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/aio.h>
19 #include <linux/falloc.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;
30 req = fuse_get_req_nopages(fc);
34 memset(&inarg, 0, sizeof(inarg));
35 inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
36 if (!fc->atomic_o_trunc)
37 inarg.flags &= ~O_TRUNC;
38 req->in.h.opcode = opcode;
39 req->in.h.nodeid = nodeid;
41 req->in.args[0].size = sizeof(inarg);
42 req->in.args[0].value = &inarg;
44 req->out.args[0].size = sizeof(*outargp);
45 req->out.args[0].value = outargp;
46 fuse_request_send(fc, req);
47 err = req->out.h.error;
48 fuse_put_request(fc, req);
53 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
57 ff = kmalloc(sizeof(struct fuse_file), GFP_KERNEL);
62 ff->reserved_req = fuse_request_alloc(0);
63 if (unlikely(!ff->reserved_req)) {
68 INIT_LIST_HEAD(&ff->write_entry);
69 atomic_set(&ff->count, 0);
70 RB_CLEAR_NODE(&ff->polled_node);
71 init_waitqueue_head(&ff->poll_wait);
75 spin_unlock(&fc->lock);
80 void fuse_file_free(struct fuse_file *ff)
82 fuse_request_free(ff->reserved_req);
86 struct fuse_file *fuse_file_get(struct fuse_file *ff)
88 atomic_inc(&ff->count);
92 static void fuse_release_async(struct work_struct *work)
98 req = container_of(work, struct fuse_req, misc.release.work);
99 path = req->misc.release.path;
100 fc = get_fuse_conn(path.dentry->d_inode);
102 fuse_put_request(fc, req);
106 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
108 if (fc->destroy_req) {
110 * If this is a fuseblk mount, then it's possible that
111 * releasing the path will result in releasing the
112 * super block and sending the DESTROY request. If
113 * the server is single threaded, this would hang.
114 * For this reason do the path_put() in a separate
117 atomic_inc(&req->count);
118 INIT_WORK(&req->misc.release.work, fuse_release_async);
119 schedule_work(&req->misc.release.work);
121 path_put(&req->misc.release.path);
125 static void fuse_file_put(struct fuse_file *ff, bool sync)
127 if (atomic_dec_and_test(&ff->count)) {
128 struct fuse_req *req = ff->reserved_req;
132 fuse_request_send(ff->fc, req);
133 path_put(&req->misc.release.path);
134 fuse_put_request(ff->fc, req);
136 req->end = fuse_release_end;
138 fuse_request_send_background(ff->fc, req);
144 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
147 struct fuse_open_out outarg;
148 struct fuse_file *ff;
150 int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
152 ff = fuse_file_alloc(fc);
156 err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
163 outarg.open_flags &= ~FOPEN_DIRECT_IO;
167 ff->open_flags = outarg.open_flags;
168 file->private_data = fuse_file_get(ff);
172 EXPORT_SYMBOL_GPL(fuse_do_open);
174 void fuse_finish_open(struct inode *inode, struct file *file)
176 struct fuse_file *ff = file->private_data;
177 struct fuse_conn *fc = get_fuse_conn(inode);
179 if (ff->open_flags & FOPEN_DIRECT_IO)
180 file->f_op = &fuse_direct_io_file_operations;
181 if (!(ff->open_flags & FOPEN_KEEP_CACHE))
182 invalidate_inode_pages2(inode->i_mapping);
183 if (ff->open_flags & FOPEN_NONSEEKABLE)
184 nonseekable_open(inode, file);
185 if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
186 struct fuse_inode *fi = get_fuse_inode(inode);
188 spin_lock(&fc->lock);
189 fi->attr_version = ++fc->attr_version;
190 i_size_write(inode, 0);
191 spin_unlock(&fc->lock);
192 fuse_invalidate_attr(inode);
196 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
198 struct fuse_conn *fc = get_fuse_conn(inode);
201 err = generic_file_open(inode, file);
205 err = fuse_do_open(fc, get_node_id(inode), file, isdir);
209 fuse_finish_open(inode, file);
214 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
216 struct fuse_conn *fc = ff->fc;
217 struct fuse_req *req = ff->reserved_req;
218 struct fuse_release_in *inarg = &req->misc.release.in;
220 spin_lock(&fc->lock);
221 list_del(&ff->write_entry);
222 if (!RB_EMPTY_NODE(&ff->polled_node))
223 rb_erase(&ff->polled_node, &fc->polled_files);
224 spin_unlock(&fc->lock);
226 wake_up_interruptible_all(&ff->poll_wait);
229 inarg->flags = flags;
230 req->in.h.opcode = opcode;
231 req->in.h.nodeid = ff->nodeid;
233 req->in.args[0].size = sizeof(struct fuse_release_in);
234 req->in.args[0].value = inarg;
237 void fuse_release_common(struct file *file, int opcode)
239 struct fuse_file *ff;
240 struct fuse_req *req;
242 ff = file->private_data;
246 req = ff->reserved_req;
247 fuse_prepare_release(ff, file->f_flags, opcode);
250 struct fuse_release_in *inarg = &req->misc.release.in;
251 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
252 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
255 /* Hold vfsmount and dentry until release is finished */
256 path_get(&file->f_path);
257 req->misc.release.path = file->f_path;
260 * Normally this will send the RELEASE request, however if
261 * some asynchronous READ or WRITE requests are outstanding,
262 * the sending will be delayed.
264 * Make the release synchronous if this is a fuseblk mount,
265 * synchronous RELEASE is allowed (and desirable) in this case
266 * because the server can be trusted not to screw up.
268 fuse_file_put(ff, ff->fc->destroy_req != NULL);
271 static int fuse_open(struct inode *inode, struct file *file)
273 return fuse_open_common(inode, file, false);
276 static int fuse_release(struct inode *inode, struct file *file)
278 fuse_release_common(file, FUSE_RELEASE);
280 /* return value is ignored by VFS */
284 void fuse_sync_release(struct fuse_file *ff, int flags)
286 WARN_ON(atomic_read(&ff->count) > 1);
287 fuse_prepare_release(ff, flags, FUSE_RELEASE);
288 ff->reserved_req->force = 1;
289 ff->reserved_req->background = 0;
290 fuse_request_send(ff->fc, ff->reserved_req);
291 fuse_put_request(ff->fc, ff->reserved_req);
294 EXPORT_SYMBOL_GPL(fuse_sync_release);
297 * Scramble the ID space with XTEA, so that the value of the files_struct
298 * pointer is not exposed to userspace.
300 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
302 u32 *k = fc->scramble_key;
303 u64 v = (unsigned long) id;
309 for (i = 0; i < 32; i++) {
310 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
312 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
315 return (u64) v0 + ((u64) v1 << 32);
319 * Check if page is under writeback
321 * This is currently done by walking the list of writepage requests
322 * for the inode, which can be pretty inefficient.
324 static bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
326 struct fuse_conn *fc = get_fuse_conn(inode);
327 struct fuse_inode *fi = get_fuse_inode(inode);
328 struct fuse_req *req;
331 spin_lock(&fc->lock);
332 list_for_each_entry(req, &fi->writepages, writepages_entry) {
335 BUG_ON(req->inode != inode);
336 curr_index = req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
337 if (curr_index <= index &&
338 index < curr_index + req->num_pages) {
343 spin_unlock(&fc->lock);
349 * Wait for page writeback to be completed.
351 * Since fuse doesn't rely on the VM writeback tracking, this has to
352 * use some other means.
354 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
356 struct fuse_inode *fi = get_fuse_inode(inode);
358 wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
362 static int fuse_flush(struct file *file, fl_owner_t id)
364 struct inode *inode = file_inode(file);
365 struct fuse_conn *fc = get_fuse_conn(inode);
366 struct fuse_file *ff = file->private_data;
367 struct fuse_req *req;
368 struct fuse_flush_in inarg;
371 if (is_bad_inode(inode))
377 req = fuse_get_req_nofail_nopages(fc, file);
378 memset(&inarg, 0, sizeof(inarg));
380 inarg.lock_owner = fuse_lock_owner_id(fc, id);
381 req->in.h.opcode = FUSE_FLUSH;
382 req->in.h.nodeid = get_node_id(inode);
384 req->in.args[0].size = sizeof(inarg);
385 req->in.args[0].value = &inarg;
387 fuse_request_send(fc, req);
388 err = req->out.h.error;
389 fuse_put_request(fc, req);
390 if (err == -ENOSYS) {
398 * Wait for all pending writepages on the inode to finish.
400 * This is currently done by blocking further writes with FUSE_NOWRITE
401 * and waiting for all sent writes to complete.
