2 FUSE: Filesystem in Userspace
3 Copyright (C) 2001-2008 Miklos Szeredi <miklos@szeredi.hu>
5 This program can be distributed under the terms of the GNU GPL.
11 #include <linux/pagemap.h>
12 #include <linux/slab.h>
13 #include <linux/kernel.h>
14 #include <linux/sched.h>
15 #include <linux/module.h>
16 #include <linux/compat.h>
17 #include <linux/swap.h>
18 #include <linux/falloc.h>
19 #include <linux/uio.h>
22 static const struct file_operations fuse_direct_io_file_operations;
24 static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
25 int opcode, struct fuse_open_out *outargp)
27 struct fuse_open_in inarg;
30 memset(&inarg, 0, sizeof(inarg));
31 inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
32 if (!fc->atomic_o_trunc)
33 inarg.flags &= ~O_TRUNC;
34 args.in.h.opcode = opcode;
35 args.in.h.nodeid = nodeid;
37 args.in.args[0].size = sizeof(inarg);
38 args.in.args[0].value = &inarg;
40 args.out.args[0].size = sizeof(*outargp);
41 args.out.args[0].value = outargp;
43 return fuse_simple_request(fc, &args);
46 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
50 ff = kzalloc(sizeof(struct fuse_file), GFP_KERNEL);
55 ff->reserved_req = fuse_request_alloc(0);
56 if (unlikely(!ff->reserved_req)) {
61 INIT_LIST_HEAD(&ff->write_entry);
62 refcount_set(&ff->count, 1);
63 RB_CLEAR_NODE(&ff->polled_node);
64 init_waitqueue_head(&ff->poll_wait);
68 spin_unlock(&fc->lock);
73 void fuse_file_free(struct fuse_file *ff)
75 fuse_request_free(ff->reserved_req);
79 static struct fuse_file *fuse_file_get(struct fuse_file *ff)
81 refcount_inc(&ff->count);
85 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
87 iput(req->misc.release.inode);
90 static void fuse_file_put(struct fuse_file *ff, bool sync, bool isdir)
92 if (refcount_dec_and_test(&ff->count)) {
93 struct fuse_req *req = ff->reserved_req;
95 if (ff->fc->no_open && !isdir) {
97 * Drop the release request when client does not
100 __clear_bit(FR_BACKGROUND, &req->flags);
101 iput(req->misc.release.inode);
102 fuse_put_request(ff->fc, req);
104 __set_bit(FR_FORCE, &req->flags);
105 __clear_bit(FR_BACKGROUND, &req->flags);
106 fuse_request_send(ff->fc, req);
107 iput(req->misc.release.inode);
108 fuse_put_request(ff->fc, req);
110 req->end = fuse_release_end;
111 __set_bit(FR_BACKGROUND, &req->flags);
112 fuse_request_send_background(ff->fc, req);
118 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
121 struct fuse_file *ff;
122 int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
124 ff = fuse_file_alloc(fc);
129 ff->open_flags = FOPEN_KEEP_CACHE; /* Default for no-open */
130 if (!fc->no_open || isdir) {
131 struct fuse_open_out outarg;
134 err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
137 ff->open_flags = outarg.open_flags;
139 } else if (err != -ENOSYS || isdir) {
148 ff->open_flags &= ~FOPEN_DIRECT_IO;
151 file->private_data = ff;
155 EXPORT_SYMBOL_GPL(fuse_do_open);
157 static void fuse_link_write_file(struct file *file)
159 struct inode *inode = file_inode(file);
160 struct fuse_conn *fc = get_fuse_conn(inode);
161 struct fuse_inode *fi = get_fuse_inode(inode);
162 struct fuse_file *ff = file->private_data;
164 * file may be written through mmap, so chain it onto the
165 * inodes's write_file list
167 spin_lock(&fc->lock);
168 if (list_empty(&ff->write_entry))
169 list_add(&ff->write_entry, &fi->write_files);
170 spin_unlock(&fc->lock);
173 void fuse_finish_open(struct inode *inode, struct file *file)
175 struct fuse_file *ff = file->private_data;
176 struct fuse_conn *fc = get_fuse_conn(inode);
178 if (ff->open_flags & FOPEN_DIRECT_IO)
179 file->f_op = &fuse_direct_io_file_operations;
180 if (!(ff->open_flags & FOPEN_KEEP_CACHE))
181 invalidate_inode_pages2(inode->i_mapping);
182 if (ff->open_flags & FOPEN_STREAM)
183 stream_open(inode, file);
184 else if (ff->open_flags & FOPEN_NONSEEKABLE)
185 nonseekable_open(inode, file);
186 if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
187 struct fuse_inode *fi = get_fuse_inode(inode);
189 spin_lock(&fc->lock);
190 fi->attr_version = ++fc->attr_version;
191 i_size_write(inode, 0);
192 spin_unlock(&fc->lock);
193 fuse_invalidate_attr(inode);
194 if (fc->writeback_cache)
195 file_update_time(file);
197 if ((file->f_mode & FMODE_WRITE) && fc->writeback_cache)
198 fuse_link_write_file(file);
201 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
203 struct fuse_conn *fc = get_fuse_conn(inode);
205 bool is_wb_truncate = (file->f_flags & O_TRUNC) &&
206 fc->atomic_o_trunc &&
209 if (fuse_is_bad(inode))
212 err = generic_file_open(inode, file);
216 if (is_wb_truncate) {
218 fuse_set_nowrite(inode);
221 err = fuse_do_open(fc, get_node_id(inode), file, isdir);
224 fuse_finish_open(inode, file);
226 if (is_wb_truncate) {
227 fuse_release_nowrite(inode);
234 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
236 struct fuse_conn *fc = ff->fc;
237 struct fuse_req *req = ff->reserved_req;
238 struct fuse_release_in *inarg = &req->misc.release.in;
240 spin_lock(&fc->lock);
241 list_del(&ff->write_entry);
242 if (!RB_EMPTY_NODE(&ff->polled_node))
243 rb_erase(&ff->polled_node, &fc->polled_files);
244 spin_unlock(&fc->lock);
246 wake_up_interruptible_all(&ff->poll_wait);
249 inarg->flags = flags;
250 req->in.h.opcode = opcode;
251 req->in.h.nodeid = ff->nodeid;
253 req->in.args[0].size = sizeof(struct fuse_release_in);
254 req->in.args[0].value = inarg;
257 void fuse_release_common(struct file *file, bool isdir)
259 struct fuse_file *ff = file->private_data;
260 struct fuse_req *req = ff->reserved_req;
261 int opcode = isdir ? FUSE_RELEASEDIR : FUSE_RELEASE;
263 fuse_prepare_release(ff, file->f_flags, opcode);
266 struct fuse_release_in *inarg = &req->misc.release.in;
267 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
268 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
271 /* Hold inode until release is finished */
272 req->misc.release.inode = igrab(file_inode(file));
275 * Normally this will send the RELEASE request, however if
276 * some asynchronous READ or WRITE requests are outstanding,
277 * the sending will be delayed.
279 * Make the release synchronous if this is a fuseblk mount,
280 * synchronous RELEASE is allowed (and desirable) in this case
281 * because the server can be trusted not to screw up.
283 fuse_file_put(ff, ff->fc->destroy_req != NULL, isdir);
286 static int fuse_open(struct inode *inode, struct file *file)
288 return fuse_open_common(inode, file, false);
291 static int fuse_release(struct inode *inode, struct file *file)
293 struct fuse_conn *fc = get_fuse_conn(inode);
295 /* see fuse_vma_close() for !writeback_cache case */
296 if (fc->writeback_cache)
297 write_inode_now(inode, 1);
299 fuse_release_common(file, false);
301 /* return value is ignored by VFS */
305 void fuse_sync_release(struct fuse_file *ff, int flags)
307 WARN_ON(refcount_read(&ff->count) > 1);
308 fuse_prepare_release(ff, flags, FUSE_RELEASE);
310 * iput(NULL) is a no-op and since the refcount is 1 and everything's
311 * synchronous, we are fine with not doing igrab() here"
313 fuse_file_put(ff, true, false);
315 EXPORT_SYMBOL_GPL(fuse_sync_release);
318 * Scramble the ID space with XTEA, so that the value of the files_struct
319 * pointer is not exposed to userspace.
321 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
323 u32 *k = fc->scramble_key;
324 u64 v = (unsigned long) id;
330 for (i = 0; i < 32; i++) {
331 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
333 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
336 return (u64) v0 + ((u64) v1 << 32);
340 * Check if any page in a range is under writeback
342 * This is currently done by walking the list of writepage requests
343 * for the inode, which can be pretty inefficient.
345 static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
348 struct fuse_conn *fc = get_fuse_conn(inode);
349 struct fuse_inode *fi = get_fuse_inode(inode);
350 struct fuse_req *req;
353 spin_lock(&fc->lock);
354 list_for_each_entry(req, &fi->writepages, writepages_entry) {
357 BUG_ON(req->inode != inode);
358 curr_index = req->misc.write.in.offset >> PAGE_SHIFT;
359 if (idx_from < curr_index + req->num_pages &&
360 curr_index <= idx_to) {
365 spin_unlock(&fc->lock);
370 static inline bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
372 return fuse_range_is_writeback(inode, index, index);
376 * Wait for page writeback to be completed.
