xref: /openbmc/linux/fs/fuse/file.c (revision b98b3581)
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 #include <linux/swap.h>
18 
19 static const struct file_operations fuse_direct_io_file_operations;
20 
21 static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
22 			  int opcode, struct fuse_open_out *outargp)
23 {
24 	struct fuse_open_in inarg;
25 	struct fuse_req *req;
26 	int err;
27 
28 	req = fuse_get_req(fc);
29 	if (IS_ERR(req))
30 		return PTR_ERR(req);
31 
32 	memset(&inarg, 0, sizeof(inarg));
33 	inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
34 	if (!fc->atomic_o_trunc)
35 		inarg.flags &= ~O_TRUNC;
36 	req->in.h.opcode = opcode;
37 	req->in.h.nodeid = nodeid;
38 	req->in.numargs = 1;
39 	req->in.args[0].size = sizeof(inarg);
40 	req->in.args[0].value = &inarg;
41 	req->out.numargs = 1;
42 	req->out.args[0].size = sizeof(*outargp);
43 	req->out.args[0].value = outargp;
44 	fuse_request_send(fc, req);
45 	err = req->out.h.error;
46 	fuse_put_request(fc, req);
47 
48 	return err;
49 }
50 
51 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
52 {
53 	struct fuse_file *ff;
54 
55 	ff = kmalloc(sizeof(struct fuse_file), GFP_KERNEL);
56 	if (unlikely(!ff))
57 		return NULL;
58 
59 	ff->fc = fc;
60 	ff->reserved_req = fuse_request_alloc();
61 	if (unlikely(!ff->reserved_req)) {
62 		kfree(ff);
63 		return NULL;
64 	}
65 
66 	INIT_LIST_HEAD(&ff->write_entry);
67 	atomic_set(&ff->count, 0);
68 	RB_CLEAR_NODE(&ff->polled_node);
69 	init_waitqueue_head(&ff->poll_wait);
70 
71 	spin_lock(&fc->lock);
72 	ff->kh = ++fc->khctr;
73 	spin_unlock(&fc->lock);
74 
75 	return ff;
76 }
77 
78 void fuse_file_free(struct fuse_file *ff)
79 {
80 	fuse_request_free(ff->reserved_req);
81 	kfree(ff);
82 }
83 
84 struct fuse_file *fuse_file_get(struct fuse_file *ff)
85 {
86 	atomic_inc(&ff->count);
87 	return ff;
88 }
89 
90 static void fuse_release_async(struct work_struct *work)
91 {
92 	struct fuse_req *req;
93 	struct fuse_conn *fc;
94 	struct path path;
95 
96 	req = container_of(work, struct fuse_req, misc.release.work);
97 	path = req->misc.release.path;
98 	fc = get_fuse_conn(path.dentry->d_inode);
99 
100 	fuse_put_request(fc, req);
101 	path_put(&path);
102 }
103 
104 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
105 {
106 	if (fc->destroy_req) {
107 		/*
108 		 * If this is a fuseblk mount, then it's possible that
109 		 * releasing the path will result in releasing the
110 		 * super block and sending the DESTROY request.  If
111 		 * the server is single threaded, this would hang.
112 		 * For this reason do the path_put() in a separate
113 		 * thread.
114 		 */
115 		atomic_inc(&req->count);
116 		INIT_WORK(&req->misc.release.work, fuse_release_async);
117 		schedule_work(&req->misc.release.work);
118 	} else {
119 		path_put(&req->misc.release.path);
120 	}
121 }
122 
123 static void fuse_file_put(struct fuse_file *ff, bool sync)
124 {
125 	if (atomic_dec_and_test(&ff->count)) {
126 		struct fuse_req *req = ff->reserved_req;
127 
128 		if (sync) {
129 			fuse_request_send(ff->fc, req);
130 			path_put(&req->misc.release.path);
131 			fuse_put_request(ff->fc, req);
132 		} else {
133 			req->end = fuse_release_end;
134 			fuse_request_send_background(ff->fc, req);
135 		}
136 		kfree(ff);
137 	}
138 }
139 
140 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
141 		 bool isdir)
142 {
143 	struct fuse_open_out outarg;
144 	struct fuse_file *ff;
145 	int err;
146 	int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
147 
148 	ff = fuse_file_alloc(fc);
149 	if (!ff)
150 		return -ENOMEM;
151 
152 	err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
153 	if (err) {
154 		fuse_file_free(ff);
155 		return err;
156 	}
157 
158 	if (isdir)
159 		outarg.open_flags &= ~FOPEN_DIRECT_IO;
160 
161 	ff->fh = outarg.fh;
162 	ff->nodeid = nodeid;
163 	ff->open_flags = outarg.open_flags;
164 	file->private_data = fuse_file_get(ff);
165 
166 	return 0;
167 }
168 EXPORT_SYMBOL_GPL(fuse_do_open);
169 
170 void fuse_finish_open(struct inode *inode, struct file *file)
171 {
172 	struct fuse_file *ff = file->private_data;
173 	struct fuse_conn *fc = get_fuse_conn(inode);
174 
175 	if (ff->open_flags & FOPEN_DIRECT_IO)
176 		file->f_op = &fuse_direct_io_file_operations;
177 	if (!(ff->open_flags & FOPEN_KEEP_CACHE))
178 		invalidate_inode_pages2(inode->i_mapping);
179 	if (ff->open_flags & FOPEN_NONSEEKABLE)
180 		nonseekable_open(inode, file);
181 	if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
182 		struct fuse_inode *fi = get_fuse_inode(inode);
183 
184 		spin_lock(&fc->lock);
185 		fi->attr_version = ++fc->attr_version;
186 		i_size_write(inode, 0);
187 		spin_unlock(&fc->lock);
188 		fuse_invalidate_attr(inode);
189 	}
190 }
191 
192 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
193 {
194 	struct fuse_conn *fc = get_fuse_conn(inode);
195 	int err;
196 
197 	err = generic_file_open(inode, file);
198 	if (err)
199 		return err;
200 
201 	err = fuse_do_open(fc, get_node_id(inode), file, isdir);
202 	if (err)
203 		return err;
204 
205 	fuse_finish_open(inode, file);
206 
207 	return 0;
208 }
209 
210 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
211 {
212 	struct fuse_conn *fc = ff->fc;
213 	struct fuse_req *req = ff->reserved_req;
214 	struct fuse_release_in *inarg = &req->misc.release.in;
215 
216 	spin_lock(&fc->lock);
217 	list_del(&ff->write_entry);
218 	if (!RB_EMPTY_NODE(&ff->polled_node))
219 		rb_erase(&ff->polled_node, &fc->polled_files);
220 	spin_unlock(&fc->lock);
221 
222 	wake_up_interruptible_all(&ff->poll_wait);
223 
224 	inarg->fh = ff->fh;
225 	inarg->flags = flags;
226 	req->in.h.opcode = opcode;
227 	req->in.h.nodeid = ff->nodeid;
228 	req->in.numargs = 1;
229 	req->in.args[0].size = sizeof(struct fuse_release_in);
230 	req->in.args[0].value = inarg;
231 }
232 
233 void fuse_release_common(struct file *file, int opcode)
234 {
235 	struct fuse_file *ff;
236 	struct fuse_req *req;
237 
238 	ff = file->private_data;
239 	if (unlikely(!ff))
240 		return;
241 
242 	req = ff->reserved_req;
243 	fuse_prepare_release(ff, file->f_flags, opcode);
244 
245 	if (ff->flock) {
246 		struct fuse_release_in *inarg = &req->misc.release.in;
247 		inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
248 		inarg->lock_owner = fuse_lock_owner_id(ff->fc,
249 						       (fl_owner_t) file);
250 	}
251 	/* Hold vfsmount and dentry until release is finished */
252 	path_get(&file->f_path);
253 	req->misc.release.path = file->f_path;
254 
255 	/*
256 	 * Normally this will send the RELEASE request, however if
257 	 * some asynchronous READ or WRITE requests are outstanding,
258 	 * the sending will be delayed.
