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>
19 static const struct file_operations fuse_direct_io_file_operations;
21 static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
22 int opcode, struct fuse_open_out *outargp)
24 struct fuse_open_in inarg;
28 req = fuse_get_req_nopages(fc);
32 memset(&inarg, 0, sizeof(inarg));
33 inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
34 if (!fc->atomic_o_trunc)
35 inarg.flags &= ~O_TRUNC;
36 req->in.h.opcode = opcode;
37 req->in.h.nodeid = nodeid;
39 req->in.args[0].size = sizeof(inarg);
40 req->in.args[0].value = &inarg;
42 req->out.args[0].size = sizeof(*outargp);
43 req->out.args[0].value = outargp;
44 fuse_request_send(fc, req);
45 err = req->out.h.error;
46 fuse_put_request(fc, req);
51 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
55 ff = kmalloc(sizeof(struct fuse_file), GFP_KERNEL);
60 ff->reserved_req = fuse_request_alloc(0);
61 if (unlikely(!ff->reserved_req)) {
66 INIT_LIST_HEAD(&ff->write_entry);
67 atomic_set(&ff->count, 0);
68 RB_CLEAR_NODE(&ff->polled_node);
69 init_waitqueue_head(&ff->poll_wait);
73 spin_unlock(&fc->lock);
78 void fuse_file_free(struct fuse_file *ff)
80 fuse_request_free(ff->reserved_req);
84 struct fuse_file *fuse_file_get(struct fuse_file *ff)
86 atomic_inc(&ff->count);
90 static void fuse_release_async(struct work_struct *work)
96 req = container_of(work, struct fuse_req, misc.release.work);
97 path = req->misc.release.path;
98 fc = get_fuse_conn(path.dentry->d_inode);
100 fuse_put_request(fc, req);
104 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
106 if (fc->destroy_req) {
108 * If this is a fuseblk mount, then it's possible that
109 * releasing the path will result in releasing the
110 * super block and sending the DESTROY request. If
111 * the server is single threaded, this would hang.
112 * For this reason do the path_put() in a separate
115 atomic_inc(&req->count);
116 INIT_WORK(&req->misc.release.work, fuse_release_async);
117 schedule_work(&req->misc.release.work);
119 path_put(&req->misc.release.path);
123 static void fuse_file_put(struct fuse_file *ff, bool sync)
125 if (atomic_dec_and_test(&ff->count)) {
126 struct fuse_req *req = ff->reserved_req;
130 fuse_request_send(ff->fc, req);
131 path_put(&req->misc.release.path);
132 fuse_put_request(ff->fc, req);
134 req->end = fuse_release_end;
136 fuse_request_send_background(ff->fc, req);
142 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
145 struct fuse_open_out outarg;
146 struct fuse_file *ff;
148 int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
150 ff = fuse_file_alloc(fc);
154 err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
161 outarg.open_flags &= ~FOPEN_DIRECT_IO;
165 ff->open_flags = outarg.open_flags;
166 file->private_data = fuse_file_get(ff);
170 EXPORT_SYMBOL_GPL(fuse_do_open);
172 void fuse_finish_open(struct inode *inode, struct file *file)
174 struct fuse_file *ff = file->private_data;
175 struct fuse_conn *fc = get_fuse_conn(inode);
177 if (ff->open_flags & FOPEN_DIRECT_IO)
178 file->f_op = &fuse_direct_io_file_operations;
179 if (!(ff->open_flags & FOPEN_KEEP_CACHE))
180 invalidate_inode_pages2(inode->i_mapping);
181 if (ff->open_flags & FOPEN_NONSEEKABLE)
182 nonseekable_open(inode, file);
183 if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
184 struct fuse_inode *fi = get_fuse_inode(inode);
186 spin_lock(&fc->lock);
187 fi->attr_version = ++fc->attr_version;
188 i_size_write(inode, 0);
189 spin_unlock(&fc->lock);
190 fuse_invalidate_attr(inode);
194 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
196 struct fuse_conn *fc = get_fuse_conn(inode);
199 err = generic_file_open(inode, file);
203 err = fuse_do_open(fc, get_node_id(inode), file, isdir);
207 fuse_finish_open(inode, file);
212 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
214 struct fuse_conn *fc = ff->fc;
215 struct fuse_req *req = ff->reserved_req;
216 struct fuse_release_in *inarg = &req->misc.release.in;
218 spin_lock(&fc->lock);
219 list_del(&ff->write_entry);
220 if (!RB_EMPTY_NODE(&ff->polled_node))
221 rb_erase(&ff->polled_node, &fc->polled_files);
222 spin_unlock(&fc->lock);
224 wake_up_interruptible_all(&ff->poll_wait);
227 inarg->flags = flags;
228 req->in.h.opcode = opcode;
229 req->in.h.nodeid = ff->nodeid;
231 req->in.args[0].size = sizeof(struct fuse_release_in);
232 req->in.args[0].value = inarg;
235 void fuse_release_common(struct file *file, int opcode)
237 struct fuse_file *ff;
238 struct fuse_req *req;
240 ff = file->private_data;
244 req = ff->reserved_req;
245 fuse_prepare_release(ff, file->f_flags, opcode);
248 struct fuse_release_in *inarg = &req->misc.release.in;
249 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
250 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
253 /* Hold vfsmount and dentry until release is finished */
254 path_get(&file->f_path);
255 req->misc.release.path = file->f_path;
258 * Normally this will send the RELEASE request, however if
259 * some asynchronous READ or WRITE requests are outstanding,
260 * the sending will be delayed.
262 * Make the release synchronous if this is a fuseblk mount,
263 * synchronous RELEASE is allowed (and desirable) in this case
264 * because the server can be trusted not to screw up.
266 fuse_file_put(ff, ff->fc->destroy_req != NULL);
269 static int fuse_open(struct inode *inode, struct file *file)
271 return fuse_open_common(inode, file, false);
274 static int fuse_release(struct inode *inode, struct file *file)
276 fuse_release_common(file, FUSE_RELEASE);
278 /* return value is ignored by VFS */
282 void fuse_sync_release(struct fuse_file *ff, int flags)
284 WARN_ON(atomic_read(&ff->count) > 1);
285 fuse_prepare_release(ff, flags, FUSE_RELEASE);
286 ff->reserved_req->force = 1;
287 ff->reserved_req->background = 0;
288 fuse_request_send(ff->fc, ff->reserved_req);
289 fuse_put_request(ff->fc, ff->reserved_req);
292 EXPORT_SYMBOL_GPL(fuse_sync_release);
295 * Scramble the ID space with XTEA, so that the value of the files_struct
296 * pointer is not exposed to userspace.
298 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
300 u32 *k = fc->scramble_key;
301 u64 v = (unsigned long) id;
307 for (i = 0; i < 32; i++) {
308 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
310 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
313 return (u64) v0 + ((u64) v1 << 32);
317 * Check if page is under writeback
319 * This is currently done by walking the list of writepage requests
320 * for the inode, which can be pretty inefficient.
322 static bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
324 struct fuse_conn *fc = get_fuse_conn(inode);
325 struct fuse_inode *fi = get_fuse_inode(inode);
326 struct fuse_req *req;
329 spin_lock(&fc->lock);
330 list_for_each_entry(req, &fi->writepages, writepages_entry) {
333 BUG_ON(req->inode != inode);
334 curr_index = req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
335 if (curr_index == index) {
340 spin_unlock(&fc->lock);
346 * Wait for page writeback to be completed.
