2 FUSE: Filesystem in Userspace
3 Copyright (C) 2001-2008 Miklos Szeredi <miklos@szeredi.hu>
5 This program can be distributed under the terms of the GNU GPL.
11 #include <linux/pagemap.h>
12 #include <linux/slab.h>
13 #include <linux/kernel.h>
14 #include <linux/sched.h>
15 #include <linux/module.h>
16 #include <linux/compat.h>
17 #include <linux/swap.h>
18 #include <linux/falloc.h>
19 #include <linux/uio.h>
21 static const struct file_operations fuse_direct_io_file_operations;
23 static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
24 int opcode, struct fuse_open_out *outargp)
26 struct fuse_open_in inarg;
29 memset(&inarg, 0, sizeof(inarg));
30 inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
31 if (!fc->atomic_o_trunc)
32 inarg.flags &= ~O_TRUNC;
33 args.in.h.opcode = opcode;
34 args.in.h.nodeid = nodeid;
36 args.in.args[0].size = sizeof(inarg);
37 args.in.args[0].value = &inarg;
39 args.out.args[0].size = sizeof(*outargp);
40 args.out.args[0].value = outargp;
42 return fuse_simple_request(fc, &args);
45 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
49 ff = kmalloc(sizeof(struct fuse_file), GFP_KERNEL);
54 ff->reserved_req = fuse_request_alloc(0);
55 if (unlikely(!ff->reserved_req)) {
60 INIT_LIST_HEAD(&ff->write_entry);
61 atomic_set(&ff->count, 0);
62 RB_CLEAR_NODE(&ff->polled_node);
63 init_waitqueue_head(&ff->poll_wait);
67 spin_unlock(&fc->lock);
72 void fuse_file_free(struct fuse_file *ff)
74 fuse_request_free(ff->reserved_req);
78 struct fuse_file *fuse_file_get(struct fuse_file *ff)
80 atomic_inc(&ff->count);
84 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
86 iput(req->misc.release.inode);
89 static void fuse_file_put(struct fuse_file *ff, bool sync)
91 if (atomic_dec_and_test(&ff->count)) {
92 struct fuse_req *req = ff->reserved_req;
94 if (ff->fc->no_open) {
96 * Drop the release request when client does not
99 __clear_bit(FR_BACKGROUND, &req->flags);
100 iput(req->misc.release.inode);
101 fuse_put_request(ff->fc, req);
103 __clear_bit(FR_BACKGROUND, &req->flags);
104 fuse_request_send(ff->fc, req);
105 iput(req->misc.release.inode);
106 fuse_put_request(ff->fc, req);
108 req->end = fuse_release_end;
109 __set_bit(FR_BACKGROUND, &req->flags);
110 fuse_request_send_background(ff->fc, req);
116 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
119 struct fuse_file *ff;
120 int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
122 ff = fuse_file_alloc(fc);
127 ff->open_flags = FOPEN_KEEP_CACHE; /* Default for no-open */
128 if (!fc->no_open || isdir) {
129 struct fuse_open_out outarg;
132 err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
135 ff->open_flags = outarg.open_flags;
137 } else if (err != -ENOSYS || isdir) {
146 ff->open_flags &= ~FOPEN_DIRECT_IO;
149 file->private_data = fuse_file_get(ff);
153 EXPORT_SYMBOL_GPL(fuse_do_open);
155 static void fuse_link_write_file(struct file *file)
157 struct inode *inode = file_inode(file);
158 struct fuse_conn *fc = get_fuse_conn(inode);
159 struct fuse_inode *fi = get_fuse_inode(inode);
160 struct fuse_file *ff = file->private_data;
162 * file may be written through mmap, so chain it onto the
163 * inodes's write_file list
165 spin_lock(&fc->lock);
166 if (list_empty(&ff->write_entry))
167 list_add(&ff->write_entry, &fi->write_files);
168 spin_unlock(&fc->lock);
171 void fuse_finish_open(struct inode *inode, struct file *file)
173 struct fuse_file *ff = file->private_data;
174 struct fuse_conn *fc = get_fuse_conn(inode);
176 if (ff->open_flags & FOPEN_DIRECT_IO)
177 file->f_op = &fuse_direct_io_file_operations;
178 if (!(ff->open_flags & FOPEN_KEEP_CACHE))
179 invalidate_inode_pages2(inode->i_mapping);
180 if (ff->open_flags & FOPEN_NONSEEKABLE)
181 nonseekable_open(inode, file);
182 if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
183 struct fuse_inode *fi = get_fuse_inode(inode);
185 spin_lock(&fc->lock);
186 fi->attr_version = ++fc->attr_version;
187 i_size_write(inode, 0);
188 spin_unlock(&fc->lock);
189 fuse_invalidate_attr(inode);
190 if (fc->writeback_cache)
191 file_update_time(file);
193 if ((file->f_mode & FMODE_WRITE) && fc->writeback_cache)
194 fuse_link_write_file(file);
197 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
199 struct fuse_conn *fc = get_fuse_conn(inode);
201 bool lock_inode = (file->f_flags & O_TRUNC) &&
202 fc->atomic_o_trunc &&
205 err = generic_file_open(inode, file);
210 mutex_lock(&inode->i_mutex);
212 err = fuse_do_open(fc, get_node_id(inode), file, isdir);
215 fuse_finish_open(inode, file);
218 mutex_unlock(&inode->i_mutex);
223 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
225 struct fuse_conn *fc = ff->fc;
226 struct fuse_req *req = ff->reserved_req;
227 struct fuse_release_in *inarg = &req->misc.release.in;
229 spin_lock(&fc->lock);
230 list_del(&ff->write_entry);
231 if (!RB_EMPTY_NODE(&ff->polled_node))
232 rb_erase(&ff->polled_node, &fc->polled_files);
233 spin_unlock(&fc->lock);
235 wake_up_interruptible_all(&ff->poll_wait);
238 inarg->flags = flags;
239 req->in.h.opcode = opcode;
240 req->in.h.nodeid = ff->nodeid;
242 req->in.args[0].size = sizeof(struct fuse_release_in);
243 req->in.args[0].value = inarg;
246 void fuse_release_common(struct file *file, int opcode)
248 struct fuse_file *ff;
249 struct fuse_req *req;
251 ff = file->private_data;
255 req = ff->reserved_req;
256 fuse_prepare_release(ff, file->f_flags, opcode);
259 struct fuse_release_in *inarg = &req->misc.release.in;
260 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
261 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
264 /* Hold inode until release is finished */
265 req->misc.release.inode = igrab(file_inode(file));
268 * Normally this will send the RELEASE request, however if
269 * some asynchronous READ or WRITE requests are outstanding,
270 * the sending will be delayed.
272 * Make the release synchronous if this is a fuseblk mount,
273 * synchronous RELEASE is allowed (and desirable) in this case
274 * because the server can be trusted not to screw up.
276 fuse_file_put(ff, ff->fc->destroy_req != NULL);
279 static int fuse_open(struct inode *inode, struct file *file)
281 return fuse_open_common(inode, file, false);
284 static int fuse_release(struct inode *inode, struct file *file)
286 struct fuse_conn *fc = get_fuse_conn(inode);
288 /* see fuse_vma_close() for !writeback_cache case */
289 if (fc->writeback_cache)
290 write_inode_now(inode, 1);
292 fuse_release_common(file, FUSE_RELEASE);
294 /* return value is ignored by VFS */
298 void fuse_sync_release(struct fuse_file *ff, int flags)
300 WARN_ON(atomic_read(&ff->count) > 1);
301 fuse_prepare_release(ff, flags, FUSE_RELEASE);
302 __set_bit(FR_FORCE, &ff->reserved_req->flags);
303 __clear_bit(FR_BACKGROUND, &ff->reserved_req->flags);
304 fuse_request_send(ff->fc, ff->reserved_req);
305 fuse_put_request(ff->fc, ff->reserved_req);
308 EXPORT_SYMBOL_GPL(fuse_sync_release);
311 * Scramble the ID space with XTEA, so that the value of the files_struct
312 * pointer is not exposed to userspace.
314 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
316 u32 *k = fc->scramble_key;
317 u64 v = (unsigned long) id;
323 for (i = 0; i < 32; i++) {
324 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
326 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
329 return (u64) v0 + ((u64) v1 << 32);
333 * Check if any page in a range is under writeback
335 * This is currently done by walking the list of writepage requests
336 * for the inode, which can be pretty inefficient.
338 static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
341 struct fuse_conn *fc = get_fuse_conn(inode);
342 struct fuse_inode *fi = get_fuse_inode(inode);
343 struct fuse_req *req;
346 spin_lock(&fc->lock);
347 list_for_each_entry(req, &fi->writepages, writepages_entry) {
350 BUG_ON(req->inode != inode);
351 curr_index = req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
352 if (idx_from < curr_index + req->num_pages &&
353 curr_index <= idx_to) {
358 spin_unlock(&fc->lock);
363 static inline bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
365 return fuse_range_is_writeback(inode, index, index);
369 * Wait for page writeback to be completed.
