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);
212 err = fuse_do_open(fc, get_node_id(inode), file, isdir);
215 fuse_finish_open(inode, file);
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_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);
417 fuse_sync_writes(inode);
420 err = filemap_check_errors(file->f_mapping);
424 req = fuse_get_req_nofail_nopages(fc, file);
425 memset(&inarg, 0, sizeof(inarg));
427 inarg.lock_owner = fuse_lock_owner_id(fc, id);
428 req->in.h.opcode = FUSE_FLUSH;
429 req->in.h.nodeid = get_node_id(inode);
431 req->in.args[0].size = sizeof(inarg);
432 req->in.args[0].value = &inarg;
433 __set_bit(FR_FORCE, &req->flags);
434 fuse_request_send(fc, req);
435 err = req->out.h.error;
436 fuse_put_request(fc, req);
437 if (err == -ENOSYS) {
444 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
445 int datasync, int isdir)
447 struct inode *inode = file->f_mapping->host;
448 struct fuse_conn *fc = get_fuse_conn(inode);
449 struct fuse_file *ff = file->private_data;
451 struct fuse_fsync_in inarg;
454 if (is_bad_inode(inode))
460 * Start writeback against all dirty pages of the inode, then
461 * wait for all outstanding writes, before sending the FSYNC
464 err = filemap_write_and_wait_range(inode->i_mapping, start, end);
468 fuse_sync_writes(inode);
471 * Due to implementation of fuse writeback
472 * filemap_write_and_wait_range() does not catch errors.
473 * We have to do this directly after fuse_sync_writes()
475 err = filemap_check_errors(file->f_mapping);
479 err = sync_inode_metadata(inode, 1);
483 if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
486 memset(&inarg, 0, sizeof(inarg));
488 inarg.fsync_flags = datasync ? 1 : 0;
489 args.in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
490 args.in.h.nodeid = get_node_id(inode);
492 args.in.args[0].size = sizeof(inarg);
493 args.in.args[0].value = &inarg;
494 err = fuse_simple_request(fc, &args);
495 if (err == -ENOSYS) {
507 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
510 return fuse_fsync_common(file, start, end, datasync, 0);
513 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
514 size_t count, int opcode)
516 struct fuse_read_in *inarg = &req->misc.read.in;
517 struct fuse_file *ff = file->private_data;
522 inarg->flags = file->f_flags;
523 req->in.h.opcode = opcode;
524 req->in.h.nodeid = ff->nodeid;
526 req->in.args[0].size = sizeof(struct fuse_read_in);
527 req->in.args[0].value = inarg;
529 req->out.numargs = 1;
530 req->out.args[0].size = count;
533 static void fuse_release_user_pages(struct fuse_req *req, bool should_dirty)
537 for (i = 0; i < req->num_pages; i++) {
538 struct page *page = req->pages[i];
540 set_page_dirty_lock(page);
545 static void fuse_io_release(struct kref *kref)
547 kfree(container_of(kref, struct fuse_io_priv, refcnt));
550 static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
555 if (io->bytes >= 0 && io->write)
558 return io->bytes < 0 ? io->size : io->bytes;
562 * In case of short read, the caller sets 'pos' to the position of
563 * actual end of fuse request in IO request. Otherwise, if bytes_requested
564 * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
567 * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
568 * both submitted asynchronously. The first of them was ACKed by userspace as
569 * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
570 * second request was ACKed as short, e.g. only 1K was read, resulting in
573 * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
574 * will be equal to the length of the longest contiguous fragment of
575 * transferred data starting from the beginning of IO request.
577 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
581 spin_lock(&io->lock);
583 io->err = io->err ? : err;
584 else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
588 if (!left && io->blocking)
590 spin_unlock(&io->lock);
592 if (!left && !io->blocking) {
593 ssize_t res = fuse_get_res_by_io(io);
596 struct inode *inode = file_inode(io->iocb->ki_filp);
597 struct fuse_conn *fc = get_fuse_conn(inode);
598 struct fuse_inode *fi = get_fuse_inode(inode);
600 spin_lock(&fc->lock);
601 fi->attr_version = ++fc->attr_version;
602 spin_unlock(&fc->lock);
605 io->iocb->ki_complete(io->iocb, res, 0);
608 kref_put(&io->refcnt, fuse_io_release);
611 static void fuse_aio_complete_req(struct fuse_conn *fc, struct fuse_req *req)
613 struct fuse_io_priv *io = req->io;
616 fuse_release_user_pages(req, !io->write);
619 if (req->misc.write.in.size != req->misc.write.out.size)
620 pos = req->misc.write.in.offset - io->offset +
621 req->misc.write.out.size;
623 if (req->misc.read.in.size != req->out.args[0].size)
624 pos = req->misc.read.in.offset - io->offset +
625 req->out.args[0].size;
628 fuse_aio_complete(io, req->out.h.error, pos);
631 static size_t fuse_async_req_send(struct fuse_conn *fc, struct fuse_req *req,
632 size_t num_bytes, struct fuse_io_priv *io)
634 spin_lock(&io->lock);
635 kref_get(&io->refcnt);
636 io->size += num_bytes;
638 spin_unlock(&io->lock);
641 req->end = fuse_aio_complete_req;
643 __fuse_get_request(req);
644 fuse_request_send_background(fc, req);
649 static size_t fuse_send_read(struct fuse_req *req, struct fuse_io_priv *io,
650 loff_t pos, size_t count, fl_owner_t owner)
652 struct file *file = io->file;
653 struct fuse_file *ff = file->private_data;
654 struct fuse_conn *fc = ff->fc;
656 fuse_read_fill(req, file, pos, count, FUSE_READ);
658 struct fuse_read_in *inarg = &req->misc.read.in;
660 inarg->read_flags |= FUSE_READ_LOCKOWNER;
661 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
665 return fuse_async_req_send(fc, req, count, io);
667 fuse_request_send(fc, req);
668 return req->out.args[0].size;
671 static void fuse_read_update_size(struct inode *inode, loff_t size,
674 struct fuse_conn *fc = get_fuse_conn(inode);
675 struct fuse_inode *fi = get_fuse_inode(inode);
677 spin_lock(&fc->lock);
678 if (attr_ver == fi->attr_version && size < inode->i_size &&
679 !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
680 fi->attr_version = ++fc->attr_version;
681 i_size_write(inode, size);
683 spin_unlock(&fc->lock);
686 static void fuse_short_read(struct fuse_req *req, struct inode *inode,
689 size_t num_read = req->out.args[0].size;
690 struct fuse_conn *fc = get_fuse_conn(inode);
692 if (fc->writeback_cache) {
694 * A hole in a file. Some data after the hole are in page cache,
695 * but have not reached the client fs yet. So, the hole is not
699 int start_idx = num_read >> PAGE_SHIFT;
700 size_t off = num_read & (PAGE_SIZE - 1);
702 for (i = start_idx; i < req->num_pages; i++) {
703 zero_user_segment(req->pages[i], off, PAGE_SIZE);
707 loff_t pos = page_offset(req->pages[0]) + num_read;
708 fuse_read_update_size(inode, pos, attr_ver);
712 static int fuse_do_readpage(struct file *file, struct page *page)
714 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
715 struct inode *inode = page->mapping->host;
716 struct fuse_conn *fc = get_fuse_conn(inode);
717 struct fuse_req *req;
719 loff_t pos = page_offset(page);
720 size_t count = PAGE_SIZE;
725 * Page writeback can extend beyond the lifetime of the
726 * page-cache page, so make sure we read a properly synced
729 fuse_wait_on_page_writeback(inode, page->index);
731 req = fuse_get_req(fc, 1);
735 attr_ver = fuse_get_attr_version(fc);
737 req->out.page_zeroing = 1;
738 req->out.argpages = 1;
740 req->pages[0] = page;
741 req->page_descs[0].length = count;
742 num_read = fuse_send_read(req, &io, pos, count, NULL);
743 err = req->out.h.error;
747 * Short read means EOF. If file size is larger, truncate it
749 if (num_read < count)
750 fuse_short_read(req, inode, attr_ver);
752 SetPageUptodate(page);
755 fuse_put_request(fc, req);
