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>
18 static const struct file_operations fuse_direct_io_file_operations;
20 static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
21 int opcode, struct fuse_open_out *outargp)
23 struct fuse_open_in inarg;
27 req = fuse_get_req(fc);
31 memset(&inarg, 0, sizeof(inarg));
32 inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
33 if (!fc->atomic_o_trunc)
34 inarg.flags &= ~O_TRUNC;
35 req->in.h.opcode = opcode;
36 req->in.h.nodeid = nodeid;
38 req->in.args[0].size = sizeof(inarg);
39 req->in.args[0].value = &inarg;
41 req->out.args[0].size = sizeof(*outargp);
42 req->out.args[0].value = outargp;
43 fuse_request_send(fc, req);
44 err = req->out.h.error;
45 fuse_put_request(fc, req);
50 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
54 ff = kmalloc(sizeof(struct fuse_file), GFP_KERNEL);
59 ff->reserved_req = fuse_request_alloc();
60 if (unlikely(!ff->reserved_req)) {
65 INIT_LIST_HEAD(&ff->write_entry);
66 atomic_set(&ff->count, 0);
67 RB_CLEAR_NODE(&ff->polled_node);
68 init_waitqueue_head(&ff->poll_wait);
72 spin_unlock(&fc->lock);
77 void fuse_file_free(struct fuse_file *ff)
79 fuse_request_free(ff->reserved_req);
83 struct fuse_file *fuse_file_get(struct fuse_file *ff)
85 atomic_inc(&ff->count);
89 static void fuse_release_async(struct work_struct *work)
95 req = container_of(work, struct fuse_req, misc.release.work);
96 path = req->misc.release.path;
97 fc = get_fuse_conn(path.dentry->d_inode);
99 fuse_put_request(fc, req);
103 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
105 if (fc->destroy_req) {
107 * If this is a fuseblk mount, then it's possible that
108 * releasing the path will result in releasing the
109 * super block and sending the DESTROY request. If
110 * the server is single threaded, this would hang.
111 * For this reason do the path_put() in a separate
114 atomic_inc(&req->count);
115 INIT_WORK(&req->misc.release.work, fuse_release_async);
116 schedule_work(&req->misc.release.work);
118 path_put(&req->misc.release.path);
122 static void fuse_file_put(struct fuse_file *ff, bool sync)
124 if (atomic_dec_and_test(&ff->count)) {
125 struct fuse_req *req = ff->reserved_req;
128 fuse_request_send(ff->fc, req);
129 path_put(&req->misc.release.path);
130 fuse_put_request(ff->fc, req);
132 req->end = fuse_release_end;
133 fuse_request_send_background(ff->fc, req);
139 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
142 struct fuse_open_out outarg;
143 struct fuse_file *ff;
145 int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
147 ff = fuse_file_alloc(fc);
151 err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
158 outarg.open_flags &= ~FOPEN_DIRECT_IO;
162 ff->open_flags = outarg.open_flags;
163 file->private_data = fuse_file_get(ff);
167 EXPORT_SYMBOL_GPL(fuse_do_open);
169 void fuse_finish_open(struct inode *inode, struct file *file)
171 struct fuse_file *ff = file->private_data;
172 struct fuse_conn *fc = get_fuse_conn(inode);
174 if (ff->open_flags & FOPEN_DIRECT_IO)
175 file->f_op = &fuse_direct_io_file_operations;
176 if (!(ff->open_flags & FOPEN_KEEP_CACHE))
177 invalidate_inode_pages2(inode->i_mapping);
178 if (ff->open_flags & FOPEN_NONSEEKABLE)
179 nonseekable_open(inode, file);
180 if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
181 struct fuse_inode *fi = get_fuse_inode(inode);
183 spin_lock(&fc->lock);
184 fi->attr_version = ++fc->attr_version;
185 i_size_write(inode, 0);
186 spin_unlock(&fc->lock);
187 fuse_invalidate_attr(inode);
191 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
193 struct fuse_conn *fc = get_fuse_conn(inode);
196 /* VFS checks this, but only _after_ ->open() */
197 if (file->f_flags & O_DIRECT)
200 err = generic_file_open(inode, file);
204 err = fuse_do_open(fc, get_node_id(inode), file, isdir);
208 fuse_finish_open(inode, file);
213 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
215 struct fuse_conn *fc = ff->fc;
216 struct fuse_req *req = ff->reserved_req;
217 struct fuse_release_in *inarg = &req->misc.release.in;
219 spin_lock(&fc->lock);
220 list_del(&ff->write_entry);
221 if (!RB_EMPTY_NODE(&ff->polled_node))
222 rb_erase(&ff->polled_node, &fc->polled_files);
223 spin_unlock(&fc->lock);
225 wake_up_interruptible_all(&ff->poll_wait);
228 inarg->flags = flags;
229 req->in.h.opcode = opcode;
230 req->in.h.nodeid = ff->nodeid;
232 req->in.args[0].size = sizeof(struct fuse_release_in);
233 req->in.args[0].value = inarg;
236 void fuse_release_common(struct file *file, int opcode)
238 struct fuse_file *ff;
239 struct fuse_req *req;
241 ff = file->private_data;
245 req = ff->reserved_req;
246 fuse_prepare_release(ff, file->f_flags, opcode);
249 struct fuse_release_in *inarg = &req->misc.release.in;
250 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
251 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
254 /* Hold vfsmount and dentry until release is finished */
255 path_get(&file->f_path);
256 req->misc.release.path = file->f_path;
259 * Normally this will send the RELEASE request, however if
260 * some asynchronous READ or WRITE requests are outstanding,
261 * the sending will be delayed.
263 * Make the release synchronous if this is a fuseblk mount,
264 * synchronous RELEASE is allowed (and desirable) in this case
265 * because the server can be trusted not to screw up.