371 * Since fuse doesn't rely on the VM writeback tracking, this has to
372 * use some other means.
374 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
376 struct fuse_inode *fi = get_fuse_inode(inode);
378 wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
383 * Wait for all pending writepages on the inode to finish.
385 * This is currently done by blocking further writes with FUSE_NOWRITE
386 * and waiting for all sent writes to complete.
388 * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
389 * could conflict with truncation.
391 static void fuse_sync_writes(struct inode *inode)
393 fuse_set_nowrite(inode);
394 fuse_release_nowrite(inode);
397 static int fuse_flush(struct file *file, fl_owner_t id)
399 struct inode *inode = file_inode(file);
400 struct fuse_conn *fc = get_fuse_conn(inode);
401 struct fuse_file *ff = file->private_data;
402 struct fuse_req *req;
403 struct fuse_flush_in inarg;
406 if (is_bad_inode(inode))
412 err = write_inode_now(inode, 1);
416 mutex_lock(&inode->i_mutex);
417 fuse_sync_writes(inode);
418 mutex_unlock(&inode->i_mutex);
420 req = fuse_get_req_nofail_nopages(fc, file);
421 memset(&inarg, 0, sizeof(inarg));
423 inarg.lock_owner = fuse_lock_owner_id(fc, id);
424 req->in.h.opcode = FUSE_FLUSH;
425 req->in.h.nodeid = get_node_id(inode);
427 req->in.args[0].size = sizeof(inarg);
428 req->in.args[0].value = &inarg;
429 __set_bit(FR_FORCE, &req->flags);
430 fuse_request_send(fc, req);
431 err = req->out.h.error;
432 fuse_put_request(fc, req);
433 if (err == -ENOSYS) {
440 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
441 int datasync, int isdir)
443 struct inode *inode = file->f_mapping->host;
444 struct fuse_conn *fc = get_fuse_conn(inode);
445 struct fuse_file *ff = file->private_data;
447 struct fuse_fsync_in inarg;
450 if (is_bad_inode(inode))
453 mutex_lock(&inode->i_mutex);
456 * Start writeback against all dirty pages of the inode, then
457 * wait for all outstanding writes, before sending the FSYNC
460 err = filemap_write_and_wait_range(inode->i_mapping, start, end);
464 fuse_sync_writes(inode);
467 * Due to implementation of fuse writeback
468 * filemap_write_and_wait_range() does not catch errors.
469 * We have to do this directly after fuse_sync_writes()
471 if (test_bit(AS_ENOSPC, &file->f_mapping->flags) &&
472 test_and_clear_bit(AS_ENOSPC, &file->f_mapping->flags))
474 if (test_bit(AS_EIO, &file->f_mapping->flags) &&
475 test_and_clear_bit(AS_EIO, &file->f_mapping->flags))
480 err = sync_inode_metadata(inode, 1);
484 if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
487 memset(&inarg, 0, sizeof(inarg));
489 inarg.fsync_flags = datasync ? 1 : 0;
490 args.in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
491 args.in.h.nodeid = get_node_id(inode);
493 args.in.args[0].size = sizeof(inarg);
494 args.in.args[0].value = &inarg;
495 err = fuse_simple_request(fc, &args);
496 if (err == -ENOSYS) {
504 mutex_unlock(&inode->i_mutex);
508 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
511 return fuse_fsync_common(file, start, end, datasync, 0);
514 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
515 size_t count, int opcode)
517 struct fuse_read_in *inarg = &req->misc.read.in;
518 struct fuse_file *ff = file->private_data;
523 inarg->flags = file->f_flags;
524 req->in.h.opcode = opcode;
525 req->in.h.nodeid = ff->nodeid;
527 req->in.args[0].size = sizeof(struct fuse_read_in);
528 req->in.args[0].value = inarg;
530 req->out.numargs = 1;
531 req->out.args[0].size = count;
534 static void fuse_release_user_pages(struct fuse_req *req, int write)
538 for (i = 0; i < req->num_pages; i++) {
539 struct page *page = req->pages[i];
541 set_page_dirty_lock(page);
546 static void fuse_io_release(struct kref *kref)
548 kfree(container_of(kref, struct fuse_io_priv, refcnt));
551 static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
556 if (io->bytes >= 0 && io->write)
559 return io->bytes < 0 ? io->size : io->bytes;
563 * In case of short read, the caller sets 'pos' to the position of
564 * actual end of fuse request in IO request. Otherwise, if bytes_requested
565 * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
568 * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
569 * both submitted asynchronously. The first of them was ACKed by userspace as
570 * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
571 * second request was ACKed as short, e.g. only 1K was read, resulting in
574 * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
575 * will be equal to the length of the longest contiguous fragment of
576 * transferred data starting from the beginning of IO request.
578 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
580 bool is_sync = is_sync_kiocb(io->iocb);
583 spin_lock(&io->lock);
585 io->err = io->err ? : err;
586 else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
590 if (!left && is_sync)
592 spin_unlock(&io->lock);
594 if (!left && !is_sync) {
595 ssize_t res = fuse_get_res_by_io(io);
598 struct inode *inode = file_inode(io->iocb->ki_filp);
599 struct fuse_conn *fc = get_fuse_conn(inode);
600 struct fuse_inode *fi = get_fuse_inode(inode);
602 spin_lock(&fc->lock);
603 fi->attr_version = ++fc->attr_version;
604 spin_unlock(&fc->lock);
607 io->iocb->ki_complete(io->iocb, res, 0);
610 kref_put(&io->refcnt, fuse_io_release);
613 static void fuse_aio_complete_req(struct fuse_conn *fc, struct fuse_req *req)
615 struct fuse_io_priv *io = req->io;
618 fuse_release_user_pages(req, !io->write);
621 if (req->misc.write.in.size != req->misc.write.out.size)
622 pos = req->misc.write.in.offset - io->offset +
623 req->misc.write.out.size;
625 if (req->misc.read.in.size != req->out.args[0].size)
626 pos = req->misc.read.in.offset - io->offset +
627 req->out.args[0].size;
630 fuse_aio_complete(io, req->out.h.error, pos);
633 static size_t fuse_async_req_send(struct fuse_conn *fc, struct fuse_req *req,
634 size_t num_bytes, struct fuse_io_priv *io)
636 spin_lock(&io->lock);
637 kref_get(&io->refcnt);
638 io->size += num_bytes;
640 spin_unlock(&io->lock);
643 req->end = fuse_aio_complete_req;
645 __fuse_get_request(req);
646 fuse_request_send_background(fc, req);
651 static size_t fuse_send_read(struct fuse_req *req, struct fuse_io_priv *io,
652 loff_t pos, size_t count, fl_owner_t owner)
654 struct file *file = io->file;
655 struct fuse_file *ff = file->private_data;
656 struct fuse_conn *fc = ff->fc;
658 fuse_read_fill(req, file, pos, count, FUSE_READ);
660 struct fuse_read_in *inarg = &req->misc.read.in;
662 inarg->read_flags |= FUSE_READ_LOCKOWNER;
663 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
667 return fuse_async_req_send(fc, req, count, io);
669 fuse_request_send(fc, req);
670 return req->out.args[0].size;
673 static void fuse_read_update_size(struct inode *inode, loff_t size,
676 struct fuse_conn *fc = get_fuse_conn(inode);
677 struct fuse_inode *fi = get_fuse_inode(inode);
679 spin_lock(&fc->lock);
680 if (attr_ver == fi->attr_version && size < inode->i_size &&
681 !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
682 fi->attr_version = ++fc->attr_version;
683 i_size_write(inode, size);
685 spin_unlock(&fc->lock);
688 static void fuse_short_read(struct fuse_req *req, struct inode *inode,
691 size_t num_read = req->out.args[0].size;
692 struct fuse_conn *fc = get_fuse_conn(inode);
694 if (fc->writeback_cache) {
696 * A hole in a file. Some data after the hole are in page cache,
697 * but have not reached the client fs yet. So, the hole is not
701 int start_idx = num_read >> PAGE_CACHE_SHIFT;
702 size_t off = num_read & (PAGE_CACHE_SIZE - 1);
704 for (i = start_idx; i < req->num_pages; i++) {
705 zero_user_segment(req->pages[i], off, PAGE_CACHE_SIZE);
709 loff_t pos = page_offset(req->pages[0]) + num_read;
710 fuse_read_update_size(inode, pos, attr_ver);
714 static int fuse_do_readpage(struct file *file, struct page *page)
716 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
717 struct inode *inode = page->mapping->host;
718 struct fuse_conn *fc = get_fuse_conn(inode);
719 struct fuse_req *req;
721 loff_t pos = page_offset(page);
722 size_t count = PAGE_CACHE_SIZE;
727 * Page writeback can extend beyond the lifetime of the
