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 size_t fuse_send_read(struct fuse_req *req, struct file *file,
498 loff_t pos, size_t count, fl_owner_t owner)
500 struct fuse_file *ff = file->private_data;
501 struct fuse_conn *fc = ff->fc;
503 fuse_read_fill(req, file, pos, count, FUSE_READ);
505 struct fuse_read_in *inarg = &req->misc.read.in;
507 inarg->read_flags |= FUSE_READ_LOCKOWNER;
508 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
510 fuse_request_send(fc, req);
511 return req->out.args[0].size;
514 static void fuse_read_update_size(struct inode *inode, loff_t size,
517 struct fuse_conn *fc = get_fuse_conn(inode);
518 struct fuse_inode *fi = get_fuse_inode(inode);
520 spin_lock(&fc->lock);
521 if (attr_ver == fi->attr_version && size < inode->i_size) {
522 fi->attr_version = ++fc->attr_version;
523 i_size_write(inode, size);
525 spin_unlock(&fc->lock);
528 static int fuse_readpage(struct file *file, struct page *page)
530 struct inode *inode = page->mapping->host;
531 struct fuse_conn *fc = get_fuse_conn(inode);
532 struct fuse_req *req;
534 loff_t pos = page_offset(page);
535 size_t count = PAGE_CACHE_SIZE;
540 if (is_bad_inode(inode))
544 * Page writeback can extend beyond the lifetime of the
545 * page-cache page, so make sure we read a properly synced
548 fuse_wait_on_page_writeback(inode, page->index);
550 req = fuse_get_req(fc, 1);
555 attr_ver = fuse_get_attr_version(fc);
557 req->out.page_zeroing = 1;
558 req->out.argpages = 1;
560 req->pages[0] = page;
561 req->page_descs[0].length = count;
562 num_read = fuse_send_read(req, file, pos, count, NULL);
563 err = req->out.h.error;
564 fuse_put_request(fc, req);
568 * Short read means EOF. If file size is larger, truncate it
570 if (num_read < count)
571 fuse_read_update_size(inode, pos + num_read, attr_ver);
573 SetPageUptodate(page);
576 fuse_invalidate_attr(inode); /* atime changed */
582 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
585 size_t count = req->misc.read.in.size;
586 size_t num_read = req->out.args[0].size;
587 struct address_space *mapping = NULL;
589 for (i = 0; mapping == NULL && i < req->num_pages; i++)
590 mapping = req->pages[i]->mapping;
593 struct inode *inode = mapping->host;
596 * Short read means EOF. If file size is larger, truncate it
598 if (!req->out.h.error && num_read < count) {
601 pos = page_offset(req->pages[0]) + num_read;
602 fuse_read_update_size(inode, pos,
603 req->misc.read.attr_ver);
605 fuse_invalidate_attr(inode); /* atime changed */
608 for (i = 0; i < req->num_pages; i++) {
609 struct page *page = req->pages[i];
610 if (!req->out.h.error)
611 SetPageUptodate(page);
615 page_cache_release(page);
618 fuse_file_put(req->ff, false);
621 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
623 struct fuse_file *ff = file->private_data;
624 struct fuse_conn *fc = ff->fc;
625 loff_t pos = page_offset(req->pages[0]);
626 size_t count = req->num_pages << PAGE_CACHE_SHIFT;
628 req->out.argpages = 1;
629 req->out.page_zeroing = 1;
630 req->out.page_replace = 1;
631 fuse_read_fill(req, file, pos, count, FUSE_READ);
632 req->misc.read.attr_ver = fuse_get_attr_version(fc);
633 if (fc->async_read) {
634 req->ff = fuse_file_get(ff);
635 req->end = fuse_readpages_end;
636 fuse_request_send_background(fc, req);
638 fuse_request_send(fc, req);
639 fuse_readpages_end(fc, req);
640 fuse_put_request(fc, req);
644 struct fuse_fill_data {
645 struct fuse_req *req;
651 static int fuse_readpages_fill(void *_data, struct page *page)
653 struct fuse_fill_data *data = _data;
654 struct fuse_req *req = data->req;
655 struct inode *inode = data->inode;
656 struct fuse_conn *fc = get_fuse_conn(inode);
658 fuse_wait_on_page_writeback(inode, page->index);
660 if (req->num_pages &&
661 (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
662 (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
663 req->pages[req->num_pages - 1]->index + 1 != page->index)) {
664 int nr_alloc = min_t(unsigned, data->nr_pages,
665 FUSE_MAX_PAGES_PER_REQ);
666 fuse_send_readpages(req, data->file);
668 req = fuse_get_req_for_background(fc, nr_alloc);
670 req = fuse_get_req(fc, nr_alloc);
679 if (WARN_ON(req->num_pages >= req->max_pages)) {
680 fuse_put_request(fc, req);
684 page_cache_get(page);
685 req->pages[req->num_pages] = page;
686 req->page_descs[req->num_pages].length = PAGE_SIZE;
692 static int fuse_readpages(struct file *file, struct address_space *mapping,
693 struct list_head *pages, unsigned nr_pages)
695 struct inode *inode = mapping->host;
696 struct fuse_conn *fc = get_fuse_conn(inode);
697 struct fuse_fill_data data;
699 int nr_alloc = min_t(unsigned, nr_pages, FUSE_MAX_PAGES_PER_REQ);
702 if (is_bad_inode(inode))
708 data.req = fuse_get_req_for_background(fc, nr_alloc);
710 data.req = fuse_get_req(fc, nr_alloc);
711 data.nr_pages = nr_pages;
712 err = PTR_ERR(data.req);
713 if (IS_ERR(data.req))
716 err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
718 if (data.req->num_pages)
719 fuse_send_readpages(data.req, file);
721 fuse_put_request(fc, data.req);
727 static ssize_t fuse_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
728 unsigned long nr_segs, loff_t pos)
730 struct inode *inode = iocb->ki_filp->f_mapping->host;
731 struct fuse_conn *fc = get_fuse_conn(inode);
734 * In auto invalidate mode, always update attributes on read.
