xref: /openbmc/linux/fs/fuse/file.c (revision 81d67439)
1 /*
2   FUSE: Filesystem in Userspace
3   Copyright (C) 2001-2008  Miklos Szeredi <miklos@szeredi.hu>
4 
5   This program can be distributed under the terms of the GNU GPL.
6   See the file COPYING.
7 */
8 
9 #include "fuse_i.h"
10 
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 
18 static const struct file_operations fuse_direct_io_file_operations;
19 
20 static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
21 			  int opcode, struct fuse_open_out *outargp)
22 {
23 	struct fuse_open_in inarg;
24 	struct fuse_req *req;
25 	int err;
26 
27 	req = fuse_get_req(fc);
28 	if (IS_ERR(req))
29 		return PTR_ERR(req);
30 
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;
37 	req->in.numargs = 1;
38 	req->in.args[0].size = sizeof(inarg);
39 	req->in.args[0].value = &inarg;
40 	req->out.numargs = 1;
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);
46 
47 	return err;
48 }
49 
50 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
51 {
52 	struct fuse_file *ff;
53 
54 	ff = kmalloc(sizeof(struct fuse_file), GFP_KERNEL);
55 	if (unlikely(!ff))
56 		return NULL;
57 
58 	ff->fc = fc;
59 	ff->reserved_req = fuse_request_alloc();
60 	if (unlikely(!ff->reserved_req)) {
61 		kfree(ff);
62 		return NULL;
63 	}
64 
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);
69 
70 	spin_lock(&fc->lock);
71 	ff->kh = ++fc->khctr;
72 	spin_unlock(&fc->lock);
73 
74 	return ff;
75 }
76 
77 void fuse_file_free(struct fuse_file *ff)
78 {
79 	fuse_request_free(ff->reserved_req);
80 	kfree(ff);
81 }
82 
83 struct fuse_file *fuse_file_get(struct fuse_file *ff)
84 {
85 	atomic_inc(&ff->count);
86 	return ff;
87 }
88 
89 static void fuse_release_async(struct work_struct *work)
90 {
91 	struct fuse_req *req;
92 	struct fuse_conn *fc;
93 	struct path path;
94 
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);
98 
99 	fuse_put_request(fc, req);
100 	path_put(&path);
101 }
102 
103 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
104 {
105 	if (fc->destroy_req) {
106 		/*
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
112 		 * thread.
113 		 */
114 		atomic_inc(&req->count);
115 		INIT_WORK(&req->misc.release.work, fuse_release_async);
116 		schedule_work(&req->misc.release.work);
117 	} else {
118 		path_put(&req->misc.release.path);
119 	}
120 }
121 
122 static void fuse_file_put(struct fuse_file *ff, bool sync)
123 {
124 	if (atomic_dec_and_test(&ff->count)) {
125 		struct fuse_req *req = ff->reserved_req;
126 
127 		if (sync) {
128 			fuse_request_send(ff->fc, req);
129 			path_put(&req->misc.release.path);
130 			fuse_put_request(ff->fc, req);
131 		} else {
132 			req->end = fuse_release_end;
133 			fuse_request_send_background(ff->fc, req);
134 		}
135 		kfree(ff);
136 	}
137 }
138 
139 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
140 		 bool isdir)
141 {
142 	struct fuse_open_out outarg;
143 	struct fuse_file *ff;
144 	int err;
145 	int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
146 
147 	ff = fuse_file_alloc(fc);
148 	if (!ff)
149 		return -ENOMEM;
150 
151 	err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
152 	if (err) {
153 		fuse_file_free(ff);
154 		return err;
155 	}
156 
157 	if (isdir)
158 		outarg.open_flags &= ~FOPEN_DIRECT_IO;
159 
160 	ff->fh = outarg.fh;
161 	ff->nodeid = nodeid;
162 	ff->open_flags = outarg.open_flags;
163 	file->private_data = fuse_file_get(ff);
164 
165 	return 0;
166 }
167 EXPORT_SYMBOL_GPL(fuse_do_open);
168 
169 void fuse_finish_open(struct inode *inode, struct file *file)
170 {
171 	struct fuse_file *ff = file->private_data;
172 	struct fuse_conn *fc = get_fuse_conn(inode);
173 
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);
182 
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);
188 	}
189 }
190 
191 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
192 {
193 	struct fuse_conn *fc = get_fuse_conn(inode);
194 	int err;
195 
196 	/* VFS checks this, but only _after_ ->open() */
197 	if (file->f_flags & O_DIRECT)
198 		return -EINVAL;
199 
200 	err = generic_file_open(inode, file);
201 	if (err)
202 		return err;
203 
204 	err = fuse_do_open(fc, get_node_id(inode), file, isdir);
205 	if (err)
206 		return err;
207 
208 	fuse_finish_open(inode, file);
209 
210 	return 0;
211 }
212 
213 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
214 {
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;
218 
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);
224 
225 	wake_up_interruptible_all(&ff->poll_wait);
226 
227 	inarg->fh = ff->fh;
228 	inarg->flags = flags;
229 	req->in.h.opcode = opcode;
230 	req->in.h.nodeid = ff->nodeid;
231 	req->in.numargs = 1;
232 	req->in.args[0].size = sizeof(struct fuse_release_in);
233 	req->in.args[0].value = inarg;
234 }
235 
236 void fuse_release_common(struct file *file, int opcode)
237 {
238 	struct fuse_file *ff;
239 	struct fuse_req *req;
240 
241 	ff = file->private_data;
242 	if (unlikely(!ff))
243 		return;
244 
245 	req = ff->reserved_req;
246 	fuse_prepare_release(ff, file->f_flags, opcode);
247 
248 	/* Hold vfsmount and dentry until release is finished */
249 	path_get(&file->f_path);
250 	req->misc.release.path = file->f_path;
251 
252 	/*
253 	 * Normally this will send the RELEASE request, however if
254 	 * some asynchronous READ or WRITE requests are outstanding,
255 	 * the sending will be delayed.
256 	 *
257 	 * Make the release synchronous if this is a fuseblk mount,
258 	 * synchronous RELEASE is allowed (and desirable) in this case
259 	 * because the server can be trusted not to screw up.
260 	 */
261 	fuse_file_put(ff, ff->fc->destroy_req != NULL);
262 }
263 
264 static int fuse_open(struct inode *inode, struct file *file)
265 {
266 	return fuse_open_common(inode, file, false);
267 }
268 
269 static int fuse_release(struct inode *inode, struct file *file)
270 {
271 	fuse_release_common(file, FUSE_RELEASE);
272 
273 	/* return value is ignored by VFS */
274 	return 0;
275 }
276 
277 void fuse_sync_release(struct fuse_file *ff, int flags)
278 {
279 	WARN_ON(atomic_read(&ff->count) > 1);
280 	fuse_prepare_release(ff, flags, FUSE_RELEASE);
281 	ff->reserved_req->force = 1;
282 	fuse_request_send(ff->fc, ff->reserved_req);
283 	fuse_put_request(ff->fc, ff->reserved_req);
284 	kfree(ff);
285 }
286 EXPORT_SYMBOL_GPL(fuse_sync_release);
287 
288 /*
289  * Scramble the ID space with XTEA, so that the value of the files_struct
290  * pointer is not exposed to userspace.
291  */
292 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
293 {
294 	u32 *k = fc->scramble_key;
295 	u64 v = (unsigned long) id;
296 	u32 v0 = v;
297 	u32 v1 = v >> 32;
298 	u32 sum = 0;
299 	int i;
300 
301 	for (i = 0; i < 32; i++) {
302 		v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
303 		sum += 0x9E3779B9;
304 		v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
305 	}
306 
307 	return (u64) v0 + ((u64) v1 << 32);
308 }
309 
310 /*
311  * Check if page is under writeback
312  *
313  * This is currently done by walking the list of writepage requests
314  * for the inode, which can be pretty inefficient.
315  */
316 static bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
317 {
318 	struct fuse_conn *fc = get_fuse_conn(inode);
319 	struct fuse_inode *fi = get_fuse_inode(inode);
320 	struct fuse_req *req;
321 	bool found = false;
322 
323 	spin_lock(&fc->lock);
324 	list_for_each_entry(req, &fi->writepages, writepages_entry) {
325 		pgoff_t curr_index;
326 
327 		BUG_ON(req->inode != inode);
328 		curr_index = req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
329 		if (curr_index == index) {
330 			found = true;
331 			break;
332 		}
333 	}
334 	spin_unlock(&fc->lock);
335 
336 	return found;
337 }
338 
339 /*
340  * Wait for page writeback to be completed.
341  *
342  * Since fuse doesn't rely on the VM writeback tracking, this has to
343  * use some other means.