728 * page-cache page, so make sure we read a properly synced
731 fuse_wait_on_page_writeback(inode, page->index);
733 req = fuse_get_req(fc, 1);
737 attr_ver = fuse_get_attr_version(fc);
739 req->out.page_zeroing = 1;
740 req->out.argpages = 1;
742 req->pages[0] = page;
743 req->page_descs[0].length = count;
744 num_read = fuse_send_read(req, &io, pos, count, NULL);
745 err = req->out.h.error;
749 * Short read means EOF. If file size is larger, truncate it
751 if (num_read < count)
752 fuse_short_read(req, inode, attr_ver);
754 SetPageUptodate(page);
757 fuse_put_request(fc, req);
762 static int fuse_readpage(struct file *file, struct page *page)
764 struct inode *inode = page->mapping->host;
768 if (is_bad_inode(inode))
771 err = fuse_do_readpage(file, page);
772 fuse_invalidate_atime(inode);
778 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
781 size_t count = req->misc.read.in.size;
782 size_t num_read = req->out.args[0].size;
783 struct address_space *mapping = NULL;
785 for (i = 0; mapping == NULL && i < req->num_pages; i++)
786 mapping = req->pages[i]->mapping;
789 struct inode *inode = mapping->host;
792 * Short read means EOF. If file size is larger, truncate it
794 if (!req->out.h.error && num_read < count)
795 fuse_short_read(req, inode, req->misc.read.attr_ver);
797 fuse_invalidate_atime(inode);
800 for (i = 0; i < req->num_pages; i++) {
801 struct page *page = req->pages[i];
802 if (!req->out.h.error)
803 SetPageUptodate(page);
807 page_cache_release(page);
810 fuse_file_put(req->ff, false);
813 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
815 struct fuse_file *ff = file->private_data;
816 struct fuse_conn *fc = ff->fc;
817 loff_t pos = page_offset(req->pages[0]);
818 size_t count = req->num_pages << PAGE_CACHE_SHIFT;
820 req->out.argpages = 1;
821 req->out.page_zeroing = 1;
822 req->out.page_replace = 1;
823 fuse_read_fill(req, file, pos, count, FUSE_READ);
824 req->misc.read.attr_ver = fuse_get_attr_version(fc);
825 if (fc->async_read) {
826 req->ff = fuse_file_get(ff);
827 req->end = fuse_readpages_end;
828 fuse_request_send_background(fc, req);
830 fuse_request_send(fc, req);
831 fuse_readpages_end(fc, req);
832 fuse_put_request(fc, req);
836 struct fuse_fill_data {
837 struct fuse_req *req;
843 static int fuse_readpages_fill(void *_data, struct page *page)
845 struct fuse_fill_data *data = _data;
846 struct fuse_req *req = data->req;
847 struct inode *inode = data->inode;
848 struct fuse_conn *fc = get_fuse_conn(inode);
850 fuse_wait_on_page_writeback(inode, page->index);
852 if (req->num_pages &&
853 (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
854 (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
855 req->pages[req->num_pages - 1]->index + 1 != page->index)) {
856 int nr_alloc = min_t(unsigned, data->nr_pages,
857 FUSE_MAX_PAGES_PER_REQ);
858 fuse_send_readpages(req, data->file);
860 req = fuse_get_req_for_background(fc, nr_alloc);
862 req = fuse_get_req(fc, nr_alloc);
871 if (WARN_ON(req->num_pages >= req->max_pages)) {
872 fuse_put_request(fc, req);
876 page_cache_get(page);
877 req->pages[req->num_pages] = page;
878 req->page_descs[req->num_pages].length = PAGE_SIZE;
884 static int fuse_readpages(struct file *file, struct address_space *mapping,
885 struct list_head *pages, unsigned nr_pages)
887 struct inode *inode = mapping->host;
888 struct fuse_conn *fc = get_fuse_conn(inode);
889 struct fuse_fill_data data;
891 int nr_alloc = min_t(unsigned, nr_pages, FUSE_MAX_PAGES_PER_REQ);
894 if (is_bad_inode(inode))
900 data.req = fuse_get_req_for_background(fc, nr_alloc);
902 data.req = fuse_get_req(fc, nr_alloc);
903 data.nr_pages = nr_pages;
904 err = PTR_ERR(data.req);
905 if (IS_ERR(data.req))
908 err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
910 if (data.req->num_pages)
911 fuse_send_readpages(data.req, file);
913 fuse_put_request(fc, data.req);
919 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
921 struct inode *inode = iocb->ki_filp->f_mapping->host;
922 struct fuse_conn *fc = get_fuse_conn(inode);
925 * In auto invalidate mode, always update attributes on read.
926 * Otherwise, only update if we attempt to read past EOF (to ensure
927 * i_size is up to date).
929 if (fc->auto_inval_data ||
930 (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
932 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
937 return generic_file_read_iter(iocb, to);
940 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
941 loff_t pos, size_t count)
943 struct fuse_write_in *inarg = &req->misc.write.in;
944 struct fuse_write_out *outarg = &req->misc.write.out;
949 req->in.h.opcode = FUSE_WRITE;
950 req->in.h.nodeid = ff->nodeid;
952 if (ff->fc->minor < 9)
953 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
955 req->in.args[0].size = sizeof(struct fuse_write_in);
956 req->in.args[0].value = inarg;
957 req->in.args[1].size = count;
958 req->out.numargs = 1;
959 req->out.args[0].size = sizeof(struct fuse_write_out);
960 req->out.args[0].value = outarg;
963 static size_t fuse_send_write(struct fuse_req *req, struct fuse_io_priv *io,
964 loff_t pos, size_t count, fl_owner_t owner)
966 struct file *file = io->file;
967 struct fuse_file *ff = file->private_data;
968 struct fuse_conn *fc = ff->fc;
969 struct fuse_write_in *inarg = &req->misc.write.in;
971 fuse_write_fill(req, ff, pos, count);
972 inarg->flags = file->f_flags;
974 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
975 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
979 return fuse_async_req_send(fc, req, count, io);
981 fuse_request_send(fc, req);
982 return req->misc.write.out.size;
985 bool fuse_write_update_size(struct inode *inode, loff_t pos)
987 struct fuse_conn *fc = get_fuse_conn(inode);
988 struct fuse_inode *fi = get_fuse_inode(inode);
991 spin_lock(&fc->lock);
992 fi->attr_version = ++fc->attr_version;
993 if (pos > inode->i_size) {
994 i_size_write(inode, pos);
997 spin_unlock(&fc->lock);
1002 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
1003 struct inode *inode, loff_t pos,
1009 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
1011 for (i = 0; i < req->num_pages; i++)
1012 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
1014 res = fuse_send_write(req, &io, pos, count, NULL);
1016 offset = req->page_descs[0].offset;
1018 for (i = 0; i < req->num_pages; i++) {
1019 struct page *page = req->pages[i];
1021 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
1022 SetPageUptodate(page);
1024 if (count > PAGE_CACHE_SIZE - offset)
1025 count -= PAGE_CACHE_SIZE - offset;
1031 page_cache_release(page);
1037 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
1038 struct address_space *mapping,
1039 struct iov_iter *ii, loff_t pos)
1041 struct fuse_conn *fc = get_fuse_conn(mapping->host);
1042 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
1046 req->in.argpages = 1;
1047 req->page_descs[0].offset = offset;
1052 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1053 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
1054 iov_iter_count(ii));
1056 bytes = min_t(size_t, bytes, fc->max_write - count);
1060 if (iov_iter_fault_in_readable(ii, bytes))
1064 page = grab_cache_page_write_begin(mapping, index, 0);
1068 if (mapping_writably_mapped(mapping))
1069 flush_dcache_page(page);
1071 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
1072 flush_dcache_page(page);
1074 iov_iter_advance(ii, tmp);
1077 page_cache_release(page);
1078 bytes = min(bytes, iov_iter_single_seg_count(ii));
1083 req->pages[req->num_pages] = page;
1084 req->page_descs[req->num_pages].length = tmp;
1090 if (offset == PAGE_CACHE_SIZE)
1093 if (!fc->big_writes)
1095 } while (iov_iter_count(ii) && count < fc->max_write &&
1096 req->num_pages < req->max_pages && offset == 0);
1098 return count > 0 ? count : err;
1101 static inline unsigned fuse_wr_pages(loff_t pos, size_t len)
1103 return min_t(unsigned,
1104 ((pos + len - 1) >> PAGE_CACHE_SHIFT) -
1105 (pos >> PAGE_CACHE_SHIFT) + 1,
1106 FUSE_MAX_PAGES_PER_REQ);
1109 static ssize_t fuse_perform_write(struct file *file,
1110 struct address_space *mapping,
1111 struct iov_iter *ii, loff_t pos)
1113 struct inode *inode = mapping->host;