259 	 *
260 	 * Make the release synchronous if this is a fuseblk mount,
261 	 * synchronous RELEASE is allowed (and desirable) in this case
262 	 * because the server can be trusted not to screw up.
263 	 */
264 	fuse_file_put(ff, ff->fc->destroy_req != NULL);
265 }
266 
267 static int fuse_open(struct inode *inode, struct file *file)
268 {
269 	return fuse_open_common(inode, file, false);
270 }
271 
272 static int fuse_release(struct inode *inode, struct file *file)
273 {
274 	fuse_release_common(file, FUSE_RELEASE);
275 
276 	/* return value is ignored by VFS */
277 	return 0;
278 }
279 
280 void fuse_sync_release(struct fuse_file *ff, int flags)
281 {
282 	WARN_ON(atomic_read(&ff->count) > 1);
283 	fuse_prepare_release(ff, flags, FUSE_RELEASE);
284 	ff->reserved_req->force = 1;
285 	fuse_request_send(ff->fc, ff->reserved_req);
286 	fuse_put_request(ff->fc, ff->reserved_req);
287 	kfree(ff);
288 }
289 EXPORT_SYMBOL_GPL(fuse_sync_release);
290 
291 /*
292  * Scramble the ID space with XTEA, so that the value of the files_struct
293  * pointer is not exposed to userspace.
294  */
295 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
296 {
297 	u32 *k = fc->scramble_key;
298 	u64 v = (unsigned long) id;
299 	u32 v0 = v;
300 	u32 v1 = v >> 32;
301 	u32 sum = 0;
302 	int i;
303 
304 	for (i = 0; i < 32; i++) {
305 		v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
306 		sum += 0x9E3779B9;
307 		v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
308 	}
309 
310 	return (u64) v0 + ((u64) v1 << 32);
311 }
312 
313 /*
314  * Check if page is under writeback
315  *
316  * This is currently done by walking the list of writepage requests
317  * for the inode, which can be pretty inefficient.
318  */
319 static bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
320 {
321 	struct fuse_conn *fc = get_fuse_conn(inode);
322 	struct fuse_inode *fi = get_fuse_inode(inode);
323 	struct fuse_req *req;
324 	bool found = false;
325 
326 	spin_lock(&fc->lock);
327 	list_for_each_entry(req, &fi->writepages, writepages_entry) {
328 		pgoff_t curr_index;
329 
330 		BUG_ON(req->inode != inode);
331 		curr_index = req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
332 		if (curr_index == index) {
333 			found = true;
334 			break;
335 		}
336 	}
337 	spin_unlock(&fc->lock);
338 
339 	return found;
340 }
341 
342 /*
343  * Wait for page writeback to be completed.
344  *
345  * Since fuse doesn't rely on the VM writeback tracking, this has to
346  * use some other means.
347  */
348 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
349 {
350 	struct fuse_inode *fi = get_fuse_inode(inode);
351 
352 	wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
353 	return 0;
354 }
355 
356 static int fuse_flush(struct file *file, fl_owner_t id)
357 {
358 	struct inode *inode = file->f_path.dentry->d_inode;
359 	struct fuse_conn *fc = get_fuse_conn(inode);
360 	struct fuse_file *ff = file->private_data;
361 	struct fuse_req *req;
362 	struct fuse_flush_in inarg;
363 	int err;
364 
365 	if (is_bad_inode(inode))
366 		return -EIO;
367 
368 	if (fc->no_flush)
369 		return 0;
370 
371 	req = fuse_get_req_nofail(fc, file);
372 	memset(&inarg, 0, sizeof(inarg));
373 	inarg.fh = ff->fh;
374 	inarg.lock_owner = fuse_lock_owner_id(fc, id);
375 	req->in.h.opcode = FUSE_FLUSH;
376 	req->in.h.nodeid = get_node_id(inode);
377 	req->in.numargs = 1;
378 	req->in.args[0].size = sizeof(inarg);
379 	req->in.args[0].value = &inarg;
380 	req->force = 1;
381 	fuse_request_send(fc, req);
382 	err = req->out.h.error;
383 	fuse_put_request(fc, req);
384 	if (err == -ENOSYS) {
385 		fc->no_flush = 1;
386 		err = 0;
387 	}
388 	return err;
389 }
390 
391 /*
392  * Wait for all pending writepages on the inode to finish.
393  *
394  * This is currently done by blocking further writes with FUSE_NOWRITE
395  * and waiting for all sent writes to complete.
396  *
397  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
398  * could conflict with truncation.
399  */
400 static void fuse_sync_writes(struct inode *inode)
401 {
402 	fuse_set_nowrite(inode);
403 	fuse_release_nowrite(inode);
404 }
405 
406 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
407 		      int datasync, int isdir)
408 {
409 	struct inode *inode = file->f_mapping->host;
410 	struct fuse_conn *fc = get_fuse_conn(inode);
411 	struct fuse_file *ff = file->private_data;
412 	struct fuse_req *req;
413 	struct fuse_fsync_in inarg;
414 	int err;
415 
416 	if (is_bad_inode(inode))
417 		return -EIO;
418 
419 	err = filemap_write_and_wait_range(inode->i_mapping, start, end);
420 	if (err)
421 		return err;
422 
423 	if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
424 		return 0;
425 
426 	mutex_lock(&inode->i_mutex);
427 
428 	/*
429 	 * Start writeback against all dirty pages of the inode, then
430 	 * wait for all outstanding writes, before sending the FSYNC
431 	 * request.