344  */
345 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
346 {
347 	struct fuse_inode *fi = get_fuse_inode(inode);
348 
349 	wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
350 	return 0;
351 }
352 
353 static int fuse_flush(struct file *file, fl_owner_t id)
354 {
355 	struct inode *inode = file->f_path.dentry->d_inode;
356 	struct fuse_conn *fc = get_fuse_conn(inode);
357 	struct fuse_file *ff = file->private_data;
358 	struct fuse_req *req;
359 	struct fuse_flush_in inarg;
360 	int err;
361 
362 	if (is_bad_inode(inode))
363 		return -EIO;
364 
365 	if (fc->no_flush)
366 		return 0;
367 
368 	req = fuse_get_req_nofail(fc, file);
369 	memset(&inarg, 0, sizeof(inarg));
370 	inarg.fh = ff->fh;
371 	inarg.lock_owner = fuse_lock_owner_id(fc, id);
372 	req->in.h.opcode = FUSE_FLUSH;
373 	req->in.h.nodeid = get_node_id(inode);
374 	req->in.numargs = 1;
375 	req->in.args[0].size = sizeof(inarg);
376 	req->in.args[0].value = &inarg;
377 	req->force = 1;
378 	fuse_request_send(fc, req);
379 	err = req->out.h.error;
380 	fuse_put_request(fc, req);
381 	if (err == -ENOSYS) {
382 		fc->no_flush = 1;
383 		err = 0;
384 	}
385 	return err;
386 }
387 
388 /*
389  * Wait for all pending writepages on the inode to finish.
390  *
391  * This is currently done by blocking further writes with FUSE_NOWRITE
392  * and waiting for all sent writes to complete.
393  *
394  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
395  * could conflict with truncation.
396  */
397 static void fuse_sync_writes(struct inode *inode)
398 {
399 	fuse_set_nowrite(inode);
400 	fuse_release_nowrite(inode);
401 }
402 
403 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
404 		      int datasync, int isdir)
405 {
406 	struct inode *inode = file->f_mapping->host;
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_fsync_in inarg;
411 	int err;
412 
413 	if (is_bad_inode(inode))
414 		return -EIO;
415 
416 	err = filemap_write_and_wait_range(inode->i_mapping, start, end);
417 	if (err)
418 		return err;
419 
420 	if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
421 		return 0;
422 
423 	mutex_lock(&inode->i_mutex);
424 
425 	/*
426 	 * Start writeback against all dirty pages of the inode, then
427 	 * wait for all outstanding writes, before sending the FSYNC
428 	 * request.
429 	 */
430 	err = write_inode_now(inode, 0);
431 	if (err)
432 		goto out;
433 
434 	fuse_sync_writes(inode);
435 
436 	req = fuse_get_req(fc);
437 	if (IS_ERR(req)) {
438 		err = PTR_ERR(req);
439 		goto out;
440 	}
441 
442 	memset(&inarg, 0, sizeof(inarg));
443 	inarg.fh = ff->fh;
444 	inarg.fsync_flags = datasync ? 1 : 0;
445 	req->in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
446 	req->in.h.nodeid = get_node_id(inode);
447 	req->in.numargs = 1;
448 	req->in.args[0].size = sizeof(inarg);
449 	req->in.args[0].value = &inarg;
450 	fuse_request_send(fc, req);
451 	err = req->out.h.error;
452 	fuse_put_request(fc, req);
453 	if (err == -ENOSYS) {
454 		if (isdir)
455 			fc->no_fsyncdir = 1;
456 		else
457 			fc->no_fsync = 1;
458 		err = 0;
459 	}
460 out:
461 	mutex_unlock(&inode->i_mutex);
462 	return err;
463 }
464 
465 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
466 		      int datasync)
467 {
468 	return fuse_fsync_common(file, start, end, datasync, 0);
469 }
470 
471 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
472 		    size_t count, int opcode)
473 {
474 	struct fuse_read_in *inarg = &req->misc.read.in;
475 	struct fuse_file *ff = file->private_data;
476 
477 	inarg->fh = ff->fh;
478 	inarg->offset = pos;
479 	inarg->size = count;
480 	inarg->flags = file->f_flags;
481 	req->in.h.opcode = opcode;
482 	req->in.h.nodeid = ff->nodeid;
483 	req->in.numargs = 1;
484 	req->in.args[0].size = sizeof(struct fuse_read_in);
485 	req->in.args[0].value = inarg;
486 	req->out.argvar = 1;
487 	req->out.numargs = 1;
488 	req->out.args[0].size = count;
489 }
490 
491 static size_t fuse_send_read(struct fuse_req *req, struct file *file,
492 			     loff_t pos, size_t count, fl_owner_t owner)
493 {
494 	struct fuse_file *ff = file->private_data;
495 	struct fuse_conn *fc = ff->fc;
496 
497 	fuse_read_fill(req, file, pos, count, FUSE_READ);
498 	if (owner != NULL) {
499 		struct fuse_read_in *inarg = &req->misc.read.in;
500 
501 		inarg->read_flags |= FUSE_READ_LOCKOWNER;
502 		inarg->lock_owner = fuse_lock_owner_id(fc, owner);
503 	}
504 	fuse_request_send(fc, req);
505 	return req->out.args[0].size;
506 }
507 
508 static void fuse_read_update_size(struct inode *inode, loff_t size,
509 				  u64 attr_ver)
510 {
511 	struct fuse_conn *fc = get_fuse_conn(inode);
512 	struct fuse_inode *fi = get_fuse_inode(inode);
513 
514 	spin_lock(&fc->lock);
515 	if (attr_ver == fi->attr_version && size < inode->i_size) {
516 		fi->attr_version = ++fc->attr_version;
517 		i_size_write(inode, size);
518 	}
519 	spin_unlock(&fc->lock);
520 }
521 
522 static int fuse_readpage(struct file *file, struct page *page)
523 {
524 	struct inode *inode = page->mapping->host;
525 	struct fuse_conn *fc = get_fuse_conn(inode);
526 	struct fuse_req *req;
527 	size_t num_read;
528 	loff_t pos = page_offset(page);
529 	size_t count = PAGE_CACHE_SIZE;
530 	u64 attr_ver;
531 	int err;
532 
533 	err = -EIO;
534 	if (is_bad_inode(inode))
535 		goto out;
536 
537 	/*
538 	 * Page writeback can extend beyond the lifetime of the
539 	 * page-cache page, so make sure we read a properly synced
540 	 * page.
541 	 */
542 	fuse_wait_on_page_writeback(inode, page->index);
543 
544 	req = fuse_get_req(fc);
545 	err = PTR_ERR(req);
546 	if (IS_ERR(req))
547 		goto out;
548 
549 	attr_ver = fuse_get_attr_version(fc);
550 
551 	req->out.page_zeroing = 1;
552 	req->out.argpages = 1;
553 	req->num_pages = 1;
554 	req->pages[0] = page;
555 	num_read = fuse_send_read(req, file, pos, count, NULL);
556 	err = req->out.h.error;
557 	fuse_put_request(fc, req);
558 
559 	if (!err) {
560 		/*
561 		 * Short read means EOF.  If file size is larger, truncate it
562 		 */
563 		if (num_read < count)
564 			fuse_read_update_size(inode, pos + num_read, attr_ver);
565 
566 		SetPageUptodate(page);
567 	}
568 
569 	fuse_invalidate_attr(inode); /* atime changed */
570  out:
571 	unlock_page(page);
572 	return err;
573 }
574 
575 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
576 {
577 	int i;
578 	size_t count = req->misc.read.in.size;
579 	size_t num_read = req->out.args[0].size;
580 	struct address_space *mapping = NULL;
581 
582 	for (i = 0; mapping == NULL && i < req->num_pages; i++)
583 		mapping = req->pages[i]->mapping;
584 
585 	if (mapping) {
586 		struct inode *inode = mapping->host;
587 
588 		/*
589 		 * Short read means EOF. If file size is larger, truncate it
590 		 */
591 		if (!req->out.h.error && num_read < count) {
592 			loff_t pos;
593 
594 			pos = page_offset(req->pages[0]) + num_read;
595 			fuse_read_update_size(inode, pos,
596 					      req->misc.read.attr_ver);
597 		}
598 		fuse_invalidate_attr(inode); /* atime changed */
599 	}
600 
601 	for (i = 0; i < req->num_pages; i++) {
602 		struct page *page = req->pages[i];
603 		if (!req->out.h.error)
604 			SetPageUptodate(page);
605 		else
606 			SetPageError(page);
607 		unlock_page(page);
608 		page_cache_release(page);
609 	}
610 	if (req->ff)
611 		fuse_file_put(req->ff, false);
612 }
613 
614 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
615 {
616 	struct fuse_file *ff = file->private_data;
617 	struct fuse_conn *fc = ff->fc;
618 	loff_t pos = page_offset(req->pages[0]);
619 	size_t count = req->num_pages << PAGE_CACHE_SHIFT;
620 
621 	req->out.argpages = 1;
622 	req->out.page_zeroing = 1;
623 	req->out.page_replace = 1;
624 	fuse_read_fill(req, file, pos, count, FUSE_READ);
625 	req->misc.read.attr_ver = fuse_get_attr_version(fc);
626 	if (fc->async_read) {
627 		req->ff = fuse_file_get(ff);
628 		req->end = fuse_readpages_end;
629 		fuse_request_send_background(fc, req);
630 	} else {
631 		fuse_request_send(fc, req);
632 		fuse_readpages_end(fc, req);
633 		fuse_put_request(fc, req);
634 	}
635 }
636 
637 struct fuse_fill_data {
638 	struct fuse_req *req;
639 	struct file *file;
640 	struct inode *inode;
641 };
642 
643 static int fuse_readpages_fill(void *_data, struct page *page)
644 {
645 	struct fuse_fill_data *data = _data;
646 	struct fuse_req *req = data->req;
647 	struct inode *inode = data->inode;
648 	struct fuse_conn *fc = get_fuse_conn(inode);
649 
650 	fuse_wait_on_page_writeback(inode, page->index);
651 
652 	if (req->num_pages &&
653 	    (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
654 	     (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
655 	     req->pages[req->num_pages - 1]->index + 1 != page->index)) {
656 		fuse_send_readpages(req, data->file);
657 		data->req = req = fuse_get_req(fc);
658 		if (IS_ERR(req)) {
659 			unlock_page(page);
660 			return PTR_ERR(req);
661 		}
662 	}
663 	page_cache_get(page);
664 	req->pages[req->num_pages] = page;
665 	req->num_pages++;
666 	return 0;
667 }
668 
669 static int fuse_readpages(struct file *file, struct address_space *mapping,
670 			  struct list_head *pages, unsigned nr_pages)
671 {
672 	struct inode *inode = mapping->host;
673 	struct fuse_conn *fc = get_fuse_conn(inode);
674 	struct fuse_fill_data data;
675 	int err;
676 
677 	err = -EIO;
678 	if (is_bad_inode(inode))
679 		goto out;
680 
681 	data.file = file;
682 	data.inode = inode;
683 	data.req = fuse_get_req(fc);
684 	err = PTR_ERR(data.req);
685 	if (IS_ERR(data.req))
686 		goto out;
687 
688 	err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
689 	if (!err) {
690 		if (data.req->num_pages)
691 			fuse_send_readpages(data.req, file);
692 		else
693 			fuse_put_request(fc, data.req);
694 	}
695 out:
696 	return err;
697 }
698 
699 static ssize_t fuse_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
700 				  unsigned long nr_segs, loff_t pos)
701 {
702 	struct inode *inode = iocb->ki_filp->f_mapping->host;
703 
704 	if (pos + iov_length(iov, nr_segs) > i_size_read(inode)) {
705 		int err;
706 		/*
707 		 * If trying to read past EOF, make sure the i_size
708 		 * attribute is up-to-date.