1114 struct fuse_conn *fc = get_fuse_conn(inode);
1115 struct fuse_inode *fi = get_fuse_inode(inode);
1119 if (is_bad_inode(inode))
1122 if (inode->i_size < pos + iov_iter_count(ii))
1123 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1126 struct fuse_req *req;
1128 unsigned nr_pages = fuse_wr_pages(pos, iov_iter_count(ii));
1130 req = fuse_get_req(fc, nr_pages);
1136 count = fuse_fill_write_pages(req, mapping, ii, pos);
1142 num_written = fuse_send_write_pages(req, file, inode,
1144 err = req->out.h.error;
1149 /* break out of the loop on short write */
1150 if (num_written != count)
1154 fuse_put_request(fc, req);
1155 } while (!err && iov_iter_count(ii));
1158 fuse_write_update_size(inode, pos);
1160 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1161 fuse_invalidate_attr(inode);
1163 return res > 0 ? res : err;
1166 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1168 struct file *file = iocb->ki_filp;
1169 struct address_space *mapping = file->f_mapping;
1170 ssize_t written = 0;
1171 ssize_t written_buffered = 0;
1172 struct inode *inode = mapping->host;
1176 if (get_fuse_conn(inode)->writeback_cache) {
1177 /* Update size (EOF optimization) and mode (SUID clearing) */
1178 err = fuse_update_attributes(mapping->host, NULL, file, NULL);
1182 return generic_file_write_iter(iocb, from);
1185 mutex_lock(&inode->i_mutex);
1187 /* We can write back this queue in page reclaim */
1188 current->backing_dev_info = inode_to_bdi(inode);
1190 err = generic_write_checks(iocb, from);
1194 err = file_remove_privs(file);
1198 err = file_update_time(file);
1202 if (iocb->ki_flags & IOCB_DIRECT) {
1203 loff_t pos = iocb->ki_pos;
1204 written = generic_file_direct_write(iocb, from, pos);
1205 if (written < 0 || !iov_iter_count(from))
1210 written_buffered = fuse_perform_write(file, mapping, from, pos);
1211 if (written_buffered < 0) {
1212 err = written_buffered;
1215 endbyte = pos + written_buffered - 1;
1217 err = filemap_write_and_wait_range(file->f_mapping, pos,
1222 invalidate_mapping_pages(file->f_mapping,
1223 pos >> PAGE_CACHE_SHIFT,
1224 endbyte >> PAGE_CACHE_SHIFT);
1226 written += written_buffered;
1227 iocb->ki_pos = pos + written_buffered;
1229 written = fuse_perform_write(file, mapping, from, iocb->ki_pos);
1231 iocb->ki_pos += written;
1234 current->backing_dev_info = NULL;
1235 mutex_unlock(&inode->i_mutex);
1237 return written ? written : err;
1240 static inline void fuse_page_descs_length_init(struct fuse_req *req,
1241 unsigned index, unsigned nr_pages)
1245 for (i = index; i < index + nr_pages; i++)
1246 req->page_descs[i].length = PAGE_SIZE -
1247 req->page_descs[i].offset;
1250 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1252 return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1255 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1258 return min(iov_iter_single_seg_count(ii), max_size);
1261 static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1262 size_t *nbytesp, int write)
1264 size_t nbytes = 0; /* # bytes already packed in req */
1266 /* Special case for kernel I/O: can copy directly into the buffer */
1267 if (ii->type & ITER_KVEC) {
1268 unsigned long user_addr = fuse_get_user_addr(ii);
1269 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1272 req->in.args[1].value = (void *) user_addr;
1274 req->out.args[0].value = (void *) user_addr;
1276 iov_iter_advance(ii, frag_size);
1277 *nbytesp = frag_size;
1281 while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1284 ssize_t ret = iov_iter_get_pages(ii,
1285 &req->pages[req->num_pages],
1287 req->max_pages - req->num_pages,
1292 iov_iter_advance(ii, ret);
1296 npages = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
1298 req->page_descs[req->num_pages].offset = start;
1299 fuse_page_descs_length_init(req, req->num_pages, npages);
1301 req->num_pages += npages;
1302 req->page_descs[req->num_pages - 1].length -=
1303 (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1307 req->in.argpages = 1;
1309 req->out.argpages = 1;
1316 static inline int fuse_iter_npages(const struct iov_iter *ii_p)
1318 return iov_iter_npages(ii_p, FUSE_MAX_PAGES_PER_REQ);
1321 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1322 loff_t *ppos, int flags)
1324 int write = flags & FUSE_DIO_WRITE;
1325 int cuse = flags & FUSE_DIO_CUSE;
1326 struct file *file = io->file;
1327 struct inode *inode = file->f_mapping->host;
1328 struct fuse_file *ff = file->private_data;
1329 struct fuse_conn *fc = ff->fc;
1330 size_t nmax = write ? fc->max_write : fc->max_read;
1332 size_t count = iov_iter_count(iter);
1333 pgoff_t idx_from = pos >> PAGE_CACHE_SHIFT;
1334 pgoff_t idx_to = (pos + count - 1) >> PAGE_CACHE_SHIFT;
1336 struct fuse_req *req;
1339 req = fuse_get_req_for_background(fc, fuse_iter_npages(iter));
1341 req = fuse_get_req(fc, fuse_iter_npages(iter));
1343 return PTR_ERR(req);
1345 if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1347 mutex_lock(&inode->i_mutex);
1348 fuse_sync_writes(inode);
1350 mutex_unlock(&inode->i_mutex);
1355 fl_owner_t owner = current->files;
1356 size_t nbytes = min(count, nmax);
1357 int err = fuse_get_user_pages(req, iter, &nbytes, write);
1364 nres = fuse_send_write(req, io, pos, nbytes, owner);
1366 nres = fuse_send_read(req, io, pos, nbytes, owner);
1369 fuse_release_user_pages(req, !write);
1370 if (req->out.h.error) {
1372 res = req->out.h.error;
1374 } else if (nres > nbytes) {
1384 fuse_put_request(fc, req);
1386 req = fuse_get_req_for_background(fc,
1387 fuse_iter_npages(iter));
1389 req = fuse_get_req(fc, fuse_iter_npages(iter));
1395 fuse_put_request(fc, req);
1401 EXPORT_SYMBOL_GPL(fuse_direct_io);
1403 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1404 struct iov_iter *iter,
1408 struct file *file = io->file;
1409 struct inode *inode = file_inode(file);
1411 if (is_bad_inode(inode))
1414 res = fuse_direct_io(io, iter, ppos, 0);
1416 fuse_invalidate_attr(inode);
1421 static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1423 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb->ki_filp);
1424 return __fuse_direct_read(&io, to, &iocb->ki_pos);
1427 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1429 struct file *file = iocb->ki_filp;
1430 struct inode *inode = file_inode(file);
1431 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
1434 if (is_bad_inode(inode))
1437 /* Don't allow parallel writes to the same file */
1438 mutex_lock(&inode->i_mutex);
1439 res = generic_write_checks(iocb, from);
1441 res = fuse_direct_io(&io, from, &iocb->ki_pos, FUSE_DIO_WRITE);
1442 fuse_invalidate_attr(inode);
1444 fuse_write_update_size(inode, iocb->ki_pos);
1445 mutex_unlock(&inode->i_mutex);
1450 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1454 for (i = 0; i < req->num_pages; i++)
1455 __free_page(req->pages[i]);
1458 fuse_file_put(req->ff, false);
1461 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1463 struct inode *inode = req->inode;
1464 struct fuse_inode *fi = get_fuse_inode(inode);
1465 struct backing_dev_info *bdi = inode_to_bdi(inode);
1468 list_del(&req->writepages_entry);
1469 for (i = 0; i < req->num_pages; i++) {
1470 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1471 dec_zone_page_state(req->pages[i], NR_WRITEBACK_TEMP);
1472 wb_writeout_inc(&bdi->wb);
1474 wake_up(&fi->page_waitq);
1477 /* Called under fc->lock, may release and reacquire it */
1478 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req,
1480 __releases(fc->lock)
1481 __acquires(fc->lock)
1483 struct fuse_inode *fi = get_fuse_inode(req->inode);
1484 struct fuse_write_in *inarg = &req->misc.write.in;
1485 __u64 data_size = req->num_pages * PAGE_CACHE_SIZE;
1490 if (inarg->offset + data_size <= size) {
1491 inarg->size = data_size;
1492 } else if (inarg->offset < size) {
1493 inarg->size = size - inarg->offset;
1495 /* Got truncated off completely */
1499 req->in.args[1].size = inarg->size;
1501 fuse_request_send_background_locked(fc, req);
1505 fuse_writepage_finish(fc, req);
1506 spin_unlock(&fc->lock);
1507 fuse_writepage_free(fc, req);
1508 fuse_put_request(fc, req);
1509 spin_lock(&fc->lock);
1513 * If fi->writectr is positive (no truncate or fsync going on) send
1514 * all queued writepage requests.
