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;
129 fuse_request_send(ff->fc, req);
130 path_put(&req->misc.release.path);
131 fuse_put_request(ff->fc, req);
133 req->end = fuse_release_end;
134 fuse_request_send_background(ff->fc, req);
140 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
143 struct fuse_open_out outarg;
144 struct fuse_file *ff;
146 int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
148 ff = fuse_file_alloc(fc);
152 err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
159 outarg.open_flags &= ~FOPEN_DIRECT_IO;
163 ff->open_flags = outarg.open_flags;
164 file->private_data = fuse_file_get(ff);
168 EXPORT_SYMBOL_GPL(fuse_do_open);
170 void fuse_finish_open(struct inode *inode, struct file *file)
172 struct fuse_file *ff = file->private_data;
173 struct fuse_conn *fc = get_fuse_conn(inode);
175 if (ff->open_flags & FOPEN_DIRECT_IO)
176 file->f_op = &fuse_direct_io_file_operations;
177 if (!(ff->open_flags & FOPEN_KEEP_CACHE))
178 invalidate_inode_pages2(inode->i_mapping);
179 if (ff->open_flags & FOPEN_NONSEEKABLE)
180 nonseekable_open(inode, file);
181 if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
182 struct fuse_inode *fi = get_fuse_inode(inode);
184 spin_lock(&fc->lock);
185 fi->attr_version = ++fc->attr_version;
186 i_size_write(inode, 0);
187 spin_unlock(&fc->lock);
188 fuse_invalidate_attr(inode);
192 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
194 struct fuse_conn *fc = get_fuse_conn(inode);
197 err = generic_file_open(inode, file);
201 err = fuse_do_open(fc, get_node_id(inode), file, isdir);
205 fuse_finish_open(inode, file);
210 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
212 struct fuse_conn *fc = ff->fc;
213 struct fuse_req *req = ff->reserved_req;
214 struct fuse_release_in *inarg = &req->misc.release.in;
216 spin_lock(&fc->lock);
217 list_del(&ff->write_entry);
218 if (!RB_EMPTY_NODE(&ff->polled_node))
219 rb_erase(&ff->polled_node, &fc->polled_files);
220 spin_unlock(&fc->lock);
222 wake_up_interruptible_all(&ff->poll_wait);
225 inarg->flags = flags;
226 req->in.h.opcode = opcode;
227 req->in.h.nodeid = ff->nodeid;
229 req->in.args[0].size = sizeof(struct fuse_release_in);
230 req->in.args[0].value = inarg;
233 void fuse_release_common(struct file *file, int opcode)
235 struct fuse_file *ff;
236 struct fuse_req *req;
238 ff = file->private_data;
242 req = ff->reserved_req;
243 fuse_prepare_release(ff, file->f_flags, opcode);
246 struct fuse_release_in *inarg = &req->misc.release.in;
247 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
248 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
251 /* Hold vfsmount and dentry until release is finished */
252 path_get(&file->f_path);
253 req->misc.release.path = file->f_path;
256 * Normally this will send the RELEASE request, however if
257 * some asynchronous READ or WRITE requests are outstanding,
258 * the sending will be delayed.
260 * Make the release synchronous if this is a fuseblk mount,
261 * synchronous RELEASE is allowed (and desirable) in this case
262 * because the server can be trusted not to screw up.
264 fuse_file_put(ff, ff->fc->destroy_req != NULL);
267 static int fuse_open(struct inode *inode, struct file *file)
269 return fuse_open_common(inode, file, false);
272 static int fuse_release(struct inode *inode, struct file *file)
274 fuse_release_common(file, FUSE_RELEASE);
276 /* return value is ignored by VFS */
280 void fuse_sync_release(struct fuse_file *ff, int flags)
282 WARN_ON(atomic_read(&ff->count) > 1);
283 fuse_prepare_release(ff, flags, FUSE_RELEASE);
284 ff->reserved_req->force = 1;
285 fuse_request_send(ff->fc, ff->reserved_req);
286 fuse_put_request(ff->fc, ff->reserved_req);
289 EXPORT_SYMBOL_GPL(fuse_sync_release);
292 * Scramble the ID space with XTEA, so that the value of the files_struct
293 * pointer is not exposed to userspace.
295 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
297 u32 *k = fc->scramble_key;
298 u64 v = (unsigned long) id;
304 for (i = 0; i < 32; i++) {
305 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
307 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
310 return (u64) v0 + ((u64) v1 << 32);
314 * Check if page is under writeback
316 * This is currently done by walking the list of writepage requests
317 * for the inode, which can be pretty inefficient.
319 static bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
321 struct fuse_conn *fc = get_fuse_conn(inode);
322 struct fuse_inode *fi = get_fuse_inode(inode);
323 struct fuse_req *req;
326 spin_lock(&fc->lock);
327 list_for_each_entry(req, &fi->writepages, writepages_entry) {
330 BUG_ON(req->inode != inode);
331 curr_index = req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
332 if (curr_index == index) {
337 spin_unlock(&fc->lock);
343 * Wait for page writeback to be completed.
345 * Since fuse doesn't rely on the VM writeback tracking, this has to
346 * use some other means.
348 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
350 struct fuse_inode *fi = get_fuse_inode(inode);
352 wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
356 static int fuse_flush(struct file *file, fl_owner_t id)
358 struct inode *inode = file_inode(file);
359 struct fuse_conn *fc = get_fuse_conn(inode);
360 struct fuse_file *ff = file->private_data;
361 struct fuse_req *req;
362 struct fuse_flush_in inarg;
365 if (is_bad_inode(inode))
371 req = fuse_get_req_nofail_nopages(fc, file);
372 memset(&inarg, 0, sizeof(inarg));
374 inarg.lock_owner = fuse_lock_owner_id(fc, id);
375 req->in.h.opcode = FUSE_FLUSH;
376 req->in.h.nodeid = get_node_id(inode);
378 req->in.args[0].size = sizeof(inarg);
379 req->in.args[0].value = &inarg;
381 fuse_request_send(fc, req);
382 err = req->out.h.error;
383 fuse_put_request(fc, req);
384 if (err == -ENOSYS) {
392 * Wait for all pending writepages on the inode to finish.
394 * This is currently done by blocking further writes with FUSE_NOWRITE
395 * and waiting for all sent writes to complete.
397 * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
398 * could conflict with truncation.