403 * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
404 * could conflict with truncation.
406 static void fuse_sync_writes(struct inode *inode)
408 fuse_set_nowrite(inode);
409 fuse_release_nowrite(inode);
412 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
413 int datasync, int isdir)
415 struct inode *inode = file->f_mapping->host;
416 struct fuse_conn *fc = get_fuse_conn(inode);
417 struct fuse_file *ff = file->private_data;
418 struct fuse_req *req;
419 struct fuse_fsync_in inarg;
422 if (is_bad_inode(inode))
425 err = filemap_write_and_wait_range(inode->i_mapping, start, end);
429 if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
432 mutex_lock(&inode->i_mutex);
435 * Start writeback against all dirty pages of the inode, then
436 * wait for all outstanding writes, before sending the FSYNC
439 err = write_inode_now(inode, 0);
443 fuse_sync_writes(inode);
445 req = fuse_get_req_nopages(fc);
451 memset(&inarg, 0, sizeof(inarg));
453 inarg.fsync_flags = datasync ? 1 : 0;
454 req->in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
455 req->in.h.nodeid = get_node_id(inode);
457 req->in.args[0].size = sizeof(inarg);
458 req->in.args[0].value = &inarg;
459 fuse_request_send(fc, req);
460 err = req->out.h.error;
461 fuse_put_request(fc, req);
462 if (err == -ENOSYS) {
470 mutex_unlock(&inode->i_mutex);
474 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
477 return fuse_fsync_common(file, start, end, datasync, 0);
480 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
481 size_t count, int opcode)
483 struct fuse_read_in *inarg = &req->misc.read.in;
484 struct fuse_file *ff = file->private_data;
489 inarg->flags = file->f_flags;
490 req->in.h.opcode = opcode;
491 req->in.h.nodeid = ff->nodeid;
493 req->in.args[0].size = sizeof(struct fuse_read_in);
494 req->in.args[0].value = inarg;
496 req->out.numargs = 1;
497 req->out.args[0].size = count;
500 static void fuse_release_user_pages(struct fuse_req *req, int write)
504 for (i = 0; i < req->num_pages; i++) {
505 struct page *page = req->pages[i];
507 set_page_dirty_lock(page);
513 * In case of short read, the caller sets 'pos' to the position of
514 * actual end of fuse request in IO request. Otherwise, if bytes_requested
515 * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
518 * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
519 * both submitted asynchronously. The first of them was ACKed by userspace as
520 * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
521 * second request was ACKed as short, e.g. only 1K was read, resulting in
524 * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
525 * will be equal to the length of the longest contiguous fragment of
526 * transferred data starting from the beginning of IO request.
528 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
532 spin_lock(&io->lock);
534 io->err = io->err ? : err;
535 else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
539 spin_unlock(&io->lock);
546 else if (io->bytes >= 0 && io->write)
549 res = io->bytes < 0 ? io->size : io->bytes;
551 if (!is_sync_kiocb(io->iocb)) {
552 struct inode *inode = file_inode(io->iocb->ki_filp);
553 struct fuse_conn *fc = get_fuse_conn(inode);
554 struct fuse_inode *fi = get_fuse_inode(inode);
556 spin_lock(&fc->lock);
557 fi->attr_version = ++fc->attr_version;
558 spin_unlock(&fc->lock);
562 aio_complete(io->iocb, res, 0);
567 static void fuse_aio_complete_req(struct fuse_conn *fc, struct fuse_req *req)
569 struct fuse_io_priv *io = req->io;
572 fuse_release_user_pages(req, !io->write);
575 if (req->misc.write.in.size != req->misc.write.out.size)
576 pos = req->misc.write.in.offset - io->offset +
577 req->misc.write.out.size;
579 if (req->misc.read.in.size != req->out.args[0].size)
580 pos = req->misc.read.in.offset - io->offset +
581 req->out.args[0].size;
584 fuse_aio_complete(io, req->out.h.error, pos);
587 static size_t fuse_async_req_send(struct fuse_conn *fc, struct fuse_req *req,
588 size_t num_bytes, struct fuse_io_priv *io)
590 spin_lock(&io->lock);
591 io->size += num_bytes;
593 spin_unlock(&io->lock);
596 req->end = fuse_aio_complete_req;
598 __fuse_get_request(req);
599 fuse_request_send_background(fc, req);
604 static size_t fuse_send_read(struct fuse_req *req, struct fuse_io_priv *io,
605 loff_t pos, size_t count, fl_owner_t owner)
607 struct file *file = io->file;
608 struct fuse_file *ff = file->private_data;
609 struct fuse_conn *fc = ff->fc;
611 fuse_read_fill(req, file, pos, count, FUSE_READ);
613 struct fuse_read_in *inarg = &req->misc.read.in;
615 inarg->read_flags |= FUSE_READ_LOCKOWNER;
616 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
620 return fuse_async_req_send(fc, req, count, io);
622 fuse_request_send(fc, req);
623 return req->out.args[0].size;
626 static void fuse_read_update_size(struct inode *inode, loff_t size,
629 struct fuse_conn *fc = get_fuse_conn(inode);
630 struct fuse_inode *fi = get_fuse_inode(inode);
632 spin_lock(&fc->lock);
633 if (attr_ver == fi->attr_version && size < inode->i_size &&
634 !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
635 fi->attr_version = ++fc->attr_version;
636 i_size_write(inode, size);
638 spin_unlock(&fc->lock);
641 static int fuse_readpage(struct file *file, struct page *page)
643 struct fuse_io_priv io = { .async = 0, .file = file };
644 struct inode *inode = page->mapping->host;
645 struct fuse_conn *fc = get_fuse_conn(inode);
646 struct fuse_req *req;
648 loff_t pos = page_offset(page);
649 size_t count = PAGE_CACHE_SIZE;
654 if (is_bad_inode(inode))
658 * Page writeback can extend beyond the lifetime of the
659 * page-cache page, so make sure we read a properly synced
662 fuse_wait_on_page_writeback(inode, page->index);
664 req = fuse_get_req(fc, 1);
669 attr_ver = fuse_get_attr_version(fc);
671 req->out.page_zeroing = 1;
672 req->out.argpages = 1;
674 req->pages[0] = page;
675 req->page_descs[0].length = count;
676 num_read = fuse_send_read(req, &io, pos, count, NULL);
677 err = req->out.h.error;
678 fuse_put_request(fc, req);
682 * Short read means EOF. If file size is larger, truncate it
684 if (num_read < count)
685 fuse_read_update_size(inode, pos + num_read, attr_ver);
687 SetPageUptodate(page);
690 fuse_invalidate_attr(inode); /* atime changed */
696 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
699 size_t count = req->misc.read.in.size;
700 size_t num_read = req->out.args[0].size;
701 struct address_space *mapping = NULL;
703 for (i = 0; mapping == NULL && i < req->num_pages; i++)
704 mapping = req->pages[i]->mapping;
707 struct inode *inode = mapping->host;
710 * Short read means EOF. If file size is larger, truncate it
712 if (!req->out.h.error && num_read < count) {
715 pos = page_offset(req->pages[0]) + num_read;
716 fuse_read_update_size(inode, pos,
717 req->misc.read.attr_ver);
719 fuse_invalidate_attr(inode); /* atime changed */
722 for (i = 0; i < req->num_pages; i++) {
723 struct page *page = req->pages[i];
724 if (!req->out.h.error)
725 SetPageUptodate(page);
729 page_cache_release(page);
732 fuse_file_put(req->ff, false);
735 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
737 struct fuse_file *ff = file->private_data;
738 struct fuse_conn *fc = ff->fc;
739 loff_t pos = page_offset(req->pages[0]);
740 size_t count = req->num_pages << PAGE_CACHE_SHIFT;
742 req->out.argpages = 1;
743 req->out.page_zeroing = 1;
744 req->out.page_replace = 1;
745 fuse_read_fill(req, file, pos, count, FUSE_READ);
746 req->misc.read.attr_ver = fuse_get_attr_version(fc);
747 if (fc->async_read) {
748 req->ff = fuse_file_get(ff);
749 req->end = fuse_readpages_end;
750 fuse_request_send_background(fc, req);
752 fuse_request_send(fc, req);
753 fuse_readpages_end(fc, req);
754 fuse_put_request(fc, req);
758 struct fuse_fill_data {
759 struct fuse_req *req;
765 static int fuse_readpages_fill(void *_data, struct page *page)
767 struct fuse_fill_data *data = _data;
768 struct fuse_req *req = data->req;
769 struct inode *inode = data->inode;
770 struct fuse_conn *fc = get_fuse_conn(inode);
772 fuse_wait_on_page_writeback(inode, page->index);
774 if (req->num_pages &&
775 (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
776 (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
777 req->pages[req->num_pages - 1]->index + 1 != page->index)) {
778 int nr_alloc = min_t(unsigned, data->nr_pages,
779 FUSE_MAX_PAGES_PER_REQ);
780 fuse_send_readpages(req, data->file);
782 req = fuse_get_req_for_background(fc, nr_alloc);
784 req = fuse_get_req(fc, nr_alloc);
793 if (WARN_ON(req->num_pages >= req->max_pages)) {
794 fuse_put_request(fc, req);
798 page_cache_get(page);
799 req->pages[req->num_pages] = page;
800 req->page_descs[req->num_pages].length = PAGE_SIZE;
806 static int fuse_readpages(struct file *file, struct address_space *mapping,
807 struct list_head *pages, unsigned nr_pages)
809 struct inode *inode = mapping->host;
810 struct fuse_conn *fc = get_fuse_conn(inode);
811 struct fuse_fill_data data;
813 int nr_alloc = min_t(unsigned, nr_pages, FUSE_MAX_PAGES_PER_REQ);
816 if (is_bad_inode(inode))
822 data.req = fuse_get_req_for_background(fc, nr_alloc);
824 data.req = fuse_get_req(fc, nr_alloc);
825 data.nr_pages = nr_pages;
826 err = PTR_ERR(data.req);
827 if (IS_ERR(data.req))
830 err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
832 if (data.req->num_pages)
833 fuse_send_readpages(data.req, file);
835 fuse_put_request(fc, data.req);
841 static ssize_t fuse_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
842 unsigned long nr_segs, loff_t pos)
844 struct inode *inode = iocb->ki_filp->f_mapping->host;
845 struct fuse_conn *fc = get_fuse_conn(inode);
848 * In auto invalidate mode, always update attributes on read.
