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>
22 static const struct file_operations fuse_direct_io_file_operations;
24 static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
25 int opcode, struct fuse_open_out *outargp)
27 struct fuse_open_in inarg;
30 memset(&inarg, 0, sizeof(inarg));
31 inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
32 if (!fc->atomic_o_trunc)
33 inarg.flags &= ~O_TRUNC;
34 args.in.h.opcode = opcode;
35 args.in.h.nodeid = nodeid;
37 args.in.args[0].size = sizeof(inarg);
38 args.in.args[0].value = &inarg;
40 args.out.args[0].size = sizeof(*outargp);
41 args.out.args[0].value = outargp;
43 return fuse_simple_request(fc, &args);
46 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
50 ff = kzalloc(sizeof(struct fuse_file), GFP_KERNEL);
55 ff->reserved_req = fuse_request_alloc(0);
56 if (unlikely(!ff->reserved_req)) {
61 INIT_LIST_HEAD(&ff->write_entry);
62 atomic_set(&ff->count, 0);
63 RB_CLEAR_NODE(&ff->polled_node);
64 init_waitqueue_head(&ff->poll_wait);
68 spin_unlock(&fc->lock);
73 void fuse_file_free(struct fuse_file *ff)
75 fuse_request_free(ff->reserved_req);
79 struct fuse_file *fuse_file_get(struct fuse_file *ff)
81 atomic_inc(&ff->count);
85 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
87 iput(req->misc.release.inode);
90 static void fuse_file_put(struct fuse_file *ff, bool sync)
92 if (atomic_dec_and_test(&ff->count)) {
93 struct fuse_req *req = ff->reserved_req;
95 if (ff->fc->no_open) {
97 * Drop the release request when client does not
100 __clear_bit(FR_BACKGROUND, &req->flags);
101 iput(req->misc.release.inode);
102 fuse_put_request(ff->fc, req);
104 __set_bit(FR_FORCE, &req->flags);
105 __clear_bit(FR_BACKGROUND, &req->flags);
106 fuse_request_send(ff->fc, req);
107 iput(req->misc.release.inode);
108 fuse_put_request(ff->fc, req);
110 req->end = fuse_release_end;
111 __set_bit(FR_BACKGROUND, &req->flags);
112 fuse_request_send_background(ff->fc, req);
118 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
121 struct fuse_file *ff;
122 int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
124 ff = fuse_file_alloc(fc);
129 ff->open_flags = FOPEN_KEEP_CACHE; /* Default for no-open */
130 if (!fc->no_open || isdir) {
131 struct fuse_open_out outarg;
134 err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
137 ff->open_flags = outarg.open_flags;
139 } else if (err != -ENOSYS || isdir) {
148 ff->open_flags &= ~FOPEN_DIRECT_IO;
151 file->private_data = fuse_file_get(ff);
155 EXPORT_SYMBOL_GPL(fuse_do_open);
157 static void fuse_link_write_file(struct file *file)
159 struct inode *inode = file_inode(file);
160 struct fuse_conn *fc = get_fuse_conn(inode);
161 struct fuse_inode *fi = get_fuse_inode(inode);
162 struct fuse_file *ff = file->private_data;
164 * file may be written through mmap, so chain it onto the
165 * inodes's write_file list
167 spin_lock(&fc->lock);
168 if (list_empty(&ff->write_entry))
169 list_add(&ff->write_entry, &fi->write_files);
170 spin_unlock(&fc->lock);
173 void fuse_finish_open(struct inode *inode, struct file *file)
175 struct fuse_file *ff = file->private_data;
176 struct fuse_conn *fc = get_fuse_conn(inode);
178 if (ff->open_flags & FOPEN_DIRECT_IO)
179 file->f_op = &fuse_direct_io_file_operations;
180 if (!(ff->open_flags & FOPEN_KEEP_CACHE))
181 invalidate_inode_pages2(inode->i_mapping);
182 if (ff->open_flags & FOPEN_STREAM)
183 stream_open(inode, file);
184 else if (ff->open_flags & FOPEN_NONSEEKABLE)
185 nonseekable_open(inode, file);
186 if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
187 struct fuse_inode *fi = get_fuse_inode(inode);
189 spin_lock(&fc->lock);
190 fi->attr_version = ++fc->attr_version;
191 i_size_write(inode, 0);
192 spin_unlock(&fc->lock);
193 fuse_invalidate_attr(inode);
194 if (fc->writeback_cache)
195 file_update_time(file);
197 if ((file->f_mode & FMODE_WRITE) && fc->writeback_cache)
198 fuse_link_write_file(file);
201 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
203 struct fuse_conn *fc = get_fuse_conn(inode);
205 bool is_wb_truncate = (file->f_flags & O_TRUNC) &&
206 fc->atomic_o_trunc &&
209 if (fuse_is_bad(inode))
212 err = generic_file_open(inode, file);
216 if (is_wb_truncate) {
218 fuse_set_nowrite(inode);
221 err = fuse_do_open(fc, get_node_id(inode), file, isdir);
224 fuse_finish_open(inode, file);
226 if (is_wb_truncate) {
227 fuse_release_nowrite(inode);
234 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
236 struct fuse_conn *fc = ff->fc;
237 struct fuse_req *req = ff->reserved_req;
238 struct fuse_release_in *inarg = &req->misc.release.in;
240 spin_lock(&fc->lock);
241 list_del(&ff->write_entry);
242 if (!RB_EMPTY_NODE(&ff->polled_node))
243 rb_erase(&ff->polled_node, &fc->polled_files);
244 spin_unlock(&fc->lock);
246 wake_up_interruptible_all(&ff->poll_wait);
249 inarg->flags = flags;
250 req->in.h.opcode = opcode;
251 req->in.h.nodeid = ff->nodeid;
253 req->in.args[0].size = sizeof(struct fuse_release_in);
254 req->in.args[0].value = inarg;
257 void fuse_release_common(struct file *file, int opcode)
259 struct fuse_file *ff;
260 struct fuse_req *req;
262 ff = file->private_data;
266 req = ff->reserved_req;
267 fuse_prepare_release(ff, file->f_flags, opcode);
270 struct fuse_release_in *inarg = &req->misc.release.in;
271 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
272 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
275 /* Hold inode until release is finished */
276 req->misc.release.inode = igrab(file_inode(file));
279 * Normally this will send the RELEASE request, however if
280 * some asynchronous READ or WRITE requests are outstanding,
281 * the sending will be delayed.
283 * Make the release synchronous if this is a fuseblk mount,
284 * synchronous RELEASE is allowed (and desirable) in this case
285 * because the server can be trusted not to screw up.
287 fuse_file_put(ff, ff->fc->destroy_req != NULL);
290 static int fuse_open(struct inode *inode, struct file *file)
292 return fuse_open_common(inode, file, false);
295 static int fuse_release(struct inode *inode, struct file *file)
297 struct fuse_conn *fc = get_fuse_conn(inode);
299 /* see fuse_vma_close() for !writeback_cache case */
300 if (fc->writeback_cache)
301 write_inode_now(inode, 1);
303 fuse_release_common(file, FUSE_RELEASE);
305 /* return value is ignored by VFS */
309 void fuse_sync_release(struct fuse_file *ff, int flags)
311 WARN_ON(atomic_read(&ff->count) > 1);
312 fuse_prepare_release(ff, flags, FUSE_RELEASE);
313 __set_bit(FR_FORCE, &ff->reserved_req->flags);
314 __clear_bit(FR_BACKGROUND, &ff->reserved_req->flags);
315 fuse_request_send(ff->fc, ff->reserved_req);
316 fuse_put_request(ff->fc, ff->reserved_req);
319 EXPORT_SYMBOL_GPL(fuse_sync_release);
322 * Scramble the ID space with XTEA, so that the value of the files_struct
323 * pointer is not exposed to userspace.
325 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
327 u32 *k = fc->scramble_key;
328 u64 v = (unsigned long) id;
334 for (i = 0; i < 32; i++) {
335 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
337 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
340 return (u64) v0 + ((u64) v1 << 32);
344 * Check if any page in a range is under writeback
346 * This is currently done by walking the list of writepage requests
347 * for the inode, which can be pretty inefficient.
349 static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
352 struct fuse_conn *fc = get_fuse_conn(inode);
353 struct fuse_inode *fi = get_fuse_inode(inode);
354 struct fuse_req *req;
357 spin_lock(&fc->lock);
358 list_for_each_entry(req, &fi->writepages, writepages_entry) {
361 BUG_ON(req->inode != inode);
362 curr_index = req->misc.write.in.offset >> PAGE_SHIFT;
363 if (idx_from < curr_index + req->num_pages &&
364 curr_index <= idx_to) {
369 spin_unlock(&fc->lock);
374 static inline bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
376 return fuse_range_is_writeback(inode, index, index);
380 * Wait for page writeback to be completed.
382 * Since fuse doesn't rely on the VM writeback tracking, this has to
383 * use some other means.
385 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
387 struct fuse_inode *fi = get_fuse_inode(inode);
389 wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
394 * Wait for all pending writepages on the inode to finish.