400 static void fuse_sync_writes(struct inode *inode)
402 fuse_set_nowrite(inode);
403 fuse_release_nowrite(inode);
406 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
407 int datasync, int isdir)
409 struct inode *inode = file->f_mapping->host;
410 struct fuse_conn *fc = get_fuse_conn(inode);
411 struct fuse_file *ff = file->private_data;
412 struct fuse_req *req;
413 struct fuse_fsync_in inarg;
416 if (is_bad_inode(inode))
419 err = filemap_write_and_wait_range(inode->i_mapping, start, end);
423 if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
426 mutex_lock(&inode->i_mutex);
429 * Start writeback against all dirty pages of the inode, then
430 * wait for all outstanding writes, before sending the FSYNC
433 err = write_inode_now(inode, 0);
437 fuse_sync_writes(inode);
439 req = fuse_get_req_nopages(fc);
445 memset(&inarg, 0, sizeof(inarg));
447 inarg.fsync_flags = datasync ? 1 : 0;
448 req->in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
449 req->in.h.nodeid = get_node_id(inode);
451 req->in.args[0].size = sizeof(inarg);
452 req->in.args[0].value = &inarg;
453 fuse_request_send(fc, req);
454 err = req->out.h.error;
455 fuse_put_request(fc, req);
456 if (err == -ENOSYS) {
464 mutex_unlock(&inode->i_mutex);
468 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
471 return fuse_fsync_common(file, start, end, datasync, 0);
474 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
475 size_t count, int opcode)
477 struct fuse_read_in *inarg = &req->misc.read.in;
478 struct fuse_file *ff = file->private_data;
483 inarg->flags = file->f_flags;
484 req->in.h.opcode = opcode;
485 req->in.h.nodeid = ff->nodeid;
487 req->in.args[0].size = sizeof(struct fuse_read_in);
488 req->in.args[0].value = inarg;
490 req->out.numargs = 1;
491 req->out.args[0].size = count;
494 static size_t fuse_send_read(struct fuse_req *req, struct file *file,
495 loff_t pos, size_t count, fl_owner_t owner)
497 struct fuse_file *ff = file->private_data;
498 struct fuse_conn *fc = ff->fc;
500 fuse_read_fill(req, file, pos, count, FUSE_READ);
502 struct fuse_read_in *inarg = &req->misc.read.in;
504 inarg->read_flags |= FUSE_READ_LOCKOWNER;
505 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
507 fuse_request_send(fc, req);
508 return req->out.args[0].size;
511 static void fuse_read_update_size(struct inode *inode, loff_t size,
514 struct fuse_conn *fc = get_fuse_conn(inode);
515 struct fuse_inode *fi = get_fuse_inode(inode);
517 spin_lock(&fc->lock);
518 if (attr_ver == fi->attr_version && size < inode->i_size) {
519 fi->attr_version = ++fc->attr_version;
520 i_size_write(inode, size);
522 spin_unlock(&fc->lock);
525 static int fuse_readpage(struct file *file, struct page *page)
527 struct inode *inode = page->mapping->host;
528 struct fuse_conn *fc = get_fuse_conn(inode);
529 struct fuse_req *req;
531 loff_t pos = page_offset(page);
532 size_t count = PAGE_CACHE_SIZE;
537 if (is_bad_inode(inode))
541 * Page writeback can extend beyond the lifetime of the
542 * page-cache page, so make sure we read a properly synced
545 fuse_wait_on_page_writeback(inode, page->index);
547 req = fuse_get_req(fc, 1);
552 attr_ver = fuse_get_attr_version(fc);
554 req->out.page_zeroing = 1;
555 req->out.argpages = 1;
557 req->pages[0] = page;
558 req->page_descs[0].length = count;
559 num_read = fuse_send_read(req, file, pos, count, NULL);
560 err = req->out.h.error;
561 fuse_put_request(fc, req);
565 * Short read means EOF. If file size is larger, truncate it
567 if (num_read < count)
568 fuse_read_update_size(inode, pos + num_read, attr_ver);
570 SetPageUptodate(page);
573 fuse_invalidate_attr(inode); /* atime changed */
579 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
582 size_t count = req->misc.read.in.size;
583 size_t num_read = req->out.args[0].size;
584 struct address_space *mapping = NULL;
586 for (i = 0; mapping == NULL && i < req->num_pages; i++)
587 mapping = req->pages[i]->mapping;
590 struct inode *inode = mapping->host;
593 * Short read means EOF. If file size is larger, truncate it
595 if (!req->out.h.error && num_read < count) {
598 pos = page_offset(req->pages[0]) + num_read;
599 fuse_read_update_size(inode, pos,
600 req->misc.read.attr_ver);
602 fuse_invalidate_attr(inode); /* atime changed */
605 for (i = 0; i < req->num_pages; i++) {
606 struct page *page = req->pages[i];
607 if (!req->out.h.error)
608 SetPageUptodate(page);
612 page_cache_release(page);
615 fuse_file_put(req->ff, false);
618 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
620 struct fuse_file *ff = file->private_data;
621 struct fuse_conn *fc = ff->fc;
622 loff_t pos = page_offset(req->pages[0]);
623 size_t count = req->num_pages << PAGE_CACHE_SHIFT;
625 req->out.argpages = 1;
626 req->out.page_zeroing = 1;
627 req->out.page_replace = 1;
628 fuse_read_fill(req, file, pos, count, FUSE_READ);
629 req->misc.read.attr_ver = fuse_get_attr_version(fc);
630 if (fc->async_read) {
631 req->ff = fuse_file_get(ff);
632 req->end = fuse_readpages_end;
633 fuse_request_send_background(fc, req);
635 fuse_request_send(fc, req);
636 fuse_readpages_end(fc, req);
637 fuse_put_request(fc, req);
641 struct fuse_fill_data {
642 struct fuse_req *req;
648 static int fuse_readpages_fill(void *_data, struct page *page)
650 struct fuse_fill_data *data = _data;
651 struct fuse_req *req = data->req;
652 struct inode *inode = data->inode;
653 struct fuse_conn *fc = get_fuse_conn(inode);
655 fuse_wait_on_page_writeback(inode, page->index);
657 if (req->num_pages &&
658 (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
659 (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
660 req->pages[req->num_pages - 1]->index + 1 != page->index)) {
661 int nr_alloc = min_t(unsigned, data->nr_pages,
662 FUSE_MAX_PAGES_PER_REQ);
663 fuse_send_readpages(req, data->file);
664 data->req = req = fuse_get_req(fc, nr_alloc);
671 if (WARN_ON(req->num_pages >= req->max_pages)) {
672 fuse_put_request(fc, req);
676 page_cache_get(page);
677 req->pages[req->num_pages] = page;
678 req->page_descs[req->num_pages].length = PAGE_SIZE;
684 static int fuse_readpages(struct file *file, struct address_space *mapping,
685 struct list_head *pages, unsigned nr_pages)
687 struct inode *inode = mapping->host;
688 struct fuse_conn *fc = get_fuse_conn(inode);
689 struct fuse_fill_data data;
691 int nr_alloc = min_t(unsigned, nr_pages, FUSE_MAX_PAGES_PER_REQ);
694 if (is_bad_inode(inode))
699 data.req = fuse_get_req(fc, nr_alloc);
700 data.nr_pages = nr_pages;
701 err = PTR_ERR(data.req);
702 if (IS_ERR(data.req))
705 err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
707 if (data.req->num_pages)
708 fuse_send_readpages(data.req, file);
710 fuse_put_request(fc, data.req);
716 static ssize_t fuse_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
717 unsigned long nr_segs, loff_t pos)
719 struct inode *inode = iocb->ki_filp->f_mapping->host;
720 struct fuse_conn *fc = get_fuse_conn(inode);
723 * In auto invalidate mode, always update attributes on read.