760 static int fuse_readpage(struct file *file, struct page *page)
762 struct inode *inode = page->mapping->host;
766 if (is_bad_inode(inode))
769 err = fuse_do_readpage(file, page);
770 fuse_invalidate_atime(inode);
776 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
779 size_t count = req->misc.read.in.size;
780 size_t num_read = req->out.args[0].size;
781 struct address_space *mapping = NULL;
783 for (i = 0; mapping == NULL && i < req->num_pages; i++)
784 mapping = req->pages[i]->mapping;
787 struct inode *inode = mapping->host;
790 * Short read means EOF. If file size is larger, truncate it
792 if (!req->out.h.error && num_read < count)
793 fuse_short_read(req, inode, req->misc.read.attr_ver);
795 fuse_invalidate_atime(inode);
798 for (i = 0; i < req->num_pages; i++) {
799 struct page *page = req->pages[i];
800 if (!req->out.h.error)
801 SetPageUptodate(page);
808 fuse_file_put(req->ff, false);
811 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
813 struct fuse_file *ff = file->private_data;
814 struct fuse_conn *fc = ff->fc;
815 loff_t pos = page_offset(req->pages[0]);
816 size_t count = req->num_pages << PAGE_SHIFT;
818 req->out.argpages = 1;
819 req->out.page_zeroing = 1;
820 req->out.page_replace = 1;
821 fuse_read_fill(req, file, pos, count, FUSE_READ);
822 req->misc.read.attr_ver = fuse_get_attr_version(fc);
823 if (fc->async_read) {
824 req->ff = fuse_file_get(ff);
825 req->end = fuse_readpages_end;
826 fuse_request_send_background(fc, req);
828 fuse_request_send(fc, req);
829 fuse_readpages_end(fc, req);
830 fuse_put_request(fc, req);
834 struct fuse_fill_data {
835 struct fuse_req *req;
841 static int fuse_readpages_fill(void *_data, struct page *page)
843 struct fuse_fill_data *data = _data;
844 struct fuse_req *req = data->req;
845 struct inode *inode = data->inode;
846 struct fuse_conn *fc = get_fuse_conn(inode);
848 fuse_wait_on_page_writeback(inode, page->index);
850 if (req->num_pages &&
851 (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
852 (req->num_pages + 1) * PAGE_SIZE > fc->max_read ||
853 req->pages[req->num_pages - 1]->index + 1 != page->index)) {
854 int nr_alloc = min_t(unsigned, data->nr_pages,
855 FUSE_MAX_PAGES_PER_REQ);
856 fuse_send_readpages(req, data->file);
858 req = fuse_get_req_for_background(fc, nr_alloc);
860 req = fuse_get_req(fc, nr_alloc);
869 if (WARN_ON(req->num_pages >= req->max_pages)) {
870 fuse_put_request(fc, req);
875 req->pages[req->num_pages] = page;
876 req->page_descs[req->num_pages].length = PAGE_SIZE;
882 static int fuse_readpages(struct file *file, struct address_space *mapping,
883 struct list_head *pages, unsigned nr_pages)
885 struct inode *inode = mapping->host;
886 struct fuse_conn *fc = get_fuse_conn(inode);
887 struct fuse_fill_data data;
889 int nr_alloc = min_t(unsigned, nr_pages, FUSE_MAX_PAGES_PER_REQ);
892 if (is_bad_inode(inode))
898 data.req = fuse_get_req_for_background(fc, nr_alloc);
900 data.req = fuse_get_req(fc, nr_alloc);
901 data.nr_pages = nr_pages;
902 err = PTR_ERR(data.req);
903 if (IS_ERR(data.req))
906 err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
908 if (data.req->num_pages)
909 fuse_send_readpages(data.req, file);
911 fuse_put_request(fc, data.req);
917 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
919 struct inode *inode = iocb->ki_filp->f_mapping->host;
920 struct fuse_conn *fc = get_fuse_conn(inode);
923 * In auto invalidate mode, always update attributes on read.
924 * Otherwise, only update if we attempt to read past EOF (to ensure
925 * i_size is up to date).
927 if (fc->auto_inval_data ||
928 (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
930 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
935 return generic_file_read_iter(iocb, to);
938 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
939 loff_t pos, size_t count)
941 struct fuse_write_in *inarg = &req->misc.write.in;
942 struct fuse_write_out *outarg = &req->misc.write.out;
947 req->in.h.opcode = FUSE_WRITE;
948 req->in.h.nodeid = ff->nodeid;
950 if (ff->fc->minor < 9)
951 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
953 req->in.args[0].size = sizeof(struct fuse_write_in);
954 req->in.args[0].value = inarg;
955 req->in.args[1].size = count;
956 req->out.numargs = 1;
957 req->out.args[0].size = sizeof(struct fuse_write_out);
958 req->out.args[0].value = outarg;
961 static size_t fuse_send_write(struct fuse_req *req, struct fuse_io_priv *io,
962 loff_t pos, size_t count, fl_owner_t owner)
964 struct file *file = io->file;
965 struct fuse_file *ff = file->private_data;
966 struct fuse_conn *fc = ff->fc;
967 struct fuse_write_in *inarg = &req->misc.write.in;
969 fuse_write_fill(req, ff, pos, count);
970 inarg->flags = file->f_flags;
972 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
973 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
977 return fuse_async_req_send(fc, req, count, io);
979 fuse_request_send(fc, req);
980 return req->misc.write.out.size;
983 bool fuse_write_update_size(struct inode *inode, loff_t pos)
985 struct fuse_conn *fc = get_fuse_conn(inode);
986 struct fuse_inode *fi = get_fuse_inode(inode);
989 spin_lock(&fc->lock);
990 fi->attr_version = ++fc->attr_version;
991 if (pos > inode->i_size) {
992 i_size_write(inode, pos);
995 spin_unlock(&fc->lock);
1000 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
1001 struct inode *inode, loff_t pos,
1007 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
1009 for (i = 0; i < req->num_pages; i++)
1010 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
1012 res = fuse_send_write(req, &io, pos, count, NULL);
1014 offset = req->page_descs[0].offset;
1016 for (i = 0; i < req->num_pages; i++) {
1017 struct page *page = req->pages[i];
1019 if (!req->out.h.error && !offset && count >= PAGE_SIZE)
1020 SetPageUptodate(page);
1022 if (count > PAGE_SIZE - offset)
1023 count -= PAGE_SIZE - offset;
1035 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
1036 struct address_space *mapping,
1037 struct iov_iter *ii, loff_t pos)
1039 struct fuse_conn *fc = get_fuse_conn(mapping->host);
1040 unsigned offset = pos & (PAGE_SIZE - 1);
1044 req->in.argpages = 1;
1045 req->page_descs[0].offset = offset;
1050 pgoff_t index = pos >> PAGE_SHIFT;
1051 size_t bytes = min_t(size_t, PAGE_SIZE - offset,
1052 iov_iter_count(ii));
1054 bytes = min_t(size_t, bytes, fc->max_write - count);
1058 if (iov_iter_fault_in_readable(ii, bytes))
1062 page = grab_cache_page_write_begin(mapping, index, 0);
1066 if (mapping_writably_mapped(mapping))
1067 flush_dcache_page(page);
1069 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
1070 flush_dcache_page(page);
1072 iov_iter_advance(ii, tmp);
1076 bytes = min(bytes, iov_iter_single_seg_count(ii));
1081 req->pages[req->num_pages] = page;
1082 req->page_descs[req->num_pages].length = tmp;
1088 if (offset == PAGE_SIZE)
1091 if (!fc->big_writes)
1093 } while (iov_iter_count(ii) && count < fc->max_write &&
1094 req->num_pages < req->max_pages && offset == 0);
1096 return count > 0 ? count : err;
1099 static inline unsigned fuse_wr_pages(loff_t pos, size_t len)
1101 return min_t(unsigned,
1102 ((pos + len - 1) >> PAGE_SHIFT) -
1103 (pos >> PAGE_SHIFT) + 1,
1104 FUSE_MAX_PAGES_PER_REQ);
1107 static ssize_t fuse_perform_write(struct file *file,
1108 struct address_space *mapping,
1109 struct iov_iter *ii, loff_t pos)
1111 struct inode *inode = mapping->host;
1112 struct fuse_conn *fc = get_fuse_conn(inode);
1113 struct fuse_inode *fi = get_fuse_inode(inode);
1117 if (is_bad_inode(inode))
1120 if (inode->i_size < pos + iov_iter_count(ii))
1121 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1124 struct fuse_req *req;
1126 unsigned nr_pages = fuse_wr_pages(pos, iov_iter_count(ii));
1128 req = fuse_get_req(fc, nr_pages);
