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) {
342 spin_unlock(&fc->lock);
348 * Wait for page writeback to be completed.
350 * Since fuse doesn't rely on the VM writeback tracking, this has to
351 * use some other means.
353 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
355 struct fuse_inode *fi = get_fuse_inode(inode);
357 wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
361 static int fuse_flush(struct file *file, fl_owner_t id)
363 struct inode *inode = file_inode(file);
364 struct fuse_conn *fc = get_fuse_conn(inode);
365 struct fuse_file *ff = file->private_data;
366 struct fuse_req *req;
367 struct fuse_flush_in inarg;
370 if (is_bad_inode(inode))
376 req = fuse_get_req_nofail_nopages(fc, file);
377 memset(&inarg, 0, sizeof(inarg));
379 inarg.lock_owner = fuse_lock_owner_id(fc, id);
380 req->in.h.opcode = FUSE_FLUSH;
381 req->in.h.nodeid = get_node_id(inode);
383 req->in.args[0].size = sizeof(inarg);
384 req->in.args[0].value = &inarg;
386 fuse_request_send(fc, req);
387 err = req->out.h.error;
388 fuse_put_request(fc, req);
389 if (err == -ENOSYS) {
397 * Wait for all pending writepages on the inode to finish.
399 * This is currently done by blocking further writes with FUSE_NOWRITE
400 * and waiting for all sent writes to complete.
402 * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
403 * could conflict with truncation.
405 static void fuse_sync_writes(struct inode *inode)
407 fuse_set_nowrite(inode);
408 fuse_release_nowrite(inode);
411 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
412 int datasync, int isdir)
414 struct inode *inode = file->f_mapping->host;
415 struct fuse_conn *fc = get_fuse_conn(inode);
416 struct fuse_file *ff = file->private_data;
417 struct fuse_req *req;
418 struct fuse_fsync_in inarg;
421 if (is_bad_inode(inode))
424 err = filemap_write_and_wait_range(inode->i_mapping, start, end);
428 if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
431 mutex_lock(&inode->i_mutex);
434 * Start writeback against all dirty pages of the inode, then
435 * wait for all outstanding writes, before sending the FSYNC
438 err = write_inode_now(inode, 0);
442 fuse_sync_writes(inode);
444 req = fuse_get_req_nopages(fc);
450 memset(&inarg, 0, sizeof(inarg));
452 inarg.fsync_flags = datasync ? 1 : 0;
453 req->in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
454 req->in.h.nodeid = get_node_id(inode);
456 req->in.args[0].size = sizeof(inarg);
457 req->in.args[0].value = &inarg;
458 fuse_request_send(fc, req);
459 err = req->out.h.error;
460 fuse_put_request(fc, req);
461 if (err == -ENOSYS) {
469 mutex_unlock(&inode->i_mutex);
473 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
476 return fuse_fsync_common(file, start, end, datasync, 0);
479 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
480 size_t count, int opcode)
482 struct fuse_read_in *inarg = &req->misc.read.in;
483 struct fuse_file *ff = file->private_data;
488 inarg->flags = file->f_flags;
489 req->in.h.opcode = opcode;
490 req->in.h.nodeid = ff->nodeid;
492 req->in.args[0].size = sizeof(struct fuse_read_in);
493 req->in.args[0].value = inarg;
495 req->out.numargs = 1;
496 req->out.args[0].size = count;
499 static void fuse_release_user_pages(struct fuse_req *req, int write)
503 for (i = 0; i < req->num_pages; i++) {
504 struct page *page = req->pages[i];
506 set_page_dirty_lock(page);
512 * In case of short read, the caller sets 'pos' to the position of
513 * actual end of fuse request in IO request. Otherwise, if bytes_requested
514 * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
517 * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
518 * both submitted asynchronously. The first of them was ACKed by userspace as
519 * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
520 * second request was ACKed as short, e.g. only 1K was read, resulting in
523 * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
524 * will be equal to the length of the longest contiguous fragment of
525 * transferred data starting from the beginning of IO request.
527 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
531 spin_lock(&io->lock);
533 io->err = io->err ? : err;
534 else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
538 spin_unlock(&io->lock);
545 else if (io->bytes >= 0 && io->write)
548 res = io->bytes < 0 ? io->size : io->bytes;
550 if (!is_sync_kiocb(io->iocb)) {
551 struct path *path = &io->iocb->ki_filp->f_path;
552 struct inode *inode = path->dentry->d_inode;
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 fi->attr_version = ++fc->attr_version;
635 i_size_write(inode, size);
637 spin_unlock(&fc->lock);
640 static int fuse_readpage(struct file *file, struct page *page)
642 struct fuse_io_priv io = { .async = 0, .file = file };
643 struct inode *inode = page->mapping->host;
644 struct fuse_conn *fc = get_fuse_conn(inode);
645 struct fuse_req *req;
647 loff_t pos = page_offset(page);
648 size_t count = PAGE_CACHE_SIZE;
653 if (is_bad_inode(inode))
657 * Page writeback can extend beyond the lifetime of the
658 * page-cache page, so make sure we read a properly synced
661 fuse_wait_on_page_writeback(inode, page->index);
663 req = fuse_get_req(fc, 1);
668 attr_ver = fuse_get_attr_version(fc);
670 req->out.page_zeroing = 1;
671 req->out.argpages = 1;
673 req->pages[0] = page;
674 req->page_descs[0].length = count;
675 num_read = fuse_send_read(req, &io, pos, count, NULL);
676 err = req->out.h.error;
677 fuse_put_request(fc, req);
681 * Short read means EOF. If file size is larger, truncate it
