENOSYS) {
2927 fc->no_fallocate = 1;
2933 /* we could have extended the file */
2934 if (!(mode & FALLOC_FL_KEEP_SIZE)) {
2935 bool changed = fuse_write_update_size(inode, offset + length);
2937 if (changed && fc->writeback_cache)
2938 file_update_time(file);
2941 if (mode & FALLOC_FL_PUNCH_HOLE)
2942 truncate_pagecache_range(inode, offset, offset + length - 1);
2944 fuse_invalidate_attr(inode);
2947 if (!(mode & FALLOC_FL_KEEP_SIZE))
2948 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2951 mutex_unlock(&inode->i_mutex);
2956 static const struct file_operations fuse_file_operations = {
2957 .llseek = fuse_file_llseek,
2958 .read_iter = fuse_file_read_iter,
2959 .write_iter = fuse_file_write_iter,
2960 .mmap = fuse_file_mmap,
2962 .flush = fuse_flush,
2963 .release = fuse_release,
2964 .fsync = fuse_fsync,
2965 .lock = fuse_file_lock,
2966 .flock = fuse_file_flock,
2967 .splice_read = generic_file_splice_read,
2968 .unlocked_ioctl = fuse_file_ioctl,
2969 .compat_ioctl = fuse_file_compat_ioctl,
2970 .poll = fuse_file_poll,
2971 .fallocate = fuse_file_fallocate,
2974 static const struct file_operations fuse_direct_io_file_operations = {
2975 .llseek = fuse_file_llseek,
2976 .read_iter = fuse_direct_read_iter,
2977 .write_iter = fuse_direct_write_iter,
2978 .mmap = fuse_direct_mmap,
2980 .flush = fuse_flush,
2981 .release = fuse_release,
2982 .fsync = fuse_fsync,
2983 .lock = fuse_file_lock,
2984 .flock = fuse_file_flock,
2985 .unlocked_ioctl = fuse_file_ioctl,
2986 .compat_ioctl = fuse_file_compat_ioctl,
2987 .poll = fuse_file_poll,
2988 .fallocate = fuse_file_fallocate,
2989 /* no splice_read */
2992 static const struct address_space_operations fuse_file_aops = {
2993 .readpage = fuse_readpage,
2994 .writepage = fuse_writepage,
2995 .writepages = fuse_writepages,
2996 .launder_page = fuse_launder_page,
2997 .readpages = fuse_readpages,
2998 .set_page_dirty = __set_page_dirty_nobuffers,
3000 .direct_IO = fuse_direct_IO,
3001 .write_begin = fuse_write_begin,
3002 .write_end = fuse_write_end,
3005 void fuse_init_file_inode(struct inode *inode)
3007 inode->i_fop = &fuse_file_operations;
3008 inode->i_data.a_ops = &fuse_file_aops;