1134 count = fuse_fill_write_pages(req, mapping, ii, pos);
1140 num_written = fuse_send_write_pages(req, file, inode,
1142 err = req->out.h.error;
1147 /* break out of the loop on short write */
1148 if (num_written != count)
1152 fuse_put_request(fc, req);
1153 } while (!err && iov_iter_count(ii));
1156 fuse_write_update_size(inode, pos);
1158 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1159 fuse_invalidate_attr(inode);
1161 return res > 0 ? res : err;
1164 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1166 struct file *file = iocb->ki_filp;
1167 struct address_space *mapping = file->f_mapping;
1168 ssize_t written = 0;
1169 ssize_t written_buffered = 0;
1170 struct inode *inode = mapping->host;
1174 if (get_fuse_conn(inode)->writeback_cache) {
1175 /* Update size (EOF optimization) and mode (SUID clearing) */
1176 err = fuse_update_attributes(mapping->host, NULL, file, NULL);
1180 return generic_file_write_iter(iocb, from);
1185 /* We can write back this queue in page reclaim */
1186 current->backing_dev_info = inode_to_bdi(inode);
1188 err = generic_write_checks(iocb, from);
1192 err = file_remove_privs(file);
1196 err = file_update_time(file);
1200 if (iocb->ki_flags & IOCB_DIRECT) {
1201 loff_t pos = iocb->ki_pos;
1202 written = generic_file_direct_write(iocb, from);
1203 if (written < 0 || !iov_iter_count(from))
1208 written_buffered = fuse_perform_write(file, mapping, from, pos);
1209 if (written_buffered < 0) {
1210 err = written_buffered;
1213 endbyte = pos + written_buffered - 1;
1215 err = filemap_write_and_wait_range(file->f_mapping, pos,
1220 invalidate_mapping_pages(file->f_mapping,
1222 endbyte >> PAGE_SHIFT);
1224 written += written_buffered;
1225 iocb->ki_pos = pos + written_buffered;
1227 written = fuse_perform_write(file, mapping, from, iocb->ki_pos);
1229 iocb->ki_pos += written;
1232 current->backing_dev_info = NULL;
1233 inode_unlock(inode);
1235 return written ? written : err;
1238 static inline void fuse_page_descs_length_init(struct fuse_req *req,
1239 unsigned index, unsigned nr_pages)
1243 for (i = index; i < index + nr_pages; i++)
1244 req->page_descs[i].length = PAGE_SIZE -
1245 req->page_descs[i].offset;
1248 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1250 return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1253 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1256 return min(iov_iter_single_seg_count(ii), max_size);
1259 static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1260 size_t *nbytesp, int write)
1262 size_t nbytes = 0; /* # bytes already packed in req */
1265 /* Special case for kernel I/O: can copy directly into the buffer */
1266 if (ii->type & ITER_KVEC) {
1267 unsigned long user_addr = fuse_get_user_addr(ii);
1268 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1271 req->in.args[1].value = (void *) user_addr;
1273 req->out.args[0].value = (void *) user_addr;
1275 iov_iter_advance(ii, frag_size);
1276 *nbytesp = frag_size;
1280 while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1283 ret = iov_iter_get_pages(ii, &req->pages[req->num_pages],
1285 req->max_pages - req->num_pages,
1290 iov_iter_advance(ii, ret);
1294 npages = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
1296 req->page_descs[req->num_pages].offset = start;
1297 fuse_page_descs_length_init(req, req->num_pages, npages);
1299 req->num_pages += npages;
1300 req->page_descs[req->num_pages - 1].length -=
1301 (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1305 req->in.argpages = 1;
1307 req->out.argpages = 1;
1311 return ret < 0 ? ret : 0;
1314 static inline int fuse_iter_npages(const struct iov_iter *ii_p)
1316 return iov_iter_npages(ii_p, FUSE_MAX_PAGES_PER_REQ);
1319 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1320 loff_t *ppos, int flags)
1322 int write = flags & FUSE_DIO_WRITE;
1323 bool should_dirty = !write && iter_is_iovec(iter);
1324 int cuse = flags & FUSE_DIO_CUSE;
1325 struct file *file = io->file;
1326 struct inode *inode = file->f_mapping->host;
1327 struct fuse_file *ff = file->private_data;
1328 struct fuse_conn *fc = ff->fc;
1329 size_t nmax = write ? fc->max_write : fc->max_read;
1331 size_t count = iov_iter_count(iter);
1332 pgoff_t idx_from = pos >> PAGE_SHIFT;
1333 pgoff_t idx_to = (pos + count - 1) >> PAGE_SHIFT;
1335 struct fuse_req *req;
1339 req = fuse_get_req_for_background(fc, fuse_iter_npages(iter));
1341 req = fuse_get_req(fc, fuse_iter_npages(iter));
1343 return PTR_ERR(req);
1345 if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1348 fuse_sync_writes(inode);
1350 inode_unlock(inode);
1355 fl_owner_t owner = current->files;
1356 size_t nbytes = min(count, nmax);
1357 err = fuse_get_user_pages(req, iter, &nbytes, write);
1362 nres = fuse_send_write(req, io, pos, nbytes, owner);
1364 nres = fuse_send_read(req, io, pos, nbytes, owner);
1367 fuse_release_user_pages(req, should_dirty);
1368 if (req->out.h.error) {
1369 err = req->out.h.error;
1371 } else if (nres > nbytes) {
1382 fuse_put_request(fc, req);
1384 req = fuse_get_req_for_background(fc,
1385 fuse_iter_npages(iter));
1387 req = fuse_get_req(fc, fuse_iter_npages(iter));
1393 fuse_put_request(fc, req);
1397 return res > 0 ? res : err;
1399 EXPORT_SYMBOL_GPL(fuse_direct_io);
1401 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1402 struct iov_iter *iter,
1406 struct file *file = io->file;
1407 struct inode *inode = file_inode(file);
1409 if (is_bad_inode(inode))
1412 res = fuse_direct_io(io, iter, ppos, 0);
1414 fuse_invalidate_attr(inode);
1419 static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1421 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb->ki_filp);
1422 return __fuse_direct_read(&io, to, &iocb->ki_pos);
1425 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1427 struct file *file = iocb->ki_filp;
1428 struct inode *inode = file_inode(file);
1429 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
1432 if (is_bad_inode(inode))
1435 /* Don't allow parallel writes to the same file */
1437 res = generic_write_checks(iocb, from);
1439 res = fuse_direct_io(&io, from, &iocb->ki_pos, FUSE_DIO_WRITE);
1440 fuse_invalidate_attr(inode);
1442 fuse_write_update_size(inode, iocb->ki_pos);
1443 inode_unlock(inode);
1448 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1452 for (i = 0; i < req->num_pages; i++)
1453 __free_page(req->pages[i]);
1456 fuse_file_put(req->ff, false);
1459 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1461 struct inode *inode = req->inode;
1462 struct fuse_inode *fi = get_fuse_inode(inode);
1463 struct backing_dev_info *bdi = inode_to_bdi(inode);
1466 list_del(&req->writepages_entry);
1467 for (i = 0; i < req->num_pages; i++) {
1468 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1469 dec_node_page_state(req->pages[i], NR_WRITEBACK_TEMP);
1470 wb_writeout_inc(&bdi->wb);
1472 wake_up(&fi->page_waitq);
1475 /* Called under fc->lock, may release and reacquire it */
1476 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req,
1478 __releases(fc->lock)
1479 __acquires(fc->lock)
1481 struct fuse_inode *fi = get_fuse_inode(req->inode);
1482 struct fuse_write_in *inarg = &req->misc.write.in;
1483 __u64 data_size = req->num_pages * PAGE_SIZE;
1488 if (inarg->offset + data_size <= size) {
1489 inarg->size = data_size;
1490 } else if (inarg->offset < size) {
1491 inarg->size = size - inarg->offset;
1493 /* Got truncated off completely */
1497 req->in.args[1].size = inarg->size;
1499 fuse_request_send_background_locked(fc, req);
1503 fuse_writepage_finish(fc, req);
1504 spin_unlock(&fc->lock);
1505 fuse_writepage_free(fc, req);
1506 fuse_put_request(fc, req);
1507 spin_lock(&fc->lock);
1511 * If fi->writectr is positive (no truncate or fsync going on) send
1512 * all queued writepage requests.