396 * This is currently done by blocking further writes with FUSE_NOWRITE
397 * and waiting for all sent writes to complete.
399 * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
400 * could conflict with truncation.
402 static void fuse_sync_writes(struct inode *inode)
404 fuse_set_nowrite(inode);
405 fuse_release_nowrite(inode);
408 static int fuse_flush(struct file *file, fl_owner_t id)
410 struct inode *inode = file_inode(file);
411 struct fuse_conn *fc = get_fuse_conn(inode);
412 struct fuse_file *ff = file->private_data;
413 struct fuse_req *req;
414 struct fuse_flush_in inarg;
417 if (fuse_is_bad(inode))
423 err = write_inode_now(inode, 1);
428 fuse_sync_writes(inode);
431 err = filemap_check_errors(file->f_mapping);
435 req = fuse_get_req_nofail_nopages(fc, file);
436 memset(&inarg, 0, sizeof(inarg));
438 inarg.lock_owner = fuse_lock_owner_id(fc, id);
439 req->in.h.opcode = FUSE_FLUSH;
440 req->in.h.nodeid = get_node_id(inode);
442 req->in.args[0].size = sizeof(inarg);
443 req->in.args[0].value = &inarg;
444 __set_bit(FR_FORCE, &req->flags);
445 fuse_request_send(fc, req);
446 err = req->out.h.error;
447 fuse_put_request(fc, req);
448 if (err == -ENOSYS) {
455 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
456 int datasync, int isdir)
458 struct inode *inode = file->f_mapping->host;
459 struct fuse_conn *fc = get_fuse_conn(inode);
460 struct fuse_file *ff = file->private_data;
462 struct fuse_fsync_in inarg;
465 if (fuse_is_bad(inode))
471 * Start writeback against all dirty pages of the inode, then
472 * wait for all outstanding writes, before sending the FSYNC
475 err = filemap_write_and_wait_range(inode->i_mapping, start, end);
479 fuse_sync_writes(inode);
482 * Due to implementation of fuse writeback
483 * filemap_write_and_wait_range() does not catch errors.
484 * We have to do this directly after fuse_sync_writes()
486 err = filemap_check_errors(file->f_mapping);
490 err = sync_inode_metadata(inode, 1);
494 if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
497 memset(&inarg, 0, sizeof(inarg));
499 inarg.fsync_flags = datasync ? 1 : 0;
500 args.in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
501 args.in.h.nodeid = get_node_id(inode);
503 args.in.args[0].size = sizeof(inarg);
504 args.in.args[0].value = &inarg;
505 err = fuse_simple_request(fc, &args);
506 if (err == -ENOSYS) {
518 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
521 return fuse_fsync_common(file, start, end, datasync, 0);
524 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
525 size_t count, int opcode)
527 struct fuse_read_in *inarg = &req->misc.read.in;
528 struct fuse_file *ff = file->private_data;
533 inarg->flags = file->f_flags;
534 req->in.h.opcode = opcode;
535 req->in.h.nodeid = ff->nodeid;
537 req->in.args[0].size = sizeof(struct fuse_read_in);
538 req->in.args[0].value = inarg;
540 req->out.numargs = 1;
541 req->out.args[0].size = count;
544 static void fuse_release_user_pages(struct fuse_req *req, bool should_dirty)
548 for (i = 0; i < req->num_pages; i++) {
549 struct page *page = req->pages[i];
551 set_page_dirty_lock(page);
556 static void fuse_io_release(struct kref *kref)
558 kfree(container_of(kref, struct fuse_io_priv, refcnt));
561 static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
566 if (io->bytes >= 0 && io->write)
569 return io->bytes < 0 ? io->size : io->bytes;
573 * In case of short read, the caller sets 'pos' to the position of
574 * actual end of fuse request in IO request. Otherwise, if bytes_requested
575 * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
578 * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
579 * both submitted asynchronously. The first of them was ACKed by userspace as
580 * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
581 * second request was ACKed as short, e.g. only 1K was read, resulting in
584 * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
585 * will be equal to the length of the longest contiguous fragment of
586 * transferred data starting from the beginning of IO request.
588 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
592 spin_lock(&io->lock);
594 io->err = io->err ? : err;
595 else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
599 if (!left && io->blocking)
601 spin_unlock(&io->lock);
603 if (!left && !io->blocking) {
604 ssize_t res = fuse_get_res_by_io(io);
607 struct inode *inode = file_inode(io->iocb->ki_filp);
608 struct fuse_conn *fc = get_fuse_conn(inode);
609 struct fuse_inode *fi = get_fuse_inode(inode);
611 spin_lock(&fc->lock);
612 fi->attr_version = ++fc->attr_version;
613 spin_unlock(&fc->lock);
616 io->iocb->ki_complete(io->iocb, res, 0);
619 kref_put(&io->refcnt, fuse_io_release);
622 static void fuse_aio_complete_req(struct fuse_conn *fc, struct fuse_req *req)
624 struct fuse_io_priv *io = req->io;
627 fuse_release_user_pages(req, !io->write);
630 if (req->misc.write.in.size != req->misc.write.out.size)
631 pos = req->misc.write.in.offset - io->offset +
632 req->misc.write.out.size;
634 if (req->misc.read.in.size != req->out.args[0].size)
635 pos = req->misc.read.in.offset - io->offset +
636 req->out.args[0].size;
639 fuse_aio_complete(io, req->out.h.error, pos);
642 static size_t fuse_async_req_send(struct fuse_conn *fc, struct fuse_req *req,
643 size_t num_bytes, struct fuse_io_priv *io)
645 spin_lock(&io->lock);
646 kref_get(&io->refcnt);
647 io->size += num_bytes;
649 spin_unlock(&io->lock);
652 req->end = fuse_aio_complete_req;
654 __fuse_get_request(req);
655 fuse_request_send_background(fc, req);
660 static size_t fuse_send_read(struct fuse_req *req, struct fuse_io_priv *io,
661 loff_t pos, size_t count, fl_owner_t owner)
663 struct file *file = io->file;
664 struct fuse_file *ff = file->private_data;
665 struct fuse_conn *fc = ff->fc;
667 fuse_read_fill(req, file, pos, count, FUSE_READ);
669 struct fuse_read_in *inarg = &req->misc.read.in;
671 inarg->read_flags |= FUSE_READ_LOCKOWNER;
672 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
676 return fuse_async_req_send(fc, req, count, io);
678 fuse_request_send(fc, req);
679 return req->out.args[0].size;
682 static void fuse_read_update_size(struct inode *inode, loff_t size,
685 struct fuse_conn *fc = get_fuse_conn(inode);
686 struct fuse_inode *fi = get_fuse_inode(inode);
688 spin_lock(&fc->lock);
689 if (attr_ver == fi->attr_version && size < inode->i_size &&
690 !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
691 fi->attr_version = ++fc->attr_version;
692 i_size_write(inode, size);
694 spin_unlock(&fc->lock);
697 static void fuse_short_read(struct fuse_req *req, struct inode *inode,
700 size_t num_read = req->out.args[0].size;
701 struct fuse_conn *fc = get_fuse_conn(inode);
703 if (fc->writeback_cache) {
705 * A hole in a file. Some data after the hole are in page cache,
706 * but have not reached the client fs yet. So, the hole is not
710 int start_idx = num_read >> PAGE_SHIFT;
711 size_t off = num_read & (PAGE_SIZE - 1);
713 for (i = start_idx; i < req->num_pages; i++) {
714 zero_user_segment(req->pages[i], off, PAGE_SIZE);
718 loff_t pos = page_offset(req->pages[0]) + num_read;
719 fuse_read_update_size(inode, pos, attr_ver);
723 static int fuse_do_readpage(struct file *file, struct page *page)
725 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
726 struct inode *inode = page->mapping->host;
727 struct fuse_conn *fc = get_fuse_conn(inode);
728 struct fuse_req *req;
730 loff_t pos = page_offset(page);
731 size_t count = PAGE_SIZE;
736 * Page writeback can extend beyond the lifetime of the
737 * page-cache page, so make sure we read a properly synced
740 fuse_wait_on_page_writeback(inode, page->index);
742 req = fuse_get_req(fc, 1);
746 attr_ver = fuse_get_attr_version(fc);
748 req->out.page_zeroing = 1;
749 req->out.argpages = 1;
751 req->pages[0] = page;
752 req->page_descs[0].length = count;
753 num_read = fuse_send_read(req, &io, pos, count, NULL);
754 err = req->out.h.error;
758 * Short read means EOF. If file size is larger, truncate it
760 if (num_read < count)
761 fuse_short_read(req, inode, attr_ver);
763 SetPageUptodate(page);
766 fuse_put_request(fc, req);
771 static int fuse_readpage(struct file *file, struct page *page)
773 struct inode *inode = page->mapping->host;
777 if (fuse_is_bad(inode))
780 err = fuse_do_readpage(file, page);