709 		 */
710 		err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
711 		if (err)
712 			return err;
713 	}
714 
715 	return generic_file_aio_read(iocb, iov, nr_segs, pos);
716 }
717 
718 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
719 			    loff_t pos, size_t count)
720 {
721 	struct fuse_write_in *inarg = &req->misc.write.in;
722 	struct fuse_write_out *outarg = &req->misc.write.out;
723 
724 	inarg->fh = ff->fh;
725 	inarg->offset = pos;
726 	inarg->size = count;
727 	req->in.h.opcode = FUSE_WRITE;
728 	req->in.h.nodeid = ff->nodeid;
729 	req->in.numargs = 2;
730 	if (ff->fc->minor < 9)
731 		req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
732 	else
733 		req->in.args[0].size = sizeof(struct fuse_write_in);
734 	req->in.args[0].value = inarg;
735 	req->in.args[1].size = count;
736 	req->out.numargs = 1;
737 	req->out.args[0].size = sizeof(struct fuse_write_out);
738 	req->out.args[0].value = outarg;
739 }
740 
741 static size_t fuse_send_write(struct fuse_req *req, struct file *file,
742 			      loff_t pos, size_t count, fl_owner_t owner)
743 {
744 	struct fuse_file *ff = file->private_data;
745 	struct fuse_conn *fc = ff->fc;
746 	struct fuse_write_in *inarg = &req->misc.write.in;
747 
748 	fuse_write_fill(req, ff, pos, count);
749 	inarg->flags = file->f_flags;
750 	if (owner != NULL) {
751 		inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
752 		inarg->lock_owner = fuse_lock_owner_id(fc, owner);
753 	}
754 	fuse_request_send(fc, req);
755 	return req->misc.write.out.size;
756 }
757 
758 static int fuse_write_begin(struct file *file, struct address_space *mapping,
759 			loff_t pos, unsigned len, unsigned flags,
760 			struct page **pagep, void **fsdata)
761 {
762 	pgoff_t index = pos >> PAGE_CACHE_SHIFT;
763 
764 	*pagep = grab_cache_page_write_begin(mapping, index, flags);
765 	if (!*pagep)
766 		return -ENOMEM;
767 	return 0;
768 }
769 
770 void fuse_write_update_size(struct inode *inode, loff_t pos)
771 {
772 	struct fuse_conn *fc = get_fuse_conn(inode);
773 	struct fuse_inode *fi = get_fuse_inode(inode);
774 
775 	spin_lock(&fc->lock);
776 	fi->attr_version = ++fc->attr_version;
777 	if (pos > inode->i_size)
778 		i_size_write(inode, pos);
779 	spin_unlock(&fc->lock);
780 }
781 
782 static int fuse_buffered_write(struct file *file, struct inode *inode,
783 			       loff_t pos, unsigned count, struct page *page)
784 {
785 	int err;
786 	size_t nres;
787 	struct fuse_conn *fc = get_fuse_conn(inode);
788 	unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
789 	struct fuse_req *req;
790 
791 	if (is_bad_inode(inode))
792 		return -EIO;
793 
794 	/*
795 	 * Make sure writepages on the same page are not mixed up with
796 	 * plain writes.
797 	 */
798 	fuse_wait_on_page_writeback(inode, page->index);
799 
800 	req = fuse_get_req(fc);
801 	if (IS_ERR(req))
802 		return PTR_ERR(req);
803 
804 	req->in.argpages = 1;
805 	req->num_pages = 1;
806 	req->pages[0] = page;
807 	req->page_offset = offset;
808 	nres = fuse_send_write(req, file, pos, count, NULL);
809 	err = req->out.h.error;
810 	fuse_put_request(fc, req);
811 	if (!err && !nres)
812 		err = -EIO;
813 	if (!err) {
814 		pos += nres;
815 		fuse_write_update_size(inode, pos);
816 		if (count == PAGE_CACHE_SIZE)
817 			SetPageUptodate(page);
818 	}
819 	fuse_invalidate_attr(inode);
820 	return err ? err : nres;
821 }
822 
823 static int fuse_write_end(struct file *file, struct address_space *mapping,
824 			loff_t pos, unsigned len, unsigned copied,
825 			struct page *page, void *fsdata)
826 {
827 	struct inode *inode = mapping->host;
828 	int res = 0;
829 
830 	if (copied)
831 		res = fuse_buffered_write(file, inode, pos, copied, page);
832 
833 	unlock_page(page);
834 	page_cache_release(page);
835 	return res;
836 }
837 
838 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
839 				    struct inode *inode, loff_t pos,
840 				    size_t count)
841 {
842 	size_t res;
843 	unsigned offset;
844 	unsigned i;
845 
846 	for (i = 0; i < req->num_pages; i++)
847 		fuse_wait_on_page_writeback(inode, req->pages[i]->index);
848 
849 	res = fuse_send_write(req, file, pos, count, NULL);
850 
851 	offset = req->page_offset;
852 	count = res;
853 	for (i = 0; i < req->num_pages; i++) {
854 		struct page *page = req->pages[i];
855 
856 		if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
857 			SetPageUptodate(page);
858 
859 		if (count > PAGE_CACHE_SIZE - offset)
860 			count -= PAGE_CACHE_SIZE - offset;
861 		else
862 			count = 0;
863 		offset = 0;
864 
865 		unlock_page(page);
866 		page_cache_release(page);
867 	}
868 
869 	return res;
870 }
871 
872 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
873 			       struct address_space *mapping,
874 			       struct iov_iter *ii, loff_t pos)
875 {
876 	struct fuse_conn *fc = get_fuse_conn(mapping->host);
877 	unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
878 	size_t count = 0;
879 	int err;
880 
881 	req->in.argpages = 1;
882 	req->page_offset = offset;
883 
884 	do {
885 		size_t tmp;
886 		struct page *page;
887 		pgoff_t index = pos >> PAGE_CACHE_SHIFT;
888 		size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
889 				     iov_iter_count(ii));
890 
891 		bytes = min_t(size_t, bytes, fc->max_write - count);
892 
893  again:
894 		err = -EFAULT;
895 		if (iov_iter_fault_in_readable(ii, bytes))
896 			break;
897 
898 		err = -ENOMEM;
899 		page = grab_cache_page_write_begin(mapping, index, 0);
900 		if (!page)
901 			break;
902 
903 		if (mapping_writably_mapped(mapping))
904 			flush_dcache_page(page);
905 
906 		pagefault_disable();
907 		tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
908 		pagefault_enable();
909 		flush_dcache_page(page);
910 
911 		if (!tmp) {
912 			unlock_page(page);
913 			page_cache_release(page);
914 			bytes = min(bytes, iov_iter_single_seg_count(ii));
915 			goto again;
916 		}
917 
918 		err = 0;
919 		req->pages[req->num_pages] = page;
920 		req->num_pages++;
921 
922 		iov_iter_advance(ii, tmp);
923 		count += tmp;
924 		pos += tmp;
925 		offset += tmp;
926 		if (offset == PAGE_CACHE_SIZE)
927 			offset = 0;
928 
929 		if (!fc->big_writes)
930 			break;
931 	} while (iov_iter_count(ii) && count < fc->max_write &&
932 		 req->num_pages < FUSE_MAX_PAGES_PER_REQ && offset == 0);
933 
934 	return count > 0 ? count : err;
935 }
936 
937 static ssize_t fuse_perform_write(struct file *file,
938 				  struct address_space *mapping,
939 				  struct iov_iter *ii, loff_t pos)
940 {
941 	struct inode *inode = mapping->host;
942 	struct fuse_conn *fc = get_fuse_conn(inode);
943 	int err = 0;
944 	ssize_t res = 0;
945 
946 	if (is_bad_inode(inode))
947 		return -EIO;
948 
949 	do {
950 		struct fuse_req *req;
951 		ssize_t count;
952 
953 		req = fuse_get_req(fc);
954 		if (IS_ERR(req)) {
955 			err = PTR_ERR(req);
956 			break;
957 		}
958 
959 		count = fuse_fill_write_pages(req, mapping, ii, pos);
960 		if (count <= 0) {
961 			err = count;
962 		} else {
963 			size_t num_written;
964 
965 			num_written = fuse_send_write_pages(req, file, inode,
966 							    pos, count);
967 			err = req->out.h.error;
968 			if (!err) {
969 				res += num_written;
970 				pos += num_written;
971 
972 				/* break out of the loop on short write */
973 				if (num_written != count)
974 					err = -EIO;
975 			}
976 		}