267 fuse_file_put(ff, ff->fc->destroy_req != NULL);
270 static int fuse_open(struct inode *inode, struct file *file)
272 return fuse_open_common(inode, file, false);
275 static int fuse_release(struct inode *inode, struct file *file)
277 fuse_release_common(file, FUSE_RELEASE);
279 /* return value is ignored by VFS */
283 void fuse_sync_release(struct fuse_file *ff, int flags)
285 WARN_ON(atomic_read(&ff->count) > 1);
286 fuse_prepare_release(ff, flags, FUSE_RELEASE);
287 ff->reserved_req->force = 1;
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->f_path.dentry->d_inode;
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(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, int datasync, int isdir)
411 struct inode *inode = file->f_mapping->host;
412 struct fuse_conn *fc = get_fuse_conn(inode);
413 struct fuse_file *ff = file->private_data;
414 struct fuse_req *req;
415 struct fuse_fsync_in inarg;
418 if (is_bad_inode(inode))
421 if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
425 * Start writeback against all dirty pages of the inode, then
426 * wait for all outstanding writes, before sending the FSYNC
429 err = write_inode_now(inode, 0);
433 fuse_sync_writes(inode);
435 req = fuse_get_req(fc);
439 memset(&inarg, 0, sizeof(inarg));
441 inarg.fsync_flags = datasync ? 1 : 0;
442 req->in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
443 req->in.h.nodeid = get_node_id(inode);
445 req->in.args[0].size = sizeof(inarg);
446 req->in.args[0].value = &inarg;
447 fuse_request_send(fc, req);
448 err = req->out.h.error;
449 fuse_put_request(fc, req);
450 if (err == -ENOSYS) {
460 static int fuse_fsync(struct file *file, int datasync)
462 return fuse_fsync_common(file, datasync, 0);
465 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
466 size_t count, int opcode)
468 struct fuse_read_in *inarg = &req->misc.read.in;
469 struct fuse_file *ff = file->private_data;
474 inarg->flags = file->f_flags;
475 req->in.h.opcode = opcode;
476 req->in.h.nodeid = ff->nodeid;
478 req->in.args[0].size = sizeof(struct fuse_read_in);
479 req->in.args[0].value = inarg;
481 req->out.numargs = 1;
482 req->out.args[0].size = count;
485 static size_t fuse_send_read(struct fuse_req *req, struct file *file,
486 loff_t pos, size_t count, fl_owner_t owner)
488 struct fuse_file *ff = file->private_data;
489 struct fuse_conn *fc = ff->fc;
491 fuse_read_fill(req, file, pos, count, FUSE_READ);
493 struct fuse_read_in *inarg = &req->misc.read.in;
495 inarg->read_flags |= FUSE_READ_LOCKOWNER;
496 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
498 fuse_request_send(fc, req);
499 return req->out.args[0].size;
502 static void fuse_read_update_size(struct inode *inode, loff_t size,
505 struct fuse_conn *fc = get_fuse_conn(inode);
506 struct fuse_inode *fi = get_fuse_inode(inode);
508 spin_lock(&fc->lock);
509 if (attr_ver == fi->attr_version && size < inode->i_size) {
510 fi->attr_version = ++fc->attr_version;
511 i_size_write(inode, size);
513 spin_unlock(&fc->lock);
516 static int fuse_readpage(struct file *file, struct page *page)
518 struct inode *inode = page->mapping->host;
519 struct fuse_conn *fc = get_fuse_conn(inode);
520 struct fuse_req *req;
522 loff_t pos = page_offset(page);
523 size_t count = PAGE_CACHE_SIZE;
528 if (is_bad_inode(inode))
532 * Page writeback can extend beyond the lifetime of the
533 * page-cache page, so make sure we read a properly synced
536 fuse_wait_on_page_writeback(inode, page->index);
538 req = fuse_get_req(fc);
543 attr_ver = fuse_get_attr_version(fc);
545 req->out.page_zeroing = 1;
546 req->out.argpages = 1;
548 req->pages[0] = page;
549 num_read = fuse_send_read(req, file, pos, count, NULL);
550 err = req->out.h.error;
551 fuse_put_request(fc, req);
555 * Short read means EOF. If file size is larger, truncate it
557 if (num_read < count)
558 fuse_read_update_size(inode, pos + num_read, attr_ver);
560 SetPageUptodate(page);
563 fuse_invalidate_attr(inode); /* atime changed */
569 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
572 size_t count = req->misc.read.in.size;
573 size_t num_read = req->out.args[0].size;
574 struct address_space *mapping = NULL;
576 for (i = 0; mapping == NULL && i < req->num_pages; i++)
577 mapping = req->pages[i]->mapping;
580 struct inode *inode = mapping->host;
583 * Short read means EOF. If file size is larger, truncate it
585 if (!req->out.h.error && num_read < count) {
588 pos = page_offset(req->pages[0]) + num_read;
589 fuse_read_update_size(inode, pos,
590 req->misc.read.attr_ver);
592 fuse_invalidate_attr(inode); /* atime changed */
595 for (i = 0; i < req->num_pages; i++) {
596 struct page *page = req->pages[i];
597 if (!req->out.h.error)
598 SetPageUptodate(page);
602 page_cache_release(page);
605 fuse_file_put(req->ff, false);
608 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
610 struct fuse_file *ff = file->private_data;
611 struct fuse_conn *fc = ff->fc;
612 loff_t pos = page_offset(req->pages[0]);
613 size_t count = req->num_pages << PAGE_CACHE_SHIFT;
615 req->out.argpages = 1;
616 req->out.page_zeroing = 1;
617 req->out.page_replace = 1;
618 fuse_read_fill(req, file, pos, count, FUSE_READ);
619 req->misc.read.attr_ver = fuse_get_attr_version(fc);
620 if (fc->async_read) {
621 req->ff = fuse_file_get(ff);
622 req->end = fuse_readpages_end;
623 fuse_request_send_background(fc, req);
625 fuse_request_send(fc, req);
626 fuse_readpages_end(fc, req);
627 fuse_put_request(fc, req);
631 struct fuse_fill_data {
632 struct fuse_req *req;
637 static int fuse_readpages_fill(void *_data, struct page *page)
639 struct fuse_fill_data *data = _data;
640 struct fuse_req *req = data->req;
641 struct inode *inode = data->inode;
642 struct fuse_conn *fc = get_fuse_conn(inode);
644 fuse_wait_on_page_writeback(inode, page->index);
646 if (req->num_pages &&
647 (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
648 (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
649 req->pages[req->num_pages - 1]->index + 1 != page->index)) {
650 fuse_send_readpages(req, data->file);
651 data->req = req = fuse_get_req(fc);
657 page_cache_get(page);
658 req->pages[req->num_pages] = page;
663 static int fuse_readpages(struct file *file, struct address_space *mapping,
664 struct list_head *pages, unsigned nr_pages)
666 struct inode *inode = mapping->host;
667 struct fuse_conn *fc = get_fuse_conn(inode);
668 struct fuse_fill_data data;
672 if (is_bad_inode(inode))
677 data.req = fuse_get_req(fc);
678 err = PTR_ERR(data.req);
679 if (IS_ERR(data.req))
682 err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
684 if (data.req->num_pages)
685 fuse_send_readpages(data.req, file);
687 fuse_put_request(fc, data.req);
693 static ssize_t fuse_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
694 unsigned long nr_segs, loff_t pos)
696 struct inode *inode = iocb->ki_filp->f_mapping->host;
698 if (pos + iov_length(iov, nr_segs) > i_size_read(inode)) {
701 * If trying to read past EOF, make sure the i_size
702 * attribute is up-to-date.