1516 * Called with fc->lock
1518 void fuse_flush_writepages(struct inode *inode)
1519 __releases(fc->lock)
1520 __acquires(fc->lock)
1522 struct fuse_conn *fc = get_fuse_conn(inode);
1523 struct fuse_inode *fi = get_fuse_inode(inode);
1524 size_t crop = i_size_read(inode);
1525 struct fuse_req *req;
1527 while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1528 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1529 list_del_init(&req->list);
1530 fuse_send_writepage(fc, req, crop);
1534 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1536 struct inode *inode = req->inode;
1537 struct fuse_inode *fi = get_fuse_inode(inode);
1539 mapping_set_error(inode->i_mapping, req->out.h.error);
1540 spin_lock(&fc->lock);
1541 while (req->misc.write.next) {
1542 struct fuse_conn *fc = get_fuse_conn(inode);
1543 struct fuse_write_in *inarg = &req->misc.write.in;
1544 struct fuse_req *next = req->misc.write.next;
1545 req->misc.write.next = next->misc.write.next;
1546 next->misc.write.next = NULL;
1547 next->ff = fuse_file_get(req->ff);
1548 list_add(&next->writepages_entry, &fi->writepages);
1551 * Skip fuse_flush_writepages() to make it easy to crop requests
1552 * based on primary request size.
1554 * 1st case (trivial): there are no concurrent activities using
1555 * fuse_set/release_nowrite. Then we're on safe side because
1556 * fuse_flush_writepages() would call fuse_send_writepage()
1559 * 2nd case: someone called fuse_set_nowrite and it is waiting
1560 * now for completion of all in-flight requests. This happens
1561 * rarely and no more than once per page, so this should be
1564 * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1565 * of fuse_set_nowrite..fuse_release_nowrite section. The fact
1566 * that fuse_set_nowrite returned implies that all in-flight
1567 * requests were completed along with all of their secondary
1568 * requests. Further primary requests are blocked by negative
1569 * writectr. Hence there cannot be any in-flight requests and
1570 * no invocations of fuse_writepage_end() while we're in
1571 * fuse_set_nowrite..fuse_release_nowrite section.
1573 fuse_send_writepage(fc, next, inarg->offset + inarg->size);
1576 fuse_writepage_finish(fc, req);
1577 spin_unlock(&fc->lock);
1578 fuse_writepage_free(fc, req);
1581 static struct fuse_file *__fuse_write_file_get(struct fuse_conn *fc,
1582 struct fuse_inode *fi)
1584 struct fuse_file *ff = NULL;
1586 spin_lock(&fc->lock);
1587 if (!list_empty(&fi->write_files)) {
1588 ff = list_entry(fi->write_files.next, struct fuse_file,
1592 spin_unlock(&fc->lock);
1597 static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1598 struct fuse_inode *fi)
1600 struct fuse_file *ff = __fuse_write_file_get(fc, fi);
1605 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1607 struct fuse_conn *fc = get_fuse_conn(inode);
1608 struct fuse_inode *fi = get_fuse_inode(inode);
1609 struct fuse_file *ff;
1612 ff = __fuse_write_file_get(fc, fi);
1613 err = fuse_flush_times(inode, ff);
1615 fuse_file_put(ff, 0);
1620 static int fuse_writepage_locked(struct page *page)
1622 struct address_space *mapping = page->mapping;
1623 struct inode *inode = mapping->host;
1624 struct fuse_conn *fc = get_fuse_conn(inode);
1625 struct fuse_inode *fi = get_fuse_inode(inode);
1626 struct fuse_req *req;
1627 struct page *tmp_page;
1628 int error = -ENOMEM;
1630 set_page_writeback(page);
1632 req = fuse_request_alloc_nofs(1);
1636 /* writeback always goes to bg_queue */
1637 __set_bit(FR_BACKGROUND, &req->flags);
1638 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1643 req->ff = fuse_write_file_get(fc, fi);
1647 fuse_write_fill(req, req->ff, page_offset(page), 0);
1649 copy_highpage(tmp_page, page);
1650 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1651 req->misc.write.next = NULL;
1652 req->in.argpages = 1;
1654 req->pages[0] = tmp_page;
1655 req->page_descs[0].offset = 0;
1656 req->page_descs[0].length = PAGE_SIZE;
1657 req->end = fuse_writepage_end;
1660 inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1661 inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1663 spin_lock(&fc->lock);
1664 list_add(&req->writepages_entry, &fi->writepages);
1665 list_add_tail(&req->list, &fi->queued_writes);
1666 fuse_flush_writepages(inode);
1667 spin_unlock(&fc->lock);
1669 end_page_writeback(page);
1674 __free_page(tmp_page);
1676 fuse_request_free(req);
1678 end_page_writeback(page);
1682 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1686 if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1688 * ->writepages() should be called for sync() and friends. We
1689 * should only get here on direct reclaim and then we are
1690 * allowed to skip a page which is already in flight
1692 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1694 redirty_page_for_writepage(wbc, page);
1698 err = fuse_writepage_locked(page);
1704 struct fuse_fill_wb_data {
1705 struct fuse_req *req;
1706 struct fuse_file *ff;
1707 struct inode *inode;
1708 struct page **orig_pages;
1711 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1713 struct fuse_req *req = data->req;
1714 struct inode *inode = data->inode;
1715 struct fuse_conn *fc = get_fuse_conn(inode);
1716 struct fuse_inode *fi = get_fuse_inode(inode);
1717 int num_pages = req->num_pages;
1720 req->ff = fuse_file_get(data->ff);
1721 spin_lock(&fc->lock);
1722 list_add_tail(&req->list, &fi->queued_writes);
1723 fuse_flush_writepages(inode);
1724 spin_unlock(&fc->lock);
1726 for (i = 0; i < num_pages; i++)
1727 end_page_writeback(data->orig_pages[i]);
1730 static bool fuse_writepage_in_flight(struct fuse_req *new_req,
1733 struct fuse_conn *fc = get_fuse_conn(new_req->inode);
1734 struct fuse_inode *fi = get_fuse_inode(new_req->inode);
1735 struct fuse_req *tmp;
1736 struct fuse_req *old_req;
1740 BUG_ON(new_req->num_pages != 0);
1742 spin_lock(&fc->lock);
1743 list_del(&new_req->writepages_entry);
1744 list_for_each_entry(old_req, &fi->writepages, writepages_entry) {
1745 BUG_ON(old_req->inode != new_req->inode);
1746 curr_index = old_req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
1747 if (curr_index <= page->index &&
1748 page->index < curr_index + old_req->num_pages) {
1754 list_add(&new_req->writepages_entry, &fi->writepages);
1758 new_req->num_pages = 1;
1759 for (tmp = old_req; tmp != NULL; tmp = tmp->misc.write.next) {
1760 BUG_ON(tmp->inode != new_req->inode);
1761 curr_index = tmp->misc.write.in.offset >> PAGE_CACHE_SHIFT;
1762 if (tmp->num_pages == 1 &&
1763 curr_index == page->index) {
1768 if (old_req->num_pages == 1 && test_bit(FR_PENDING, &old_req->flags)) {
1769 struct backing_dev_info *bdi = inode_to_bdi(page->mapping->host);
1771 copy_highpage(old_req->pages[0], page);
1772 spin_unlock(&fc->lock);
1774 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1775 dec_zone_page_state(page, NR_WRITEBACK_TEMP);
1776 wb_writeout_inc(&bdi->wb);
1777 fuse_writepage_free(fc, new_req);
1778 fuse_request_free(new_req);
1781 new_req->misc.write.next = old_req->misc.write.next;
1782 old_req->misc.write.next = new_req;
1785 spin_unlock(&fc->lock);
1790 static int fuse_writepages_fill(struct page *page,
1791 struct writeback_control *wbc, void *_data)
1793 struct fuse_fill_wb_data *data = _data;
1794 struct fuse_req *req = data->req;
1795 struct inode *inode = data->inode;
1796 struct fuse_conn *fc = get_fuse_conn(inode);
1797 struct page *tmp_page;
1803 data->ff = fuse_write_file_get(fc, get_fuse_inode(inode));
1809 * Being under writeback is unlikely but possible. For example direct
1810 * read to an mmaped fuse file will set the page dirty twice; once when
1811 * the pages are faulted with get_user_pages(), and then after the read
1814 is_writeback = fuse_page_is_writeback(inode, page->index);
1816 if (req && req->num_pages &&
1817 (is_writeback || req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
1818 (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_write ||
1819 data->orig_pages[req->num_pages - 1]->index + 1 != page->index)) {
1820 fuse_writepages_send(data);
1824 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1829 * The page must not be redirtied until the writeout is completed
1830 * (i.e. userspace has sent a reply to the write request). Otherwise
1831 * there could be more than one temporary page instance for each real
1834 * This is ensured by holding the page lock in page_mkwrite() while
1835 * checking fuse_page_is_writeback(). We already hold the page lock
1836 * since clear_page_dirty_for_io() and keep it held until we add the
1837 * request to the fi->writepages list and increment req->num_pages.
1838 * After this fuse_page_is_writeback() will indicate that the page is
1839 * under writeback, so we can release the page lock.