432 	 */
433 	err = write_inode_now(inode, 0);
434 	if (err)
435 		goto out;
436 
437 	fuse_sync_writes(inode);
438 
439 	req = fuse_get_req(fc);
440 	if (IS_ERR(req)) {
441 		err = PTR_ERR(req);
442 		goto out;
443 	}
444 
445 	memset(&inarg, 0, sizeof(inarg));
446 	inarg.fh = ff->fh;
447 	inarg.fsync_flags = datasync ? 1 : 0;
448 	req->in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
449 	req->in.h.nodeid = get_node_id(inode);
450 	req->in.numargs = 1;
451 	req->in.args[0].size = sizeof(inarg);
452 	req->in.args[0].value = &inarg;
453 	fuse_request_send(fc, req);
454 	err = req->out.h.error;
455 	fuse_put_request(fc, req);
456 	if (err == -ENOSYS) {
457 		if (isdir)
458 			fc->no_fsyncdir = 1;
459 		else
460 			fc->no_fsync = 1;
461 		err = 0;
462 	}
463 out:
464 	mutex_unlock(&inode->i_mutex);
465 	return err;
466 }
467 
468 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
469 		      int datasync)
470 {
471 	return fuse_fsync_common(file, start, end, datasync, 0);
472 }
473 
474 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
475 		    size_t count, int opcode)
476 {
477 	struct fuse_read_in *inarg = &req->misc.read.in;
478 	struct fuse_file *ff = file->private_data;
479 
480 	inarg->fh = ff->fh;
481 	inarg->offset = pos;
482 	inarg->size = count;
483 	inarg->flags = file->f_flags;
484 	req->in.h.opcode = opcode;
485 	req->in.h.nodeid = ff->nodeid;
486 	req->in.numargs = 1;
487 	req->in.args[0].size = sizeof(struct fuse_read_in);
488 	req->in.args[0].value = inarg;
489 	req->out.argvar = 1;
490 	req->out.numargs = 1;
491 	req->out.args[0].size = count;
492 }
493 
494 static size_t fuse_send_read(struct fuse_req *req, struct file *file,
495 			     loff_t pos, size_t count, fl_owner_t owner)
496 {
497 	struct fuse_file *ff = file->private_data;
498 	struct fuse_conn *fc = ff->fc;
499 
500 	fuse_read_fill(req, file, pos, count, FUSE_READ);
501 	if (owner != NULL) {
502 		struct fuse_read_in *inarg = &req->misc.read.in;
503 
504 		inarg->read_flags |= FUSE_READ_LOCKOWNER;
505 		inarg->lock_owner = fuse_lock_owner_id(fc, owner);
506 	}
507 	fuse_request_send(fc, req);
508 	return req->out.args[0].size;
509 }
510 
511 static void fuse_read_update_size(struct inode *inode, loff_t size,
512 				  u64 attr_ver)
513 {
514 	struct fuse_conn *fc = get_fuse_conn(inode);
515 	struct fuse_inode *fi = get_fuse_inode(inode);
516 
517 	spin_lock(&fc->lock);
518 	if (attr_ver == fi->attr_version && size < inode->i_size) {
519 		fi->attr_version = ++fc->attr_version;
520 		i_size_write(inode, size);
521 	}
522 	spin_unlock(&fc->lock);
523 }
524 
525 static int fuse_readpage(struct file *file, struct page *page)
526 {
527 	struct inode *inode = page->mapping->host;
528 	struct fuse_conn *fc = get_fuse_conn(inode);
529 	struct fuse_req *req;
530 	size_t num_read;
531 	loff_t pos = page_offset(page);
532 	size_t count = PAGE_CACHE_SIZE;
533 	u64 attr_ver;
534 	int err;
535 
536 	err = -EIO;
537 	if (is_bad_inode(inode))
538 		goto out;
539 
540 	/*
541 	 * Page writeback can extend beyond the lifetime of the
542 	 * page-cache page, so make sure we read a properly synced
543 	 * page.
544 	 */
545 	fuse_wait_on_page_writeback(inode, page->index);
546 
547 	req = fuse_get_req(fc);
548 	err = PTR_ERR(req);
549 	if (IS_ERR(req))
550 		goto out;
551 
552 	attr_ver = fuse_get_attr_version(fc);
553 
554 	req->out.page_zeroing = 1;
555 	req->out.argpages = 1;
556 	req->num_pages = 1;
557 	req->pages[0] = page;
558 	num_read = fuse_send_read(req, file, pos, count, NULL);
559 	err = req->out.h.error;
560 	fuse_put_request(fc, req);
561 
562 	if (!err) {
563 		/*
564 		 * Short read means EOF.  If file size is larger, truncate it
565 		 */
566 		if (num_read < count)
567 			fuse_read_update_size(inode, pos + num_read, attr_ver);
568 
569 		SetPageUptodate(page);
570 	}
571 
572 	fuse_invalidate_attr(inode); /* atime changed */
573  out:
574 	unlock_page(page);
575 	return err;
576 }
577 
578 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
579 {
580 	int i;
581 	size_t count = req->misc.read.in.size;
582 	size_t num_read = req->out.args[0].size;
583 	struct address_space *mapping = NULL;
584 
585 	for (i = 0; mapping == NULL && i < req->num_pages; i++)
586 		mapping = req->pages[i]->mapping;
587 
588 	if (mapping) {
589 		struct inode *inode = mapping->host;
590 
591 		/*
592 		 * Short read means EOF. If file size is larger, truncate it
593 		 */
594 		if (!req->out.h.error && num_read < count) {
595 			loff_t pos;
596 
597 			pos = page_offset(req->pages[0]) + num_read;
598 			fuse_read_update_size(inode, pos,
599 					      req->misc.read.attr_ver);
600 		}
601 		fuse_invalidate_attr(inode); /* atime changed */
602 	}
603 
604 	for (i = 0; i < req->num_pages; i++) {
605 		struct page *page = req->pages[i];
606 		if (!req->out.h.error)
607 			SetPageUptodate(page);
608 		else
609 			SetPageError(page);
610 		unlock_page(page);
611 		page_cache_release(page);
612 	}
613 	if (req->ff)
614 		fuse_file_put(req->ff, false);
615 }
616 
617 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
618 {
619 	struct fuse_file *ff = file->private_data;
620 	struct fuse_conn *fc = ff->fc;
621 	loff_t pos = page_offset(req->pages[0]);
622 	size_t count = req->num_pages << PAGE_CACHE_SHIFT;
623 
624 	req->out.argpages = 1;
625 	req->out.page_zeroing = 1;
626 	req->out.page_replace = 1;
627 	fuse_read_fill(req, file, pos, count, FUSE_READ);
628 	req->misc.read.attr_ver = fuse_get_attr_version(fc);
629 	if (fc->async_read) {
630 		req->ff = fuse_file_get(ff);
631 		req->end = fuse_readpages_end;
632 		fuse_request_send_background(fc, req);
633 	} else {
634 		fuse_request_send(fc, req);
635 		fuse_readpages_end(fc, req);
636 		fuse_put_request(fc, req);
637 	}
638 }
639 
640 struct fuse_fill_data {
641 	struct fuse_req *req;
642 	struct file *file;
643 	struct inode *inode;
644 };
645 
646 static int fuse_readpages_fill(void *_data, struct page *page)
647 {
648 	struct fuse_fill_data *data = _data;
649 	struct fuse_req *req = data->req;
650 	struct inode *inode = data->inode;
651 	struct fuse_conn *fc = get_fuse_conn(inode);
652 
653 	fuse_wait_on_page_writeback(inode, page->index);
654 
655 	if (req->num_pages &&
656 	    (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
657 	     (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
658 	     req->pages[req->num_pages - 1]->index + 1 != page->index)) {
659 		fuse_send_readpages(req, data->file);
660 		data->req = req = fuse_get_req(fc);
661 		if (IS_ERR(req)) {
662 			unlock_page(page);
663 			return PTR_ERR(req);
664 		}
665 	}
666 	page_cache_get(page);
667 	req->pages[req->num_pages] = page;
668 	req->num_pages++;
669 	return 0;
670 }
671 
672 static int fuse_readpages(struct file *file, struct address_space *mapping,
673 			  struct list_head *pages, unsigned nr_pages)
674 {
675 	struct inode *inode = mapping->host;
676 	struct fuse_conn *fc = get_fuse_conn(inode);
677 	struct fuse_fill_data data;
678 	int err;
679 
680 	err = -EIO;
681 	if (is_bad_inode(inode))
682 		goto out;
683 
684 	data.file = file;
685 	data.inode = inode;
686 	data.req = fuse_get_req(fc);
687 	err = PTR_ERR(data.req);
688 	if (IS_ERR(data.req))
689 		goto out;
690 
691 	err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
692 	if (!err) {
693 		if (data.req->num_pages)
694 			fuse_send_readpages(data.req, file);
695 		else
696 			fuse_put_request(fc, data.req);
697 	}
698 out:
699 	return err;
700 }
701 
702 static ssize_t fuse_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
703 				  unsigned long nr_segs, loff_t pos)
704 {
705 	struct inode *inode = iocb->ki_filp->f_mapping->host;
706 
707 	if (pos + iov_length(iov, nr_segs) > i_size_read(inode)) {
708 		int err;
709 		/*
710 		 * If trying to read past EOF, make sure the i_size
711 		 * attribute is up-to-date.