683 if (num_read < count)
684 fuse_read_update_size(inode, pos + num_read, attr_ver);
686 SetPageUptodate(page);
689 fuse_invalidate_attr(inode); /* atime changed */
695 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
698 size_t count = req->misc.read.in.size;
699 size_t num_read = req->out.args[0].size;
700 struct address_space *mapping = NULL;
702 for (i = 0; mapping == NULL && i < req->num_pages; i++)
703 mapping = req->pages[i]->mapping;
706 struct inode *inode = mapping->host;
709 * Short read means EOF. If file size is larger, truncate it
711 if (!req->out.h.error && num_read < count) {
714 pos = page_offset(req->pages[0]) + num_read;
715 fuse_read_update_size(inode, pos,
716 req->misc.read.attr_ver);
718 fuse_invalidate_attr(inode); /* atime changed */
721 for (i = 0; i < req->num_pages; i++) {
722 struct page *page = req->pages[i];
723 if (!req->out.h.error)
724 SetPageUptodate(page);
728 page_cache_release(page);
731 fuse_file_put(req->ff, false);
734 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
736 struct fuse_file *ff = file->private_data;
737 struct fuse_conn *fc = ff->fc;
738 loff_t pos = page_offset(req->pages[0]);
739 size_t count = req->num_pages << PAGE_CACHE_SHIFT;
741 req->out.argpages = 1;
742 req->out.page_zeroing = 1;
743 req->out.page_replace = 1;
744 fuse_read_fill(req, file, pos, count, FUSE_READ);
745 req->misc.read.attr_ver = fuse_get_attr_version(fc);
746 if (fc->async_read) {
747 req->ff = fuse_file_get(ff);
748 req->end = fuse_readpages_end;
749 fuse_request_send_background(fc, req);
751 fuse_request_send(fc, req);
752 fuse_readpages_end(fc, req);
753 fuse_put_request(fc, req);
757 struct fuse_fill_data {
758 struct fuse_req *req;
764 static int fuse_readpages_fill(void *_data, struct page *page)
766 struct fuse_fill_data *data = _data;
767 struct fuse_req *req = data->req;
768 struct inode *inode = data->inode;
769 struct fuse_conn *fc = get_fuse_conn(inode);
771 fuse_wait_on_page_writeback(inode, page->index);
773 if (req->num_pages &&
774 (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
775 (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
776 req->pages[req->num_pages - 1]->index + 1 != page->index)) {
777 int nr_alloc = min_t(unsigned, data->nr_pages,
778 FUSE_MAX_PAGES_PER_REQ);
779 fuse_send_readpages(req, data->file);
781 req = fuse_get_req_for_background(fc, nr_alloc);
783 req = fuse_get_req(fc, nr_alloc);
792 if (WARN_ON(req->num_pages >= req->max_pages)) {
793 fuse_put_request(fc, req);
797 page_cache_get(page);
798 req->pages[req->num_pages] = page;
799 req->page_descs[req->num_pages].length = PAGE_SIZE;
805 static int fuse_readpages(struct file *file, struct address_space *mapping,
806 struct list_head *pages, unsigned nr_pages)
808 struct inode *inode = mapping->host;
809 struct fuse_conn *fc = get_fuse_conn(inode);
810 struct fuse_fill_data data;
812 int nr_alloc = min_t(unsigned, nr_pages, FUSE_MAX_PAGES_PER_REQ);
815 if (is_bad_inode(inode))
821 data.req = fuse_get_req_for_background(fc, nr_alloc);
823 data.req = fuse_get_req(fc, nr_alloc);
824 data.nr_pages = nr_pages;
825 err = PTR_ERR(data.req);
826 if (IS_ERR(data.req))
829 err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
831 if (data.req->num_pages)
832 fuse_send_readpages(data.req, file);
834 fuse_put_request(fc, data.req);
840 static ssize_t fuse_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
841 unsigned long nr_segs, loff_t pos)
843 struct inode *inode = iocb->ki_filp->f_mapping->host;
844 struct fuse_conn *fc = get_fuse_conn(inode);
847 * In auto invalidate mode, always update attributes on read.
848 * Otherwise, only update if we attempt to read past EOF (to ensure
849 * i_size is up to date).
851 if (fc->auto_inval_data ||
852 (pos + iov_length(iov, nr_segs) > i_size_read(inode))) {
854 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
859 return generic_file_aio_read(iocb, iov, nr_segs, pos);
862 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
863 loff_t pos, size_t count)
865 struct fuse_write_in *inarg = &req->misc.write.in;
866 struct fuse_write_out *outarg = &req->misc.write.out;
871 req->in.h.opcode = FUSE_WRITE;
872 req->in.h.nodeid = ff->nodeid;
874 if (ff->fc->minor < 9)
875 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
877 req->in.args[0].size = sizeof(struct fuse_write_in);
878 req->in.args[0].value = inarg;
879 req->in.args[1].size = count;
880 req->out.numargs = 1;
881 req->out.args[0].size = sizeof(struct fuse_write_out);
882 req->out.args[0].value = outarg;
885 static size_t fuse_send_write(struct fuse_req *req, struct fuse_io_priv *io,
886 loff_t pos, size_t count, fl_owner_t owner)
888 struct file *file = io->file;
889 struct fuse_file *ff = file->private_data;
890 struct fuse_conn *fc = ff->fc;
891 struct fuse_write_in *inarg = &req->misc.write.in;
893 fuse_write_fill(req, ff, pos, count);
894 inarg->flags = file->f_flags;
896 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
897 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
901 return fuse_async_req_send(fc, req, count, io);
903 fuse_request_send(fc, req);
904 return req->misc.write.out.size;
907 void fuse_write_update_size(struct inode *inode, loff_t pos)
909 struct fuse_conn *fc = get_fuse_conn(inode);
910 struct fuse_inode *fi = get_fuse_inode(inode);
912 spin_lock(&fc->lock);
913 fi->attr_version = ++fc->attr_version;
914 if (pos > inode->i_size)
915 i_size_write(inode, pos);
916 spin_unlock(&fc->lock);
919 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