724 * Otherwise, only update if we attempt to read past EOF (to ensure
725 * i_size is up to date).
727 if (fc->auto_inval_data ||
728 (pos + iov_length(iov, nr_segs) > i_size_read(inode))) {
730 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
735 return generic_file_aio_read(iocb, iov, nr_segs, pos);
738 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
739 loff_t pos, size_t count)
741 struct fuse_write_in *inarg = &req->misc.write.in;
742 struct fuse_write_out *outarg = &req->misc.write.out;
747 req->in.h.opcode = FUSE_WRITE;
748 req->in.h.nodeid = ff->nodeid;
750 if (ff->fc->minor < 9)
751 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
753 req->in.args[0].size = sizeof(struct fuse_write_in);
754 req->in.args[0].value = inarg;
755 req->in.args[1].size = count;
756 req->out.numargs = 1;
757 req->out.args[0].size = sizeof(struct fuse_write_out);
758 req->out.args[0].value = outarg;
761 static size_t fuse_send_write(struct fuse_req *req, struct file *file,
762 loff_t pos, size_t count, fl_owner_t owner)
764 struct fuse_file *ff = file->private_data;
765 struct fuse_conn *fc = ff->fc;
766 struct fuse_write_in *inarg = &req->misc.write.in;
768 fuse_write_fill(req, ff, pos, count);
769 inarg->flags = file->f_flags;
771 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
772 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
774 fuse_request_send(fc, req);
775 return req->misc.write.out.size;
778 void fuse_write_update_size(struct inode *inode, loff_t pos)
780 struct fuse_conn *fc = get_fuse_conn(inode);
781 struct fuse_inode *fi = get_fuse_inode(inode);
783 spin_lock(&fc->lock);
784 fi->attr_version = ++fc->attr_version;
785 if (pos > inode->i_size)
786 i_size_write(inode, pos);
787 spin_unlock(&fc->lock);
790 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
791 struct inode *inode, loff_t pos,
798 for (i = 0; i < req->num_pages; i++)
799 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
801 res = fuse_send_write(req, file, pos, count, NULL);
803 offset = req->page_descs[0].offset;
805 for (i = 0; i < req->num_pages; i++) {
806 struct page *page = req->pages[i];
808 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
809 SetPageUptodate(page);
811 if (count > PAGE_CACHE_SIZE - offset)
812 count -= PAGE_CACHE_SIZE - offset;
818 page_cache_release(page);
824 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
825 struct address_space *mapping,
826 struct iov_iter *ii, loff_t pos)
828 struct fuse_conn *fc = get_fuse_conn(mapping->host);
829 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
833 req->in.argpages = 1;
834 req->page_descs[0].offset = offset;
839 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
840 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
843 bytes = min_t(size_t, bytes, fc->max_write - count);
847 if (iov_iter_fault_in_readable(ii, bytes))
851 page = grab_cache_page_write_begin(mapping, index, 0);
855 if (mapping_writably_mapped(mapping))
856 flush_dcache_page(page);
859 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
861 flush_dcache_page(page);
863 mark_page_accessed(page);
867 page_cache_release(page);
868 bytes = min(bytes, iov_iter_single_seg_count(ii));
873 req->pages[req->num_pages] = page;
874 req->page_descs[req->num_pages].length = tmp;
877 iov_iter_advance(ii, tmp);
881 if (offset == PAGE_CACHE_SIZE)
886 } while (iov_iter_count(ii) && count < fc->max_write &&
887 req->num_pages < req->max_pages && offset == 0);
889 return count > 0 ? count : err;
892 static inline unsigned fuse_wr_pages(loff_t pos, size_t len)
894 return min_t(unsigned,
895 ((pos + len - 1) >> PAGE_CACHE_SHIFT) -
896 (pos >> PAGE_CACHE_SHIFT) + 1,
897 FUSE_MAX_PAGES_PER_REQ);
900 static ssize_t fuse_perform_write(struct file *file,
901 struct address_space *mapping,
902 struct iov_iter *ii, loff_t pos)
904 struct inode *inode = mapping->host;
905 struct fuse_conn *fc = get_fuse_conn(inode);
909 if (is_bad_inode(inode))
913 struct fuse_req *req;
915 unsigned nr_pages = fuse_wr_pages(pos, iov_iter_count(ii));
917 req = fuse_get_req(fc, nr_pages);
923 count = fuse_fill_write_pages(req, mapping, ii, pos);
929 num_written = fuse_send_write_pages(req, file, inode,
931 err = req->out.h.error;
936 /* break out of the loop on short write */
937 if (num_written != count)
941 fuse_put_request(fc, req);
942 } while (!err && iov_iter_count(ii));
945 fuse_write_update_size(inode, pos);
947 fuse_invalidate_attr(inode);
949 return res > 0 ? res : err;
952 static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
953 unsigned long nr_segs, loff_t pos)
955 struct file *file = iocb->ki_filp;
956 struct address_space *mapping = file->f_mapping;
960 ssize_t written_buffered = 0;
961 struct inode *inode = mapping->host;
966 WARN_ON(iocb->ki_pos != pos);
969 err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
974 mutex_lock(&inode->i_mutex);
976 /* We can write back this queue in page reclaim */
977 current->backing_dev_info = mapping->backing_dev_info;
979 err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
986 err = file_remove_suid(file);
990 err = file_update_time(file);
994 if (file->f_flags & O_DIRECT) {
995 written = generic_file_direct_write(iocb, iov, &nr_segs,
998 if (written < 0 || written == count)
1004 iov_iter_init(&i, iov, nr_segs, count, written);
1005 written_buffered = fuse_perform_write(file, mapping, &i, pos);
1006 if (written_buffered < 0) {
1007 err = written_buffered;
1010 endbyte = pos + written_buffered - 1;
1012 err = filemap_write_and_wait_range(file->f_mapping, pos,
1017 invalidate_mapping_pages(file->f_mapping,
1018 pos >> PAGE_CACHE_SHIFT,
1019 endbyte >> PAGE_CACHE_SHIFT);
1021 written += written_buffered;
1022 iocb->ki_pos = pos + written_buffered;
1024 iov_iter_init(&i, iov, nr_segs, count, 0);
1025 written = fuse_perform_write(file, mapping, &i, pos);
1027 iocb->ki_pos = pos + written;
1030 current->backing_dev_info = NULL;
1031 mutex_unlock(&inode->i_mutex);
1033 return written ? written : err;
1036 static void fuse_release_user_pages(struct fuse_req *req, int write)
1040 for (i = 0; i < req->num_pages; i++) {
1041 struct page *page = req->pages[i];
1043 set_page_dirty_lock(page);
1048 static inline void fuse_page_descs_length_init(struct fuse_req *req,