781 fuse_invalidate_atime(inode);
787 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
790 size_t count = req->misc.read.in.size;
791 size_t num_read = req->out.args[0].size;
792 struct address_space *mapping = NULL;
794 for (i = 0; mapping == NULL && i < req->num_pages; i++)
795 mapping = req->pages[i]->mapping;
798 struct inode *inode = mapping->host;
801 * Short read means EOF. If file size is larger, truncate it
803 if (!req->out.h.error && num_read < count)
804 fuse_short_read(req, inode, req->misc.read.attr_ver);
806 fuse_invalidate_atime(inode);
809 for (i = 0; i < req->num_pages; i++) {
810 struct page *page = req->pages[i];
811 if (!req->out.h.error)
812 SetPageUptodate(page);
819 fuse_file_put(req->ff, false);
822 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
824 struct fuse_file *ff = file->private_data;
825 struct fuse_conn *fc = ff->fc;
826 loff_t pos = page_offset(req->pages[0]);
827 size_t count = req->num_pages << PAGE_SHIFT;
829 req->out.argpages = 1;
830 req->out.page_zeroing = 1;
831 req->out.page_replace = 1;
832 fuse_read_fill(req, file, pos, count, FUSE_READ);
833 req->misc.read.attr_ver = fuse_get_attr_version(fc);
834 if (fc->async_read) {
835 req->ff = fuse_file_get(ff);
836 req->end = fuse_readpages_end;
837 fuse_request_send_background(fc, req);
839 fuse_request_send(fc, req);
840 fuse_readpages_end(fc, req);
841 fuse_put_request(fc, req);
845 struct fuse_fill_data {
846 struct fuse_req *req;
852 static int fuse_readpages_fill(void *_data, struct page *page)
854 struct fuse_fill_data *data = _data;
855 struct fuse_req *req = data->req;
856 struct inode *inode = data->inode;
857 struct fuse_conn *fc = get_fuse_conn(inode);
859 fuse_wait_on_page_writeback(inode, page->index);
861 if (req->num_pages &&
862 (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
863 (req->num_pages + 1) * PAGE_SIZE > fc->max_read ||
864 req->pages[req->num_pages - 1]->index + 1 != page->index)) {
865 int nr_alloc = min_t(unsigned, data->nr_pages,
866 FUSE_MAX_PAGES_PER_REQ);
867 fuse_send_readpages(req, data->file);
869 req = fuse_get_req_for_background(fc, nr_alloc);
871 req = fuse_get_req(fc, nr_alloc);
880 if (WARN_ON(req->num_pages >= req->max_pages)) {
882 fuse_put_request(fc, req);
887 req->pages[req->num_pages] = page;
888 req->page_descs[req->num_pages].length = PAGE_SIZE;
894 static int fuse_readpages(struct file *file, struct address_space *mapping,
895 struct list_head *pages, unsigned nr_pages)
897 struct inode *inode = mapping->host;
898 struct fuse_conn *fc = get_fuse_conn(inode);
899 struct fuse_fill_data data;
901 int nr_alloc = min_t(unsigned, nr_pages, FUSE_MAX_PAGES_PER_REQ);
904 if (fuse_is_bad(inode))
910 data.req = fuse_get_req_for_background(fc, nr_alloc);
912 data.req = fuse_get_req(fc, nr_alloc);
913 data.nr_pages = nr_pages;
914 err = PTR_ERR(data.req);
915 if (IS_ERR(data.req))
918 err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
920 if (data.req->num_pages)
921 fuse_send_readpages(data.req, file);
923 fuse_put_request(fc, data.req);
929 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
931 struct inode *inode = iocb->ki_filp->f_mapping->host;
932 struct fuse_conn *fc = get_fuse_conn(inode);
934 if (fuse_is_bad(inode))
938 * In auto invalidate mode, always update attributes on read.
939 * Otherwise, only update if we attempt to read past EOF (to ensure
940 * i_size is up to date).
942 if (fc->auto_inval_data ||
943 (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
945 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
950 return generic_file_read_iter(iocb, to);
953 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
954 loff_t pos, size_t count)
956 struct fuse_write_in *inarg = &req->misc.write.in;
957 struct fuse_write_out *outarg = &req->misc.write.out;
962 req->in.h.opcode = FUSE_WRITE;
963 req->in.h.nodeid = ff->nodeid;
965 if (ff->fc->minor < 9)
966 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
968 req->in.args[0].size = sizeof(struct fuse_write_in);
969 req->in.args[0].value = inarg;
970 req->in.args[1].size = count;
971 req->out.numargs = 1;
972 req->out.args[0].size = sizeof(struct fuse_write_out);
973 req->out.args[0].value = outarg;
976 static size_t fuse_send_write(struct fuse_req *req, struct fuse_io_priv *io,
977 loff_t pos, size_t count, fl_owner_t owner)
979 struct file *file = io->file;
980 struct fuse_file *ff = file->private_data;
981 struct fuse_conn *fc = ff->fc;
982 struct fuse_write_in *inarg = &req->misc.write.in;
984 fuse_write_fill(req, ff, pos, count);
985 inarg->flags = file->f_flags;
987 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
988 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
992 return fuse_async_req_send(fc, req, count, io);
994 fuse_request_send(fc, req);
995 return req->misc.write.out.size;
998 bool fuse_write_update_size(struct inode *inode, loff_t pos)
1000 struct fuse_conn *fc = get_fuse_conn(inode);
1001 struct fuse_inode *fi = get_fuse_inode(inode);
1004 spin_lock(&fc->lock);
1005 fi->attr_version = ++fc->attr_version;
1006 if (pos > inode->i_size) {
1007 i_size_write(inode, pos);
1010 spin_unlock(&fc->lock);
1015 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
1016 struct inode *inode, loff_t pos,
1022 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
1024 for (i = 0; i < req->num_pages; i++)
1025 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
1027 res = fuse_send_write(req, &io, pos, count, NULL);
1029 offset = req->page_descs[0].offset;
1031 for (i = 0; i < req->num_pages; i++) {
1032 struct page *page = req->pages[i];
1034 if (!req->out.h.error && !offset && count >= PAGE_SIZE)
1035 SetPageUptodate(page);
1037 if (count > PAGE_SIZE - offset)
1038 count -= PAGE_SIZE - offset;
1050 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
1051 struct address_space *mapping,
1052 struct iov_iter *ii, loff_t pos)
1054 struct fuse_conn *fc = get_fuse_conn(mapping->host);
1055 unsigned offset = pos & (PAGE_SIZE - 1);
1059 req->in.argpages = 1;
1060 req->page_descs[0].offset = offset;
1065 pgoff_t index = pos >> PAGE_SHIFT;
1066 size_t bytes = min_t(size_t, PAGE_SIZE - offset,
1067 iov_iter_count(ii));
1069 bytes = min_t(size_t, bytes, fc->max_write - count);
1073 if (iov_iter_fault_in_readable(ii, bytes))
1077 page = grab_cache_page_write_begin(mapping, index, 0);
1081 if (mapping_writably_mapped(mapping))
1082 flush_dcache_page(page);
1084 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
1085 flush_dcache_page(page);
1087 iov_iter_advance(ii, tmp);
1091 bytes = min(bytes, iov_iter_single_seg_count(ii));
1096 req->pages[req->num_pages] = page;
1097 req->page_descs[req->num_pages].length = tmp;
1103 if (offset == PAGE_SIZE)
1106 if (!fc->big_writes)
1108 } while (iov_iter_count(ii) && count < fc->max_write &&
1109 req->num_pages < req->max_pages && offset == 0);
1111 return count > 0 ? count : err;
1114 static inline unsigned fuse_wr_pages(loff_t pos, size_t len)
1116 return min_t(unsigned,
1117 ((pos + len - 1) >> PAGE_SHIFT) -
1118 (pos >> PAGE_SHIFT) + 1,
1119 FUSE_MAX_PAGES_PER_REQ);
1122 static ssize_t fuse_perform_write(struct file *file,
1123 struct address_space *mapping,
1124 struct iov_iter *ii, loff_t pos)
1126 struct inode *inode = mapping->host;
1127 struct fuse_conn *fc = get_fuse_conn(inode);
1128 struct fuse_inode *fi = get_fuse_inode(inode);
1132 if (fuse_is_bad(inode))
1135 if (inode->i_size < pos + iov_iter_count(ii))
1136 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1139 struct fuse_req *req;
1141 unsigned nr_pages = fuse_wr_pages(pos, iov_iter_count(ii));
1143 req = fuse_get_req(fc, nr_pages);
1149 count = fuse_fill_write_pages(req, mapping, ii, pos);
1155 num_written = fuse_send_write_pages(req, file, inode,
1157 err = req->out.h.error;
1162 /* break out of the loop on short write */
1163 if (num_written != count)
1167 fuse_put_request(fc, req);
1168 } while (!err && iov_iter_count(ii));
1171 fuse_write_update_size(inode, pos);
1173 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1174 fuse_invalidate_attr(inode);
1176 return res > 0 ? res : err;
1179 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1181 struct file *file = iocb->ki_filp;
1182 struct address_space *mapping = file->f_mapping;
1183 ssize_t written = 0;
1184 ssize_t written_buffered = 0;