920 struct inode *inode, loff_t pos,
926 struct fuse_io_priv io = { .async = 0, .file = file };
928 for (i = 0; i < req->num_pages; i++)
929 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
931 res = fuse_send_write(req, &io, pos, count, NULL);
933 offset = req->page_descs[0].offset;
935 for (i = 0; i < req->num_pages; i++) {
936 struct page *page = req->pages[i];
938 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
939 SetPageUptodate(page);
941 if (count > PAGE_CACHE_SIZE - offset)
942 count -= PAGE_CACHE_SIZE - offset;
948 page_cache_release(page);
954 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
955 struct address_space *mapping,
956 struct iov_iter *ii, loff_t pos)
958 struct fuse_conn *fc = get_fuse_conn(mapping->host);
959 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
963 req->in.argpages = 1;
964 req->page_descs[0].offset = offset;
969 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
970 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
973 bytes = min_t(size_t, bytes, fc->max_write - count);
977 if (iov_iter_fault_in_readable(ii, bytes))
981 page = grab_cache_page_write_begin(mapping, index, 0);
985 if (mapping_writably_mapped(mapping))
986 flush_dcache_page(page);
989 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
991 flush_dcache_page(page);
993 mark_page_accessed(page);
997 page_cache_release(page);
998 bytes = min(bytes, iov_iter_single_seg_count(ii));
1003 req->pages[req->num_pages] = page;
1004 req->page_descs[req->num_pages].length = tmp;
1007 iov_iter_advance(ii, tmp);
1011 if (offset == PAGE_CACHE_SIZE)
1014 if (!fc->big_writes)
1016 } while (iov_iter_count(ii) && count < fc->max_write &&
1017 req->num_pages < req->max_pages && offset == 0);
1019 return count > 0 ? count : err;
1022 static inline unsigned fuse_wr_pages(loff_t pos, size_t len)
1024 return min_t(unsigned,
1025 ((pos + len - 1) >> PAGE_CACHE_SHIFT) -
1026 (pos >> PAGE_CACHE_SHIFT) + 1,
1027 FUSE_MAX_PAGES_PER_REQ);
1030 static ssize_t fuse_perform_write(struct file *file,
1031 struct address_space *mapping,
1032 struct iov_iter *ii, loff_t pos)
1034 struct inode *inode = mapping->host;
1035 struct fuse_conn *fc = get_fuse_conn(inode);
1039 if (is_bad_inode(inode))
1043 struct fuse_req *req;
1045 unsigned nr_pages = fuse_wr_pages(pos, iov_iter_count(ii));
1047 req = fuse_get_req(fc, nr_pages);
1053 count = fuse_fill_write_pages(req, mapping, ii, pos);
1059 num_written = fuse_send_write_pages(req, file, inode,
1061 err = req->out.h.error;
1066 /* break out of the loop on short write */
1067 if (num_written != count)
1071 fuse_put_request(fc, req);
1072 } while (!err && iov_iter_count(ii));
1075 fuse_write_update_size(inode, pos);
1077 fuse_invalidate_attr(inode);
1079 return res > 0 ? res : err;
1082 static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
1083 unsigned long nr_segs, loff_t pos)
1085 struct file *file = iocb->ki_filp;
1086 struct address_space *mapping = file->f_mapping;
1089 ssize_t written = 0;
1090 ssize_t written_buffered = 0;
1091 struct inode *inode = mapping->host;
1096 WARN_ON(iocb->ki_pos != pos);
1099 err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
1104 mutex_lock(&inode->i_mutex);
1106 /* We can write back this queue in page reclaim */
1107 current->backing_dev_info = mapping->backing_dev_info;
1109 err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
1116 err = file_remove_suid(file);
1120 err = file_update_time(file);
1124 if (file->f_flags & O_DIRECT) {
1125 written = generic_file_direct_write(iocb, iov, &nr_segs,
1128 if (written < 0 || written == count)
1134 iov_iter_init(&i, iov, nr_segs, count, written);
1135 written_buffered = fuse_perform_write(file, mapping, &i, pos);
1136 if (written_buffered < 0) {
1137 err = written_buffered;
1140 endbyte = pos + written_buffered - 1;
1142 err = filemap_write_and_wait_range(file->f_mapping, pos,
1147 invalidate_mapping_pages(file->f_mapping,
1148 pos >> PAGE_CACHE_SHIFT,
1149 endbyte >> PAGE_CACHE_SHIFT);
1151 written += written_buffered;
1152 iocb->ki_pos = pos + written_buffered;
1154 iov_iter_init(&i, iov, nr_segs, count, 0);
1155 written = fuse_perform_write(file, mapping, &i, pos);
1157 iocb->ki_pos = pos + written;
1160 current->backing_dev_info = NULL;
1161 mutex_unlock(&inode->i_mutex);
1163 return written ? written : err;
1166 static inline void fuse_page_descs_length_init(struct fuse_req *req,
1167 unsigned index, unsigned nr_pages)
1171 for (i = index; i < index + nr_pages; i++)
1172 req->page_descs[i].length = PAGE_SIZE -
1173 req->page_descs[i].offset;
1176 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1178 return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1181 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1184 return min(iov_iter_single_seg_count(ii), max_size);
1187 static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1188 size_t *nbytesp, int write)
1190 size_t nbytes = 0; /* # bytes already packed in req */
1192 /* Special case for kernel I/O: can copy directly into the buffer */
1193 if (segment_eq(get_fs(), KERNEL_DS)) {
1194 unsigned long user_addr = fuse_get_user_addr(ii);
1195 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1198 req->in.args[1].value = (void *) user_addr;
1200 req->out.args[0].value = (void *) user_addr;
1202 iov_iter_advance(ii, frag_size);
1203 *nbytesp = frag_size;
1207 while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1209 unsigned long user_addr = fuse_get_user_addr(ii);
1210 unsigned offset = user_addr & ~PAGE_MASK;