977 		fuse_put_request(fc, req);
978 	} while (!err && iov_iter_count(ii));
979 
980 	if (res > 0)
981 		fuse_write_update_size(inode, pos);
982 
983 	fuse_invalidate_attr(inode);
984 
985 	return res > 0 ? res : err;
986 }
987 
988 static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
989 				   unsigned long nr_segs, loff_t pos)
990 {
991 	struct file *file = iocb->ki_filp;
992 	struct address_space *mapping = file->f_mapping;
993 	size_t count = 0;
994 	ssize_t written = 0;
995 	struct inode *inode = mapping->host;
996 	ssize_t err;
997 	struct iov_iter i;
998 
999 	WARN_ON(iocb->ki_pos != pos);
1000 
1001 	err = generic_segment_checks(iov, &nr_segs, &count, VERIFY_READ);
1002 	if (err)
1003 		return err;
1004 
1005 	mutex_lock(&inode->i_mutex);
1006 	vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
1007 
1008 	/* We can write back this queue in page reclaim */
1009 	current->backing_dev_info = mapping->backing_dev_info;
1010 
1011 	err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
1012 	if (err)
1013 		goto out;
1014 
1015 	if (count == 0)
1016 		goto out;
1017 
1018 	err = file_remove_suid(file);
1019 	if (err)
1020 		goto out;
1021 
1022 	file_update_time(file);
1023 
1024 	iov_iter_init(&i, iov, nr_segs, count, 0);
1025 	written = fuse_perform_write(file, mapping, &i, pos);
1026 	if (written >= 0)
1027 		iocb->ki_pos = pos + written;
1028 
1029 out:
1030 	current->backing_dev_info = NULL;
1031 	mutex_unlock(&inode->i_mutex);
1032 
1033 	return written ? written : err;
1034 }
1035 
1036 static void fuse_release_user_pages(struct fuse_req *req, int write)
1037 {
1038 	unsigned i;
1039 
1040 	for (i = 0; i < req->num_pages; i++) {
1041 		struct page *page = req->pages[i];
1042 		if (write)
1043 			set_page_dirty_lock(page);
1044 		put_page(page);
1045 	}
1046 }
1047 
1048 static int fuse_get_user_pages(struct fuse_req *req, const char __user *buf,
1049 			       size_t *nbytesp, int write)
1050 {
1051 	size_t nbytes = *nbytesp;
1052 	unsigned long user_addr = (unsigned long) buf;
1053 	unsigned offset = user_addr & ~PAGE_MASK;
1054 	int npages;
1055 
1056 	/* Special case for kernel I/O: can copy directly into the buffer */
1057 	if (segment_eq(get_fs(), KERNEL_DS)) {
1058 		if (write)
1059 			req->in.args[1].value = (void *) user_addr;
1060 		else
1061 			req->out.args[0].value = (void *) user_addr;
1062 
1063 		return 0;
1064 	}
1065 
1066 	nbytes = min_t(size_t, nbytes, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
1067 	npages = (nbytes + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1068 	npages = clamp(npages, 1, FUSE_MAX_PAGES_PER_REQ);
1069 	npages = get_user_pages_fast(user_addr, npages, !write, req->pages);
1070 	if (npages < 0)
1071 		return npages;
1072 
1073 	req->num_pages = npages;
1074 	req->page_offset = offset;
1075 
1076 	if (write)
1077 		req->in.argpages = 1;
1078 	else
1079 		req->out.argpages = 1;
1080 
1081 	nbytes = (req->num_pages << PAGE_SHIFT) - req->page_offset;
1082 	*nbytesp = min(*nbytesp, nbytes);
1083 
1084 	return 0;
1085 }
1086 
1087 ssize_t fuse_direct_io(struct file *file, const char __user *buf,
1088 		       size_t count, loff_t *ppos, int write)
1089 {
1090 	struct fuse_file *ff = file->private_data;
1091 	struct fuse_conn *fc = ff->fc;
1092 	size_t nmax = write ? fc->max_write : fc->max_read;
1093 	loff_t pos = *ppos;
1094 	ssize_t res = 0;
1095 	struct fuse_req *req;
1096 
1097 	req = fuse_get_req(fc);
1098 	if (IS_ERR(req))
1099 		return PTR_ERR(req);
1100 
1101 	while (count) {
1102 		size_t nres;
1103 		fl_owner_t owner = current->files;
1104 		size_t nbytes = min(count, nmax);
1105 		int err = fuse_get_user_pages(req, buf, &nbytes, write);
1106 		if (err) {
1107 			res = err;
1108 			break;
1109 		}
1110 
1111 		if (write)
1112 			nres = fuse_send_write(req, file, pos, nbytes, owner);
1113 		else
1114 			nres = fuse_send_read(req, file, pos, nbytes, owner);
1115 
1116 		fuse_release_user_pages(req, !write);
1117 		if (req->out.h.error) {
1118 			if (!res)
1119 				res = req->out.h.error;
1120 			break;
1121 		} else if (nres > nbytes) {
1122 			res = -EIO;
1123 			break;
1124 		}
1125 		count -= nres;
1126 		res += nres;
1127 		pos += nres;
1128 		buf += nres;
1129 		if (nres != nbytes)
1130 			break;
1131 		if (count) {
1132 			fuse_put_request(fc, req);
1133 			req = fuse_get_req(fc);
1134 			if (IS_ERR(req))
1135 				break;
1136 		}
1137 	}
1138 	if (!IS_ERR(req))
1139 		fuse_put_request(fc, req);
1140 	if (res > 0)
1141 		*ppos = pos;
1142 
1143 	return res;
1144 }
1145 EXPORT_SYMBOL_GPL(fuse_direct_io);
1146 
1147 static ssize_t fuse_direct_read(struct file *file, char __user *buf,
1148 				     size_t count, loff_t *ppos)
1149 {
1150 	ssize_t res;
1151 	struct inode *inode = file->f_path.dentry->d_inode;
1152 
1153 	if (is_bad_inode(inode))
1154 		return -EIO;
1155 
1156 	res = fuse_direct_io(file, buf, count, ppos, 0);
1157 
1158 	fuse_invalidate_attr(inode);
1159 
1160 	return res;
1161 }
1162 
1163 static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
1164 				 size_t count, loff_t *ppos)
1165 {
1166 	struct inode *inode = file->f_path.dentry->d_inode;
1167 	ssize_t res;
1168 
1169 	if (is_bad_inode(inode))
1170 		return -EIO;
1171 
1172 	/* Don't allow parallel writes to the same file */
1173 	mutex_lock(&inode->i_mutex);
1174 	res = generic_write_checks(file, ppos, &count, 0);
1175 	if (!res) {
1176 		res = fuse_direct_io(file, buf, count, ppos, 1);
1177 		if (res > 0)
1178 			fuse_write_update_size(inode, *ppos);
1179 	}
1180 	mutex_unlock(&inode->i_mutex);
1181 
1182 	fuse_invalidate_attr(inode);
1183 
1184 	return res;
1185 }
1186 
1187 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1188 {
1189 	__free_page(req->pages[0]);
1190 	fuse_file_put(req->ff, false);
1191 }
1192 
1193 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1194 {
1195 	struct inode *inode = req->inode;
1196 	struct fuse_inode *fi = get_fuse_inode(inode);
1197 	struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
1198 
1199 	list_del(&req->writepages_entry);
1200 	dec_bdi_stat(bdi, BDI_WRITEBACK);
1201 	dec_zone_page_state(req->pages[0], NR_WRITEBACK_TEMP);
1202 	bdi_writeout_inc(bdi);
1203 	wake_up(&fi->page_waitq);
1204 }
1205 
1206 /* Called under fc->lock, may release and reacquire it */
1207 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req)
1208 __releases(fc->lock)
1209 __acquires(fc->lock)
1210 {
1211 	struct fuse_inode *fi = get_fuse_inode(req->inode);
1212 	loff_t size = i_size_read(req->inode);
1213 	struct fuse_write_in *inarg = &req->misc.write.in;
1214 
1215 	if (!fc->connected)
1216 		goto out_free;
1217 
1218 	if (inarg->offset + PAGE_CACHE_SIZE <= size) {
1219 		inarg->size = PAGE_CACHE_SIZE;
1220 	} else if (inarg->offset < size) {
1221 		inarg->size = size & (PAGE_CACHE_SIZE - 1);
1222 	} else {
1223 		/* Got truncated off completely */
1224 		goto out_free;
1225 	}
1226 
1227 	req->in.args[1].size = inarg->size;
1228 	fi->writectr++;
1229 	fuse_request_send_background_locked(fc, req);
1230 	return;
1231 
1232  out_free:
1233 	fuse_writepage_finish(fc, req);
1234 	spin_unlock(&fc->lock);
1235 	fuse_writepage_free(fc, req);
1236 	fuse_put_request(fc, req);
1237 	spin_lock(&fc->lock);
1238 }
1239 
1240 /*
1241  * If fi->writectr is positive (no truncate or fsync going on) send
1242  * all queued writepage requests.
