2 FUSE: Filesystem in Userspace
3 Copyright (C) 2001-2008 Miklos Szeredi <miklos@szeredi.hu>
5 This program can be distributed under the terms of the GNU GPL.
11 #include <linux/pagemap.h>
12 #include <linux/slab.h>
13 #include <linux/kernel.h>
14 #include <linux/sched.h>
15 #include <linux/module.h>
16 #include <linux/compat.h>
17 #include <linux/swap.h>
18 #include <linux/falloc.h>
19 #include <linux/uio.h>
21 static const struct file_operations fuse_direct_io_file_operations;
23 static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
24 int opcode, struct fuse_open_out *outargp)
26 struct fuse_open_in inarg;
29 memset(&inarg, 0, sizeof(inarg));
30 inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
31 if (!fc->atomic_o_trunc)
32 inarg.flags &= ~O_TRUNC;
33 args.in.h.opcode = opcode;
34 args.in.h.nodeid = nodeid;
36 args.in.args[0].size = sizeof(inarg);
37 args.in.args[0].value = &inarg;
39 args.out.args[0].size = sizeof(*outargp);
40 args.out.args[0].value = outargp;
42 return fuse_simple_request(fc, &args);
45 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
49 ff = kmalloc(sizeof(struct fuse_file), GFP_KERNEL);
54 ff->reserved_req = fuse_request_alloc(0);
55 if (unlikely(!ff->reserved_req)) {
60 INIT_LIST_HEAD(&ff->write_entry);
61 atomic_set(&ff->count, 0);
62 RB_CLEAR_NODE(&ff->polled_node);
63 init_waitqueue_head(&ff->poll_wait);
67 spin_unlock(&fc->lock);
72 void fuse_file_free(struct fuse_file *ff)
74 fuse_request_free(ff->reserved_req);
78 struct fuse_file *fuse_file_get(struct fuse_file *ff)
80 atomic_inc(&ff->count);
84 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
86 iput(req->misc.release.inode);
89 static void fuse_file_put(struct fuse_file *ff, bool sync)
91 if (atomic_dec_and_test(&ff->count)) {
92 struct fuse_req *req = ff->reserved_req;
94 if (ff->fc->no_open) {
96 * Drop the release request when client does not
99 __clear_bit(FR_BACKGROUND, &req->flags);
100 iput(req->misc.release.inode);
101 fuse_put_request(ff->fc, req);
103 __clear_bit(FR_BACKGROUND, &req->flags);
104 fuse_request_send(ff->fc, req);
105 iput(req->misc.release.inode);
106 fuse_put_request(ff->fc, req);
108 req->end = fuse_release_end;
109 __set_bit(FR_BACKGROUND, &req->flags);
110 fuse_request_send_background(ff->fc, req);
116 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
119 struct fuse_file *ff;
120 int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
122 ff = fuse_file_alloc(fc);
127 ff->open_flags = FOPEN_KEEP_CACHE; /* Default for no-open */
128 if (!fc->no_open || isdir) {
129 struct fuse_open_out outarg;
132 err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
135 ff->open_flags = outarg.open_flags;
137 } else if (err != -ENOSYS || isdir) {
146 ff->open_flags &= ~FOPEN_DIRECT_IO;
149 file->private_data = fuse_file_get(ff);
153 EXPORT_SYMBOL_GPL(fuse_do_open);
155 static void fuse_link_write_file(struct file *file)
157 struct inode *inode = file_inode(file);
158 struct fuse_conn *fc = get_fuse_conn(inode);
159 struct fuse_inode *fi = get_fuse_inode(inode);
160 struct fuse_file *ff = file->private_data;
162 * file may be written through mmap, so chain it onto the
163 * inodes's write_file list
165 spin_lock(&fc->lock);
166 if (list_empty(&ff->write_entry))
167 list_add(&ff->write_entry, &fi->write_files);
168 spin_unlock(&fc->lock);
171 void fuse_finish_open(struct inode *inode, struct file *file)
173 struct fuse_file *ff = file->private_data;
174 struct fuse_conn *fc = get_fuse_conn(inode);
176 if (ff->open_flags & FOPEN_DIRECT_IO)
177 file->f_op = &fuse_direct_io_file_operations;
178 if (!(ff->open_flags & FOPEN_KEEP_CACHE))
179 invalidate_inode_pages2(inode->i_mapping);
180 if (ff->open_flags & FOPEN_NONSEEKABLE)
181 nonseekable_open(inode, file);
182 if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
183 struct fuse_inode *fi = get_fuse_inode(inode);
185 spin_lock(&fc->lock);
186 fi->attr_version = ++fc->attr_version;
187 i_size_write(inode, 0);
188 spin_unlock(&fc->lock);
189 fuse_invalidate_attr(inode);
190 if (fc->writeback_cache)
191 file_update_time(file);
193 if ((file->f_mode & FMODE_WRITE) && fc->writeback_cache)
194 fuse_link_write_file(file);
197 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
199 struct fuse_conn *fc = get_fuse_conn(inode);
201 bool lock_inode = (file->f_flags & O_TRUNC) &&
202 fc->atomic_o_trunc &&
205 err = generic_file_open(inode, file);
210 mutex_lock(&inode->i_mutex);
212 err = fuse_do_open(fc, get_node_id(inode), file, isdir);
215 fuse_finish_open(inode, file);
218 mutex_unlock(&inode->i_mutex);
223 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
225 struct fuse_conn *fc = ff->fc;
226 struct fuse_req *req = ff->reserved_req;
227 struct fuse_release_in *inarg = &req->misc.release.in;
229 spin_lock(&fc->lock);
230 list_del(&ff->write_entry);
231 if (!RB_EMPTY_NODE(&ff->polled_node))
232 rb_erase(&ff->polled_node, &fc->polled_files);
233 spin_unlock(&fc->lock);
235 wake_up_interruptible_all(&ff->poll_wait);
238 inarg->flags = flags;
239 req->in.h.opcode = opcode;
240 req->in.h.nodeid = ff->nodeid;
242 req->in.args[0].size = sizeof(struct fuse_release_in);
243 req->in.args[0].value = inarg;
246 void fuse_release_common(struct file *file, int opcode)
248 struct fuse_file *ff;
249 struct fuse_req *req;
251 ff = file->private_data;
255 req = ff->reserved_req;
256 fuse_prepare_release(ff, file->f_flags, opcode);
259 struct fuse_release_in *inarg = &req->misc.release.in;
260 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
261 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
264 /* Hold inode until release is finished */
265 req->misc.release.inode = igrab(file_inode(file));
268 * Normally this will send the RELEASE request, however if
269 * some asynchronous READ or WRITE requests are outstanding,
270 * the sending will be delayed.
272 * Make the release synchronous if this is a fuseblk mount,
273 * synchronous RELEASE is allowed (and desirable) in this case
274 * because the server can be trusted not to screw up.
276 fuse_file_put(ff, ff->fc->destroy_req != NULL);
279 static int fuse_open(struct inode *inode, struct file *file)
281 return fuse_open_common(inode, file, false);
284 static int fuse_release(struct inode *inode, struct file *file)
286 struct fuse_conn *fc = get_fuse_conn(inode);
288 /* see fuse_vma_close() for !writeback_cache case */
289 if (fc->writeback_cache)
290 write_inode_now(inode, 1);
292 fuse_release_common(file, FUSE_RELEASE);
294 /* return value is ignored by VFS */
298 void fuse_sync_release(struct fuse_file *ff, int flags)
300 WARN_ON(atomic_read(&ff->count) > 1);
301 fuse_prepare_release(ff, flags, FUSE_RELEASE);
302 __set_bit(FR_FORCE, &ff->reserved_req->flags);
303 __clear_bit(FR_BACKGROUND, &ff->reserved_req->flags);
304 fuse_request_send(ff->fc, ff->reserved_req);
305 fuse_put_request(ff->fc, ff->reserved_req);
308 EXPORT_SYMBOL_GPL(fuse_sync_release);
311 * Scramble the ID space with XTEA, so that the value of the files_struct
312 * pointer is not exposed to userspace.