735 * Otherwise, only update if we attempt to read past EOF (to ensure
736 * i_size is up to date).
738 if (fc->auto_inval_data ||
739 (pos + iov_length(iov, nr_segs) > i_size_read(inode))) {
741 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
746 return generic_file_aio_read(iocb, iov, nr_segs, pos);
749 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
750 loff_t pos, size_t count)
752 struct fuse_write_in *inarg = &req->misc.write.in;
753 struct fuse_write_out *outarg = &req->misc.write.out;
758 req->in.h.opcode = FUSE_WRITE;
759 req->in.h.nodeid = ff->nodeid;
761 if (ff->fc->minor < 9)
762 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
764 req->in.args[0].size = sizeof(struct fuse_write_in);
765 req->in.args[0].value = inarg;
766 req->in.args[1].size = count;
767 req->out.numargs = 1;
768 req->out.args[0].size = sizeof(struct fuse_write_out);
769 req->out.args[0].value = outarg;
772 static size_t fuse_send_write(struct fuse_req *req, struct file *file,
773 loff_t pos, size_t count, fl_owner_t owner)
775 struct fuse_file *ff = file->private_data;
776 struct fuse_conn *fc = ff->fc;
777 struct fuse_write_in *inarg = &req->misc.write.in;
779 fuse_write_fill(req, ff, pos, count);
780 inarg->flags = file->f_flags;
782 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
783 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
785 fuse_request_send(fc, req);
786 return req->misc.write.out.size;
789 void fuse_write_update_size(struct inode *inode, loff_t pos)
791 struct fuse_conn *fc = get_fuse_conn(inode);
792 struct fuse_inode *fi = get_fuse_inode(inode);
794 spin_lock(&fc->lock);
795 fi->attr_version = ++fc->attr_version;
796 if (pos > inode->i_size)
797 i_size_write(inode, pos);
798 spin_unlock(&fc->lock);
801 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
802 struct inode *inode, loff_t pos,
809 for (i = 0; i < req->num_pages; i++)
810 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
812 res = fuse_send_write(req, file, pos, count, NULL);
814 offset = req->page_descs[0].offset;
816 for (i = 0; i < req->num_pages; i++) {
817 struct page *page = req->pages[i];
819 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
820 SetPageUptodate(page);
822 if (count > PAGE_CACHE_SIZE - offset)
823 count -= PAGE_CACHE_SIZE - offset;
829 page_cache_release(page);
835 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
836 struct address_space *mapping,
837 struct iov_iter *ii, loff_t pos)
839 struct fuse_conn *fc = get_fuse_conn(mapping->host);
840 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
844 req->in.argpages = 1;
845 req->page_descs[0].offset = offset;
850 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
851 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
854 bytes = min_t(size_t, bytes, fc->max_write - count);
858 if (iov_iter_fault_in_readable(ii, bytes))
862 page = grab_cache_page_write_begin(mapping, index, 0);
866 if (mapping_writably_mapped(mapping))
867 flush_dcache_page(page);
870 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
872 flush_dcache_page(page);
874 mark_page_accessed(page);
878 page_cache_release(page);
879 bytes = min(bytes, iov_iter_single_seg_count(ii));
884 req->pages[req->num_pages] = page;
885 req->page_descs[req->num_pages].length = tmp;
888 iov_iter_advance(ii, tmp);
892 if (offset == PAGE_CACHE_SIZE)
897 } while (iov_iter_count(ii) && count < fc->max_write &&
898 req->num_pages < req->max_pages && offset == 0);
900 return count > 0 ? count : err;
903 static inline unsigned fuse_wr_pages(loff_t pos, size_t len)
905 return min_t(unsigned,
906 ((pos + len - 1) >> PAGE_CACHE_SHIFT) -
907 (pos >> PAGE_CACHE_SHIFT) + 1,
908 FUSE_MAX_PAGES_PER_REQ);
911 static ssize_t fuse_perform_write(struct file *file,
912 struct address_space *mapping,
913 struct iov_iter *ii, loff_t pos)
915 struct inode *inode = mapping->host;
916 struct fuse_conn *fc = get_fuse_conn(inode);
920 if (is_bad_inode(inode))
924 struct fuse_req *req;
926 unsigned nr_pages = fuse_wr_pages(pos, iov_iter_count(ii));
928 req = fuse_get_req(fc, nr_pages);
934 count = fuse_fill_write_pages(req, mapping, ii, pos);
940 num_written = fuse_send_write_pages(req, file, inode,
942 err = req->out.h.error;
947 /* break out of the loop on short write */
948 if (num_written != count)
952 fuse_put_request(fc, req);
953 } while (!err && iov_iter_count(ii));
956 fuse_write_update_size(inode, pos);
958 fuse_invalidate_attr(inode);
960 return res > 0 ? res : err;
963 static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
964 unsigned long nr_segs, loff_t pos)
966 struct file *file = iocb->ki_filp;
967 struct address_space *mapping = file->f_mapping;
971 ssize_t written_buffered = 0;
972 struct inode *inode = mapping->host;
977 WARN_ON(iocb->ki_pos != pos);
980 err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
985 sb_start_write(inode->i_sb);
986 mutex_lock(&inode->i_mutex);
988 /* We can write back this queue in page reclaim */
989 current->backing_dev_info = mapping->backing_dev_info;
991 err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
998 err = file_remove_suid(file);
1002 err = file_update_time(file);
1006 if (file->f_flags & O_DIRECT) {