1211 size_t frag_size = fuse_get_frag_size(ii, *nbytesp - nbytes);
1214 unsigned n = req->max_pages - req->num_pages;
1215 frag_size = min_t(size_t, frag_size, n << PAGE_SHIFT);
1217 npages = (frag_size + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1218 npages = clamp(npages, 1U, n);
1220 ret = get_user_pages_fast(user_addr, npages, !write,
1221 &req->pages[req->num_pages]);
1226 frag_size = min_t(size_t, frag_size,
1227 (npages << PAGE_SHIFT) - offset);
1228 iov_iter_advance(ii, frag_size);
1230 req->page_descs[req->num_pages].offset = offset;
1231 fuse_page_descs_length_init(req, req->num_pages, npages);
1233 req->num_pages += npages;
1234 req->page_descs[req->num_pages - 1].length -=
1235 (npages << PAGE_SHIFT) - offset - frag_size;
1237 nbytes += frag_size;
1241 req->in.argpages = 1;
1243 req->out.argpages = 1;
1250 static inline int fuse_iter_npages(const struct iov_iter *ii_p)
1252 struct iov_iter ii = *ii_p;
1255 while (iov_iter_count(&ii) && npages < FUSE_MAX_PAGES_PER_REQ) {
1256 unsigned long user_addr = fuse_get_user_addr(&ii);
1257 unsigned offset = user_addr & ~PAGE_MASK;
1258 size_t frag_size = iov_iter_single_seg_count(&ii);
1260 npages += (frag_size + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1261 iov_iter_advance(&ii, frag_size);
1264 return min(npages, FUSE_MAX_PAGES_PER_REQ);
1267 ssize_t fuse_direct_io(struct fuse_io_priv *io, const struct iovec *iov,
1268 unsigned long nr_segs, size_t count, loff_t *ppos,
1271 struct file *file = io->file;
1272 struct fuse_file *ff = file->private_data;
1273 struct fuse_conn *fc = ff->fc;
1274 size_t nmax = write ? fc->max_write : fc->max_read;
1277 struct fuse_req *req;
1280 iov_iter_init(&ii, iov, nr_segs, count, 0);
1283 req = fuse_get_req_for_background(fc, fuse_iter_npages(&ii));
1285 req = fuse_get_req(fc, fuse_iter_npages(&ii));
1287 return PTR_ERR(req);
1291 fl_owner_t owner = current->files;
1292 size_t nbytes = min(count, nmax);
1293 int err = fuse_get_user_pages(req, &ii, &nbytes, write);
1300 nres = fuse_send_write(req, io, pos, nbytes, owner);
1302 nres = fuse_send_read(req, io, pos, nbytes, owner);
1305 fuse_release_user_pages(req, !write);
1306 if (req->out.h.error) {
1308 res = req->out.h.error;
1310 } else if (nres > nbytes) {
1320 fuse_put_request(fc, req);
1322 req = fuse_get_req_for_background(fc,
1323 fuse_iter_npages(&ii));
1325 req = fuse_get_req(fc, fuse_iter_npages(&ii));
1331 fuse_put_request(fc, req);
1337 EXPORT_SYMBOL_GPL(fuse_direct_io);
1339 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1340 const struct iovec *iov,
1341 unsigned long nr_segs, loff_t *ppos,
1345 struct file *file = io->file;
1346 struct inode *inode = file_inode(file);
1348 if (is_bad_inode(inode))
1351 res = fuse_direct_io(io, iov, nr_segs, count, ppos, 0);
1353 fuse_invalidate_attr(inode);
1358 static ssize_t fuse_direct_read(struct file *file, char __user *buf,
1359 size_t count, loff_t *ppos)
1361 struct fuse_io_priv io = { .async = 0, .file = file };
1362 struct iovec iov = { .iov_base = buf, .iov_len = count };
1363 return __fuse_direct_read(&io, &iov, 1, ppos, count);
1366 static ssize_t __fuse_direct_write(struct fuse_io_priv *io,
1367 const struct iovec *iov,
1368 unsigned long nr_segs, loff_t *ppos)
1370 struct file *file = io->file;
1371 struct inode *inode = file_inode(file);
1372 size_t count = iov_length(iov, nr_segs);
1375 res = generic_write_checks(file, ppos, &count, 0);
1377 res = fuse_direct_io(io, iov, nr_segs, count, ppos, 1);
1379 fuse_invalidate_attr(inode);
1384 static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
1385 size_t count, loff_t *ppos)
1387 struct iovec iov = { .iov_base = (void __user *)buf, .iov_len = count };
1388 struct inode *inode = file_inode(file);
1390 struct fuse_io_priv io = { .async = 0, .file = file };
1392 if (is_bad_inode(inode))
1395 /* Don't allow parallel writes to the same file */
1396 mutex_lock(&inode->i_mutex);
1397 res = __fuse_direct_write(&io, &iov, 1, ppos);
1399 fuse_write_update_size(inode, *ppos);
1400 mutex_unlock(&inode->i_mutex);
1405 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1407 __free_page(req->pages[0]);
1408 fuse_file_put(req->ff, false);
1411 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1413 struct inode *inode = req->inode;
1414 struct fuse_inode *fi = get_fuse_inode(inode);
1415 struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
1417 list_del(&req->writepages_entry);
1418 dec_bdi_stat(bdi, BDI_WRITEBACK);
1419 dec_zone_page_state(req->pages[0], NR_WRITEBACK_TEMP);
1420 bdi_writeout_inc(bdi);
1421 wake_up(&fi->page_waitq);
1424 /* Called under fc->lock, may release and reacquire it */
1425 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req)
1426 __releases(fc->lock)
1427 __acquires(fc->lock)
1429 struct fuse_inode *fi = get_fuse_inode(req->inode);
1430 loff_t size = i_size_read(req->inode);
1431 struct fuse_write_in *inarg = &req->misc.write.in;
1436 if (inarg->offset + PAGE_CACHE_SIZE <= size) {
1437 inarg->size = PAGE_CACHE_SIZE;
1438 } else if (inarg->offset < size) {
1439 inarg->size = size & (PAGE_CACHE_SIZE - 1);
1441 /* Got truncated off completely */
1445 req->in.args[1].size = inarg->size;
1447 fuse_request_send_background_locked(fc, req);
1451 fuse_writepage_finish(fc, req);
1452 spin_unlock(&fc->lock);
1453 fuse_writepage_free(fc, req);
1454 fuse_put_request(fc, req);
1455 spin_lock(&fc->lock);
1459 * If fi->writectr is positive (no truncate or fsync going on) send
1460 * all queued writepage requests.