1185 struct inode *inode = mapping->host;
1189 if (fuse_is_bad(inode))
1192 if (get_fuse_conn(inode)->writeback_cache) {
1193 /* Update size (EOF optimization) and mode (SUID clearing) */
1194 err = fuse_update_attributes(mapping->host, NULL, file, NULL);
1198 return generic_file_write_iter(iocb, from);
1203 /* We can write back this queue in page reclaim */
1204 current->backing_dev_info = inode_to_bdi(inode);
1206 err = generic_write_checks(iocb, from);
1210 err = file_remove_privs(file);
1214 err = file_update_time(file);
1218 if (iocb->ki_flags & IOCB_DIRECT) {
1219 loff_t pos = iocb->ki_pos;
1220 written = generic_file_direct_write(iocb, from);
1221 if (written < 0 || !iov_iter_count(from))
1226 written_buffered = fuse_perform_write(file, mapping, from, pos);
1227 if (written_buffered < 0) {
1228 err = written_buffered;
1231 endbyte = pos + written_buffered - 1;
1233 err = filemap_write_and_wait_range(file->f_mapping, pos,
1238 invalidate_mapping_pages(file->f_mapping,
1240 endbyte >> PAGE_SHIFT);
1242 written += written_buffered;
1243 iocb->ki_pos = pos + written_buffered;
1245 written = fuse_perform_write(file, mapping, from, iocb->ki_pos);
1247 iocb->ki_pos += written;
1250 current->backing_dev_info = NULL;
1251 inode_unlock(inode);
1253 return written ? written : err;
1256 static inline void fuse_page_descs_length_init(struct fuse_req *req,
1257 unsigned index, unsigned nr_pages)
1261 for (i = index; i < index + nr_pages; i++)
1262 req->page_descs[i].length = PAGE_SIZE -
1263 req->page_descs[i].offset;
1266 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1268 return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1271 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1274 return min(iov_iter_single_seg_count(ii), max_size);
1277 static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1278 size_t *nbytesp, int write)
1280 size_t nbytes = 0; /* # bytes already packed in req */
1283 /* Special case for kernel I/O: can copy directly into the buffer */
1284 if (ii->type & ITER_KVEC) {
1285 unsigned long user_addr = fuse_get_user_addr(ii);
1286 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1289 req->in.args[1].value = (void *) user_addr;
1291 req->out.args[0].value = (void *) user_addr;
1293 iov_iter_advance(ii, frag_size);
1294 *nbytesp = frag_size;
1298 while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1301 ret = iov_iter_get_pages(ii, &req->pages[req->num_pages],
1303 req->max_pages - req->num_pages,
1308 iov_iter_advance(ii, ret);
1312 npages = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
1314 req->page_descs[req->num_pages].offset = start;
1315 fuse_page_descs_length_init(req, req->num_pages, npages);
1317 req->num_pages += npages;
1318 req->page_descs[req->num_pages - 1].length -=
1319 (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1323 req->in.argpages = 1;
1325 req->out.argpages = 1;
1329 return ret < 0 ? ret : 0;
1332 static inline int fuse_iter_npages(const struct iov_iter *ii_p)
1334 return iov_iter_npages(ii_p, FUSE_MAX_PAGES_PER_REQ);
1337 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1338 loff_t *ppos, int flags)
1340 int write = flags & FUSE_DIO_WRITE;
1341 bool should_dirty = !write && iter_is_iovec(iter);
1342 int cuse = flags & FUSE_DIO_CUSE;
1343 struct file *file = io->file;
1344 struct inode *inode = file->f_mapping->host;
1345 struct fuse_file *ff = file->private_data;
1346 struct fuse_conn *fc = ff->fc;
1347 size_t nmax = write ? fc->max_write : fc->max_read;
1349 size_t count = iov_iter_count(iter);
1350 pgoff_t idx_from = pos >> PAGE_SHIFT;
1351 pgoff_t idx_to = (pos + count - 1) >> PAGE_SHIFT;
1353 struct fuse_req *req;
1357 req = fuse_get_req_for_background(fc, fuse_iter_npages(iter));
1359 req = fuse_get_req(fc, fuse_iter_npages(iter));
1361 return PTR_ERR(req);
1363 if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1366 fuse_sync_writes(inode);
1368 inode_unlock(inode);
1373 fl_owner_t owner = current->files;
1374 size_t nbytes = min(count, nmax);
1375 err = fuse_get_user_pages(req, iter, &nbytes, write);
1380 nres = fuse_send_write(req, io, pos, nbytes, owner);
1382 nres = fuse_send_read(req, io, pos, nbytes, owner);
1385 fuse_release_user_pages(req, should_dirty);
1386 if (req->out.h.error) {
1387 err = req->out.h.error;
1389 } else if (nres > nbytes) {
1400 fuse_put_request(fc, req);
1402 req = fuse_get_req_for_background(fc,
1403 fuse_iter_npages(iter));
1405 req = fuse_get_req(fc, fuse_iter_npages(iter));
1411 fuse_put_request(fc, req);
1415 return res > 0 ? res : err;
1417 EXPORT_SYMBOL_GPL(fuse_direct_io);
1419 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1420 struct iov_iter *iter,
1424 struct file *file = io->file;
1425 struct inode *inode = file_inode(file);
1427 if (fuse_is_bad(inode))
1430 res = fuse_direct_io(io, iter, ppos, 0);
1432 fuse_invalidate_attr(inode);
1437 static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1439 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb->ki_filp);
1440 return __fuse_direct_read(&io, to, &iocb->ki_pos);
1443 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1445 struct file *file = iocb->ki_filp;
1446 struct inode *inode = file_inode(file);
1447 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
1450 if (fuse_is_bad(inode))
1453 /* Don't allow parallel writes to the same file */
1455 res = generic_write_checks(iocb, from);
1457 res = fuse_direct_io(&io, from, &iocb->ki_pos, FUSE_DIO_WRITE);
1458 fuse_invalidate_attr(inode);
1460 fuse_write_update_size(inode, iocb->ki_pos);
1461 inode_unlock(inode);
1466 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1470 for (i = 0; i < req->num_pages; i++)
1471 __free_page(req->pages[i]);
1474 fuse_file_put(req->ff, false);
1477 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1479 struct inode *inode = req->inode;
1480 struct fuse_inode *fi = get_fuse_inode(inode);
1481 struct backing_dev_info *bdi = inode_to_bdi(inode);
1484 list_del(&req->writepages_entry);
1485 for (i = 0; i < req->num_pages; i++) {
1486 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1487 dec_node_page_state(req->pages[i], NR_WRITEBACK_TEMP);
1488 wb_writeout_inc(&bdi->wb);
1490 wake_up(&fi->page_waitq);
1493 /* Called under fc->lock, may release and reacquire it */
1494 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req,
1496 __releases(fc->lock)
1497 __acquires(fc->lock)
1499 struct fuse_inode *fi = get_fuse_inode(req->inode);
1500 struct fuse_write_in *inarg = &req->misc.write.in;
1501 __u64 data_size = req->num_pages * PAGE_SIZE;
1506 if (inarg->offset + data_size <= size) {
1507 inarg->size = data_size;
1508 } else if (inarg->offset < size) {
1509 inarg->size = size - inarg->offset;
1511 /* Got truncated off completely */
1515 req->in.args[1].size = inarg->size;
1517 fuse_request_send_background_locked(fc, req);
1521 fuse_writepage_finish(fc, req);
1522 spin_unlock(&fc->lock);
1523 fuse_writepage_free(fc, req);
1524 fuse_put_request(fc, req);
1525 spin_lock(&fc->lock);
1529 * If fi->writectr is positive (no truncate or fsync going on) send
1530 * all queued writepage requests.
1532 * Called with fc->lock
1534 void fuse_flush_writepages(struct inode *inode)
1535 __releases(fc->lock)
1536 __acquires(fc->lock)
1538 struct fuse_conn *fc = get_fuse_conn(inode);
1539 struct fuse_inode *fi = get_fuse_inode(inode);
1540 loff_t crop = i_size_read(inode);
1541 struct fuse_req *req;
1543 while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1544 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1545 list_del_init(&req->list);
1546 fuse_send_writepage(fc, req, crop);
1550 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1552 struct inode *inode = req->inode;
1553 struct fuse_inode *fi = get_fuse_inode(inode);
1555 mapping_set_error(inode->i_mapping, req->out.h.error);
1556 spin_lock(&fc->lock);
1557 while (req->misc.write.next) {
1558 struct fuse_conn *fc = get_fuse_conn(inode);
1559 struct fuse_write_in *inarg = &req->misc.write.in;
1560 struct fuse_req *next = req->misc.write.next;
1561 req->misc.write.next = next->misc.write.next;
1562 next->misc.write.next = NULL;
1563 next->ff = fuse_file_get(req->ff);
1564 list_add(&next->writepages_entry, &fi->writepages);
1567 * Skip fuse_flush_writepages() to make it easy to crop requests
1568 * based on primary request size.