314 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
316 u32 *k = fc->scramble_key;
317 u64 v = (unsigned long) id;
323 for (i = 0; i < 32; i++) {
324 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
326 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
329 return (u64) v0 + ((u64) v1 << 32);
333 * Check if any page in a range is under writeback
335 * This is currently done by walking the list of writepage requests
336 * for the inode, which can be pretty inefficient.
338 static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
341 struct fuse_conn *fc = get_fuse_conn(inode);
342 struct fuse_inode *fi = get_fuse_inode(inode);
343 struct fuse_req *req;
346 spin_lock(&fc->lock);
347 list_for_each_entry(req, &fi->writepages, writepages_entry) {
350 BUG_ON(req->inode != inode);
351 curr_index = req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
352 if (idx_from < curr_index + req->num_pages &&
353 curr_index <= idx_to) {
358 spin_unlock(&fc->lock);
363 static inline bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
365 return fuse_range_is_writeback(inode, index, index);
369 * Wait for page writeback to be completed.
371 * Since fuse doesn't rely on the VM writeback tracking, this has to
372 * use some other means.
374 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
376 struct fuse_inode *fi = get_fuse_inode(inode);
378 wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
383 * Wait for all pending writepages on the inode to finish.
385 * This is currently done by blocking further writes with FUSE_NOWRITE
386 * and waiting for all sent writes to complete.
388 * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
389 * could conflict with truncation.
391 static void fuse_sync_writes(struct inode *inode)
393 fuse_set_nowrite(inode);
394 fuse_release_nowrite(inode);
397 static int fuse_flush(struct file *file, fl_owner_t id)
399 struct inode *inode = file_inode(file);
400 struct fuse_conn *fc = get_fuse_conn(inode);
401 struct fuse_file *ff = file->private_data;
402 struct fuse_req *req;
403 struct fuse_flush_in inarg;
406 if (is_bad_inode(inode))
412 err = write_inode_now(inode, 1);
416 mutex_lock(&inode->i_mutex);
417 fuse_sync_writes(inode);
418 mutex_unlock(&inode->i_mutex);
420 if (test_bit(AS_ENOSPC, &file->f_mapping->flags) &&
421 test_and_clear_bit(AS_ENOSPC, &file->f_mapping->flags))
423 if (test_bit(AS_EIO, &file->f_mapping->flags) &&
424 test_and_clear_bit(AS_EIO, &file->f_mapping->flags))
429 req = fuse_get_req_nofail_nopages(fc, file);
430 memset(&inarg, 0, sizeof(inarg));
432 inarg.lock_owner = fuse_lock_owner_id(fc, id);
433 req->in.h.opcode = FUSE_FLUSH;
434 req->in.h.nodeid = get_node_id(inode);
436 req->in.args[0].size = sizeof(inarg);
437 req->in.args[0].value = &inarg;
438 __set_bit(FR_FORCE, &req->flags);
439 fuse_request_send(fc, req);
440 err = req->out.h.error;
441 fuse_put_request(fc, req);
442 if (err == -ENOSYS) {
449 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
450 int datasync, int isdir)
452 struct inode *inode = file->f_mapping->host;
453 struct fuse_conn *fc = get_fuse_conn(inode);
454 struct fuse_file *ff = file->private_data;
456 struct fuse_fsync_in inarg;
459 if (is_bad_inode(inode))
462 mutex_lock(&inode->i_mutex);
465 * Start writeback against all dirty pages of the inode, then
466 * wait for all outstanding writes, before sending the FSYNC
469 err = filemap_write_and_wait_range(inode->i_mapping, start, end);
473 fuse_sync_writes(inode);
476 * Due to implementation of fuse writeback
477 * filemap_write_and_wait_range() does not catch errors.
478 * We have to do this directly after fuse_sync_writes()
480 if (test_bit(AS_ENOSPC, &file->f_mapping->flags) &&
481 test_and_clear_bit(AS_ENOSPC, &file->f_mapping->flags))
483 if (test_bit(AS_EIO, &file->f_mapping->flags) &&
484 test_and_clear_bit(AS_EIO, &file->f_mapping->flags))
489 err = sync_inode_metadata(inode, 1);
493 if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
496 memset(&inarg, 0, sizeof(inarg));
498 inarg.fsync_flags = datasync ? 1 : 0;
499 args.in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
500 args.in.h.nodeid = get_node_id(inode);
502 args.in.args[0].size = sizeof(inarg);
503 args.in.args[0].value = &inarg;
504 err = fuse_simple_request(fc, &args);
505 if (err == -ENOSYS) {
513 mutex_unlock(&inode->i_mutex);
517 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
520 return fuse_fsync_common(file, start, end, datasync, 0);
523 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
524 size_t count, int opcode)
526 struct fuse_read_in *inarg = &req->misc.read.in;
527 struct fuse_file *ff = file->private_data;
532 inarg->flags = file->f_flags;
533 req->in.h.opcode = opcode;
534 req->in.h.nodeid = ff->nodeid;
536 req->in.args[0].size = sizeof(struct fuse_read_in);
537 req->in.args[0].value = inarg;
539 req->out.numargs = 1;
540 req->out.args[0].size = count;
543 static void fuse_release_user_pages(struct fuse_req *req, bool should_dirty)
547 for (i = 0; i < req->num_pages; i++) {
548 struct page *page = req->pages[i];
550 set_page_dirty_lock(page);
555 static void fuse_io_release(struct kref *kref)
557 kfree(container_of(kref, struct fuse_io_priv, refcnt));
560 static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
565 if (io->bytes >= 0 && io->write)
568 return io->bytes < 0 ? io->size : io->bytes;
572 * In case of short read, the caller sets 'pos' to the position of
573 * actual end of fuse request in IO request. Otherwise, if bytes_requested
574 * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
577 * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
578 * both submitted asynchronously. The first of them was ACKed by userspace as
579 * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
580 * second request was ACKed as short, e.g. only 1K was read, resulting in
583 * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
584 * will be equal to the length of the longest contiguous fragment of
585 * transferred data starting from the beginning of IO request.
587 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
589 bool is_sync = is_sync_kiocb(io->iocb);
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 && is_sync)
601 spin_unlock(&io->lock);
603 if (!left && !is_sync) {
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_CACHE_SHIFT;
711 size_t off = num_read & (PAGE_CACHE_SIZE - 1);
713 for (i = start_idx; i < req->num_pages; i++) {
714 zero_user_segment(req->pages[i], off, PAGE_CACHE_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_CACHE_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 (is_bad_inode(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);
816 page_cache_release(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_CACHE_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_CACHE_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)) {
881 fuse_put_request(fc, req);
885 page_cache_get(page);
886 req->pages[req->num_pages] = page;
887 req->page_descs[req->num_pages].length = PAGE_SIZE;
893 static int fuse_readpages(struct file *file, struct address_space *mapping,
894 struct list_head *pages, unsigned nr_pages)
896 struct inode *inode = mapping->host;
897 struct fuse_conn *fc = get_fuse_conn(inode);
898 struct fuse_fill_data data;
900 int nr_alloc = min_t(unsigned, nr_pages, FUSE_MAX_PAGES_PER_REQ);
903 if (is_bad_inode(inode))
909 data.req = fuse_get_req_for_background(fc, nr_alloc);
911 data.req = fuse_get_req(fc, nr_alloc);
912 data.nr_pages = nr_pages;
913 err = PTR_ERR(data.req);
914 if (IS_ERR(data.req))
917 err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
919 if (data.req->num_pages)
920 fuse_send_readpages(data.req, file);
922 fuse_put_request(fc, data.req);
928 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
930 struct inode *inode = iocb->ki_filp->f_mapping->host;
931 struct fuse_conn *fc = get_fuse_conn(inode);
934 * In auto invalidate mode, always update attributes on read.
935 * Otherwise, only update if we attempt to read past EOF (to ensure
936 * i_size is up to date).