1462 * Called with fc->lock
1464 void fuse_flush_writepages(struct inode *inode)
1465 __releases(fc->lock)
1466 __acquires(fc->lock)
1468 struct fuse_conn *fc = get_fuse_conn(inode);
1469 struct fuse_inode *fi = get_fuse_inode(inode);
1470 struct fuse_req *req;
1472 while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1473 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1474 list_del_init(&req->list);
1475 fuse_send_writepage(fc, req);
1479 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1481 struct inode *inode = req->inode;
1482 struct fuse_inode *fi = get_fuse_inode(inode);
1484 mapping_set_error(inode->i_mapping, req->out.h.error);
1485 spin_lock(&fc->lock);
1487 fuse_writepage_finish(fc, req);
1488 spin_unlock(&fc->lock);
1489 fuse_writepage_free(fc, req);
1492 static int fuse_writepage_locked(struct page *page)
1494 struct address_space *mapping = page->mapping;
1495 struct inode *inode = mapping->host;
1496 struct fuse_conn *fc = get_fuse_conn(inode);
1497 struct fuse_inode *fi = get_fuse_inode(inode);
1498 struct fuse_req *req;
1499 struct fuse_file *ff;
1500 struct page *tmp_page;
1502 set_page_writeback(page);
1504 req = fuse_request_alloc_nofs(1);
1508 req->background = 1; /* writeback always goes to bg_queue */
1509 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1513 spin_lock(&fc->lock);
1514 BUG_ON(list_empty(&fi->write_files));
1515 ff = list_entry(fi->write_files.next, struct fuse_file, write_entry);
1516 req->ff = fuse_file_get(ff);
1517 spin_unlock(&fc->lock);
1519 fuse_write_fill(req, ff, page_offset(page), 0);
1521 copy_highpage(tmp_page, page);
1522 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1523 req->in.argpages = 1;
1525 req->pages[0] = tmp_page;
1526 req->page_descs[0].offset = 0;
1527 req->page_descs[0].length = PAGE_SIZE;
1528 req->end = fuse_writepage_end;
1531 inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
1532 inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1533 end_page_writeback(page);
1535 spin_lock(&fc->lock);
1536 list_add(&req->writepages_entry, &fi->writepages);
1537 list_add_tail(&req->list, &fi->queued_writes);
1538 fuse_flush_writepages(inode);
1539 spin_unlock(&fc->lock);
1544 fuse_request_free(req);
1546 end_page_writeback(page);
1550 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1554 err = fuse_writepage_locked(page);
1560 static int fuse_launder_page(struct page *page)
1563 if (clear_page_dirty_for_io(page)) {
1564 struct inode *inode = page->mapping->host;
1565 err = fuse_writepage_locked(page);
1567 fuse_wait_on_page_writeback(inode, page->index);
1573 * Write back dirty pages now, because there may not be any suitable
1576 static void fuse_vma_close(struct vm_area_struct *vma)
1578 filemap_write_and_wait(vma->vm_file->f_mapping);
1582 * Wait for writeback against this page to complete before allowing it
1583 * to be marked dirty again, and hence written back again, possibly
1584 * before the previous writepage completed.
1586 * Block here, instead of in ->writepage(), so that the userspace fs
1587 * can only block processes actually operating on the filesystem.
1589 * Otherwise unprivileged userspace fs would be able to block
1594 * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
1596 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1598 struct page *page = vmf->page;
1600 * Don't use page->mapping as it may become NULL from a
1601 * concurrent truncate.
1603 struct inode *inode = vma->vm_file->f_mapping->host;
1605 fuse_wait_on_page_writeback(inode, page->index);
1609 static const struct vm_operations_struct fuse_file_vm_ops = {
1610 .close = fuse_vma_close,
1611 .fault = filemap_fault,
1612 .page_mkwrite = fuse_page_mkwrite,
1613 .remap_pages = generic_file_remap_pages,
1616 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
1618 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE)) {
1619 struct inode *inode = file_inode(file);
1620 struct fuse_conn *fc = get_fuse_conn(inode);
1621 struct fuse_inode *fi = get_fuse_inode(inode);
1622 struct fuse_file *ff = file->private_data;
1624 * file may be written through mmap, so chain it onto the
1625 * inodes's write_file list
1627 spin_lock(&fc->lock);
1628 if (list_empty(&ff->write_entry))
1629 list_add(&ff->write_entry, &fi->write_files);
1630 spin_unlock(&fc->lock);
1632 file_accessed(file);
1633 vma->vm_ops = &fuse_file_vm_ops;
1637 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
1639 /* Can't provide the coherency needed for MAP_SHARED */
1640 if (vma->vm_flags & VM_MAYSHARE)
1643 invalidate_inode_pages2(file->f_mapping);
1645 return generic_file_mmap(file, vma);
1648 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
1649 struct file_lock *fl)
1651 switch (ffl->type) {
1657 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
1658 ffl->end < ffl->start)
1661 fl->fl_start = ffl->start;
1662 fl->fl_end = ffl->end;
1663 fl->fl_pid = ffl->pid;
1669 fl->fl_type = ffl->type;
1673 static void fuse_lk_fill(struct fuse_req *req, struct file *file,
1674 const struct file_lock *fl, int opcode, pid_t pid,
1677 struct inode *inode = file_inode(file);
1678 struct fuse_conn *fc = get_fuse_conn(inode);
1679 struct fuse_file *ff = file->private_data;
1680 struct fuse_lk_in *arg = &req->misc.lk_in;
1683 arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
1684 arg->lk.start = fl->fl_start;
1685 arg->lk.end = fl->fl_end;
1686 arg->lk.type = fl->fl_type;
1689 arg->lk_flags |= FUSE_LK_FLOCK;
1690 req->in.h.opcode = opcode;
1691 req->in.h.nodeid = get_node_id(inode);
1692 req->in.numargs = 1;
1693 req->in.args[0].size = sizeof(*arg);
1694 req->in.args[0].value = arg;
1697 static int fuse_getlk(struct file *file, struct file_lock *fl)
1699 struct inode *inode = file_inode(file);
1700 struct fuse_conn *fc = get_fuse_conn(inode);
1701 struct fuse_req *req;
1702 struct fuse_lk_out outarg;
1705 req = fuse_get_req_nopages(fc);
1707 return PTR_ERR(req);
1709 fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
1710 req->out.numargs = 1;
1711 req->out.args[0].size = sizeof(outarg);
1712 req->out.args[0].value = &outarg;
1713 fuse_request_send(fc, req);
1714 err = req->out.h.error;
1715 fuse_put_request(fc, req);
1717 err = convert_fuse_file_lock(&outarg.lk, fl);
1722 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
1724 struct inode *inode = file_inode(file);
1725 struct fuse_conn *fc = get_fuse_conn(inode);
1726 struct fuse_req *req;
1727 int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
1728 pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
1731 if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
1732 /* NLM needs asynchronous locks, which we don't support yet */