849 * Otherwise, only update if we attempt to read past EOF (to ensure
850 * i_size is up to date).
852 if (fc->auto_inval_data ||
853 (pos + iov_length(iov, nr_segs) > i_size_read(inode))) {
855 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
860 return generic_file_aio_read(iocb, iov, nr_segs, pos);
863 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
864 loff_t pos, size_t count)
866 struct fuse_write_in *inarg = &req->misc.write.in;
867 struct fuse_write_out *outarg = &req->misc.write.out;
872 req->in.h.opcode = FUSE_WRITE;
873 req->in.h.nodeid = ff->nodeid;
875 if (ff->fc->minor < 9)
876 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
878 req->in.args[0].size = sizeof(struct fuse_write_in);
879 req->in.args[0].value = inarg;
880 req->in.args[1].size = count;
881 req->out.numargs = 1;
882 req->out.args[0].size = sizeof(struct fuse_write_out);
883 req->out.args[0].value = outarg;
886 static size_t fuse_send_write(struct fuse_req *req, struct fuse_io_priv *io,
887 loff_t pos, size_t count, fl_owner_t owner)
889 struct file *file = io->file;
890 struct fuse_file *ff = file->private_data;
891 struct fuse_conn *fc = ff->fc;
892 struct fuse_write_in *inarg = &req->misc.write.in;
894 fuse_write_fill(req, ff, pos, count);
895 inarg->flags = file->f_flags;
897 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
898 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
902 return fuse_async_req_send(fc, req, count, io);
904 fuse_request_send(fc, req);
905 return req->misc.write.out.size;
908 void fuse_write_update_size(struct inode *inode, loff_t pos)
910 struct fuse_conn *fc = get_fuse_conn(inode);
911 struct fuse_inode *fi = get_fuse_inode(inode);
913 spin_lock(&fc->lock);
914 fi->attr_version = ++fc->attr_version;
915 if (pos > inode->i_size)
916 i_size_write(inode, pos);
917 spin_unlock(&fc->lock);
920 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
921 struct inode *inode, loff_t pos,
927 struct fuse_io_priv io = { .async = 0, .file = file };
929 for (i = 0; i < req->num_pages; i++)
930 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
932 res = fuse_send_write(req, &io, pos, count, NULL);
934 offset = req->page_descs[0].offset;
936 for (i = 0; i < req->num_pages; i++) {
937 struct page *page = req->pages[i];
939 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
940 SetPageUptodate(page);
942 if (count > PAGE_CACHE_SIZE - offset)
943 count -= PAGE_CACHE_SIZE - offset;
949 page_cache_release(page);
955 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
956 struct address_space *mapping,
957 struct iov_iter *ii, loff_t pos)
959 struct fuse_conn *fc = get_fuse_conn(mapping->host);
960 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
964 req->in.argpages = 1;
965 req->page_descs[0].offset = offset;
970 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
971 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
974 bytes = min_t(size_t, bytes, fc->max_write - count);
978 if (iov_iter_fault_in_readable(ii, bytes))
982 page = grab_cache_page_write_begin(mapping, index, 0);
986 if (mapping_writably_mapped(mapping))
987 flush_dcache_page(page);
990 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
992 flush_dcache_page(page);
994 mark_page_accessed(page);
998 page_cache_release(page);
999 bytes = min(bytes, iov_iter_single_seg_count(ii));
1004 req->pages[req->num_pages] = page;
1005 req->page_descs[req->num_pages].length = tmp;
1008 iov_iter_advance(ii, tmp);
1012 if (offset == PAGE_CACHE_SIZE)
1015 if (!fc->big_writes)
1017 } while (iov_iter_count(ii) && count < fc->max_write &&
1018 req->num_pages < req->max_pages && offset == 0);
1020 return count > 0 ? count : err;
1023 static inline unsigned fuse_wr_pages(loff_t pos, size_t len)
1025 return min_t(unsigned,
1026 ((pos + len - 1) >> PAGE_CACHE_SHIFT) -
1027 (pos >> PAGE_CACHE_SHIFT) + 1,
1028 FUSE_MAX_PAGES_PER_REQ);
1031 static ssize_t fuse_perform_write(struct file *file,
1032 struct address_space *mapping,
1033 struct iov_iter *ii, loff_t pos)
1035 struct inode *inode = mapping->host;
1036 struct fuse_conn *fc = get_fuse_conn(inode);
1037 struct fuse_inode *fi = get_fuse_inode(inode);
1041 if (is_bad_inode(inode))
1044 if (inode->i_size < pos + iov_iter_count(ii))
1045 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1048 struct fuse_req *req;
1050 unsigned nr_pages = fuse_wr_pages(pos, iov_iter_count(ii));
1052 req = fuse_get_req(fc, nr_pages);
1058 count = fuse_fill_write_pages(req, mapping, ii, pos);
1064 num_written = fuse_send_write_pages(req, file, inode,
1066 err = req->out.h.error;
1071 /* break out of the loop on short write */
1072 if (num_written != count)
1076 fuse_put_request(fc, req);
1077 } while (!err && iov_iter_count(ii));
1080 fuse_write_update_size(inode, pos);
1082 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1083 fuse_invalidate_attr(inode);
1085 return res > 0 ? res : err;
1088 static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
1089 unsigned long nr_segs, loff_t pos)
1091 struct file *file = iocb->ki_filp;
1092 struct address_space *mapping = file->f_mapping;
1095 ssize_t written = 0;
1096 ssize_t written_buffered = 0;
1097 struct inode *inode = mapping->host;
1102 WARN_ON(iocb->ki_pos != pos);
1105 err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
1110 mutex_lock(&inode->i_mutex);
1112 /* We can write back this queue in page reclaim */
1113 current->backing_dev_info = mapping->backing_dev_info;
1115 err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
1122 err = file_remove_suid(file);
1126 err = file_update_time(file);
1130 if (file->f_flags & O_DIRECT) {
1131 written = generic_file_direct_write(iocb, iov, &nr_segs,
1134 if (written < 0 || written == count)
1140 iov_iter_init(&i, iov, nr_segs, count, written);
1141 written_buffered = fuse_perform_write(file, mapping, &i, pos);
1142 if (written_buffered < 0) {
1143 err = written_buffered;
1146 endbyte = pos + written_buffered - 1;
1148 err = filemap_write_and_wait_range(file->f_mapping, pos,
1153 invalidate_mapping_pages(file->f_mapping,
1154 pos >> PAGE_CACHE_SHIFT,
1155 endbyte >> PAGE_CACHE_SHIFT);
1157 written += written_buffered;
1158 iocb->ki_pos = pos + written_buffered;
1160 iov_iter_init(&i, iov, nr_segs, count, 0);
1161 written = fuse_perform_write(file, mapping, &i, pos);
1163 iocb->ki_pos = pos + written;
1166 current->backing_dev_info = NULL;
1167 mutex_unlock(&inode->i_mutex);
1169 return written ? written : err;
1172 static inline void fuse_page_descs_length_init(struct fuse_req *req,
1173 unsigned index, unsigned nr_pages)
1177 for (i = index; i < index + nr_pages; i++)