938 if (fc->auto_inval_data ||
939 (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
941 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
946 return generic_file_read_iter(iocb, to);
949 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
950 loff_t pos, size_t count)
952 struct fuse_write_in *inarg = &req->misc.write.in;
953 struct fuse_write_out *outarg = &req->misc.write.out;
958 req->in.h.opcode = FUSE_WRITE;
959 req->in.h.nodeid = ff->nodeid;
961 if (ff->fc->minor < 9)
962 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
964 req->in.args[0].size = sizeof(struct fuse_write_in);
965 req->in.args[0].value = inarg;
966 req->in.args[1].size = count;
967 req->out.numargs = 1;
968 req->out.args[0].size = sizeof(struct fuse_write_out);
969 req->out.args[0].value = outarg;
972 static size_t fuse_send_write(struct fuse_req *req, struct fuse_io_priv *io,
973 loff_t pos, size_t count, fl_owner_t owner)
975 struct file *file = io->file;
976 struct fuse_file *ff = file->private_data;
977 struct fuse_conn *fc = ff->fc;
978 struct fuse_write_in *inarg = &req->misc.write.in;
980 fuse_write_fill(req, ff, pos, count);
981 inarg->flags = file->f_flags;
983 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
984 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
988 return fuse_async_req_send(fc, req, count, io);
990 fuse_request_send(fc, req);
991 return req->misc.write.out.size;
994 bool fuse_write_update_size(struct inode *inode, loff_t pos)
996 struct fuse_conn *fc = get_fuse_conn(inode);
997 struct fuse_inode *fi = get_fuse_inode(inode);
1000 spin_lock(&fc->lock);
1001 fi->attr_version = ++fc->attr_version;
1002 if (pos > inode->i_size) {
1003 i_size_write(inode, pos);
1006 spin_unlock(&fc->lock);
1011 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
1012 struct inode *inode, loff_t pos,
1018 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
1020 for (i = 0; i < req->num_pages; i++)
1021 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
1023 res = fuse_send_write(req, &io, pos, count, NULL);
1025 offset = req->page_descs[0].offset;
1027 for (i = 0; i < req->num_pages; i++) {
1028 struct page *page = req->pages[i];
1030 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
1031 SetPageUptodate(page);
1033 if (count > PAGE_CACHE_SIZE - offset)
1034 count -= PAGE_CACHE_SIZE - offset;
1040 page_cache_release(page);
1046 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
1047 struct address_space *mapping,
1048 struct iov_iter *ii, loff_t pos)
1050 struct fuse_conn *fc = get_fuse_conn(mapping->host);
1051 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
1055 req->in.argpages = 1;
1056 req->page_descs[0].offset = offset;
1061 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1062 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
1063 iov_iter_count(ii));
1065 bytes = min_t(size_t, bytes, fc->max_write - count);
1069 if (iov_iter_fault_in_readable(ii, bytes))
1073 page = grab_cache_page_write_begin(mapping, index, 0);
1077 if (mapping_writably_mapped(mapping))
1078 flush_dcache_page(page);
1080 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
1081 flush_dcache_page(page);
1083 iov_iter_advance(ii, tmp);
1086 page_cache_release(page);
1087 bytes = min(bytes, iov_iter_single_seg_count(ii));
1092 req->pages[req->num_pages] = page;
1093 req->page_descs[req->num_pages].length = tmp;
1099 if (offset == PAGE_CACHE_SIZE)
1102 if (!fc->big_writes)
1104 } while (iov_iter_count(ii) && count < fc->max_write &&
1105 req->num_pages < req->max_pages && offset == 0);
1107 return count > 0 ? count : err;
1110 static inline unsigned fuse_wr_pages(loff_t pos, size_t len)
1112 return min_t(unsigned,
1113 ((pos + len - 1) >> PAGE_CACHE_SHIFT) -
1114 (pos >> PAGE_CACHE_SHIFT) + 1,
1115 FUSE_MAX_PAGES_PER_REQ);
1118 static ssize_t fuse_perform_write(struct file *file,
1119 struct address_space *mapping,
1120 struct iov_iter *ii, loff_t pos)
1122 struct inode *inode = mapping->host;
1123 struct fuse_conn *fc = get_fuse_conn(inode);
1124 struct fuse_inode *fi = get_fuse_inode(inode);
1128 if (is_bad_inode(inode))
1131 if (inode->i_size < pos + iov_iter_count(ii))
1132 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1135 struct fuse_req *req;
1137 unsigned nr_pages = fuse_wr_pages(pos, iov_iter_count(ii));
1139 req = fuse_get_req(fc, nr_pages);
1145 count = fuse_fill_write_pages(req, mapping, ii, pos);
1151 num_written = fuse_send_write_pages(req, file, inode,
1153 err = req->out.h.error;
1158 /* break out of the loop on short write */
1159 if (num_written != count)
1163 fuse_put_request(fc, req);
1164 } while (!err && iov_iter_count(ii));
1167 fuse_write_update_size(inode, pos);
1169 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1170 fuse_invalidate_attr(inode);
1172 return res > 0 ? res : err;
1175 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1177 struct file *file = iocb->ki_filp;
1178 struct address_space *mapping = file->f_mapping;
1179 ssize_t written = 0;
1180 ssize_t written_buffered = 0;
1181 struct inode *inode = mapping->host;
1185 if (get_fuse_conn(inode)->writeback_cache) {
1186 /* Update size (EOF optimization) and mode (SUID clearing) */
1187 err = fuse_update_attributes(mapping->host, NULL, file, NULL);
1191 return generic_file_write_iter(iocb, from);
1194 mutex_lock(&inode->i_mutex);
1196 /* We can write back this queue in page reclaim */
1197 current->backing_dev_info = inode_to_bdi(inode);
1199 err = generic_write_checks(iocb, from);
1203 err = file_remove_privs(file);
1207 err = file_update_time(file);
1211 if (iocb->ki_flags & IOCB_DIRECT) {
1212 loff_t pos = iocb->ki_pos;
1213 written = generic_file_direct_write(iocb, from, pos);
1214 if (written < 0 || !iov_iter_count(from))
1219 written_buffered = fuse_perform_write(file, mapping, from, pos);
1220 if (written_buffered < 0) {
1221 err = written_buffered;
1224 endbyte = pos + written_buffered - 1;
1226 err = filemap_write_and_wait_range(file->f_mapping, pos,
1231 invalidate_mapping_pages(file->f_mapping,
1232 pos >> PAGE_CACHE_SHIFT,
1233 endbyte >> PAGE_CACHE_SHIFT);
1235 written += written_buffered;
1236 iocb->ki_pos = pos + written_buffered;
1238 written = fuse_perform_write(file, mapping, from, iocb->ki_pos);
1240 iocb->ki_pos += written;
1243 current->backing_dev_info = NULL;
1244 mutex_unlock(&inode->i_mutex);
1246 return written ? written : err;
1249 static inline void fuse_page_descs_length_init(struct fuse_req *req,
1250 unsigned index, unsigned nr_pages)
1254 for (i = index; i < index + nr_pages; i++)
1255 req->page_descs[i].length = PAGE_SIZE -
1256 req->page_descs[i].offset;
1259 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1261 return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1264 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1267 return min(iov_iter_single_seg_count(ii), max_size);
1270 static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1271 size_t *nbytesp, int write)
1273 size_t nbytes = 0; /* # bytes already packed in req */
1275 /* Special case for kernel I/O: can copy directly into the buffer */
1276 if (ii->type & ITER_KVEC) {
1277 unsigned long user_addr = fuse_get_user_addr(ii);