1007 written = generic_file_direct_write(iocb, iov, &nr_segs,
1010 if (written < 0 || written == count)
1016 iov_iter_init(&i, iov, nr_segs, count, written);
1017 written_buffered = fuse_perform_write(file, mapping, &i, pos);
1018 if (written_buffered < 0) {
1019 err = written_buffered;
1022 endbyte = pos + written_buffered - 1;
1024 err = filemap_write_and_wait_range(file->f_mapping, pos,
1029 invalidate_mapping_pages(file->f_mapping,
1030 pos >> PAGE_CACHE_SHIFT,
1031 endbyte >> PAGE_CACHE_SHIFT);
1033 written += written_buffered;
1034 iocb->ki_pos = pos + written_buffered;
1036 iov_iter_init(&i, iov, nr_segs, count, 0);
1037 written = fuse_perform_write(file, mapping, &i, pos);
1039 iocb->ki_pos = pos + written;
1042 current->backing_dev_info = NULL;
1043 mutex_unlock(&inode->i_mutex);
1044 sb_end_write(inode->i_sb);
1046 return written ? written : err;
1049 static void fuse_release_user_pages(struct fuse_req *req, int write)
1053 for (i = 0; i < req->num_pages; i++) {
1054 struct page *page = req->pages[i];
1056 set_page_dirty_lock(page);
1061 static inline void fuse_page_descs_length_init(struct fuse_req *req,
1062 unsigned index, unsigned nr_pages)
1066 for (i = index; i < index + nr_pages; i++)
1067 req->page_descs[i].length = PAGE_SIZE -
1068 req->page_descs[i].offset;
1071 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1073 return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1076 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1079 return min(iov_iter_single_seg_count(ii), max_size);
1082 static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1083 size_t *nbytesp, int write)
1085 size_t nbytes = 0; /* # bytes already packed in req */
1087 /* Special case for kernel I/O: can copy directly into the buffer */
1088 if (segment_eq(get_fs(), KERNEL_DS)) {
1089 unsigned long user_addr = fuse_get_user_addr(ii);
1090 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1093 req->in.args[1].value = (void *) user_addr;
1095 req->out.args[0].value = (void *) user_addr;
1097 iov_iter_advance(ii, frag_size);
1098 *nbytesp = frag_size;
1102 while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1104 unsigned long user_addr = fuse_get_user_addr(ii);
1105 unsigned offset = user_addr & ~PAGE_MASK;
1106 size_t frag_size = fuse_get_frag_size(ii, *nbytesp - nbytes);
1109 unsigned n = req->max_pages - req->num_pages;
1110 frag_size = min_t(size_t, frag_size, n << PAGE_SHIFT);
1112 npages = (frag_size + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1113 npages = clamp(npages, 1U, n);
1115 ret = get_user_pages_fast(user_addr, npages, !write,
1116 &req->pages[req->num_pages]);
1121 frag_size = min_t(size_t, frag_size,
1122 (npages << PAGE_SHIFT) - offset);
1123 iov_iter_advance(ii, frag_size);
1125 req->page_descs[req->num_pages].offset = offset;
1126 fuse_page_descs_length_init(req, req->num_pages, npages);
1128 req->num_pages += npages;
1129 req->page_descs[req->num_pages - 1].length -=
1130 (npages << PAGE_SHIFT) - offset - frag_size;
1132 nbytes += frag_size;
1136 req->in.argpages = 1;
1138 req->out.argpages = 1;
1145 static inline int fuse_iter_npages(const struct iov_iter *ii_p)
1147 struct iov_iter ii = *ii_p;
1150 while (iov_iter_count(&ii) && npages < FUSE_MAX_PAGES_PER_REQ) {
1151 unsigned long user_addr = fuse_get_user_addr(&ii);
1152 unsigned offset = user_addr & ~PAGE_MASK;
1153 size_t frag_size = iov_iter_single_seg_count(&ii);
1155 npages += (frag_size + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1156 iov_iter_advance(&ii, frag_size);
1159 return min(npages, FUSE_MAX_PAGES_PER_REQ);
1162 ssize_t fuse_direct_io(struct file *file, const struct iovec *iov,
1163 unsigned long nr_segs, size_t count, loff_t *ppos,
1166 struct fuse_file *ff = file->private_data;
1167 struct fuse_conn *fc = ff->fc;
1168 size_t nmax = write ? fc->max_write : fc->max_read;
1171 struct fuse_req *req;
1174 iov_iter_init(&ii, iov, nr_segs, count, 0);
1176 req = fuse_get_req(fc, fuse_iter_npages(&ii));
1178 return PTR_ERR(req);
1182 fl_owner_t owner = current->files;
1183 size_t nbytes = min(count, nmax);
1184 int err = fuse_get_user_pages(req, &ii, &nbytes, write);
1191 nres = fuse_send_write(req, file, pos, nbytes, owner);
1193 nres = fuse_send_read(req, file, pos, nbytes, owner);
1195 fuse_release_user_pages(req, !write);
1196 if (req->out.h.error) {
1198 res = req->out.h.error;
1200 } else if (nres > nbytes) {
1210 fuse_put_request(fc, req);
1211 req = fuse_get_req(fc, fuse_iter_npages(&ii));
1217 fuse_put_request(fc, req);
1223 EXPORT_SYMBOL_GPL(fuse_direct_io);
1225 static ssize_t __fuse_direct_read(struct file *file, const struct iovec *iov,
1226 unsigned long nr_segs, loff_t *ppos)
1229 struct inode *inode = file_inode(file);
1231 if (is_bad_inode(inode))
1234 res = fuse_direct_io(file, iov, nr_segs, iov_length(iov, nr_segs),
1237 fuse_invalidate_attr(inode);
1242 static ssize_t fuse_direct_read(struct file *file, char __user *buf,
1243 size_t count, loff_t *ppos)
1245 struct iovec iov = { .iov_base = buf, .iov_len = count };
1246 return __fuse_direct_read(file, &iov, 1, ppos);
1249 static ssize_t __fuse_direct_write(struct file *file, const struct iovec *iov,