348 * Since fuse doesn't rely on the VM writeback tracking, this has to
349 * use some other means.
351 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
353 struct fuse_inode *fi = get_fuse_inode(inode);
355 wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
359 static int fuse_flush(struct file *file, fl_owner_t id)
361 struct inode *inode = file_inode(file);
362 struct fuse_conn *fc = get_fuse_conn(inode);
363 struct fuse_file *ff = file->private_data;
364 struct fuse_req *req;
365 struct fuse_flush_in inarg;
368 if (is_bad_inode(inode))
374 req = fuse_get_req_nofail_nopages(fc, file);
375 memset(&inarg, 0, sizeof(inarg));
377 inarg.lock_owner = fuse_lock_owner_id(fc, id);
378 req->in.h.opcode = FUSE_FLUSH;
379 req->in.h.nodeid = get_node_id(inode);
381 req->in.args[0].size = sizeof(inarg);
382 req->in.args[0].value = &inarg;
384 fuse_request_send(fc, req);
385 err = req->out.h.error;
386 fuse_put_request(fc, req);
387 if (err == -ENOSYS) {
395 * Wait for all pending writepages on the inode to finish.
397 * This is currently done by blocking further writes with FUSE_NOWRITE
398 * and waiting for all sent writes to complete.
400 * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
401 * could conflict with truncation.
403 static void fuse_sync_writes(struct inode *inode)
405 fuse_set_nowrite(inode);
406 fuse_release_nowrite(inode);
409 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
410 int datasync, int isdir)
412 struct inode *inode = file->f_mapping->host;
413 struct fuse_conn *fc = get_fuse_conn(inode);
414 struct fuse_file *ff = file->private_data;
415 struct fuse_req *req;
416 struct fuse_fsync_in inarg;
419 if (is_bad_inode(inode))
422 err = filemap_write_and_wait_range(inode->i_mapping, start, end);
426 if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
429 mutex_lock(&inode->i_mutex);
432 * Start writeback against all dirty pages of the inode, then
433 * wait for all outstanding writes, before sending the FSYNC
436 err = write_inode_now(inode, 0);
440 fuse_sync_writes(inode);
442 req = fuse_get_req_nopages(fc);
448 memset(&inarg, 0, sizeof(inarg));
450 inarg.fsync_flags = datasync ? 1 : 0;
451 req->in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
452 req->in.h.nodeid = get_node_id(inode);
454 req->in.args[0].size = sizeof(inarg);
455 req->in.args[0].value = &inarg;
456 fuse_request_send(fc, req);
457 err = req->out.h.error;
458 fuse_put_request(fc, req);
459 if (err == -ENOSYS) {
467 mutex_unlock(&inode->i_mutex);
471 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
474 return fuse_fsync_common(file, start, end, datasync, 0);
477 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
478 size_t count, int opcode)
480 struct fuse_read_in *inarg = &req->misc.read.in;
481 struct fuse_file *ff = file->private_data;
486 inarg->flags = file->f_flags;
487 req->in.h.opcode = opcode;
488 req->in.h.nodeid = ff->nodeid;
490 req->in.args[0].size = sizeof(struct fuse_read_in);
491 req->in.args[0].value = inarg;
493 req->out.numargs = 1;
494 req->out.args[0].size = count;
497 static void fuse_release_user_pages(struct fuse_req *req, int write)
501 for (i = 0; i < req->num_pages; i++) {
502 struct page *page = req->pages[i];
504 set_page_dirty_lock(page);
510 * In case of short read, the caller sets 'pos' to the position of
511 * actual end of fuse request in IO request. Otherwise, if bytes_requested
512 * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
515 * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
516 * both submitted asynchronously. The first of them was ACKed by userspace as
517 * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
518 * second request was ACKed as short, e.g. only 1K was read, resulting in
521 * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
522 * will be equal to the length of the longest contiguous fragment of
523 * transferred data starting from the beginning of IO request.
525 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
529 spin_lock(&io->lock);
531 io->err = io->err ? : err;
532 else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
536 spin_unlock(&io->lock);
543 else if (io->bytes >= 0 && io->write)
546 res = io->bytes < 0 ? io->size : io->bytes;
548 if (!is_sync_kiocb(io->iocb)) {
549 struct path *path = &io->iocb->ki_filp->f_path;
550 struct inode *inode = path->dentry->d_inode;
551 struct fuse_conn *fc = get_fuse_conn(inode);
552 struct fuse_inode *fi = get_fuse_inode(inode);
554 spin_lock(&fc->lock);
555 fi->attr_version = ++fc->attr_version;
556 spin_unlock(&fc->lock);
560 aio_complete(io->iocb, res, 0);
565 static void fuse_aio_complete_req(struct fuse_conn *fc, struct fuse_req *req)
567 struct fuse_io_priv *io = req->io;
570 fuse_release_user_pages(req, !io->write);
573 if (req->misc.write.in.size != req->misc.write.out.size)
574 pos = req->misc.write.in.offset - io->offset +
575 req->misc.write.out.size;
577 if (req->misc.read.in.size != req->out.args[0].size)
578 pos = req->misc.read.in.offset - io->offset +
579 req->out.args[0].size;
582 fuse_aio_complete(io, req->out.h.error, pos);
585 static size_t fuse_async_req_send(struct fuse_conn *fc, struct fuse_req *req,
586 size_t num_bytes, struct fuse_io_priv *io)
588 spin_lock(&io->lock);
589 io->size += num_bytes;
591 spin_unlock(&io->lock);
594 req->end = fuse_aio_complete_req;
596 __fuse_get_request(req);
597 fuse_request_send_background(fc, req);
602 static size_t fuse_send_read(struct fuse_req *req, struct fuse_io_priv *io,
603 loff_t pos, size_t count, fl_owner_t owner)
605 struct file *file = io->file;
606 struct fuse_file *ff = file->private_data;
607 struct fuse_conn *fc = ff->fc;
609 fuse_read_fill(req, file, pos, count, FUSE_READ);
611 struct fuse_read_in *inarg = &req->misc.read.in;
613 inarg->read_flags |= FUSE_READ_LOCKOWNER;
614 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
618 return fuse_async_req_send(fc, req, count, io);
620 fuse_request_send(fc, req);
621 return req->out.args[0].size;
624 static void fuse_read_update_size(struct inode *inode, loff_t size,
627 struct fuse_conn *fc = get_fuse_conn(inode);
628 struct fuse_inode *fi = get_fuse_inode(inode);
630 spin_lock(&fc->lock);
631 if (attr_ver == fi->attr_version && size < inode->i_size) {
632 fi->attr_version = ++fc->attr_version;
633 i_size_write(inode, size);
635 spin_unlock(&fc->lock);
638 static int fuse_readpage(struct file *file, struct page *page)
640 struct fuse_io_priv io = { .async = 0, .file = file };
641 struct inode *inode = page->mapping->host;
642 struct fuse_conn *fc = get_fuse_conn(inode);
643 struct fuse_req *req;
645 loff_t pos = page_offset(page);
646 size_t count = PAGE_CACHE_SIZE;
651 if (is_bad_inode(inode))
655 * Page writeback can extend beyond the lifetime of the