378 * Since fuse doesn't rely on the VM writeback tracking, this has to
379 * use some other means.
381 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
383 struct fuse_inode *fi = get_fuse_inode(inode);
385 wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
390 * Wait for all pending writepages on the inode to finish.
392 * This is currently done by blocking further writes with FUSE_NOWRITE
393 * and waiting for all sent writes to complete.
395 * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
396 * could conflict with truncation.
398 static void fuse_sync_writes(struct inode *inode)
400 fuse_set_nowrite(inode);
401 fuse_release_nowrite(inode);
404 static int fuse_flush(struct file *file, fl_owner_t id)
406 struct inode *inode = file_inode(file);
407 struct fuse_conn *fc = get_fuse_conn(inode);
408 struct fuse_file *ff = file->private_data;
409 struct fuse_req *req;
410 struct fuse_flush_in inarg;
413 if (fuse_is_bad(inode))
419 err = write_inode_now(inode, 1);
424 fuse_sync_writes(inode);
427 err = filemap_check_errors(file->f_mapping);
431 req = fuse_get_req_nofail_nopages(fc, file);
432 memset(&inarg, 0, sizeof(inarg));
434 inarg.lock_owner = fuse_lock_owner_id(fc, id);
435 req->in.h.opcode = FUSE_FLUSH;
436 req->in.h.nodeid = get_node_id(inode);
438 req->in.args[0].size = sizeof(inarg);
439 req->in.args[0].value = &inarg;
440 __set_bit(FR_FORCE, &req->flags);
441 fuse_request_send(fc, req);
442 err = req->out.h.error;
443 fuse_put_request(fc, req);
444 if (err == -ENOSYS) {
451 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
452 int datasync, int isdir)
454 struct inode *inode = file->f_mapping->host;
455 struct fuse_conn *fc = get_fuse_conn(inode);
456 struct fuse_file *ff = file->private_data;
458 struct fuse_fsync_in inarg;
461 if (fuse_is_bad(inode))
467 * Start writeback against all dirty pages of the inode, then
468 * wait for all outstanding writes, before sending the FSYNC
471 err = file_write_and_wait_range(file, start, end);
475 fuse_sync_writes(inode);
478 * Due to implementation of fuse writeback
479 * file_write_and_wait_range() does not catch errors.
480 * We have to do this directly after fuse_sync_writes()
482 err = file_check_and_advance_wb_err(file);
486 err = sync_inode_metadata(inode, 1);
490 if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
493 memset(&inarg, 0, sizeof(inarg));
495 inarg.fsync_flags = datasync ? 1 : 0;
496 args.in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
497 args.in.h.nodeid = get_node_id(inode);
499 args.in.args[0].size = sizeof(inarg);
500 args.in.args[0].value = &inarg;
501 err = fuse_simple_request(fc, &args);
502 if (err == -ENOSYS) {
514 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
517 return fuse_fsync_common(file, start, end, datasync, 0);
520 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
521 size_t count, int opcode)
523 struct fuse_read_in *inarg = &req->misc.read.in;
524 struct fuse_file *ff = file->private_data;
529 inarg->flags = file->f_flags;
530 req->in.h.opcode = opcode;
531 req->in.h.nodeid = ff->nodeid;
533 req->in.args[0].size = sizeof(struct fuse_read_in);
534 req->in.args[0].value = inarg;
536 req->out.numargs = 1;
537 req->out.args[0].size = count;
540 static void fuse_release_user_pages(struct fuse_req *req, bool should_dirty)
544 for (i = 0; i < req->num_pages; i++) {
545 struct page *page = req->pages[i];
547 set_page_dirty_lock(page);
552 static void fuse_io_release(struct kref *kref)
554 kfree(container_of(kref, struct fuse_io_priv, refcnt));
557 static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
562 if (io->bytes >= 0 && io->write)
565 return io->bytes < 0 ? io->size : io->bytes;
569 * In case of short read, the caller sets 'pos' to the position of
570 * actual end of fuse request in IO request. Otherwise, if bytes_requested
571 * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
574 * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
575 * both submitted asynchronously. The first of them was ACKed by userspace as
576 * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
577 * second request was ACKed as short, e.g. only 1K was read, resulting in
580 * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
581 * will be equal to the length of the longest contiguous fragment of
582 * transferred data starting from the beginning of IO request.
584 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
588 spin_lock(&io->lock);
590 io->err = io->err ? : err;
591 else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
595 if (!left && io->blocking)
597 spin_unlock(&io->lock);
599 if (!left && !io->blocking) {
600 ssize_t res = fuse_get_res_by_io(io);
603 struct inode *inode = file_inode(io->iocb->ki_filp);
604 struct fuse_conn *fc = get_fuse_conn(inode);
605 struct fuse_inode *fi = get_fuse_inode(inode);
607 spin_lock(&fc->lock);
608 fi->attr_version = ++fc->attr_version;
609 spin_unlock(&fc->lock);
612 io->iocb->ki_complete(io->iocb, res, 0);
615 kref_put(&io->refcnt, fuse_io_release);
618 static void fuse_aio_complete_req(struct fuse_conn *fc, struct fuse_req *req)
620 struct fuse_io_priv *io = req->io;
623 fuse_release_user_pages(req, io->should_dirty);
626 if (req->misc.write.in.size != req->misc.write.out.size)
627 pos = req->misc.write.in.offset - io->offset +
628 req->misc.write.out.size;
630 if (req->misc.read.in.size != req->out.args[0].size)
631 pos = req->misc.read.in.offset - io->offset +
632 req->out.args[0].size;
635 fuse_aio_complete(io, req->out.h.error, pos);
638 static size_t fuse_async_req_send(struct fuse_conn *fc, struct fuse_req *req,
639 size_t num_bytes, struct fuse_io_priv *io)
641 spin_lock(&io->lock);
642 kref_get(&io->refcnt);
643 io->size += num_bytes;
645 spin_unlock(&io->lock);
648 req->end = fuse_aio_complete_req;
650 __fuse_get_request(req);
651 fuse_request_send_background(fc, req);
656 static size_t fuse_send_read(struct fuse_req *req, struct fuse_io_priv *io,
657 loff_t pos, size_t count, fl_owner_t owner)
659 struct file *file = io->iocb->ki_filp;
660 struct fuse_file *ff = file->private_data;
661 struct fuse_conn *fc = ff->fc;
663 fuse_read_fill(req, file, pos, count, FUSE_READ);
665 struct fuse_read_in *inarg = &req->misc.read.in;
667 inarg->read_flags |= FUSE_READ_LOCKOWNER;
668 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
672 return fuse_async_req_send(fc, req, count, io);
674 fuse_request_send(fc, req);
675 return req->out.args[0].size;
678 static void fuse_read_update_size(struct inode *inode, loff_t size,
681 struct fuse_conn *fc = get_fuse_conn(inode);
682 struct fuse_inode *fi = get_fuse_inode(inode);
684 spin_lock(&fc->lock);
685 if (attr_ver == fi->attr_version && size < inode->i_size &&
686 !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
687 fi->attr_version = ++fc->attr_version;
688 i_size_write(inode, size);
690 spin_unlock(&fc->lock);
693 static void fuse_short_read(struct fuse_req *req, struct inode *inode,
696 size_t num_read = req->out.args[0].size;
697 struct fuse_conn *fc = get_fuse_conn(inode);
699 if (fc->writeback_cache) {
701 * A hole in a file. Some data after the hole are in page cache,
702 * but have not reached the client fs yet. So, the hole is not
706 int start_idx = num_read >> PAGE_SHIFT;
707 size_t off = num_read & (PAGE_SIZE - 1);
709 for (i = start_idx; i < req->num_pages; i++) {
710 zero_user_segment(req->pages[i], off, PAGE_SIZE);
714 loff_t pos = page_offset(req->pages[0]) + num_read;
715 fuse_read_update_size(inode, pos, attr_ver);
719 static int fuse_do_readpage(struct file *file, struct page *page)
722 struct fuse_io_priv io;
723 struct inode *inode = page->mapping->host;
724 struct fuse_conn *fc = get_fuse_conn(inode);
725 struct fuse_req *req;
727 loff_t pos = page_offset(page);
728 size_t count = PAGE_SIZE;
733 * Page writeback can extend beyond the lifetime of the
734 * page-cache page, so make sure we read a properly synced
737 fuse_wait_on_page_writeback(inode, page->index);
739 req = fuse_get_req(fc, 1);
743 attr_ver = fuse_get_attr_version(fc);
745 req->out.page_zeroing = 1;
746 req->out.argpages = 1;
748 req->pages[0] = page;
749 req->page_descs[0].length = count;
750 init_sync_kiocb(&iocb, file);
751 io = (struct fuse_io_priv) FUSE_IO_PRIV_SYNC(&iocb);
752 num_read = fuse_send_read(req, &io, pos, count, NULL);
753 err = req->out.h.error;
757 * Short read means EOF. If file size is larger, truncate it
759 if (num_read < count)
760 fuse_short_read(req, inode, attr_ver);
762 SetPageUptodate(page);
765 fuse_put_request(fc, req);