1250 unsigned long nr_segs, loff_t *ppos)
1252 struct inode *inode = file_inode(file);
1253 size_t count = iov_length(iov, nr_segs);
1256 res = generic_write_checks(file, ppos, &count, 0);
1258 res = fuse_direct_io(file, iov, nr_segs, count, ppos, 1);
1260 fuse_write_update_size(inode, *ppos);
1263 fuse_invalidate_attr(inode);
1268 static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
1269 size_t count, loff_t *ppos)
1271 struct iovec iov = { .iov_base = (void __user *)buf, .iov_len = count };
1272 struct inode *inode = file_inode(file);
1275 if (is_bad_inode(inode))
1278 /* Don't allow parallel writes to the same file */
1279 mutex_lock(&inode->i_mutex);
1280 res = __fuse_direct_write(file, &iov, 1, ppos);
1281 mutex_unlock(&inode->i_mutex);
1286 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1288 __free_page(req->pages[0]);
1289 fuse_file_put(req->ff, false);
1292 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1294 struct inode *inode = req->inode;
1295 struct fuse_inode *fi = get_fuse_inode(inode);
1296 struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
1298 list_del(&req->writepages_entry);
1299 dec_bdi_stat(bdi, BDI_WRITEBACK);
1300 dec_zone_page_state(req->pages[0], NR_WRITEBACK_TEMP);
1301 bdi_writeout_inc(bdi);
1302 wake_up(&fi->page_waitq);
1305 /* Called under fc->lock, may release and reacquire it */
1306 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req)
1307 __releases(fc->lock)
1308 __acquires(fc->lock)
1310 struct fuse_inode *fi = get_fuse_inode(req->inode);
1311 loff_t size = i_size_read(req->inode);
1312 struct fuse_write_in *inarg = &req->misc.write.in;
1317 if (inarg->offset + PAGE_CACHE_SIZE <= size) {
1318 inarg->size = PAGE_CACHE_SIZE;
1319 } else if (inarg->offset < size) {
1320 inarg->size = size & (PAGE_CACHE_SIZE - 1);
1322 /* Got truncated off completely */
1326 req->in.args[1].size = inarg->size;
1328 fuse_request_send_background_locked(fc, req);
1332 fuse_writepage_finish(fc, req);
1333 spin_unlock(&fc->lock);
1334 fuse_writepage_free(fc, req);
1335 fuse_put_request(fc, req);
1336 spin_lock(&fc->lock);
1340 * If fi->writectr is positive (no truncate or fsync going on) send
1341 * all queued writepage requests.
1343 * Called with fc->lock
1345 void fuse_flush_writepages(struct inode *inode)
1346 __releases(fc->lock)
1347 __acquires(fc->lock)
1349 struct fuse_conn *fc = get_fuse_conn(inode);
1350 struct fuse_inode *fi = get_fuse_inode(inode);
1351 struct fuse_req *req;
1353 while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1354 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1355 list_del_init(&req->list);
1356 fuse_send_writepage(fc, req);
1360 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1362 struct inode *inode = req->inode;
1363 struct fuse_inode *fi = get_fuse_inode(inode);
1365 mapping_set_error(inode->i_mapping, req->out.h.error);
1366 spin_lock(&fc->lock);
1368 fuse_writepage_finish(fc, req);
1369 spin_unlock(&fc->lock);
1370 fuse_writepage_free(fc, req);
1373 static int fuse_writepage_locked(struct page *page)
1375 struct address_space *mapping = page->mapping;
1376 struct inode *inode = mapping->host;
1377 struct fuse_conn *fc = get_fuse_conn(inode);
1378 struct fuse_inode *fi = get_fuse_inode(inode);
1379 struct fuse_req *req;
1380 struct fuse_file *ff;
1381 struct page *tmp_page;
1383 set_page_writeback(page);
1385 req = fuse_request_alloc_nofs(1);
1389 req->background = 1; /* writeback always goes to bg_queue */
1390 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1394 spin_lock(&fc->lock);
1395 BUG_ON(list_empty(&fi->write_files));
1396 ff = list_entry(fi->write_files.next, struct fuse_file, write_entry);
1397 req->ff = fuse_file_get(ff);
1398 spin_unlock(&fc->lock);
1400 fuse_write_fill(req, ff, page_offset(page), 0);
1402 copy_highpage(tmp_page, page);
1403 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1404 req->in.argpages = 1;
1406 req->pages[0] = tmp_page;
1407 req->page_descs[0].offset = 0;
1408 req->page_descs[0].length = PAGE_SIZE;
1409 req->end = fuse_writepage_end;
1412 inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
1413 inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1414 end_page_writeback(page);
1416 spin_lock(&fc->lock);
1417 list_add(&req->writepages_entry, &fi->writepages);
1418 list_add_tail(&req->list, &fi->queued_writes);
1419 fuse_flush_writepages(inode);
1420 spin_unlock(&fc->lock);
1425 fuse_request_free(req);
1427 end_page_writeback(page);
1431 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1435 err = fuse_writepage_locked(page);
1441 static int fuse_launder_page(struct page *page)
1444 if (clear_page_dirty_for_io(page)) {
1445 struct inode *inode = page->mapping->host;
1446 err = fuse_writepage_locked(page);
1448 fuse_wait_on_page_writeback(inode, page->index);
1454 * Write back dirty pages now, because there may not be any suitable
1457 static void fuse_vma_close(struct vm_area_struct *vma)
1459 filemap_write_and_wait(vma->vm_file->f_mapping);
1463 * Wait for writeback against this page to complete before allowing it
1464 * to be marked dirty again, and hence written back again, possibly
1465 * before the previous writepage completed.