656 * page-cache page, so make sure we read a properly synced
659 fuse_wait_on_page_writeback(inode, page->index);
661 req = fuse_get_req(fc, 1);
666 attr_ver = fuse_get_attr_version(fc);
668 req->out.page_zeroing = 1;
669 req->out.argpages = 1;
671 req->pages[0] = page;
672 req->page_descs[0].length = count;
673 num_read = fuse_send_read(req, &io, pos, count, NULL);
674 err = req->out.h.error;
675 fuse_put_request(fc, req);
679 * Short read means EOF. If file size is larger, truncate it
681 if (num_read < count)
682 fuse_read_update_size(inode, pos + num_read, attr_ver);
684 SetPageUptodate(page);
687 fuse_invalidate_attr(inode); /* atime changed */
693 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
696 size_t count = req->misc.read.in.size;
697 size_t num_read = req->out.args[0].size;
698 struct address_space *mapping = NULL;
700 for (i = 0; mapping == NULL && i < req->num_pages; i++)
701 mapping = req->pages[i]->mapping;
704 struct inode *inode = mapping->host;
707 * Short read means EOF. If file size is larger, truncate it
709 if (!req->out.h.error && num_read < count) {
712 pos = page_offset(req->pages[0]) + num_read;
713 fuse_read_update_size(inode, pos,
714 req->misc.read.attr_ver);
716 fuse_invalidate_attr(inode); /* atime changed */
719 for (i = 0; i < req->num_pages; i++) {
720 struct page *page = req->pages[i];
721 if (!req->out.h.error)
722 SetPageUptodate(page);
726 page_cache_release(page);
729 fuse_file_put(req->ff, false);
732 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
734 struct fuse_file *ff = file->private_data;
735 struct fuse_conn *fc = ff->fc;
736 loff_t pos = page_offset(req->pages[0]);
737 size_t count = req->num_pages << PAGE_CACHE_SHIFT;
739 req->out.argpages = 1;
740 req->out.page_zeroing = 1;
741 req->out.page_replace = 1;
742 fuse_read_fill(req, file, pos, count, FUSE_READ);
743 req->misc.read.attr_ver = fuse_get_attr_version(fc);
744 if (fc->async_read) {
745 req->ff = fuse_file_get(ff);
746 req->end = fuse_readpages_end;
747 fuse_request_send_background(fc, req);
749 fuse_request_send(fc, req);
750 fuse_readpages_end(fc, req);
751 fuse_put_request(fc, req);
755 struct fuse_fill_data {
756 struct fuse_req *req;
762 static int fuse_readpages_fill(void *_data, struct page *page)
764 struct fuse_fill_data *data = _data;
765 struct fuse_req *req = data->req;
766 struct inode *inode = data->inode;
767 struct fuse_conn *fc = get_fuse_conn(inode);
769 fuse_wait_on_page_writeback(inode, page->index);
771 if (req->num_pages &&
772 (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
773 (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
774 req->pages[req->num_pages - 1]->index + 1 != page->index)) {
775 int nr_alloc = min_t(unsigned, data->nr_pages,
776 FUSE_MAX_PAGES_PER_REQ);
777 fuse_send_readpages(req, data->file);
779 req = fuse_get_req_for_background(fc, nr_alloc);
781 req = fuse_get_req(fc, nr_alloc);
790 if (WARN_ON(req->num_pages >= req->max_pages)) {
791 fuse_put_request(fc, req);
795 page_cache_get(page);
796 req->pages[req->num_pages] = page;
797 req->page_descs[req->num_pages].length = PAGE_SIZE;
803 static int fuse_readpages(struct file *file, struct address_space *mapping,
804 struct list_head *pages, unsigned nr_pages)
806 struct inode *inode = mapping->host;
807 struct fuse_conn *fc = get_fuse_conn(inode);
808 struct fuse_fill_data data;
810 int nr_alloc = min_t(unsigned, nr_pages, FUSE_MAX_PAGES_PER_REQ);
813 if (is_bad_inode(inode))
819 data.req = fuse_get_req_for_background(fc, nr_alloc);
821 data.req = fuse_get_req(fc, nr_alloc);
822 data.nr_pages = nr_pages;
823 err = PTR_ERR(data.req);
824 if (IS_ERR(data.req))
827 err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
829 if (data.req->num_pages)
830 fuse_send_readpages(data.req, file);
832 fuse_put_request(fc, data.req);
838 static ssize_t fuse_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
839 unsigned long nr_segs, loff_t pos)
841 struct inode *inode = iocb->ki_filp->f_mapping->host;
842 struct fuse_conn *fc = get_fuse_conn(inode);
845 * In auto invalidate mode, always update attributes on read.
846 * Otherwise, only update if we attempt to read past EOF (to ensure
847 * i_size is up to date).
849 if (fc->auto_inval_data ||
850 (pos + iov_length(iov, nr_segs) > i_size_read(inode))) {
852 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
857 return generic_file_aio_read(iocb, iov, nr_segs, pos);
860 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
861 loff_t pos, size_t count)
863 struct fuse_write_in *inarg = &req->misc.write.in;
864 struct fuse_write_out *outarg = &req->misc.write.out;
869 req->in.h.opcode = FUSE_WRITE;
870 req->in.h.nodeid = ff->nodeid;
872 if (ff->fc->minor < 9)
873 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
875 req->in.args[0].size = sizeof(struct fuse_write_in);
876 req->in.args[0].value = inarg;
877 req->in.args[1].size = count;
878 req->out.numargs = 1;
879 req->out.args[0].size = sizeof(struct fuse_write_out);
880 req->out.args[0].value = outarg;
883 static size_t fuse_send_write(struct fuse_req *req, struct fuse_io_priv *io,
884 loff_t pos, size_t count, fl_owner_t owner)
886 struct file *file = io->file;
887 struct fuse_file *ff = file->private_data;
888 struct fuse_conn *fc = ff->fc;
889 struct fuse_write_in *inarg = &req->misc.write.in;
891 fuse_write_fill(req, ff, pos, count);
892 inarg->flags = file->f_flags;
894 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
895 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
899 return fuse_async_req_send(fc, req, count, io);
901 fuse_request_send(fc, req);
902 return req->misc.write.out.size;
905 void fuse_write_update_size(struct inode *inode, loff_t pos)
907 struct fuse_conn *fc = get_fuse_conn(inode);
908 struct fuse_inode *fi = get_fuse_inode(inode);
910 spin_lock(&fc->lock);
911 fi->attr_version = ++fc->attr_version;
912 if (pos > inode->i_size)
913 i_size_write(inode, pos);
914 spin_unlock(&fc->lock);
917 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
918 struct inode *inode, loff_t pos,
924 struct fuse_io_priv io = { .async = 0, .file = file };
926 for (i = 0; i < req->num_pages; i++)
927 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
929 res = fuse_send_write(req, &io, pos, count, NULL);
931 offset = req->page_descs[0].offset;
933 for (i = 0; i < req->num_pages; i++) {
934 struct page *page = req->pages[i];
936 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
937 SetPageUptodate(page);
939 if (count > PAGE_CACHE_SIZE - offset)
940 count -= PAGE_CACHE_SIZE - offset;
946 page_cache_release(page);
952 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
953 struct address_space *mapping,
954 struct iov_iter *ii, loff_t pos)
956 struct fuse_conn *fc = get_fuse_conn(mapping->host);