712 		 */
713 		err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
714 		if (err)
715 			return err;
716 	}
717 
718 	return generic_file_aio_read(iocb, iov, nr_segs, pos);
719 }
720 
721 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
722 			    loff_t pos, size_t count)
723 {
724 	struct fuse_write_in *inarg = &req->misc.write.in;
725 	struct fuse_write_out *outarg = &req->misc.write.out;
726 
727 	inarg->fh = ff->fh;
728 	inarg->offset = pos;
729 	inarg->size = count;
730 	req->in.h.opcode = FUSE_WRITE;
731 	req->in.h.nodeid = ff->nodeid;
732 	req->in.numargs = 2;
733 	if (ff->fc->minor < 9)
734 		req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
735 	else
736 		req->in.args[0].size = sizeof(struct fuse_write_in);
737 	req->in.args[0].value = inarg;
738 	req->in.args[1].size = count;
739 	req->out.numargs = 1;
740 	req->out.args[0].size = sizeof(struct fuse_write_out);
741 	req->out.args[0].value = outarg;
742 }
743 
744 static size_t fuse_send_write(struct fuse_req *req, struct file *file,
745 			      loff_t pos, size_t count, fl_owner_t owner)
746 {
747 	struct fuse_file *ff = file->private_data;
748 	struct fuse_conn *fc = ff->fc;
749 	struct fuse_write_in *inarg = &req->misc.write.in;
750 
751 	fuse_write_fill(req, ff, pos, count);
752 	inarg->flags = file->f_flags;
753 	if (owner != NULL) {
754 		inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
755 		inarg->lock_owner = fuse_lock_owner_id(fc, owner);
756 	}
757 	fuse_request_send(fc, req);
758 	return req->misc.write.out.size;
759 }
760 
761 void fuse_write_update_size(struct inode *inode, loff_t pos)
762 {
763 	struct fuse_conn *fc = get_fuse_conn(inode);
764 	struct fuse_inode *fi = get_fuse_inode(inode);
765 
766 	spin_lock(&fc->lock);
767 	fi->attr_version = ++fc->attr_version;
768 	if (pos > inode->i_size)
769 		i_size_write(inode, pos);
770 	spin_unlock(&fc->lock);
771 }
772 
773 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
774 				    struct inode *inode, loff_t pos,
775 				    size_t count)
776 {
777 	size_t res;
778 	unsigned offset;
779 	unsigned i;
780 
781 	for (i = 0; i < req->num_pages; i++)
782 		fuse_wait_on_page_writeback(inode, req->pages[i]->index);
783 
784 	res = fuse_send_write(req, file, pos, count, NULL);
785 
786 	offset = req->page_offset;
787 	count = res;
788 	for (i = 0; i < req->num_pages; i++) {
789 		struct page *page = req->pages[i];
790 
791 		if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
792 			SetPageUptodate(page);
793 
794 		if (count > PAGE_CACHE_SIZE - offset)
795 			count -= PAGE_CACHE_SIZE - offset;
796 		else
797 			count = 0;
798 		offset = 0;
799 
800 		unlock_page(page);
801 		page_cache_release(page);
802 	}
803 
804 	return res;
805 }
806 
807 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
808 			       struct address_space *mapping,
809 			       struct iov_iter *ii, loff_t pos)
810 {
811 	struct fuse_conn *fc = get_fuse_conn(mapping->host);
812 	unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
813 	size_t count = 0;
814 	int err;
815 
816 	req->in.argpages = 1;
817 	req->page_offset = offset;
818 
819 	do {
820 		size_t tmp;
821 		struct page *page;
822 		pgoff_t index = pos >> PAGE_CACHE_SHIFT;
823 		size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
824 				     iov_iter_count(ii));
825 
826 		bytes = min_t(size_t, bytes, fc->max_write - count);
827 
828  again:
829 		err = -EFAULT;
830 		if (iov_iter_fault_in_readable(ii, bytes))
831 			break;
832 
833 		err = -ENOMEM;
834 		page = grab_cache_page_write_begin(mapping, index, 0);
835 		if (!page)
836 			break;
837 
838 		if (mapping_writably_mapped(mapping))
839 			flush_dcache_page(page);
840 
841 		pagefault_disable();
842 		tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
843 		pagefault_enable();
844 		flush_dcache_page(page);
845 
846 		mark_page_accessed(page);
847 
848 		if (!tmp) {
849 			unlock_page(page);
850 			page_cache_release(page);
851 			bytes = min(bytes, iov_iter_single_seg_count(ii));
852 			goto again;
853 		}
854 
855 		err = 0;
856 		req->pages[req->num_pages] = page;
857 		req->num_pages++;
858 
859 		iov_iter_advance(ii, tmp);
860 		count += tmp;
861 		pos += tmp;
862 		offset += tmp;
863 		if (offset == PAGE_CACHE_SIZE)
864 			offset = 0;
865 
866 		if (!fc->big_writes)
867 			break;
868 	} while (iov_iter_count(ii) && count < fc->max_write &&
869 		 req->num_pages < FUSE_MAX_PAGES_PER_REQ && offset == 0);
870 
871 	return count > 0 ? count : err;
872 }
873 
874 static ssize_t fuse_perform_write(struct file *file,
875 				  struct address_space *mapping,
876 				  struct iov_iter *ii, loff_t pos)
877 {
878 	struct inode *inode = mapping->host;
879 	struct fuse_conn *fc = get_fuse_conn(inode);
880 	int err = 0;
881 	ssize_t res = 0;
882 
883 	if (is_bad_inode(inode))
884 		return -EIO;
885 
886 	do {
887 		struct fuse_req *req;
888 		ssize_t count;
889 
890 		req = fuse_get_req(fc);
891 		if (IS_ERR(req)) {
892 			err = PTR_ERR(req);
893 			break;
894 		}
895 
896 		count = fuse_fill_write_pages(req, mapping, ii, pos);
897 		if (count <= 0) {
898 			err = count;
899 		} else {
900 			size_t num_written;
901 
902 			num_written = fuse_send_write_pages(req, file, inode,
903 							    pos, count);
904 			err = req->out.h.error;
905 			if (!err) {
906 				res += num_written;
907 				pos += num_written;
908 
909 				/* break out of the loop on short write */
910 				if (num_written != count)
911 					err = -EIO;
912 			}
913 		}
914 		fuse_put_request(fc, req);
915 	} while (!err && iov_iter_count(ii));
916 
917 	if (res > 0)
918 		fuse_write_update_size(inode, pos);
919 
920 	fuse_invalidate_attr(inode);
921 
922 	return res > 0 ? res : err;
923 }
924 
925 static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
926 				   unsigned long nr_segs, loff_t pos)
927 {
928 	struct file *file = iocb->ki_filp;
929 	struct address_space *mapping = file->f_mapping;
930 	size_t count = 0;
931 	size_t ocount = 0;
932 	ssize_t written = 0;
933 	ssize_t written_buffered = 0;
934 	struct inode *inode = mapping->host;
935 	ssize_t err;
936 	struct iov_iter i;
937 	loff_t endbyte = 0;
938 
939 	WARN_ON(iocb->ki_pos != pos);
940 
941 	ocount = 0;
942 	err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
943 	if (err)
944 		return err;
945 
946 	count = ocount;
947 
948 	mutex_lock(&inode->i_mutex);
949 	vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
950 
951 	/* We can write back this queue in page reclaim */
952 	current->backing_dev_info = mapping->backing_dev_info;
953 
954 	err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
955 	if (err)
956 		goto out;
957 
958 	if (count == 0)
959 		goto out;
960 
961 	err = file_remove_suid(file);
962 	if (err)
963 		goto out;
964 
965 	file_update_time(file);
966 
967 	if (file->f_flags & O_DIRECT) {
968 		written = generic_file_direct_write(iocb, iov, &nr_segs,
969 						    pos, &iocb->ki_pos,
970 						    count, ocount);
971 		if (written < 0 || written == count)
972 			goto out;
973 
974 		pos += written;
975 		count -= written;
976 