1467 * Block here, instead of in ->writepage(), so that the userspace fs
1468 * can only block processes actually operating on the filesystem.
1470 * Otherwise unprivileged userspace fs would be able to block
1475 * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
1477 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1479 struct page *page = vmf->page;
1481 * Don't use page->mapping as it may become NULL from a
1482 * concurrent truncate.
1484 struct inode *inode = vma->vm_file->f_mapping->host;
1486 fuse_wait_on_page_writeback(inode, page->index);
1490 static const struct vm_operations_struct fuse_file_vm_ops = {
1491 .close = fuse_vma_close,
1492 .fault = filemap_fault,
1493 .page_mkwrite = fuse_page_mkwrite,
1494 .remap_pages = generic_file_remap_pages,
1497 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
1499 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE)) {
1500 struct inode *inode = file_inode(file);
1501 struct fuse_conn *fc = get_fuse_conn(inode);
1502 struct fuse_inode *fi = get_fuse_inode(inode);
1503 struct fuse_file *ff = file->private_data;
1505 * file may be written through mmap, so chain it onto the
1506 * inodes's write_file list
1508 spin_lock(&fc->lock);
1509 if (list_empty(&ff->write_entry))
1510 list_add(&ff->write_entry, &fi->write_files);
1511 spin_unlock(&fc->lock);
1513 file_accessed(file);
1514 vma->vm_ops = &fuse_file_vm_ops;
1518 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
1520 /* Can't provide the coherency needed for MAP_SHARED */
1521 if (vma->vm_flags & VM_MAYSHARE)
1524 invalidate_inode_pages2(file->f_mapping);
1526 return generic_file_mmap(file, vma);
1529 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
1530 struct file_lock *fl)
1532 switch (ffl->type) {
1538 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
1539 ffl->end < ffl->start)
1542 fl->fl_start = ffl->start;
1543 fl->fl_end = ffl->end;
1544 fl->fl_pid = ffl->pid;
1550 fl->fl_type = ffl->type;
1554 static void fuse_lk_fill(struct fuse_req *req, struct file *file,
1555 const struct file_lock *fl, int opcode, pid_t pid,
1558 struct inode *inode = file_inode(file);
1559 struct fuse_conn *fc = get_fuse_conn(inode);
1560 struct fuse_file *ff = file->private_data;
1561 struct fuse_lk_in *arg = &req->misc.lk_in;
1564 arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
1565 arg->lk.start = fl->fl_start;
1566 arg->lk.end = fl->fl_end;
1567 arg->lk.type = fl->fl_type;
1570 arg->lk_flags |= FUSE_LK_FLOCK;
1571 req->in.h.opcode = opcode;
1572 req->in.h.nodeid = get_node_id(inode);
1573 req->in.numargs = 1;
1574 req->in.args[0].size = sizeof(*arg);
1575 req->in.args[0].value = arg;
1578 static int fuse_getlk(struct file *file, struct file_lock *fl)
1580 struct inode *inode = file_inode(file);
1581 struct fuse_conn *fc = get_fuse_conn(inode);
1582 struct fuse_req *req;
1583 struct fuse_lk_out outarg;
1586 req = fuse_get_req_nopages(fc);
1588 return PTR_ERR(req);
1590 fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
1591 req->out.numargs = 1;
1592 req->out.args[0].size = sizeof(outarg);
1593 req->out.args[0].value = &outarg;
1594 fuse_request_send(fc, req);
1595 err = req->out.h.error;
1596 fuse_put_request(fc, req);
1598 err = convert_fuse_file_lock(&outarg.lk, fl);
1603 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
1605 struct inode *inode = file_inode(file);
1606 struct fuse_conn *fc = get_fuse_conn(inode);
1607 struct fuse_req *req;
1608 int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
1609 pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
1612 if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
1613 /* NLM needs asynchronous locks, which we don't support yet */
1617 /* Unlock on close is handled by the flush method */
1618 if (fl->fl_flags & FL_CLOSE)
1621 req = fuse_get_req_nopages(fc);
1623 return PTR_ERR(req);
1625 fuse_lk_fill(req, file, fl, opcode, pid, flock);
1626 fuse_request_send(fc, req);
1627 err = req->out.h.error;
1628 /* locking is restartable */
1631 fuse_put_request(fc, req);
1635 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
1637 struct inode *inode = file_inode(file);
1638 struct fuse_conn *fc = get_fuse_conn(inode);
1641 if (cmd == F_CANCELLK) {
1643 } else if (cmd == F_GETLK) {
1645 posix_test_lock(file, fl);
1648 err = fuse_getlk(file, fl);
1651 err = posix_lock_file(file, fl, NULL);
1653 err = fuse_setlk(file, fl, 0);
1658 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
1660 struct inode *inode = file_inode(file);
1661 struct fuse_conn *fc = get_fuse_conn(inode);
1665 err = flock_lock_file_wait(file, fl);
1667 struct fuse_file *ff = file->private_data;
1669 /* emulate flock with POSIX locks */
1670 fl->fl_owner = (fl_owner_t) file;
1672 err = fuse_setlk(file, fl, 1);
1678 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
1680 struct inode *inode = mapping->host;
1681 struct fuse_conn *fc = get_fuse_conn(inode);
1682 struct fuse_req *req;
1683 struct fuse_bmap_in inarg;