1514 * Called with fc->lock
1516 void fuse_flush_writepages(struct inode *inode)
1517 __releases(fc->lock)
1518 __acquires(fc->lock)
1520 struct fuse_conn *fc = get_fuse_conn(inode);
1521 struct fuse_inode *fi = get_fuse_inode(inode);
1522 size_t crop = i_size_read(inode);
1523 struct fuse_req *req;
1525 while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1526 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1527 list_del_init(&req->list);
1528 fuse_send_writepage(fc, req, crop);
1532 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1534 struct inode *inode = req->inode;
1535 struct fuse_inode *fi = get_fuse_inode(inode);
1537 mapping_set_error(inode->i_mapping, req->out.h.error);
1538 spin_lock(&fc->lock);
1539 while (req->misc.write.next) {
1540 struct fuse_conn *fc = get_fuse_conn(inode);
1541 struct fuse_write_in *inarg = &req->misc.write.in;
1542 struct fuse_req *next = req->misc.write.next;
1543 req->misc.write.next = next->misc.write.next;
1544 next->misc.write.next = NULL;
1545 next->ff = fuse_file_get(req->ff);
1546 list_add(&next->writepages_entry, &fi->writepages);
1549 * Skip fuse_flush_writepages() to make it easy to crop requests
1550 * based on primary request size.
1552 * 1st case (trivial): there are no concurrent activities using
1553 * fuse_set/release_nowrite. Then we're on safe side because
1554 * fuse_flush_writepages() would call fuse_send_writepage()
1557 * 2nd case: someone called fuse_set_nowrite and it is waiting
1558 * now for completion of all in-flight requests. This happens
1559 * rarely and no more than once per page, so this should be
1562 * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1563 * of fuse_set_nowrite..fuse_release_nowrite section. The fact
1564 * that fuse_set_nowrite returned implies that all in-flight
1565 * requests were completed along with all of their secondary
1566 * requests. Further primary requests are blocked by negative
1567 * writectr. Hence there cannot be any in-flight requests and
1568 * no invocations of fuse_writepage_end() while we're in
1569 * fuse_set_nowrite..fuse_release_nowrite section.
1571 fuse_send_writepage(fc, next, inarg->offset + inarg->size);
1574 fuse_writepage_finish(fc, req);
1575 spin_unlock(&fc->lock);
1576 fuse_writepage_free(fc, req);
1579 static struct fuse_file *__fuse_write_file_get(struct fuse_conn *fc,
1580 struct fuse_inode *fi)
1582 struct fuse_file *ff = NULL;
1584 spin_lock(&fc->lock);
1585 if (!list_empty(&fi->write_files)) {
1586 ff = list_entry(fi->write_files.next, struct fuse_file,
1590 spin_unlock(&fc->lock);
1595 static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1596 struct fuse_inode *fi)
1598 struct fuse_file *ff = __fuse_write_file_get(fc, fi);
1603 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1605 struct fuse_conn *fc = get_fuse_conn(inode);
1606 struct fuse_inode *fi = get_fuse_inode(inode);
1607 struct fuse_file *ff;
1610 ff = __fuse_write_file_get(fc, fi);
1611 err = fuse_flush_times(inode, ff);
1613 fuse_file_put(ff, 0);
1618 static int fuse_writepage_locked(struct page *page)
1620 struct address_space *mapping = page->mapping;
1621 struct inode *inode = mapping->host;
1622 struct fuse_conn *fc = get_fuse_conn(inode);
1623 struct fuse_inode *fi = get_fuse_inode(inode);
1624 struct fuse_req *req;
1625 struct page *tmp_page;
1626 int error = -ENOMEM;
1628 set_page_writeback(page);
1630 req = fuse_request_alloc_nofs(1);
1634 /* writeback always goes to bg_queue */
1635 __set_bit(FR_BACKGROUND, &req->flags);
1636 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1641 req->ff = fuse_write_file_get(fc, fi);
1645 fuse_write_fill(req, req->ff, page_offset(page), 0);
1647 copy_highpage(tmp_page, page);
1648 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1649 req->misc.write.next = NULL;
1650 req->in.argpages = 1;
1652 req->pages[0] = tmp_page;
1653 req->page_descs[0].offset = 0;
1654 req->page_descs[0].length = PAGE_SIZE;
1655 req->end = fuse_writepage_end;
1658 inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1659 inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1661 spin_lock(&fc->lock);
1662 list_add(&req->writepages_entry, &fi->writepages);
1663 list_add_tail(&req->list, &fi->queued_writes);
1664 fuse_flush_writepages(inode);
1665 spin_unlock(&fc->lock);
1667 end_page_writeback(page);
1672 __free_page(tmp_page);
1674 fuse_request_free(req);
1676 end_page_writeback(page);
1680 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1684 if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1686 * ->writepages() should be called for sync() and friends. We
1687 * should only get here on direct reclaim and then we are
1688 * allowed to skip a page which is already in flight
1690 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1692 redirty_page_for_writepage(wbc, page);
1696 err = fuse_writepage_locked(page);
1702 struct fuse_fill_wb_data {
1703 struct fuse_req *req;
1704 struct fuse_file *ff;
1705 struct inode *inode;
1706 struct page **orig_pages;
1709 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1711 struct fuse_req *req = data->req;
1712 struct inode *inode = data->inode;
1713 struct fuse_conn *fc = get_fuse_conn(inode);
1714 struct fuse_inode *fi = get_fuse_inode(inode);
1715 int num_pages = req->num_pages;
1718 req->ff = fuse_file_get(data->ff);
1719 spin_lock(&fc->lock);
1720 list_add_tail(&req->list, &fi->queued_writes);
1721 fuse_flush_writepages(inode);
1722 spin_unlock(&fc->lock);
1724 for (i = 0; i < num_pages; i++)
1725 end_page_writeback(data->orig_pages[i]);
1728 static bool fuse_writepage_in_flight(struct fuse_req *new_req,
1731 struct fuse_conn *fc = get_fuse_conn(new_req->inode);
1732 struct fuse_inode *fi = get_fuse_inode(new_req->inode);
1733 struct fuse_req *tmp;
1734 struct fuse_req *old_req;
1738 BUG_ON(new_req->num_pages != 0);
1740 spin_lock(&fc->lock);
1741 list_del(&new_req->writepages_entry);
1742 list_for_each_entry(old_req, &fi->writepages, writepages_entry) {
1743 BUG_ON(old_req->inode != new_req->inode);
1744 curr_index = old_req->misc.write.in.offset >> PAGE_SHIFT;
1745 if (curr_index <= page->index &&
1746 page->index < curr_index + old_req->num_pages) {
1752 list_add(&new_req->writepages_entry, &fi->writepages);
1756 new_req->num_pages = 1;
1757 for (tmp = old_req; tmp != NULL; tmp = tmp->misc.write.next) {