1841 if (data->req == NULL) {
1842 struct fuse_inode *fi = get_fuse_inode(inode);
1845 req = fuse_request_alloc_nofs(FUSE_MAX_PAGES_PER_REQ);
1847 __free_page(tmp_page);
1851 fuse_write_fill(req, data->ff, page_offset(page), 0);
1852 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1853 req->misc.write.next = NULL;
1854 req->in.argpages = 1;
1855 __set_bit(FR_BACKGROUND, &req->flags);
1857 req->end = fuse_writepage_end;
1860 spin_lock(&fc->lock);
1861 list_add(&req->writepages_entry, &fi->writepages);
1862 spin_unlock(&fc->lock);
1866 set_page_writeback(page);
1868 copy_highpage(tmp_page, page);
1869 req->pages[req->num_pages] = tmp_page;
1870 req->page_descs[req->num_pages].offset = 0;
1871 req->page_descs[req->num_pages].length = PAGE_SIZE;
1873 inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1874 inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1877 if (is_writeback && fuse_writepage_in_flight(req, page)) {
1878 end_page_writeback(page);
1882 data->orig_pages[req->num_pages] = page;
1885 * Protected by fc->lock against concurrent access by
1886 * fuse_page_is_writeback().
1888 spin_lock(&fc->lock);
1890 spin_unlock(&fc->lock);
1898 static int fuse_writepages(struct address_space *mapping,
1899 struct writeback_control *wbc)
1901 struct inode *inode = mapping->host;
1902 struct fuse_fill_wb_data data;
1906 if (is_bad_inode(inode))
1914 data.orig_pages = kcalloc(FUSE_MAX_PAGES_PER_REQ,
1915 sizeof(struct page *),
1917 if (!data.orig_pages)
1920 err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
1922 /* Ignore errors if we can write at least one page */
1923 BUG_ON(!data.req->num_pages);
1924 fuse_writepages_send(&data);
1928 fuse_file_put(data.ff, false);
1930 kfree(data.orig_pages);
1936 * It's worthy to make sure that space is reserved on disk for the write,
1937 * but how to implement it without killing performance need more thinking.
1939 static int fuse_write_begin(struct file *file, struct address_space *mapping,
1940 loff_t pos, unsigned len, unsigned flags,
1941 struct page **pagep, void **fsdata)
1943 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1944 struct fuse_conn *fc = get_fuse_conn(file_inode(file));
1949 WARN_ON(!fc->writeback_cache);
1951 page = grab_cache_page_write_begin(mapping, index, flags);
1955 fuse_wait_on_page_writeback(mapping->host, page->index);
1957 if (PageUptodate(page) || len == PAGE_CACHE_SIZE)
1960 * Check if the start this page comes after the end of file, in which
1961 * case the readpage can be optimized away.
1963 fsize = i_size_read(mapping->host);
1964 if (fsize <= (pos & PAGE_CACHE_MASK)) {
1965 size_t off = pos & ~PAGE_CACHE_MASK;
1967 zero_user_segment(page, 0, off);
1970 err = fuse_do_readpage(file, page);
1979 page_cache_release(page);
1984 static int fuse_write_end(struct file *file, struct address_space *mapping,
1985 loff_t pos, unsigned len, unsigned copied,
1986 struct page *page, void *fsdata)
1988 struct inode *inode = page->mapping->host;
1990 if (!PageUptodate(page)) {
1991 /* Zero any unwritten bytes at the end of the page */
1992 size_t endoff = (pos + copied) & ~PAGE_CACHE_MASK;
1994 zero_user_segment(page, endoff, PAGE_CACHE_SIZE);
1995 SetPageUptodate(page);
1998 fuse_write_update_size(inode, pos + copied);
1999 set_page_dirty(page);
2001 page_cache_release(page);
2006 static int fuse_launder_page(struct page *page)
2009 if (clear_page_dirty_for_io(page)) {
2010 struct inode *inode = page->mapping->host;
2011 err = fuse_writepage_locked(page);
2013 fuse_wait_on_page_writeback(inode, page->index);
2019 * Write back dirty pages now, because there may not be any suitable
2022 static void fuse_vma_close(struct vm_area_struct *vma)
2024 filemap_write_and_wait(vma->vm_file->f_mapping);
2028 * Wait for writeback against this page to complete before allowing it
2029 * to be marked dirty again, and hence written back again, possibly
2030 * before the previous writepage completed.
2032 * Block here, instead of in ->writepage(), so that the userspace fs
2033 * can only block processes actually operating on the filesystem.
2035 * Otherwise unprivileged userspace fs would be able to block
2040 * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2042 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
2044 struct page *page = vmf->page;
2045 struct inode *inode = file_inode(vma->vm_file);
2047 file_update_time(vma->vm_file);
2049 if (page->mapping != inode->i_mapping) {
2051 return VM_FAULT_NOPAGE;
2054 fuse_wait_on_page_writeback(inode, page->index);
2055 return VM_FAULT_LOCKED;
2058 static const struct vm_operations_struct fuse_file_vm_ops = {
2059 .close = fuse_vma_close,
2060 .fault = filemap_fault,
2061 .map_pages = filemap_map_pages,
2062 .page_mkwrite = fuse_page_mkwrite,
2065 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2067 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2068 fuse_link_write_file(file);
2070 file_accessed(file);
2071 vma->vm_ops = &fuse_file_vm_ops;
2075 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
2077 /* Can't provide the coherency needed for MAP_SHARED */
2078 if (vma->vm_flags & VM_MAYSHARE)
2081 invalidate_inode_pages2(file->f_mapping);
2083 return generic_file_mmap(file, vma);
2086 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
2087 struct file_lock *fl)
2089 switch (ffl->type) {
2095 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2096 ffl->end < ffl->start)
2099 fl->fl_start = ffl->start;
2100 fl->fl_end = ffl->end;
2101 fl->fl_pid = ffl->pid;
2107 fl->fl_type = ffl->type;
2111 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2112 const struct file_lock *fl, int opcode, pid_t pid,
2113 int flock, struct fuse_lk_in *inarg)
2115 struct inode *inode = file_inode(file);
2116 struct fuse_conn *fc = get_fuse_conn(inode);
2117 struct fuse_file *ff = file->private_data;
2119 memset(inarg, 0, sizeof(*inarg));
2121 inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2122 inarg->lk.start = fl->fl_start;
2123 inarg->lk.end = fl->fl_end;
2124 inarg->lk.type = fl->fl_type;
2125 inarg->lk.pid = pid;
2127 inarg->lk_flags |= FUSE_LK_FLOCK;
2128 args->in.h.opcode = opcode;
2129 args->in.h.nodeid = get_node_id(inode);
2130 args->in.numargs = 1;
2131 args->in.args[0].size = sizeof(*inarg);
2132 args->in.args[0].value = inarg;
2135 static int fuse_getlk(struct file *file, struct file_lock *fl)
2137 struct inode *inode = file_inode(file);
2138 struct fuse_conn *fc = get_fuse_conn(inode);
2140 struct fuse_lk_in inarg;
2141 struct fuse_lk_out outarg;
2144 fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2145 args.out.numargs = 1;
2146 args.out.args[0].size = sizeof(outarg);
2147 args.out.args[0].value = &outarg;
2148 err = fuse_simple_request(fc, &args);
2150 err = convert_fuse_file_lock(&outarg.lk, fl);
2155 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2157 struct inode *inode = file_inode(file);
2158 struct fuse_conn *fc = get_fuse_conn(inode);
2160 struct fuse_lk_in inarg;
2161 int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2162 pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
2165 if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2166 /* NLM needs asynchronous locks, which we don't support yet */
2170 /* Unlock on close is handled by the flush method */
2171 if (fl->fl_flags & FL_CLOSE)
2174 fuse_lk_fill(&args, file, fl, opcode, pid, flock, &inarg);
2175 err = fuse_simple_request(fc, &args);
2177 /* locking is restartable */
2184 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2186 struct inode *inode = file_inode(file);
2187 struct fuse_conn *fc = get_fuse_conn(inode);
2190 if (cmd == F_CANCELLK) {
2192 } else if (cmd == F_GETLK) {
2194 posix_test_lock(file, fl);
2197 err = fuse_getlk(file, fl);
2200 err = posix_lock_file(file, fl, NULL);
2202 err = fuse_setlk(file, fl, 0);
2207 static int fuse_file_flock(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);
2214 err = locks_lock_file_wait(file, fl);
2216 struct fuse_file *ff = file->private_data;
2218 /* emulate flock with POSIX locks */
2220 err = fuse_setlk(file, fl, 1);
2226 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2228 struct inode *inode = mapping->host;
2229 struct fuse_conn *fc = get_fuse_conn(inode);
2231 struct fuse_bmap_in inarg;
2232 struct fuse_bmap_out outarg;
2235 if (!inode->i_sb->s_bdev || fc->no_bmap)
2238 memset(&inarg, 0, sizeof(inarg));
2239 inarg.block = block;
2240 inarg.blocksize = inode->i_sb->s_blocksize;
2241 args.in.h.opcode = FUSE_BMAP;
2242 args.in.h.nodeid = get_node_id(inode);
2243 args.in.numargs = 1;
2244 args.in.args[0].size = sizeof(inarg);
2245 args.in.args[0].value = &inarg;
2246 args.out.numargs = 1;
2247 args.out.args[0].size = sizeof(outarg);
2248 args.out.args[0].value = &outarg;
2249 err = fuse_simple_request(fc, &args);
2253 return err ? 0 : outarg.block;
2256 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2259 struct inode *inode = file_inode(file);
2261 /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2262 if (whence == SEEK_CUR || whence == SEEK_SET)
2263 return generic_file_llseek(file, offset, whence);
2265 mutex_lock(&inode->i_mutex);
2266 retval = fuse_update_attributes(inode, NULL, file, NULL);
2268 retval = generic_file_llseek(file, offset, whence);
2269 mutex_unlock(&inode->i_mutex);
2274 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
2275 unsigned int nr_segs, size_t bytes, bool to_user)
2283 iov_iter_init(&ii, to_user ? READ : WRITE, iov, nr_segs, bytes);
2285 while (iov_iter_count(&ii)) {
2286 struct page *page = pages[page_idx++];
2287 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
2293 char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
2294 size_t iov_len = ii.iov->iov_len - ii.iov_offset;
2295 size_t copy = min(todo, iov_len);
2299 left = copy_from_user(kaddr, uaddr, copy);
2301 left = copy_to_user(uaddr, kaddr, copy);
2306 iov_iter_advance(&ii, copy);
2318 * CUSE servers compiled on 32bit broke on 64bit kernels because the
2319 * ABI was defined to be 'struct iovec' which is different on 32bit
2320 * and 64bit. Fortunately we can determine which structure the server
2321 * used from the size of the reply.