957 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
961 req->in.argpages = 1;
962 req->page_descs[0].offset = offset;
967 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
968 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
971 bytes = min_t(size_t, bytes, fc->max_write - count);
975 if (iov_iter_fault_in_readable(ii, bytes))
979 page = grab_cache_page_write_begin(mapping, index, 0);
983 if (mapping_writably_mapped(mapping))
984 flush_dcache_page(page);
987 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
989 flush_dcache_page(page);
991 mark_page_accessed(page);
995 page_cache_release(page);
996 bytes = min(bytes, iov_iter_single_seg_count(ii));
1001 req->pages[req->num_pages] = page;
1002 req->page_descs[req->num_pages].length = tmp;
1005 iov_iter_advance(ii, tmp);
1009 if (offset == PAGE_CACHE_SIZE)
1012 if (!fc->big_writes)
1014 } while (iov_iter_count(ii) && count < fc->max_write &&
1015 req->num_pages < req->max_pages && offset == 0);
1017 return count > 0 ? count : err;
1020 static inline unsigned fuse_wr_pages(loff_t pos, size_t len)
1022 return min_t(unsigned,
1023 ((pos + len - 1) >> PAGE_CACHE_SHIFT) -
1024 (pos >> PAGE_CACHE_SHIFT) + 1,
1025 FUSE_MAX_PAGES_PER_REQ);
1028 static ssize_t fuse_perform_write(struct file *file,
1029 struct address_space *mapping,
1030 struct iov_iter *ii, loff_t pos)
1032 struct inode *inode = mapping->host;
1033 struct fuse_conn *fc = get_fuse_conn(inode);
1037 if (is_bad_inode(inode))
1041 struct fuse_req *req;
1043 unsigned nr_pages = fuse_wr_pages(pos, iov_iter_count(ii));
1045 req = fuse_get_req(fc, nr_pages);
1051 count = fuse_fill_write_pages(req, mapping, ii, pos);
1057 num_written = fuse_send_write_pages(req, file, inode,
1059 err = req->out.h.error;
1064 /* break out of the loop on short write */
1065 if (num_written != count)
1069 fuse_put_request(fc, req);
1070 } while (!err && iov_iter_count(ii));
1073 fuse_write_update_size(inode, pos);
1075 fuse_invalidate_attr(inode);
1077 return res > 0 ? res : err;
1080 static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
1081 unsigned long nr_segs, loff_t pos)
1083 struct file *file = iocb->ki_filp;
1084 struct address_space *mapping = file->f_mapping;
1087 ssize_t written = 0;
1088 ssize_t written_buffered = 0;
1089 struct inode *inode = mapping->host;
1094 WARN_ON(iocb->ki_pos != pos);
1097 err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
1102 mutex_lock(&inode->i_mutex);
1104 /* We can write back this queue in page reclaim */
1105 current->backing_dev_info = mapping->backing_dev_info;
1107 err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
1114 err = file_remove_suid(file);
1118 err = file_update_time(file);
1122 if (file->f_flags & O_DIRECT) {
1123 written = generic_file_direct_write(iocb, iov, &nr_segs,
1126 if (written < 0 || written == count)
1132 iov_iter_init(&i, iov, nr_segs, count, written);
1133 written_buffered = fuse_perform_write(file, mapping, &i, pos);
1134 if (written_buffered < 0) {
1135 err = written_buffered;
1138 endbyte = pos + written_buffered - 1;
1140 err = filemap_write_and_wait_range(file->f_mapping, pos,
1145 invalidate_mapping_pages(file->f_mapping,
1146 pos >> PAGE_CACHE_SHIFT,
1147 endbyte >> PAGE_CACHE_SHIFT);
1149 written += written_buffered;
1150 iocb->ki_pos = pos + written_buffered;
1152 iov_iter_init(&i, iov, nr_segs, count, 0);
1153 written = fuse_perform_write(file, mapping, &i, pos);
1155 iocb->ki_pos = pos + written;
1158 current->backing_dev_info = NULL;
1159 mutex_unlock(&inode->i_mutex);
1161 return written ? written : err;
1164 static inline void fuse_page_descs_length_init(struct fuse_req *req,
1165 unsigned index, unsigned nr_pages)
1169 for (i = index; i < index + nr_pages; i++)
1170 req->page_descs[i].length = PAGE_SIZE -
1171 req->page_descs[i].offset;
1174 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1176 return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1179 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1182 return min(iov_iter_single_seg_count(ii), max_size);
1185 static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1186 size_t *nbytesp, int write)
1188 size_t nbytes = 0; /* # bytes already packed in req */
1190 /* Special case for kernel I/O: can copy directly into the buffer */
1191 if (segment_eq(get_fs(), KERNEL_DS)) {
1192 unsigned long user_addr = fuse_get_user_addr(ii);
1193 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1196 req->in.args[1].value = (void *) user_addr;
1198 req->out.args[0].value = (void *) user_addr;
1200 iov_iter_advance(ii, frag_size);
1201 *nbytesp = frag_size;
1205 while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1207 unsigned long user_addr = fuse_get_user_addr(ii);
1208 unsigned offset = user_addr & ~PAGE_MASK;
1209 size_t frag_size = fuse_get_frag_size(ii, *nbytesp - nbytes);
1212 unsigned n = req->max_pages - req->num_pages;
1213 frag_size = min_t(size_t, frag_size, n << PAGE_SHIFT);
1215 npages = (frag_size + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1216 npages = clamp(npages, 1U, n);
1218 ret = get_user_pages_fast(user_addr, npages, !write,
1219 &req->pages[req->num_pages]);
1224 frag_size = min_t(size_t, frag_size,
1225 (npages << PAGE_SHIFT) - offset);
1226 iov_iter_advance(ii, frag_size);
1228 req->page_descs[req->num_pages].offset = offset;
1229 fuse_page_descs_length_init(req, req->num_pages, npages);
1231 req->num_pages += npages;
1232 req->page_descs[req->num_pages - 1].length -=
1233 (npages << PAGE_SHIFT) - offset - frag_size;
1235 nbytes += frag_size;
1239 req->in.argpages = 1;
1241 req->out.argpages = 1;
1248 static inline int fuse_iter_npages(const struct iov_iter *ii_p)
1250 struct iov_iter ii = *ii_p;
1253 while (iov_iter_count(&ii) && npages < FUSE_MAX_PAGES_PER_REQ) {
1254 unsigned long user_addr = fuse_get_user_addr(&ii);
1255 unsigned offset = user_addr & ~PAGE_MASK;
1256 size_t frag_size = iov_iter_single_seg_count(&ii);
1258 npages += (frag_size + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1259 iov_iter_advance(&ii, frag_size);
1262 return min(npages, FUSE_MAX_PAGES_PER_REQ);
1265 ssize_t fuse_direct_io(struct fuse_io_priv *io, const struct iovec *iov,
1266 unsigned long nr_segs, size_t count, loff_t *ppos,
1269 struct file *file = io->file;
1270 struct fuse_file *ff = file->private_data;
1271 struct fuse_conn *fc = ff->fc;
1272 size_t nmax = write ? fc->max_write : fc->max_read;
1275 struct fuse_req *req;
1278 iov_iter_init(&ii, iov, nr_segs, count, 0);
1280 req = fuse_get_req(fc, fuse_iter_npages(&ii));
1282 return PTR_ERR(req);
1286 fl_owner_t owner = current->files;
1287 size_t nbytes = min(count, nmax);
1288 int err = fuse_get_user_pages(req, &ii, &nbytes, write);
1295 nres = fuse_send_write(req, io, pos, nbytes, owner);