1178 req->page_descs[i].length = PAGE_SIZE -
1179 req->page_descs[i].offset;
1182 static inline unsigned long fuse_get_user_addr(struct iov_iter *ii)
1184 struct iovec *iov = iov_iter_iovec(ii);
1185 return (unsigned long)iov->iov_base + ii->iov_offset;
1188 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1191 return min(iov_iter_single_seg_count(ii), max_size);
1194 static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1195 size_t *nbytesp, int write)
1197 size_t nbytes = 0; /* # bytes already packed in req */
1199 /* Special case for kernel I/O: can copy directly into the buffer */
1200 if (segment_eq(get_fs(), KERNEL_DS)) {
1201 unsigned long user_addr = fuse_get_user_addr(ii);
1202 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1205 req->in.args[1].value = (void *) user_addr;
1207 req->out.args[0].value = (void *) user_addr;
1209 iov_iter_advance(ii, frag_size);
1210 *nbytesp = frag_size;
1214 while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1216 unsigned long user_addr = fuse_get_user_addr(ii);
1217 unsigned offset = user_addr & ~PAGE_MASK;
1218 size_t frag_size = fuse_get_frag_size(ii, *nbytesp - nbytes);
1221 unsigned n = req->max_pages - req->num_pages;
1222 frag_size = min_t(size_t, frag_size, n << PAGE_SHIFT);
1224 npages = (frag_size + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1225 npages = clamp(npages, 1U, n);
1227 ret = get_user_pages_fast(user_addr, npages, !write,
1228 &req->pages[req->num_pages]);
1233 frag_size = min_t(size_t, frag_size,
1234 (npages << PAGE_SHIFT) - offset);
1235 iov_iter_advance(ii, frag_size);
1237 req->page_descs[req->num_pages].offset = offset;
1238 fuse_page_descs_length_init(req, req->num_pages, npages);
1240 req->num_pages += npages;
1241 req->page_descs[req->num_pages - 1].length -=
1242 (npages << PAGE_SHIFT) - offset - frag_size;
1244 nbytes += frag_size;
1248 req->in.argpages = 1;
1250 req->out.argpages = 1;
1257 static inline int fuse_iter_npages(const struct iov_iter *ii_p)
1259 struct iov_iter ii = *ii_p;
1262 while (iov_iter_count(&ii) && npages < FUSE_MAX_PAGES_PER_REQ) {
1263 unsigned long user_addr = fuse_get_user_addr(&ii);
1264 unsigned offset = user_addr & ~PAGE_MASK;
1265 size_t frag_size = iov_iter_single_seg_count(&ii);
1267 npages += (frag_size + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1268 iov_iter_advance(&ii, frag_size);
1271 return min(npages, FUSE_MAX_PAGES_PER_REQ);
1274 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *ii,
1275 size_t count, loff_t *ppos, int write)
1277 struct file *file = io->file;
1278 struct fuse_file *ff = file->private_data;
1279 struct fuse_conn *fc = ff->fc;
1280 size_t nmax = write ? fc->max_write : fc->max_read;
1283 struct fuse_req *req;
1286 req = fuse_get_req_for_background(fc, fuse_iter_npages(ii));
1288 req = fuse_get_req(fc, fuse_iter_npages(ii));
1290 return PTR_ERR(req);
1294 fl_owner_t owner = current->files;
1295 size_t nbytes = min(count, nmax);
1296 int err = fuse_get_user_pages(req, ii, &nbytes, write);
1303 nres = fuse_send_write(req, io, pos, nbytes, owner);
1305 nres = fuse_send_read(req, io, pos, nbytes, owner);
1308 fuse_release_user_pages(req, !write);
1309 if (req->out.h.error) {
1311 res = req->out.h.error;
1313 } else if (nres > nbytes) {
1323 fuse_put_request(fc, req);
1325 req = fuse_get_req_for_background(fc,
1326 fuse_iter_npages(ii));
1328 req = fuse_get_req(fc, fuse_iter_npages(ii));
1334 fuse_put_request(fc, req);
1340 EXPORT_SYMBOL_GPL(fuse_direct_io);
1342 static ssize_t __fuse_direct_read(struct fuse_io_priv *io, struct iov_iter *ii,
1343 loff_t *ppos, size_t count)
1346 struct file *file = io->file;
1347 struct inode *inode = file_inode(file);
1349 if (is_bad_inode(inode))
1352 res = fuse_direct_io(io, ii, count, ppos, 0);
1354 fuse_invalidate_attr(inode);
1359 static ssize_t fuse_direct_read(struct file *file, char __user *buf,
1360 size_t count, loff_t *ppos)
1362 struct fuse_io_priv io = { .async = 0, .file = file };
1363 struct iovec iov = { .iov_base = buf, .iov_len = count };
1366 iov_iter_init(&ii, &iov, 1, count, 0);
1368 return __fuse_direct_read(&io, &ii, ppos, count);
1371 static ssize_t __fuse_direct_write(struct fuse_io_priv *io, struct iov_iter *ii,
1374 struct file *file = io->file;
1375 struct inode *inode = file_inode(file);
1376 size_t count = iov_iter_count(ii);
1379 res = generic_write_checks(file, ppos, &count, 0);
1381 res = fuse_direct_io(io, ii, count, ppos, 1);
1383 fuse_invalidate_attr(inode);
1388 static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
1389 size_t count, loff_t *ppos)
1391 struct iovec iov = { .iov_base = (void __user *)buf, .iov_len = count };
1393 struct inode *inode = file_inode(file);
1395 struct fuse_io_priv io = { .async = 0, .file = file };
1397 if (is_bad_inode(inode))
1400 iov_iter_init(&ii, &iov, 1, count, 0);
1402 /* Don't allow parallel writes to the same file */
1403 mutex_lock(&inode->i_mutex);
1404 res = __fuse_direct_write(&io, &ii, ppos);
1406 fuse_write_update_size(inode, *ppos);
1407 mutex_unlock(&inode->i_mutex);
1412 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1416 for (i = 0; i < req->num_pages; i++)
1417 __free_page(req->pages[i]);
1420 fuse_file_put(req->ff, false);
1423 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1425 struct inode *inode = req->inode;
1426 struct fuse_inode *fi = get_fuse_inode(inode);
1427 struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
1430 list_del(&req->writepages_entry);
1431 for (i = 0; i < req->num_pages; i++) {
1432 dec_bdi_stat(bdi, BDI_WRITEBACK);
1433 dec_zone_page_state(req->pages[i], NR_WRITEBACK_TEMP);
1434 bdi_writeout_inc(bdi);
1436 wake_up(&fi->page_waitq);
1439 /* Called under fc->lock, may release and reacquire it */
1440 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req,
1442 __releases(fc->lock)
1443 __acquires(fc->lock)
1445 struct fuse_inode *fi = get_fuse_inode(req->inode);
1446 struct fuse_write_in *inarg = &req->misc.write.in;
1447 __u64 data_size = req->num_pages * PAGE_CACHE_SIZE;
1452 if (inarg->offset + data_size <= size) {
1453 inarg->size = data_size;
1454 } else if (inarg->offset < size) {
1455 inarg->size = size - inarg->offset;
1457 /* Got truncated off completely */
1461 req->in.args[1].size = inarg->size;
1463 fuse_request_send_background_locked(fc, req);
1467 fuse_writepage_finish(fc, req);
1468 spin_unlock(&fc->lock);
1469 fuse_writepage_free(fc, req);
1470 fuse_put_request(fc, req);
1471 spin_lock(&fc->lock);
1475 * If fi->writectr is positive (no truncate or fsync going on) send
1476 * all queued writepage requests.