770 static int fuse_readpage(struct file *file, struct page *page)
772 struct inode *inode = page->mapping->host;
776 if (fuse_is_bad(inode))
779 err = fuse_do_readpage(file, page);
780 fuse_invalidate_atime(inode);
786 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
789 size_t count = req->misc.read.in.size;
790 size_t num_read = req->out.args[0].size;
791 struct address_space *mapping = NULL;
793 for (i = 0; mapping == NULL && i < req->num_pages; i++)
794 mapping = req->pages[i]->mapping;
797 struct inode *inode = mapping->host;
800 * Short read means EOF. If file size is larger, truncate it
802 if (!req->out.h.error && num_read < count)
803 fuse_short_read(req, inode, req->misc.read.attr_ver);
805 fuse_invalidate_atime(inode);
808 for (i = 0; i < req->num_pages; i++) {
809 struct page *page = req->pages[i];
810 if (!req->out.h.error)
811 SetPageUptodate(page);
818 fuse_file_put(req->ff, false, false);
821 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
823 struct fuse_file *ff = file->private_data;
824 struct fuse_conn *fc = ff->fc;
825 loff_t pos = page_offset(req->pages[0]);
826 size_t count = req->num_pages << PAGE_SHIFT;
828 req->out.argpages = 1;
829 req->out.page_zeroing = 1;
830 req->out.page_replace = 1;
831 fuse_read_fill(req, file, pos, count, FUSE_READ);
832 req->misc.read.attr_ver = fuse_get_attr_version(fc);
833 if (fc->async_read) {
834 req->ff = fuse_file_get(ff);
835 req->end = fuse_readpages_end;
836 fuse_request_send_background(fc, req);
838 fuse_request_send(fc, req);
839 fuse_readpages_end(fc, req);
840 fuse_put_request(fc, req);
844 struct fuse_fill_data {
845 struct fuse_req *req;
851 static int fuse_readpages_fill(void *_data, struct page *page)
853 struct fuse_fill_data *data = _data;
854 struct fuse_req *req = data->req;
855 struct inode *inode = data->inode;
856 struct fuse_conn *fc = get_fuse_conn(inode);
858 fuse_wait_on_page_writeback(inode, page->index);
860 if (req->num_pages &&
861 (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
862 (req->num_pages + 1) * PAGE_SIZE > fc->max_read ||
863 req->pages[req->num_pages - 1]->index + 1 != page->index)) {
864 int nr_alloc = min_t(unsigned, data->nr_pages,
865 FUSE_MAX_PAGES_PER_REQ);
866 fuse_send_readpages(req, data->file);
868 req = fuse_get_req_for_background(fc, nr_alloc);
870 req = fuse_get_req(fc, nr_alloc);
879 if (WARN_ON(req->num_pages >= req->max_pages)) {
881 fuse_put_request(fc, req);
886 req->pages[req->num_pages] = page;
887 req->page_descs[req->num_pages].length = PAGE_SIZE;
893 static int fuse_readpages(struct file *file, struct address_space *mapping,
894 struct list_head *pages, unsigned nr_pages)
896 struct inode *inode = mapping->host;
897 struct fuse_conn *fc = get_fuse_conn(inode);
898 struct fuse_fill_data data;
900 int nr_alloc = min_t(unsigned, nr_pages, FUSE_MAX_PAGES_PER_REQ);
903 if (fuse_is_bad(inode))
909 data.req = fuse_get_req_for_background(fc, nr_alloc);
911 data.req = fuse_get_req(fc, nr_alloc);
912 data.nr_pages = nr_pages;
913 err = PTR_ERR(data.req);
914 if (IS_ERR(data.req))
917 err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
919 if (data.req->num_pages)
920 fuse_send_readpages(data.req, file);
922 fuse_put_request(fc, data.req);
928 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
930 struct inode *inode = iocb->ki_filp->f_mapping->host;
931 struct fuse_conn *fc = get_fuse_conn(inode);
933 if (fuse_is_bad(inode))
937 * In auto invalidate mode, always update attributes on read.
938 * Otherwise, only update if we attempt to read past EOF (to ensure
939 * i_size is up to date).
941 if (fc->auto_inval_data ||
942 (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
944 err = fuse_update_attributes(inode, iocb->ki_filp);
949 return generic_file_read_iter(iocb, to);
952 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
953 loff_t pos, size_t count)
955 struct fuse_write_in *inarg = &req->misc.write.in;
956 struct fuse_write_out *outarg = &req->misc.write.out;
961 req->in.h.opcode = FUSE_WRITE;
962 req->in.h.nodeid = ff->nodeid;
964 if (ff->fc->minor < 9)
965 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
967 req->in.args[0].size = sizeof(struct fuse_write_in);
968 req->in.args[0].value = inarg;
969 req->in.args[1].size = count;
970 req->out.numargs = 1;
971 req->out.args[0].size = sizeof(struct fuse_write_out);
972 req->out.args[0].value = outarg;
975 static size_t fuse_send_write(struct fuse_req *req, struct fuse_io_priv *io,
976 loff_t pos, size_t count, fl_owner_t owner)
978 struct kiocb *iocb = io->iocb;
979 struct file *file = iocb->ki_filp;
980 struct fuse_file *ff = file->private_data;
981 struct fuse_conn *fc = ff->fc;
982 struct fuse_write_in *inarg = &req->misc.write.in;
984 fuse_write_fill(req, ff, pos, count);
985 inarg->flags = file->f_flags;
986 if (iocb->ki_flags & IOCB_DSYNC)
987 inarg->flags |= O_DSYNC;
988 if (iocb->ki_flags & IOCB_SYNC)
989 inarg->flags |= O_SYNC;
991 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
992 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
996 return fuse_async_req_send(fc, req, count, io);
998 fuse_request_send(fc, req);
999 return req->misc.write.out.size;
1002 bool fuse_write_update_size(struct inode *inode, loff_t pos)
1004 struct fuse_conn *fc = get_fuse_conn(inode);
1005 struct fuse_inode *fi = get_fuse_inode(inode);
1008 spin_lock(&fc->lock);
1009 fi->attr_version = ++fc->attr_version;
1010 if (pos > inode->i_size) {
1011 i_size_write(inode, pos);
1014 spin_unlock(&fc->lock);
1019 static size_t fuse_send_write_pages(struct fuse_req *req, struct kiocb *iocb,
1020 struct inode *inode, loff_t pos,
1026 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1028 for (i = 0; i < req->num_pages; i++)
1029 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
1031 res = fuse_send_write(req, &io, pos, count, NULL);
1033 offset = req->page_descs[0].offset;
1035 for (i = 0; i < req->num_pages; i++) {
1036 struct page *page = req->pages[i];
1038 if (!req->out.h.error && !offset && count >= PAGE_SIZE)
1039 SetPageUptodate(page);
1041 if (count > PAGE_SIZE - offset)
1042 count -= PAGE_SIZE - offset;
1054 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
1055 struct address_space *mapping,
1056 struct iov_iter *ii, loff_t pos)
1058 struct fuse_conn *fc = get_fuse_conn(mapping->host);
1059 unsigned offset = pos & (PAGE_SIZE - 1);
1063 req->in.argpages = 1;
1064 req->page_descs[0].offset = offset;
1069 pgoff_t index = pos >> PAGE_SHIFT;
1070 size_t bytes = min_t(size_t, PAGE_SIZE - offset,
1071 iov_iter_count(ii));
1073 bytes = min_t(size_t, bytes, fc->max_write - count);
1077 if (iov_iter_fault_in_readable(ii, bytes))
1081 page = grab_cache_page_write_begin(mapping, index, 0);
1085 if (mapping_writably_mapped(mapping))
1086 flush_dcache_page(page);
1088 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
1089 flush_dcache_page(page);
1091 iov_iter_advance(ii, tmp);
1095 bytes = min(bytes, iov_iter_single_seg_count(ii));
1100 req->pages[req->num_pages] = page;
1101 req->page_descs[req->num_pages].length = tmp;
1107 if (offset == PAGE_SIZE)
1110 if (!fc->big_writes)
1112 } while (iov_iter_count(ii) && count < fc->max_write &&
1113 req->num_pages < req->max_pages && offset == 0);
1115 return count > 0 ? count : err;
1118 static inline unsigned fuse_wr_pages(loff_t pos, size_t len)
1120 return min_t(unsigned,
1121 ((pos + len - 1) >> PAGE_SHIFT) -
1122 (pos >> PAGE_SHIFT) + 1,
1123 FUSE_MAX_PAGES_PER_REQ);
1126 static ssize_t fuse_perform_write(struct kiocb *iocb,
1127 struct address_space *mapping,
1128 struct iov_iter *ii, loff_t pos)
1130 struct inode *inode = mapping->host;
1131 struct fuse_conn *fc = get_fuse_conn(inode);
1132 struct fuse_inode *fi = get_fuse_inode(inode);
1136 if (fuse_is_bad(inode))
1139 if (inode->i_size < pos + iov_iter_count(ii))
1140 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1143 struct fuse_req *req;
1145 unsigned nr_pages = fuse_wr_pages(pos, iov_iter_count(ii));
1147 req = fuse_get_req(fc, nr_pages);
1153 count = fuse_fill_write_pages(req, mapping, ii, pos);
1159 num_written = fuse_send_write_pages(req, iocb, inode,
1161 err = req->out.h.error;
1166 /* break out of the loop on short write */
1167 if (num_written != count)
1171 fuse_put_request(fc, req);
1172 } while (!err && iov_iter_count(ii));
1175 fuse_write_update_size(inode, pos);
1177 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1178 fuse_invalidate_attr(inode);
1180 return res > 0 ? res : err;