1684 struct fuse_bmap_out outarg;
1687 if (!inode->i_sb->s_bdev || fc->no_bmap)
1690 req = fuse_get_req_nopages(fc);
1694 memset(&inarg, 0, sizeof(inarg));
1695 inarg.block = block;
1696 inarg.blocksize = inode->i_sb->s_blocksize;
1697 req->in.h.opcode = FUSE_BMAP;
1698 req->in.h.nodeid = get_node_id(inode);
1699 req->in.numargs = 1;
1700 req->in.args[0].size = sizeof(inarg);
1701 req->in.args[0].value = &inarg;
1702 req->out.numargs = 1;
1703 req->out.args[0].size = sizeof(outarg);
1704 req->out.args[0].value = &outarg;
1705 fuse_request_send(fc, req);
1706 err = req->out.h.error;
1707 fuse_put_request(fc, req);
1711 return err ? 0 : outarg.block;
1714 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
1717 struct inode *inode = file_inode(file);
1719 /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
1720 if (whence == SEEK_CUR || whence == SEEK_SET)
1721 return generic_file_llseek(file, offset, whence);
1723 mutex_lock(&inode->i_mutex);
1724 retval = fuse_update_attributes(inode, NULL, file, NULL);
1726 retval = generic_file_llseek(file, offset, whence);
1727 mutex_unlock(&inode->i_mutex);
1732 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
1733 unsigned int nr_segs, size_t bytes, bool to_user)
1741 iov_iter_init(&ii, iov, nr_segs, bytes, 0);
1743 while (iov_iter_count(&ii)) {
1744 struct page *page = pages[page_idx++];
1745 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
1751 char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
1752 size_t iov_len = ii.iov->iov_len - ii.iov_offset;
1753 size_t copy = min(todo, iov_len);
1757 left = copy_from_user(kaddr, uaddr, copy);
1759 left = copy_to_user(uaddr, kaddr, copy);
1764 iov_iter_advance(&ii, copy);
1776 * CUSE servers compiled on 32bit broke on 64bit kernels because the
1777 * ABI was defined to be 'struct iovec' which is different on 32bit
1778 * and 64bit. Fortunately we can determine which structure the server
1779 * used from the size of the reply.
1781 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
1782 size_t transferred, unsigned count,
1785 #ifdef CONFIG_COMPAT
1786 if (count * sizeof(struct compat_iovec) == transferred) {
1787 struct compat_iovec *ciov = src;
1791 * With this interface a 32bit server cannot support
1792 * non-compat (i.e. ones coming from 64bit apps) ioctl
1798 for (i = 0; i < count; i++) {
1799 dst[i].iov_base = compat_ptr(ciov[i].iov_base);
1800 dst[i].iov_len = ciov[i].iov_len;
1806 if (count * sizeof(struct iovec) != transferred)
1809 memcpy(dst, src, transferred);
1813 /* Make sure iov_length() won't overflow */
1814 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
1817 u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
1819 for (n = 0; n < count; n++, iov++) {
1820 if (iov->iov_len > (size_t) max)
1822 max -= iov->iov_len;
1827 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
1828 void *src, size_t transferred, unsigned count,
1832 struct fuse_ioctl_iovec *fiov = src;
1834 if (fc->minor < 16) {
1835 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
1839 if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
1842 for (i = 0; i < count; i++) {
1843 /* Did the server supply an inappropriate value? */
1844 if (fiov[i].base != (unsigned long) fiov[i].base ||
1845 fiov[i].len != (unsigned long) fiov[i].len)
1848 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
1849 dst[i].iov_len = (size_t) fiov[i].len;
1851 #ifdef CONFIG_COMPAT
1853 (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
1854 (compat_size_t) dst[i].iov_len != fiov[i].len))
1864 * For ioctls, there is no generic way to determine how much memory
1865 * needs to be read and/or written. Furthermore, ioctls are allowed
1866 * to dereference the passed pointer, so the parameter requires deep
1867 * copying but FUSE has no idea whatsoever about what to copy in or
1870 * This is solved by allowing FUSE server to retry ioctl with
1871 * necessary in/out iovecs. Let's assume the ioctl implementation
1872 * needs to read in the following structure.
1879 * On the first callout to FUSE server, inarg->in_size and
1880 * inarg->out_size will be NULL; then, the server completes the ioctl
1881 * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
1882 * the actual iov array to
1884 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) } }
1886 * which tells FUSE to copy in the requested area and retry the ioctl.
1887 * On the second round, the server has access to the structure and
1888 * from that it can tell what to look for next, so on the invocation,
1889 * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
1891 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) },
1892 * { .iov_base = a.buf, .iov_len = a.buflen } }
1894 * FUSE will copy both struct a and the pointed buffer from the
1895 * process doing the ioctl and retry ioctl with both struct a and the
1898 * This time, FUSE server has everything it needs and completes ioctl
1899 * without FUSE_IOCTL_RETRY which finishes the ioctl call.