1278 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1281 req->in.args[1].value = (void *) user_addr;
1283 req->out.args[0].value = (void *) user_addr;
1285 iov_iter_advance(ii, frag_size);
1286 *nbytesp = frag_size;
1290 while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1293 ssize_t ret = iov_iter_get_pages(ii,
1294 &req->pages[req->num_pages],
1296 req->max_pages - req->num_pages,
1301 iov_iter_advance(ii, ret);
1305 npages = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
1307 req->page_descs[req->num_pages].offset = start;
1308 fuse_page_descs_length_init(req, req->num_pages, npages);
1310 req->num_pages += npages;
1311 req->page_descs[req->num_pages - 1].length -=
1312 (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1316 req->in.argpages = 1;
1318 req->out.argpages = 1;
1325 static inline int fuse_iter_npages(const struct iov_iter *ii_p)
1327 return iov_iter_npages(ii_p, FUSE_MAX_PAGES_PER_REQ);
1330 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1331 loff_t *ppos, int flags)
1333 int write = flags & FUSE_DIO_WRITE;
1334 bool should_dirty = !write && iter_is_iovec(iter);
1335 int cuse = flags & FUSE_DIO_CUSE;
1336 struct file *file = io->file;
1337 struct inode *inode = file->f_mapping->host;
1338 struct fuse_file *ff = file->private_data;
1339 struct fuse_conn *fc = ff->fc;
1340 size_t nmax = write ? fc->max_write : fc->max_read;
1342 size_t count = iov_iter_count(iter);
1343 pgoff_t idx_from = pos >> PAGE_CACHE_SHIFT;
1344 pgoff_t idx_to = (pos + count - 1) >> PAGE_CACHE_SHIFT;
1346 struct fuse_req *req;
1349 req = fuse_get_req_for_background(fc, fuse_iter_npages(iter));
1351 req = fuse_get_req(fc, fuse_iter_npages(iter));
1353 return PTR_ERR(req);
1355 if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1357 mutex_lock(&inode->i_mutex);
1358 fuse_sync_writes(inode);
1360 mutex_unlock(&inode->i_mutex);
1365 fl_owner_t owner = current->files;
1366 size_t nbytes = min(count, nmax);
1367 int err = fuse_get_user_pages(req, iter, &nbytes, write);
1374 nres = fuse_send_write(req, io, pos, nbytes, owner);
1376 nres = fuse_send_read(req, io, pos, nbytes, owner);
1379 fuse_release_user_pages(req, should_dirty);
1380 if (req->out.h.error) {
1382 res = req->out.h.error;
1384 } else if (nres > nbytes) {
1394 fuse_put_request(fc, req);
1396 req = fuse_get_req_for_background(fc,
1397 fuse_iter_npages(iter));
1399 req = fuse_get_req(fc, fuse_iter_npages(iter));
1405 fuse_put_request(fc, req);
1411 EXPORT_SYMBOL_GPL(fuse_direct_io);
1413 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1414 struct iov_iter *iter,
1418 struct file *file = io->file;
1419 struct inode *inode = file_inode(file);
1421 if (is_bad_inode(inode))
1424 res = fuse_direct_io(io, iter, ppos, 0);
1426 fuse_invalidate_attr(inode);
1431 static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1433 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb->ki_filp);
1434 return __fuse_direct_read(&io, to, &iocb->ki_pos);
1437 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1439 struct file *file = iocb->ki_filp;
1440 struct inode *inode = file_inode(file);
1441 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
1444 if (is_bad_inode(inode))
1447 /* Don't allow parallel writes to the same file */
1448 mutex_lock(&inode->i_mutex);
1449 res = generic_write_checks(iocb, from);
1451 res = fuse_direct_io(&io, from, &iocb->ki_pos, FUSE_DIO_WRITE);
1452 fuse_invalidate_attr(inode);
1454 fuse_write_update_size(inode, iocb->ki_pos);
1455 mutex_unlock(&inode->i_mutex);
1460 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1464 for (i = 0; i < req->num_pages; i++)
1465 __free_page(req->pages[i]);
1468 fuse_file_put(req->ff, false);
1471 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1473 struct inode *inode = req->inode;
1474 struct fuse_inode *fi = get_fuse_inode(inode);
1475 struct backing_dev_info *bdi = inode_to_bdi(inode);
1478 list_del(&req->writepages_entry);
1479 for (i = 0; i < req->num_pages; i++) {
1480 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1481 dec_zone_page_state(req->pages[i], NR_WRITEBACK_TEMP);
1482 wb_writeout_inc(&bdi->wb);
1484 wake_up(&fi->page_waitq);
1487 /* Called under fc->lock, may release and reacquire it */
1488 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req,
1490 __releases(fc->lock)
1491 __acquires(fc->lock)
1493 struct fuse_inode *fi = get_fuse_inode(req->inode);
1494 struct fuse_write_in *inarg = &req->misc.write.in;
1495 __u64 data_size = req->num_pages * PAGE_CACHE_SIZE;
1500 if (inarg->offset + data_size <= size) {
1501 inarg->size = data_size;
1502 } else if (inarg->offset < size) {
1503 inarg->size = size - inarg->offset;
1505 /* Got truncated off completely */
1509 req->in.args[1].size = inarg->size;
1511 fuse_request_send_background_locked(fc, req);
1515 fuse_writepage_finish(fc, req);
1516 spin_unlock(&fc->lock);
1517 fuse_writepage_free(fc, req);
1518 fuse_put_request(fc, req);
1519 spin_lock(&fc->lock);
1523 * If fi->writectr is positive (no truncate or fsync going on) send
1524 * all queued writepage requests.
1526 * Called with fc->lock
1528 void fuse_flush_writepages(struct inode *inode)
1529 __releases(fc->lock)
1530 __acquires(fc->lock)
1532 struct fuse_conn *fc = get_fuse_conn(inode);
1533 struct fuse_inode *fi = get_fuse_inode(inode);
1534 size_t crop = i_size_read(inode);
1535 struct fuse_req *req;
1537 while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1538 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1539 list_del_init(&req->list);
1540 fuse_send_writepage(fc, req, crop);
1544 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1546 struct inode *inode = req->inode;
1547 struct fuse_inode *fi = get_fuse_inode(inode);
1549 mapping_set_error(inode->i_mapping, req->out.h.error);
1550 spin_lock(&fc->lock);
1551 while (req->misc.write.next) {
1552 struct fuse_conn *fc = get_fuse_conn(inode);
1553 struct fuse_write_in *inarg = &req->misc.write.in;
1554 struct fuse_req *next = req->misc.write.next;
1555 req->misc.write.next = next->misc.write.next;
1556 next->misc.write.next = NULL;
1557 next->ff = fuse_file_get(req->ff);
1558 list_add(&next->writepages_entry, &fi->writepages);
1561 * Skip fuse_flush_writepages() to make it easy to crop requests
1562 * based on primary request size.
1564 * 1st case (trivial): there are no concurrent activities using
1565 * fuse_set/release_nowrite. Then we're on safe side because
1566 * fuse_flush_writepages() would call fuse_send_writepage()
1569 * 2nd case: someone called fuse_set_nowrite and it is waiting
1570 * now for completion of all in-flight requests. This happens
1571 * rarely and no more than once per page, so this should be
1574 * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1575 * of fuse_set_nowrite..fuse_release_nowrite section. The fact
1576 * that fuse_set_nowrite returned implies that all in-flight
1577 * requests were completed along with all of their secondary
1578 * requests. Further primary requests are blocked by negative
1579 * writectr. Hence there cannot be any in-flight requests and
1580 * no invocations of fuse_writepage_end() while we're in
1581 * fuse_set_nowrite..fuse_release_nowrite section.