1570 * 1st case (trivial): there are no concurrent activities using
1571 * fuse_set/release_nowrite. Then we're on safe side because
1572 * fuse_flush_writepages() would call fuse_send_writepage()
1575 * 2nd case: someone called fuse_set_nowrite and it is waiting
1576 * now for completion of all in-flight requests. This happens
1577 * rarely and no more than once per page, so this should be
1580 * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1581 * of fuse_set_nowrite..fuse_release_nowrite section. The fact
1582 * that fuse_set_nowrite returned implies that all in-flight
1583 * requests were completed along with all of their secondary
1584 * requests. Further primary requests are blocked by negative
1585 * writectr. Hence there cannot be any in-flight requests and
1586 * no invocations of fuse_writepage_end() while we're in
1587 * fuse_set_nowrite..fuse_release_nowrite section.
1589 fuse_send_writepage(fc, next, inarg->offset + inarg->size);
1592 fuse_writepage_finish(fc, req);
1593 spin_unlock(&fc->lock);
1594 fuse_writepage_free(fc, req);
1597 static struct fuse_file *__fuse_write_file_get(struct fuse_conn *fc,
1598 struct fuse_inode *fi)
1600 struct fuse_file *ff = NULL;
1602 spin_lock(&fc->lock);
1603 if (!list_empty(&fi->write_files)) {
1604 ff = list_entry(fi->write_files.next, struct fuse_file,
1608 spin_unlock(&fc->lock);
1613 static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1614 struct fuse_inode *fi)
1616 struct fuse_file *ff = __fuse_write_file_get(fc, fi);
1621 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1623 struct fuse_conn *fc = get_fuse_conn(inode);
1624 struct fuse_inode *fi = get_fuse_inode(inode);
1625 struct fuse_file *ff;
1628 ff = __fuse_write_file_get(fc, fi);
1629 err = fuse_flush_times(inode, ff);
1631 fuse_file_put(ff, 0);
1636 static int fuse_writepage_locked(struct page *page)
1638 struct address_space *mapping = page->mapping;
1639 struct inode *inode = mapping->host;
1640 struct fuse_conn *fc = get_fuse_conn(inode);
1641 struct fuse_inode *fi = get_fuse_inode(inode);
1642 struct fuse_req *req;
1643 struct page *tmp_page;
1644 int error = -ENOMEM;
1646 set_page_writeback(page);
1648 req = fuse_request_alloc_nofs(1);
1652 /* writeback always goes to bg_queue */
1653 __set_bit(FR_BACKGROUND, &req->flags);
1654 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1659 req->ff = fuse_write_file_get(fc, fi);
1663 fuse_write_fill(req, req->ff, page_offset(page), 0);
1665 copy_highpage(tmp_page, page);
1666 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1667 req->misc.write.next = NULL;
1668 req->in.argpages = 1;
1670 req->pages[0] = tmp_page;
1671 req->page_descs[0].offset = 0;
1672 req->page_descs[0].length = PAGE_SIZE;
1673 req->end = fuse_writepage_end;
1676 inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1677 inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1679 spin_lock(&fc->lock);
1680 list_add(&req->writepages_entry, &fi->writepages);
1681 list_add_tail(&req->list, &fi->queued_writes);
1682 fuse_flush_writepages(inode);
1683 spin_unlock(&fc->lock);
1685 end_page_writeback(page);
1690 __free_page(tmp_page);
1692 fuse_request_free(req);
1694 end_page_writeback(page);
1698 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1702 if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1704 * ->writepages() should be called for sync() and friends. We
1705 * should only get here on direct reclaim and then we are
1706 * allowed to skip a page which is already in flight
1708 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1710 redirty_page_for_writepage(wbc, page);
1715 err = fuse_writepage_locked(page);
1721 struct fuse_fill_wb_data {
1722 struct fuse_req *req;
1723 struct fuse_file *ff;
1724 struct inode *inode;
1725 struct page **orig_pages;
1728 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1730 struct fuse_req *req = data->req;
1731 struct inode *inode = data->inode;
1732 struct fuse_conn *fc = get_fuse_conn(inode);
1733 struct fuse_inode *fi = get_fuse_inode(inode);
1734 int num_pages = req->num_pages;
1737 req->ff = fuse_file_get(data->ff);
1738 spin_lock(&fc->lock);
1739 list_add_tail(&req->list, &fi->queued_writes);
1740 fuse_flush_writepages(inode);
1741 spin_unlock(&fc->lock);
1743 for (i = 0; i < num_pages; i++)
1744 end_page_writeback(data->orig_pages[i]);
1747 static bool fuse_writepage_in_flight(struct fuse_req *new_req,
1750 struct fuse_conn *fc = get_fuse_conn(new_req->inode);
1751 struct fuse_inode *fi = get_fuse_inode(new_req->inode);
1752 struct fuse_req *tmp;
1753 struct fuse_req *old_req;
1757 BUG_ON(new_req->num_pages != 0);
1759 spin_lock(&fc->lock);
1760 list_del(&new_req->writepages_entry);
1761 list_for_each_entry(old_req, &fi->writepages, writepages_entry) {
1762 BUG_ON(old_req->inode != new_req->inode);
1763 curr_index = old_req->misc.write.in.offset >> PAGE_SHIFT;
1764 if (curr_index <= page->index &&
1765 page->index < curr_index + old_req->num_pages) {
1771 list_add(&new_req->writepages_entry, &fi->writepages);
1775 new_req->num_pages = 1;
1776 for (tmp = old_req; tmp != NULL; tmp = tmp->misc.write.next) {
1777 BUG_ON(tmp->inode != new_req->inode);
1778 curr_index = tmp->misc.write.in.offset >> PAGE_SHIFT;
1779 if (tmp->num_pages == 1 &&
1780 curr_index == page->index) {
1785 if (old_req->num_pages == 1 && test_bit(FR_PENDING, &old_req->flags)) {
1786 struct backing_dev_info *bdi = inode_to_bdi(page->mapping->host);
1788 copy_highpage(old_req->pages[0], page);
1789 spin_unlock(&fc->lock);
1791 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1792 dec_node_page_state(new_req->pages[0], NR_WRITEBACK_TEMP);
1793 wb_writeout_inc(&bdi->wb);
1794 fuse_writepage_free(fc, new_req);
1795 fuse_request_free(new_req);
1798 new_req->misc.write.next = old_req->misc.write.next;
1799 old_req->misc.write.next = new_req;
1802 spin_unlock(&fc->lock);
1807 static int fuse_writepages_fill(struct page *page,
1808 struct writeback_control *wbc, void *_data)
1810 struct fuse_fill_wb_data *data = _data;
1811 struct fuse_req *req = data->req;
1812 struct inode *inode = data->inode;
1813 struct fuse_conn *fc = get_fuse_conn(inode);
1814 struct page *tmp_page;
1820 data->ff = fuse_write_file_get(fc, get_fuse_inode(inode));
1826 * Being under writeback is unlikely but possible. For example direct
1827 * read to an mmaped fuse file will set the page dirty twice; once when
1828 * the pages are faulted with get_user_pages(), and then after the read
1831 is_writeback = fuse_page_is_writeback(inode, page->index);
1833 if (req && req->num_pages &&
1834 (is_writeback || req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
1835 (req->num_pages + 1) * PAGE_SIZE > fc->max_write ||
1836 data->orig_pages[req->num_pages - 1]->index + 1 != page->index)) {
1837 fuse_writepages_send(data);
1841 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1846 * The page must not be redirtied until the writeout is completed
1847 * (i.e. userspace has sent a reply to the write request). Otherwise
1848 * there could be more than one temporary page instance for each real
1851 * This is ensured by holding the page lock in page_mkwrite() while
1852 * checking fuse_page_is_writeback(). We already hold the page lock
1853 * since clear_page_dirty_for_io() and keep it held until we add the
1854 * request to the fi->writepages list and increment req->num_pages.
1855 * After this fuse_page_is_writeback() will indicate that the page is
1856 * under writeback, so we can release the page lock.
1858 if (data->req == NULL) {
1859 struct fuse_inode *fi = get_fuse_inode(inode);
1862 req = fuse_request_alloc_nofs(FUSE_MAX_PAGES_PER_REQ);
1864 __free_page(tmp_page);
1868 fuse_write_fill(req, data->ff, page_offset(page), 0);
1869 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1870 req->misc.write.next = NULL;
1871 req->in.argpages = 1;
1872 __set_bit(FR_BACKGROUND, &req->flags);
1874 req->end = fuse_writepage_end;
1877 spin_lock(&fc->lock);
1878 list_add(&req->writepages_entry, &fi->writepages);
1879 spin_unlock(&fc->lock);
1883 set_page_writeback(page);
1885 copy_highpage(tmp_page, page);
1886 req->pages[req->num_pages] = tmp_page;
1887 req->page_descs[req->num_pages].offset = 0;
1888 req->page_descs[req->num_pages].length = PAGE_SIZE;
1890 inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1891 inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1894 if (is_writeback && fuse_writepage_in_flight(req, page)) {
1895 end_page_writeback(page);
1899 data->orig_pages[req->num_pages] = page;
1902 * Protected by fc->lock against concurrent access by
1903 * fuse_page_is_writeback().