1758 BUG_ON(tmp->inode != new_req->inode);
1759 curr_index = tmp->misc.write.in.offset >> PAGE_SHIFT;
1760 if (tmp->num_pages == 1 &&
1761 curr_index == page->index) {
1766 if (old_req->num_pages == 1 && test_bit(FR_PENDING, &old_req->flags)) {
1767 struct backing_dev_info *bdi = inode_to_bdi(page->mapping->host);
1769 copy_highpage(old_req->pages[0], page);
1770 spin_unlock(&fc->lock);
1772 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1773 dec_node_page_state(page, NR_WRITEBACK_TEMP);
1774 wb_writeout_inc(&bdi->wb);
1775 fuse_writepage_free(fc, new_req);
1776 fuse_request_free(new_req);
1779 new_req->misc.write.next = old_req->misc.write.next;
1780 old_req->misc.write.next = new_req;
1783 spin_unlock(&fc->lock);
1788 static int fuse_writepages_fill(struct page *page,
1789 struct writeback_control *wbc, void *_data)
1791 struct fuse_fill_wb_data *data = _data;
1792 struct fuse_req *req = data->req;
1793 struct inode *inode = data->inode;
1794 struct fuse_conn *fc = get_fuse_conn(inode);
1795 struct page *tmp_page;
1801 data->ff = fuse_write_file_get(fc, get_fuse_inode(inode));
1807 * Being under writeback is unlikely but possible. For example direct
1808 * read to an mmaped fuse file will set the page dirty twice; once when
1809 * the pages are faulted with get_user_pages(), and then after the read
1812 is_writeback = fuse_page_is_writeback(inode, page->index);
1814 if (req && req->num_pages &&
1815 (is_writeback || req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
1816 (req->num_pages + 1) * PAGE_SIZE > fc->max_write ||
1817 data->orig_pages[req->num_pages - 1]->index + 1 != page->index)) {
1818 fuse_writepages_send(data);
1822 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1827 * The page must not be redirtied until the writeout is completed
1828 * (i.e. userspace has sent a reply to the write request). Otherwise
1829 * there could be more than one temporary page instance for each real
1832 * This is ensured by holding the page lock in page_mkwrite() while
1833 * checking fuse_page_is_writeback(). We already hold the page lock
1834 * since clear_page_dirty_for_io() and keep it held until we add the
1835 * request to the fi->writepages list and increment req->num_pages.
1836 * After this fuse_page_is_writeback() will indicate that the page is
1837 * under writeback, so we can release the page lock.
1839 if (data->req == NULL) {
1840 struct fuse_inode *fi = get_fuse_inode(inode);
1843 req = fuse_request_alloc_nofs(FUSE_MAX_PAGES_PER_REQ);
1845 __free_page(tmp_page);
1849 fuse_write_fill(req, data->ff, page_offset(page), 0);
1850 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1851 req->misc.write.next = NULL;
1852 req->in.argpages = 1;
1853 __set_bit(FR_BACKGROUND, &req->flags);
1855 req->end = fuse_writepage_end;
1858 spin_lock(&fc->lock);
1859 list_add(&req->writepages_entry, &fi->writepages);
1860 spin_unlock(&fc->lock);
1864 set_page_writeback(page);
1866 copy_highpage(tmp_page, page);
1867 req->pages[req->num_pages] = tmp_page;
1868 req->page_descs[req->num_pages].offset = 0;
1869 req->page_descs[req->num_pages].length = PAGE_SIZE;
1871 inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1872 inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1875 if (is_writeback && fuse_writepage_in_flight(req, page)) {
1876 end_page_writeback(page);
1880 data->orig_pages[req->num_pages] = page;
1883 * Protected by fc->lock against concurrent access by
1884 * fuse_page_is_writeback().
1886 spin_lock(&fc->lock);
1888 spin_unlock(&fc->lock);
1896 static int fuse_writepages(struct address_space *mapping,
1897 struct writeback_control *wbc)
1899 struct inode *inode = mapping->host;
1900 struct fuse_fill_wb_data data;
1904 if (is_bad_inode(inode))
1912 data.orig_pages = kcalloc(FUSE_MAX_PAGES_PER_REQ,
1913 sizeof(struct page *),
1915 if (!data.orig_pages)
1918 err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
1920 /* Ignore errors if we can write at least one page */
1921 BUG_ON(!data.req->num_pages);
1922 fuse_writepages_send(&data);
1926 fuse_file_put(data.ff, false);
1928 kfree(data.orig_pages);
1934 * It's worthy to make sure that space is reserved on disk for the write,
1935 * but how to implement it without killing performance need more thinking.
1937 static int fuse_write_begin(struct file *file, struct address_space *mapping,
1938 loff_t pos, unsigned len, unsigned flags,
1939 struct page **pagep, void **fsdata)
1941 pgoff_t index = pos >> PAGE_SHIFT;
1942 struct fuse_conn *fc = get_fuse_conn(file_inode(file));
1947 WARN_ON(!fc->writeback_cache);
1949 page = grab_cache_page_write_begin(mapping, index, flags);
1953 fuse_wait_on_page_writeback(mapping->host, page->index);
1955 if (PageUptodate(page) || len == PAGE_SIZE)
1958 * Check if the start this page comes after the end of file, in which
1959 * case the readpage can be optimized away.
1961 fsize = i_size_read(mapping->host);
1962 if (fsize <= (pos & PAGE_MASK)) {
1963 size_t off = pos & ~PAGE_MASK;
1965 zero_user_segment(page, 0, off);
1968 err = fuse_do_readpage(file, page);
1982 static int fuse_write_end(struct file *file, struct address_space *mapping,
1983 loff_t pos, unsigned len, unsigned copied,
1984 struct page *page, void *fsdata)
1986 struct inode *inode = page->mapping->host;
1988 if (!PageUptodate(page)) {
1989 /* Zero any unwritten bytes at the end of the page */
1990 size_t endoff = (pos + copied) & ~PAGE_MASK;
1992 zero_user_segment(page, endoff, PAGE_SIZE);
1993 SetPageUptodate(page);
1996 fuse_write_update_size(inode, pos + copied);
1997 set_page_dirty(page);
2004 static int fuse_launder_page(struct page *page)
2007 if (clear_page_dirty_for_io(page)) {
2008 struct inode *inode = page->mapping->host;
2009 err = fuse_writepage_locked(page);
2011 fuse_wait_on_page_writeback(inode, page->index);
2017 * Write back dirty pages now, because there may not be any suitable
2020 static void fuse_vma_close(struct vm_area_struct *vma)
2022 filemap_write_and_wait(vma->vm_file->f_mapping);
2026 * Wait for writeback against this page to complete before allowing it
2027 * to be marked dirty again, and hence written back again, possibly
2028 * before the previous writepage completed.
2030 * Block here, instead of in ->writepage(), so that the userspace fs
2031 * can only block processes actually operating on the filesystem.