977 		iov_iter_init(&i, iov, nr_segs, count, written);
978 		written_buffered = fuse_perform_write(file, mapping, &i, pos);
979 		if (written_buffered < 0) {
980 			err = written_buffered;
981 			goto out;
982 		}
983 		endbyte = pos + written_buffered - 1;
984 
985 		err = filemap_write_and_wait_range(file->f_mapping, pos,
986 						   endbyte);
987 		if (err)
988 			goto out;
989 
990 		invalidate_mapping_pages(file->f_mapping,
991 					 pos >> PAGE_CACHE_SHIFT,
992 					 endbyte >> PAGE_CACHE_SHIFT);
993 
994 		written += written_buffered;
995 		iocb->ki_pos = pos + written_buffered;
996 	} else {
997 		iov_iter_init(&i, iov, nr_segs, count, 0);
998 		written = fuse_perform_write(file, mapping, &i, pos);
999 		if (written >= 0)
1000 			iocb->ki_pos = pos + written;
1001 	}
1002 out:
1003 	current->backing_dev_info = NULL;
1004 	mutex_unlock(&inode->i_mutex);
1005 
1006 	return written ? written : err;
1007 }
1008 
1009 static void fuse_release_user_pages(struct fuse_req *req, int write)
1010 {
1011 	unsigned i;
1012 
1013 	for (i = 0; i < req->num_pages; i++) {
1014 		struct page *page = req->pages[i];
1015 		if (write)
1016 			set_page_dirty_lock(page);
1017 		put_page(page);
1018 	}
1019 }
1020 
1021 static int fuse_get_user_pages(struct fuse_req *req, const char __user *buf,
1022 			       size_t *nbytesp, int write)
1023 {
1024 	size_t nbytes = *nbytesp;
1025 	unsigned long user_addr = (unsigned long) buf;
1026 	unsigned offset = user_addr & ~PAGE_MASK;
1027 	int npages;
1028 
1029 	/* Special case for kernel I/O: can copy directly into the buffer */
1030 	if (segment_eq(get_fs(), KERNEL_DS)) {
1031 		if (write)
1032 			req->in.args[1].value = (void *) user_addr;
1033 		else
1034 			req->out.args[0].value = (void *) user_addr;
1035 
1036 		return 0;
1037 	}
1038 
1039 	nbytes = min_t(size_t, nbytes, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
1040 	npages = (nbytes + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1041 	npages = clamp(npages, 1, FUSE_MAX_PAGES_PER_REQ);
1042 	npages = get_user_pages_fast(user_addr, npages, !write, req->pages);
1043 	if (npages < 0)
1044 		return npages;
1045 
1046 	req->num_pages = npages;
1047 	req->page_offset = offset;
1048 
1049 	if (write)
1050 		req->in.argpages = 1;
1051 	else
1052 		req->out.argpages = 1;
1053 
1054 	nbytes = (req->num_pages << PAGE_SHIFT) - req->page_offset;
1055 	*nbytesp = min(*nbytesp, nbytes);
1056 
1057 	return 0;
1058 }
1059 
1060 ssize_t fuse_direct_io(struct file *file, const char __user *buf,
1061 		       size_t count, loff_t *ppos, int write)
1062 {
1063 	struct fuse_file *ff = file->private_data;
1064 	struct fuse_conn *fc = ff->fc;
1065 	size_t nmax = write ? fc->max_write : fc->max_read;
1066 	loff_t pos = *ppos;
1067 	ssize_t res = 0;
1068 	struct fuse_req *req;
1069 
1070 	req = fuse_get_req(fc);
1071 	if (IS_ERR(req))
1072 		return PTR_ERR(req);
1073 
1074 	while (count) {
1075 		size_t nres;
1076 		fl_owner_t owner = current->files;
1077 		size_t nbytes = min(count, nmax);
1078 		int err = fuse_get_user_pages(req, buf, &nbytes, write);
1079 		if (err) {
1080 			res = err;
1081 			break;
1082 		}
1083 
1084 		if (write)
1085 			nres = fuse_send_write(req, file, pos, nbytes, owner);
1086 		else
1087 			nres = fuse_send_read(req, file, pos, nbytes, owner);
1088 
1089 		fuse_release_user_pages(req, !write);
1090 		if (req->out.h.error) {
1091 			if (!res)
1092 				res = req->out.h.error;
1093 			break;
1094 		} else if (nres > nbytes) {
1095 			res = -EIO;
1096 			break;
1097 		}
1098 		count -= nres;
1099 		res += nres;
1100 		pos += nres;
1101 		buf += nres;
1102 		if (nres != nbytes)
1103 			break;
1104 		if (count) {
1105 			fuse_put_request(fc, req);
1106 			req = fuse_get_req(fc);
1107 			if (IS_ERR(req))
1108 				break;
1109 		}
1110 	}
1111 	if (!IS_ERR(req))
1112 		fuse_put_request(fc, req);
1113 	if (res > 0)
1114 		*ppos = pos;
1115 
1116 	return res;
1117 }
1118 EXPORT_SYMBOL_GPL(fuse_direct_io);
1119 
1120 static ssize_t fuse_direct_read(struct file *file, char __user *buf,
1121 				     size_t count, loff_t *ppos)
1122 {
1123 	ssize_t res;
1124 	struct inode *inode = file->f_path.dentry->d_inode;
1125 
1126 	if (is_bad_inode(inode))
1127 		return -EIO;
1128 
1129 	res = fuse_direct_io(file, buf, count, ppos, 0);
1130 
1131 	fuse_invalidate_attr(inode);
1132 
1133 	return res;
1134 }
1135 
1136 static ssize_t __fuse_direct_write(struct file *file, const char __user *buf,
1137 				   size_t count, loff_t *ppos)
1138 {
1139 	struct inode *inode = file->f_path.dentry->d_inode;
1140 	ssize_t res;
1141 
1142 	res = generic_write_checks(file, ppos, &count, 0);
1143 	if (!res) {
1144 		res = fuse_direct_io(file, buf, count, ppos, 1);
1145 		if (res > 0)
1146 			fuse_write_update_size(inode, *ppos);
1147 	}
1148 
1149 	fuse_invalidate_attr(inode);
1150 
1151 	return res;
1152 }
1153 
1154 static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
1155 				 size_t count, loff_t *ppos)
1156 {
1157 	struct inode *inode = file->f_path.dentry->d_inode;
1158 	ssize_t res;
1159 
1160 	if (is_bad_inode(inode))
1161 		return -EIO;
1162 
1163 	/* Don't allow parallel writes to the same file */
1164 	mutex_lock(&inode->i_mutex);
1165 	res = __fuse_direct_write(file, buf, count, ppos);
1166 	mutex_unlock(&inode->i_mutex);
1167 
1168 	return res;
1169 }
1170 
1171 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1172 {
1173 	__free_page(req->pages[0]);
1174 	fuse_file_put(req->ff, false);
1175 }
1176 
1177 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1178 {
1179 	struct inode *inode = req->inode;
1180 	struct fuse_inode *fi = get_fuse_inode(inode);
1181 	struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
1182 
1183 	list_del(&req->writepages_entry);
1184 	dec_bdi_stat(bdi, BDI_WRITEBACK);
1185 	dec_zone_page_state(req->pages[0], NR_WRITEBACK_TEMP);
1186 	bdi_writeout_inc(bdi);
1187 	wake_up(&fi->page_waitq);
1188 }
1189 
1190 /* Called under fc->lock, may release and reacquire it */
1191 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req)
1192 __releases(fc->lock)
1193 __acquires(fc->lock)
1194 {
1195 	struct fuse_inode *fi = get_fuse_inode(req->inode);
1196 	loff_t size = i_size_read(req->inode);
1197 	struct fuse_write_in *inarg = &req->misc.write.in;
1198 
1199 	if (!fc->connected)
1200 		goto out_free;
1201 
1202 	if (inarg->offset + PAGE_CACHE_SIZE <= size) {
1203 		inarg->size = PAGE_CACHE_SIZE;
1204 	} else if (inarg->offset < size) {
1205 		inarg->size = size & (PAGE_CACHE_SIZE - 1);
1206 	} else {
1207 		/* Got truncated off completely */
1208 		goto out_free;
1209 	}
1210 
1211 	req->in.args[1].size = inarg->size;
1212 	fi->writectr++;
1213 	fuse_request_send_background_locked(fc, req);
1214 	return;
1215 
1216  out_free:
1217 	fuse_writepage_finish(fc, req);
1218 	spin_unlock(&fc->lock);
1219 	fuse_writepage_free(fc, req);
1220 	fuse_put_request(fc, req);
1221 	spin_lock(&fc->lock);
1222 }
1223 
1224 /*
1225  * If fi->writectr is positive (no truncate or fsync going on) send
1226  * all queued writepage requests.