1243  *
1244  * Called with fc->lock
1245  */
1246 void fuse_flush_writepages(struct inode *inode)
1247 __releases(fc->lock)
1248 __acquires(fc->lock)
1249 {
1250 	struct fuse_conn *fc = get_fuse_conn(inode);
1251 	struct fuse_inode *fi = get_fuse_inode(inode);
1252 	struct fuse_req *req;
1253 
1254 	while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1255 		req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1256 		list_del_init(&req->list);
1257 		fuse_send_writepage(fc, req);
1258 	}
1259 }
1260 
1261 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1262 {
1263 	struct inode *inode = req->inode;
1264 	struct fuse_inode *fi = get_fuse_inode(inode);
1265 
1266 	mapping_set_error(inode->i_mapping, req->out.h.error);
1267 	spin_lock(&fc->lock);
1268 	fi->writectr--;
1269 	fuse_writepage_finish(fc, req);
1270 	spin_unlock(&fc->lock);
1271 	fuse_writepage_free(fc, req);
1272 }
1273 
1274 static int fuse_writepage_locked(struct page *page)
1275 {
1276 	struct address_space *mapping = page->mapping;
1277 	struct inode *inode = mapping->host;
1278 	struct fuse_conn *fc = get_fuse_conn(inode);
1279 	struct fuse_inode *fi = get_fuse_inode(inode);
1280 	struct fuse_req *req;
1281 	struct fuse_file *ff;
1282 	struct page *tmp_page;
1283 
1284 	set_page_writeback(page);
1285 
1286 	req = fuse_request_alloc_nofs();
1287 	if (!req)
1288 		goto err;
1289 
1290 	tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1291 	if (!tmp_page)
1292 		goto err_free;
1293 
1294 	spin_lock(&fc->lock);
1295 	BUG_ON(list_empty(&fi->write_files));
1296 	ff = list_entry(fi->write_files.next, struct fuse_file, write_entry);
1297 	req->ff = fuse_file_get(ff);
1298 	spin_unlock(&fc->lock);
1299 
1300 	fuse_write_fill(req, ff, page_offset(page), 0);
1301 
1302 	copy_highpage(tmp_page, page);
1303 	req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1304 	req->in.argpages = 1;
1305 	req->num_pages = 1;
1306 	req->pages[0] = tmp_page;
1307 	req->page_offset = 0;
1308 	req->end = fuse_writepage_end;
1309 	req->inode = inode;
1310 
1311 	inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
1312 	inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1313 	end_page_writeback(page);
1314 
1315 	spin_lock(&fc->lock);
1316 	list_add(&req->writepages_entry, &fi->writepages);
1317 	list_add_tail(&req->list, &fi->queued_writes);
1318 	fuse_flush_writepages(inode);
1319 	spin_unlock(&fc->lock);
1320 
1321 	return 0;
1322 
1323 err_free:
1324 	fuse_request_free(req);
1325 err:
1326 	end_page_writeback(page);
1327 	return -ENOMEM;
1328 }
1329 
1330 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1331 {
1332 	int err;
1333 
1334 	err = fuse_writepage_locked(page);
1335 	unlock_page(page);
1336 
1337 	return err;
1338 }
1339 
1340 static int fuse_launder_page(struct page *page)
1341 {
1342 	int err = 0;
1343 	if (clear_page_dirty_for_io(page)) {
1344 		struct inode *inode = page->mapping->host;
1345 		err = fuse_writepage_locked(page);
1346 		if (!err)
1347 			fuse_wait_on_page_writeback(inode, page->index);
1348 	}
1349 	return err;
1350 }
1351 
1352 /*
1353  * Write back dirty pages now, because there may not be any suitable
1354  * open files later
1355  */
1356 static void fuse_vma_close(struct vm_area_struct *vma)
1357 {
1358 	filemap_write_and_wait(vma->vm_file->f_mapping);
1359 }
1360 
1361 /*
1362  * Wait for writeback against this page to complete before allowing it
1363  * to be marked dirty again, and hence written back again, possibly
1364  * before the previous writepage completed.
1365  *
1366  * Block here, instead of in ->writepage(), so that the userspace fs
1367  * can only block processes actually operating on the filesystem.
1368  *
1369  * Otherwise unprivileged userspace fs would be able to block
1370  * unrelated:
1371  *
1372  * - page migration
1373  * - sync(2)
1374  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
1375  */
1376 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1377 {
1378 	struct page *page = vmf->page;
1379 	/*
1380 	 * Don't use page->mapping as it may become NULL from a
1381 	 * concurrent truncate.