1478 * Called with fc->lock
1480 void fuse_flush_writepages(struct inode *inode)
1481 __releases(fc->lock)
1482 __acquires(fc->lock)
1484 struct fuse_conn *fc = get_fuse_conn(inode);
1485 struct fuse_inode *fi = get_fuse_inode(inode);
1486 size_t crop = i_size_read(inode);
1487 struct fuse_req *req;
1489 while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1490 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1491 list_del_init(&req->list);
1492 fuse_send_writepage(fc, req, crop);
1496 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1498 struct inode *inode = req->inode;
1499 struct fuse_inode *fi = get_fuse_inode(inode);
1501 mapping_set_error(inode->i_mapping, req->out.h.error);
1502 spin_lock(&fc->lock);
1503 while (req->misc.write.next) {
1504 struct fuse_conn *fc = get_fuse_conn(inode);
1505 struct fuse_write_in *inarg = &req->misc.write.in;
1506 struct fuse_req *next = req->misc.write.next;
1507 req->misc.write.next = next->misc.write.next;
1508 next->misc.write.next = NULL;
1509 next->ff = fuse_file_get(req->ff);
1510 list_add(&next->writepages_entry, &fi->writepages);
1513 * Skip fuse_flush_writepages() to make it easy to crop requests
1514 * based on primary request size.
1516 * 1st case (trivial): there are no concurrent activities using
1517 * fuse_set/release_nowrite. Then we're on safe side because
1518 * fuse_flush_writepages() would call fuse_send_writepage()
1521 * 2nd case: someone called fuse_set_nowrite and it is waiting
1522 * now for completion of all in-flight requests. This happens
1523 * rarely and no more than once per page, so this should be
1526 * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1527 * of fuse_set_nowrite..fuse_release_nowrite section. The fact
1528 * that fuse_set_nowrite returned implies that all in-flight
1529 * requests were completed along with all of their secondary
1530 * requests. Further primary requests are blocked by negative
1531 * writectr. Hence there cannot be any in-flight requests and
1532 * no invocations of fuse_writepage_end() while we're in
1533 * fuse_set_nowrite..fuse_release_nowrite section.
1535 fuse_send_writepage(fc, next, inarg->offset + inarg->size);
1538 fuse_writepage_finish(fc, req);
1539 spin_unlock(&fc->lock);
1540 fuse_writepage_free(fc, req);
1543 static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1544 struct fuse_inode *fi)
1546 struct fuse_file *ff = NULL;
1548 spin_lock(&fc->lock);
1549 if (!WARN_ON(list_empty(&fi->write_files))) {
1550 ff = list_entry(fi->write_files.next, struct fuse_file,
1554 spin_unlock(&fc->lock);
1559 static int fuse_writepage_locked(struct page *page)
1561 struct address_space *mapping = page->mapping;
1562 struct inode *inode = mapping->host;
1563 struct fuse_conn *fc = get_fuse_conn(inode);
1564 struct fuse_inode *fi = get_fuse_inode(inode);
1565 struct fuse_req *req;
1566 struct page *tmp_page;
1567 int error = -ENOMEM;
1569 set_page_writeback(page);
1571 req = fuse_request_alloc_nofs(1);
1575 req->background = 1; /* writeback always goes to bg_queue */
1576 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1581 req->ff = fuse_write_file_get(fc, fi);
1585 fuse_write_fill(req, req->ff, page_offset(page), 0);
1587 copy_highpage(tmp_page, page);
1588 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1589 req->misc.write.next = NULL;
1590 req->in.argpages = 1;
1592 req->pages[0] = tmp_page;
1593 req->page_descs[0].offset = 0;
1594 req->page_descs[0].length = PAGE_SIZE;
1595 req->end = fuse_writepage_end;
1598 inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
1599 inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1601 spin_lock(&fc->lock);
1602 list_add(&req->writepages_entry, &fi->writepages);
1603 list_add_tail(&req->list, &fi->queued_writes);
1604 fuse_flush_writepages(inode);
1605 spin_unlock(&fc->lock);
1607 end_page_writeback(page);
1612 fuse_request_free(req);
1614 end_page_writeback(page);
1618 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1622 if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1624 * ->writepages() should be called for sync() and friends. We
1625 * should only get here on direct reclaim and then we are
1626 * allowed to skip a page which is already in flight
1628 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1630 redirty_page_for_writepage(wbc, page);
1634 err = fuse_writepage_locked(page);
1640 struct fuse_fill_wb_data {
1641 struct fuse_req *req;
1642 struct fuse_file *ff;
1643 struct inode *inode;
1644 struct page **orig_pages;
1647 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1649 struct fuse_req *req = data->req;
1650 struct inode *inode = data->inode;
1651 struct fuse_conn *fc = get_fuse_conn(inode);
1652 struct fuse_inode *fi = get_fuse_inode(inode);
1653 int num_pages = req->num_pages;
1656 req->ff = fuse_file_get(data->ff);
1657 spin_lock(&fc->lock);
1658 list_add_tail(&req->list, &fi->queued_writes);
1659 fuse_flush_writepages(inode);
1660 spin_unlock(&fc->lock);
1662 for (i = 0; i < num_pages; i++)
1663 end_page_writeback(data->orig_pages[i]);
1666 static bool fuse_writepage_in_flight(struct fuse_req *new_req,
1669 struct fuse_conn *fc = get_fuse_conn(new_req->inode);
1670 struct fuse_inode *fi = get_fuse_inode(new_req->inode);
1671 struct fuse_req *tmp;
1672 struct fuse_req *old_req;
1676 BUG_ON(new_req->num_pages != 0);
1678 spin_lock(&fc->lock);
1679 list_del(&new_req->writepages_entry);
1680 list_for_each_entry(old_req, &fi->writepages, writepages_entry) {
1681 BUG_ON(old_req->inode != new_req->inode);
1682 curr_index = old_req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
1683 if (curr_index <= page->index &&
1684 page->index < curr_index + old_req->num_pages) {
1690 list_add(&new_req->writepages_entry, &fi->writepages);
1694 new_req->num_pages = 1;
1695 for (tmp = old_req; tmp != NULL; tmp = tmp->misc.write.next) {
1696 BUG_ON(tmp->inode != new_req->inode);
1697 curr_index = tmp->misc.write.in.offset >> PAGE_CACHE_SHIFT;
1698 if (tmp->num_pages == 1 &&
1699 curr_index == page->index) {
1704 if (old_req->num_pages == 1 && (old_req->state == FUSE_REQ_INIT ||
1705 old_req->state == FUSE_REQ_PENDING)) {
1706 struct backing_dev_info *bdi = page->mapping->backing_dev_info;
1708 copy_highpage(old_req->pages[0], page);
1709 spin_unlock(&fc->lock);
1711 dec_bdi_stat(bdi, BDI_WRITEBACK);
1712 dec_zone_page_state(page, NR_WRITEBACK_TEMP);
1713 bdi_writeout_inc(bdi);
1714 fuse_writepage_free(fc, new_req);
1715 fuse_request_free(new_req);
1718 new_req->misc.write.next = old_req->misc.write.next;
1719 old_req->misc.write.next = new_req;
1722 spin_unlock(&fc->lock);
1727 static int fuse_writepages_fill(struct page *page,
1728 struct writeback_control *wbc, void *_data)
1730 struct fuse_fill_wb_data *data = _data;
1731 struct fuse_req *req = data->req;
1732 struct inode *inode = data->inode;
1733 struct fuse_conn *fc = get_fuse_conn(inode);
1734 struct page *tmp_page;
1740 data->ff = fuse_write_file_get(fc, get_fuse_inode(inode));
1746 * Being under writeback is unlikely but possible. For example direct
1747 * read to an mmaped fuse file will set the page dirty twice; once when
1748 * the pages are faulted with get_user_pages(), and then after the read
1751 is_writeback = fuse_page_is_writeback(inode, page->index);
1753 if (req && req->num_pages &&
1754 (is_writeback || req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
1755 (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_write ||
1756 data->orig_pages[req->num_pages - 1]->index + 1 != page->index)) {
1757 fuse_writepages_send(data);
1761 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1766 * The page must not be redirtied until the writeout is completed
1767 * (i.e. userspace has sent a reply to the write request). Otherwise
1768 * there could be more than one temporary page instance for each real
1771 * This is ensured by holding the page lock in page_mkwrite() while
1772 * checking fuse_page_is_writeback(). We already hold the page lock
1773 * since clear_page_dirty_for_io() and keep it held until we add the
1774 * request to the fi->writepages list and increment req->num_pages.
1775 * After this fuse_page_is_writeback() will indicate that the page is
1776 * under writeback, so we can release the page lock.