1297 nres = fuse_send_read(req, io, pos, nbytes, owner);
1300 fuse_release_user_pages(req, !write);
1301 if (req->out.h.error) {
1303 res = req->out.h.error;
1305 } else if (nres > nbytes) {
1315 fuse_put_request(fc, req);
1316 req = fuse_get_req(fc, fuse_iter_npages(&ii));
1322 fuse_put_request(fc, req);
1328 EXPORT_SYMBOL_GPL(fuse_direct_io);
1330 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1331 const struct iovec *iov,
1332 unsigned long nr_segs, loff_t *ppos,
1336 struct file *file = io->file;
1337 struct inode *inode = file_inode(file);
1339 if (is_bad_inode(inode))
1342 res = fuse_direct_io(io, iov, nr_segs, count, ppos, 0);
1344 fuse_invalidate_attr(inode);
1349 static ssize_t fuse_direct_read(struct file *file, char __user *buf,
1350 size_t count, loff_t *ppos)
1352 struct fuse_io_priv io = { .async = 0, .file = file };
1353 struct iovec iov = { .iov_base = buf, .iov_len = count };
1354 return __fuse_direct_read(&io, &iov, 1, ppos, count);
1357 static ssize_t __fuse_direct_write(struct fuse_io_priv *io,
1358 const struct iovec *iov,
1359 unsigned long nr_segs, loff_t *ppos)
1361 struct file *file = io->file;
1362 struct inode *inode = file_inode(file);
1363 size_t count = iov_length(iov, nr_segs);
1366 res = generic_write_checks(file, ppos, &count, 0);
1368 res = fuse_direct_io(io, iov, nr_segs, count, ppos, 1);
1370 fuse_invalidate_attr(inode);
1375 static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
1376 size_t count, loff_t *ppos)
1378 struct iovec iov = { .iov_base = (void __user *)buf, .iov_len = count };
1379 struct inode *inode = file_inode(file);
1381 struct fuse_io_priv io = { .async = 0, .file = file };
1383 if (is_bad_inode(inode))
1386 /* Don't allow parallel writes to the same file */
1387 mutex_lock(&inode->i_mutex);
1388 res = __fuse_direct_write(&io, &iov, 1, ppos);
1390 fuse_write_update_size(inode, *ppos);
1391 mutex_unlock(&inode->i_mutex);
1396 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1398 __free_page(req->pages[0]);
1399 fuse_file_put(req->ff, false);
1402 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1404 struct inode *inode = req->inode;
1405 struct fuse_inode *fi = get_fuse_inode(inode);
1406 struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
1408 list_del(&req->writepages_entry);
1409 dec_bdi_stat(bdi, BDI_WRITEBACK);
1410 dec_zone_page_state(req->pages[0], NR_WRITEBACK_TEMP);
1411 bdi_writeout_inc(bdi);
1412 wake_up(&fi->page_waitq);
1415 /* Called under fc->lock, may release and reacquire it */
1416 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req)
1417 __releases(fc->lock)
1418 __acquires(fc->lock)
1420 struct fuse_inode *fi = get_fuse_inode(req->inode);
1421 loff_t size = i_size_read(req->inode);
1422 struct fuse_write_in *inarg = &req->misc.write.in;
1427 if (inarg->offset + PAGE_CACHE_SIZE <= size) {
1428 inarg->size = PAGE_CACHE_SIZE;
1429 } else if (inarg->offset < size) {
1430 inarg->size = size & (PAGE_CACHE_SIZE - 1);
1432 /* Got truncated off completely */
1436 req->in.args[1].size = inarg->size;
1438 fuse_request_send_background_locked(fc, req);
1442 fuse_writepage_finish(fc, req);
1443 spin_unlock(&fc->lock);
1444 fuse_writepage_free(fc, req);
1445 fuse_put_request(fc, req);
1446 spin_lock(&fc->lock);
1450 * If fi->writectr is positive (no truncate or fsync going on) send
1451 * all queued writepage requests.
1453 * Called with fc->lock
1455 void fuse_flush_writepages(struct inode *inode)
1456 __releases(fc->lock)
1457 __acquires(fc->lock)
1459 struct fuse_conn *fc = get_fuse_conn(inode);
1460 struct fuse_inode *fi = get_fuse_inode(inode);
1461 struct fuse_req *req;
1463 while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1464 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1465 list_del_init(&req->list);
1466 fuse_send_writepage(fc, req);
1470 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1472 struct inode *inode = req->inode;
1473 struct fuse_inode *fi = get_fuse_inode(inode);
1475 mapping_set_error(inode->i_mapping, req->out.h.error);
1476 spin_lock(&fc->lock);
1478 fuse_writepage_finish(fc, req);
1479 spin_unlock(&fc->lock);
1480 fuse_writepage_free(fc, req);
1483 static int fuse_writepage_locked(struct page *page)
1485 struct address_space *mapping = page->mapping;
1486 struct inode *inode = mapping->host;
1487 struct fuse_conn *fc = get_fuse_conn(inode);
1488 struct fuse_inode *fi = get_fuse_inode(inode);
1489 struct fuse_req *req;
1490 struct fuse_file *ff;
1491 struct page *tmp_page;
1493 set_page_writeback(page);
1495 req = fuse_request_alloc_nofs(1);
1499 req->background = 1; /* writeback always goes to bg_queue */
1500 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1504 spin_lock(&fc->lock);
1505 BUG_ON(list_empty(&fi->write_files));
1506 ff = list_entry(fi->write_files.next, struct fuse_file, write_entry);
1507 req->ff = fuse_file_get(ff);
1508 spin_unlock(&fc->lock);
1510 fuse_write_fill(req, ff, page_offset(page), 0);
1512 copy_highpage(tmp_page, page);
1513 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1514 req->in.argpages = 1;
1516 req->pages[0] = tmp_page;
1517 req->page_descs[0].offset = 0;
1518 req->page_descs[0].length = PAGE_SIZE;
1519 req->end = fuse_writepage_end;
1522 inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
1523 inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1524 end_page_writeback(page);
1526 spin_lock(&fc->lock);
1527 list_add(&req->writepages_entry, &fi->writepages);
1528 list_add_tail(&req->list, &fi->queued_writes);
1529 fuse_flush_writepages(inode);
1530 spin_unlock(&fc->lock);
1535 fuse_request_free(req);
1537 end_page_writeback(page);
1541 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1545 err = fuse_writepage_locked(page);
1551 static int fuse_launder_page(struct page *page)
1554 if (clear_page_dirty_for_io(page)) {
1555 struct inode *inode = page->mapping->host;
1556 err = fuse_writepage_locked(page);
1558 fuse_wait_on_page_writeback(inode, page->index);
1564 * Write back dirty pages now, because there may not be any suitable
1567 static void fuse_vma_close(struct vm_area_struct *vma)
1569 filemap_write_and_wait(vma->vm_file->f_mapping);
1573 * Wait for writeback against this page to complete before allowing it
1574 * to be marked dirty again, and hence written back again, possibly
1575 * before the previous writepage completed.
1577 * Block here, instead of in ->writepage(), so that the userspace fs
1578 * can only block processes actually operating on the filesystem.
1580 * Otherwise unprivileged userspace fs would be able to block
1585 * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
1587 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1589 struct page *page = vmf->page;
1591 * Don't use page->mapping as it may become NULL from a
1592 * concurrent truncate.