2033 * Otherwise unprivileged userspace fs would be able to block
2038 * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2040 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
2042 struct page *page = vmf->page;
2043 struct inode *inode = file_inode(vma->vm_file);
2045 file_update_time(vma->vm_file);
2047 if (page->mapping != inode->i_mapping) {
2049 return VM_FAULT_NOPAGE;
2052 fuse_wait_on_page_writeback(inode, page->index);
2053 return VM_FAULT_LOCKED;
2056 static const struct vm_operations_struct fuse_file_vm_ops = {
2057 .close = fuse_vma_close,
2058 .fault = filemap_fault,
2059 .map_pages = filemap_map_pages,
2060 .page_mkwrite = fuse_page_mkwrite,
2063 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2065 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2066 fuse_link_write_file(file);
2068 file_accessed(file);
2069 vma->vm_ops = &fuse_file_vm_ops;
2073 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
2075 /* Can't provide the coherency needed for MAP_SHARED */
2076 if (vma->vm_flags & VM_MAYSHARE)
2079 invalidate_inode_pages2(file->f_mapping);
2081 return generic_file_mmap(file, vma);
2084 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
2085 struct file_lock *fl)
2087 switch (ffl->type) {
2093 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2094 ffl->end < ffl->start)
2097 fl->fl_start = ffl->start;
2098 fl->fl_end = ffl->end;
2099 fl->fl_pid = ffl->pid;
2105 fl->fl_type = ffl->type;
2109 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2110 const struct file_lock *fl, int opcode, pid_t pid,
2111 int flock, struct fuse_lk_in *inarg)
2113 struct inode *inode = file_inode(file);
2114 struct fuse_conn *fc = get_fuse_conn(inode);
2115 struct fuse_file *ff = file->private_data;
2117 memset(inarg, 0, sizeof(*inarg));
2119 inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2120 inarg->lk.start = fl->fl_start;
2121 inarg->lk.end = fl->fl_end;
2122 inarg->lk.type = fl->fl_type;
2123 inarg->lk.pid = pid;
2125 inarg->lk_flags |= FUSE_LK_FLOCK;
2126 args->in.h.opcode = opcode;
2127 args->in.h.nodeid = get_node_id(inode);
2128 args->in.numargs = 1;
2129 args->in.args[0].size = sizeof(*inarg);
2130 args->in.args[0].value = inarg;
2133 static int fuse_getlk(struct file *file, struct file_lock *fl)
2135 struct inode *inode = file_inode(file);
2136 struct fuse_conn *fc = get_fuse_conn(inode);
2138 struct fuse_lk_in inarg;
2139 struct fuse_lk_out outarg;
2142 fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2143 args.out.numargs = 1;
2144 args.out.args[0].size = sizeof(outarg);
2145 args.out.args[0].value = &outarg;
2146 err = fuse_simple_request(fc, &args);
2148 err = convert_fuse_file_lock(&outarg.lk, fl);
2153 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2155 struct inode *inode = file_inode(file);
2156 struct fuse_conn *fc = get_fuse_conn(inode);
2158 struct fuse_lk_in inarg;
2159 int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2160 pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
2163 if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2164 /* NLM needs asynchronous locks, which we don't support yet */
2168 /* Unlock on close is handled by the flush method */
2169 if (fl->fl_flags & FL_CLOSE)
2172 fuse_lk_fill(&args, file, fl, opcode, pid, flock, &inarg);
2173 err = fuse_simple_request(fc, &args);
2175 /* locking is restartable */
2182 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2184 struct inode *inode = file_inode(file);
2185 struct fuse_conn *fc = get_fuse_conn(inode);
2188 if (cmd == F_CANCELLK) {
2190 } else if (cmd == F_GETLK) {
2192 posix_test_lock(file, fl);
2195 err = fuse_getlk(file, fl);
2198 err = posix_lock_file(file, fl, NULL);
2200 err = fuse_setlk(file, fl, 0);
2205 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2207 struct inode *inode = file_inode(file);
2208 struct fuse_conn *fc = get_fuse_conn(inode);
2212 err = locks_lock_file_wait(file, fl);
2214 struct fuse_file *ff = file->private_data;
2216 /* emulate flock with POSIX locks */
2218 err = fuse_setlk(file, fl, 1);
2224 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2226 struct inode *inode = mapping->host;
2227 struct fuse_conn *fc = get_fuse_conn(inode);
2229 struct fuse_bmap_in inarg;
2230 struct fuse_bmap_out outarg;
2233 if (!inode->i_sb->s_bdev || fc->no_bmap)
2236 memset(&inarg, 0, sizeof(inarg));
2237 inarg.block = block;
2238 inarg.blocksize = inode->i_sb->s_blocksize;
2239 args.in.h.opcode = FUSE_BMAP;
2240 args.in.h.nodeid = get_node_id(inode);
2241 args.in.numargs = 1;
2242 args.in.args[0].size = sizeof(inarg);
2243 args.in.args[0].value = &inarg;
2244 args.out.numargs = 1;
2245 args.out.args[0].size = sizeof(outarg);
2246 args.out.args[0].value = &outarg;
2247 err = fuse_simple_request(fc, &args);
2251 return err ? 0 : outarg.block;
2254 static loff_t fuse_lseek(struct file *file, loff_t offset, int whence)
2256 struct inode *inode = file->f_mapping->host;
2257 struct fuse_conn *fc = get_fuse_conn(inode);
2258 struct fuse_file *ff = file->private_data;
2260 struct fuse_lseek_in inarg = {
2265 struct fuse_lseek_out outarg;
2271 args.in.h.opcode = FUSE_LSEEK;
2272 args.in.h.nodeid = ff->nodeid;
2273 args.in.numargs = 1;
2274 args.in.args[0].size = sizeof(inarg);
2275 args.in.args[0].value = &inarg;
2276 args.out.numargs = 1;
2277 args.out.args[0].size = sizeof(outarg);
2278 args.out.args[0].value = &outarg;
2279 err = fuse_simple_request(fc, &args);
2281 if (err == -ENOSYS) {
2288 return vfs_setpos(file, outarg.offset, inode->i_sb->s_maxbytes);
2291 err = fuse_update_attributes(inode, NULL, file, NULL);
2293 return generic_file_llseek(file, offset, whence);
2298 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2301 struct inode *inode = file_inode(file);
2306 /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2307 retval = generic_file_llseek(file, offset, whence);
2311 retval = fuse_update_attributes(inode, NULL, file, NULL);
2313 retval = generic_file_llseek(file, offset, whence);
2314 inode_unlock(inode);
2319 retval = fuse_lseek(file, offset, whence);
2320 inode_unlock(inode);
2330 * CUSE servers compiled on 32bit broke on 64bit kernels because the
2331 * ABI was defined to be 'struct iovec' which is different on 32bit
2332 * and 64bit. Fortunately we can determine which structure the server
2333 * used from the size of the reply.
2335 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2336 size_t transferred, unsigned count,
2339 #ifdef CONFIG_COMPAT
2340 if (count * sizeof(struct compat_iovec) == transferred) {
2341 struct compat_iovec *ciov = src;
2345 * With this interface a 32bit server cannot support
2346 * non-compat (i.e. ones coming from 64bit apps) ioctl
2352 for (i = 0; i < count; i++) {
2353 dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2354 dst[i].iov_len = ciov[i].iov_len;
2360 if (count * sizeof(struct iovec) != transferred)
2363 memcpy(dst, src, transferred);
2367 /* Make sure iov_length() won't overflow */
2368 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
2371 u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
2373 for (n = 0; n < count; n++, iov++) {
2374 if (iov->iov_len > (size_t) max)
2376 max -= iov->iov_len;
2381 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2382 void *src, size_t transferred, unsigned count,
2386 struct fuse_ioctl_iovec *fiov = src;
2388 if (fc->minor < 16) {
2389 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2393 if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2396 for (i = 0; i < count; i++) {
2397 /* Did the server supply an inappropriate value? */
2398 if (fiov[i].base != (unsigned long) fiov[i].base ||
2399 fiov[i].len != (unsigned long) fiov[i].len)
2402 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2403 dst[i].iov_len = (size_t) fiov[i].len;
2405 #ifdef CONFIG_COMPAT
2407 (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2408 (compat_size_t) dst[i].iov_len != fiov[i].len))
2418 * For ioctls, there is no generic way to determine how much memory
2419 * needs to be read and/or written. Furthermore, ioctls are allowed
2420 * to dereference the passed pointer, so the parameter requires deep
2421 * copying but FUSE has no idea whatsoever about what to copy in or
2424 * This is solved by allowing FUSE server to retry ioctl with
2425 * necessary in/out iovecs. Let's assume the ioctl implementation
2426 * needs to read in the following structure.