1778 if (data->req == NULL) {
1779 struct fuse_inode *fi = get_fuse_inode(inode);
1782 req = fuse_request_alloc_nofs(FUSE_MAX_PAGES_PER_REQ);
1784 __free_page(tmp_page);
1788 fuse_write_fill(req, data->ff, page_offset(page), 0);
1789 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1790 req->misc.write.next = NULL;
1791 req->in.argpages = 1;
1792 req->background = 1;
1794 req->end = fuse_writepage_end;
1797 spin_lock(&fc->lock);
1798 list_add(&req->writepages_entry, &fi->writepages);
1799 spin_unlock(&fc->lock);
1803 set_page_writeback(page);
1805 copy_highpage(tmp_page, page);
1806 req->pages[req->num_pages] = tmp_page;
1807 req->page_descs[req->num_pages].offset = 0;
1808 req->page_descs[req->num_pages].length = PAGE_SIZE;
1810 inc_bdi_stat(page->mapping->backing_dev_info, BDI_WRITEBACK);
1811 inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1814 if (is_writeback && fuse_writepage_in_flight(req, page)) {
1815 end_page_writeback(page);
1819 data->orig_pages[req->num_pages] = page;
1822 * Protected by fc->lock against concurrent access by
1823 * fuse_page_is_writeback().
1825 spin_lock(&fc->lock);
1827 spin_unlock(&fc->lock);
1835 static int fuse_writepages(struct address_space *mapping,
1836 struct writeback_control *wbc)
1838 struct inode *inode = mapping->host;
1839 struct fuse_fill_wb_data data;
1843 if (is_bad_inode(inode))
1851 data.orig_pages = kzalloc(sizeof(struct page *) *
1852 FUSE_MAX_PAGES_PER_REQ,
1854 if (!data.orig_pages)
1857 err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
1859 /* Ignore errors if we can write at least one page */
1860 BUG_ON(!data.req->num_pages);
1861 fuse_writepages_send(&data);
1865 fuse_file_put(data.ff, false);
1867 kfree(data.orig_pages);
1872 static int fuse_launder_page(struct page *page)
1875 if (clear_page_dirty_for_io(page)) {
1876 struct inode *inode = page->mapping->host;
1877 err = fuse_writepage_locked(page);
1879 fuse_wait_on_page_writeback(inode, page->index);
1885 * Write back dirty pages now, because there may not be any suitable
1888 static void fuse_vma_close(struct vm_area_struct *vma)
1890 filemap_write_and_wait(vma->vm_file->f_mapping);
1894 * Wait for writeback against this page to complete before allowing it
1895 * to be marked dirty again, and hence written back again, possibly
1896 * before the previous writepage completed.
1898 * Block here, instead of in ->writepage(), so that the userspace fs
1899 * can only block processes actually operating on the filesystem.
1901 * Otherwise unprivileged userspace fs would be able to block
1906 * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
1908 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1910 struct page *page = vmf->page;
1911 struct inode *inode = file_inode(vma->vm_file);
1913 file_update_time(vma->vm_file);
1915 if (page->mapping != inode->i_mapping) {
1917 return VM_FAULT_NOPAGE;
1920 fuse_wait_on_page_writeback(inode, page->index);
1921 return VM_FAULT_LOCKED;
1924 static const struct vm_operations_struct fuse_file_vm_ops = {
1925 .close = fuse_vma_close,
1926 .fault = filemap_fault,
1927 .page_mkwrite = fuse_page_mkwrite,
1928 .remap_pages = generic_file_remap_pages,
1931 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
1933 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE)) {
1934 struct inode *inode = file_inode(file);
1935 struct fuse_conn *fc = get_fuse_conn(inode);
1936 struct fuse_inode *fi = get_fuse_inode(inode);
1937 struct fuse_file *ff = file->private_data;
1939 * file may be written through mmap, so chain it onto the
1940 * inodes's write_file list
1942 spin_lock(&fc->lock);
1943 if (list_empty(&ff->write_entry))
1944 list_add(&ff->write_entry, &fi->write_files);
1945 spin_unlock(&fc->lock);
1947 file_accessed(file);
1948 vma->vm_ops = &fuse_file_vm_ops;
1952 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
1954 /* Can't provide the coherency needed for MAP_SHARED */
1955 if (vma->vm_flags & VM_MAYSHARE)
1958 invalidate_inode_pages2(file->f_mapping);
1960 return generic_file_mmap(file, vma);
1963 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
1964 struct file_lock *fl)
1966 switch (ffl->type) {
1972 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
1973 ffl->end < ffl->start)
1976 fl->fl_start = ffl->start;
1977 fl->fl_end = ffl->end;
1978 fl->fl_pid = ffl->pid;
1984 fl->fl_type = ffl->type;
1988 static void fuse_lk_fill(struct fuse_req *req, struct file *file,
1989 const struct file_lock *fl, int opcode, pid_t pid,
1992 struct inode *inode = file_inode(file);
1993 struct fuse_conn *fc = get_fuse_conn(inode);
1994 struct fuse_file *ff = file->private_data;
1995 struct fuse_lk_in *arg = &req->misc.lk_in;
1998 arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
1999 arg->lk.start = fl->fl_start;
2000 arg->lk.end = fl->fl_end;
2001 arg->lk.type = fl->fl_type;
2004 arg->lk_flags |= FUSE_LK_FLOCK;
2005 req->in.h.opcode = opcode;
2006 req->in.h.nodeid = get_node_id(inode);
2007 req->in.numargs = 1;
2008 req->in.args[0].size = sizeof(*arg);
2009 req->in.args[0].value = arg;
2012 static int fuse_getlk(struct file *file, struct file_lock *fl)
2014 struct inode *inode = file_inode(file);
2015 struct fuse_conn *fc = get_fuse_conn(inode);
2016 struct fuse_req *req;
2017 struct fuse_lk_out outarg;
2020 req = fuse_get_req_nopages(fc);
2022 return PTR_ERR(req);
2024 fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
2025 req->out.numargs = 1;
2026 req->out.args[0].size = sizeof(outarg);
2027 req->out.args[0].value = &outarg;
2028 fuse_request_send(fc, req);
2029 err = req->out.h.error;
2030 fuse_put_request(fc, req);
2032 err = convert_fuse_file_lock(&outarg.lk, fl);
2037 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2039 struct inode *inode = file_inode(file);
2040 struct fuse_conn *fc = get_fuse_conn(inode);
2041 struct fuse_req *req;
2042 int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2043 pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
2046 if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2047 /* NLM needs asynchronous locks, which we don't support yet */
2051 /* Unlock on close is handled by the flush method */
2052 if (fl->fl_flags & FL_CLOSE)
2055 req = fuse_get_req_nopages(fc);
2057 return PTR_ERR(req);
2059 fuse_lk_fill(req, file, fl, opcode, pid, flock);
2060 fuse_request_send(fc, req);
2061 err = req->out.h.error;
2062 /* locking is restartable */
2065 fuse_put_request(fc, req);
2069 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2071 struct inode *inode = file_inode(file);
2072 struct fuse_conn *fc = get_fuse_conn(inode);
2075 if (cmd == F_CANCELLK) {
2077 } else if (cmd == F_GETLK) {
2079 posix_test_lock(file, fl);
2082 err = fuse_getlk(file, fl);
2085 err = posix_lock_file(file, fl, NULL);
2087 err = fuse_setlk(file, fl, 0);
2092 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2094 struct inode *inode = file_inode(file);
2095 struct fuse_conn *fc = get_fuse_conn(inode);
2099 err = flock_lock_file_wait(file, fl);
2101 struct fuse_file *ff = file->private_data;
2103 /* emulate flock with POSIX locks */
2104 fl->fl_owner = (fl_owner_t) file;
2106 err = fuse_setlk(file, fl, 1);
2112 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2114 struct inode *inode = mapping->host;
2115 struct fuse_conn *fc = get_fuse_conn(inode);
2116 struct fuse_req *req;
2117 struct fuse_bmap_in inarg;
2118 struct fuse_bmap_out outarg;
2121 if (!inode->i_sb->s_bdev || fc->no_bmap)
2124 req = fuse_get_req_nopages(fc);
2128 memset(&inarg, 0, sizeof(inarg));
2129 inarg.block = block;
2130 inarg.blocksize = inode->i_sb->s_blocksize;
2131 req->in.h.opcode = FUSE_BMAP;
2132 req->in.h.nodeid = get_node_id(inode);
2133 req->in.numargs = 1;
2134 req->in.args[0].size = sizeof(inarg);
2135 req->in.args[0].value = &inarg;
2136 req->out.numargs = 1;
2137 req->out.args[0].size = sizeof(outarg);
2138 req->out.args[0].value = &outarg;
2139 fuse_request_send(fc, req);
2140 err = req->out.h.error;
2141 fuse_put_request(fc, req);
2145 return err ? 0 : outarg.block;
2148 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2151 struct inode *inode = file_inode(file);
2153 /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2154 if (whence == SEEK_CUR || whence == SEEK_SET)
2155 return generic_file_llseek(file, offset, whence);
2157 mutex_lock(&inode->i_mutex);
2158 retval = fuse_update_attributes(inode, NULL, file, NULL);
2160 retval = generic_file_llseek(file, offset, whence);
2161 mutex_unlock(&inode->i_mutex);
2166 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
2167 unsigned int nr_segs, size_t bytes, bool to_user)
2175 iov_iter_init(&ii, iov, nr_segs, bytes, 0);
2177 while (iov_iter_count(&ii)) {
2178 struct page *page = pages[page_idx++];
2179 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
2183 left = iov_iter_copy_from_user(page, &ii, 0, todo);
2185 left = iov_iter_copy_to_user(page, &ii, 0, todo);
2190 iov_iter_advance(&ii, todo);
2197 * CUSE servers compiled on 32bit broke on 64bit kernels because the
2198 * ABI was defined to be 'struct iovec' which is different on 32bit
2199 * and 64bit. Fortunately we can determine which structure the server
2200 * used from the size of the reply.