1049 unsigned index, unsigned nr_pages)
1053 for (i = index; i < index + nr_pages; i++)
1054 req->page_descs[i].length = PAGE_SIZE -
1055 req->page_descs[i].offset;
1058 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1060 return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1063 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1066 return min(iov_iter_single_seg_count(ii), max_size);
1069 static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1070 size_t *nbytesp, int write)
1072 size_t nbytes = 0; /* # bytes already packed in req */
1074 /* Special case for kernel I/O: can copy directly into the buffer */
1075 if (segment_eq(get_fs(), KERNEL_DS)) {
1076 unsigned long user_addr = fuse_get_user_addr(ii);
1077 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1080 req->in.args[1].value = (void *) user_addr;
1082 req->out.args[0].value = (void *) user_addr;
1084 iov_iter_advance(ii, frag_size);
1085 *nbytesp = frag_size;
1089 while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1091 unsigned long user_addr = fuse_get_user_addr(ii);
1092 unsigned offset = user_addr & ~PAGE_MASK;
1093 size_t frag_size = fuse_get_frag_size(ii, *nbytesp - nbytes);
1096 unsigned n = req->max_pages - req->num_pages;
1097 frag_size = min_t(size_t, frag_size, n << PAGE_SHIFT);
1099 npages = (frag_size + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1100 npages = clamp(npages, 1U, n);
1102 ret = get_user_pages_fast(user_addr, npages, !write,
1103 &req->pages[req->num_pages]);
1108 frag_size = min_t(size_t, frag_size,
1109 (npages << PAGE_SHIFT) - offset);
1110 iov_iter_advance(ii, frag_size);
1112 req->page_descs[req->num_pages].offset = offset;
1113 fuse_page_descs_length_init(req, req->num_pages, npages);
1115 req->num_pages += npages;
1116 req->page_descs[req->num_pages - 1].length -=
1117 (npages << PAGE_SHIFT) - offset - frag_size;
1119 nbytes += frag_size;
1123 req->in.argpages = 1;
1125 req->out.argpages = 1;
1132 static inline int fuse_iter_npages(const struct iov_iter *ii_p)
1134 struct iov_iter ii = *ii_p;
1137 while (iov_iter_count(&ii) && npages < FUSE_MAX_PAGES_PER_REQ) {
1138 unsigned long user_addr = fuse_get_user_addr(&ii);
1139 unsigned offset = user_addr & ~PAGE_MASK;
1140 size_t frag_size = iov_iter_single_seg_count(&ii);
1142 npages += (frag_size + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1143 iov_iter_advance(&ii, frag_size);
1146 return min(npages, FUSE_MAX_PAGES_PER_REQ);
1149 ssize_t fuse_direct_io(struct file *file, const struct iovec *iov,
1150 unsigned long nr_segs, size_t count, loff_t *ppos,
1153 struct fuse_file *ff = file->private_data;
1154 struct fuse_conn *fc = ff->fc;
1155 size_t nmax = write ? fc->max_write : fc->max_read;
1158 struct fuse_req *req;
1161 iov_iter_init(&ii, iov, nr_segs, count, 0);
1163 req = fuse_get_req(fc, fuse_iter_npages(&ii));
1165 return PTR_ERR(req);
1169 fl_owner_t owner = current->files;
1170 size_t nbytes = min(count, nmax);
1171 int err = fuse_get_user_pages(req, &ii, &nbytes, write);
1178 nres = fuse_send_write(req, file, pos, nbytes, owner);
1180 nres = fuse_send_read(req, file, pos, nbytes, owner);
1182 fuse_release_user_pages(req, !write);
1183 if (req->out.h.error) {
1185 res = req->out.h.error;
1187 } else if (nres > nbytes) {
1197 fuse_put_request(fc, req);
1198 req = fuse_get_req(fc, fuse_iter_npages(&ii));
1204 fuse_put_request(fc, req);
1210 EXPORT_SYMBOL_GPL(fuse_direct_io);
1212 static ssize_t __fuse_direct_read(struct file *file, const struct iovec *iov,
1213 unsigned long nr_segs, loff_t *ppos)
1216 struct inode *inode = file_inode(file);
1218 if (is_bad_inode(inode))
1221 res = fuse_direct_io(file, iov, nr_segs, iov_length(iov, nr_segs),
1224 fuse_invalidate_attr(inode);
1229 static ssize_t fuse_direct_read(struct file *file, char __user *buf,
1230 size_t count, loff_t *ppos)
1232 struct iovec iov = { .iov_base = buf, .iov_len = count };
1233 return __fuse_direct_read(file, &iov, 1, ppos);
1236 static ssize_t __fuse_direct_write(struct file *file, const struct iovec *iov,
1237 unsigned long nr_segs, loff_t *ppos)
1239 struct inode *inode = file_inode(file);
1240 size_t count = iov_length(iov, nr_segs);
1243 res = generic_write_checks(file, ppos, &count, 0);
1245 res = fuse_direct_io(file, iov, nr_segs, count, ppos, 1);
1247 fuse_write_update_size(inode, *ppos);
1250 fuse_invalidate_attr(inode);
1255 static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
1256 size_t count, loff_t *ppos)
1258 struct iovec iov = { .iov_base = (void __user *)buf, .iov_len = count };
1259 struct inode *inode = file_inode(file);
1262 if (is_bad_inode(inode))
1265 /* Don't allow parallel writes to the same file */
1266 mutex_lock(&inode->i_mutex);
1267 res = __fuse_direct_write(file, &iov, 1, ppos);
1268 mutex_unlock(&inode->i_mutex);
1273 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1275 __free_page(req->pages[0]);
1276 fuse_file_put(req->ff, false);
1279 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1281 struct inode *inode = req->inode;
1282 struct fuse_inode *fi = get_fuse_inode(inode);
1283 struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
1285 list_del(&req->writepages_entry);
1286 dec_bdi_stat(bdi, BDI_WRITEBACK);
1287 dec_zone_page_state(req->pages[0], NR_WRITEBACK_TEMP);
1288 bdi_writeout_inc(bdi);
1289 wake_up(&fi->page_waitq);
1292 /* Called under fc->lock, may release and reacquire it */
1293 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req)
1294 __releases(fc->lock)
1295 __acquires(fc->lock)
1297 struct fuse_inode *fi = get_fuse_inode(req->inode);
1298 loff_t size = i_size_read(req->inode);
1299 struct fuse_write_in *inarg = &req->misc.write.in;
1304 if (inarg->offset + PAGE_CACHE_SIZE <= size) {
1305 inarg->size = PAGE_CACHE_SIZE;
1306 } else if (inarg->offset < size) {
1307 inarg->size = size & (PAGE_CACHE_SIZE - 1);
1309 /* Got truncated off completely */
1313 req->in.args[1].size = inarg->size;
1315 fuse_request_send_background_locked(fc, req);
1319 fuse_writepage_finish(fc, req);
1320 spin_unlock(&fc->lock);
1321 fuse_writepage_free(fc, req);
1322 fuse_put_request(fc, req);
1323 spin_lock(&fc->lock);
1327 * If fi->writectr is positive (no truncate or fsync going on) send
1328 * all queued writepage requests.