1183 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1185 struct file *file = iocb->ki_filp;
1186 struct address_space *mapping = file->f_mapping;
1187 ssize_t written = 0;
1188 ssize_t written_buffered = 0;
1189 struct inode *inode = mapping->host;
1193 if (fuse_is_bad(inode))
1196 if (get_fuse_conn(inode)->writeback_cache) {
1197 /* Update size (EOF optimization) and mode (SUID clearing) */
1198 err = fuse_update_attributes(mapping->host, file);
1202 return generic_file_write_iter(iocb, from);
1207 /* We can write back this queue in page reclaim */
1208 current->backing_dev_info = inode_to_bdi(inode);
1210 err = generic_write_checks(iocb, from);
1214 err = file_remove_privs(file);
1218 err = file_update_time(file);
1222 if (iocb->ki_flags & IOCB_DIRECT) {
1223 loff_t pos = iocb->ki_pos;
1224 written = generic_file_direct_write(iocb, from);
1225 if (written < 0 || !iov_iter_count(from))
1230 written_buffered = fuse_perform_write(iocb, mapping, from, pos);
1231 if (written_buffered < 0) {
1232 err = written_buffered;
1235 endbyte = pos + written_buffered - 1;
1237 err = filemap_write_and_wait_range(file->f_mapping, pos,
1242 invalidate_mapping_pages(file->f_mapping,
1244 endbyte >> PAGE_SHIFT);
1246 written += written_buffered;
1247 iocb->ki_pos = pos + written_buffered;
1249 written = fuse_perform_write(iocb, mapping, from, iocb->ki_pos);
1251 iocb->ki_pos += written;
1254 current->backing_dev_info = NULL;
1255 inode_unlock(inode);
1257 written = generic_write_sync(iocb, written);
1259 return written ? written : err;
1262 static inline void fuse_page_descs_length_init(struct fuse_req *req,
1263 unsigned index, unsigned nr_pages)
1267 for (i = index; i < index + nr_pages; i++)
1268 req->page_descs[i].length = PAGE_SIZE -
1269 req->page_descs[i].offset;
1272 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1274 return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1277 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1280 return min(iov_iter_single_seg_count(ii), max_size);
1283 static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1284 size_t *nbytesp, int write)
1286 size_t nbytes = 0; /* # bytes already packed in req */
1289 /* Special case for kernel I/O: can copy directly into the buffer */
1290 if (ii->type & ITER_KVEC) {
1291 unsigned long user_addr = fuse_get_user_addr(ii);
1292 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1295 req->in.args[1].value = (void *) user_addr;
1297 req->out.args[0].value = (void *) user_addr;
1299 iov_iter_advance(ii, frag_size);
1300 *nbytesp = frag_size;
1304 while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1307 ret = iov_iter_get_pages(ii, &req->pages[req->num_pages],
1309 req->max_pages - req->num_pages,
1314 iov_iter_advance(ii, ret);
1318 npages = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
1320 req->page_descs[req->num_pages].offset = start;
1321 fuse_page_descs_length_init(req, req->num_pages, npages);
1323 req->num_pages += npages;
1324 req->page_descs[req->num_pages - 1].length -=
1325 (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1328 req->user_pages = true;
1330 req->in.argpages = 1;
1332 req->out.argpages = 1;
1336 return ret < 0 ? ret : 0;
1339 static inline int fuse_iter_npages(const struct iov_iter *ii_p)
1341 return iov_iter_npages(ii_p, FUSE_MAX_PAGES_PER_REQ);
1344 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1345 loff_t *ppos, int flags)
1347 int write = flags & FUSE_DIO_WRITE;
1348 int cuse = flags & FUSE_DIO_CUSE;
1349 struct file *file = io->iocb->ki_filp;
1350 struct inode *inode = file->f_mapping->host;
1351 struct fuse_file *ff = file->private_data;
1352 struct fuse_conn *fc = ff->fc;
1353 size_t nmax = write ? fc->max_write : fc->max_read;
1355 size_t count = iov_iter_count(iter);
1356 pgoff_t idx_from = pos >> PAGE_SHIFT;
1357 pgoff_t idx_to = (pos + count - 1) >> PAGE_SHIFT;
1359 struct fuse_req *req;
1363 req = fuse_get_req_for_background(fc, fuse_iter_npages(iter));
1365 req = fuse_get_req(fc, fuse_iter_npages(iter));
1367 return PTR_ERR(req);
1369 if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1372 fuse_sync_writes(inode);
1374 inode_unlock(inode);
1377 io->should_dirty = !write && iter_is_iovec(iter);
1380 fl_owner_t owner = current->files;
1381 size_t nbytes = min(count, nmax);
1382 err = fuse_get_user_pages(req, iter, &nbytes, write);
1387 nres = fuse_send_write(req, io, pos, nbytes, owner);
1389 nres = fuse_send_read(req, io, pos, nbytes, owner);
1392 fuse_release_user_pages(req, io->should_dirty);
1393 if (req->out.h.error) {
1394 err = req->out.h.error;
1396 } else if (nres > nbytes) {
1407 fuse_put_request(fc, req);
1409 req = fuse_get_req_for_background(fc,
1410 fuse_iter_npages(iter));
1412 req = fuse_get_req(fc, fuse_iter_npages(iter));
1418 fuse_put_request(fc, req);
1422 return res > 0 ? res : err;
1424 EXPORT_SYMBOL_GPL(fuse_direct_io);
1426 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1427 struct iov_iter *iter,
1431 struct inode *inode = file_inode(io->iocb->ki_filp);
1433 if (fuse_is_bad(inode))
1436 res = fuse_direct_io(io, iter, ppos, 0);
1438 fuse_invalidate_attr(inode);
1443 static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1445 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1446 return __fuse_direct_read(&io, to, &iocb->ki_pos);
1449 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1451 struct inode *inode = file_inode(iocb->ki_filp);
1452 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1455 if (fuse_is_bad(inode))
1458 /* Don't allow parallel writes to the same file */
1460 res = generic_write_checks(iocb, from);
1462 res = fuse_direct_io(&io, from, &iocb->ki_pos, FUSE_DIO_WRITE);
1463 fuse_invalidate_attr(inode);
1465 fuse_write_update_size(inode, iocb->ki_pos);
1466 inode_unlock(inode);
1471 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1475 for (i = 0; i < req->num_pages; i++)
1476 __free_page(req->pages[i]);
1479 fuse_file_put(req->ff, false, false);
1482 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1484 struct inode *inode = req->inode;
1485 struct fuse_inode *fi = get_fuse_inode(inode);
1486 struct backing_dev_info *bdi = inode_to_bdi(inode);
1489 list_del(&req->writepages_entry);
1490 for (i = 0; i < req->num_pages; i++) {
1491 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1492 dec_node_page_state(req->pages[i], NR_WRITEBACK_TEMP);
1493 wb_writeout_inc(&bdi->wb);
1495 wake_up(&fi->page_waitq);
1498 /* Called under fc->lock, may release and reacquire it */
1499 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req,
1501 __releases(fc->lock)
1502 __acquires(fc->lock)
1504 struct fuse_inode *fi = get_fuse_inode(req->inode);
1505 struct fuse_write_in *inarg = &req->misc.write.in;
1506 __u64 data_size = req->num_pages * PAGE_SIZE;
1511 if (inarg->offset + data_size <= size) {
1512 inarg->size = data_size;
1513 } else if (inarg->offset < size) {
1514 inarg->size = size - inarg->offset;
1516 /* Got truncated off completely */
1520 req->in.args[1].size = inarg->size;
1522 fuse_request_send_background_locked(fc, req);
1526 fuse_writepage_finish(fc, req);
1527 spin_unlock(&fc->lock);
1528 fuse_writepage_free(fc, req);
1529 fuse_put_request(fc, req);
1530 spin_lock(&fc->lock);
1534 * If fi->writectr is positive (no truncate or fsync going on) send
1535 * all queued writepage requests.
1537 * Called with fc->lock
1539 void fuse_flush_writepages(struct inode *inode)
1540 __releases(fc->lock)
1541 __acquires(fc->lock)
1543 struct fuse_conn *fc = get_fuse_conn(inode);
1544 struct fuse_inode *fi = get_fuse_inode(inode);
1545 loff_t crop = i_size_read(inode);
1546 struct fuse_req *req;
1548 while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1549 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1550 list_del_init(&req->list);
1551 fuse_send_writepage(fc, req, crop);
1555 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1557 struct inode *inode = req->inode;
1558 struct fuse_inode *fi = get_fuse_inode(inode);
1560 mapping_set_error(inode->i_mapping, req->out.h.error);
1561 spin_lock(&fc->lock);
1562 while (req->misc.write.next) {
1563 struct fuse_conn *fc = get_fuse_conn(inode);
1564 struct fuse_write_in *inarg = &req->misc.write.in;
1565 struct fuse_req *next = req->misc.write.next;
1566 req->misc.write.next = next->misc.write.next;
1567 next->misc.write.next = NULL;
1568 next->ff = fuse_file_get(req->ff);
1569 list_add(&next->writepages_entry, &fi->writepages);
1572 * Skip fuse_flush_writepages() to make it easy to crop requests
1573 * based on primary request size.