704 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
709 return generic_file_aio_read(iocb, iov, nr_segs, pos);
712 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
713 loff_t pos, size_t count)
715 struct fuse_write_in *inarg = &req->misc.write.in;
716 struct fuse_write_out *outarg = &req->misc.write.out;
721 req->in.h.opcode = FUSE_WRITE;
722 req->in.h.nodeid = ff->nodeid;
724 if (ff->fc->minor < 9)
725 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
727 req->in.args[0].size = sizeof(struct fuse_write_in);
728 req->in.args[0].value = inarg;
729 req->in.args[1].size = count;
730 req->out.numargs = 1;
731 req->out.args[0].size = sizeof(struct fuse_write_out);
732 req->out.args[0].value = outarg;
735 static size_t fuse_send_write(struct fuse_req *req, struct file *file,
736 loff_t pos, size_t count, fl_owner_t owner)
738 struct fuse_file *ff = file->private_data;
739 struct fuse_conn *fc = ff->fc;
740 struct fuse_write_in *inarg = &req->misc.write.in;
742 fuse_write_fill(req, ff, pos, count);
743 inarg->flags = file->f_flags;
745 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
746 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
748 fuse_request_send(fc, req);
749 return req->misc.write.out.size;
752 static int fuse_write_begin(struct file *file, struct address_space *mapping,
753 loff_t pos, unsigned len, unsigned flags,
754 struct page **pagep, void **fsdata)
756 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
758 *pagep = grab_cache_page_write_begin(mapping, index, flags);
764 void fuse_write_update_size(struct inode *inode, loff_t pos)
766 struct fuse_conn *fc = get_fuse_conn(inode);
767 struct fuse_inode *fi = get_fuse_inode(inode);
769 spin_lock(&fc->lock);
770 fi->attr_version = ++fc->attr_version;
771 if (pos > inode->i_size)
772 i_size_write(inode, pos);
773 spin_unlock(&fc->lock);
776 static int fuse_buffered_write(struct file *file, struct inode *inode,
777 loff_t pos, unsigned count, struct page *page)
781 struct fuse_conn *fc = get_fuse_conn(inode);
782 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
783 struct fuse_req *req;
785 if (is_bad_inode(inode))
789 * Make sure writepages on the same page are not mixed up with
792 fuse_wait_on_page_writeback(inode, page->index);
794 req = fuse_get_req(fc);
798 req->in.argpages = 1;
800 req->pages[0] = page;
801 req->page_offset = offset;
802 nres = fuse_send_write(req, file, pos, count, NULL);
803 err = req->out.h.error;
804 fuse_put_request(fc, req);
809 fuse_write_update_size(inode, pos);
810 if (count == PAGE_CACHE_SIZE)
811 SetPageUptodate(page);
813 fuse_invalidate_attr(inode);
814 return err ? err : nres;
817 static int fuse_write_end(struct file *file, struct address_space *mapping,
818 loff_t pos, unsigned len, unsigned copied,
819 struct page *page, void *fsdata)
821 struct inode *inode = mapping->host;
825 res = fuse_buffered_write(file, inode, pos, copied, page);
828 page_cache_release(page);
832 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
833 struct inode *inode, loff_t pos,
840 for (i = 0; i < req->num_pages; i++)
841 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
843 res = fuse_send_write(req, file, pos, count, NULL);
845 offset = req->page_offset;
847 for (i = 0; i < req->num_pages; i++) {
848 struct page *page = req->pages[i];
850 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
851 SetPageUptodate(page);
853 if (count > PAGE_CACHE_SIZE - offset)
854 count -= PAGE_CACHE_SIZE - offset;
860 page_cache_release(page);
866 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
867 struct address_space *mapping,
868 struct iov_iter *ii, loff_t pos)
870 struct fuse_conn *fc = get_fuse_conn(mapping->host);
871 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
875 req->in.argpages = 1;
876 req->page_offset = offset;
881 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
882 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
885 bytes = min_t(size_t, bytes, fc->max_write - count);
889 if (iov_iter_fault_in_readable(ii, bytes))
893 page = grab_cache_page_write_begin(mapping, index, 0);
897 if (mapping_writably_mapped(mapping))
898 flush_dcache_page(page);
901 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
903 flush_dcache_page(page);
907 page_cache_release(page);
908 bytes = min(bytes, iov_iter_single_seg_count(ii));
913 req->pages[req->num_pages] = page;
916 iov_iter_advance(ii, tmp);
920 if (offset == PAGE_CACHE_SIZE)
925 } while (iov_iter_count(ii) && count < fc->max_write &&
926 req->num_pages < FUSE_MAX_PAGES_PER_REQ && offset == 0);
928 return count > 0 ? count : err;
931 static ssize_t fuse_perform_write(struct file *file,
932 struct address_space *mapping,
933 struct iov_iter *ii, loff_t pos)
935 struct inode *inode = mapping->host;
936 struct fuse_conn *fc = get_fuse_conn(inode);
940 if (is_bad_inode(inode))
944 struct fuse_req *req;
947 req = fuse_get_req(fc);
953 count = fuse_fill_write_pages(req, mapping, ii, pos);
959 num_written = fuse_send_write_pages(req, file, inode,
961 err = req->out.h.error;
966 /* break out of the loop on short write */
967 if (num_written != count)
971 fuse_put_request(fc, req);
972 } while (!err && iov_iter_count(ii));
975 fuse_write_update_size(inode, pos);