1330 * Called with fc->lock
1332 void fuse_flush_writepages(struct inode *inode)
1333 __releases(fc->lock)
1334 __acquires(fc->lock)
1336 struct fuse_conn *fc = get_fuse_conn(inode);
1337 struct fuse_inode *fi = get_fuse_inode(inode);
1338 struct fuse_req *req;
1340 while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1341 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1342 list_del_init(&req->list);
1343 fuse_send_writepage(fc, req);
1347 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1349 struct inode *inode = req->inode;
1350 struct fuse_inode *fi = get_fuse_inode(inode);
1352 mapping_set_error(inode->i_mapping, req->out.h.error);
1353 spin_lock(&fc->lock);
1355 fuse_writepage_finish(fc, req);
1356 spin_unlock(&fc->lock);
1357 fuse_writepage_free(fc, req);
1360 static int fuse_writepage_locked(struct page *page)
1362 struct address_space *mapping = page->mapping;
1363 struct inode *inode = mapping->host;
1364 struct fuse_conn *fc = get_fuse_conn(inode);
1365 struct fuse_inode *fi = get_fuse_inode(inode);
1366 struct fuse_req *req;
1367 struct fuse_file *ff;
1368 struct page *tmp_page;
1370 set_page_writeback(page);
1372 req = fuse_request_alloc_nofs(1);
1376 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1380 spin_lock(&fc->lock);
1381 BUG_ON(list_empty(&fi->write_files));
1382 ff = list_entry(fi->write_files.next, struct fuse_file, write_entry);
1383 req->ff = fuse_file_get(ff);
1384 spin_unlock(&fc->lock);
1386 fuse_write_fill(req, ff, page_offset(page), 0);
1388 copy_highpage(tmp_page, page);
1389 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1390 req->in.argpages = 1;
1392 req->pages[0] = tmp_page;
1393 req->page_descs[0].offset = 0;
1394 req->page_descs[0].length = PAGE_SIZE;
1395 req->end = fuse_writepage_end;
1398 inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
1399 inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1400 end_page_writeback(page);
1402 spin_lock(&fc->lock);
1403 list_add(&req->writepages_entry, &fi->writepages);
1404 list_add_tail(&req->list, &fi->queued_writes);
1405 fuse_flush_writepages(inode);
1406 spin_unlock(&fc->lock);
1411 fuse_request_free(req);
1413 end_page_writeback(page);
1417 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1421 err = fuse_writepage_locked(page);
1427 static int fuse_launder_page(struct page *page)
1430 if (clear_page_dirty_for_io(page)) {
1431 struct inode *inode = page->mapping->host;
1432 err = fuse_writepage_locked(page);
1434 fuse_wait_on_page_writeback(inode, page->index);
1440 * Write back dirty pages now, because there may not be any suitable
1443 static void fuse_vma_close(struct vm_area_struct *vma)
1445 filemap_write_and_wait(vma->vm_file->f_mapping);
1449 * Wait for writeback against this page to complete before allowing it
1450 * to be marked dirty again, and hence written back again, possibly
1451 * before the previous writepage completed.
1453 * Block here, instead of in ->writepage(), so that the userspace fs
1454 * can only block processes actually operating on the filesystem.
1456 * Otherwise unprivileged userspace fs would be able to block
1461 * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
1463 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1465 struct page *page = vmf->page;
1467 * Don't use page->mapping as it may become NULL from a
1468 * concurrent truncate.