1575 * 1st case (trivial): there are no concurrent activities using
1576 * fuse_set/release_nowrite. Then we're on safe side because
1577 * fuse_flush_writepages() would call fuse_send_writepage()
1580 * 2nd case: someone called fuse_set_nowrite and it is waiting
1581 * now for completion of all in-flight requests. This happens
1582 * rarely and no more than once per page, so this should be
1585 * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1586 * of fuse_set_nowrite..fuse_release_nowrite section. The fact
1587 * that fuse_set_nowrite returned implies that all in-flight
1588 * requests were completed along with all of their secondary
1589 * requests. Further primary requests are blocked by negative
1590 * writectr. Hence there cannot be any in-flight requests and
1591 * no invocations of fuse_writepage_end() while we're in
1592 * fuse_set_nowrite..fuse_release_nowrite section.
1594 fuse_send_writepage(fc, next, inarg->offset + inarg->size);
1597 fuse_writepage_finish(fc, req);
1598 spin_unlock(&fc->lock);
1599 fuse_writepage_free(fc, req);
1602 static struct fuse_file *__fuse_write_file_get(struct fuse_conn *fc,
1603 struct fuse_inode *fi)
1605 struct fuse_file *ff = NULL;
1607 spin_lock(&fc->lock);
1608 if (!list_empty(&fi->write_files)) {
1609 ff = list_entry(fi->write_files.next, struct fuse_file,
1613 spin_unlock(&fc->lock);
1618 static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1619 struct fuse_inode *fi)
1621 struct fuse_file *ff = __fuse_write_file_get(fc, fi);
1626 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1628 struct fuse_conn *fc = get_fuse_conn(inode);
1629 struct fuse_inode *fi = get_fuse_inode(inode);
1630 struct fuse_file *ff;
1633 ff = __fuse_write_file_get(fc, fi);
1634 err = fuse_flush_times(inode, ff);
1636 fuse_file_put(ff, false, false);
1641 static int fuse_writepage_locked(struct page *page)
1643 struct address_space *mapping = page->mapping;
1644 struct inode *inode = mapping->host;
1645 struct fuse_conn *fc = get_fuse_conn(inode);
1646 struct fuse_inode *fi = get_fuse_inode(inode);
1647 struct fuse_req *req;
1648 struct page *tmp_page;
1649 int error = -ENOMEM;
1651 set_page_writeback(page);
1653 req = fuse_request_alloc_nofs(1);
1657 /* writeback always goes to bg_queue */
1658 __set_bit(FR_BACKGROUND, &req->flags);
1659 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1664 req->ff = fuse_write_file_get(fc, fi);
1668 fuse_write_fill(req, req->ff, page_offset(page), 0);
1670 copy_highpage(tmp_page, page);
1671 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1672 req->misc.write.next = NULL;
1673 req->in.argpages = 1;
1675 req->pages[0] = tmp_page;
1676 req->page_descs[0].offset = 0;
1677 req->page_descs[0].length = PAGE_SIZE;
1678 req->end = fuse_writepage_end;
1681 inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1682 inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1684 spin_lock(&fc->lock);
1685 list_add(&req->writepages_entry, &fi->writepages);
1686 list_add_tail(&req->list, &fi->queued_writes);
1687 fuse_flush_writepages(inode);
1688 spin_unlock(&fc->lock);
1690 end_page_writeback(page);
1695 __free_page(tmp_page);
1697 fuse_request_free(req);
1699 mapping_set_error(page->mapping, error);
1700 end_page_writeback(page);
1704 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1708 if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1710 * ->writepages() should be called for sync() and friends. We
1711 * should only get here on direct reclaim and then we are
1712 * allowed to skip a page which is already in flight
1714 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1716 redirty_page_for_writepage(wbc, page);
1721 err = fuse_writepage_locked(page);
1727 struct fuse_fill_wb_data {
1728 struct fuse_req *req;
1729 struct fuse_file *ff;
1730 struct inode *inode;
1731 struct page **orig_pages;
1734 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1736 struct fuse_req *req = data->req;
1737 struct inode *inode = data->inode;
1738 struct fuse_conn *fc = get_fuse_conn(inode);
1739 struct fuse_inode *fi = get_fuse_inode(inode);
1740 int num_pages = req->num_pages;
1743 req->ff = fuse_file_get(data->ff);
1744 spin_lock(&fc->lock);
1745 list_add_tail(&req->list, &fi->queued_writes);
1746 fuse_flush_writepages(inode);
1747 spin_unlock(&fc->lock);
1749 for (i = 0; i < num_pages; i++)
1750 end_page_writeback(data->orig_pages[i]);
1753 static bool fuse_writepage_in_flight(struct fuse_req *new_req,
1756 struct fuse_conn *fc = get_fuse_conn(new_req->inode);
1757 struct fuse_inode *fi = get_fuse_inode(new_req->inode);
1758 struct fuse_req *tmp;
1759 struct fuse_req *old_req;
1763 BUG_ON(new_req->num_pages != 0);
1765 spin_lock(&fc->lock);
1766 list_del(&new_req->writepages_entry);
1767 list_for_each_entry(old_req, &fi->writepages, writepages_entry) {
1768 BUG_ON(old_req->inode != new_req->inode);
1769 curr_index = old_req->misc.write.in.offset >> PAGE_SHIFT;
1770 if (curr_index <= page->index &&
1771 page->index < curr_index + old_req->num_pages) {
1777 list_add(&new_req->writepages_entry, &fi->writepages);
1781 new_req->num_pages = 1;
1782 for (tmp = old_req; tmp != NULL; tmp = tmp->misc.write.next) {
1783 BUG_ON(tmp->inode != new_req->inode);
1784 curr_index = tmp->misc.write.in.offset >> PAGE_SHIFT;
1785 if (tmp->num_pages == 1 &&
1786 curr_index == page->index) {
1791 if (old_req->num_pages == 1 && test_bit(FR_PENDING, &old_req->flags)) {
1792 struct backing_dev_info *bdi = inode_to_bdi(page->mapping->host);
1794 copy_highpage(old_req->pages[0], page);
1795 spin_unlock(&fc->lock);
1797 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1798 dec_node_page_state(new_req->pages[0], NR_WRITEBACK_TEMP);
1799 wb_writeout_inc(&bdi->wb);
1800 fuse_writepage_free(fc, new_req);
1801 fuse_request_free(new_req);
1804 new_req->misc.write.next = old_req->misc.write.next;
1805 old_req->misc.write.next = new_req;
1808 spin_unlock(&fc->lock);
1813 static int fuse_writepages_fill(struct page *page,
1814 struct writeback_control *wbc, void *_data)
1816 struct fuse_fill_wb_data *data = _data;
1817 struct fuse_req *req = data->req;
1818 struct inode *inode = data->inode;
1819 struct fuse_conn *fc = get_fuse_conn(inode);
1820 struct page *tmp_page;
1826 data->ff = fuse_write_file_get(fc, get_fuse_inode(inode));
1832 * Being under writeback is unlikely but possible. For example direct
1833 * read to an mmaped fuse file will set the page dirty twice; once when
1834 * the pages are faulted with get_user_pages(), and then after the read
1837 is_writeback = fuse_page_is_writeback(inode, page->index);
1839 if (req && req->num_pages &&
1840 (is_writeback || req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
1841 (req->num_pages + 1) * PAGE_SIZE > fc->max_write ||
1842 data->orig_pages[req->num_pages - 1]->index + 1 != page->index)) {
1843 fuse_writepages_send(data);
1847 tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1852 * The page must not be redirtied until the writeout is completed
1853 * (i.e. userspace has sent a reply to the write request). Otherwise
1854 * there could be more than one temporary page instance for each real
1857 * This is ensured by holding the page lock in page_mkwrite() while
1858 * checking fuse_page_is_writeback(). We already hold the page lock
1859 * since clear_page_dirty_for_io() and keep it held until we add the
1860 * request to the fi->writepages list and increment req->num_pages.
1861 * After this fuse_page_is_writeback() will indicate that the page is
1862 * under writeback, so we can release the page lock.
1864 if (data->req == NULL) {
1865 struct fuse_inode *fi = get_fuse_inode(inode);
1868 req = fuse_request_alloc_nofs(FUSE_MAX_PAGES_PER_REQ);
1870 __free_page(tmp_page);
1874 fuse_write_fill(req, data->ff, page_offset(page), 0);
1875 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1876 req->misc.write.next = NULL;
1877 req->in.argpages = 1;
1878 __set_bit(FR_BACKGROUND, &req->flags);
1880 req->end = fuse_writepage_end;
1883 spin_lock(&fc->lock);
1884 list_add(&req->writepages_entry, &fi->writepages);
1885 spin_unlock(&fc->lock);
1889 set_page_writeback(page);
1891 copy_highpage(tmp_page, page);
1892 req->pages[req->num_pages] = tmp_page;
1893 req->page_descs[req->num_pages].offset = 0;
1894 req->page_descs[req->num_pages].length = PAGE_SIZE;
1896 inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1897 inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1900 if (is_writeback && fuse_writepage_in_flight(req, page)) {
1901 end_page_writeback(page);
1905 data->orig_pages[req->num_pages] = page;
1908 * Protected by fc->lock against concurrent access by
1909 * fuse_page_is_writeback().