977 fuse_invalidate_attr(inode);
979 return res > 0 ? res : err;
982 static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
983 unsigned long nr_segs, loff_t pos)
985 struct file *file = iocb->ki_filp;
986 struct address_space *mapping = file->f_mapping;
989 struct inode *inode = mapping->host;
993 WARN_ON(iocb->ki_pos != pos);
995 err = generic_segment_checks(iov, &nr_segs, &count, VERIFY_READ);
999 mutex_lock(&inode->i_mutex);
1000 vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
1002 /* We can write back this queue in page reclaim */
1003 current->backing_dev_info = mapping->backing_dev_info;
1005 err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
1012 err = file_remove_suid(file);
1016 file_update_time(file);
1018 iov_iter_init(&i, iov, nr_segs, count, 0);
1019 written = fuse_perform_write(file, mapping, &i, pos);
1021 iocb->ki_pos = pos + written;
1024 current->backing_dev_info = NULL;
1025 mutex_unlock(&inode->i_mutex);
1027 return written ? written : err;
1030 static void fuse_release_user_pages(struct fuse_req *req, int write)
1034 for (i = 0; i < req->num_pages; i++) {
1035 struct page *page = req->pages[i];
1037 set_page_dirty_lock(page);
1042 static int fuse_get_user_pages(struct fuse_req *req, const char __user *buf,
1043 size_t *nbytesp, int write)
1045 size_t nbytes = *nbytesp;
1046 unsigned long user_addr = (unsigned long) buf;
1047 unsigned offset = user_addr & ~PAGE_MASK;
1050 /* Special case for kernel I/O: can copy directly into the buffer */
1051 if (segment_eq(get_fs(), KERNEL_DS)) {
1053 req->in.args[1].value = (void *) user_addr;
1055 req->out.args[0].value = (void *) user_addr;
1060 nbytes = min_t(size_t, nbytes, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
1061 npages = (nbytes + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1062 npages = clamp(npages, 1, FUSE_MAX_PAGES_PER_REQ);
1063 npages = get_user_pages_fast(user_addr, npages, !write, req->pages);
1067 req->num_pages = npages;
1068 req->page_offset = offset;
1071 req->in.argpages = 1;
1073 req->out.argpages = 1;
1075 nbytes = (req->num_pages << PAGE_SHIFT) - req->page_offset;
1076 *nbytesp = min(*nbytesp, nbytes);
1081 ssize_t fuse_direct_io(struct file *file, const char __user *buf,
1082 size_t count, loff_t *ppos, int write)
1084 struct fuse_file *ff = file->private_data;
1085 struct fuse_conn *fc = ff->fc;
1086 size_t nmax = write ? fc->max_write : fc->max_read;
1089 struct fuse_req *req;
1091 req = fuse_get_req(fc);
1093 return PTR_ERR(req);
1097 fl_owner_t owner = current->files;
1098 size_t nbytes = min(count, nmax);
1099 int err = fuse_get_user_pages(req, buf, &nbytes, write);
1106 nres = fuse_send_write(req, file, pos, nbytes, owner);
1108 nres = fuse_send_read(req, file, pos, nbytes, owner);
1110 fuse_release_user_pages(req, !write);
1111 if (req->out.h.error) {
1113 res = req->out.h.error;
1115 } else if (nres > nbytes) {
1126 fuse_put_request(fc, req);
1127 req = fuse_get_req(fc);
1133 fuse_put_request(fc, req);
1139 EXPORT_SYMBOL_GPL(fuse_direct_io);
1141 static ssize_t fuse_direct_read(struct file *file, char __user *buf,
1142 size_t count, loff_t *ppos)
1145 struct inode *inode = file->f_path.dentry->d_inode;
1147 if (is_bad_inode(inode))
1150 res = fuse_direct_io(file, buf, count, ppos, 0);
1152 fuse_invalidate_attr(inode);
1157 static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
1158 size_t count, loff_t *ppos)
1160 struct inode *inode = file->f_path.dentry->d_inode;
1163 if (is_bad_inode(inode))
1166 /* Don't allow parallel writes to the same file */
1167 mutex_lock(&inode->i_mutex);
1168 res = generic_write_checks(file, ppos, &count, 0);
1170 res = fuse_direct_io(file, buf, count, ppos, 1);
1172 fuse_write_update_size(inode, *ppos);
1174 mutex_unlock(&inode->i_mutex);
1176 fuse_invalidate_attr(inode);
1181 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1183 __free_page(req->pages[0]);
1184 fuse_file_put(req->ff, false);
1187 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1189 struct inode *inode = req->inode;
1190 struct fuse_inode *fi = get_fuse_inode(inode);
1191 struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
1193 list_del(&req->writepages_entry);
1194 dec_bdi_stat(bdi, BDI_WRITEBACK);
1195 dec_zone_page_state(req->pages[0], NR_WRITEBACK_TEMP);
1196 bdi_writeout_inc(bdi);
1197 wake_up(&fi->page_waitq);
1200 /* Called under fc->lock, may release and reacquire it */
1201 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req)
1202 __releases(fc->lock)
1203 __acquires(fc->lock)
1205 struct fuse_inode *fi = get_fuse_inode(req->inode);
1206 loff_t size = i_size_read(req->inode);
1207 struct fuse_write_in *inarg = &req->misc.write.in;
1212 if (inarg->offset + PAGE_CACHE_SIZE <= size) {
1213 inarg->size = PAGE_CACHE_SIZE;
1214 } else if (inarg->offset < size) {
1215 inarg->size = size & (PAGE_CACHE_SIZE - 1);
1217 /* Got truncated off completely */
1221 req->in.args[1].size = inarg->size;
1223 fuse_request_send_background_locked(fc, req);
1227 fuse_writepage_finish(fc, req);
1228 spin_unlock(&fc->lock);
1229 fuse_writepage_free(fc, req);
1230 fuse_put_request(fc, req);
1231 spin_lock(&fc->lock);
1235 * If fi->writectr is positive (no truncate or fsync going on) send
1236 * all queued writepage requests.