1583 fuse_send_writepage(fc, next, inarg->offset + inarg->size);
1586 fuse_writepage_finish(fc, req);
1587 spin_unlock(&fc->lock);
1588 fuse_writepage_free(fc, req);
1591 static struct fuse_file *__fuse_write_file_get(struct fuse_conn *fc,
1592 struct fuse_inode *fi)
1594 struct fuse_file *ff = NULL;
1596 spin_lock(&fc->lock);
1597 if (!list_empty(&fi->write_files)) {
1598 ff = list_entry(fi->write_files.next, struct fuse_file,
1602 spin_unlock(&fc->lock);
1607 static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1608 struct fuse_inode *fi)
1610 struct fuse_file *ff = __fuse_write_file_get(fc, fi);
1615 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1617 struct fuse_conn *fc = get_fuse_conn(inode);
1618 struct fuse_inode *fi = get_fuse_inode(inode);
1619 struct fuse_file *ff;
1622 ff = __fuse_write_file_get(fc, fi);
1623 err = fuse_flush_times(inode, ff);
1625 fuse_file_put(ff, 0);
1630 static int fuse_writepage_locked(struct page *page)
1632 struct address_space *mapping = page->mapping;
1633 struct inode *inode = mapping->host;
1634 struct fuse_conn *fc = get_fuse_conn(inode);
1635 struct fuse_inode *fi = get_fuse_inode(inode);
1636 struct fuse_req *req;
1637 struct page *tmp_page;
1638 int error = -ENOMEM;
1640 set_page_writeback(page);
1642 req = fuse_request_alloc_nofs(1);
1646 /* writeback always goes to bg_queue */
1647 __set_bit(FR_BACKGROUND, &req->flags);
1648 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1653 req->ff = fuse_write_file_get(fc, fi);
1657 fuse_write_fill(req, req->ff, page_offset(page), 0);
1659 copy_highpage(tmp_page, page);
1660 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1661 req->misc.write.next = NULL;
1662 req->in.argpages = 1;
1664 req->pages[0] = tmp_page;
1665 req->page_descs[0].offset = 0;
1666 req->page_descs[0].length = PAGE_SIZE;
1667 req->end = fuse_writepage_end;
1670 inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1671 inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1673 spin_lock(&fc->lock);
1674 list_add(&req->writepages_entry, &fi->writepages);
1675 list_add_tail(&req->list, &fi->queued_writes);
1676 fuse_flush_writepages(inode);
1677 spin_unlock(&fc->lock);
1679 end_page_writeback(page);
1684 __free_page(tmp_page);
1686 fuse_request_free(req);
1688 end_page_writeback(page);
1692 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1696 if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1698 * ->writepages() should be called for sync() and friends. We
1699 * should only get here on direct reclaim and then we are
1700 * allowed to skip a page which is already in flight
1702 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1704 redirty_page_for_writepage(wbc, page);
1708 err = fuse_writepage_locked(page);
1714 struct fuse_fill_wb_data {
1715 struct fuse_req *req;
1716 struct fuse_file *ff;
1717 struct inode *inode;
1718 struct page **orig_pages;
1721 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1723 struct fuse_req *req = data->req;
1724 struct inode *inode = data->inode;
1725 struct fuse_conn *fc = get_fuse_conn(inode);
1726 struct fuse_inode *fi = get_fuse_inode(inode);
1727 int num_pages = req->num_pages;
1730 req->ff = fuse_file_get(data->ff);
1731 spin_lock(&fc->lock);
1732 list_add_tail(&req->list, &fi->queued_writes);
1733 fuse_flush_writepages(inode);
1734 spin_unlock(&fc->lock);
1736 for (i = 0; i < num_pages; i++)
1737 end_page_writeback(data->orig_pages[i]);
1740 static bool fuse_writepage_in_flight(struct fuse_req *new_req,
1743 struct fuse_conn *fc = get_fuse_conn(new_req->inode);
1744 struct fuse_inode *fi = get_fuse_inode(new_req->inode);
1745 struct fuse_req *tmp;
1746 struct fuse_req *old_req;
1750 BUG_ON(new_req->num_pages != 0);
1752 spin_lock(&fc->lock);
1753 list_del(&new_req->writepages_entry);
1754 list_for_each_entry(old_req, &fi->writepages, writepages_entry) {
1755 BUG_ON(old_req->inode != new_req->inode);
1756 curr_index = old_req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
1757 if (curr_index <= page->index &&
1758 page->index < curr_index + old_req->num_pages) {
1764 list_add(&new_req->writepages_entry, &fi->writepages);
1768 new_req->num_pages = 1;
1769 for (tmp = old_req; tmp != NULL; tmp = tmp->misc.write.next) {
1770 BUG_ON(tmp->inode != new_req->inode);
1771 curr_index = tmp->misc.write.in.offset >> PAGE_CACHE_SHIFT;
1772 if (tmp->num_pages == 1 &&
1773 curr_index == page->index) {
1778 if (old_req->num_pages == 1 && test_bit(FR_PENDING, &old_req->flags)) {
1779 struct backing_dev_info *bdi = inode_to_bdi(page->mapping->host);
1781 copy_highpage(old_req->pages[0], page);
1782 spin_unlock(&fc->lock);
1784 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1785 dec_zone_page_state(page, NR_WRITEBACK_TEMP);
1786 wb_writeout_inc(&bdi->wb);
1787 fuse_writepage_free(fc, new_req);
1788 fuse_request_free(new_req);
1791 new_req->misc.write.next = old_req->misc.write.next;
1792 old_req->misc.write.next = new_req;
1795 spin_unlock(&fc->lock);
1800 static int fuse_writepages_fill(struct page *page,
1801 struct writeback_control *wbc, void *_data)
1803 struct fuse_fill_wb_data *data = _data;
1804 struct fuse_req *req = data->req;
1805 struct inode *inode = data->inode;
1806 struct fuse_conn *fc = get_fuse_conn(inode);
1807 struct page *tmp_page;
1813 data->ff = fuse_write_file_get(fc, get_fuse_inode(inode));
1819 * Being under writeback is unlikely but possible. For example direct
1820 * read to an mmaped fuse file will set the page dirty twice; once when
1821 * the pages are faulted with get_user_pages(), and then after the read
1824 is_writeback = fuse_page_is_writeback(inode, page->index);
1826 if (req && req->num_pages &&
1827 (is_writeback || req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
1828 (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_write ||
1829 data->orig_pages[req->num_pages - 1]->index + 1 != page->index)) {
1830 fuse_writepages_send(data);
1834 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1839 * The page must not be redirtied until the writeout is completed
1840 * (i.e. userspace has sent a reply to the write request). Otherwise
1841 * there could be more than one temporary page instance for each real
1844 * This is ensured by holding the page lock in page_mkwrite() while
1845 * checking fuse_page_is_writeback(). We already hold the page lock
1846 * since clear_page_dirty_for_io() and keep it held until we add the
1847 * request to the fi->writepages list and increment req->num_pages.
1848 * After this fuse_page_is_writeback() will indicate that the page is
1849 * under writeback, so we can release the page lock.
1851 if (data->req == NULL) {
1852 struct fuse_inode *fi = get_fuse_inode(inode);
1855 req = fuse_request_alloc_nofs(FUSE_MAX_PAGES_PER_REQ);
1857 __free_page(tmp_page);
1861 fuse_write_fill(req, data->ff, page_offset(page), 0);
1862 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1863 req->misc.write.next = NULL;
1864 req->in.argpages = 1;
1865 __set_bit(FR_BACKGROUND, &req->flags);
1867 req->end = fuse_writepage_end;
1870 spin_lock(&fc->lock);
1871 list_add(&req->writepages_entry, &fi->writepages);
1872 spin_unlock(&fc->lock);
1876 set_page_writeback(page);
1878 copy_highpage(tmp_page, page);
1879 req->pages[req->num_pages] = tmp_page;
1880 req->page_descs[req->num_pages].offset = 0;
1881 req->page_descs[req->num_pages].length = PAGE_SIZE;
1883 inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1884 inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1887 if (is_writeback && fuse_writepage_in_flight(req, page)) {
1888 end_page_writeback(page);
1892 data->orig_pages[req->num_pages] = page;
1895 * Protected by fc->lock against concurrent access by
1896 * fuse_page_is_writeback().
1898 spin_lock(&fc->lock);
1900 spin_unlock(&fc->lock);
1908 static int fuse_writepages(struct address_space *mapping,
1909 struct writeback_control *wbc)
1911 struct inode *inode = mapping->host;
1912 struct fuse_fill_wb_data data;
1916 if (is_bad_inode(inode))
1924 data.orig_pages = kcalloc(FUSE_MAX_PAGES_PER_REQ,
1925 sizeof(struct page *),
1927 if (!data.orig_pages)
1930 err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
1932 /* Ignore errors if we can write at least one page */
1933 BUG_ON(!data.req->num_pages);
1934 fuse_writepages_send(&data);
1938 fuse_file_put(data.ff, false);
1940 kfree(data.orig_pages);
1946 * It's worthy to make sure that space is reserved on disk for the write,
1947 * but how to implement it without killing performance need more thinking.