1736 /* Unlock on close is handled by the flush method */
1737 if (fl->fl_flags & FL_CLOSE)
1740 req = fuse_get_req_nopages(fc);
1742 return PTR_ERR(req);
1744 fuse_lk_fill(req, file, fl, opcode, pid, flock);
1745 fuse_request_send(fc, req);
1746 err = req->out.h.error;
1747 /* locking is restartable */
1750 fuse_put_request(fc, req);
1754 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
1756 struct inode *inode = file_inode(file);
1757 struct fuse_conn *fc = get_fuse_conn(inode);
1760 if (cmd == F_CANCELLK) {
1762 } else if (cmd == F_GETLK) {
1764 posix_test_lock(file, fl);
1767 err = fuse_getlk(file, fl);
1770 err = posix_lock_file(file, fl, NULL);
1772 err = fuse_setlk(file, fl, 0);
1777 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
1779 struct inode *inode = file_inode(file);
1780 struct fuse_conn *fc = get_fuse_conn(inode);
1784 err = flock_lock_file_wait(file, fl);
1786 struct fuse_file *ff = file->private_data;
1788 /* emulate flock with POSIX locks */
1789 fl->fl_owner = (fl_owner_t) file;
1791 err = fuse_setlk(file, fl, 1);
1797 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
1799 struct inode *inode = mapping->host;
1800 struct fuse_conn *fc = get_fuse_conn(inode);
1801 struct fuse_req *req;
1802 struct fuse_bmap_in inarg;
1803 struct fuse_bmap_out outarg;
1806 if (!inode->i_sb->s_bdev || fc->no_bmap)
1809 req = fuse_get_req_nopages(fc);
1813 memset(&inarg, 0, sizeof(inarg));
1814 inarg.block = block;
1815 inarg.blocksize = inode->i_sb->s_blocksize;
1816 req->in.h.opcode = FUSE_BMAP;
1817 req->in.h.nodeid = get_node_id(inode);
1818 req->in.numargs = 1;
1819 req->in.args[0].size = sizeof(inarg);
1820 req->in.args[0].value = &inarg;
1821 req->out.numargs = 1;
1822 req->out.args[0].size = sizeof(outarg);
1823 req->out.args[0].value = &outarg;
1824 fuse_request_send(fc, req);
1825 err = req->out.h.error;
1826 fuse_put_request(fc, req);
1830 return err ? 0 : outarg.block;
1833 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
1836 struct inode *inode = file_inode(file);
1838 /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
1839 if (whence == SEEK_CUR || whence == SEEK_SET)
1840 return generic_file_llseek(file, offset, whence);
1842 mutex_lock(&inode->i_mutex);
1843 retval = fuse_update_attributes(inode, NULL, file, NULL);
1845 retval = generic_file_llseek(file, offset, whence);
1846 mutex_unlock(&inode->i_mutex);
1851 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
1852 unsigned int nr_segs, size_t bytes, bool to_user)
1860 iov_iter_init(&ii, iov, nr_segs, bytes, 0);
1862 while (iov_iter_count(&ii)) {
1863 struct page *page = pages[page_idx++];
1864 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
1870 char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
1871 size_t iov_len = ii.iov->iov_len - ii.iov_offset;
1872 size_t copy = min(todo, iov_len);
1876 left = copy_from_user(kaddr, uaddr, copy);
1878 left = copy_to_user(uaddr, kaddr, copy);
1883 iov_iter_advance(&ii, copy);
1895 * CUSE servers compiled on 32bit broke on 64bit kernels because the
1896 * ABI was defined to be 'struct iovec' which is different on 32bit
1897 * and 64bit. Fortunately we can determine which structure the server
1898 * used from the size of the reply.
1900 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
1901 size_t transferred, unsigned count,
1904 #ifdef CONFIG_COMPAT
1905 if (count * sizeof(struct compat_iovec) == transferred) {
1906 struct compat_iovec *ciov = src;
1910 * With this interface a 32bit server cannot support
1911 * non-compat (i.e. ones coming from 64bit apps) ioctl
1917 for (i = 0; i < count; i++) {
1918 dst[i].iov_base = compat_ptr(ciov[i].iov_base);
1919 dst[i].iov_len = ciov[i].iov_len;
1925 if (count * sizeof(struct iovec) != transferred)
1928 memcpy(dst, src, transferred);
1932 /* Make sure iov_length() won't overflow */
1933 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
1936 u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
1938 for (n = 0; n < count; n++, iov++) {
1939 if (iov->iov_len > (size_t) max)
1941 max -= iov->iov_len;
1946 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
1947 void *src, size_t transferred, unsigned count,
1951 struct fuse_ioctl_iovec *fiov = src;
1953 if (fc->minor < 16) {
1954 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
1958 if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
1961 for (i = 0; i < count; i++) {
1962 /* Did the server supply an inappropriate value? */
1963 if (fiov[i].base != (unsigned long) fiov[i].base ||
1964 fiov[i].len != (unsigned long) fiov[i].len)
1967 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
1968 dst[i].iov_len = (size_t) fiov[i].len;
1970 #ifdef CONFIG_COMPAT
1972 (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
1973 (compat_size_t) dst[i].iov_len != fiov[i].len))
1983 * For ioctls, there is no generic way to determine how much memory
1984 * needs to be read and/or written. Furthermore, ioctls are allowed
1985 * to dereference the passed pointer, so the parameter requires deep
1986 * copying but FUSE has no idea whatsoever about what to copy in or
1989 * This is solved by allowing FUSE server to retry ioctl with
1990 * necessary in/out iovecs. Let's assume the ioctl implementation
1991 * needs to read in the following structure.
1998 * On the first callout to FUSE server, inarg->in_size and
1999 * inarg->out_size will be NULL; then, the server completes the ioctl
2000 * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2001 * the actual iov array to
2003 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) } }
2005 * which tells FUSE to copy in the requested area and retry the ioctl.
2006 * On the second round, the server has access to the structure and
2007 * from that it can tell what to look for next, so on the invocation,
2008 * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2010 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) },
2011 * { .iov_base = a.buf, .iov_len = a.buflen } }
2013 * FUSE will copy both struct a and the pointed buffer from the
2014 * process doing the ioctl and retry ioctl with both struct a and the
2017 * This time, FUSE server has everything it needs and completes ioctl
2018 * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2020 * Copying data out works the same way.
2022 * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2023 * automatically initializes in and out iovs by decoding @cmd with
2024 * _IOC_* macros and the server is not allowed to request RETRY. This
2025 * limits ioctl data transfers to well-formed ioctls and is the forced
2026 * behavior for all FUSE servers.