2433 * On the first callout to FUSE server, inarg->in_size and
2434 * inarg->out_size will be NULL; then, the server completes the ioctl
2435 * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2436 * the actual iov array to
2438 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) } }
2440 * which tells FUSE to copy in the requested area and retry the ioctl.
2441 * On the second round, the server has access to the structure and
2442 * from that it can tell what to look for next, so on the invocation,
2443 * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2445 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) },
2446 * { .iov_base = a.buf, .iov_len = a.buflen } }
2448 * FUSE will copy both struct a and the pointed buffer from the
2449 * process doing the ioctl and retry ioctl with both struct a and the
2452 * This time, FUSE server has everything it needs and completes ioctl
2453 * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2455 * Copying data out works the same way.
2457 * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2458 * automatically initializes in and out iovs by decoding @cmd with
2459 * _IOC_* macros and the server is not allowed to request RETRY. This
2460 * limits ioctl data transfers to well-formed ioctls and is the forced
2461 * behavior for all FUSE servers.
2463 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2466 struct fuse_file *ff = file->private_data;
2467 struct fuse_conn *fc = ff->fc;
2468 struct fuse_ioctl_in inarg = {
2474 struct fuse_ioctl_out outarg;
2475 struct fuse_req *req = NULL;
2476 struct page **pages = NULL;
2477 struct iovec *iov_page = NULL;
2478 struct iovec *in_iov = NULL, *out_iov = NULL;
2479 unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
2480 size_t in_size, out_size, transferred, c;
2484 #if BITS_PER_LONG == 32
2485 inarg.flags |= FUSE_IOCTL_32BIT;
2487 if (flags & FUSE_IOCTL_COMPAT)
2488 inarg.flags |= FUSE_IOCTL_32BIT;
2491 /* assume all the iovs returned by client always fits in a page */
2492 BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2495 pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
2496 iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2497 if (!pages || !iov_page)
2501 * If restricted, initialize IO parameters as encoded in @cmd.
2502 * RETRY from server is not allowed.
2504 if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2505 struct iovec *iov = iov_page;
2507 iov->iov_base = (void __user *)arg;
2508 iov->iov_len = _IOC_SIZE(cmd);
2510 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2515 if (_IOC_DIR(cmd) & _IOC_READ) {
2522 inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2523 inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2526 * Out data can be used either for actual out data or iovs,
2527 * make sure there always is at least one page.
2529 out_size = max_t(size_t, out_size, PAGE_SIZE);
2530 max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2532 /* make sure there are enough buffer pages and init request with them */
2534 if (max_pages > FUSE_MAX_PAGES_PER_REQ)
2536 while (num_pages < max_pages) {
2537 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2538 if (!pages[num_pages])
2543 req = fuse_get_req(fc, num_pages);
2549 memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
2550 req->num_pages = num_pages;
2551 fuse_page_descs_length_init(req, 0, req->num_pages);
2553 /* okay, let's send it to the client */
2554 req->in.h.opcode = FUSE_IOCTL;
2555 req->in.h.nodeid = ff->nodeid;
2556 req->in.numargs = 1;
2557 req->in.args[0].size = sizeof(inarg);
2558 req->in.args[0].value = &inarg;
2561 req->in.args[1].size = in_size;
2562 req->in.argpages = 1;
2565 iov_iter_init(&ii, WRITE, in_iov, in_iovs, in_size);
2566 for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= num_pages); i++) {
2567 c = copy_page_from_iter(pages[i], 0, PAGE_SIZE, &ii);
2568 if (c != PAGE_SIZE && iov_iter_count(&ii))
2573 req->out.numargs = 2;
2574 req->out.args[0].size = sizeof(outarg);
2575 req->out.args[0].value = &outarg;
2576 req->out.args[1].size = out_size;
2577 req->out.argpages = 1;
2578 req->out.argvar = 1;
2580 fuse_request_send(fc, req);
2581 err = req->out.h.error;
2582 transferred = req->out.args[1].size;
2583 fuse_put_request(fc, req);
2588 /* did it ask for retry? */
2589 if (outarg.flags & FUSE_IOCTL_RETRY) {
2592 /* no retry if in restricted mode */
2594 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2597 in_iovs = outarg.in_iovs;
2598 out_iovs = outarg.out_iovs;
2601 * Make sure things are in boundary, separate checks
2602 * are to protect against overflow.
2605 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2606 out_iovs > FUSE_IOCTL_MAX_IOV ||
2607 in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2610 vaddr = kmap_atomic(pages[0]);
2611 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2612 transferred, in_iovs + out_iovs,
2613 (flags & FUSE_IOCTL_COMPAT) != 0);
2614 kunmap_atomic(vaddr);
2619 out_iov = in_iov + in_iovs;
2621 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
2625 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
2633 if (transferred > inarg.out_size)
2637 iov_iter_init(&ii, READ, out_iov, out_iovs, transferred);
2638 for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= num_pages); i++) {
2639 c = copy_page_to_iter(pages[i], 0, PAGE_SIZE, &ii);
2640 if (c != PAGE_SIZE && iov_iter_count(&ii))
2646 fuse_put_request(fc, req);
2647 free_page((unsigned long) iov_page);
2649 __free_page(pages[--num_pages]);
2652 return err ? err : outarg.result;
2654 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2656 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2657 unsigned long arg, unsigned int flags)
2659 struct inode *inode = file_inode(file);
2660 struct fuse_conn *fc = get_fuse_conn(inode);
2662 if (!fuse_allow_current_process(fc))
2665 if (is_bad_inode(inode))
2668 return fuse_do_ioctl(file, cmd, arg, flags);
2671 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2674 return fuse_ioctl_common(file, cmd, arg, 0);
2677 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2680 return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2684 * All files which have been polled are linked to RB tree
2685 * fuse_conn->polled_files which is indexed by kh. Walk the tree and
2686 * find the matching one.
2688 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2689 struct rb_node **parent_out)
2691 struct rb_node **link = &fc->polled_files.rb_node;
2692 struct rb_node *last = NULL;
2695 struct fuse_file *ff;
2698 ff = rb_entry(last, struct fuse_file, polled_node);
2701 link = &last->rb_left;
2702 else if (kh > ff->kh)
2703 link = &last->rb_right;
2714 * The file is about to be polled. Make sure it's on the polled_files
2715 * RB tree. Note that files once added to the polled_files tree are
2716 * not removed before the file is released. This is because a file
2717 * polled once is likely to be polled again.
2719 static void fuse_register_polled_file(struct fuse_conn *fc,
2720 struct fuse_file *ff)
2722 spin_lock(&fc->lock);
2723 if (RB_EMPTY_NODE(&ff->polled_node)) {
2724 struct rb_node **link, *uninitialized_var(parent);
2726 link = fuse_find_polled_node(fc, ff->kh, &parent);
2728 rb_link_node(&ff->polled_node, parent, link);
2729 rb_insert_color(&ff->polled_node, &fc->polled_files);
2731 spin_unlock(&fc->lock);
2734 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2736 struct fuse_file *ff = file->private_data;
2737 struct fuse_conn *fc = ff->fc;
2738 struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2739 struct fuse_poll_out outarg;
2744 return DEFAULT_POLLMASK;
2746 poll_wait(file, &ff->poll_wait, wait);
2747 inarg.events = (__u32)poll_requested_events(wait);
2750 * Ask for notification iff there's someone waiting for it.
2751 * The client may ignore the flag and always notify.