1382 	 */
1383 	struct inode *inode = vma->vm_file->f_mapping->host;
1384 
1385 	fuse_wait_on_page_writeback(inode, page->index);
1386 	return 0;
1387 }
1388 
1389 static const struct vm_operations_struct fuse_file_vm_ops = {
1390 	.close		= fuse_vma_close,
1391 	.fault		= filemap_fault,
1392 	.page_mkwrite	= fuse_page_mkwrite,
1393 };
1394 
1395 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
1396 {
1397 	if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE)) {
1398 		struct inode *inode = file->f_dentry->d_inode;
1399 		struct fuse_conn *fc = get_fuse_conn(inode);
1400 		struct fuse_inode *fi = get_fuse_inode(inode);
1401 		struct fuse_file *ff = file->private_data;
1402 		/*
1403 		 * file may be written through mmap, so chain it onto the
1404 		 * inodes's write_file list
1405 		 */
1406 		spin_lock(&fc->lock);
1407 		if (list_empty(&ff->write_entry))
1408 			list_add(&ff->write_entry, &fi->write_files);
1409 		spin_unlock(&fc->lock);
1410 	}
1411 	file_accessed(file);
1412 	vma->vm_ops = &fuse_file_vm_ops;
1413 	return 0;
1414 }
1415 
1416 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
1417 {
1418 	/* Can't provide the coherency needed for MAP_SHARED */
1419 	if (vma->vm_flags & VM_MAYSHARE)
1420 		return -ENODEV;
1421 
1422 	invalidate_inode_pages2(file->f_mapping);
1423 
1424 	return generic_file_mmap(file, vma);
1425 }
1426 
1427 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
1428 				  struct file_lock *fl)
1429 {
1430 	switch (ffl->type) {
1431 	case F_UNLCK:
1432 		break;
1433 
1434 	case F_RDLCK:
1435 	case F_WRLCK:
1436 		if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
1437 		    ffl->end < ffl->start)
1438 			return -EIO;
1439 
1440 		fl->fl_start = ffl->start;
1441 		fl->fl_end = ffl->end;
1442 		fl->fl_pid = ffl->pid;
1443 		break;
1444 
1445 	default:
1446 		return -EIO;
1447 	}
1448 	fl->fl_type = ffl->type;
1449 	return 0;
1450 }
1451 
1452 static void fuse_lk_fill(struct fuse_req *req, struct file *file,
1453 			 const struct file_lock *fl, int opcode, pid_t pid,
1454 			 int flock)
1455 {
1456 	struct inode *inode = file->f_path.dentry->d_inode;
1457 	struct fuse_conn *fc = get_fuse_conn(inode);
1458 	struct fuse_file *ff = file->private_data;
1459 	struct fuse_lk_in *arg = &req->misc.lk_in;
1460 
1461 	arg->fh = ff->fh;
1462 	arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
1463 	arg->lk.start = fl->fl_start;
1464 	arg->lk.end = fl->fl_end;
1465 	arg->lk.type = fl->fl_type;
1466 	arg->lk.pid = pid;
1467 	if (flock)
1468 		arg->lk_flags |= FUSE_LK_FLOCK;
1469 	req->in.h.opcode = opcode;
1470 	req->in.h.nodeid = get_node_id(inode);
1471 	req->in.numargs = 1;
1472 	req->in.args[0].size = sizeof(*arg);
1473 	req->in.args[0].value = arg;
1474 }
1475 
1476 static int fuse_getlk(struct file *file, struct file_lock *fl)
1477 {
1478 	struct inode *inode = file->f_path.dentry->d_inode;
1479 	struct fuse_conn *fc = get_fuse_conn(inode);
1480 	struct fuse_req *req;
1481 	struct fuse_lk_out outarg;
1482 	int err;
1483 
1484 	req = fuse_get_req(fc);
1485 	if (IS_ERR(req))
1486 		return PTR_ERR(req);
1487 
1488 	fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
1489 	req->out.numargs = 1;
1490 	req->out.args[0].size = sizeof(outarg);
1491 	req->out.args[0].value = &outarg;
1492 	fuse_request_send(fc, req);
1493 	err = req->out.h.error;
1494 	fuse_put_request(fc, req);
1495 	if (!err)
1496 		err = convert_fuse_file_lock(&outarg.lk, fl);
1497 
1498 	return err;
1499 }
1500 
1501 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
1502 {
1503 	struct inode *inode = file->f_path.dentry->d_inode;
1504 	struct fuse_conn *fc = get_fuse_conn(inode);
1505 	struct fuse_req *req;
1506 	int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
1507 	pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
1508 	int err;
1509 
1510 	if (fl->fl_lmops && fl->fl_lmops->fl_grant) {
1511 		/* NLM needs asynchronous locks, which we don't support yet */
1512 		return -ENOLCK;
1513 	}
1514 
1515 	/* Unlock on close is handled by the flush method */
1516 	if (fl->fl_flags & FL_CLOSE)
1517 		return 0;
1518 
1519 	req = fuse_get_req(fc);
1520 	if (IS_ERR(req))
1521 		return PTR_ERR(req);
1522 
1523 	fuse_lk_fill(req, file, fl, opcode, pid, flock);
1524 	fuse_request_send(fc, req);
1525 	err = req->out.h.error;
1526 	/* locking is restartable */
1527 	if (err == -EINTR)
1528 		err = -ERESTARTSYS;
1529 	fuse_put_request(fc, req);
1530 	return err;
1531 }
1532 
1533 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
1534 {
1535 	struct inode *inode = file->f_path.dentry->d_inode;
1536 	struct fuse_conn *fc = get_fuse_conn(inode);
1537 	int err;
1538 
1539 	if (cmd == F_CANCELLK) {
1540 		err = 0;
1541 	} else if (cmd == F_GETLK) {
1542 		if (fc->no_lock) {
1543 			posix_test_lock(file, fl);
1544 			err = 0;
1545 		} else
1546 			err = fuse_getlk(file, fl);
1547 	} else {
1548 		if (fc->no_lock)
1549 			err = posix_lock_file(file, fl, NULL);
1550 		else
1551 			err = fuse_setlk(file, fl, 0);
1552 	}
1553 	return err;
1554 }
1555 
1556 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
1557 {
1558 	struct inode *inode = file->f_path.dentry->d_inode;
1559 	struct fuse_conn *fc = get_fuse_conn(inode);
1560 	int err;
1561 
1562 	if (fc->no_lock) {
1563 		err = flock_lock_file_wait(file, fl);
1564 	} else {
1565 		/* emulate flock with POSIX locks */
1566 		fl->fl_owner = (fl_owner_t) file;
1567 		err = fuse_setlk(file, fl, 1);
1568 	}
1569 
1570 	return err;
1571 }
1572 
1573 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
1574 {
1575 	struct inode *inode = mapping->host;
1576 	struct fuse_conn *fc = get_fuse_conn(inode);
1577 	struct fuse_req *req;
1578 	struct fuse_bmap_in inarg;
1579 	struct fuse_bmap_out outarg;
1580 	int err;
1581 
1582 	if (!inode->i_sb->s_bdev || fc->no_bmap)
1583 		return 0;
1584 
1585 	req = fuse_get_req(fc);
1586 	if (IS_ERR(req))
1587 		return 0;
1588 
1589 	memset(&inarg, 0, sizeof(inarg));
1590 	inarg.block = block;
1591 	inarg.blocksize = inode->i_sb->s_blocksize;
1592 	req->in.h.opcode = FUSE_BMAP;
1593 	req->in.h.nodeid = get_node_id(inode);
1594 	req->in.numargs = 1;
1595 	req->in.args[0].size = sizeof(inarg);
1596 	req->in.args[0].value = &inarg;
1597 	req->out.numargs = 1;
1598 	req->out.args[0].size = sizeof(outarg);
1599 	req->out.args[0].value = &outarg;
1600 	fuse_request_send(fc, req);
1601 	err = req->out.h.error;
1602 	fuse_put_request(fc, req);
1603 	if (err == -ENOSYS)
1604 		fc->no_bmap = 1;
1605 
1606 	return err ? 0 : outarg.block;
1607 }
1608 
1609 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int origin)
1610 {
1611 	loff_t retval;
1612 	struct inode *inode = file->f_path.dentry->d_inode;
1613 
1614 	mutex_lock(&inode->i_mutex);
1615 	if (origin != SEEK_CUR || origin != SEEK_SET) {
1616 		retval = fuse_update_attributes(inode, NULL, file, NULL);
1617 		if (retval)
1618 			goto exit;
1619 	}
1620 
1621 	switch (origin) {
1622 	case SEEK_END:
1623 		offset += i_size_read(inode);
1624 		break;
1625 	case SEEK_CUR:
1626 		offset += file->f_pos;
1627 		break;
1628 	case SEEK_DATA:
1629 		if (offset >= i_size_read(inode)) {
1630 			retval = -ENXIO;
1631 			goto exit;
1632 		}
1633 		break;
1634 	case SEEK_HOLE:
1635 		if (offset >= i_size_read(inode)) {
1636 			retval = -ENXIO;
1637 			goto exit;
1638 		}
1639 		offset = i_size_read(inode);
1640 		break;
1641 	}
1642 	retval = -EINVAL;
1643 	if (offset >= 0 && offset <= inode->i_sb->s_maxbytes) {
1644 		if (offset != file->f_pos) {
1645 			file->f_pos = offset;
1646 			file->f_version = 0;
1647 		}
1648 		retval = offset;
1649 	}
1650 exit:
1651 	mutex_unlock(&inode->i_mutex);
1652 	return retval;
1653 }
1654 
1655 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
1656 			unsigned int nr_segs, size_t bytes, bool to_user)
1657 {
1658 	struct iov_iter ii;
1659 	int page_idx = 0;
1660 
1661 	if (!bytes)
1662 		return 0;
1663 
1664 	iov_iter_init(&ii, iov, nr_segs, bytes, 0);
1665 
1666 	while (iov_iter_count(&ii)) {
1667 		struct page *page = pages[page_idx++];
1668 		size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
1669 		void *kaddr;
1670 
1671 		kaddr = kmap(page);
1672 
1673 		while (todo) {
1674 			char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
1675 			size_t iov_len = ii.iov->iov_len - ii.iov_offset;
1676 			size_t copy = min(todo, iov_len);
1677 			size_t left;
1678 
1679 			if (!to_user)
1680 				left = copy_from_user(kaddr, uaddr, copy);
1681 			else
1682 				left = copy_to_user(uaddr, kaddr, copy);
1683 
1684 			if (unlikely(left))
1685 				return -EFAULT;
1686 
1687 			iov_iter_advance(&ii, copy);
1688 			todo -= copy;
1689 			kaddr += copy;
1690 		}
1691 
1692 		kunmap(page);
1693 	}
1694 
1695 	return 0;
1696 }
1697 
1698 /*
1699  * CUSE servers compiled on 32bit broke on 64bit kernels because the
1700  * ABI was defined to be 'struct iovec' which is different on 32bit
1701  * and 64bit.  Fortunately we can determine which structure the server
1702  * used from the size of the reply.