1901 * Copying data out works the same way.
1903 * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
1904 * automatically initializes in and out iovs by decoding @cmd with
1905 * _IOC_* macros and the server is not allowed to request RETRY. This
1906 * limits ioctl data transfers to well-formed ioctls and is the forced
1907 * behavior for all FUSE servers.
1909 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
1912 struct fuse_file *ff = file->private_data;
1913 struct fuse_conn *fc = ff->fc;
1914 struct fuse_ioctl_in inarg = {
1920 struct fuse_ioctl_out outarg;
1921 struct fuse_req *req = NULL;
1922 struct page **pages = NULL;
1923 struct iovec *iov_page = NULL;
1924 struct iovec *in_iov = NULL, *out_iov = NULL;
1925 unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
1926 size_t in_size, out_size, transferred;
1929 #if BITS_PER_LONG == 32
1930 inarg.flags |= FUSE_IOCTL_32BIT;
1932 if (flags & FUSE_IOCTL_COMPAT)
1933 inarg.flags |= FUSE_IOCTL_32BIT;
1936 /* assume all the iovs returned by client always fits in a page */
1937 BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
1940 pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
1941 iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
1942 if (!pages || !iov_page)
1946 * If restricted, initialize IO parameters as encoded in @cmd.
1947 * RETRY from server is not allowed.
1949 if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
1950 struct iovec *iov = iov_page;
1952 iov->iov_base = (void __user *)arg;
1953 iov->iov_len = _IOC_SIZE(cmd);
1955 if (_IOC_DIR(cmd) & _IOC_WRITE) {
1960 if (_IOC_DIR(cmd) & _IOC_READ) {
1967 inarg.in_size = in_size = iov_length(in_iov, in_iovs);
1968 inarg.out_size = out_size = iov_length(out_iov, out_iovs);
1971 * Out data can be used either for actual out data or iovs,
1972 * make sure there always is at least one page.
1974 out_size = max_t(size_t, out_size, PAGE_SIZE);
1975 max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
1977 /* make sure there are enough buffer pages and init request with them */
1979 if (max_pages > FUSE_MAX_PAGES_PER_REQ)
1981 while (num_pages < max_pages) {
1982 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
1983 if (!pages[num_pages])
1988 req = fuse_get_req(fc, num_pages);
1994 memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
1995 req->num_pages = num_pages;
1996 fuse_page_descs_length_init(req, 0, req->num_pages);
1998 /* okay, let's send it to the client */
1999 req->in.h.opcode = FUSE_IOCTL;
2000 req->in.h.nodeid = ff->nodeid;
2001 req->in.numargs = 1;
2002 req->in.args[0].size = sizeof(inarg);
2003 req->in.args[0].value = &inarg;
2006 req->in.args[1].size = in_size;
2007 req->in.argpages = 1;
2009 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
2015 req->out.numargs = 2;
2016 req->out.args[0].size = sizeof(outarg);
2017 req->out.args[0].value = &outarg;
2018 req->out.args[1].size = out_size;
2019 req->out.argpages = 1;
2020 req->out.argvar = 1;
2022 fuse_request_send(fc, req);
2023 err = req->out.h.error;
2024 transferred = req->out.args[1].size;
2025 fuse_put_request(fc, req);
2030 /* did it ask for retry? */
2031 if (outarg.flags & FUSE_IOCTL_RETRY) {
2034 /* no retry if in restricted mode */
2036 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2039 in_iovs = outarg.in_iovs;
2040 out_iovs = outarg.out_iovs;
2043 * Make sure things are in boundary, separate checks
2044 * are to protect against overflow.
2047 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2048 out_iovs > FUSE_IOCTL_MAX_IOV ||
2049 in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2052 vaddr = kmap_atomic(pages[0]);
2053 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2054 transferred, in_iovs + out_iovs,
2055 (flags & FUSE_IOCTL_COMPAT) != 0);
2056 kunmap_atomic(vaddr);
2061 out_iov = in_iov + in_iovs;
2063 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
2067 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
2075 if (transferred > inarg.out_size)
2078 err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
2081 fuse_put_request(fc, req);
2082 free_page((unsigned long) iov_page);
2084 __free_page(pages[--num_pages]);
2087 return err ? err : outarg.result;
2089 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2091 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2092 unsigned long arg, unsigned int flags)
2094 struct inode *inode = file_inode(file);
2095 struct fuse_conn *fc = get_fuse_conn(inode);
2097 if (!fuse_allow_current_process(fc))
2100 if (is_bad_inode(inode))
2103 return fuse_do_ioctl(file, cmd, arg, flags);
2106 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2109 return fuse_ioctl_common(file, cmd, arg, 0);
2112 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2115 return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2119 * All files which have been polled are linked to RB tree
2120 * fuse_conn->polled_files which is indexed by kh. Walk the tree and
2121 * find the matching one.
2123 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2124 struct rb_node **parent_out)
2126 struct rb_node **link = &fc->polled_files.rb_node;
2127 struct rb_node *last = NULL;
2130 struct fuse_file *ff;
2133 ff = rb_entry(last, struct fuse_file, polled_node);
2136 link = &last->rb_left;
2137 else if (kh > ff->kh)
2138 link = &last->rb_right;
2149 * The file is about to be polled. Make sure it's on the polled_files
2150 * RB tree. Note that files once added to the polled_files tree are
2151 * not removed before the file is released. This is because a file
2152 * polled once is likely to be polled again.