2202 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2203 size_t transferred, unsigned count,
2206 #ifdef CONFIG_COMPAT
2207 if (count * sizeof(struct compat_iovec) == transferred) {
2208 struct compat_iovec *ciov = src;
2212 * With this interface a 32bit server cannot support
2213 * non-compat (i.e. ones coming from 64bit apps) ioctl
2219 for (i = 0; i < count; i++) {
2220 dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2221 dst[i].iov_len = ciov[i].iov_len;
2227 if (count * sizeof(struct iovec) != transferred)
2230 memcpy(dst, src, transferred);
2234 /* Make sure iov_length() won't overflow */
2235 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
2238 u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
2240 for (n = 0; n < count; n++, iov++) {
2241 if (iov->iov_len > (size_t) max)
2243 max -= iov->iov_len;
2248 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2249 void *src, size_t transferred, unsigned count,
2253 struct fuse_ioctl_iovec *fiov = src;
2255 if (fc->minor < 16) {
2256 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2260 if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2263 for (i = 0; i < count; i++) {
2264 /* Did the server supply an inappropriate value? */
2265 if (fiov[i].base != (unsigned long) fiov[i].base ||
2266 fiov[i].len != (unsigned long) fiov[i].len)
2269 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2270 dst[i].iov_len = (size_t) fiov[i].len;
2272 #ifdef CONFIG_COMPAT
2274 (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2275 (compat_size_t) dst[i].iov_len != fiov[i].len))
2285 * For ioctls, there is no generic way to determine how much memory
2286 * needs to be read and/or written. Furthermore, ioctls are allowed
2287 * to dereference the passed pointer, so the parameter requires deep
2288 * copying but FUSE has no idea whatsoever about what to copy in or
2291 * This is solved by allowing FUSE server to retry ioctl with
2292 * necessary in/out iovecs. Let's assume the ioctl implementation
2293 * needs to read in the following structure.
2300 * On the first callout to FUSE server, inarg->in_size and
2301 * inarg->out_size will be NULL; then, the server completes the ioctl
2302 * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2303 * the actual iov array to
2305 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) } }
2307 * which tells FUSE to copy in the requested area and retry the ioctl.
2308 * On the second round, the server has access to the structure and
2309 * from that it can tell what to look for next, so on the invocation,
2310 * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2312 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) },
2313 * { .iov_base = a.buf, .iov_len = a.buflen } }
2315 * FUSE will copy both struct a and the pointed buffer from the
2316 * process doing the ioctl and retry ioctl with both struct a and the
2319 * This time, FUSE server has everything it needs and completes ioctl
2320 * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2322 * Copying data out works the same way.
2324 * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2325 * automatically initializes in and out iovs by decoding @cmd with
2326 * _IOC_* macros and the server is not allowed to request RETRY. This
2327 * limits ioctl data transfers to well-formed ioctls and is the forced
2328 * behavior for all FUSE servers.
2330 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2333 struct fuse_file *ff = file->private_data;
2334 struct fuse_conn *fc = ff->fc;
2335 struct fuse_ioctl_in inarg = {
2341 struct fuse_ioctl_out outarg;
2342 struct fuse_req *req = NULL;
2343 struct page **pages = NULL;
2344 struct iovec *iov_page = NULL;
2345 struct iovec *in_iov = NULL, *out_iov = NULL;
2346 unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
2347 size_t in_size, out_size, transferred;
2350 #if BITS_PER_LONG == 32
2351 inarg.flags |= FUSE_IOCTL_32BIT;
2353 if (flags & FUSE_IOCTL_COMPAT)
2354 inarg.flags |= FUSE_IOCTL_32BIT;
2357 /* assume all the iovs returned by client always fits in a page */
2358 BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2361 pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
2362 iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2363 if (!pages || !iov_page)
2367 * If restricted, initialize IO parameters as encoded in @cmd.
2368 * RETRY from server is not allowed.
2370 if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2371 struct iovec *iov = iov_page;
2373 iov->iov_base = (void __user *)arg;
2374 iov->iov_len = _IOC_SIZE(cmd);
2376 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2381 if (_IOC_DIR(cmd) & _IOC_READ) {
2388 inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2389 inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2392 * Out data can be used either for actual out data or iovs,
2393 * make sure there always is at least one page.
2395 out_size = max_t(size_t, out_size, PAGE_SIZE);
2396 max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2398 /* make sure there are enough buffer pages and init request with them */
2400 if (max_pages > FUSE_MAX_PAGES_PER_REQ)
2402 while (num_pages < max_pages) {
2403 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2404 if (!pages[num_pages])
2409 req = fuse_get_req(fc, num_pages);
2415 memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
2416 req->num_pages = num_pages;
2417 fuse_page_descs_length_init(req, 0, req->num_pages);
2419 /* okay, let's send it to the client */
2420 req->in.h.opcode = FUSE_IOCTL;
2421 req->in.h.nodeid = ff->nodeid;
2422 req->in.numargs = 1;
2423 req->in.args[0].size = sizeof(inarg);
2424 req->in.args[0].value = &inarg;
2427 req->in.args[1].size = in_size;
2428 req->in.argpages = 1;
2430 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
2436 req->out.numargs = 2;
2437 req->out.args[0].size = sizeof(outarg);
2438 req->out.args[0].value = &outarg;
2439 req->out.args[1].size = out_size;
2440 req->out.argpages = 1;
2441 req->out.argvar = 1;
2443 fuse_request_send(fc, req);
2444 err = req->out.h.error;
2445 transferred = req->out.args[1].size;
2446 fuse_put_request(fc, req);
2451 /* did it ask for retry? */
2452 if (outarg.flags & FUSE_IOCTL_RETRY) {
2455 /* no retry if in restricted mode */
2457 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2460 in_iovs = outarg.in_iovs;
2461 out_iovs = outarg.out_iovs;
2464 * Make sure things are in boundary, separate checks
2465 * are to protect against overflow.
2468 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2469 out_iovs > FUSE_IOCTL_MAX_IOV ||
2470 in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2473 vaddr = kmap_atomic(pages[0]);
2474 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2475 transferred, in_iovs + out_iovs,
2476 (flags & FUSE_IOCTL_COMPAT) != 0);
2477 kunmap_atomic(vaddr);
2482 out_iov = in_iov + in_iovs;
2484 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
2488 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
2496 if (transferred > inarg.out_size)
2499 err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
2502 fuse_put_request(fc, req);
2503 free_page((unsigned long) iov_page);
2505 __free_page(pages[--num_pages]);
2508 return err ? err : outarg.result;
2510 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2512 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2513 unsigned long arg, unsigned int flags)
2515 struct inode *inode = file_inode(file);
2516 struct fuse_conn *fc = get_fuse_conn(inode);
2518 if (!fuse_allow_current_process(fc))
2521 if (is_bad_inode(inode))
2524 return fuse_do_ioctl(file, cmd, arg, flags);
2527 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2530 return fuse_ioctl_common(file, cmd, arg, 0);
2533 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2536 return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2540 * All files which have been polled are linked to RB tree
2541 * fuse_conn->polled_files which is indexed by kh. Walk the tree and
2542 * find the matching one.
2544 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2545 struct rb_node **parent_out)
2547 struct rb_node **link = &fc->polled_files.rb_node;
2548 struct rb_node *last = NULL;
2551 struct fuse_file *ff;
2554 ff = rb_entry(last, struct fuse_file, polled_node);
2557 link = &last->rb_left;
2558 else if (kh > ff->kh)
2559 link = &last->rb_right;
2570 * The file is about to be polled. Make sure it's on the polled_files
2571 * RB tree. Note that files once added to the polled_files tree are
2572 * not removed before the file is released. This is because a file
2573 * polled once is likely to be polled again.