2323 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2324 size_t transferred, unsigned count,
2327 #ifdef CONFIG_COMPAT
2328 if (count * sizeof(struct compat_iovec) == transferred) {
2329 struct compat_iovec *ciov = src;
2333 * With this interface a 32bit server cannot support
2334 * non-compat (i.e. ones coming from 64bit apps) ioctl
2340 for (i = 0; i < count; i++) {
2341 dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2342 dst[i].iov_len = ciov[i].iov_len;
2348 if (count * sizeof(struct iovec) != transferred)
2351 memcpy(dst, src, transferred);
2355 /* Make sure iov_length() won't overflow */
2356 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
2359 u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
2361 for (n = 0; n < count; n++, iov++) {
2362 if (iov->iov_len > (size_t) max)
2364 max -= iov->iov_len;
2369 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2370 void *src, size_t transferred, unsigned count,
2374 struct fuse_ioctl_iovec *fiov = src;
2376 if (fc->minor < 16) {
2377 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2381 if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2384 for (i = 0; i < count; i++) {
2385 /* Did the server supply an inappropriate value? */
2386 if (fiov[i].base != (unsigned long) fiov[i].base ||
2387 fiov[i].len != (unsigned long) fiov[i].len)
2390 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2391 dst[i].iov_len = (size_t) fiov[i].len;
2393 #ifdef CONFIG_COMPAT
2395 (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2396 (compat_size_t) dst[i].iov_len != fiov[i].len))
2406 * For ioctls, there is no generic way to determine how much memory
2407 * needs to be read and/or written. Furthermore, ioctls are allowed
2408 * to dereference the passed pointer, so the parameter requires deep
2409 * copying but FUSE has no idea whatsoever about what to copy in or
2412 * This is solved by allowing FUSE server to retry ioctl with
2413 * necessary in/out iovecs. Let's assume the ioctl implementation
2414 * needs to read in the following structure.
2421 * On the first callout to FUSE server, inarg->in_size and
2422 * inarg->out_size will be NULL; then, the server completes the ioctl
2423 * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2424 * the actual iov array to
2426 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) } }
2428 * which tells FUSE to copy in the requested area and retry the ioctl.
2429 * On the second round, the server has access to the structure and
2430 * from that it can tell what to look for next, so on the invocation,
2431 * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2433 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) },
2434 * { .iov_base = a.buf, .iov_len = a.buflen } }
2436 * FUSE will copy both struct a and the pointed buffer from the
2437 * process doing the ioctl and retry ioctl with both struct a and the
2440 * This time, FUSE server has everything it needs and completes ioctl
2441 * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2443 * Copying data out works the same way.
2445 * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2446 * automatically initializes in and out iovs by decoding @cmd with
2447 * _IOC_* macros and the server is not allowed to request RETRY. This
2448 * limits ioctl data transfers to well-formed ioctls and is the forced
2449 * behavior for all FUSE servers.
2451 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2454 struct fuse_file *ff = file->private_data;
2455 struct fuse_conn *fc = ff->fc;
2456 struct fuse_ioctl_in inarg = {
2462 struct fuse_ioctl_out outarg;
2463 struct fuse_req *req = NULL;
2464 struct page **pages = NULL;
2465 struct iovec *iov_page = NULL;
2466 struct iovec *in_iov = NULL, *out_iov = NULL;
2467 unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
2468 size_t in_size, out_size, transferred;
2471 #if BITS_PER_LONG == 32
2472 inarg.flags |= FUSE_IOCTL_32BIT;
2474 if (flags & FUSE_IOCTL_COMPAT)
2475 inarg.flags |= FUSE_IOCTL_32BIT;
2478 /* assume all the iovs returned by client always fits in a page */
2479 BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2482 pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
2483 iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2484 if (!pages || !iov_page)
2488 * If restricted, initialize IO parameters as encoded in @cmd.
2489 * RETRY from server is not allowed.
2491 if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2492 struct iovec *iov = iov_page;
2494 iov->iov_base = (void __user *)arg;
2495 iov->iov_len = _IOC_SIZE(cmd);
2497 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2502 if (_IOC_DIR(cmd) & _IOC_READ) {
2509 inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2510 inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2513 * Out data can be used either for actual out data or iovs,
2514 * make sure there always is at least one page.
2516 out_size = max_t(size_t, out_size, PAGE_SIZE);
2517 max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2519 /* make sure there are enough buffer pages and init request with them */
2521 if (max_pages > FUSE_MAX_PAGES_PER_REQ)
2523 while (num_pages < max_pages) {
2524 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2525 if (!pages[num_pages])
2530 req = fuse_get_req(fc, num_pages);
2536 memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
2537 req->num_pages = num_pages;
2538 fuse_page_descs_length_init(req, 0, req->num_pages);
2540 /* okay, let's send it to the client */
2541 req->in.h.opcode = FUSE_IOCTL;
2542 req->in.h.nodeid = ff->nodeid;
2543 req->in.numargs = 1;
2544 req->in.args[0].size = sizeof(inarg);
2545 req->in.args[0].value = &inarg;
2548 req->in.args[1].size = in_size;
2549 req->in.argpages = 1;
2551 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
2557 req->out.numargs = 2;
2558 req->out.args[0].size = sizeof(outarg);
2559 req->out.args[0].value = &outarg;
2560 req->out.args[1].size = out_size;
2561 req->out.argpages = 1;
2562 req->out.argvar = 1;
2564 fuse_request_send(fc, req);
2565 err = req->out.h.error;
2566 transferred = req->out.args[1].size;
2567 fuse_put_request(fc, req);
2572 /* did it ask for retry? */
2573 if (outarg.flags & FUSE_IOCTL_RETRY) {
2576 /* no retry if in restricted mode */
2578 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2581 in_iovs = outarg.in_iovs;
2582 out_iovs = outarg.out_iovs;
2585 * Make sure things are in boundary, separate checks
2586 * are to protect against overflow.
2589 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2590 out_iovs > FUSE_IOCTL_MAX_IOV ||
2591 in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2594 vaddr = kmap_atomic(pages[0]);
2595 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2596 transferred, in_iovs + out_iovs,
2597 (flags & FUSE_IOCTL_COMPAT) != 0);
2598 kunmap_atomic(vaddr);
2603 out_iov = in_iov + in_iovs;
2605 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
2609 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
2617 if (transferred > inarg.out_size)
2620 err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
2623 fuse_put_request(fc, req);
2624 free_page((unsigned long) iov_page);
2626 __free_page(pages[--num_pages]);
2629 return err ? err : outarg.result;
2631 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2633 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2634 unsigned long arg, unsigned int flags)
2636 struct inode *inode = file_inode(file);
2637 struct fuse_conn *fc = get_fuse_conn(inode);
2639 if (!fuse_allow_current_process(fc))
2642 if (is_bad_inode(inode))
2645 return fuse_do_ioctl(file, cmd, arg, flags);
2648 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2651 return fuse_ioctl_common(file, cmd, arg, 0);
2654 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2657 return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2661 * All files which have been polled are linked to RB tree
2662 * fuse_conn->polled_files which is indexed by kh. Walk the tree and
2663 * find the matching one.
2665 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2666 struct rb_node **parent_out)
2668 struct rb_node **link = &fc->polled_files.rb_node;
2669 struct rb_node *last = NULL;
2672 struct fuse_file *ff;
2675 ff = rb_entry(last, struct fuse_file, polled_node);
2678 link = &last->rb_left;
2679 else if (kh > ff->kh)
2680 link = &last->rb_right;
2691 * The file is about to be polled. Make sure it's on the polled_files
2692 * RB tree. Note that files once added to the polled_files tree are
2693 * not removed before the file is released. This is because a file
2694 * polled once is likely to be polled again.