1949 static int fuse_write_begin(struct file *file, struct address_space *mapping,
1950 loff_t pos, unsigned len, unsigned flags,
1951 struct page **pagep, void **fsdata)
1953 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1954 struct fuse_conn *fc = get_fuse_conn(file_inode(file));
1959 WARN_ON(!fc->writeback_cache);
1961 page = grab_cache_page_write_begin(mapping, index, flags);
1965 fuse_wait_on_page_writeback(mapping->host, page->index);
1967 if (PageUptodate(page) || len == PAGE_CACHE_SIZE)
1970 * Check if the start this page comes after the end of file, in which
1971 * case the readpage can be optimized away.
1973 fsize = i_size_read(mapping->host);
1974 if (fsize <= (pos & PAGE_CACHE_MASK)) {
1975 size_t off = pos & ~PAGE_CACHE_MASK;
1977 zero_user_segment(page, 0, off);
1980 err = fuse_do_readpage(file, page);
1989 page_cache_release(page);
1994 static int fuse_write_end(struct file *file, struct address_space *mapping,
1995 loff_t pos, unsigned len, unsigned copied,
1996 struct page *page, void *fsdata)
1998 struct inode *inode = page->mapping->host;
2000 /* Haven't copied anything? Skip zeroing, size extending, dirtying. */
2004 if (!PageUptodate(page)) {
2005 /* Zero any unwritten bytes at the end of the page */
2006 size_t endoff = (pos + copied) & ~PAGE_CACHE_MASK;
2008 zero_user_segment(page, endoff, PAGE_CACHE_SIZE);
2009 SetPageUptodate(page);
2012 fuse_write_update_size(inode, pos + copied);
2013 set_page_dirty(page);
2017 page_cache_release(page);
2022 static int fuse_launder_page(struct page *page)
2025 if (clear_page_dirty_for_io(page)) {
2026 struct inode *inode = page->mapping->host;
2027 err = fuse_writepage_locked(page);
2029 fuse_wait_on_page_writeback(inode, page->index);
2035 * Write back dirty pages now, because there may not be any suitable
2038 static void fuse_vma_close(struct vm_area_struct *vma)
2040 filemap_write_and_wait(vma->vm_file->f_mapping);
2044 * Wait for writeback against this page to complete before allowing it
2045 * to be marked dirty again, and hence written back again, possibly
2046 * before the previous writepage completed.
2048 * Block here, instead of in ->writepage(), so that the userspace fs
2049 * can only block processes actually operating on the filesystem.
2051 * Otherwise unprivileged userspace fs would be able to block
2056 * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2058 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
2060 struct page *page = vmf->page;
2061 struct inode *inode = file_inode(vma->vm_file);
2063 file_update_time(vma->vm_file);
2065 if (page->mapping != inode->i_mapping) {
2067 return VM_FAULT_NOPAGE;
2070 fuse_wait_on_page_writeback(inode, page->index);
2071 return VM_FAULT_LOCKED;
2074 static const struct vm_operations_struct fuse_file_vm_ops = {
2075 .close = fuse_vma_close,
2076 .fault = filemap_fault,
2077 .map_pages = filemap_map_pages,
2078 .page_mkwrite = fuse_page_mkwrite,
2081 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2083 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2084 fuse_link_write_file(file);
2086 file_accessed(file);
2087 vma->vm_ops = &fuse_file_vm_ops;
2091 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
2093 /* Can't provide the coherency needed for MAP_SHARED */
2094 if (vma->vm_flags & VM_MAYSHARE)
2097 invalidate_inode_pages2(file->f_mapping);
2099 return generic_file_mmap(file, vma);
2102 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
2103 struct file_lock *fl)
2105 switch (ffl->type) {
2111 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2112 ffl->end < ffl->start)
2115 fl->fl_start = ffl->start;
2116 fl->fl_end = ffl->end;
2117 fl->fl_pid = ffl->pid;
2123 fl->fl_type = ffl->type;
2127 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2128 const struct file_lock *fl, int opcode, pid_t pid,
2129 int flock, struct fuse_lk_in *inarg)
2131 struct inode *inode = file_inode(file);
2132 struct fuse_conn *fc = get_fuse_conn(inode);
2133 struct fuse_file *ff = file->private_data;
2135 memset(inarg, 0, sizeof(*inarg));
2137 inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2138 inarg->lk.start = fl->fl_start;
2139 inarg->lk.end = fl->fl_end;
2140 inarg->lk.type = fl->fl_type;
2141 inarg->lk.pid = pid;
2143 inarg->lk_flags |= FUSE_LK_FLOCK;
2144 args->in.h.opcode = opcode;
2145 args->in.h.nodeid = get_node_id(inode);
2146 args->in.numargs = 1;
2147 args->in.args[0].size = sizeof(*inarg);
2148 args->in.args[0].value = inarg;
2151 static int fuse_getlk(struct file *file, struct file_lock *fl)
2153 struct inode *inode = file_inode(file);
2154 struct fuse_conn *fc = get_fuse_conn(inode);
2156 struct fuse_lk_in inarg;
2157 struct fuse_lk_out outarg;
2160 fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2161 args.out.numargs = 1;
2162 args.out.args[0].size = sizeof(outarg);
2163 args.out.args[0].value = &outarg;
2164 err = fuse_simple_request(fc, &args);
2166 err = convert_fuse_file_lock(&outarg.lk, fl);
2171 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2173 struct inode *inode = file_inode(file);
2174 struct fuse_conn *fc = get_fuse_conn(inode);
2176 struct fuse_lk_in inarg;
2177 int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2178 pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
2181 if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2182 /* NLM needs asynchronous locks, which we don't support yet */
2186 /* Unlock on close is handled by the flush method */
2187 if (fl->fl_flags & FL_CLOSE)
2190 fuse_lk_fill(&args, file, fl, opcode, pid, flock, &inarg);
2191 err = fuse_simple_request(fc, &args);
2193 /* locking is restartable */
2200 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2202 struct inode *inode = file_inode(file);
2203 struct fuse_conn *fc = get_fuse_conn(inode);
2206 if (cmd == F_CANCELLK) {
2208 } else if (cmd == F_GETLK) {
2210 posix_test_lock(file, fl);
2213 err = fuse_getlk(file, fl);
2216 err = posix_lock_file(file, fl, NULL);
2218 err = fuse_setlk(file, fl, 0);
2223 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2225 struct inode *inode = file_inode(file);
2226 struct fuse_conn *fc = get_fuse_conn(inode);
2230 err = locks_lock_file_wait(file, fl);
2232 struct fuse_file *ff = file->private_data;
2234 /* emulate flock with POSIX locks */
2236 err = fuse_setlk(file, fl, 1);
2242 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2244 struct inode *inode = mapping->host;
2245 struct fuse_conn *fc = get_fuse_conn(inode);
2247 struct fuse_bmap_in inarg;
2248 struct fuse_bmap_out outarg;
2251 if (!inode->i_sb->s_bdev || fc->no_bmap)
2254 memset(&inarg, 0, sizeof(inarg));
2255 inarg.block = block;
2256 inarg.blocksize = inode->i_sb->s_blocksize;
2257 args.in.h.opcode = FUSE_BMAP;
2258 args.in.h.nodeid = get_node_id(inode);
2259 args.in.numargs = 1;
2260 args.in.args[0].size = sizeof(inarg);
2261 args.in.args[0].value = &inarg;
2262 args.out.numargs = 1;
2263 args.out.args[0].size = sizeof(outarg);
2264 args.out.args[0].value = &outarg;
2265 err = fuse_simple_request(fc, &args);
2269 return err ? 0 : outarg.block;
2272 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2275 struct inode *inode = file_inode(file);
2277 /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2278 if (whence == SEEK_CUR || whence == SEEK_SET)
2279 return generic_file_llseek(file, offset, whence);
2281 mutex_lock(&inode->i_mutex);
2282 retval = fuse_update_attributes(inode, NULL, file, NULL);
2284 retval = generic_file_llseek(file, offset, whence);
2285 mutex_unlock(&inode->i_mutex);
2290 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
2291 unsigned int nr_segs, size_t bytes, bool to_user)
2299 iov_iter_init(&ii, to_user ? READ : WRITE, iov, nr_segs, bytes);
2301 while (iov_iter_count(&ii)) {
2302 struct page *page = pages[page_idx++];
2303 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
2309 char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
2310 size_t iov_len = ii.iov->iov_len - ii.iov_offset;
2311 size_t copy = min(todo, iov_len);
2315 left = copy_from_user(kaddr, uaddr, copy);
2317 left = copy_to_user(uaddr, kaddr, copy);
2322 iov_iter_advance(&ii, copy);
2334 * CUSE servers compiled on 32bit broke on 64bit kernels because the
2335 * ABI was defined to be 'struct iovec' which is different on 32bit
2336 * and 64bit. Fortunately we can determine which structure the server
2337 * used from the size of the reply.