1238 * Called with fc->lock
1240 void fuse_flush_writepages(struct inode *inode)
1241 __releases(fc->lock)
1242 __acquires(fc->lock)
1244 struct fuse_conn *fc = get_fuse_conn(inode);
1245 struct fuse_inode *fi = get_fuse_inode(inode);
1246 struct fuse_req *req;
1248 while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1249 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1250 list_del_init(&req->list);
1251 fuse_send_writepage(fc, req);
1255 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1257 struct inode *inode = req->inode;
1258 struct fuse_inode *fi = get_fuse_inode(inode);
1260 mapping_set_error(inode->i_mapping, req->out.h.error);
1261 spin_lock(&fc->lock);
1263 fuse_writepage_finish(fc, req);
1264 spin_unlock(&fc->lock);
1265 fuse_writepage_free(fc, req);
1268 static int fuse_writepage_locked(struct page *page)
1270 struct address_space *mapping = page->mapping;
1271 struct inode *inode = mapping->host;
1272 struct fuse_conn *fc = get_fuse_conn(inode);
1273 struct fuse_inode *fi = get_fuse_inode(inode);
1274 struct fuse_req *req;
1275 struct fuse_file *ff;
1276 struct page *tmp_page;
1278 set_page_writeback(page);
1280 req = fuse_request_alloc_nofs();
1284 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1288 spin_lock(&fc->lock);
1289 BUG_ON(list_empty(&fi->write_files));
1290 ff = list_entry(fi->write_files.next, struct fuse_file, write_entry);
1291 req->ff = fuse_file_get(ff);
1292 spin_unlock(&fc->lock);
1294 fuse_write_fill(req, ff, page_offset(page), 0);
1296 copy_highpage(tmp_page, page);
1297 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1298 req->in.argpages = 1;
1300 req->pages[0] = tmp_page;
1301 req->page_offset = 0;
1302 req->end = fuse_writepage_end;
1305 inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
1306 inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1307 end_page_writeback(page);
1309 spin_lock(&fc->lock);
1310 list_add(&req->writepages_entry, &fi->writepages);
1311 list_add_tail(&req->list, &fi->queued_writes);
1312 fuse_flush_writepages(inode);
1313 spin_unlock(&fc->lock);
1318 fuse_request_free(req);
1320 end_page_writeback(page);
1324 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1328 err = fuse_writepage_locked(page);
1334 static int fuse_launder_page(struct page *page)
1337 if (clear_page_dirty_for_io(page)) {
1338 struct inode *inode = page->mapping->host;
1339 err = fuse_writepage_locked(page);
1341 fuse_wait_on_page_writeback(inode, page->index);
1347 * Write back dirty pages now, because there may not be any suitable
1350 static void fuse_vma_close(struct vm_area_struct *vma)
1352 filemap_write_and_wait(vma->vm_file->f_mapping);
1356 * Wait for writeback against this page to complete before allowing it
1357 * to be marked dirty again, and hence written back again, possibly
1358 * before the previous writepage completed.
1360 * Block here, instead of in ->writepage(), so that the userspace fs
1361 * can only block processes actually operating on the filesystem.
1363 * Otherwise unprivileged userspace fs would be able to block
1368 * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
1370 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1372 struct page *page = vmf->page;
1374 * Don't use page->mapping as it may become NULL from a
1375 * concurrent truncate.
1377 struct inode *inode = vma->vm_file->f_mapping->host;
1379 fuse_wait_on_page_writeback(inode, page->index);
1383 static const struct vm_operations_struct fuse_file_vm_ops = {
1384 .close = fuse_vma_close,
1385 .fault = filemap_fault,
1386 .page_mkwrite = fuse_page_mkwrite,
1389 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
1391 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE)) {
1392 struct inode *inode = file->f_dentry->d_inode;
1393 struct fuse_conn *fc = get_fuse_conn(inode);
1394 struct fuse_inode *fi = get_fuse_inode(inode);
1395 struct fuse_file *ff = file->private_data;
1397 * file may be written through mmap, so chain it onto the
1398 * inodes's write_file list
1400 spin_lock(&fc->lock);
1401 if (list_empty(&ff->write_entry))
1402 list_add(&ff->write_entry, &fi->write_files);
1403 spin_unlock(&fc->lock);
1405 file_accessed(file);
1406 vma->vm_ops = &fuse_file_vm_ops;
1410 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
1412 /* Can't provide the coherency needed for MAP_SHARED */
1413 if (vma->vm_flags & VM_MAYSHARE)
1416 invalidate_inode_pages2(file->f_mapping);
1418 return generic_file_mmap(file, vma);
1421 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
1422 struct file_lock *fl)
1424 switch (ffl->type) {
1430 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
1431 ffl->end < ffl->start)
1434 fl->fl_start = ffl->start;
1435 fl->fl_end = ffl->end;
1436 fl->fl_pid = ffl->pid;
1442 fl->fl_type = ffl->type;
1446 static void fuse_lk_fill(struct fuse_req *req, struct file *file,
1447 const struct file_lock *fl, int opcode, pid_t pid,