2753 if (waitqueue_active(&ff->poll_wait)) {
2754 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2755 fuse_register_polled_file(fc, ff);
2758 args.in.h.opcode = FUSE_POLL;
2759 args.in.h.nodeid = ff->nodeid;
2760 args.in.numargs = 1;
2761 args.in.args[0].size = sizeof(inarg);
2762 args.in.args[0].value = &inarg;
2763 args.out.numargs = 1;
2764 args.out.args[0].size = sizeof(outarg);
2765 args.out.args[0].value = &outarg;
2766 err = fuse_simple_request(fc, &args);
2769 return outarg.revents;
2770 if (err == -ENOSYS) {
2772 return DEFAULT_POLLMASK;
2776 EXPORT_SYMBOL_GPL(fuse_file_poll);
2779 * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2780 * wakes up the poll waiters.
2782 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2783 struct fuse_notify_poll_wakeup_out *outarg)
2785 u64 kh = outarg->kh;
2786 struct rb_node **link;
2788 spin_lock(&fc->lock);
2790 link = fuse_find_polled_node(fc, kh, NULL);
2792 struct fuse_file *ff;
2794 ff = rb_entry(*link, struct fuse_file, polled_node);
2795 wake_up_interruptible_sync(&ff->poll_wait);
2798 spin_unlock(&fc->lock);
2802 static void fuse_do_truncate(struct file *file)
2804 struct inode *inode = file->f_mapping->host;
2807 attr.ia_valid = ATTR_SIZE;
2808 attr.ia_size = i_size_read(inode);
2810 attr.ia_file = file;
2811 attr.ia_valid |= ATTR_FILE;
2813 fuse_do_setattr(inode, &attr, file);
2816 static inline loff_t fuse_round_up(loff_t off)
2818 return round_up(off, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
2822 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
2824 DECLARE_COMPLETION_ONSTACK(wait);
2826 struct file *file = iocb->ki_filp;
2827 struct fuse_file *ff = file->private_data;
2828 bool async_dio = ff->fc->async_dio;
2830 struct inode *inode;
2832 size_t count = iov_iter_count(iter);
2833 loff_t offset = iocb->ki_pos;
2834 struct fuse_io_priv *io;
2837 inode = file->f_mapping->host;
2838 i_size = i_size_read(inode);
2840 if ((iov_iter_rw(iter) == READ) && (offset > i_size))
2843 /* optimization for short read */
2844 if (async_dio && iov_iter_rw(iter) != WRITE && offset + count > i_size) {
2845 if (offset >= i_size)
2847 iov_iter_truncate(iter, fuse_round_up(i_size - offset));
2848 count = iov_iter_count(iter);
2851 io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2854 spin_lock_init(&io->lock);
2855 kref_init(&io->refcnt);
2859 io->offset = offset;
2860 io->write = (iov_iter_rw(iter) == WRITE);
2864 * By default, we want to optimize all I/Os with async request
2865 * submission to the client filesystem if supported.
2867 io->async = async_dio;
2869 io->blocking = is_sync_kiocb(iocb);
2872 * We cannot asynchronously extend the size of a file.
2873 * In such case the aio will behave exactly like sync io.
2875 if ((offset + count > i_size) && iov_iter_rw(iter) == WRITE)
2876 io->blocking = true;
2878 if (io->async && io->blocking) {
2880 * Additional reference to keep io around after
2881 * calling fuse_aio_complete()
2883 kref_get(&io->refcnt);
2887 if (iov_iter_rw(iter) == WRITE) {
2888 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
2889 fuse_invalidate_attr(inode);
2891 ret = __fuse_direct_read(io, iter, &pos);
2895 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2897 /* we have a non-extending, async request, so return */
2899 return -EIOCBQUEUED;
2901 wait_for_completion(&wait);
2902 ret = fuse_get_res_by_io(io);
2905 kref_put(&io->refcnt, fuse_io_release);
2907 if (iov_iter_rw(iter) == WRITE) {
2909 fuse_write_update_size(inode, pos);
2910 else if (ret < 0 && offset + count > i_size)
2911 fuse_do_truncate(file);
2917 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2920 struct fuse_file *ff = file->private_data;
2921 struct inode *inode = file_inode(file);
2922 struct fuse_inode *fi = get_fuse_inode(inode);
2923 struct fuse_conn *fc = ff->fc;
2925 struct fuse_fallocate_in inarg = {
2932 bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
2933 (mode & FALLOC_FL_PUNCH_HOLE);
2935 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2938 if (fc->no_fallocate)
2943 if (mode & FALLOC_FL_PUNCH_HOLE) {
2944 loff_t endbyte = offset + length - 1;
2945 err = filemap_write_and_wait_range(inode->i_mapping,
2950 fuse_sync_writes(inode);
2954 if (!(mode & FALLOC_FL_KEEP_SIZE))
2955 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2957 args.in.h.opcode = FUSE_FALLOCATE;
2958 args.in.h.nodeid = ff->nodeid;
2959 args.in.numargs = 1;
2960 args.in.args[0].size = sizeof(inarg);
2961 args.in.args[0].value = &inarg;
2962 err = fuse_simple_request(fc, &args);
2963 if (err == -ENOSYS) {
2964 fc->no_fallocate = 1;
2970 /* we could have extended the file */
2971 if (!(mode & FALLOC_FL_KEEP_SIZE)) {
2972 bool changed = fuse_write_update_size(inode, offset + length);
2974 if (changed && fc->writeback_cache)
2975 file_update_time(file);
2978 if (mode & FALLOC_FL_PUNCH_HOLE)
2979 truncate_pagecache_range(inode, offset, offset + length - 1);
2981 fuse_invalidate_attr(inode);
2984 if (!(mode & FALLOC_FL_KEEP_SIZE))
2985 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2988 inode_unlock(inode);
2993 static const struct file_operations fuse_file_operations = {
2994 .llseek = fuse_file_llseek,
2995 .read_iter = fuse_file_read_iter,
2996 .write_iter = fuse_file_write_iter,
2997 .mmap = fuse_file_mmap,
2999 .flush = fuse_flush,
3000 .release = fuse_release,
3001 .fsync = fuse_fsync,
3002 .lock = fuse_file_lock,
3003 .flock = fuse_file_flock,
3004 .splice_read = generic_file_splice_read,
3005 .unlocked_ioctl = fuse_file_ioctl,
3006 .compat_ioctl = fuse_file_compat_ioctl,
3007 .poll = fuse_file_poll,
3008 .fallocate = fuse_file_fallocate,
3011 static const struct file_operations fuse_direct_io_file_operations = {
3012 .llseek = fuse_file_llseek,
3013 .read_iter = fuse_direct_read_iter,
3014 .write_iter = fuse_direct_write_iter,
3015 .mmap = fuse_direct_mmap,
3017 .flush = fuse_flush,
3018 .release = fuse_release,
3019 .fsync = fuse_fsync,
3020 .lock = fuse_file_lock,
3021 .flock = fuse_file_flock,
3022 .unlocked_ioctl = fuse_file_ioctl,
3023 .compat_ioctl = fuse_file_compat_ioctl,
3024 .poll = fuse_file_poll,
3025 .fallocate = fuse_file_fallocate,
3026 /* no splice_read */
3029 static const struct address_space_operations fuse_file_aops = {
3030 .readpage = fuse_readpage,
3031 .writepage = fuse_writepage,
3032 .writepages = fuse_writepages,
3033 .launder_page = fuse_launder_page,
3034 .readpages = fuse_readpages,
3035 .set_page_dirty = __set_page_dirty_nobuffers,
3037 .direct_IO = fuse_direct_IO,
3038 .write_begin = fuse_write_begin,
3039 .write_end = fuse_write_end,
3042 void fuse_init_file_inode(struct inode *inode)
3044 inode->i_fop = &fuse_file_operations;
3045 inode->i_data.a_ops = &fuse_file_aops;