1227  *
1228  * Called with fc->lock
1229  */
1230 void fuse_flush_writepages(struct inode *inode)
1231 __releases(fc->lock)
1232 __acquires(fc->lock)
1233 {
1234 	struct fuse_conn *fc = get_fuse_conn(inode);
1235 	struct fuse_inode *fi = get_fuse_inode(inode);
1236 	struct fuse_req *req;
1237 
1238 	while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1239 		req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1240 		list_del_init(&req->list);
1241 		fuse_send_writepage(fc, req);
1242 	}
1243 }
1244 
1245 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1246 {
1247 	struct inode *inode = req->inode;
1248 	struct fuse_inode *fi = get_fuse_inode(inode);
1249 
1250 	mapping_set_error(inode->i_mapping, req->out.h.error);
1251 	spin_lock(&fc->lock);
1252 	fi->writectr--;
1253 	fuse_writepage_finish(fc, req);
1254 	spin_unlock(&fc->lock);
1255 	fuse_writepage_free(fc, req);
1256 }
1257 
1258 static int fuse_writepage_locked(struct page *page)
1259 {
1260 	struct address_space *mapping = page->mapping;
1261 	struct inode *inode = mapping->host;
1262 	struct fuse_conn *fc = get_fuse_conn(inode);
1263 	struct fuse_inode *fi = get_fuse_inode(inode);
1264 	struct fuse_req *req;
1265 	struct fuse_file *ff;
1266 	struct page *tmp_page;
1267 
1268 	set_page_writeback(page);
1269 
1270 	req = fuse_request_alloc_nofs();
1271 	if (!req)
1272 		goto err;
1273 
1274 	tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1275 	if (!tmp_page)
1276 		goto err_free;
1277 
1278 	spin_lock(&fc->lock);
1279 	BUG_ON(list_empty(&fi->write_files));
1280 	ff = list_entry(fi->write_files.next, struct fuse_file, write_entry);
1281 	req->ff = fuse_file_get(ff);
1282 	spin_unlock(&fc->lock);
1283 
1284 	fuse_write_fill(req, ff, page_offset(page), 0);
1285 
1286 	copy_highpage(tmp_page, page);
1287 	req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1288 	req->in.argpages = 1;
1289 	req->num_pages = 1;
1290 	req->pages[0] = tmp_page;
1291 	req->page_offset = 0;
1292 	req->end = fuse_writepage_end;
1293 	req->inode = inode;
1294 
1295 	inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
1296 	inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1297 	end_page_writeback(page);
1298 
1299 	spin_lock(&fc->lock);
1300 	list_add(&req->writepages_entry, &fi->writepages);
1301 	list_add_tail(&req->list, &fi->queued_writes);
1302 	fuse_flush_writepages(inode);
1303 	spin_unlock(&fc->lock);
1304 
1305 	return 0;
1306 
1307 err_free:
1308 	fuse_request_free(req);
1309 err:
1310 	end_page_writeback(page);
1311 	return -ENOMEM;
1312 }
1313 
1314 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1315 {
1316 	int err;
1317 
1318 	err = fuse_writepage_locked(page);
1319 	unlock_page(page);
1320 
1321 	return err;
1322 }
1323 
1324 static int fuse_launder_page(struct page *page)
1325 {
1326 	int err = 0;
1327 	if (clear_page_dirty_for_io(page)) {
1328 		struct inode *inode = page->mapping->host;
1329 		err = fuse_writepage_locked(page);
1330 		if (!err)
1331 			fuse_wait_on_page_writeback(inode, page->index);
1332 	}
1333 	return err;
1334 }
1335 
1336 /*
1337  * Write back dirty pages now, because there may not be any suitable
1338  * open files later
1339  */
1340 static void fuse_vma_close(struct vm_area_struct *vma)
1341 {
1342 	filemap_write_and_wait(vma->vm_file->f_mapping);
1343 }
1344 
1345 /*
1346  * Wait for writeback against this page to complete before allowing it
1347  * to be marked dirty again, and hence written back again, possibly
1348  * before the previous writepage completed.
1349  *
1350  * Block here, instead of in ->writepage(), so that the userspace fs
1351  * can only block processes actually operating on the filesystem.
1352  *
1353  * Otherwise unprivileged userspace fs would be able to block
1354  * unrelated:
1355  *
1356  * - page migration
1357  * - sync(2)
1358  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
1359  */
1360 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1361 {
1362 	struct page *page = vmf->page;
1363 	/*
1364 	 * Don't use page->mapping as it may become NULL from a
1365 	 * concurrent truncate.