1450 struct inode *inode = file->f_path.dentry->d_inode;
1451 struct fuse_conn *fc = get_fuse_conn(inode);
1452 struct fuse_file *ff = file->private_data;
1453 struct fuse_lk_in *arg = &req->misc.lk_in;
1456 arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
1457 arg->lk.start = fl->fl_start;
1458 arg->lk.end = fl->fl_end;
1459 arg->lk.type = fl->fl_type;
1462 arg->lk_flags |= FUSE_LK_FLOCK;
1463 req->in.h.opcode = opcode;
1464 req->in.h.nodeid = get_node_id(inode);
1465 req->in.numargs = 1;
1466 req->in.args[0].size = sizeof(*arg);
1467 req->in.args[0].value = arg;
1470 static int fuse_getlk(struct file *file, struct file_lock *fl)
1472 struct inode *inode = file->f_path.dentry->d_inode;
1473 struct fuse_conn *fc = get_fuse_conn(inode);
1474 struct fuse_req *req;
1475 struct fuse_lk_out outarg;
1478 req = fuse_get_req(fc);
1480 return PTR_ERR(req);
1482 fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
1483 req->out.numargs = 1;
1484 req->out.args[0].size = sizeof(outarg);
1485 req->out.args[0].value = &outarg;
1486 fuse_request_send(fc, req);
1487 err = req->out.h.error;
1488 fuse_put_request(fc, req);
1490 err = convert_fuse_file_lock(&outarg.lk, fl);
1495 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
1497 struct inode *inode = file->f_path.dentry->d_inode;
1498 struct fuse_conn *fc = get_fuse_conn(inode);
1499 struct fuse_req *req;
1500 int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
1501 pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
1504 if (fl->fl_lmops && fl->fl_lmops->fl_grant) {
1505 /* NLM needs asynchronous locks, which we don't support yet */
1509 /* Unlock on close is handled by the flush method */
1510 if (fl->fl_flags & FL_CLOSE)
1513 req = fuse_get_req(fc);
1515 return PTR_ERR(req);
1517 fuse_lk_fill(req, file, fl, opcode, pid, flock);
1518 fuse_request_send(fc, req);
1519 err = req->out.h.error;
1520 /* locking is restartable */
1523 fuse_put_request(fc, req);
1527 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
1529 struct inode *inode = file->f_path.dentry->d_inode;
1530 struct fuse_conn *fc = get_fuse_conn(inode);
1533 if (cmd == F_CANCELLK) {
1535 } else if (cmd == F_GETLK) {
1537 posix_test_lock(file, fl);
1540 err = fuse_getlk(file, fl);
1543 err = posix_lock_file(file, fl, NULL);
1545 err = fuse_setlk(file, fl, 0);
1550 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
1552 struct inode *inode = file->f_path.dentry->d_inode;
1553 struct fuse_conn *fc = get_fuse_conn(inode);
1557 err = flock_lock_file_wait(file, fl);
1559 struct fuse_file *ff = file->private_data;
1561 /* emulate flock with POSIX locks */
1562 fl->fl_owner = (fl_owner_t) file;
1564 err = fuse_setlk(file, fl, 1);
1570 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
1572 struct inode *inode = mapping->host;
1573 struct fuse_conn *fc = get_fuse_conn(inode);
1574 struct fuse_req *req;
1575 struct fuse_bmap_in inarg;
1576 struct fuse_bmap_out outarg;
1579 if (!inode->i_sb->s_bdev || fc->no_bmap)
1582 req = fuse_get_req(fc);
1586 memset(&inarg, 0, sizeof(inarg));
1587 inarg.block = block;
1588 inarg.blocksize = inode->i_sb->s_blocksize;
1589 req->in.h.opcode = FUSE_BMAP;
1590 req->in.h.nodeid = get_node_id(inode);
1591 req->in.numargs = 1;
1592 req->in.args[0].size = sizeof(inarg);
1593 req->in.args[0].value = &inarg;
1594 req->out.numargs = 1;
1595 req->out.args[0].size = sizeof(outarg);
1596 req->out.args[0].value = &outarg;
1597 fuse_request_send(fc, req);
1598 err = req->out.h.error;
1599 fuse_put_request(fc, req);
1603 return err ? 0 : outarg.block;
1606 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int origin)
1609 struct inode *inode = file->f_path.dentry->d_inode;
1611 mutex_lock(&inode->i_mutex);
1614 retval = fuse_update_attributes(inode, NULL, file, NULL);
1617 offset += i_size_read(inode);
1620 offset += file->f_pos;
1623 if (offset >= 0 && offset <= inode->i_sb->s_maxbytes) {
1624 if (offset != file->f_pos) {
1625 file->f_pos = offset;
1626 file->f_version = 0;
1631 mutex_unlock(&inode->i_mutex);
1635 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
1636 unsigned int nr_segs, size_t bytes, bool to_user)
1644 iov_iter_init(&ii, iov, nr_segs, bytes, 0);
1646 while (iov_iter_count(&ii)) {
1647 struct page *page = pages[page_idx++];
1648 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
1654 char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
1655 size_t iov_len = ii.iov->iov_len - ii.iov_offset;
1656 size_t copy = min(todo, iov_len);
1660 left = copy_from_user(kaddr, uaddr, copy);
1662 left = copy_to_user(uaddr, kaddr, copy);
1667 iov_iter_advance(&ii, copy);
1679 * CUSE servers compiled on 32bit broke on 64bit kernels because the
1680 * ABI was defined to be 'struct iovec' which is different on 32bit
1681 * and 64bit. Fortunately we can determine which structure the server
1682 * used from the size of the reply.