1911 spin_lock(&fc->lock);
1913 spin_unlock(&fc->lock);
1921 static int fuse_writepages(struct address_space *mapping,
1922 struct writeback_control *wbc)
1924 struct inode *inode = mapping->host;
1925 struct fuse_fill_wb_data data;
1929 if (fuse_is_bad(inode))
1937 data.orig_pages = kcalloc(FUSE_MAX_PAGES_PER_REQ,
1938 sizeof(struct page *),
1940 if (!data.orig_pages)
1943 err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
1945 /* Ignore errors if we can write at least one page */
1946 BUG_ON(!data.req->num_pages);
1947 fuse_writepages_send(&data);
1951 fuse_file_put(data.ff, false, false);
1953 kfree(data.orig_pages);
1959 * It's worthy to make sure that space is reserved on disk for the write,
1960 * but how to implement it without killing performance need more thinking.
1962 static int fuse_write_begin(struct file *file, struct address_space *mapping,
1963 loff_t pos, unsigned len, unsigned flags,
1964 struct page **pagep, void **fsdata)
1966 pgoff_t index = pos >> PAGE_SHIFT;
1967 struct fuse_conn *fc = get_fuse_conn(file_inode(file));
1972 WARN_ON(!fc->writeback_cache);
1974 page = grab_cache_page_write_begin(mapping, index, flags);
1978 fuse_wait_on_page_writeback(mapping->host, page->index);
1980 if (PageUptodate(page) || len == PAGE_SIZE)
1983 * Check if the start this page comes after the end of file, in which
1984 * case the readpage can be optimized away.
1986 fsize = i_size_read(mapping->host);
1987 if (fsize <= (pos & PAGE_MASK)) {
1988 size_t off = pos & ~PAGE_MASK;
1990 zero_user_segment(page, 0, off);
1993 err = fuse_do_readpage(file, page);
2007 static int fuse_write_end(struct file *file, struct address_space *mapping,
2008 loff_t pos, unsigned len, unsigned copied,
2009 struct page *page, void *fsdata)
2011 struct inode *inode = page->mapping->host;
2013 /* Haven't copied anything? Skip zeroing, size extending, dirtying. */
2017 if (!PageUptodate(page)) {
2018 /* Zero any unwritten bytes at the end of the page */
2019 size_t endoff = (pos + copied) & ~PAGE_MASK;
2021 zero_user_segment(page, endoff, PAGE_SIZE);
2022 SetPageUptodate(page);
2025 fuse_write_update_size(inode, pos + copied);
2026 set_page_dirty(page);
2035 static int fuse_launder_page(struct page *page)
2038 if (clear_page_dirty_for_io(page)) {
2039 struct inode *inode = page->mapping->host;
2040 err = fuse_writepage_locked(page);
2042 fuse_wait_on_page_writeback(inode, page->index);
2048 * Write back dirty pages now, because there may not be any suitable
2051 static void fuse_vma_close(struct vm_area_struct *vma)
2053 filemap_write_and_wait(vma->vm_file->f_mapping);
2057 * Wait for writeback against this page to complete before allowing it
2058 * to be marked dirty again, and hence written back again, possibly
2059 * before the previous writepage completed.
2061 * Block here, instead of in ->writepage(), so that the userspace fs
2062 * can only block processes actually operating on the filesystem.
2064 * Otherwise unprivileged userspace fs would be able to block
2069 * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2071 static int fuse_page_mkwrite(struct vm_fault *vmf)
2073 struct page *page = vmf->page;
2074 struct inode *inode = file_inode(vmf->vma->vm_file);
2076 file_update_time(vmf->vma->vm_file);
2078 if (page->mapping != inode->i_mapping) {
2080 return VM_FAULT_NOPAGE;
2083 fuse_wait_on_page_writeback(inode, page->index);
2084 return VM_FAULT_LOCKED;
2087 static const struct vm_operations_struct fuse_file_vm_ops = {
2088 .close = fuse_vma_close,
2089 .fault = filemap_fault,
2090 .map_pages = filemap_map_pages,
2091 .page_mkwrite = fuse_page_mkwrite,
2094 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2096 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2097 fuse_link_write_file(file);
2099 file_accessed(file);
2100 vma->vm_ops = &fuse_file_vm_ops;
2104 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
2106 /* Can't provide the coherency needed for MAP_SHARED */
2107 if (vma->vm_flags & VM_MAYSHARE)
2110 invalidate_inode_pages2(file->f_mapping);
2112 return generic_file_mmap(file, vma);
2115 static int convert_fuse_file_lock(struct fuse_conn *fc,
2116 const struct fuse_file_lock *ffl,
2117 struct file_lock *fl)
2119 switch (ffl->type) {
2125 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2126 ffl->end < ffl->start)
2129 fl->fl_start = ffl->start;
2130 fl->fl_end = ffl->end;
2133 * Convert pid into init's pid namespace. The locks API will
2134 * translate it into the caller's pid namespace.
2137 fl->fl_pid = pid_nr_ns(find_pid_ns(ffl->pid, fc->pid_ns), &init_pid_ns);
2144 fl->fl_type = ffl->type;
2148 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2149 const struct file_lock *fl, int opcode, pid_t pid,
2150 int flock, struct fuse_lk_in *inarg)
2152 struct inode *inode = file_inode(file);
2153 struct fuse_conn *fc = get_fuse_conn(inode);
2154 struct fuse_file *ff = file->private_data;
2156 memset(inarg, 0, sizeof(*inarg));
2158 inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2159 inarg->lk.start = fl->fl_start;
2160 inarg->lk.end = fl->fl_end;
2161 inarg->lk.type = fl->fl_type;
2162 inarg->lk.pid = pid;
2164 inarg->lk_flags |= FUSE_LK_FLOCK;
2165 args->in.h.opcode = opcode;
2166 args->in.h.nodeid = get_node_id(inode);
2167 args->in.numargs = 1;
2168 args->in.args[0].size = sizeof(*inarg);
2169 args->in.args[0].value = inarg;
2172 static int fuse_getlk(struct file *file, struct file_lock *fl)
2174 struct inode *inode = file_inode(file);
2175 struct fuse_conn *fc = get_fuse_conn(inode);
2177 struct fuse_lk_in inarg;
2178 struct fuse_lk_out outarg;
2181 fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2182 args.out.numargs = 1;
2183 args.out.args[0].size = sizeof(outarg);
2184 args.out.args[0].value = &outarg;
2185 err = fuse_simple_request(fc, &args);
2187 err = convert_fuse_file_lock(fc, &outarg.lk, fl);
2192 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2194 struct inode *inode = file_inode(file);
2195 struct fuse_conn *fc = get_fuse_conn(inode);
2197 struct fuse_lk_in inarg;
2198 int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2199 struct pid *pid = fl->fl_type != F_UNLCK ? task_tgid(current) : NULL;
2200 pid_t pid_nr = pid_nr_ns(pid, fc->pid_ns);
2203 if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2204 /* NLM needs asynchronous locks, which we don't support yet */
2208 /* Unlock on close is handled by the flush method */
2209 if ((fl->fl_flags & FL_CLOSE_POSIX) == FL_CLOSE_POSIX)
2212 fuse_lk_fill(&args, file, fl, opcode, pid_nr, flock, &inarg);
2213 err = fuse_simple_request(fc, &args);
2215 /* locking is restartable */
2222 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2224 struct inode *inode = file_inode(file);
2225 struct fuse_conn *fc = get_fuse_conn(inode);
2228 if (cmd == F_CANCELLK) {
2230 } else if (cmd == F_GETLK) {
2232 posix_test_lock(file, fl);
2235 err = fuse_getlk(file, fl);
2238 err = posix_lock_file(file, fl, NULL);
2240 err = fuse_setlk(file, fl, 0);
2245 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2247 struct inode *inode = file_inode(file);
2248 struct fuse_conn *fc = get_fuse_conn(inode);
2252 err = locks_lock_file_wait(file, fl);
2254 struct fuse_file *ff = file->private_data;
2256 /* emulate flock with POSIX locks */
2258 err = fuse_setlk(file, fl, 1);
2264 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2266 struct inode *inode = mapping->host;
2267 struct fuse_conn *fc = get_fuse_conn(inode);
2269 struct fuse_bmap_in inarg;
2270 struct fuse_bmap_out outarg;
2273 if (!inode->i_sb->s_bdev || fc->no_bmap)
2276 memset(&inarg, 0, sizeof(inarg));
2277 inarg.block = block;
2278 inarg.blocksize = inode->i_sb->s_blocksize;
2279 args.in.h.opcode = FUSE_BMAP;
2280 args.in.h.nodeid = get_node_id(inode);
2281 args.in.numargs = 1;
2282 args.in.args[0].size = sizeof(inarg);
2283 args.in.args[0].value = &inarg;
2284 args.out.numargs = 1;
2285 args.out.args[0].size = sizeof(outarg);
2286 args.out.args[0].value = &outarg;
2287 err = fuse_simple_request(fc, &args);
2291 return err ? 0 : outarg.block;
2294 static loff_t fuse_lseek(struct file *file, loff_t offset, int whence)
2296 struct inode *inode = file->f_mapping->host;
2297 struct fuse_conn *fc = get_fuse_conn(inode);
2298 struct fuse_file *ff = file->private_data;
2300 struct fuse_lseek_in inarg = {
2305 struct fuse_lseek_out outarg;
2311 args.in.h.opcode = FUSE_LSEEK;
2312 args.in.h.nodeid = ff->nodeid;
2313 args.in.numargs = 1;
2314 args.in.args[0].size = sizeof(inarg);
2315 args.in.args[0].value = &inarg;
2316 args.out.numargs = 1;
2317 args.out.args[0].size = sizeof(outarg);
2318 args.out.args[0].value = &outarg;
2319 err = fuse_simple_request(fc, &args);
2321 if (err == -ENOSYS) {
2328 return vfs_setpos(file, outarg.offset, inode->i_sb->s_maxbytes);
2331 err = fuse_update_attributes(inode, file);
2333 return generic_file_llseek(file, offset, whence);
2338 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2341 struct inode *inode = file_inode(file);
2346 /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2347 retval = generic_file_llseek(file, offset, whence);
2351 retval = fuse_update_attributes(inode, file);
2353 retval = generic_file_llseek(file, offset, whence);
2354 inode_unlock(inode);
2359 retval = fuse_lseek(file, offset, whence);
2360 inode_unlock(inode);
2370 * CUSE servers compiled on 32bit broke on 64bit kernels because the
2371 * ABI was defined to be 'struct iovec' which is different on 32bit
2372 * and 64bit. Fortunately we can determine which structure the server
2373 * used from the size of the reply.