1703  */
1704 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
1705 				     size_t transferred, unsigned count,
1706 				     bool is_compat)
1707 {
1708 #ifdef CONFIG_COMPAT
1709 	if (count * sizeof(struct compat_iovec) == transferred) {
1710 		struct compat_iovec *ciov = src;
1711 		unsigned i;
1712 
1713 		/*
1714 		 * With this interface a 32bit server cannot support
1715 		 * non-compat (i.e. ones coming from 64bit apps) ioctl
1716 		 * requests
1717 		 */
1718 		if (!is_compat)
1719 			return -EINVAL;
1720 
1721 		for (i = 0; i < count; i++) {
1722 			dst[i].iov_base = compat_ptr(ciov[i].iov_base);
1723 			dst[i].iov_len = ciov[i].iov_len;
1724 		}
1725 		return 0;
1726 	}
1727 #endif
1728 
1729 	if (count * sizeof(struct iovec) != transferred)
1730 		return -EIO;
1731 
1732 	memcpy(dst, src, transferred);
1733 	return 0;
1734 }
1735 
1736 /* Make sure iov_length() won't overflow */
1737 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
1738 {
1739 	size_t n;
1740 	u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
1741 
1742 	for (n = 0; n < count; n++) {
1743 		if (iov->iov_len > (size_t) max)
1744 			return -ENOMEM;
1745 		max -= iov->iov_len;
1746 	}
1747 	return 0;
1748 }
1749 
1750 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
1751 				 void *src, size_t transferred, unsigned count,
1752 				 bool is_compat)
1753 {
1754 	unsigned i;
1755 	struct fuse_ioctl_iovec *fiov = src;
1756 
1757 	if (fc->minor < 16) {
1758 		return fuse_copy_ioctl_iovec_old(dst, src, transferred,
1759 						 count, is_compat);
1760 	}
1761 
1762 	if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
1763 		return -EIO;
1764 
1765 	for (i = 0; i < count; i++) {
1766 		/* Did the server supply an inappropriate value? */
1767 		if (fiov[i].base != (unsigned long) fiov[i].base ||
1768 		    fiov[i].len != (unsigned long) fiov[i].len)
1769 			return -EIO;
1770 
1771 		dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
1772 		dst[i].iov_len = (size_t) fiov[i].len;
1773 
1774 #ifdef CONFIG_COMPAT
1775 		if (is_compat &&
1776 		    (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
1777 		     (compat_size_t) dst[i].iov_len != fiov[i].len))
1778 			return -EIO;
1779 #endif
1780 	}
1781 
1782 	return 0;
1783 }
1784 
1785 
1786 /*
1787  * For ioctls, there is no generic way to determine how much memory
1788  * needs to be read and/or written.  Furthermore, ioctls are allowed
1789  * to dereference the passed pointer, so the parameter requires deep
1790  * copying but FUSE has no idea whatsoever about what to copy in or
1791  * out.
1792  *
1793  * This is solved by allowing FUSE server to retry ioctl with
1794  * necessary in/out iovecs.  Let's assume the ioctl implementation
1795  * needs to read in the following structure.
1796  *
1797  * struct a {
1798  *	char	*buf;
1799  *	size_t	buflen;
1800  * }
1801  *
1802  * On the first callout to FUSE server, inarg->in_size and
1803  * inarg->out_size will be NULL; then, the server completes the ioctl
1804  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
1805  * the actual iov array to
1806  *
1807  * { { .iov_base = inarg.arg,	.iov_len = sizeof(struct a) } }
1808  *
1809  * which tells FUSE to copy in the requested area and retry the ioctl.
1810  * On the second round, the server has access to the structure and
1811  * from that it can tell what to look for next, so on the invocation,
1812  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
1813  *
1814  * { { .iov_base = inarg.arg,	.iov_len = sizeof(struct a)	},
1815  *   { .iov_base = a.buf,	.iov_len = a.buflen		} }
1816  *
1817  * FUSE will copy both struct a and the pointed buffer from the
1818  * process doing the ioctl and retry ioctl with both struct a and the
1819  * buffer.
1820  *
1821  * This time, FUSE server has everything it needs and completes ioctl
1822  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
1823  *
1824  * Copying data out works the same way.
1825  *
1826  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
1827  * automatically initializes in and out iovs by decoding @cmd with
1828  * _IOC_* macros and the server is not allowed to request RETRY.  This
1829  * limits ioctl data transfers to well-formed ioctls and is the forced
1830  * behavior for all FUSE servers.
1831  */
1832 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
1833 		   unsigned int flags)
1834 {
1835 	struct fuse_file *ff = file->private_data;
1836 	struct fuse_conn *fc = ff->fc;
1837 	struct fuse_ioctl_in inarg = {
1838 		.fh = ff->fh,
1839 		.cmd = cmd,
1840 		.arg = arg,
1841 		.flags = flags
1842 	};
1843 	struct fuse_ioctl_out outarg;
1844 	struct fuse_req *req = NULL;
1845 	struct page **pages = NULL;
1846 	struct iovec *iov_page = NULL;
1847 	struct iovec *in_iov = NULL, *out_iov = NULL;
1848 	unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
1849 	size_t in_size, out_size, transferred;
1850 	int err;
1851 
1852 #if BITS_PER_LONG == 32
1853 	inarg.flags |= FUSE_IOCTL_32BIT;
1854 #else
1855 	if (flags & FUSE_IOCTL_COMPAT)
1856 		inarg.flags |= FUSE_IOCTL_32BIT;
1857 #endif
1858 
1859 	/* assume all the iovs returned by client always fits in a page */
1860 	BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
1861 
1862 	err = -ENOMEM;
1863 	pages = kzalloc(sizeof(pages[0]) * FUSE_MAX_PAGES_PER_REQ, GFP_KERNEL);
1864 	iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
1865 	if (!pages || !iov_page)
1866 		goto out;
1867 
1868 	/*
1869 	 * If restricted, initialize IO parameters as encoded in @cmd.
1870 	 * RETRY from server is not allowed.
1871 	 */
1872 	if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
1873 		struct iovec *iov = iov_page;
1874 
1875 		iov->iov_base = (void __user *)arg;
1876 		iov->iov_len = _IOC_SIZE(cmd);
1877 
1878 		if (_IOC_DIR(cmd) & _IOC_WRITE) {
1879 			in_iov = iov;
1880 			in_iovs = 1;
1881 		}
1882 
1883 		if (_IOC_DIR(cmd) & _IOC_READ) {
1884 			out_iov = iov;
1885 			out_iovs = 1;
1886 		}
1887 	}
1888 
1889  retry:
1890 	inarg.in_size = in_size = iov_length(in_iov, in_iovs);
1891 	inarg.out_size = out_size = iov_length(out_iov, out_iovs);
1892 
1893 	/*
1894 	 * Out data can be used either for actual out data or iovs,
1895 	 * make sure there always is at least one page.
1896 	 */
1897 	out_size = max_t(size_t, out_size, PAGE_SIZE);
1898 	max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
1899 
1900 	/* make sure there are enough buffer pages and init request with them */
1901 	err = -ENOMEM;
1902 	if (max_pages > FUSE_MAX_PAGES_PER_REQ)
1903 		goto out;
1904 	while (num_pages < max_pages) {
1905 		pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
1906 		if (!pages[num_pages])
1907 			goto out;
1908 		num_pages++;
1909 	}
1910 
1911 	req = fuse_get_req(fc);
1912 	if (IS_ERR(req)) {
1913 		err = PTR_ERR(req);
1914 		req = NULL;
1915 		goto out;
1916 	}
1917 	memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
1918 	req->num_pages = num_pages;
1919 
1920 	/* okay, let's send it to the client */
1921 	req->in.h.opcode = FUSE_IOCTL;
1922 	req->in.h.nodeid = ff->nodeid;
1923 	req->in.numargs = 1;
1924 	req->in.args[0].size = sizeof(inarg);
1925 	req->in.args[0].value = &inarg;
1926 	if (in_size) {
1927 		req->in.numargs++;
1928 		req->in.args[1].size = in_size;
1929 		req->in.argpages = 1;
1930 
1931 		err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
1932 					   false);
1933 		if (err)
1934 			goto out;
1935 	}
1936 
1937 	req->out.numargs = 2;
1938 	req->out.args[0].size = sizeof(outarg);
1939 	req->out.args[0].value = &outarg;
1940 	req->out.args[1].size = out_size;
1941 	req->out.argpages = 1;
1942 	req->out.argvar = 1;
1943 
1944 	fuse_request_send(fc, req);
1945 	err = req->out.h.error;
1946 	transferred = req->out.args[1].size;
1947 	fuse_put_request(fc, req);
1948 	req = NULL;
1949 	if (err)
1950 		goto out;
1951 
1952 	/* did it ask for retry? */
1953 	if (outarg.flags & FUSE_IOCTL_RETRY) {
1954 		void *vaddr;
1955 
1956 		/* no retry if in restricted mode */
1957 		err = -EIO;
1958 		if (!(flags & FUSE_IOCTL_UNRESTRICTED))
1959 			goto out;
1960 
1961 		in_iovs = outarg.in_iovs;
1962 		out_iovs = outarg.out_iovs;
1963 
1964 		/*
1965 		 * Make sure things are in boundary, separate checks
1966 		 * are to protect against overflow.