1905 spin_lock(&fc->lock);
1907 spin_unlock(&fc->lock);
1915 static int fuse_writepages(struct address_space *mapping,
1916 struct writeback_control *wbc)
1918 struct inode *inode = mapping->host;
1919 struct fuse_fill_wb_data data;
1923 if (fuse_is_bad(inode))
1931 data.orig_pages = kcalloc(FUSE_MAX_PAGES_PER_REQ,
1932 sizeof(struct page *),
1934 if (!data.orig_pages)
1937 err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
1939 /* Ignore errors if we can write at least one page */
1940 BUG_ON(!data.req->num_pages);
1941 fuse_writepages_send(&data);
1945 fuse_file_put(data.ff, false);
1947 kfree(data.orig_pages);
1953 * It's worthy to make sure that space is reserved on disk for the write,
1954 * but how to implement it without killing performance need more thinking.
1956 static int fuse_write_begin(struct file *file, struct address_space *mapping,
1957 loff_t pos, unsigned len, unsigned flags,
1958 struct page **pagep, void **fsdata)
1960 pgoff_t index = pos >> PAGE_SHIFT;
1961 struct fuse_conn *fc = get_fuse_conn(file_inode(file));
1966 WARN_ON(!fc->writeback_cache);
1968 page = grab_cache_page_write_begin(mapping, index, flags);
1972 fuse_wait_on_page_writeback(mapping->host, page->index);
1974 if (PageUptodate(page) || len == PAGE_SIZE)
1977 * Check if the start this page comes after the end of file, in which
1978 * case the readpage can be optimized away.
1980 fsize = i_size_read(mapping->host);
1981 if (fsize <= (pos & PAGE_MASK)) {
1982 size_t off = pos & ~PAGE_MASK;
1984 zero_user_segment(page, 0, off);
1987 err = fuse_do_readpage(file, page);
2001 static int fuse_write_end(struct file *file, struct address_space *mapping,
2002 loff_t pos, unsigned len, unsigned copied,
2003 struct page *page, void *fsdata)
2005 struct inode *inode = page->mapping->host;
2007 /* Haven't copied anything? Skip zeroing, size extending, dirtying. */
2011 if (!PageUptodate(page)) {
2012 /* Zero any unwritten bytes at the end of the page */
2013 size_t endoff = (pos + copied) & ~PAGE_MASK;
2015 zero_user_segment(page, endoff, PAGE_SIZE);
2016 SetPageUptodate(page);
2019 fuse_write_update_size(inode, pos + copied);
2020 set_page_dirty(page);
2029 static int fuse_launder_page(struct page *page)
2032 if (clear_page_dirty_for_io(page)) {
2033 struct inode *inode = page->mapping->host;
2034 err = fuse_writepage_locked(page);
2036 fuse_wait_on_page_writeback(inode, page->index);
2042 * Write back dirty pages now, because there may not be any suitable
2045 static void fuse_vma_close(struct vm_area_struct *vma)
2047 filemap_write_and_wait(vma->vm_file->f_mapping);
2051 * Wait for writeback against this page to complete before allowing it
2052 * to be marked dirty again, and hence written back again, possibly
2053 * before the previous writepage completed.
2055 * Block here, instead of in ->writepage(), so that the userspace fs
2056 * can only block processes actually operating on the filesystem.
2058 * Otherwise unprivileged userspace fs would be able to block
2063 * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2065 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
2067 struct page *page = vmf->page;
2068 struct inode *inode = file_inode(vma->vm_file);
2070 file_update_time(vma->vm_file);
2072 if (page->mapping != inode->i_mapping) {
2074 return VM_FAULT_NOPAGE;
2077 fuse_wait_on_page_writeback(inode, page->index);
2078 return VM_FAULT_LOCKED;
2081 static const struct vm_operations_struct fuse_file_vm_ops = {
2082 .close = fuse_vma_close,
2083 .fault = filemap_fault,
2084 .map_pages = filemap_map_pages,
2085 .page_mkwrite = fuse_page_mkwrite,
2088 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2090 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2091 fuse_link_write_file(file);
2093 file_accessed(file);
2094 vma->vm_ops = &fuse_file_vm_ops;
2098 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
2100 /* Can't provide the coherency needed for MAP_SHARED */
2101 if (vma->vm_flags & VM_MAYSHARE)
2104 invalidate_inode_pages2(file->f_mapping);
2106 return generic_file_mmap(file, vma);
2109 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
2110 struct file_lock *fl)
2112 switch (ffl->type) {
2118 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2119 ffl->end < ffl->start)
2122 fl->fl_start = ffl->start;
2123 fl->fl_end = ffl->end;
2124 fl->fl_pid = ffl->pid;
2130 fl->fl_type = ffl->type;
2134 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2135 const struct file_lock *fl, int opcode, pid_t pid,
2136 int flock, struct fuse_lk_in *inarg)
2138 struct inode *inode = file_inode(file);
2139 struct fuse_conn *fc = get_fuse_conn(inode);
2140 struct fuse_file *ff = file->private_data;
2142 memset(inarg, 0, sizeof(*inarg));
2144 inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2145 inarg->lk.start = fl->fl_start;
2146 inarg->lk.end = fl->fl_end;
2147 inarg->lk.type = fl->fl_type;
2148 inarg->lk.pid = pid;
2150 inarg->lk_flags |= FUSE_LK_FLOCK;
2151 args->in.h.opcode = opcode;
2152 args->in.h.nodeid = get_node_id(inode);
2153 args->in.numargs = 1;
2154 args->in.args[0].size = sizeof(*inarg);
2155 args->in.args[0].value = inarg;
2158 static int fuse_getlk(struct file *file, struct file_lock *fl)
2160 struct inode *inode = file_inode(file);
2161 struct fuse_conn *fc = get_fuse_conn(inode);
2163 struct fuse_lk_in inarg;
2164 struct fuse_lk_out outarg;
2167 fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2168 args.out.numargs = 1;
2169 args.out.args[0].size = sizeof(outarg);
2170 args.out.args[0].value = &outarg;
2171 err = fuse_simple_request(fc, &args);
2173 err = convert_fuse_file_lock(&outarg.lk, fl);
2178 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2180 struct inode *inode = file_inode(file);
2181 struct fuse_conn *fc = get_fuse_conn(inode);
2183 struct fuse_lk_in inarg;
2184 int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2185 pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
2188 if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2189 /* NLM needs asynchronous locks, which we don't support yet */
2193 /* Unlock on close is handled by the flush method */
2194 if (fl->fl_flags & FL_CLOSE)
2197 fuse_lk_fill(&args, file, fl, opcode, pid, flock, &inarg);
2198 err = fuse_simple_request(fc, &args);
2200 /* locking is restartable */
2207 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2209 struct inode *inode = file_inode(file);
2210 struct fuse_conn *fc = get_fuse_conn(inode);
2213 if (cmd == F_CANCELLK) {
2215 } else if (cmd == F_GETLK) {
2217 posix_test_lock(file, fl);
2220 err = fuse_getlk(file, fl);
2223 err = posix_lock_file(file, fl, NULL);
2225 err = fuse_setlk(file, fl, 0);
2230 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2232 struct inode *inode = file_inode(file);
2233 struct fuse_conn *fc = get_fuse_conn(inode);
2237 err = locks_lock_file_wait(file, fl);
2239 struct fuse_file *ff = file->private_data;
2241 /* emulate flock with POSIX locks */
2243 err = fuse_setlk(file, fl, 1);
2249 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2251 struct inode *inode = mapping->host;
2252 struct fuse_conn *fc = get_fuse_conn(inode);
2254 struct fuse_bmap_in inarg;
2255 struct fuse_bmap_out outarg;
2258 if (!inode->i_sb->s_bdev || fc->no_bmap)
2261 memset(&inarg, 0, sizeof(inarg));
2262 inarg.block = block;
2263 inarg.blocksize = inode->i_sb->s_blocksize;
2264 args.in.h.opcode = FUSE_BMAP;
2265 args.in.h.nodeid = get_node_id(inode);
2266 args.in.numargs = 1;
2267 args.in.args[0].size = sizeof(inarg);
2268 args.in.args[0].value = &inarg;
2269 args.out.numargs = 1;
2270 args.out.args[0].size = sizeof(outarg);
2271 args.out.args[0].value = &outarg;
2272 err = fuse_simple_request(fc, &args);
2276 return err ? 0 : outarg.block;
2279 static loff_t fuse_lseek(struct file *file, loff_t offset, int whence)
2281 struct inode *inode = file->f_mapping->host;
2282 struct fuse_conn *fc = get_fuse_conn(inode);
2283 struct fuse_file *ff = file->private_data;
2285 struct fuse_lseek_in inarg = {
2290 struct fuse_lseek_out outarg;
2296 args.in.h.opcode = FUSE_LSEEK;
2297 args.in.h.nodeid = ff->nodeid;
2298 args.in.numargs = 1;
2299 args.in.args[0].size = sizeof(inarg);
2300 args.in.args[0].value = &inarg;
2301 args.out.numargs = 1;
2302 args.out.args[0].size = sizeof(outarg);
2303 args.out.args[0].value = &outarg;
2304 err = fuse_simple_request(fc, &args);
2306 if (err == -ENOSYS) {
2313 return vfs_setpos(file, outarg.offset, inode->i_sb->s_maxbytes);
2316 err = fuse_update_attributes(inode, NULL, file, NULL);
2318 return generic_file_llseek(file, offset, whence);
2323 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2326 struct inode *inode = file_inode(file);
2331 /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2332 retval = generic_file_llseek(file, offset, whence);
2336 retval = fuse_update_attributes(inode, NULL, file, NULL);
2338 retval = generic_file_llseek(file, offset, whence);
2339 inode_unlock(inode);
2344 retval = fuse_lseek(file, offset, whence);
2345 inode_unlock(inode);
2355 * CUSE servers compiled on 32bit broke on 64bit kernels because the
2356 * ABI was defined to be 'struct iovec' which is different on 32bit
2357 * and 64bit. Fortunately we can determine which structure the server
2358 * used from the size of the reply.