1366 	 */
1367 	struct inode *inode = vma->vm_file->f_mapping->host;
1368 
1369 	fuse_wait_on_page_writeback(inode, page->index);
1370 	return 0;
1371 }
1372 
1373 static const struct vm_operations_struct fuse_file_vm_ops = {
1374 	.close		= fuse_vma_close,
1375 	.fault		= filemap_fault,
1376 	.page_mkwrite	= fuse_page_mkwrite,
1377 };
1378 
1379 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
1380 {
1381 	if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE)) {
1382 		struct inode *inode = file->f_dentry->d_inode;
1383 		struct fuse_conn *fc = get_fuse_conn(inode);
1384 		struct fuse_inode *fi = get_fuse_inode(inode);
1385 		struct fuse_file *ff = file->private_data;
1386 		/*
1387 		 * file may be written through mmap, so chain it onto the
1388 		 * inodes's write_file list
1389 		 */
1390 		spin_lock(&fc->lock);
1391 		if (list_empty(&ff->write_entry))
1392 			list_add(&ff->write_entry, &fi->write_files);
1393 		spin_unlock(&fc->lock);
1394 	}
1395 	file_accessed(file);
1396 	vma->vm_ops = &fuse_file_vm_ops;
1397 	return 0;
1398 }
1399 
1400 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
1401 {
1402 	/* Can't provide the coherency needed for MAP_SHARED */
1403 	if (vma->vm_flags & VM_MAYSHARE)
1404 		return -ENODEV;
1405 
1406 	invalidate_inode_pages2(file->f_mapping);
1407 
1408 	return generic_file_mmap(file, vma);
1409 }
1410 
1411 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
1412 				  struct file_lock *fl)
1413 {
1414 	switch (ffl->type) {
1415 	case F_UNLCK:
1416 		break;
1417 
1418 	case F_RDLCK:
1419 	case F_WRLCK:
1420 		if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
1421 		    ffl->end < ffl->start)
1422 			return -EIO;
1423 
1424 		fl->fl_start = ffl->start;
1425 		fl->fl_end = ffl->end;
1426 		fl->fl_pid = ffl->pid;
1427 		break;
1428 
1429 	default:
1430 		return -EIO;
1431 	}
1432 	fl->fl_type = ffl->type;
1433 	return 0;
1434 }
1435 
1436 static void fuse_lk_fill(struct fuse_req *req, struct file *file,
1437 			 const struct file_lock *fl, int opcode, pid_t pid,
1438 			 int flock)
1439 {
1440 	struct inode *inode = file->f_path.dentry->d_inode;
1441 	struct fuse_conn *fc = get_fuse_conn(inode);
1442 	struct fuse_file *ff = file->private_data;
1443 	struct fuse_lk_in *arg = &req->misc.lk_in;
1444 
1445 	arg->fh = ff->fh;
1446 	arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
1447 	arg->lk.start = fl->fl_start;
1448 	arg->lk.end = fl->fl_end;
1449 	arg->lk.type = fl->fl_type;
1450 	arg->lk.pid = pid;
1451 	if (flock)
1452 		arg->lk_flags |= FUSE_LK_FLOCK;
1453 	req->in.h.opcode = opcode;
1454 	req->in.h.nodeid = get_node_id(inode);
1455 	req->in.numargs = 1;
1456 	req->in.args[0].size = sizeof(*arg);
1457 	req->in.args[0].value = arg;
1458 }
1459 
1460 static int fuse_getlk(struct file *file, struct file_lock *fl)
1461 {
1462 	struct inode *inode = file->f_path.dentry->d_inode;
1463 	struct fuse_conn *fc = get_fuse_conn(inode);
1464 	struct fuse_req *req;
1465 	struct fuse_lk_out outarg;
1466 	int err;
1467 
1468 	req = fuse_get_req(fc);
1469 	if (IS_ERR(req))
1470 		return PTR_ERR(req);
1471 
1472 	fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
1473 	req->out.numargs = 1;
1474 	req->out.args[0].size = sizeof(outarg);
1475 	req->out.args[0].value = &outarg;
1476 	fuse_request_send(fc, req);
1477 	err = req->out.h.error;
1478 	fuse_put_request(fc, req);
1479 	if (!err)
1480 		err = convert_fuse_file_lock(&outarg.lk, fl);
1481 
1482 	return err;
1483 }
1484 
1485 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
1486 {
1487 	struct inode *inode = file->f_path.dentry->d_inode;
1488 	struct fuse_conn *fc = get_fuse_conn(inode);
1489 	struct fuse_req *req;
1490 	int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
1491 	pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
1492 	int err;
1493 
1494 	if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
1495 		/* NLM needs asynchronous locks, which we don't support yet */
1496 		return -ENOLCK;
1497 	}
1498 
1499 	/* Unlock on close is handled by the flush method */
1500 	if (fl->fl_flags & FL_CLOSE)
1501 		return 0;
1502 
1503 	req = fuse_get_req(fc);
1504 	if (IS_ERR(req))
1505 		return PTR_ERR(req);
1506 
1507 	fuse_lk_fill(req, file, fl, opcode, pid, flock);
1508 	fuse_request_send(fc, req);
1509 	err = req->out.h.error;
1510 	/* locking is restartable */
1511 	if (err == -EINTR)
1512 		err = -ERESTARTSYS;
1513 	fuse_put_request(fc, req);
1514 	return err;
1515 }
1516 
1517 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
1518 {
1519 	struct inode *inode = file->f_path.dentry->d_inode;
1520 	struct fuse_conn *fc = get_fuse_conn(inode);
1521 	int err;
1522 
1523 	if (cmd == F_CANCELLK) {
1524 		err = 0;
1525 	} else if (cmd == F_GETLK) {
1526 		if (fc->no_lock) {
1527 			posix_test_lock(file, fl);
1528 			err = 0;
1529 		} else
1530 			err = fuse_getlk(file, fl);
1531 	} else {
1532 		if (fc->no_lock)
1533 			err = posix_lock_file(file, fl, NULL);
1534 		else
1535 			err = fuse_setlk(file, fl, 0);
1536 	}
1537 	return err;
1538 }
1539 
1540 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
1541 {
1542 	struct inode *inode = file->f_path.dentry->d_inode;
1543 	struct fuse_conn *fc = get_fuse_conn(inode);
1544 	int err;
1545 
1546 	if (fc->no_flock) {
1547 		err = flock_lock_file_wait(file, fl);
1548 	} else {
1549 		struct fuse_file *ff = file->private_data;
1550 
1551 		/* emulate flock with POSIX locks */
1552 		fl->fl_owner = (fl_owner_t) file;
1553 		ff->flock = true;
1554 		err = fuse_setlk(file, fl, 1);
1555 	}
1556 
1557 	return err;
1558 }
1559 
1560 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
1561 {
1562 	struct inode *inode = mapping->host;
1563 	struct fuse_conn *fc = get_fuse_conn(inode);
1564 	struct fuse_req *req;
1565 	struct fuse_bmap_in inarg;
1566 	struct fuse_bmap_out outarg;
1567 	int err;
1568 
1569 	if (!inode->i_sb->s_bdev || fc->no_bmap)
1570 		return 0;
1571 
1572 	req = fuse_get_req(fc);
1573 	if (IS_ERR(req))
1574 		return 0;
1575 
1576 	memset(&inarg, 0, sizeof(inarg));
1577 	inarg.block = block;
1578 	inarg.blocksize = inode->i_sb->s_blocksize;
1579 	req->in.h.opcode = FUSE_BMAP;
1580 	req->in.h.nodeid = get_node_id(inode);
1581 	req->in.numargs = 1;
1582 	req->in.args[0].size = sizeof(inarg);
1583 	req->in.args[0].value = &inarg;
1584 	req->out.numargs = 1;
1585 	req->out.args[0].size = sizeof(outarg);
1586 	req->out.args[0].value = &outarg;
1587 	fuse_request_send(fc, req);
1588 	err = req->out.h.error;
1589 	fuse_put_request(fc, req);
1590 	if (err == -ENOSYS)
1591 		fc->no_bmap = 1;
1592 
1593 	return err ? 0 : outarg.block;
1594 }
1595 
1596 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int origin)
1597 {
1598 	loff_t retval;
1599 	struct inode *inode = file->f_path.dentry->d_inode;
1600 
1601 	/* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
1602 	if (origin == SEEK_CUR || origin == SEEK_SET)
1603 		return generic_file_llseek(file, offset, origin);
1604 
1605 	mutex_lock(&inode->i_mutex);
1606 	retval = fuse_update_attributes(inode, NULL, file, NULL);
1607 	if (!retval)
1608 		retval = generic_file_llseek(file, offset, origin);
1609 	mutex_unlock(&inode->i_mutex);
1610 
1611 	return retval;
1612 }
1613 
1614 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
1615 			unsigned int nr_segs, size_t bytes, bool to_user)
1616 {
1617 	struct iov_iter ii;
1618 	int page_idx = 0;
1619 
1620 	if (!bytes)
1621 		return 0;
1622 
1623 	iov_iter_init(&ii, iov, nr_segs, bytes, 0);
1624 
1625 	while (iov_iter_count(&ii)) {
1626 		struct page *page = pages[page_idx++];
1627 		size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
1628 		void *kaddr;
1629 
1630 		kaddr = kmap(page);
1631 
1632 		while (todo) {
1633 			char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
1634 			size_t iov_len = ii.iov->iov_len - ii.iov_offset;
1635 			size_t copy = min(todo, iov_len);
1636 			size_t left;
1637 
1638 			if (!to_user)
1639 				left = copy_from_user(kaddr, uaddr, copy);
1640 			else
1641 				left = copy_to_user(uaddr, kaddr, copy);
1642 
1643 			if (unlikely(left))
1644 				return -EFAULT;
1645 
1646 			iov_iter_advance(&ii, copy);
1647 			todo -= copy;
1648 			kaddr += copy;
1649 		}
1650 
1651 		kunmap(page);
1652 	}
1653 
1654 	return 0;
1655 }
1656 
1657 /*
1658  * CUSE servers compiled on 32bit broke on 64bit kernels because the
1659  * ABI was defined to be 'struct iovec' which is different on 32bit
1660  * and 64bit.  Fortunately we can determine which structure the server
1661  * used from the size of the reply.