2154 static void fuse_register_polled_file(struct fuse_conn *fc,
2155 struct fuse_file *ff)
2157 spin_lock(&fc->lock);
2158 if (RB_EMPTY_NODE(&ff->polled_node)) {
2159 struct rb_node **link, *parent;
2161 link = fuse_find_polled_node(fc, ff->kh, &parent);
2163 rb_link_node(&ff->polled_node, parent, link);
2164 rb_insert_color(&ff->polled_node, &fc->polled_files);
2166 spin_unlock(&fc->lock);
2169 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2171 struct fuse_file *ff = file->private_data;
2172 struct fuse_conn *fc = ff->fc;
2173 struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2174 struct fuse_poll_out outarg;
2175 struct fuse_req *req;
2179 return DEFAULT_POLLMASK;
2181 poll_wait(file, &ff->poll_wait, wait);
2182 inarg.events = (__u32)poll_requested_events(wait);
2185 * Ask for notification iff there's someone waiting for it.
2186 * The client may ignore the flag and always notify.
2188 if (waitqueue_active(&ff->poll_wait)) {
2189 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2190 fuse_register_polled_file(fc, ff);
2193 req = fuse_get_req_nopages(fc);
2197 req->in.h.opcode = FUSE_POLL;
2198 req->in.h.nodeid = ff->nodeid;
2199 req->in.numargs = 1;
2200 req->in.args[0].size = sizeof(inarg);
2201 req->in.args[0].value = &inarg;
2202 req->out.numargs = 1;
2203 req->out.args[0].size = sizeof(outarg);
2204 req->out.args[0].value = &outarg;
2205 fuse_request_send(fc, req);
2206 err = req->out.h.error;
2207 fuse_put_request(fc, req);
2210 return outarg.revents;
2211 if (err == -ENOSYS) {
2213 return DEFAULT_POLLMASK;
2217 EXPORT_SYMBOL_GPL(fuse_file_poll);
2220 * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2221 * wakes up the poll waiters.
2223 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2224 struct fuse_notify_poll_wakeup_out *outarg)
2226 u64 kh = outarg->kh;
2227 struct rb_node **link;
2229 spin_lock(&fc->lock);
2231 link = fuse_find_polled_node(fc, kh, NULL);
2233 struct fuse_file *ff;
2235 ff = rb_entry(*link, struct fuse_file, polled_node);
2236 wake_up_interruptible_sync(&ff->poll_wait);
2239 spin_unlock(&fc->lock);
2244 fuse_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2245 loff_t offset, unsigned long nr_segs)
2248 struct file *file = NULL;
2251 file = iocb->ki_filp;
2255 ret = __fuse_direct_write(file, iov, nr_segs, &pos);
2257 ret = __fuse_direct_read(file, iov, nr_segs, &pos);
2262 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2265 struct fuse_file *ff = file->private_data;
2266 struct fuse_conn *fc = ff->fc;
2267 struct fuse_req *req;
2268 struct fuse_fallocate_in inarg = {
2276 if (fc->no_fallocate)
2279 req = fuse_get_req_nopages(fc);
2281 return PTR_ERR(req);
2283 req->in.h.opcode = FUSE_FALLOCATE;
2284 req->in.h.nodeid = ff->nodeid;
2285 req->in.numargs = 1;
2286 req->in.args[0].size = sizeof(inarg);
2287 req->in.args[0].value = &inarg;
2288 fuse_request_send(fc, req);
2289 err = req->out.h.error;
2290 if (err == -ENOSYS) {
2291 fc->no_fallocate = 1;
2294 fuse_put_request(fc, req);
2299 static const struct file_operations fuse_file_operations = {
2300 .llseek = fuse_file_llseek,
2301 .read = do_sync_read,
2302 .aio_read = fuse_file_aio_read,
2303 .write = do_sync_write,
2304 .aio_write = fuse_file_aio_write,
2305 .mmap = fuse_file_mmap,
2307 .flush = fuse_flush,
2308 .release = fuse_release,
2309 .fsync = fuse_fsync,
2310 .lock = fuse_file_lock,
2311 .flock = fuse_file_flock,
2312 .splice_read = generic_file_splice_read,
2313 .unlocked_ioctl = fuse_file_ioctl,
2314 .compat_ioctl = fuse_file_compat_ioctl,
2315 .poll = fuse_file_poll,
2316 .fallocate = fuse_file_fallocate,
2319 static const struct file_operations fuse_direct_io_file_operations = {
2320 .llseek = fuse_file_llseek,
2321 .read = fuse_direct_read,
2322 .write = fuse_direct_write,
2323 .mmap = fuse_direct_mmap,
2325 .flush = fuse_flush,
2326 .release = fuse_release,
2327 .fsync = fuse_fsync,
2328 .lock = fuse_file_lock,
2329 .flock = fuse_file_flock,
2330 .unlocked_ioctl = fuse_file_ioctl,
2331 .compat_ioctl = fuse_file_compat_ioctl,
2332 .poll = fuse_file_poll,
2333 .fallocate = fuse_file_fallocate,
2334 /* no splice_read */
2337 static const struct address_space_operations fuse_file_aops = {
2338 .readpage = fuse_readpage,
2339 .writepage = fuse_writepage,
2340 .launder_page = fuse_launder_page,
2341 .readpages = fuse_readpages,
2342 .set_page_dirty = __set_page_dirty_nobuffers,
2344 .direct_IO = fuse_direct_IO,
2347 void fuse_init_file_inode(struct inode *inode)
2349 inode->i_fop = &fuse_file_operations;
2350 inode->i_data.a_ops = &fuse_file_aops;