2339 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2340 size_t transferred, unsigned count,
2343 #ifdef CONFIG_COMPAT
2344 if (count * sizeof(struct compat_iovec) == transferred) {
2345 struct compat_iovec *ciov = src;
2349 * With this interface a 32bit server cannot support
2350 * non-compat (i.e. ones coming from 64bit apps) ioctl
2356 for (i = 0; i < count; i++) {
2357 dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2358 dst[i].iov_len = ciov[i].iov_len;
2364 if (count * sizeof(struct iovec) != transferred)
2367 memcpy(dst, src, transferred);
2371 /* Make sure iov_length() won't overflow */
2372 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
2375 u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
2377 for (n = 0; n < count; n++, iov++) {
2378 if (iov->iov_len > (size_t) max)
2380 max -= iov->iov_len;
2385 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2386 void *src, size_t transferred, unsigned count,
2390 struct fuse_ioctl_iovec *fiov = src;
2392 if (fc->minor < 16) {
2393 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2397 if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2400 for (i = 0; i < count; i++) {
2401 /* Did the server supply an inappropriate value? */
2402 if (fiov[i].base != (unsigned long) fiov[i].base ||
2403 fiov[i].len != (unsigned long) fiov[i].len)
2406 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2407 dst[i].iov_len = (size_t) fiov[i].len;
2409 #ifdef CONFIG_COMPAT
2411 (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2412 (compat_size_t) dst[i].iov_len != fiov[i].len))
2422 * For ioctls, there is no generic way to determine how much memory
2423 * needs to be read and/or written. Furthermore, ioctls are allowed
2424 * to dereference the passed pointer, so the parameter requires deep
2425 * copying but FUSE has no idea whatsoever about what to copy in or
2428 * This is solved by allowing FUSE server to retry ioctl with
2429 * necessary in/out iovecs. Let's assume the ioctl implementation
2430 * needs to read in the following structure.
2437 * On the first callout to FUSE server, inarg->in_size and
2438 * inarg->out_size will be NULL; then, the server completes the ioctl
2439 * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2440 * the actual iov array to
2442 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) } }
2444 * which tells FUSE to copy in the requested area and retry the ioctl.
2445 * On the second round, the server has access to the structure and
2446 * from that it can tell what to look for next, so on the invocation,
2447 * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2449 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) },
2450 * { .iov_base = a.buf, .iov_len = a.buflen } }
2452 * FUSE will copy both struct a and the pointed buffer from the
2453 * process doing the ioctl and retry ioctl with both struct a and the
2456 * This time, FUSE server has everything it needs and completes ioctl
2457 * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2459 * Copying data out works the same way.
2461 * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2462 * automatically initializes in and out iovs by decoding @cmd with
2463 * _IOC_* macros and the server is not allowed to request RETRY. This
2464 * limits ioctl data transfers to well-formed ioctls and is the forced
2465 * behavior for all FUSE servers.
2467 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2470 struct fuse_file *ff = file->private_data;
2471 struct fuse_conn *fc = ff->fc;
2472 struct fuse_ioctl_in inarg = {
2478 struct fuse_ioctl_out outarg;
2479 struct fuse_req *req = NULL;
2480 struct page **pages = NULL;
2481 struct iovec *iov_page = NULL;
2482 struct iovec *in_iov = NULL, *out_iov = NULL;
2483 unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
2484 size_t in_size, out_size, transferred;
2487 #if BITS_PER_LONG == 32
2488 inarg.flags |= FUSE_IOCTL_32BIT;
2490 if (flags & FUSE_IOCTL_COMPAT)
2491 inarg.flags |= FUSE_IOCTL_32BIT;
2494 /* assume all the iovs returned by client always fits in a page */
2495 BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2498 pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
2499 iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2500 if (!pages || !iov_page)
2504 * If restricted, initialize IO parameters as encoded in @cmd.
2505 * RETRY from server is not allowed.
2507 if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2508 struct iovec *iov = iov_page;
2510 iov->iov_base = (void __user *)arg;
2511 iov->iov_len = _IOC_SIZE(cmd);
2513 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2518 if (_IOC_DIR(cmd) & _IOC_READ) {
2525 inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2526 inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2529 * Out data can be used either for actual out data or iovs,
2530 * make sure there always is at least one page.
2532 out_size = max_t(size_t, out_size, PAGE_SIZE);
2533 max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2535 /* make sure there are enough buffer pages and init request with them */
2537 if (max_pages > FUSE_MAX_PAGES_PER_REQ)
2539 while (num_pages < max_pages) {
2540 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2541 if (!pages[num_pages])
2546 req = fuse_get_req(fc, num_pages);
2552 memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
2553 req->num_pages = num_pages;
2554 fuse_page_descs_length_init(req, 0, req->num_pages);
2556 /* okay, let's send it to the client */
2557 req->in.h.opcode = FUSE_IOCTL;
2558 req->in.h.nodeid = ff->nodeid;
2559 req->in.numargs = 1;
2560 req->in.args[0].size = sizeof(inarg);
2561 req->in.args[0].value = &inarg;
2564 req->in.args[1].size = in_size;
2565 req->in.argpages = 1;
2567 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
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)
2636 err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
2639 fuse_put_request(fc, req);
2640 free_page((unsigned long) iov_page);
2642 __free_page(pages[--num_pages]);
2645 return err ? err : outarg.result;
2647 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2649 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2650 unsigned long arg, unsigned int flags)
2652 struct inode *inode = file_inode(file);
2653 struct fuse_conn *fc = get_fuse_conn(inode);
2655 if (!fuse_allow_current_process(fc))
2658 if (is_bad_inode(inode))
2661 return fuse_do_ioctl(file, cmd, arg, flags);
2664 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2667 return fuse_ioctl_common(file, cmd, arg, 0);
2670 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2673 return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2677 * All files which have been polled are linked to RB tree
2678 * fuse_conn->polled_files which is indexed by kh. Walk the tree and
2679 * find the matching one.
2681 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2682 struct rb_node **parent_out)
2684 struct rb_node **link = &fc->polled_files.rb_node;
2685 struct rb_node *last = NULL;
2688 struct fuse_file *ff;
2691 ff = rb_entry(last, struct fuse_file, polled_node);
2694 link = &last->rb_left;
2695 else if (kh > ff->kh)
2696 link = &last->rb_right;
2707 * The file is about to be polled. Make sure it's on the polled_files
2708 * RB tree. Note that files once added to the polled_files tree are
2709 * not removed before the file is released. This is because a file
2710 * polled once is likely to be polled again.
2712 static void fuse_register_polled_file(struct fuse_conn *fc,
2713 struct fuse_file *ff)
2715 spin_lock(&fc->lock);
2716 if (RB_EMPTY_NODE(&ff->polled_node)) {
2717 struct rb_node **link, *uninitialized_var(parent);
2719 link = fuse_find_polled_node(fc, ff->kh, &parent);
2721 rb_link_node(&ff->polled_node, parent, link);
2722 rb_insert_color(&ff->polled_node, &fc->polled_files);
2724 spin_unlock(&fc->lock);
2727 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2729 struct fuse_file *ff = file->private_data;
2730 struct fuse_conn *fc = ff->fc;
2731 struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2732 struct fuse_poll_out outarg;
2737 return DEFAULT_POLLMASK;
2739 poll_wait(file, &ff->poll_wait, wait);
2740 inarg.events = (__u32)poll_requested_events(wait);
2743 * Ask for notification iff there's someone waiting for it.