1470 struct inode *inode = vma->vm_file->f_mapping->host;
1472 fuse_wait_on_page_writeback(inode, page->index);
1476 static const struct vm_operations_struct fuse_file_vm_ops = {
1477 .close = fuse_vma_close,
1478 .fault = filemap_fault,
1479 .page_mkwrite = fuse_page_mkwrite,
1480 .remap_pages = generic_file_remap_pages,
1483 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
1485 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE)) {
1486 struct inode *inode = file_inode(file);
1487 struct fuse_conn *fc = get_fuse_conn(inode);
1488 struct fuse_inode *fi = get_fuse_inode(inode);
1489 struct fuse_file *ff = file->private_data;
1491 * file may be written through mmap, so chain it onto the
1492 * inodes's write_file list
1494 spin_lock(&fc->lock);
1495 if (list_empty(&ff->write_entry))
1496 list_add(&ff->write_entry, &fi->write_files);
1497 spin_unlock(&fc->lock);
1499 file_accessed(file);
1500 vma->vm_ops = &fuse_file_vm_ops;
1504 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
1506 /* Can't provide the coherency needed for MAP_SHARED */
1507 if (vma->vm_flags & VM_MAYSHARE)
1510 invalidate_inode_pages2(file->f_mapping);
1512 return generic_file_mmap(file, vma);
1515 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
1516 struct file_lock *fl)
1518 switch (ffl->type) {
1524 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
1525 ffl->end < ffl->start)
1528 fl->fl_start = ffl->start;
1529 fl->fl_end = ffl->end;
1530 fl->fl_pid = ffl->pid;
1536 fl->fl_type = ffl->type;
1540 static void fuse_lk_fill(struct fuse_req *req, struct file *file,
1541 const struct file_lock *fl, int opcode, pid_t pid,
1544 struct inode *inode = file_inode(file);
1545 struct fuse_conn *fc = get_fuse_conn(inode);
1546 struct fuse_file *ff = file->private_data;
1547 struct fuse_lk_in *arg = &req->misc.lk_in;
1550 arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
1551 arg->lk.start = fl->fl_start;
1552 arg->lk.end = fl->fl_end;
1553 arg->lk.type = fl->fl_type;
1556 arg->lk_flags |= FUSE_LK_FLOCK;
1557 req->in.h.opcode = opcode;
1558 req->in.h.nodeid = get_node_id(inode);
1559 req->in.numargs = 1;
1560 req->in.args[0].size = sizeof(*arg);
1561 req->in.args[0].value = arg;
1564 static int fuse_getlk(struct file *file, struct file_lock *fl)
1566 struct inode *inode = file_inode(file);
1567 struct fuse_conn *fc = get_fuse_conn(inode);
1568 struct fuse_req *req;
1569 struct fuse_lk_out outarg;
1572 req = fuse_get_req_nopages(fc);
1574 return PTR_ERR(req);
1576 fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
1577 req->out.numargs = 1;
1578 req->out.args[0].size = sizeof(outarg);
1579 req->out.args[0].value = &outarg;
1580 fuse_request_send(fc, req);
1581 err = req->out.h.error;
1582 fuse_put_request(fc, req);
1584 err = convert_fuse_file_lock(&outarg.lk, fl);
1589 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
1591 struct inode *inode = file_inode(file);
1592 struct fuse_conn *fc = get_fuse_conn(inode);
1593 struct fuse_req *req;
1594 int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
1595 pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
1598 if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
1599 /* NLM needs asynchronous locks, which we don't support yet */
1603 /* Unlock on close is handled by the flush method */
1604 if (fl->fl_flags & FL_CLOSE)
1607 req = fuse_get_req_nopages(fc);
1609 return PTR_ERR(req);
1611 fuse_lk_fill(req, file, fl, opcode, pid, flock);
1612 fuse_request_send(fc, req);
1613 err = req->out.h.error;
1614 /* locking is restartable */
1617 fuse_put_request(fc, req);
1621 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
1623 struct inode *inode = file_inode(file);
1624 struct fuse_conn *fc = get_fuse_conn(inode);
1627 if (cmd == F_CANCELLK) {
1629 } else if (cmd == F_GETLK) {
1631 posix_test_lock(file, fl);
1634 err = fuse_getlk(file, fl);
1637 err = posix_lock_file(file, fl, NULL);
1639 err = fuse_setlk(file, fl, 0);
1644 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
1646 struct inode *inode = file_inode(file);
1647 struct fuse_conn *fc = get_fuse_conn(inode);
1651 err = flock_lock_file_wait(file, fl);
1653 struct fuse_file *ff = file->private_data;
1655 /* emulate flock with POSIX locks */
1656 fl->fl_owner = (fl_owner_t) file;
1658 err = fuse_setlk(file, fl, 1);
1664 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
1666 struct inode *inode = mapping->host;
1667 struct fuse_conn *fc = get_fuse_conn(inode);
1668 struct fuse_req *req;
1669 struct fuse_bmap_in inarg;
1670 struct fuse_bmap_out outarg;
1673 if (!inode->i_sb->s_bdev || fc->no_bmap)
1676 req = fuse_get_req_nopages(fc);
1680 memset(&inarg, 0, sizeof(inarg));
1681 inarg.block = block;
1682 inarg.blocksize = inode->i_sb->s_blocksize;
1683 req->in.h.opcode = FUSE_BMAP;
1684 req->in.h.nodeid = get_node_id(inode);
1685 req->in.numargs = 1;
1686 req->in.args[0].size = sizeof(inarg);
1687 req->in.args[0].value = &inarg;
1688 req->out.numargs = 1;
1689 req->out.args[0].size = sizeof(outarg);
1690 req->out.args[0].value = &outarg;
1691 fuse_request_send(fc, req);
1692 err = req->out.h.error;
1693 fuse_put_request(fc, req);
1697 return err ? 0 : outarg.block;
1700 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
1703 struct inode *inode = file_inode(file);
1705 /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
1706 if (whence == SEEK_CUR || whence == SEEK_SET)
1707 return generic_file_llseek(file, offset, whence);
1709 mutex_lock(&inode->i_mutex);
1710 retval = fuse_update_attributes(inode, NULL, file, NULL);
1712 retval = generic_file_llseek(file, offset, whence);
1713 mutex_unlock(&inode->i_mutex);
1718 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
1719 unsigned int nr_segs, size_t bytes, bool to_user)
1727 iov_iter_init(&ii, iov, nr_segs, bytes, 0);
1729 while (iov_iter_count(&ii)) {
1730 struct page *page = pages[page_idx++];
1731 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
1737 char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
1738 size_t iov_len = ii.iov->iov_len - ii.iov_offset;
1739 size_t copy = min(todo, iov_len);
1743 left = copy_from_user(kaddr, uaddr, copy);
1745 left = copy_to_user(uaddr, kaddr, copy);
1750 iov_iter_advance(&ii, copy);
1762 * CUSE servers compiled on 32bit broke on 64bit kernels because the
1763 * ABI was defined to be 'struct iovec' which is different on 32bit
1764 * and 64bit. Fortunately we can determine which structure the server
1765 * used from the size of the reply.