1594 struct inode *inode = vma->vm_file->f_mapping->host;
1596 fuse_wait_on_page_writeback(inode, page->index);
1600 static const struct vm_operations_struct fuse_file_vm_ops = {
1601 .close = fuse_vma_close,
1602 .fault = filemap_fault,
1603 .page_mkwrite = fuse_page_mkwrite,
1604 .remap_pages = generic_file_remap_pages,
1607 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
1609 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE)) {
1610 struct inode *inode = file_inode(file);
1611 struct fuse_conn *fc = get_fuse_conn(inode);
1612 struct fuse_inode *fi = get_fuse_inode(inode);
1613 struct fuse_file *ff = file->private_data;
1615 * file may be written through mmap, so chain it onto the
1616 * inodes's write_file list
1618 spin_lock(&fc->lock);
1619 if (list_empty(&ff->write_entry))
1620 list_add(&ff->write_entry, &fi->write_files);
1621 spin_unlock(&fc->lock);
1623 file_accessed(file);
1624 vma->vm_ops = &fuse_file_vm_ops;
1628 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
1630 /* Can't provide the coherency needed for MAP_SHARED */
1631 if (vma->vm_flags & VM_MAYSHARE)
1634 invalidate_inode_pages2(file->f_mapping);
1636 return generic_file_mmap(file, vma);
1639 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
1640 struct file_lock *fl)
1642 switch (ffl->type) {
1648 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
1649 ffl->end < ffl->start)
1652 fl->fl_start = ffl->start;
1653 fl->fl_end = ffl->end;
1654 fl->fl_pid = ffl->pid;
1660 fl->fl_type = ffl->type;
1664 static void fuse_lk_fill(struct fuse_req *req, struct file *file,
1665 const struct file_lock *fl, int opcode, pid_t pid,
1668 struct inode *inode = file_inode(file);
1669 struct fuse_conn *fc = get_fuse_conn(inode);
1670 struct fuse_file *ff = file->private_data;
1671 struct fuse_lk_in *arg = &req->misc.lk_in;
1674 arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
1675 arg->lk.start = fl->fl_start;
1676 arg->lk.end = fl->fl_end;
1677 arg->lk.type = fl->fl_type;
1680 arg->lk_flags |= FUSE_LK_FLOCK;
1681 req->in.h.opcode = opcode;
1682 req->in.h.nodeid = get_node_id(inode);
1683 req->in.numargs = 1;
1684 req->in.args[0].size = sizeof(*arg);
1685 req->in.args[0].value = arg;
1688 static int fuse_getlk(struct file *file, struct file_lock *fl)
1690 struct inode *inode = file_inode(file);
1691 struct fuse_conn *fc = get_fuse_conn(inode);
1692 struct fuse_req *req;
1693 struct fuse_lk_out outarg;
1696 req = fuse_get_req_nopages(fc);
1698 return PTR_ERR(req);
1700 fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
1701 req->out.numargs = 1;
1702 req->out.args[0].size = sizeof(outarg);
1703 req->out.args[0].value = &outarg;
1704 fuse_request_send(fc, req);
1705 err = req->out.h.error;
1706 fuse_put_request(fc, req);
1708 err = convert_fuse_file_lock(&outarg.lk, fl);
1713 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
1715 struct inode *inode = file_inode(file);
1716 struct fuse_conn *fc = get_fuse_conn(inode);
1717 struct fuse_req *req;
1718 int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
1719 pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
1722 if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
1723 /* NLM needs asynchronous locks, which we don't support yet */
1727 /* Unlock on close is handled by the flush method */
1728 if (fl->fl_flags & FL_CLOSE)
1731 req = fuse_get_req_nopages(fc);
1733 return PTR_ERR(req);
1735 fuse_lk_fill(req, file, fl, opcode, pid, flock);
1736 fuse_request_send(fc, req);
1737 err = req->out.h.error;
1738 /* locking is restartable */
1741 fuse_put_request(fc, req);
1745 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
1747 struct inode *inode = file_inode(file);
1748 struct fuse_conn *fc = get_fuse_conn(inode);
1751 if (cmd == F_CANCELLK) {
1753 } else if (cmd == F_GETLK) {
1755 posix_test_lock(file, fl);
1758 err = fuse_getlk(file, fl);
1761 err = posix_lock_file(file, fl, NULL);
1763 err = fuse_setlk(file, fl, 0);
1768 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
1770 struct inode *inode = file_inode(file);
1771 struct fuse_conn *fc = get_fuse_conn(inode);
1775 err = flock_lock_file_wait(file, fl);
1777 struct fuse_file *ff = file->private_data;
1779 /* emulate flock with POSIX locks */
1780 fl->fl_owner = (fl_owner_t) file;
1782 err = fuse_setlk(file, fl, 1);
1788 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
1790 struct inode *inode = mapping->host;
1791 struct fuse_conn *fc = get_fuse_conn(inode);
1792 struct fuse_req *req;
1793 struct fuse_bmap_in inarg;
1794 struct fuse_bmap_out outarg;
1797 if (!inode->i_sb->s_bdev || fc->no_bmap)
1800 req = fuse_get_req_nopages(fc);
1804 memset(&inarg, 0, sizeof(inarg));
1805 inarg.block = block;
1806 inarg.blocksize = inode->i_sb->s_blocksize;
1807 req->in.h.opcode = FUSE_BMAP;
1808 req->in.h.nodeid = get_node_id(inode);
1809 req->in.numargs = 1;
1810 req->in.args[0].size = sizeof(inarg);
1811 req->in.args[0].value = &inarg;
1812 req->out.numargs = 1;
1813 req->out.args[0].size = sizeof(outarg);
1814 req->out.args[0].value = &outarg;
1815 fuse_request_send(fc, req);
1816 err = req->out.h.error;
1817 fuse_put_request(fc, req);
1821 return err ? 0 : outarg.block;
1824 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
1827 struct inode *inode = file_inode(file);
1829 /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
1830 if (whence == SEEK_CUR || whence == SEEK_SET)
1831 return generic_file_llseek(file, offset, whence);
1833 mutex_lock(&inode->i_mutex);
1834 retval = fuse_update_attributes(inode, NULL, file, NULL);
1836 retval = generic_file_llseek(file, offset, whence);
1837 mutex_unlock(&inode->i_mutex);
1842 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
1843 unsigned int nr_segs, size_t bytes, bool to_user)
1851 iov_iter_init(&ii, iov, nr_segs, bytes, 0);
1853 while (iov_iter_count(&ii)) {
1854 struct page *page = pages[page_idx++];
1855 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
1861 char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
1862 size_t iov_len = ii.iov->iov_len - ii.iov_offset;
1863 size_t copy = min(todo, iov_len);
1867 left = copy_from_user(kaddr, uaddr, copy);
1869 left = copy_to_user(uaddr, kaddr, copy);
1874 iov_iter_advance(&ii, copy);
1886 * CUSE servers compiled on 32bit broke on 64bit kernels because the
1887 * ABI was defined to be 'struct iovec' which is different on 32bit
1888 * and 64bit. Fortunately we can determine which structure the server
1889 * used from the size of the reply.