2375 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2376 size_t transferred, unsigned count,
2379 #ifdef CONFIG_COMPAT
2380 if (count * sizeof(struct compat_iovec) == transferred) {
2381 struct compat_iovec *ciov = src;
2385 * With this interface a 32bit server cannot support
2386 * non-compat (i.e. ones coming from 64bit apps) ioctl
2392 for (i = 0; i < count; i++) {
2393 dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2394 dst[i].iov_len = ciov[i].iov_len;
2400 if (count * sizeof(struct iovec) != transferred)
2403 memcpy(dst, src, transferred);
2407 /* Make sure iov_length() won't overflow */
2408 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
2411 u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
2413 for (n = 0; n < count; n++, iov++) {
2414 if (iov->iov_len > (size_t) max)
2416 max -= iov->iov_len;
2421 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2422 void *src, size_t transferred, unsigned count,
2426 struct fuse_ioctl_iovec *fiov = src;
2428 if (fc->minor < 16) {
2429 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2433 if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2436 for (i = 0; i < count; i++) {
2437 /* Did the server supply an inappropriate value? */
2438 if (fiov[i].base != (unsigned long) fiov[i].base ||
2439 fiov[i].len != (unsigned long) fiov[i].len)
2442 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2443 dst[i].iov_len = (size_t) fiov[i].len;
2445 #ifdef CONFIG_COMPAT
2447 (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2448 (compat_size_t) dst[i].iov_len != fiov[i].len))
2458 * For ioctls, there is no generic way to determine how much memory
2459 * needs to be read and/or written. Furthermore, ioctls are allowed
2460 * to dereference the passed pointer, so the parameter requires deep
2461 * copying but FUSE has no idea whatsoever about what to copy in or
2464 * This is solved by allowing FUSE server to retry ioctl with
2465 * necessary in/out iovecs. Let's assume the ioctl implementation
2466 * needs to read in the following structure.
2473 * On the first callout to FUSE server, inarg->in_size and
2474 * inarg->out_size will be NULL; then, the server completes the ioctl
2475 * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2476 * the actual iov array to
2478 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) } }
2480 * which tells FUSE to copy in the requested area and retry the ioctl.
2481 * On the second round, the server has access to the structure and
2482 * from that it can tell what to look for next, so on the invocation,
2483 * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2485 * { { .iov_base = inarg.arg, .iov_len = sizeof(struct a) },
2486 * { .iov_base = a.buf, .iov_len = a.buflen } }
2488 * FUSE will copy both struct a and the pointed buffer from the
2489 * process doing the ioctl and retry ioctl with both struct a and the
2492 * This time, FUSE server has everything it needs and completes ioctl
2493 * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2495 * Copying data out works the same way.
2497 * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2498 * automatically initializes in and out iovs by decoding @cmd with
2499 * _IOC_* macros and the server is not allowed to request RETRY. This
2500 * limits ioctl data transfers to well-formed ioctls and is the forced
2501 * behavior for all FUSE servers.
2503 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2506 struct fuse_file *ff = file->private_data;
2507 struct fuse_conn *fc = ff->fc;
2508 struct fuse_ioctl_in inarg = {
2514 struct fuse_ioctl_out outarg;
2515 struct fuse_req *req = NULL;
2516 struct page **pages = NULL;
2517 struct iovec *iov_page = NULL;
2518 struct iovec *in_iov = NULL, *out_iov = NULL;
2519 unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
2520 size_t in_size, out_size, transferred, c;
2524 #if BITS_PER_LONG == 32
2525 inarg.flags |= FUSE_IOCTL_32BIT;
2527 if (flags & FUSE_IOCTL_COMPAT)
2528 inarg.flags |= FUSE_IOCTL_32BIT;
2531 /* assume all the iovs returned by client always fits in a page */
2532 BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2535 pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
2536 iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2537 if (!pages || !iov_page)
2541 * If restricted, initialize IO parameters as encoded in @cmd.
2542 * RETRY from server is not allowed.
2544 if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2545 struct iovec *iov = iov_page;
2547 iov->iov_base = (void __user *)arg;
2550 case FS_IOC_GETFLAGS:
2551 case FS_IOC_SETFLAGS:
2552 iov->iov_len = sizeof(int);
2555 iov->iov_len = _IOC_SIZE(cmd);
2559 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2564 if (_IOC_DIR(cmd) & _IOC_READ) {
2571 inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2572 inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2575 * Out data can be used either for actual out data or iovs,
2576 * make sure there always is at least one page.
2578 out_size = max_t(size_t, out_size, PAGE_SIZE);
2579 max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2581 /* make sure there are enough buffer pages and init request with them */
2583 if (max_pages > FUSE_MAX_PAGES_PER_REQ)
2585 while (num_pages < max_pages) {
2586 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2587 if (!pages[num_pages])
2592 req = fuse_get_req(fc, num_pages);
2598 memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
2599 req->num_pages = num_pages;
2600 fuse_page_descs_length_init(req, 0, req->num_pages);
2602 /* okay, let's send it to the client */
2603 req->in.h.opcode = FUSE_IOCTL;
2604 req->in.h.nodeid = ff->nodeid;
2605 req->in.numargs = 1;
2606 req->in.args[0].size = sizeof(inarg);
2607 req->in.args[0].value = &inarg;
2610 req->in.args[1].size = in_size;
2611 req->in.argpages = 1;
2614 iov_iter_init(&ii, WRITE, in_iov, in_iovs, in_size);
2615 for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= num_pages); i++) {
2616 c = copy_page_from_iter(pages[i], 0, PAGE_SIZE, &ii);
2617 if (c != PAGE_SIZE && iov_iter_count(&ii))
2622 req->out.numargs = 2;
2623 req->out.args[0].size = sizeof(outarg);
2624 req->out.args[0].value = &outarg;
2625 req->out.args[1].size = out_size;
2626 req->out.argpages = 1;
2627 req->out.argvar = 1;
2629 fuse_request_send(fc, req);
2630 err = req->out.h.error;
2631 transferred = req->out.args[1].size;
2632 fuse_put_request(fc, req);
2637 /* did it ask for retry? */
2638 if (outarg.flags & FUSE_IOCTL_RETRY) {
2641 /* no retry if in restricted mode */
2643 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2646 in_iovs = outarg.in_iovs;
2647 out_iovs = outarg.out_iovs;
2650 * Make sure things are in boundary, separate checks
2651 * are to protect against overflow.
2654 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2655 out_iovs > FUSE_IOCTL_MAX_IOV ||
2656 in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2659 vaddr = kmap_atomic(pages[0]);
2660 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2661 transferred, in_iovs + out_iovs,
2662 (flags & FUSE_IOCTL_COMPAT) != 0);
2663 kunmap_atomic(vaddr);
2668 out_iov = in_iov + in_iovs;
2670 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
2674 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
2682 if (transferred > inarg.out_size)
2686 iov_iter_init(&ii, READ, out_iov, out_iovs, transferred);
2687 for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= num_pages); i++) {
2688 c = copy_page_to_iter(pages[i], 0, PAGE_SIZE, &ii);
2689 if (c != PAGE_SIZE && iov_iter_count(&ii))
2695 fuse_put_request(fc, req);
2696 free_page((unsigned long) iov_page);
2698 __free_page(pages[--num_pages]);
2701 return err ? err : outarg.result;
2703 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2705 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2706 unsigned long arg, unsigned int flags)
2708 struct inode *inode = file_inode(file);
2709 struct fuse_conn *fc = get_fuse_conn(inode);
2711 if (!fuse_allow_current_process(fc))
2714 if (fuse_is_bad(inode))
2717 return fuse_do_ioctl(file, cmd, arg, flags);
2720 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2723 return fuse_ioctl_common(file, cmd, arg, 0);
2726 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2729 return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2733 * All files which have been polled are linked to RB tree
2734 * fuse_conn->polled_files which is indexed by kh. Walk the tree and
2735 * find the matching one.