2696 static void fuse_register_polled_file(struct fuse_conn *fc,
2697 struct fuse_file *ff)
2699 spin_lock(&fc->lock);
2700 if (RB_EMPTY_NODE(&ff->polled_node)) {
2701 struct rb_node **link, *uninitialized_var(parent);
2703 link = fuse_find_polled_node(fc, ff->kh, &parent);
2705 rb_link_node(&ff->polled_node, parent, link);
2706 rb_insert_color(&ff->polled_node, &fc->polled_files);
2708 spin_unlock(&fc->lock);
2711 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2713 struct fuse_file *ff = file->private_data;
2714 struct fuse_conn *fc = ff->fc;
2715 struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2716 struct fuse_poll_out outarg;
2721 return DEFAULT_POLLMASK;
2723 poll_wait(file, &ff->poll_wait, wait);
2724 inarg.events = (__u32)poll_requested_events(wait);
2727 * Ask for notification iff there's someone waiting for it.
2728 * The client may ignore the flag and always notify.
2730 if (waitqueue_active(&ff->poll_wait)) {
2731 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2732 fuse_register_polled_file(fc, ff);
2735 args.in.h.opcode = FUSE_POLL;
2736 args.in.h.nodeid = ff->nodeid;
2737 args.in.numargs = 1;
2738 args.in.args[0].size = sizeof(inarg);
2739 args.in.args[0].value = &inarg;
2740 args.out.numargs = 1;
2741 args.out.args[0].size = sizeof(outarg);
2742 args.out.args[0].value = &outarg;
2743 err = fuse_simple_request(fc, &args);
2746 return outarg.revents;
2747 if (err == -ENOSYS) {
2749 return DEFAULT_POLLMASK;
2753 EXPORT_SYMBOL_GPL(fuse_file_poll);
2756 * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2757 * wakes up the poll waiters.
2759 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2760 struct fuse_notify_poll_wakeup_out *outarg)
2762 u64 kh = outarg->kh;
2763 struct rb_node **link;
2765 spin_lock(&fc->lock);
2767 link = fuse_find_polled_node(fc, kh, NULL);
2769 struct fuse_file *ff;
2771 ff = rb_entry(*link, struct fuse_file, polled_node);
2772 wake_up_interruptible_sync(&ff->poll_wait);
2775 spin_unlock(&fc->lock);
2779 static void fuse_do_truncate(struct file *file)
2781 struct inode *inode = file->f_mapping->host;
2784 attr.ia_valid = ATTR_SIZE;
2785 attr.ia_size = i_size_read(inode);
2787 attr.ia_file = file;
2788 attr.ia_valid |= ATTR_FILE;
2790 fuse_do_setattr(inode, &attr, file);
2793 static inline loff_t fuse_round_up(loff_t off)
2795 return round_up(off, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
2799 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter, loff_t offset)
2801 DECLARE_COMPLETION_ONSTACK(wait);
2803 struct file *file = iocb->ki_filp;
2804 struct fuse_file *ff = file->private_data;
2805 bool async_dio = ff->fc->async_dio;
2807 struct inode *inode;
2809 size_t count = iov_iter_count(iter);
2810 struct fuse_io_priv *io;
2811 bool is_sync = is_sync_kiocb(iocb);
2814 inode = file->f_mapping->host;
2815 i_size = i_size_read(inode);
2817 if ((iov_iter_rw(iter) == READ) && (offset > i_size))
2820 /* optimization for short read */
2821 if (async_dio && iov_iter_rw(iter) != WRITE && offset + count > i_size) {
2822 if (offset >= i_size)
2824 iov_iter_truncate(iter, fuse_round_up(i_size - offset));
2825 count = iov_iter_count(iter);
2828 io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2831 spin_lock_init(&io->lock);
2832 kref_init(&io->refcnt);
2836 io->offset = offset;
2837 io->write = (iov_iter_rw(iter) == WRITE);
2841 * By default, we want to optimize all I/Os with async request
2842 * submission to the client filesystem if supported.
2844 io->async = async_dio;
2848 * We cannot asynchronously extend the size of a file. We have no method
2849 * to wait on real async I/O requests, so we must submit this request
2852 if (!is_sync && (offset + count > i_size) &&
2853 iov_iter_rw(iter) == WRITE)
2856 if (io->async && is_sync) {
2858 * Additional reference to keep io around after
2859 * calling fuse_aio_complete()
2861 kref_get(&io->refcnt);
2865 if (iov_iter_rw(iter) == WRITE) {
2866 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
2867 fuse_invalidate_attr(inode);
2869 ret = __fuse_direct_read(io, iter, &pos);
2873 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2875 /* we have a non-extending, async request, so return */
2877 return -EIOCBQUEUED;
2879 wait_for_completion(&wait);
2880 ret = fuse_get_res_by_io(io);
2883 kref_put(&io->refcnt, fuse_io_release);
2885 if (iov_iter_rw(iter) == WRITE) {
2887 fuse_write_update_size(inode, pos);
2888 else if (ret < 0 && offset + count > i_size)
2889 fuse_do_truncate(file);
2895 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2898 struct fuse_file *ff = file->private_data;
2899 struct inode *inode = file_inode(file);
2900 struct fuse_inode *fi = get_fuse_inode(inode);
2901 struct fuse_conn *fc = ff->fc;
2903 struct fuse_fallocate_in inarg = {
2910 bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
2911 (mode & FALLOC_FL_PUNCH_HOLE);
2913 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2916 if (fc->no_fallocate)
2920 mutex_lock(&inode->i_mutex);
2921 if (mode & FALLOC_FL_PUNCH_HOLE) {
2922 loff_t endbyte = offset + length - 1;
2923 err = filemap_write_and_wait_range(inode->i_mapping,
2928 fuse_sync_writes(inode);
2932 if (!(mode & FALLOC_FL_KEEP_SIZE))
2933 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2935 args.in.h.opcode = FUSE_FALLOCATE;
2936 args.in.h.nodeid = ff->nodeid;
2937 args.in.numargs = 1;
2938 args.in.args[0].size = sizeof(inarg);
2939 args.in.args[0].value = &inarg;
2940 err = fuse_simple_request(fc, &args);
2941 if (err == -ENOSYS) {
2942 fc->no_fallocate = 1;
2948 /* we could have extended the file */
2949 if (!(mode & FALLOC_FL_KEEP_SIZE)) {
2950 bool changed = fuse_write_update_size(inode, offset + length);
2952 if (changed && fc->writeback_cache)
2953 file_update_time(file);
2956 if (mode & FALLOC_FL_PUNCH_HOLE)
2957 truncate_pagecache_range(inode, offset, offset + length - 1);
2959 fuse_invalidate_attr(inode);
2962 if (!(mode & FALLOC_FL_KEEP_SIZE))
2963 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2966 mutex_unlock(&inode->i_mutex);
2971 static const struct file_operations fuse_file_operations = {
2972 .llseek = fuse_file_llseek,
2973 .read_iter = fuse_file_read_iter,
2974 .write_iter = fuse_file_write_iter,
2975 .mmap = fuse_file_mmap,
2977 .flush = fuse_flush,
2978 .release = fuse_release,
2979 .fsync = fuse_fsync,
2980 .lock = fuse_file_lock,
2981 .flock = fuse_file_flock,
2982 .splice_read = generic_file_splice_read,
2983 .unlocked_ioctl = fuse_file_ioctl,
2984 .compat_ioctl = fuse_file_compat_ioctl,
2985 .poll = fuse_file_poll,
2986 .fallocate = fuse_file_fallocate,
2989 static const struct file_operations fuse_direct_io_file_operations = {
2990 .llseek = fuse_file_llseek,
2991 .read_iter = fuse_direct_read_iter,
2992 .write_iter = fuse_direct_write_iter,
2993 .mmap = fuse_direct_mmap,
2995 .flush = fuse_flush,
2996 .release = fuse_release,
2997 .fsync = fuse_fsync,
2998 .lock = fuse_file_lock,
2999 .flock = fuse_file_flock,
3000 .unlocked_ioctl = fuse_file_ioctl,
3001 .compat_ioctl = fuse_file_compat_ioctl,
3002 .poll = fuse_file_poll,
3003 .fallocate = fuse_file_fallocate,
3004 /* no splice_read */
3007 static const struct address_space_operations fuse_file_aops = {
3008 .readpage = fuse_readpage,
3009 .writepage = fuse_writepage,
3010 .writepages = fuse_writepages,
3011 .launder_page = fuse_launder_page,
3012 .readpages = fuse_readpages,
3013 .set_page_dirty = __set_page_dirty_nobuffers,
3015 .direct_IO = fuse_direct_IO,
3016 .write_begin = fuse_write_begin,
3017 .write_end = fuse_write_end,
3020 void fuse_init_file_inode(struct inode *inode)
3022 inode->i_fop = &fuse_file_operations;
3023 inode->i_data.a_ops = &fuse_file_aops;