2575 static void fuse_register_polled_file(struct fuse_conn *fc,
2576 struct fuse_file *ff)
2578 spin_lock(&fc->lock);
2579 if (RB_EMPTY_NODE(&ff->polled_node)) {
2580 struct rb_node **link, *parent;
2582 link = fuse_find_polled_node(fc, ff->kh, &parent);
2584 rb_link_node(&ff->polled_node, parent, link);
2585 rb_insert_color(&ff->polled_node, &fc->polled_files);
2587 spin_unlock(&fc->lock);
2590 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2592 struct fuse_file *ff = file->private_data;
2593 struct fuse_conn *fc = ff->fc;
2594 struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2595 struct fuse_poll_out outarg;
2596 struct fuse_req *req;
2600 return DEFAULT_POLLMASK;
2602 poll_wait(file, &ff->poll_wait, wait);
2603 inarg.events = (__u32)poll_requested_events(wait);
2606 * Ask for notification iff there's someone waiting for it.
2607 * The client may ignore the flag and always notify.
2609 if (waitqueue_active(&ff->poll_wait)) {
2610 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2611 fuse_register_polled_file(fc, ff);
2614 req = fuse_get_req_nopages(fc);
2618 req->in.h.opcode = FUSE_POLL;
2619 req->in.h.nodeid = ff->nodeid;
2620 req->in.numargs = 1;
2621 req->in.args[0].size = sizeof(inarg);
2622 req->in.args[0].value = &inarg;
2623 req->out.numargs = 1;
2624 req->out.args[0].size = sizeof(outarg);
2625 req->out.args[0].value = &outarg;
2626 fuse_request_send(fc, req);
2627 err = req->out.h.error;
2628 fuse_put_request(fc, req);
2631 return outarg.revents;
2632 if (err == -ENOSYS) {
2634 return DEFAULT_POLLMASK;
2638 EXPORT_SYMBOL_GPL(fuse_file_poll);
2641 * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2642 * wakes up the poll waiters.
2644 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2645 struct fuse_notify_poll_wakeup_out *outarg)
2647 u64 kh = outarg->kh;
2648 struct rb_node **link;
2650 spin_lock(&fc->lock);
2652 link = fuse_find_polled_node(fc, kh, NULL);
2654 struct fuse_file *ff;
2656 ff = rb_entry(*link, struct fuse_file, polled_node);
2657 wake_up_interruptible_sync(&ff->poll_wait);
2660 spin_unlock(&fc->lock);
2664 static void fuse_do_truncate(struct file *file)
2666 struct inode *inode = file->f_mapping->host;
2669 attr.ia_valid = ATTR_SIZE;
2670 attr.ia_size = i_size_read(inode);
2672 attr.ia_file = file;
2673 attr.ia_valid |= ATTR_FILE;
2675 fuse_do_setattr(inode, &attr, file);
2678 static inline loff_t fuse_round_up(loff_t off)
2680 return round_up(off, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
2684 fuse_direct_IO(int rw, struct kiocb *iocb, struct iov_iter *ii,
2688 struct file *file = iocb->ki_filp;
2689 struct fuse_file *ff = file->private_data;
2690 bool async_dio = ff->fc->async_dio;
2692 struct inode *inode;
2694 size_t count = iov_iter_count(ii);
2695 struct fuse_io_priv *io;
2698 inode = file->f_mapping->host;
2699 i_size = i_size_read(inode);
2701 /* optimization for short read */
2702 if (async_dio && rw != WRITE && offset + count > i_size) {
2703 if (offset >= i_size)
2705 count = min_t(loff_t, count, fuse_round_up(i_size - offset));
2708 io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2711 spin_lock_init(&io->lock);
2715 io->offset = offset;
2716 io->write = (rw == WRITE);
2720 * By default, we want to optimize all I/Os with async request
2721 * submission to the client filesystem if supported.
2723 io->async = async_dio;
2727 * We cannot asynchronously extend the size of a file. We have no method
2728 * to wait on real async I/O requests, so we must submit this request
2731 if (!is_sync_kiocb(iocb) && (offset + count > i_size) && rw == WRITE)
2735 ret = __fuse_direct_write(io, ii, &pos);
2737 ret = __fuse_direct_read(io, ii, &pos, count);
2740 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2742 /* we have a non-extending, async request, so return */
2743 if (!is_sync_kiocb(iocb))
2744 return -EIOCBQUEUED;
2746 ret = wait_on_sync_kiocb(iocb);
2753 fuse_write_update_size(inode, pos);
2754 else if (ret < 0 && offset + count > i_size)
2755 fuse_do_truncate(file);
2761 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2764 struct fuse_file *ff = file->private_data;
2765 struct inode *inode = file->f_inode;
2766 struct fuse_inode *fi = get_fuse_inode(inode);
2767 struct fuse_conn *fc = ff->fc;
2768 struct fuse_req *req;
2769 struct fuse_fallocate_in inarg = {
2776 bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
2777 (mode & FALLOC_FL_PUNCH_HOLE);
2779 if (fc->no_fallocate)
2783 mutex_lock(&inode->i_mutex);
2784 if (mode & FALLOC_FL_PUNCH_HOLE) {
2785 loff_t endbyte = offset + length - 1;
2786 err = filemap_write_and_wait_range(inode->i_mapping,
2791 fuse_sync_writes(inode);
2795 if (!(mode & FALLOC_FL_KEEP_SIZE))
2796 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2798 req = fuse_get_req_nopages(fc);
2804 req->in.h.opcode = FUSE_FALLOCATE;
2805 req->in.h.nodeid = ff->nodeid;
2806 req->in.numargs = 1;
2807 req->in.args[0].size = sizeof(inarg);
2808 req->in.args[0].value = &inarg;
2809 fuse_request_send(fc, req);
2810 err = req->out.h.error;
2811 if (err == -ENOSYS) {
2812 fc->no_fallocate = 1;
2815 fuse_put_request(fc, req);
2820 /* we could have extended the file */
2821 if (!(mode & FALLOC_FL_KEEP_SIZE))
2822 fuse_write_update_size(inode, offset + length);
2824 if (mode & FALLOC_FL_PUNCH_HOLE)
2825 truncate_pagecache_range(inode, offset, offset + length - 1);
2827 fuse_invalidate_attr(inode);
2830 if (!(mode & FALLOC_FL_KEEP_SIZE))
2831 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2834 mutex_unlock(&inode->i_mutex);
2839 static const struct file_operations fuse_file_operations = {
2840 .llseek = fuse_file_llseek,
2841 .read = do_sync_read,
2842 .aio_read = fuse_file_aio_read,
2843 .write = do_sync_write,
2844 .aio_write = fuse_file_aio_write,
2845 .mmap = fuse_file_mmap,
2847 .flush = fuse_flush,
2848 .release = fuse_release,
2849 .fsync = fuse_fsync,
2850 .lock = fuse_file_lock,
2851 .flock = fuse_file_flock,
2852 .splice_read = generic_file_splice_read,
2853 .unlocked_ioctl = fuse_file_ioctl,
2854 .compat_ioctl = fuse_file_compat_ioctl,
2855 .poll = fuse_file_poll,
2856 .fallocate = fuse_file_fallocate,
2859 static const struct file_operations fuse_direct_io_file_operations = {
2860 .llseek = fuse_file_llseek,
2861 .read = fuse_direct_read,
2862 .write = fuse_direct_write,
2863 .mmap = fuse_direct_mmap,
2865 .flush = fuse_flush,
2866 .release = fuse_release,
2867 .fsync = fuse_fsync,
2868 .lock = fuse_file_lock,
2869 .flock = fuse_file_flock,
2870 .unlocked_ioctl = fuse_file_ioctl,
2871 .compat_ioctl = fuse_file_compat_ioctl,
2872 .poll = fuse_file_poll,
2873 .fallocate = fuse_file_fallocate,
2874 /* no splice_read */
2877 static const struct address_space_operations fuse_file_aops = {
2878 .readpage = fuse_readpage,
2879 .writepage = fuse_writepage,
2880 .writepages = fuse_writepages,
2881 .launder_page = fuse_launder_page,
2882 .readpages = fuse_readpages,
2883 .set_page_dirty = __set_page_dirty_nobuffers,
2885 .direct_IO = fuse_direct_IO,
2888 void fuse_init_file_inode(struct inode *inode)
2890 inode->i_fop = &fuse_file_operations;
2891 inode->i_data.a_ops = &fuse_file_aops;