1684 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
1685 size_t transferred, unsigned count,
1688 #ifdef CONFIG_COMPAT
1689 if (count * sizeof(struct compat_iovec) == transferred) {
1690 struct compat_iovec *ciov = src;
1694 * With this interface a 32bit server cannot support
1695 * non-compat (i.e. ones coming from 64bit apps) ioctl
1701 for (i = 0; i < count; i++) {
1702 dst[i].iov_base = compat_ptr(ciov[i].iov_base);
1703 dst[i].iov_len = ciov[i].iov_len;
1709 if (count * sizeof(struct iovec) != transferred)
1712 memcpy(dst, src, transferred);
1716 /* Make sure iov_length() won't overflow */
1717 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
1720 u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
1722 for (n = 0; n < count; n++) {
1723 if (iov->iov_len > (size_t) max)
1725 max -= iov->iov_len;
1730 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
1731 void *src, size_t transferred, unsigned count,
1735 struct fuse_ioctl_iovec *fiov = src;
1737 if (fc->minor < 16) {
1738 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
1742 if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
1745 for (i = 0; i < count; i++) {
1746 /* Did the server supply an inappropriate value? */
1747 if (fiov[i].base != (unsigned long) fiov[i].base ||
1748 fiov[i].len != (unsigned long) fiov[i].len)
1751 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
1752 dst[i].iov_len = (size_t) fiov[i].len;
1754 #ifdef CONFIG_COMPAT
1756 (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
1757 (compat_size_t) dst[i].iov_len != fiov[i].len))
1767 * For ioctls, there is no generic way to determine how much memory
1768 * needs to be read and/or written. Furthermore, ioctls are allowed
1769 * to dereference the passed pointer, so the parameter requires deep
1770 * copying but FUSE has no idea whatsoever about what to copy in or
1773 * This is solved by allowing FUSE server to retry ioctl with
1774 * necessary in/out iovecs. Let's assume the ioctl implementation
1775 * needs to read in the following structure.
1782 * On the first callout to FUSE server, inarg->in_size and
1783 * inarg->out_size will be NULL; then, the server completes the ioctl
1784 * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
1785 * the actual iov array to
1787 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) } }
1789 * which tells FUSE to copy in the requested area and retry the ioctl.
1790 * On the second round, the server has access to the structure and
1791 * from that it can tell what to look for next, so on the invocation,
1792 * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
1794 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) },
1795 * { .iov_base = a.buf, .iov_len = a.buflen } }
1797 * FUSE will copy both struct a and the pointed buffer from the
1798 * process doing the ioctl and retry ioctl with both struct a and the
1801 * This time, FUSE server has everything it needs and completes ioctl
1802 * without FUSE_IOCTL_RETRY which finishes the ioctl call.
1804 * Copying data out works the same way.
1806 * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
1807 * automatically initializes in and out iovs by decoding @cmd with
1808 * _IOC_* macros and the server is not allowed to request RETRY. This
1809 * limits ioctl data transfers to well-formed ioctls and is the forced
1810 * behavior for all FUSE servers.
1812 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
1815 struct fuse_file *ff = file->private_data;
1816 struct fuse_conn *fc = ff->fc;
1817 struct fuse_ioctl_in inarg = {
1823 struct fuse_ioctl_out outarg;
1824 struct fuse_req *req = NULL;
1825 struct page **pages = NULL;
1826 struct iovec *iov_page = NULL;
1827 struct iovec *in_iov = NULL, *out_iov = NULL;
1828 unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
1829 size_t in_size, out_size, transferred;
1832 #if BITS_PER_LONG == 32
1833 inarg.flags |= FUSE_IOCTL_32BIT;
1835 if (flags & FUSE_IOCTL_COMPAT)
1836 inarg.flags |= FUSE_IOCTL_32BIT;
1839 /* assume all the iovs returned by client always fits in a page */
1840 BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
1843 pages = kzalloc(sizeof(pages[0]) * FUSE_MAX_PAGES_PER_REQ, GFP_KERNEL);
1844 iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
1845 if (!pages || !iov_page)
1849 * If restricted, initialize IO parameters as encoded in @cmd.
1850 * RETRY from server is not allowed.
1852 if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
1853 struct iovec *iov = iov_page;
1855 iov->iov_base = (void __user *)arg;
1856 iov->iov_len = _IOC_SIZE(cmd);
1858 if (_IOC_DIR(cmd) & _IOC_WRITE) {
1863 if (_IOC_DIR(cmd) & _IOC_READ) {
1870 inarg.in_size = in_size = iov_length(in_iov, in_iovs);
1871 inarg.out_size = out_size = iov_length(out_iov, out_iovs);
1874 * Out data can be used either for actual out data or iovs,
1875 * make sure there always is at least one page.
1877 out_size = max_t(size_t, out_size, PAGE_SIZE);
1878 max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
1880 /* make sure there are enough buffer pages and init request with them */
1882 if (max_pages > FUSE_MAX_PAGES_PER_REQ)
1884 while (num_pages < max_pages) {
1885 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
1886 if (!pages[num_pages])
1891 req = fuse_get_req(fc);
1897 memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
1898 req->num_pages = num_pages;
1900 /* okay, let's send it to the client */
1901 req->in.h.opcode = FUSE_IOCTL;
1902 req->in.h.nodeid = ff->nodeid;
1903 req->in.numargs = 1;
1904 req->in.args[0].size = sizeof(inarg);
1905 req->in.args[0].value = &inarg;
1908 req->in.args[1].size = in_size;
1909 req->in.argpages = 1;
1911 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
1917 req->out.numargs = 2;
1918 req->out.args[0].size = sizeof(outarg);
1919 req->out.args[0].value = &outarg;
1920 req->out.args[1].size = out_size;
1921 req->out.argpages = 1;
1922 req->out.argvar = 1;
1924 fuse_request_send(fc, req);
1925 err = req->out.h.error;
1926 transferred = req->out.args[1].size;
1927 fuse_put_request(fc, req);
1932 /* did it ask for retry? */
1933 if (outarg.flags & FUSE_IOCTL_RETRY) {
1936 /* no retry if in restricted mode */
1938 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
1941 in_iovs = outarg.in_iovs;
1942 out_iovs = outarg.out_iovs;
1945 * Make sure things are in boundary, separate checks
1946 * are to protect against overflow.