2360 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2361 size_t transferred, unsigned count,
2364 #ifdef CONFIG_COMPAT
2365 if (count * sizeof(struct compat_iovec) == transferred) {
2366 struct compat_iovec *ciov = src;
2370 * With this interface a 32bit server cannot support
2371 * non-compat (i.e. ones coming from 64bit apps) ioctl
2377 for (i = 0; i < count; i++) {
2378 dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2379 dst[i].iov_len = ciov[i].iov_len;
2385 if (count * sizeof(struct iovec) != transferred)
2388 memcpy(dst, src, transferred);
2392 /* Make sure iov_length() won't overflow */
2393 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
2396 u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
2398 for (n = 0; n < count; n++, iov++) {
2399 if (iov->iov_len > (size_t) max)
2401 max -= iov->iov_len;
2406 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2407 void *src, size_t transferred, unsigned count,
2411 struct fuse_ioctl_iovec *fiov = src;
2413 if (fc->minor < 16) {
2414 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2418 if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2421 for (i = 0; i < count; i++) {
2422 /* Did the server supply an inappropriate value? */
2423 if (fiov[i].base != (unsigned long) fiov[i].base ||
2424 fiov[i].len != (unsigned long) fiov[i].len)
2427 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2428 dst[i].iov_len = (size_t) fiov[i].len;
2430 #ifdef CONFIG_COMPAT
2432 (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2433 (compat_size_t) dst[i].iov_len != fiov[i].len))
2443 * For ioctls, there is no generic way to determine how much memory
2444 * needs to be read and/or written. Furthermore, ioctls are allowed
2445 * to dereference the passed pointer, so the parameter requires deep
2446 * copying but FUSE has no idea whatsoever about what to copy in or
2449 * This is solved by allowing FUSE server to retry ioctl with
2450 * necessary in/out iovecs. Let's assume the ioctl implementation
2451 * needs to read in the following structure.
2458 * On the first callout to FUSE server, inarg->in_size and
2459 * inarg->out_size will be NULL; then, the server completes the ioctl
2460 * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2461 * the actual iov array to
2463 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) } }
2465 * which tells FUSE to copy in the requested area and retry the ioctl.
2466 * On the second round, the server has access to the structure and
2467 * from that it can tell what to look for next, so on the invocation,
2468 * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2470 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) },
2471 * { .iov_base = a.buf, .iov_len = a.buflen } }
2473 * FUSE will copy both struct a and the pointed buffer from the
2474 * process doing the ioctl and retry ioctl with both struct a and the
2477 * This time, FUSE server has everything it needs and completes ioctl
2478 * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2480 * Copying data out works the same way.
2482 * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2483 * automatically initializes in and out iovs by decoding @cmd with
2484 * _IOC_* macros and the server is not allowed to request RETRY. This
2485 * limits ioctl data transfers to well-formed ioctls and is the forced
2486 * behavior for all FUSE servers.
2488 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2491 struct fuse_file *ff = file->private_data;
2492 struct fuse_conn *fc = ff->fc;
2493 struct fuse_ioctl_in inarg = {
2499 struct fuse_ioctl_out outarg;
2500 struct fuse_req *req = NULL;
2501 struct page **pages = NULL;
2502 struct iovec *iov_page = NULL;
2503 struct iovec *in_iov = NULL, *out_iov = NULL;
2504 unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
2505 size_t in_size, out_size, transferred, c;
2509 #if BITS_PER_LONG == 32
2510 inarg.flags |= FUSE_IOCTL_32BIT;
2512 if (flags & FUSE_IOCTL_COMPAT)
2513 inarg.flags |= FUSE_IOCTL_32BIT;
2516 /* assume all the iovs returned by client always fits in a page */
2517 BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2520 pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
2521 iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2522 if (!pages || !iov_page)
2526 * If restricted, initialize IO parameters as encoded in @cmd.
2527 * RETRY from server is not allowed.
2529 if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2530 struct iovec *iov = iov_page;
2532 iov->iov_base = (void __user *)arg;
2535 case FS_IOC_GETFLAGS:
2536 case FS_IOC_SETFLAGS:
2537 iov->iov_len = sizeof(int);
2540 iov->iov_len = _IOC_SIZE(cmd);
2544 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2549 if (_IOC_DIR(cmd) & _IOC_READ) {
2556 inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2557 inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2560 * Out data can be used either for actual out data or iovs,
2561 * make sure there always is at least one page.
2563 out_size = max_t(size_t, out_size, PAGE_SIZE);
2564 max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2566 /* make sure there are enough buffer pages and init request with them */
2568 if (max_pages > FUSE_MAX_PAGES_PER_REQ)
2570 while (num_pages < max_pages) {
2571 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2572 if (!pages[num_pages])
2577 req = fuse_get_req(fc, num_pages);
2583 memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
2584 req->num_pages = num_pages;
2585 fuse_page_descs_length_init(req, 0, req->num_pages);
2587 /* okay, let's send it to the client */
2588 req->in.h.opcode = FUSE_IOCTL;
2589 req->in.h.nodeid = ff->nodeid;
2590 req->in.numargs = 1;
2591 req->in.args[0].size = sizeof(inarg);
2592 req->in.args[0].value = &inarg;
2595 req->in.args[1].size = in_size;
2596 req->in.argpages = 1;
2599 iov_iter_init(&ii, WRITE, in_iov, in_iovs, in_size);
2600 for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= num_pages); i++) {
2601 c = copy_page_from_iter(pages[i], 0, PAGE_SIZE, &ii);
2602 if (c != PAGE_SIZE && iov_iter_count(&ii))
2607 req->out.numargs = 2;
2608 req->out.args[0].size = sizeof(outarg);
2609 req->out.args[0].value = &outarg;
2610 req->out.args[1].size = out_size;
2611 req->out.argpages = 1;
2612 req->out.argvar = 1;
2614 fuse_request_send(fc, req);
2615 err = req->out.h.error;
2616 transferred = req->out.args[1].size;
2617 fuse_put_request(fc, req);
2622 /* did it ask for retry? */
2623 if (outarg.flags & FUSE_IOCTL_RETRY) {
2626 /* no retry if in restricted mode */
2628 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2631 in_iovs = outarg.in_iovs;
2632 out_iovs = outarg.out_iovs;
2635 * Make sure things are in boundary, separate checks
2636 * are to protect against overflow.
2639 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2640 out_iovs > FUSE_IOCTL_MAX_IOV ||
2641 in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2644 vaddr = kmap_atomic(pages[0]);
2645 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2646 transferred, in_iovs + out_iovs,
2647 (flags & FUSE_IOCTL_COMPAT) != 0);
2648 kunmap_atomic(vaddr);
2653 out_iov = in_iov + in_iovs;
2655 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
2659 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
2667 if (transferred > inarg.out_size)
2671 iov_iter_init(&ii, READ, out_iov, out_iovs, transferred);
2672 for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= num_pages); i++) {
2673 c = copy_page_to_iter(pages[i], 0, PAGE_SIZE, &ii);
2674 if (c != PAGE_SIZE && iov_iter_count(&ii))
2680 fuse_put_request(fc, req);
2681 free_page((unsigned long) iov_page);
2683 __free_page(pages[--num_pages]);
2686 return err ? err : outarg.result;
2688 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2690 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2691 unsigned long arg, unsigned int flags)
2693 struct inode *inode = file_inode(file);
2694 struct fuse_conn *fc = get_fuse_conn(inode);
2696 if (!fuse_allow_current_process(fc))
2699 if (fuse_is_bad(inode))
2702 return fuse_do_ioctl(file, cmd, arg, flags);
2705 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2708 return fuse_ioctl_common(file, cmd, arg, 0);
2711 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2714 return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2718 * All files which have been polled are linked to RB tree
2719 * fuse_conn->polled_files which is indexed by kh. Walk the tree and
2720 * find the matching one.
2722 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2723 struct rb_node **parent_out)
2725 struct rb_node **link = &fc->polled_files.rb_node;
2726 struct rb_node *last = NULL;
2729 struct fuse_file *ff;
2732 ff = rb_entry(last, struct fuse_file, polled_node);
2735 link = &last->rb_left;
2736 else if (kh > ff->kh)
2737 link = &last->rb_right;
2748 * The file is about to be polled. Make sure it's on the polled_files
2749 * RB tree. Note that files once added to the polled_files tree are
2750 * not removed before the file is released. This is because a file
2751 * polled once is likely to be polled again.