1891 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
1892 size_t transferred, unsigned count,
1895 #ifdef CONFIG_COMPAT
1896 if (count * sizeof(struct compat_iovec) == transferred) {
1897 struct compat_iovec *ciov = src;
1901 * With this interface a 32bit server cannot support
1902 * non-compat (i.e. ones coming from 64bit apps) ioctl
1908 for (i = 0; i < count; i++) {
1909 dst[i].iov_base = compat_ptr(ciov[i].iov_base);
1910 dst[i].iov_len = ciov[i].iov_len;
1916 if (count * sizeof(struct iovec) != transferred)
1919 memcpy(dst, src, transferred);
1923 /* Make sure iov_length() won't overflow */
1924 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
1927 u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
1929 for (n = 0; n < count; n++, iov++) {
1930 if (iov->iov_len > (size_t) max)
1932 max -= iov->iov_len;
1937 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
1938 void *src, size_t transferred, unsigned count,
1942 struct fuse_ioctl_iovec *fiov = src;
1944 if (fc->minor < 16) {
1945 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
1949 if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
1952 for (i = 0; i < count; i++) {
1953 /* Did the server supply an inappropriate value? */
1954 if (fiov[i].base != (unsigned long) fiov[i].base ||
1955 fiov[i].len != (unsigned long) fiov[i].len)
1958 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
1959 dst[i].iov_len = (size_t) fiov[i].len;
1961 #ifdef CONFIG_COMPAT
1963 (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
1964 (compat_size_t) dst[i].iov_len != fiov[i].len))
1974 * For ioctls, there is no generic way to determine how much memory
1975 * needs to be read and/or written. Furthermore, ioctls are allowed
1976 * to dereference the passed pointer, so the parameter requires deep
1977 * copying but FUSE has no idea whatsoever about what to copy in or
1980 * This is solved by allowing FUSE server to retry ioctl with
1981 * necessary in/out iovecs. Let's assume the ioctl implementation
1982 * needs to read in the following structure.
1989 * On the first callout to FUSE server, inarg->in_size and
1990 * inarg->out_size will be NULL; then, the server completes the ioctl
1991 * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
1992 * the actual iov array to
1994 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) } }
1996 * which tells FUSE to copy in the requested area and retry the ioctl.
1997 * On the second round, the server has access to the structure and
1998 * from that it can tell what to look for next, so on the invocation,
1999 * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2001 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) },
2002 * { .iov_base = a.buf, .iov_len = a.buflen } }
2004 * FUSE will copy both struct a and the pointed buffer from the
2005 * process doing the ioctl and retry ioctl with both struct a and the
2008 * This time, FUSE server has everything it needs and completes ioctl
2009 * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2011 * Copying data out works the same way.
2013 * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2014 * automatically initializes in and out iovs by decoding @cmd with
2015 * _IOC_* macros and the server is not allowed to request RETRY. This
2016 * limits ioctl data transfers to well-formed ioctls and is the forced
2017 * behavior for all FUSE servers.
2019 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2022 struct fuse_file *ff = file->private_data;
2023 struct fuse_conn *fc = ff->fc;
2024 struct fuse_ioctl_in inarg = {
2030 struct fuse_ioctl_out outarg;
2031 struct fuse_req *req = NULL;
2032 struct page **pages = NULL;
2033 struct iovec *iov_page = NULL;
2034 struct iovec *in_iov = NULL, *out_iov = NULL;
2035 unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
2036 size_t in_size, out_size, transferred;
2039 #if BITS_PER_LONG == 32
2040 inarg.flags |= FUSE_IOCTL_32BIT;
2042 if (flags & FUSE_IOCTL_COMPAT)
2043 inarg.flags |= FUSE_IOCTL_32BIT;
2046 /* assume all the iovs returned by client always fits in a page */
2047 BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2050 pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
2051 iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2052 if (!pages || !iov_page)
2056 * If restricted, initialize IO parameters as encoded in @cmd.
2057 * RETRY from server is not allowed.
2059 if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2060 struct iovec *iov = iov_page;
2062 iov->iov_base = (void __user *)arg;
2063 iov->iov_len = _IOC_SIZE(cmd);
2065 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2070 if (_IOC_DIR(cmd) & _IOC_READ) {
2077 inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2078 inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2081 * Out data can be used either for actual out data or iovs,
2082 * make sure there always is at least one page.
2084 out_size = max_t(size_t, out_size, PAGE_SIZE);
2085 max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2087 /* make sure there are enough buffer pages and init request with them */
2089 if (max_pages > FUSE_MAX_PAGES_PER_REQ)
2091 while (num_pages < max_pages) {
2092 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2093 if (!pages[num_pages])
2098 req = fuse_get_req(fc, num_pages);
2104 memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
2105 req->num_pages = num_pages;
2106 fuse_page_descs_length_init(req, 0, req->num_pages);
2108 /* okay, let's send it to the client */
2109 req->in.h.opcode = FUSE_IOCTL;
2110 req->in.h.nodeid = ff->nodeid;
2111 req->in.numargs = 1;
2112 req->in.args[0].size = sizeof(inarg);
2113 req->in.args[0].value = &inarg;
2116 req->in.args[1].size = in_size;
2117 req->in.argpages = 1;
2119 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
2125 req->out.numargs = 2;
2126 req->out.args[0].size = sizeof(outarg);
2127 req->out.args[0].value = &outarg;
2128 req->out.args[1].size = out_size;
2129 req->out.argpages = 1;
2130 req->out.argvar = 1;
2132 fuse_request_send(fc, req);
2133 err = req->out.h.error;
2134 transferred = req->out.args[1].size;
2135 fuse_put_request(fc, req);
2140 /* did it ask for retry? */
2141 if (outarg.flags & FUSE_IOCTL_RETRY) {
2144 /* no retry if in restricted mode */
2146 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2149 in_iovs = outarg.in_iovs;
2150 out_iovs = outarg.out_iovs;
2153 * Make sure things are in boundary, separate checks
2154 * are to protect against overflow.
2157 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2158 out_iovs > FUSE_IOCTL_MAX_IOV ||
2159 in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2162 vaddr = kmap_atomic(pages[0]);
2163 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2164 transferred, in_iovs + out_iovs,
2165 (flags & FUSE_IOCTL_COMPAT) != 0);
2166 kunmap_atomic(vaddr);
2171 out_iov = in_iov + in_iovs;
2173 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
2177 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
2185 if (transferred > inarg.out_size)
2188 err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
2191 fuse_put_request(fc, req);
2192 free_page((unsigned long) iov_page);
2194 __free_page(pages[--num_pages]);
2197 return err ? err : outarg.result;
2199 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2201 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2202 unsigned long arg, unsigned int flags)
2204 struct inode *inode = file_inode(file);
2205 struct fuse_conn *fc = get_fuse_conn(inode);
2207 if (!fuse_allow_current_process(fc))
2210 if (is_bad_inode(inode))
2213 return fuse_do_ioctl(file, cmd, arg, flags);
2216 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2219 return fuse_ioctl_common(file, cmd, arg, 0);
2222 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2225 return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2229 * All files which have been polled are linked to RB tree
2230 * fuse_conn->polled_files which is indexed by kh. Walk the tree and
2231 * find the matching one.
2233 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2234 struct rb_node **parent_out)
2236 struct rb_node **link = &fc->polled_files.rb_node;
2237 struct rb_node *last = NULL;
2240 struct fuse_file *ff;
2243 ff = rb_entry(last, struct fuse_file, polled_node);
2246 link = &last->rb_left;
2247 else if (kh > ff->kh)
2248 link = &last->rb_right;
2259 * The file is about to be polled. Make sure it's on the polled_files
2260 * RB tree. Note that files once added to the polled_files tree are
2261 * not removed before the file is released. This is because a file
2262 * polled once is likely to be polled again.