1967 		 */
1968 		err = -ENOMEM;
1969 		if (in_iovs > FUSE_IOCTL_MAX_IOV ||
1970 		    out_iovs > FUSE_IOCTL_MAX_IOV ||
1971 		    in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
1972 			goto out;
1973 
1974 		vaddr = kmap_atomic(pages[0], KM_USER0);
1975 		err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
1976 					    transferred, in_iovs + out_iovs,
1977 					    (flags & FUSE_IOCTL_COMPAT) != 0);
1978 		kunmap_atomic(vaddr, KM_USER0);
1979 		if (err)
1980 			goto out;
1981 
1982 		in_iov = iov_page;
1983 		out_iov = in_iov + in_iovs;
1984 
1985 		err = fuse_verify_ioctl_iov(in_iov, in_iovs);
1986 		if (err)
1987 			goto out;
1988 
1989 		err = fuse_verify_ioctl_iov(out_iov, out_iovs);
1990 		if (err)
1991 			goto out;
1992 
1993 		goto retry;
1994 	}
1995 
1996 	err = -EIO;
1997 	if (transferred > inarg.out_size)
1998 		goto out;
1999 
2000 	err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
2001  out:
2002 	if (req)
2003 		fuse_put_request(fc, req);
2004 	free_page((unsigned long) iov_page);
2005 	while (num_pages)
2006 		__free_page(pages[--num_pages]);
2007 	kfree(pages);
2008 
2009 	return err ? err : outarg.result;
2010 }
2011 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2012 
2013 static long fuse_file_ioctl_common(struct file *file, unsigned int cmd,
2014 				   unsigned long arg, unsigned int flags)
2015 {
2016 	struct inode *inode = file->f_dentry->d_inode;
2017 	struct fuse_conn *fc = get_fuse_conn(inode);
2018 
2019 	if (!fuse_allow_task(fc, current))
2020 		return -EACCES;
2021 
2022 	if (is_bad_inode(inode))
2023 		return -EIO;
2024 
2025 	return fuse_do_ioctl(file, cmd, arg, flags);
2026 }
2027 
2028 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2029 			    unsigned long arg)
2030 {
2031 	return fuse_file_ioctl_common(file, cmd, arg, 0);
2032 }
2033 
2034 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2035 				   unsigned long arg)
2036 {
2037 	return fuse_file_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2038 }
2039 
2040 /*
2041  * All files which have been polled are linked to RB tree
2042  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2043  * find the matching one.
2044  */
2045 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2046 					      struct rb_node **parent_out)
2047 {
2048 	struct rb_node **link = &fc->polled_files.rb_node;
2049 	struct rb_node *last = NULL;
2050 
2051 	while (*link) {
2052 		struct fuse_file *ff;
2053 
2054 		last = *link;
2055 		ff = rb_entry(last, struct fuse_file, polled_node);
2056 
2057 		if (kh < ff->kh)
2058 			link = &last->rb_left;
2059 		else if (kh > ff->kh)
2060 			link = &last->rb_right;
2061 		else
2062 			return link;
2063 	}
2064 
2065 	if (parent_out)
2066 		*parent_out = last;
2067 	return link;
2068 }
2069 
2070 /*
2071  * The file is about to be polled.  Make sure it's on the polled_files
2072  * RB tree.  Note that files once added to the polled_files tree are
2073  * not removed before the file is released.  This is because a file
2074  * polled once is likely to be polled again.
2075  */
2076 static void fuse_register_polled_file(struct fuse_conn *fc,
2077 				      struct fuse_file *ff)
2078 {
2079 	spin_lock(&fc->lock);
2080 	if (RB_EMPTY_NODE(&ff->polled_node)) {
2081 		struct rb_node **link, *parent;
2082 
2083 		link = fuse_find_polled_node(fc, ff->kh, &parent);
2084 		BUG_ON(*link);
2085 		rb_link_node(&ff->polled_node, parent, link);
2086 		rb_insert_color(&ff->polled_node, &fc->polled_files);
2087 	}
2088 	spin_unlock(&fc->lock);
2089 }
2090 
2091 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2092 {
2093 	struct fuse_file *ff = file->private_data;
2094 	struct fuse_conn *fc = ff->fc;
2095 	struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2096 	struct fuse_poll_out outarg;
2097 	struct fuse_req *req;
2098 	int err;
2099 
2100 	if (fc->no_poll)
2101 		return DEFAULT_POLLMASK;
2102 
2103 	poll_wait(file, &ff->poll_wait, wait);
2104 
2105 	/*
2106 	 * Ask for notification iff there's someone waiting for it.
2107 	 * The client may ignore the flag and always notify.
2108 	 */
2109 	if (waitqueue_active(&ff->poll_wait)) {
2110 		inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2111 		fuse_register_polled_file(fc, ff);
2112 	}
2113 
2114 	req = fuse_get_req(fc);
2115 	if (IS_ERR(req))
2116 		return POLLERR;
2117 
2118 	req->in.h.opcode = FUSE_POLL;
2119 	req->in.h.nodeid = ff->nodeid;
2120 	req->in.numargs = 1;
2121 	req->in.args[0].size = sizeof(inarg);
2122 	req->in.args[0].value = &inarg;
2123 	req->out.numargs = 1;
2124 	req->out.args[0].size = sizeof(outarg);
2125 	req->out.args[0].value = &outarg;
2126 	fuse_request_send(fc, req);
2127 	err = req->out.h.error;
2128 	fuse_put_request(fc, req);
2129 
2130 	if (!err)
2131 		return outarg.revents;
2132 	if (err == -ENOSYS) {
2133 		fc->no_poll = 1;
2134 		return DEFAULT_POLLMASK;
2135 	}
2136 	return POLLERR;
2137 }
2138 EXPORT_SYMBOL_GPL(fuse_file_poll);
2139 
2140 /*
2141  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2142  * wakes up the poll waiters.
2143  */
2144 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2145 			    struct fuse_notify_poll_wakeup_out *outarg)
2146 {
2147 	u64 kh = outarg->kh;
2148 	struct rb_node **link;
2149 
2150 	spin_lock(&fc->lock);
2151 
2152 	link = fuse_find_polled_node(fc, kh, NULL);
2153 	if (*link) {
2154 		struct fuse_file *ff;
2155 
2156 		ff = rb_entry(*link, struct fuse_file, polled_node);
2157 		wake_up_interruptible_sync(&ff->poll_wait);
2158 	}
2159 
2160 	spin_unlock(&fc->lock);
2161 	return 0;
2162 }
2163 
2164 static const struct file_operations fuse_file_operations = {
2165 	.llseek		= fuse_file_llseek,
2166 	.read		= do_sync_read,
2167 	.aio_read	= fuse_file_aio_read,
2168 	.write		= do_sync_write,
2169 	.aio_write	= fuse_file_aio_write,
2170 	.mmap		= fuse_file_mmap,
2171 	.open		= fuse_open,
2172 	.flush		= fuse_flush,
2173 	.release	= fuse_release,
2174 	.fsync		= fuse_fsync,
2175 	.lock		= fuse_file_lock,
2176 	.flock		= fuse_file_flock,
2177 	.splice_read	= generic_file_splice_read,
2178 	.unlocked_ioctl	= fuse_file_ioctl,
2179 	.compat_ioctl	= fuse_file_compat_ioctl,
2180 	.poll		= fuse_file_poll,
2181 };
2182 
2183 static const struct file_operations fuse_direct_io_file_operations = {
2184 	.llseek		= fuse_file_llseek,
2185 	.read		= fuse_direct_read,
2186 	.write		= fuse_direct_write,
2187 	.mmap		= fuse_direct_mmap,
2188 	.open		= fuse_open,
2189 	.flush		= fuse_flush,
2190 	.release	= fuse_release,
2191 	.fsync		= fuse_fsync,
2192 	.lock		= fuse_file_lock,
2193 	.flock		= fuse_file_flock,
2194 	.unlocked_ioctl	= fuse_file_ioctl,
2195 	.compat_ioctl	= fuse_file_compat_ioctl,
2196 	.poll		= fuse_file_poll,
2197 	/* no splice_read */
2198 };
2199 
2200 static const struct address_space_operations fuse_file_aops  = {
2201 	.readpage	= fuse_readpage,
2202 	.writepage	= fuse_writepage,
2203 	.launder_page	= fuse_launder_page,
2204 	.write_begin	= fuse_write_begin,
2205 	.write_end	= fuse_write_end,
2206 	.readpages	= fuse_readpages,
2207 	.set_page_dirty	= __set_page_dirty_nobuffers,
2208 	.bmap		= fuse_bmap,
2209 };
2210 
2211 void fuse_init_file_inode(struct inode *inode)
2212 {
2213 	inode->i_fop = &fuse_file_operations;
2214 	inode->i_data.a_ops = &fuse_file_aops;
2215 }
2216