1949 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
1950 out_iovs > FUSE_IOCTL_MAX_IOV ||
1951 in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
1954 vaddr = kmap_atomic(pages[0], KM_USER0);
1955 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
1956 transferred, in_iovs + out_iovs,
1957 (flags & FUSE_IOCTL_COMPAT) != 0);
1958 kunmap_atomic(vaddr, KM_USER0);
1963 out_iov = in_iov + in_iovs;
1965 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
1969 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
1977 if (transferred > inarg.out_size)
1980 err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
1983 fuse_put_request(fc, req);
1984 free_page((unsigned long) iov_page);
1986 __free_page(pages[--num_pages]);
1989 return err ? err : outarg.result;
1991 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
1993 static long fuse_file_ioctl_common(struct file *file, unsigned int cmd,
1994 unsigned long arg, unsigned int flags)
1996 struct inode *inode = file->f_dentry->d_inode;
1997 struct fuse_conn *fc = get_fuse_conn(inode);
1999 if (!fuse_allow_task(fc, current))
2002 if (is_bad_inode(inode))
2005 return fuse_do_ioctl(file, cmd, arg, flags);
2008 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2011 return fuse_file_ioctl_common(file, cmd, arg, 0);
2014 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2017 return fuse_file_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2021 * All files which have been polled are linked to RB tree
2022 * fuse_conn->polled_files which is indexed by kh. Walk the tree and
2023 * find the matching one.
2025 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2026 struct rb_node **parent_out)
2028 struct rb_node **link = &fc->polled_files.rb_node;
2029 struct rb_node *last = NULL;
2032 struct fuse_file *ff;
2035 ff = rb_entry(last, struct fuse_file, polled_node);
2038 link = &last->rb_left;
2039 else if (kh > ff->kh)
2040 link = &last->rb_right;
2051 * The file is about to be polled. Make sure it's on the polled_files
2052 * RB tree. Note that files once added to the polled_files tree are
2053 * not removed before the file is released. This is because a file
2054 * polled once is likely to be polled again.
2056 static void fuse_register_polled_file(struct fuse_conn *fc,
2057 struct fuse_file *ff)
2059 spin_lock(&fc->lock);
2060 if (RB_EMPTY_NODE(&ff->polled_node)) {
2061 struct rb_node **link, *parent;
2063 link = fuse_find_polled_node(fc, ff->kh, &parent);
2065 rb_link_node(&ff->polled_node, parent, link);
2066 rb_insert_color(&ff->polled_node, &fc->polled_files);
2068 spin_unlock(&fc->lock);
2071 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2073 struct fuse_file *ff = file->private_data;
2074 struct fuse_conn *fc = ff->fc;
2075 struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2076 struct fuse_poll_out outarg;
2077 struct fuse_req *req;
2081 return DEFAULT_POLLMASK;
2083 poll_wait(file, &ff->poll_wait, wait);
2086 * Ask for notification iff there's someone waiting for it.
2087 * The client may ignore the flag and always notify.
2089 if (waitqueue_active(&ff->poll_wait)) {
2090 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2091 fuse_register_polled_file(fc, ff);
2094 req = fuse_get_req(fc);
2098 req->in.h.opcode = FUSE_POLL;
2099 req->in.h.nodeid = ff->nodeid;
2100 req->in.numargs = 1;
2101 req->in.args[0].size = sizeof(inarg);
2102 req->in.args[0].value = &inarg;
2103 req->out.numargs = 1;
2104 req->out.args[0].size = sizeof(outarg);
2105 req->out.args[0].value = &outarg;
2106 fuse_request_send(fc, req);
2107 err = req->out.h.error;
2108 fuse_put_request(fc, req);
2111 return outarg.revents;
2112 if (err == -ENOSYS) {
2114 return DEFAULT_POLLMASK;
2118 EXPORT_SYMBOL_GPL(fuse_file_poll);
2121 * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2122 * wakes up the poll waiters.
2124 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2125 struct fuse_notify_poll_wakeup_out *outarg)
2127 u64 kh = outarg->kh;
2128 struct rb_node **link;
2130 spin_lock(&fc->lock);
2132 link = fuse_find_polled_node(fc, kh, NULL);
2134 struct fuse_file *ff;
2136 ff = rb_entry(*link, struct fuse_file, polled_node);
2137 wake_up_interruptible_sync(&ff->poll_wait);
2140 spin_unlock(&fc->lock);
2144 static const struct file_operations fuse_file_operations = {
2145 .llseek = fuse_file_llseek,
2146 .read = do_sync_read,
2147 .aio_read = fuse_file_aio_read,
2148 .write = do_sync_write,
2149 .aio_write = fuse_file_aio_write,
2150 .mmap = fuse_file_mmap,
2152 .flush = fuse_flush,
2153 .release = fuse_release,
2154 .fsync = fuse_fsync,
2155 .lock = fuse_file_lock,
2156 .flock = fuse_file_flock,
2157 .splice_read = generic_file_splice_read,
2158 .unlocked_ioctl = fuse_file_ioctl,
2159 .compat_ioctl = fuse_file_compat_ioctl,
2160 .poll = fuse_file_poll,
2163 static const struct file_operations fuse_direct_io_file_operations = {
2164 .llseek = fuse_file_llseek,
2165 .read = fuse_direct_read,
2166 .write = fuse_direct_write,
2167 .mmap = fuse_direct_mmap,
2169 .flush = fuse_flush,
2170 .release = fuse_release,
2171 .fsync = fuse_fsync,
2172 .lock = fuse_file_lock,
2173 .flock = fuse_file_flock,
2174 .unlocked_ioctl = fuse_file_ioctl,
2175 .compat_ioctl = fuse_file_compat_ioctl,
2176 .poll = fuse_file_poll,
2177 /* no splice_read */
2180 static const struct address_space_operations fuse_file_aops = {
2181 .readpage = fuse_readpage,
2182 .writepage = fuse_writepage,
2183 .launder_page = fuse_launder_page,
2184 .write_begin = fuse_write_begin,
2185 .write_end = fuse_write_end,
2186 .readpages = fuse_readpages,
2187 .set_page_dirty = __set_page_dirty_nobuffers,
2191 void fuse_init_file_inode(struct inode *inode)
2193 inode->i_fop = &fuse_file_operations;
2194 inode->i_data.a_ops = &fuse_file_aops;