2264 static void fuse_register_polled_file(struct fuse_conn *fc,
2265 struct fuse_file *ff)
2267 spin_lock(&fc->lock);
2268 if (RB_EMPTY_NODE(&ff->polled_node)) {
2269 struct rb_node **link, *parent;
2271 link = fuse_find_polled_node(fc, ff->kh, &parent);
2273 rb_link_node(&ff->polled_node, parent, link);
2274 rb_insert_color(&ff->polled_node, &fc->polled_files);
2276 spin_unlock(&fc->lock);
2279 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2281 struct fuse_file *ff = file->private_data;
2282 struct fuse_conn *fc = ff->fc;
2283 struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2284 struct fuse_poll_out outarg;
2285 struct fuse_req *req;
2289 return DEFAULT_POLLMASK;
2291 poll_wait(file, &ff->poll_wait, wait);
2292 inarg.events = (__u32)poll_requested_events(wait);
2295 * Ask for notification iff there's someone waiting for it.
2296 * The client may ignore the flag and always notify.
2298 if (waitqueue_active(&ff->poll_wait)) {
2299 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2300 fuse_register_polled_file(fc, ff);
2303 req = fuse_get_req_nopages(fc);
2307 req->in.h.opcode = FUSE_POLL;
2308 req->in.h.nodeid = ff->nodeid;
2309 req->in.numargs = 1;
2310 req->in.args[0].size = sizeof(inarg);
2311 req->in.args[0].value = &inarg;
2312 req->out.numargs = 1;
2313 req->out.args[0].size = sizeof(outarg);
2314 req->out.args[0].value = &outarg;
2315 fuse_request_send(fc, req);
2316 err = req->out.h.error;
2317 fuse_put_request(fc, req);
2320 return outarg.revents;
2321 if (err == -ENOSYS) {
2323 return DEFAULT_POLLMASK;
2327 EXPORT_SYMBOL_GPL(fuse_file_poll);
2330 * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2331 * wakes up the poll waiters.
2333 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2334 struct fuse_notify_poll_wakeup_out *outarg)
2336 u64 kh = outarg->kh;
2337 struct rb_node **link;
2339 spin_lock(&fc->lock);
2341 link = fuse_find_polled_node(fc, kh, NULL);
2343 struct fuse_file *ff;
2345 ff = rb_entry(*link, struct fuse_file, polled_node);
2346 wake_up_interruptible_sync(&ff->poll_wait);
2349 spin_unlock(&fc->lock);
2353 static void fuse_do_truncate(struct file *file)
2355 struct inode *inode = file->f_mapping->host;
2358 attr.ia_valid = ATTR_SIZE;
2359 attr.ia_size = i_size_read(inode);
2361 attr.ia_file = file;
2362 attr.ia_valid |= ATTR_FILE;
2364 fuse_do_setattr(inode, &attr, file);
2368 fuse_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2369 loff_t offset, unsigned long nr_segs)
2372 struct file *file = iocb->ki_filp;
2373 struct fuse_file *ff = file->private_data;
2375 struct inode *inode;
2377 size_t count = iov_length(iov, nr_segs);
2378 struct fuse_io_priv *io;
2381 inode = file->f_mapping->host;
2382 i_size = i_size_read(inode);
2384 /* optimization for short read */
2385 if (rw != WRITE && offset + count > i_size) {
2386 if (offset >= i_size)
2388 count = i_size - offset;
2391 io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2394 spin_lock_init(&io->lock);
2398 io->offset = offset;
2399 io->write = (rw == WRITE);
2403 * By default, we want to optimize all I/Os with async request
2404 * submission to the client filesystem if supported.
2406 io->async = ff->fc->async_dio;
2410 * We cannot asynchronously extend the size of a file. We have no method
2411 * to wait on real async I/O requests, so we must submit this request
2414 if (!is_sync_kiocb(iocb) && (offset + count > i_size))
2418 ret = __fuse_direct_write(io, iov, nr_segs, &pos);
2420 ret = __fuse_direct_read(io, iov, nr_segs, &pos, count);
2423 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2425 /* we have a non-extending, async request, so return */
2426 if (ret > 0 && !is_sync_kiocb(iocb))
2427 return -EIOCBQUEUED;
2429 ret = wait_on_sync_kiocb(iocb);
2436 fuse_write_update_size(inode, pos);
2437 else if (ret < 0 && offset + count > i_size)
2438 fuse_do_truncate(file);
2444 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2447 struct fuse_file *ff = file->private_data;
2448 struct fuse_conn *fc = ff->fc;
2449 struct fuse_req *req;
2450 struct fuse_fallocate_in inarg = {
2458 if (fc->no_fallocate)
2461 req = fuse_get_req_nopages(fc);
2463 return PTR_ERR(req);
2465 req->in.h.opcode = FUSE_FALLOCATE;
2466 req->in.h.nodeid = ff->nodeid;
2467 req->in.numargs = 1;
2468 req->in.args[0].size = sizeof(inarg);
2469 req->in.args[0].value = &inarg;
2470 fuse_request_send(fc, req);
2471 err = req->out.h.error;
2472 if (err == -ENOSYS) {
2473 fc->no_fallocate = 1;
2476 fuse_put_request(fc, req);
2481 static const struct file_operations fuse_file_operations = {
2482 .llseek = fuse_file_llseek,
2483 .read = do_sync_read,
2484 .aio_read = fuse_file_aio_read,
2485 .write = do_sync_write,
2486 .aio_write = fuse_file_aio_write,
2487 .mmap = fuse_file_mmap,
2489 .flush = fuse_flush,
2490 .release = fuse_release,
2491 .fsync = fuse_fsync,
2492 .lock = fuse_file_lock,
2493 .flock = fuse_file_flock,
2494 .splice_read = generic_file_splice_read,
2495 .unlocked_ioctl = fuse_file_ioctl,
2496 .compat_ioctl = fuse_file_compat_ioctl,
2497 .poll = fuse_file_poll,
2498 .fallocate = fuse_file_fallocate,
2501 static const struct file_operations fuse_direct_io_file_operations = {
2502 .llseek = fuse_file_llseek,
2503 .read = fuse_direct_read,
2504 .write = fuse_direct_write,
2505 .mmap = fuse_direct_mmap,
2507 .flush = fuse_flush,
2508 .release = fuse_release,
2509 .fsync = fuse_fsync,
2510 .lock = fuse_file_lock,
2511 .flock = fuse_file_flock,
2512 .unlocked_ioctl = fuse_file_ioctl,
2513 .compat_ioctl = fuse_file_compat_ioctl,
2514 .poll = fuse_file_poll,
2515 .fallocate = fuse_file_fallocate,
2516 /* no splice_read */
2519 static const struct address_space_operations fuse_file_aops = {
2520 .readpage = fuse_readpage,
2521 .writepage = fuse_writepage,
2522 .launder_page = fuse_launder_page,
2523 .readpages = fuse_readpages,
2524 .set_page_dirty = __set_page_dirty_nobuffers,
2526 .direct_IO = fuse_direct_IO,
2529 void fuse_init_file_inode(struct inode *inode)
2531 inode->i_fop = &fuse_file_operations;
2532 inode->i_data.a_ops = &fuse_file_aops;