2737 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2738 struct rb_node **parent_out)
2740 struct rb_node **link = &fc->polled_files.rb_node;
2741 struct rb_node *last = NULL;
2744 struct fuse_file *ff;
2747 ff = rb_entry(last, struct fuse_file, polled_node);
2750 link = &last->rb_left;
2751 else if (kh > ff->kh)
2752 link = &last->rb_right;
2763 * The file is about to be polled. Make sure it's on the polled_files
2764 * RB tree. Note that files once added to the polled_files tree are
2765 * not removed before the file is released. This is because a file
2766 * polled once is likely to be polled again.
2768 static void fuse_register_polled_file(struct fuse_conn *fc,
2769 struct fuse_file *ff)
2771 spin_lock(&fc->lock);
2772 if (RB_EMPTY_NODE(&ff->polled_node)) {
2773 struct rb_node **link, *parent;
2775 link = fuse_find_polled_node(fc, ff->kh, &parent);
2777 rb_link_node(&ff->polled_node, parent, link);
2778 rb_insert_color(&ff->polled_node, &fc->polled_files);
2780 spin_unlock(&fc->lock);
2783 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2785 struct fuse_file *ff = file->private_data;
2786 struct fuse_conn *fc = ff->fc;
2787 struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2788 struct fuse_poll_out outarg;
2793 return DEFAULT_POLLMASK;
2795 poll_wait(file, &ff->poll_wait, wait);
2796 inarg.events = (__u32)poll_requested_events(wait);
2799 * Ask for notification iff there's someone waiting for it.
2800 * The client may ignore the flag and always notify.
2802 if (waitqueue_active(&ff->poll_wait)) {
2803 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2804 fuse_register_polled_file(fc, ff);
2807 args.in.h.opcode = FUSE_POLL;
2808 args.in.h.nodeid = ff->nodeid;
2809 args.in.numargs = 1;
2810 args.in.args[0].size = sizeof(inarg);
2811 args.in.args[0].value = &inarg;
2812 args.out.numargs = 1;
2813 args.out.args[0].size = sizeof(outarg);
2814 args.out.args[0].value = &outarg;
2815 err = fuse_simple_request(fc, &args);
2818 return outarg.revents;
2819 if (err == -ENOSYS) {
2821 return DEFAULT_POLLMASK;
2825 EXPORT_SYMBOL_GPL(fuse_file_poll);
2828 * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2829 * wakes up the poll waiters.
2831 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2832 struct fuse_notify_poll_wakeup_out *outarg)
2834 u64 kh = outarg->kh;
2835 struct rb_node **link;
2837 spin_lock(&fc->lock);
2839 link = fuse_find_polled_node(fc, kh, NULL);
2841 struct fuse_file *ff;
2843 ff = rb_entry(*link, struct fuse_file, polled_node);
2844 wake_up_interruptible_sync(&ff->poll_wait);
2847 spin_unlock(&fc->lock);
2851 static void fuse_do_truncate(struct file *file)
2853 struct inode *inode = file->f_mapping->host;
2856 attr.ia_valid = ATTR_SIZE;
2857 attr.ia_size = i_size_read(inode);
2859 attr.ia_file = file;
2860 attr.ia_valid |= ATTR_FILE;
2862 fuse_do_setattr(file_dentry(file), &attr, file);
2865 static inline loff_t fuse_round_up(loff_t off)
2867 return round_up(off, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
2871 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
2873 DECLARE_COMPLETION_ONSTACK(wait);
2875 struct file *file = iocb->ki_filp;
2876 struct fuse_file *ff = file->private_data;
2877 bool async_dio = ff->fc->async_dio;
2879 struct inode *inode;
2881 size_t count = iov_iter_count(iter);
2882 loff_t offset = iocb->ki_pos;
2883 struct fuse_io_priv *io;
2886 inode = file->f_mapping->host;
2887 i_size = i_size_read(inode);
2889 if ((iov_iter_rw(iter) == READ) && (offset > i_size))
2892 /* optimization for short read */
2893 if (async_dio && iov_iter_rw(iter) != WRITE && offset + count > i_size) {
2894 if (offset >= i_size)
2896 iov_iter_truncate(iter, fuse_round_up(i_size - offset));
2897 count = iov_iter_count(iter);
2900 io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2903 spin_lock_init(&io->lock);
2904 kref_init(&io->refcnt);
2908 io->offset = offset;
2909 io->write = (iov_iter_rw(iter) == WRITE);
2912 * By default, we want to optimize all I/Os with async request
2913 * submission to the client filesystem if supported.
2915 io->async = async_dio;
2917 io->blocking = is_sync_kiocb(iocb);
2920 * We cannot asynchronously extend the size of a file.
2921 * In such case the aio will behave exactly like sync io.
2923 if ((offset + count > i_size) && iov_iter_rw(iter) == WRITE)
2924 io->blocking = true;
2926 if (io->async && io->blocking) {
2928 * Additional reference to keep io around after
2929 * calling fuse_aio_complete()
2931 kref_get(&io->refcnt);
2935 if (iov_iter_rw(iter) == WRITE) {
2936 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
2937 fuse_invalidate_attr(inode);
2939 ret = __fuse_direct_read(io, iter, &pos);
2943 bool blocking = io->blocking;
2945 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2947 /* we have a non-extending, async request, so return */
2949 return -EIOCBQUEUED;
2951 wait_for_completion(&wait);
2952 ret = fuse_get_res_by_io(io);
2955 kref_put(&io->refcnt, fuse_io_release);
2957 if (iov_iter_rw(iter) == WRITE) {
2959 fuse_write_update_size(inode, pos);
2960 else if (ret < 0 && offset + count > i_size)
2961 fuse_do_truncate(file);
2967 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2970 struct fuse_file *ff = file->private_data;
2971 struct inode *inode = file_inode(file);
2972 struct fuse_inode *fi = get_fuse_inode(inode);
2973 struct fuse_conn *fc = ff->fc;
2975 struct fuse_fallocate_in inarg = {
2982 bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
2983 (mode & FALLOC_FL_PUNCH_HOLE);
2985 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2988 if (fc->no_fallocate)
2993 if (mode & FALLOC_FL_PUNCH_HOLE) {
2994 loff_t endbyte = offset + length - 1;
2995 err = filemap_write_and_wait_range(inode->i_mapping,
3000 fuse_sync_writes(inode);
3004 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
3005 offset + length > i_size_read(inode)) {
3006 err = inode_newsize_ok(inode, offset + length);
3011 if (!(mode & FALLOC_FL_KEEP_SIZE))
3012 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3014 args.in.h.opcode = FUSE_FALLOCATE;
3015 args.in.h.nodeid = ff->nodeid;
3016 args.in.numargs = 1;
3017 args.in.args[0].size = sizeof(inarg);
3018 args.in.args[0].value = &inarg;
3019 err = fuse_simple_request(fc, &args);
3020 if (err == -ENOSYS) {
3021 fc->no_fallocate = 1;
3027 /* we could have extended the file */
3028 if (!(mode & FALLOC_FL_KEEP_SIZE)) {
3029 bool changed = fuse_write_update_size(inode, offset + length);
3031 if (changed && fc->writeback_cache)
3032 file_update_time(file);
3035 if (mode & FALLOC_FL_PUNCH_HOLE)
3036 truncate_pagecache_range(inode, offset, offset + length - 1);
3038 fuse_invalidate_attr(inode);
3041 if (!(mode & FALLOC_FL_KEEP_SIZE))
3042 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3045 inode_unlock(inode);
3050 static const struct file_operations fuse_file_operations = {
3051 .llseek = fuse_file_llseek,
3052 .read_iter = fuse_file_read_iter,
3053 .write_iter = fuse_file_write_iter,
3054 .mmap = fuse_file_mmap,
3056 .flush = fuse_flush,
3057 .release = fuse_release,
3058 .fsync = fuse_fsync,
3059 .lock = fuse_file_lock,
3060 .flock = fuse_file_flock,
3061 .splice_read = generic_file_splice_read,
3062 .unlocked_ioctl = fuse_file_ioctl,
3063 .compat_ioctl = fuse_file_compat_ioctl,
3064 .poll = fuse_file_poll,
3065 .fallocate = fuse_file_fallocate,
3068 static const struct file_operations fuse_direct_io_file_operations = {
3069 .llseek = fuse_file_llseek,
3070 .read_iter = fuse_direct_read_iter,
3071 .write_iter = fuse_direct_write_iter,
3072 .mmap = fuse_direct_mmap,
3074 .flush = fuse_flush,
3075 .release = fuse_release,
3076 .fsync = fuse_fsync,
3077 .lock = fuse_file_lock,
3078 .flock = fuse_file_flock,
3079 .unlocked_ioctl = fuse_file_ioctl,
3080 .compat_ioctl = fuse_file_compat_ioctl,
3081 .poll = fuse_file_poll,
3082 .fallocate = fuse_file_fallocate,
3083 /* no splice_read */
3086 static const struct address_space_operations fuse_file_aops = {
3087 .readpage = fuse_readpage,
3088 .writepage = fuse_writepage,
3089 .writepages = fuse_writepages,
3090 .launder_page = fuse_launder_page,
3091 .readpages = fuse_readpages,
3092 .set_page_dirty = __set_page_dirty_nobuffers,
3094 .direct_IO = fuse_direct_IO,
3095 .write_begin = fuse_write_begin,
3096 .write_end = fuse_write_end,
3099 void fuse_init_file_inode(struct inode *inode)
3101 inode->i_fop = &fuse_file_operations;
3102 inode->i_data.a_ops = &fuse_file_aops;