1767 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
1768 size_t transferred, unsigned count,
1771 #ifdef CONFIG_COMPAT
1772 if (count * sizeof(struct compat_iovec) == transferred) {
1773 struct compat_iovec *ciov = src;
1777 * With this interface a 32bit server cannot support
1778 * non-compat (i.e. ones coming from 64bit apps) ioctl
1784 for (i = 0; i < count; i++) {
1785 dst[i].iov_base = compat_ptr(ciov[i].iov_base);
1786 dst[i].iov_len = ciov[i].iov_len;
1792 if (count * sizeof(struct iovec) != transferred)
1795 memcpy(dst, src, transferred);
1799 /* Make sure iov_length() won't overflow */
1800 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
1803 u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
1805 for (n = 0; n < count; n++, iov++) {
1806 if (iov->iov_len > (size_t) max)
1808 max -= iov->iov_len;
1813 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
1814 void *src, size_t transferred, unsigned count,
1818 struct fuse_ioctl_iovec *fiov = src;
1820 if (fc->minor < 16) {
1821 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
1825 if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
1828 for (i = 0; i < count; i++) {
1829 /* Did the server supply an inappropriate value? */
1830 if (fiov[i].base != (unsigned long) fiov[i].base ||
1831 fiov[i].len != (unsigned long) fiov[i].len)
1834 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
1835 dst[i].iov_len = (size_t) fiov[i].len;
1837 #ifdef CONFIG_COMPAT
1839 (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
1840 (compat_size_t) dst[i].iov_len != fiov[i].len))
1850 * For ioctls, there is no generic way to determine how much memory
1851 * needs to be read and/or written. Furthermore, ioctls are allowed
1852 * to dereference the passed pointer, so the parameter requires deep
1853 * copying but FUSE has no idea whatsoever about what to copy in or
1856 * This is solved by allowing FUSE server to retry ioctl with
1857 * necessary in/out iovecs. Let's assume the ioctl implementation
1858 * needs to read in the following structure.
1865 * On the first callout to FUSE server, inarg->in_size and
1866 * inarg->out_size will be NULL; then, the server completes the ioctl
1867 * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
1868 * the actual iov array to
1870 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) } }
1872 * which tells FUSE to copy in the requested area and retry the ioctl.
1873 * On the second round, the server has access to the structure and
1874 * from that it can tell what to look for next, so on the invocation,
1875 * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
1877 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) },
1878 * { .iov_base = a.buf, .iov_len = a.buflen } }
1880 * FUSE will copy both struct a and the pointed buffer from the
1881 * process doing the ioctl and retry ioctl with both struct a and the
1884 * This time, FUSE server has everything it needs and completes ioctl
1885 * without FUSE_IOCTL_RETRY which finishes the ioctl call.
1887 * Copying data out works the same way.
1889 * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
1890 * automatically initializes in and out iovs by decoding @cmd with
1891 * _IOC_* macros and the server is not allowed to request RETRY. This
1892 * limits ioctl data transfers to well-formed ioctls and is the forced
1893 * behavior for all FUSE servers.
1895 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
1898 struct fuse_file *ff = file->private_data;
1899 struct fuse_conn *fc = ff->fc;
1900 struct fuse_ioctl_in inarg = {
1906 struct fuse_ioctl_out outarg;
1907 struct fuse_req *req = NULL;
1908 struct page **pages = NULL;
1909 struct iovec *iov_page = NULL;
1910 struct iovec *in_iov = NULL, *out_iov = NULL;
1911 unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
1912 size_t in_size, out_size, transferred;
1915 #if BITS_PER_LONG == 32
1916 inarg.flags |= FUSE_IOCTL_32BIT;
1918 if (flags & FUSE_IOCTL_COMPAT)
1919 inarg.flags |= FUSE_IOCTL_32BIT;
1922 /* assume all the iovs returned by client always fits in a page */
1923 BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
1926 pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
1927 iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
1928 if (!pages || !iov_page)
1932 * If restricted, initialize IO parameters as encoded in @cmd.
1933 * RETRY from server is not allowed.
1935 if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
1936 struct iovec *iov = iov_page;
1938 iov->iov_base = (void __user *)arg;
1939 iov->iov_len = _IOC_SIZE(cmd);
1941 if (_IOC_DIR(cmd) & _IOC_WRITE) {
1946 if (_IOC_DIR(cmd) & _IOC_READ) {
1953 inarg.in_size = in_size = iov_length(in_iov, in_iovs);
1954 inarg.out_size = out_size = iov_length(out_iov, out_iovs);
1957 * Out data can be used either for actual out data or iovs,
1958 * make sure there always is at least one page.
1960 out_size = max_t(size_t, out_size, PAGE_SIZE);
1961 max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
1963 /* make sure there are enough buffer pages and init request with them */
1965 if (max_pages > FUSE_MAX_PAGES_PER_REQ)
1967 while (num_pages < max_pages) {
1968 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
1969 if (!pages[num_pages])
1974 req = fuse_get_req(fc, num_pages);
1980 memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
1981 req->num_pages = num_pages;
1982 fuse_page_descs_length_init(req, 0, req->num_pages);
1984 /* okay, let's send it to the client */
1985 req->in.h.opcode = FUSE_IOCTL;
1986 req->in.h.nodeid = ff->nodeid;
1987 req->in.numargs = 1;
1988 req->in.args[0].size = sizeof(inarg);
1989 req->in.args[0].value = &inarg;
1992 req->in.args[1].size = in_size;
1993 req->in.argpages = 1;
1995 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
2001 req->out.numargs = 2;
2002 req->out.args[0].size = sizeof(outarg);
2003 req->out.args[0].value = &outarg;
2004 req->out.args[1].size = out_size;
2005 req->out.argpages = 1;
2006 req->out.argvar = 1;
2008 fuse_request_send(fc, req);
2009 err = req->out.h.error;
2010 transferred = req->out.args[1].size;
2011 fuse_put_request(fc, req);
2016 /* did it ask for retry? */
2017 if (outarg.flags & FUSE_IOCTL_RETRY) {
2020 /* no retry if in restricted mode */
2022 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2025 in_iovs = outarg.in_iovs;
2026 out_iovs = outarg.out_iovs;
2029 * Make sure things are in boundary, separate checks
2030 * are to protect against overflow.
2033 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2034 out_iovs > FUSE_IOCTL_MAX_IOV ||
2035 in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2038 vaddr = kmap_atomic(pages[0]);
2039 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2040 transferred, in_iovs + out_iovs,
2041 (flags & FUSE_IOCTL_COMPAT) != 0);
2042 kunmap_atomic(vaddr);
2047 out_iov = in_iov + in_iovs;
2049 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
2053 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
2061 if (transferred > inarg.out_size)
2064 err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
2067 fuse_put_request(fc, req);
2068 free_page((unsigned long) iov_page);
2070 __free_page(pages[--num_pages]);
2073 return err ? err : outarg.result;
2075 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2077 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2078 unsigned long arg, unsigned int flags)
2080 struct inode *inode = file_inode(file);
2081 struct fuse_conn *fc = get_fuse_conn(inode);
2083 if (!fuse_allow_current_process(fc))
2086 if (is_bad_inode(inode))
2089 return fuse_do_ioctl(file, cmd, arg, flags);
2092 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2095 return fuse_ioctl_common(file, cmd, arg, 0);
2098 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2101 return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2105 * All files which have been polled are linked to RB tree
2106 * fuse_conn->polled_files which is indexed by kh. Walk the tree and
2107 * find the matching one.