2028 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2031 struct fuse_file *ff = file->private_data;
2032 struct fuse_conn *fc = ff->fc;
2033 struct fuse_ioctl_in inarg = {
2039 struct fuse_ioctl_out outarg;
2040 struct fuse_req *req = NULL;
2041 struct page **pages = NULL;
2042 struct iovec *iov_page = NULL;
2043 struct iovec *in_iov = NULL, *out_iov = NULL;
2044 unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
2045 size_t in_size, out_size, transferred;
2048 #if BITS_PER_LONG == 32
2049 inarg.flags |= FUSE_IOCTL_32BIT;
2051 if (flags & FUSE_IOCTL_COMPAT)
2052 inarg.flags |= FUSE_IOCTL_32BIT;
2055 /* assume all the iovs returned by client always fits in a page */
2056 BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2059 pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
2060 iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2061 if (!pages || !iov_page)
2065 * If restricted, initialize IO parameters as encoded in @cmd.
2066 * RETRY from server is not allowed.
2068 if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2069 struct iovec *iov = iov_page;
2071 iov->iov_base = (void __user *)arg;
2072 iov->iov_len = _IOC_SIZE(cmd);
2074 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2079 if (_IOC_DIR(cmd) & _IOC_READ) {
2086 inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2087 inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2090 * Out data can be used either for actual out data or iovs,
2091 * make sure there always is at least one page.
2093 out_size = max_t(size_t, out_size, PAGE_SIZE);
2094 max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2096 /* make sure there are enough buffer pages and init request with them */
2098 if (max_pages > FUSE_MAX_PAGES_PER_REQ)
2100 while (num_pages < max_pages) {
2101 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2102 if (!pages[num_pages])
2107 req = fuse_get_req(fc, num_pages);
2113 memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
2114 req->num_pages = num_pages;
2115 fuse_page_descs_length_init(req, 0, req->num_pages);
2117 /* okay, let's send it to the client */
2118 req->in.h.opcode = FUSE_IOCTL;
2119 req->in.h.nodeid = ff->nodeid;
2120 req->in.numargs = 1;
2121 req->in.args[0].size = sizeof(inarg);
2122 req->in.args[0].value = &inarg;
2125 req->in.args[1].size = in_size;
2126 req->in.argpages = 1;
2128 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
2134 req->out.numargs = 2;
2135 req->out.args[0].size = sizeof(outarg);
2136 req->out.args[0].value = &outarg;
2137 req->out.args[1].size = out_size;
2138 req->out.argpages = 1;
2139 req->out.argvar = 1;
2141 fuse_request_send(fc, req);
2142 err = req->out.h.error;
2143 transferred = req->out.args[1].size;
2144 fuse_put_request(fc, req);
2149 /* did it ask for retry? */
2150 if (outarg.flags & FUSE_IOCTL_RETRY) {
2153 /* no retry if in restricted mode */
2155 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2158 in_iovs = outarg.in_iovs;
2159 out_iovs = outarg.out_iovs;
2162 * Make sure things are in boundary, separate checks
2163 * are to protect against overflow.
2166 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2167 out_iovs > FUSE_IOCTL_MAX_IOV ||
2168 in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2171 vaddr = kmap_atomic(pages[0]);
2172 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2173 transferred, in_iovs + out_iovs,
2174 (flags & FUSE_IOCTL_COMPAT) != 0);
2175 kunmap_atomic(vaddr);
2180 out_iov = in_iov + in_iovs;
2182 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
2186 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
2194 if (transferred > inarg.out_size)
2197 err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
2200 fuse_put_request(fc, req);
2201 free_page((unsigned long) iov_page);
2203 __free_page(pages[--num_pages]);
2206 return err ? err : outarg.result;
2208 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2210 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2211 unsigned long arg, unsigned int flags)
2213 struct inode *inode = file_inode(file);
2214 struct fuse_conn *fc = get_fuse_conn(inode);
2216 if (!fuse_allow_current_process(fc))
2219 if (is_bad_inode(inode))
2222 return fuse_do_ioctl(file, cmd, arg, flags);
2225 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2228 return fuse_ioctl_common(file, cmd, arg, 0);
2231 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2234 return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2238 * All files which have been polled are linked to RB tree
2239 * fuse_conn->polled_files which is indexed by kh. Walk the tree and
2240 * find the matching one.
2242 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2243 struct rb_node **parent_out)
2245 struct rb_node **link = &fc->polled_files.rb_node;
2246 struct rb_node *last = NULL;
2249 struct fuse_file *ff;
2252 ff = rb_entry(last, struct fuse_file, polled_node);
2255 link = &last->rb_left;
2256 else if (kh > ff->kh)
2257 link = &last->rb_right;
2268 * The file is about to be polled. Make sure it's on the polled_files
2269 * RB tree. Note that files once added to the polled_files tree are
2270 * not removed before the file is released. This is because a file
2271 * polled once is likely to be polled again.
2273 static void fuse_register_polled_file(struct fuse_conn *fc,
2274 struct fuse_file *ff)
2276 spin_lock(&fc->lock);
2277 if (RB_EMPTY_NODE(&ff->polled_node)) {
2278 struct rb_node **link, *parent;
2280 link = fuse_find_polled_node(fc, ff->kh, &parent);
2282 rb_link_node(&ff->polled_node, parent, link);
2283 rb_insert_color(&ff->polled_node, &fc->polled_files);
2285 spin_unlock(&fc->lock);
2288 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2290 struct fuse_file *ff = file->private_data;
2291 struct fuse_conn *fc = ff->fc;
2292 struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2293 struct fuse_poll_out outarg;
2294 struct fuse_req *req;
2298 return DEFAULT_POLLMASK;
2300 poll_wait(file, &ff->poll_wait, wait);
2301 inarg.events = (__u32)poll_requested_events(wait);
2304 * Ask for notification iff there's someone waiting for it.
2305 * The client may ignore the flag and always notify.
2307 if (waitqueue_active(&ff->poll_wait)) {
2308 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2309 fuse_register_polled_file(fc, ff);
2312 req = fuse_get_req_nopages(fc);
2316 req->in.h.opcode = FUSE_POLL;
2317 req->in.h.nodeid = ff->nodeid;
2318 req->in.numargs = 1;
2319 req->in.args[0].size = sizeof(inarg);
2320 req->in.args[0].value = &inarg;
2321 req->out.numargs = 1;
2322 req->out.args[0].size = sizeof(outarg);
2323 req->out.args[0].value = &outarg;
2324 fuse_request_send(fc, req);
2325 err = req->out.h.error;
2326 fuse_put_request(fc, req);
2329 return outarg.revents;
2330 if (err == -ENOSYS) {
2332 return DEFAULT_POLLMASK;
2336 EXPORT_SYMBOL_GPL(fuse_file_poll);
2339 * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2340 * wakes up the poll waiters.