2744 * The client may ignore the flag and always notify.
2746 if (waitqueue_active(&ff->poll_wait)) {
2747 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2748 fuse_register_polled_file(fc, ff);
2751 args.in.h.opcode = FUSE_POLL;
2752 args.in.h.nodeid = ff->nodeid;
2753 args.in.numargs = 1;
2754 args.in.args[0].size = sizeof(inarg);
2755 args.in.args[0].value = &inarg;
2756 args.out.numargs = 1;
2757 args.out.args[0].size = sizeof(outarg);
2758 args.out.args[0].value = &outarg;
2759 err = fuse_simple_request(fc, &args);
2762 return outarg.revents;
2763 if (err == -ENOSYS) {
2765 return DEFAULT_POLLMASK;
2769 EXPORT_SYMBOL_GPL(fuse_file_poll);
2772 * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2773 * wakes up the poll waiters.
2775 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2776 struct fuse_notify_poll_wakeup_out *outarg)
2778 u64 kh = outarg->kh;
2779 struct rb_node **link;
2781 spin_lock(&fc->lock);
2783 link = fuse_find_polled_node(fc, kh, NULL);
2785 struct fuse_file *ff;
2787 ff = rb_entry(*link, struct fuse_file, polled_node);
2788 wake_up_interruptible_sync(&ff->poll_wait);
2791 spin_unlock(&fc->lock);
2795 static void fuse_do_truncate(struct file *file)
2797 struct inode *inode = file->f_mapping->host;
2800 attr.ia_valid = ATTR_SIZE;
2801 attr.ia_size = i_size_read(inode);
2803 attr.ia_file = file;
2804 attr.ia_valid |= ATTR_FILE;
2806 fuse_do_setattr(inode, &attr, file);
2809 static inline loff_t fuse_round_up(loff_t off)
2811 return round_up(off, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
2815 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter, loff_t offset)
2817 DECLARE_COMPLETION_ONSTACK(wait);
2819 struct file *file = iocb->ki_filp;
2820 struct fuse_file *ff = file->private_data;
2821 bool async_dio = ff->fc->async_dio;
2823 struct inode *inode;
2825 size_t count = iov_iter_count(iter);
2826 struct fuse_io_priv *io;
2827 bool is_sync = is_sync_kiocb(iocb);
2830 inode = file->f_mapping->host;
2831 i_size = i_size_read(inode);
2833 if ((iov_iter_rw(iter) == READ) && (offset > i_size))
2836 /* optimization for short read */
2837 if (async_dio && iov_iter_rw(iter) != WRITE && offset + count > i_size) {
2838 if (offset >= i_size)
2840 iov_iter_truncate(iter, fuse_round_up(i_size - offset));
2841 count = iov_iter_count(iter);
2844 io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2847 spin_lock_init(&io->lock);
2848 kref_init(&io->refcnt);
2852 io->offset = offset;
2853 io->write = (iov_iter_rw(iter) == WRITE);
2857 * By default, we want to optimize all I/Os with async request
2858 * submission to the client filesystem if supported.
2860 io->async = async_dio;
2864 * We cannot asynchronously extend the size of a file. We have no method
2865 * to wait on real async I/O requests, so we must submit this request
2868 if (!is_sync && (offset + count > i_size) &&
2869 iov_iter_rw(iter) == WRITE)
2872 if (io->async && is_sync) {
2874 * Additional reference to keep io around after
2875 * calling fuse_aio_complete()
2877 kref_get(&io->refcnt);
2881 if (iov_iter_rw(iter) == WRITE) {
2882 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
2883 fuse_invalidate_attr(inode);
2885 ret = __fuse_direct_read(io, iter, &pos);
2889 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2891 /* we have a non-extending, async request, so return */
2893 return -EIOCBQUEUED;
2895 wait_for_completion(&wait);
2896 ret = fuse_get_res_by_io(io);
2899 kref_put(&io->refcnt, fuse_io_release);
2901 if (iov_iter_rw(iter) == WRITE) {
2903 fuse_write_update_size(inode, pos);
2904 else if (ret < 0 && offset + count > i_size)
2905 fuse_do_truncate(file);
2911 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2914 struct fuse_file *ff = file->private_data;
2915 struct inode *inode = file_inode(file);
2916 struct fuse_inode *fi = get_fuse_inode(inode);
2917 struct fuse_conn *fc = ff->fc;
2919 struct fuse_fallocate_in inarg = {
2926 bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
2927 (mode & FALLOC_FL_PUNCH_HOLE);
2929 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2932 if (fc->no_fallocate)
2936 mutex_lock(&inode->i_mutex);
2937 if (mode & FALLOC_FL_PUNCH_HOLE) {
2938 loff_t endbyte = offset + length - 1;
2939 err = filemap_write_and_wait_range(inode->i_mapping,
2944 fuse_sync_writes(inode);
2948 if (!(mode & FALLOC_FL_KEEP_SIZE))
2949 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2951 args.in.h.opcode = FUSE_FALLOCATE;
2952 args.in.h.nodeid = ff->nodeid;
2953 args.in.numargs = 1;
2954 args.in.args[0].size = sizeof(inarg);
2955 args.in.args[0].value = &inarg;
2956 err = fuse_simple_request(fc, &args);
2957 if (err == -ENOSYS) {
2958 fc->no_fallocate = 1;
2964 /* we could have extended the file */
2965 if (!(mode & FALLOC_FL_KEEP_SIZE)) {
2966 bool changed = fuse_write_update_size(inode, offset + length);
2968 if (changed && fc->writeback_cache)
2969 file_update_time(file);
2972 if (mode & FALLOC_FL_PUNCH_HOLE)
2973 truncate_pagecache_range(inode, offset, offset + length - 1);
2975 fuse_invalidate_attr(inode);
2978 if (!(mode & FALLOC_FL_KEEP_SIZE))
2979 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2982 mutex_unlock(&inode->i_mutex);
2987 static const struct file_operations fuse_file_operations = {
2988 .llseek = fuse_file_llseek,
2989 .read_iter = fuse_file_read_iter,
2990 .write_iter = fuse_file_write_iter,
2991 .mmap = fuse_file_mmap,
2993 .flush = fuse_flush,
2994 .release = fuse_release,
2995 .fsync = fuse_fsync,
2996 .lock = fuse_file_lock,
2997 .flock = fuse_file_flock,
2998 .splice_read = generic_file_splice_read,
2999 .unlocked_ioctl = fuse_file_ioctl,
3000 .compat_ioctl = fuse_file_compat_ioctl,
3001 .poll = fuse_file_poll,
3002 .fallocate = fuse_file_fallocate,
3005 static const struct file_operations fuse_direct_io_file_operations = {
3006 .llseek = fuse_file_llseek,
3007 .read_iter = fuse_direct_read_iter,
3008 .write_iter = fuse_direct_write_iter,
3009 .mmap = fuse_direct_mmap,
3011 .flush = fuse_flush,
3012 .release = fuse_release,
3013 .fsync = fuse_fsync,
3014 .lock = fuse_file_lock,
3015 .flock = fuse_file_flock,
3016 .unlocked_ioctl = fuse_file_ioctl,
3017 .compat_ioctl = fuse_file_compat_ioctl,
3018 .poll = fuse_file_poll,
3019 .fallocate = fuse_file_fallocate,
3020 /* no splice_read */
3023 static const struct address_space_operations fuse_file_aops = {
3024 .readpage = fuse_readpage,
3025 .writepage = fuse_writepage,
3026 .writepages = fuse_writepages,
3027 .launder_page = fuse_launder_page,
3028 .readpages = fuse_readpages,
3029 .set_page_dirty = __set_page_dirty_nobuffers,
3031 .direct_IO = fuse_direct_IO,
3032 .write_begin = fuse_write_begin,
3033 .write_end = fuse_write_end,
3036 void fuse_init_file_inode(struct inode *inode)
3038 inode->i_fop = &fuse_file_operations;
3039 inode->i_data.a_ops = &fuse_file_aops;