2109 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2110 struct rb_node **parent_out)
2112 struct rb_node **link = &fc->polled_files.rb_node;
2113 struct rb_node *last = NULL;
2116 struct fuse_file *ff;
2119 ff = rb_entry(last, struct fuse_file, polled_node);
2122 link = &last->rb_left;
2123 else if (kh > ff->kh)
2124 link = &last->rb_right;
2135 * The file is about to be polled. Make sure it's on the polled_files
2136 * RB tree. Note that files once added to the polled_files tree are
2137 * not removed before the file is released. This is because a file
2138 * polled once is likely to be polled again.
2140 static void fuse_register_polled_file(struct fuse_conn *fc,
2141 struct fuse_file *ff)
2143 spin_lock(&fc->lock);
2144 if (RB_EMPTY_NODE(&ff->polled_node)) {
2145 struct rb_node **link, *parent;
2147 link = fuse_find_polled_node(fc, ff->kh, &parent);
2149 rb_link_node(&ff->polled_node, parent, link);
2150 rb_insert_color(&ff->polled_node, &fc->polled_files);
2152 spin_unlock(&fc->lock);
2155 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2157 struct fuse_file *ff = file->private_data;
2158 struct fuse_conn *fc = ff->fc;
2159 struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2160 struct fuse_poll_out outarg;
2161 struct fuse_req *req;
2165 return DEFAULT_POLLMASK;
2167 poll_wait(file, &ff->poll_wait, wait);
2168 inarg.events = (__u32)poll_requested_events(wait);
2171 * Ask for notification iff there's someone waiting for it.
2172 * The client may ignore the flag and always notify.
2174 if (waitqueue_active(&ff->poll_wait)) {
2175 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2176 fuse_register_polled_file(fc, ff);
2179 req = fuse_get_req_nopages(fc);
2183 req->in.h.opcode = FUSE_POLL;
2184 req->in.h.nodeid = ff->nodeid;
2185 req->in.numargs = 1;
2186 req->in.args[0].size = sizeof(inarg);
2187 req->in.args[0].value = &inarg;
2188 req->out.numargs = 1;
2189 req->out.args[0].size = sizeof(outarg);
2190 req->out.args[0].value = &outarg;
2191 fuse_request_send(fc, req);
2192 err = req->out.h.error;
2193 fuse_put_request(fc, req);
2196 return outarg.revents;
2197 if (err == -ENOSYS) {
2199 return DEFAULT_POLLMASK;
2203 EXPORT_SYMBOL_GPL(fuse_file_poll);
2206 * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2207 * wakes up the poll waiters.
2209 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2210 struct fuse_notify_poll_wakeup_out *outarg)
2212 u64 kh = outarg->kh;
2213 struct rb_node **link;
2215 spin_lock(&fc->lock);
2217 link = fuse_find_polled_node(fc, kh, NULL);
2219 struct fuse_file *ff;
2221 ff = rb_entry(*link, struct fuse_file, polled_node);
2222 wake_up_interruptible_sync(&ff->poll_wait);
2225 spin_unlock(&fc->lock);
2230 fuse_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2231 loff_t offset, unsigned long nr_segs)
2234 struct file *file = NULL;
2237 file = iocb->ki_filp;
2241 ret = __fuse_direct_write(file, iov, nr_segs, &pos);
2243 ret = __fuse_direct_read(file, iov, nr_segs, &pos);
2248 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2251 struct fuse_file *ff = file->private_data;
2252 struct fuse_conn *fc = ff->fc;
2253 struct fuse_req *req;
2254 struct fuse_fallocate_in inarg = {
2262 if (fc->no_fallocate)
2265 req = fuse_get_req_nopages(fc);
2267 return PTR_ERR(req);
2269 req->in.h.opcode = FUSE_FALLOCATE;
2270 req->in.h.nodeid = ff->nodeid;
2271 req->in.numargs = 1;
2272 req->in.args[0].size = sizeof(inarg);
2273 req->in.args[0].value = &inarg;
2274 fuse_request_send(fc, req);
2275 err = req->out.h.error;
2276 if (err == -ENOSYS) {
2277 fc->no_fallocate = 1;
2280 fuse_put_request(fc, req);
2285 static const struct file_operations fuse_file_operations = {
2286 .llseek = fuse_file_llseek,
2287 .read = do_sync_read,
2288 .aio_read = fuse_file_aio_read,
2289 .write = do_sync_write,
2290 .aio_write = fuse_file_aio_write,
2291 .mmap = fuse_file_mmap,
2293 .flush = fuse_flush,
2294 .release = fuse_release,
2295 .fsync = fuse_fsync,
2296 .lock = fuse_file_lock,
2297 .flock = fuse_file_flock,
2298 .splice_read = generic_file_splice_read,
2299 .unlocked_ioctl = fuse_file_ioctl,
2300 .compat_ioctl = fuse_file_compat_ioctl,
2301 .poll = fuse_file_poll,
2302 .fallocate = fuse_file_fallocate,
2305 static const struct file_operations fuse_direct_io_file_operations = {
2306 .llseek = fuse_file_llseek,
2307 .read = fuse_direct_read,
2308 .write = fuse_direct_write,
2309 .mmap = fuse_direct_mmap,
2311 .flush = fuse_flush,
2312 .release = fuse_release,
2313 .fsync = fuse_fsync,
2314 .lock = fuse_file_lock,
2315 .flock = fuse_file_flock,
2316 .unlocked_ioctl = fuse_file_ioctl,
2317 .compat_ioctl = fuse_file_compat_ioctl,
2318 .poll = fuse_file_poll,
2319 .fallocate = fuse_file_fallocate,
2320 /* no splice_read */
2323 static const struct address_space_operations fuse_file_aops = {
2324 .readpage = fuse_readpage,
2325 .writepage = fuse_writepage,
2326 .launder_page = fuse_launder_page,
2327 .readpages = fuse_readpages,
2328 .set_page_dirty = __set_page_dirty_nobuffers,
2330 .direct_IO = fuse_direct_IO,
2333 void fuse_init_file_inode(struct inode *inode)
2335 inode->i_fop = &fuse_file_operations;
2336 inode->i_data.a_ops = &fuse_file_aops;