2342 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2343 struct fuse_notify_poll_wakeup_out *outarg)
2345 u64 kh = outarg->kh;
2346 struct rb_node **link;
2348 spin_lock(&fc->lock);
2350 link = fuse_find_polled_node(fc, kh, NULL);
2352 struct fuse_file *ff;
2354 ff = rb_entry(*link, struct fuse_file, polled_node);
2355 wake_up_interruptible_sync(&ff->poll_wait);
2358 spin_unlock(&fc->lock);
2362 static void fuse_do_truncate(struct file *file)
2364 struct inode *inode = file->f_mapping->host;
2367 attr.ia_valid = ATTR_SIZE;
2368 attr.ia_size = i_size_read(inode);
2370 attr.ia_file = file;
2371 attr.ia_valid |= ATTR_FILE;
2373 fuse_do_setattr(inode, &attr, file);
2376 static inline loff_t fuse_round_up(loff_t off)
2378 return round_up(off, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
2382 fuse_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2383 loff_t offset, unsigned long nr_segs)
2386 struct file *file = iocb->ki_filp;
2387 struct fuse_file *ff = file->private_data;
2388 bool async_dio = ff->fc->async_dio;
2390 struct inode *inode;
2392 size_t count = iov_length(iov, nr_segs);
2393 struct fuse_io_priv *io;
2396 inode = file->f_mapping->host;
2397 i_size = i_size_read(inode);
2399 /* optimization for short read */
2400 if (async_dio && rw != WRITE && offset + count > i_size) {
2401 if (offset >= i_size)
2403 count = min_t(loff_t, count, fuse_round_up(i_size - offset));
2406 io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2409 spin_lock_init(&io->lock);
2413 io->offset = offset;
2414 io->write = (rw == WRITE);
2418 * By default, we want to optimize all I/Os with async request
2419 * submission to the client filesystem if supported.
2421 io->async = async_dio;
2425 * We cannot asynchronously extend the size of a file. We have no method
2426 * to wait on real async I/O requests, so we must submit this request
2429 if (!is_sync_kiocb(iocb) && (offset + count > i_size) && rw == WRITE)
2433 ret = __fuse_direct_write(io, iov, nr_segs, &pos);
2435 ret = __fuse_direct_read(io, iov, nr_segs, &pos, count);
2438 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2440 /* we have a non-extending, async request, so return */
2441 if (!is_sync_kiocb(iocb))
2442 return -EIOCBQUEUED;
2444 ret = wait_on_sync_kiocb(iocb);
2451 fuse_write_update_size(inode, pos);
2452 else if (ret < 0 && offset + count > i_size)
2453 fuse_do_truncate(file);
2459 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2462 struct fuse_file *ff = file->private_data;
2463 struct inode *inode = file->f_inode;
2464 struct fuse_conn *fc = ff->fc;
2465 struct fuse_req *req;
2466 struct fuse_fallocate_in inarg = {
2474 if (fc->no_fallocate)
2477 if (mode & FALLOC_FL_PUNCH_HOLE) {
2478 mutex_lock(&inode->i_mutex);
2479 fuse_set_nowrite(inode);
2482 req = fuse_get_req_nopages(fc);
2488 req->in.h.opcode = FUSE_FALLOCATE;
2489 req->in.h.nodeid = ff->nodeid;
2490 req->in.numargs = 1;
2491 req->in.args[0].size = sizeof(inarg);
2492 req->in.args[0].value = &inarg;
2493 fuse_request_send(fc, req);
2494 err = req->out.h.error;
2495 if (err == -ENOSYS) {
2496 fc->no_fallocate = 1;
2499 fuse_put_request(fc, req);
2504 /* we could have extended the file */
2505 if (!(mode & FALLOC_FL_KEEP_SIZE))
2506 fuse_write_update_size(inode, offset + length);
2508 if (mode & FALLOC_FL_PUNCH_HOLE)
2509 truncate_pagecache_range(inode, offset, offset + length - 1);
2511 fuse_invalidate_attr(inode);
2514 if (mode & FALLOC_FL_PUNCH_HOLE) {
2515 fuse_release_nowrite(inode);
2516 mutex_unlock(&inode->i_mutex);
2522 static const struct file_operations fuse_file_operations = {
2523 .llseek = fuse_file_llseek,
2524 .read = do_sync_read,
2525 .aio_read = fuse_file_aio_read,
2526 .write = do_sync_write,
2527 .aio_write = fuse_file_aio_write,
2528 .mmap = fuse_file_mmap,
2530 .flush = fuse_flush,
2531 .release = fuse_release,
2532 .fsync = fuse_fsync,
2533 .lock = fuse_file_lock,
2534 .flock = fuse_file_flock,
2535 .splice_read = generic_file_splice_read,
2536 .unlocked_ioctl = fuse_file_ioctl,
2537 .compat_ioctl = fuse_file_compat_ioctl,
2538 .poll = fuse_file_poll,
2539 .fallocate = fuse_file_fallocate,
2542 static const struct file_operations fuse_direct_io_file_operations = {
2543 .llseek = fuse_file_llseek,
2544 .read = fuse_direct_read,
2545 .write = fuse_direct_write,
2546 .mmap = fuse_direct_mmap,
2548 .flush = fuse_flush,
2549 .release = fuse_release,
2550 .fsync = fuse_fsync,
2551 .lock = fuse_file_lock,
2552 .flock = fuse_file_flock,
2553 .unlocked_ioctl = fuse_file_ioctl,
2554 .compat_ioctl = fuse_file_compat_ioctl,
2555 .poll = fuse_file_poll,
2556 .fallocate = fuse_file_fallocate,
2557 /* no splice_read */
2560 static const struct address_space_operations fuse_file_aops = {
2561 .readpage = fuse_readpage,
2562 .writepage = fuse_writepage,
2563 .launder_page = fuse_launder_page,
2564 .readpages = fuse_readpages,
2565 .set_page_dirty = __set_page_dirty_nobuffers,
2567 .direct_IO = fuse_direct_IO,
2570 void fuse_init_file_inode(struct inode *inode)
2572 inode->i_fop = &fuse_file_operations;
2573 inode->i_data.a_ops = &fuse_file_aops;