2753 static void fuse_register_polled_file(struct fuse_conn *fc,
2754 struct fuse_file *ff)
2756 spin_lock(&fc->lock);
2757 if (RB_EMPTY_NODE(&ff->polled_node)) {
2758 struct rb_node **link, *uninitialized_var(parent);
2760 link = fuse_find_polled_node(fc, ff->kh, &parent);
2762 rb_link_node(&ff->polled_node, parent, link);
2763 rb_insert_color(&ff->polled_node, &fc->polled_files);
2765 spin_unlock(&fc->lock);
2768 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2770 struct fuse_file *ff = file->private_data;
2771 struct fuse_conn *fc = ff->fc;
2772 struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2773 struct fuse_poll_out outarg;
2778 return DEFAULT_POLLMASK;
2780 poll_wait(file, &ff->poll_wait, wait);
2781 inarg.events = (__u32)poll_requested_events(wait);
2784 * Ask for notification iff there's someone waiting for it.
2785 * The client may ignore the flag and always notify.
2787 if (waitqueue_active(&ff->poll_wait)) {
2788 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2789 fuse_register_polled_file(fc, ff);
2792 args.in.h.opcode = FUSE_POLL;
2793 args.in.h.nodeid = ff->nodeid;
2794 args.in.numargs = 1;
2795 args.in.args[0].size = sizeof(inarg);
2796 args.in.args[0].value = &inarg;
2797 args.out.numargs = 1;
2798 args.out.args[0].size = sizeof(outarg);
2799 args.out.args[0].value = &outarg;
2800 err = fuse_simple_request(fc, &args);
2803 return outarg.revents;
2804 if (err == -ENOSYS) {
2806 return DEFAULT_POLLMASK;
2810 EXPORT_SYMBOL_GPL(fuse_file_poll);
2813 * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2814 * wakes up the poll waiters.
2816 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2817 struct fuse_notify_poll_wakeup_out *outarg)
2819 u64 kh = outarg->kh;
2820 struct rb_node **link;
2822 spin_lock(&fc->lock);
2824 link = fuse_find_polled_node(fc, kh, NULL);
2826 struct fuse_file *ff;
2828 ff = rb_entry(*link, struct fuse_file, polled_node);
2829 wake_up_interruptible_sync(&ff->poll_wait);
2832 spin_unlock(&fc->lock);
2836 static void fuse_do_truncate(struct file *file)
2838 struct inode *inode = file->f_mapping->host;
2841 attr.ia_valid = ATTR_SIZE;
2842 attr.ia_size = i_size_read(inode);
2844 attr.ia_file = file;
2845 attr.ia_valid |= ATTR_FILE;
2847 fuse_do_setattr(file_dentry(file), &attr, file);
2850 static inline loff_t fuse_round_up(loff_t off)
2852 return round_up(off, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
2856 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
2858 DECLARE_COMPLETION_ONSTACK(wait);
2860 struct file *file = iocb->ki_filp;
2861 struct fuse_file *ff = file->private_data;
2862 bool async_dio = ff->fc->async_dio;
2864 struct inode *inode;
2866 size_t count = iov_iter_count(iter);
2867 loff_t offset = iocb->ki_pos;
2868 struct fuse_io_priv *io;
2871 inode = file->f_mapping->host;
2872 i_size = i_size_read(inode);
2874 if ((iov_iter_rw(iter) == READ) && (offset > i_size))
2877 /* optimization for short read */
2878 if (async_dio && iov_iter_rw(iter) != WRITE && offset + count > i_size) {
2879 if (offset >= i_size)
2881 iov_iter_truncate(iter, fuse_round_up(i_size - offset));
2882 count = iov_iter_count(iter);
2885 io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2888 spin_lock_init(&io->lock);
2889 kref_init(&io->refcnt);
2893 io->offset = offset;
2894 io->write = (iov_iter_rw(iter) == WRITE);
2898 * By default, we want to optimize all I/Os with async request
2899 * submission to the client filesystem if supported.
2901 io->async = async_dio;
2903 io->blocking = is_sync_kiocb(iocb);
2906 * We cannot asynchronously extend the size of a file.
2907 * In such case the aio will behave exactly like sync io.
2909 if ((offset + count > i_size) && iov_iter_rw(iter) == WRITE)
2910 io->blocking = true;
2912 if (io->async && io->blocking) {
2914 * Additional reference to keep io around after
2915 * calling fuse_aio_complete()
2917 kref_get(&io->refcnt);
2921 if (iov_iter_rw(iter) == WRITE) {
2922 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
2923 fuse_invalidate_attr(inode);
2925 ret = __fuse_direct_read(io, iter, &pos);
2929 bool blocking = io->blocking;
2931 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2933 /* we have a non-extending, async request, so return */
2935 return -EIOCBQUEUED;
2937 wait_for_completion(&wait);
2938 ret = fuse_get_res_by_io(io);
2941 kref_put(&io->refcnt, fuse_io_release);
2943 if (iov_iter_rw(iter) == WRITE) {
2945 fuse_write_update_size(inode, pos);
2946 else if (ret < 0 && offset + count > i_size)
2947 fuse_do_truncate(file);
2953 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2956 struct fuse_file *ff = file->private_data;
2957 struct inode *inode = file_inode(file);
2958 struct fuse_inode *fi = get_fuse_inode(inode);
2959 struct fuse_conn *fc = ff->fc;
2961 struct fuse_fallocate_in inarg = {
2968 bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
2969 (mode & FALLOC_FL_PUNCH_HOLE);
2971 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2974 if (fc->no_fallocate)
2979 if (mode & FALLOC_FL_PUNCH_HOLE) {
2980 loff_t endbyte = offset + length - 1;
2981 err = filemap_write_and_wait_range(inode->i_mapping,
2986 fuse_sync_writes(inode);
2990 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
2991 offset + length > i_size_read(inode)) {
2992 err = inode_newsize_ok(inode, offset + length);
2997 if (!(mode & FALLOC_FL_KEEP_SIZE))
2998 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3000 args.in.h.opcode = FUSE_FALLOCATE;
3001 args.in.h.nodeid = ff->nodeid;
3002 args.in.numargs = 1;
3003 args.in.args[0].size = sizeof(inarg);
3004 args.in.args[0].value = &inarg;
3005 err = fuse_simple_request(fc, &args);
3006 if (err == -ENOSYS) {
3007 fc->no_fallocate = 1;
3013 /* we could have extended the file */
3014 if (!(mode & FALLOC_FL_KEEP_SIZE)) {
3015 bool changed = fuse_write_update_size(inode, offset + length);
3017 if (changed && fc->writeback_cache)
3018 file_update_time(file);
3021 if (mode & FALLOC_FL_PUNCH_HOLE)
3022 truncate_pagecache_range(inode, offset, offset + length - 1);
3024 fuse_invalidate_attr(inode);
3027 if (!(mode & FALLOC_FL_KEEP_SIZE))
3028 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3031 inode_unlock(inode);
3036 static const struct file_operations fuse_file_operations = {
3037 .llseek = fuse_file_llseek,
3038 .read_iter = fuse_file_read_iter,
3039 .write_iter = fuse_file_write_iter,
3040 .mmap = fuse_file_mmap,
3042 .flush = fuse_flush,
3043 .release = fuse_release,
3044 .fsync = fuse_fsync,
3045 .lock = fuse_file_lock,
3046 .flock = fuse_file_flock,
3047 .splice_read = generic_file_splice_read,
3048 .unlocked_ioctl = fuse_file_ioctl,
3049 .compat_ioctl = fuse_file_compat_ioctl,
3050 .poll = fuse_file_poll,
3051 .fallocate = fuse_file_fallocate,
3054 static const struct file_operations fuse_direct_io_file_operations = {
3055 .llseek = fuse_file_llseek,
3056 .read_iter = fuse_direct_read_iter,
3057 .write_iter = fuse_direct_write_iter,
3058 .mmap = fuse_direct_mmap,
3060 .flush = fuse_flush,
3061 .release = fuse_release,
3062 .fsync = fuse_fsync,
3063 .lock = fuse_file_lock,
3064 .flock = fuse_file_flock,
3065 .unlocked_ioctl = fuse_file_ioctl,
3066 .compat_ioctl = fuse_file_compat_ioctl,
3067 .poll = fuse_file_poll,
3068 .fallocate = fuse_file_fallocate,
3069 /* no splice_read */
3072 static const struct address_space_operations fuse_file_aops = {
3073 .readpage = fuse_readpage,
3074 .writepage = fuse_writepage,
3075 .writepages = fuse_writepages,
3076 .launder_page = fuse_launder_page,
3077 .readpages = fuse_readpages,
3078 .set_page_dirty = __set_page_dirty_nobuffers,
3080 .direct_IO = fuse_direct_IO,
3081 .write_begin = fuse_write_begin,
3082 .write_end = fuse_write_end,
3085 void fuse_init_file_inode(struct inode *inode)
3087 inode->i_fop = &fuse_file_operations;
3088 inode->i_data.a_ops = &fuse_file_aops;