1662  */
1663 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
1664 				     size_t transferred, unsigned count,
1665 				     bool is_compat)
1666 {
1667 #ifdef CONFIG_COMPAT
1668 	if (count * sizeof(struct compat_iovec) == transferred) {
1669 		struct compat_iovec *ciov = src;
1670 		unsigned i;
1671 
1672 		/*
1673 		 * With this interface a 32bit server cannot support
1674 		 * non-compat (i.e. ones coming from 64bit apps) ioctl
1675 		 * requests
1676 		 */
1677 		if (!is_compat)
1678 			return -EINVAL;
1679 
1680 		for (i = 0; i < count; i++) {
1681 			dst[i].iov_base = compat_ptr(ciov[i].iov_base);
1682 			dst[i].iov_len = ciov[i].iov_len;
1683 		}
1684 		return 0;
1685 	}
1686 #endif
1687 
1688 	if (count * sizeof(struct iovec) != transferred)
1689 		return -EIO;
1690 
1691 	memcpy(dst, src, transferred);
1692 	return 0;
1693 }
1694 
1695 /* Make sure iov_length() won't overflow */
1696 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
1697 {
1698 	size_t n;
1699 	u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
1700 
1701 	for (n = 0; n < count; n++) {
1702 		if (iov->iov_len > (size_t) max)
1703 			return -ENOMEM;
1704 		max -= iov->iov_len;
1705 	}
1706 	return 0;
1707 }
1708 
1709 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
1710 				 void *src, size_t transferred, unsigned count,
1711 				 bool is_compat)
1712 {
1713 	unsigned i;
1714 	struct fuse_ioctl_iovec *fiov = src;
1715 
1716 	if (fc->minor < 16) {
1717 		return fuse_copy_ioctl_iovec_old(dst, src, transferred,
1718 						 count, is_compat);
1719 	}
1720 
1721 	if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
1722 		return -EIO;
1723 
1724 	for (i = 0; i < count; i++) {
1725 		/* Did the server supply an inappropriate value? */
1726 		if (fiov[i].base != (unsigned long) fiov[i].base ||
1727 		    fiov[i].len != (unsigned long) fiov[i].len)
1728 			return -EIO;
1729 
1730 		dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
1731 		dst[i].iov_len = (size_t) fiov[i].len;
1732 
1733 #ifdef CONFIG_COMPAT
1734 		if (is_compat &&
1735 		    (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
1736 		     (compat_size_t) dst[i].iov_len != fiov[i].len))
1737 			return -EIO;
1738 #endif
1739 	}
1740 
1741 	return 0;
1742 }
1743 
1744 
1745 /*
1746  * For ioctls, there is no generic way to determine how much memory
1747  * needs to be read and/or written.  Furthermore, ioctls are allowed
1748  * to dereference the passed pointer, so the parameter requires deep
1749  * copying but FUSE has no idea whatsoever about what to copy in or
1750  * out.
1751  *
1752  * This is solved by allowing FUSE server to retry ioctl with
1753  * necessary in/out iovecs.  Let's assume the ioctl implementation
1754  * needs to read in the following structure.
1755  *
1756  * struct a {
1757  *	char	*buf;
1758  *	size_t	buflen;
1759  * }
1760  *
1761  * On the first callout to FUSE server, inarg->in_size and
1762  * inarg->out_size will be NULL; then, the server completes the ioctl
1763  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
1764  * the actual iov array to
1765  *
1766  * { { .iov_base = inarg.arg,	.iov_len = sizeof(struct a) } }
1767  *
1768  * which tells FUSE to copy in the requested area and retry the ioctl.
1769  * On the second round, the server has access to the structure and
1770  * from that it can tell what to look for next, so on the invocation,
1771  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
1772  *
1773  * { { .iov_base = inarg.arg,	.iov_len = sizeof(struct a)	},
1774  *   { .iov_base = a.buf,	.iov_len = a.buflen		} }
1775  *
1776  * FUSE will copy both struct a and the pointed buffer from the
1777  * process doing the ioctl and retry ioctl with both struct a and the
1778  * buffer.
1779  *
1780  * This time, FUSE server has everything it needs and completes ioctl
1781  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
1782  *
1783  * Copying data out works the same way.
1784  *
1785  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
1786  * automatically initializes in and out iovs by decoding @cmd with
1787  * _IOC_* macros and the server is not allowed to request RETRY.  This
1788  * limits ioctl data transfers to well-formed ioctls and is the forced
1789  * behavior for all FUSE servers.
1790  */
1791 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
1792 		   unsigned int flags)
1793 {
1794 	struct fuse_file *ff = file->private_data;
1795 	struct fuse_conn *fc = ff->fc;
1796 	struct fuse_ioctl_in inarg = {
1797 		.fh = ff->fh,
1798 		.cmd = cmd,
1799 		.arg = arg,
1800 		.flags = flags
1801 	};
1802 	struct fuse_ioctl_out outarg;
1803 	struct fuse_req *req = NULL;
1804 	struct page **pages = NULL;
1805 	struct iovec *iov_page = NULL;
1806 	struct iovec *in_iov = NULL, *out_iov = NULL;
1807 	unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
1808 	size_t in_size, out_size, transferred;
1809 	int err;
1810 
1811 #if BITS_PER_LONG == 32
1812 	inarg.flags |= FUSE_IOCTL_32BIT;
1813 #else
1814 	if (flags & FUSE_IOCTL_COMPAT)
1815 		inarg.flags |= FUSE_IOCTL_32BIT;
1816 #endif
1817 
1818 	/* assume all the iovs returned by client always fits in a page */
1819 	BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
1820 
1821 	err = -ENOMEM;
1822 	pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
1823 	iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
1824 	if (!pages || !iov_page)
1825 		goto out;
1826 
1827 	/*
1828 	 * If restricted, initialize IO parameters as encoded in @cmd.
1829 	 * RETRY from server is not allowed.
1830 	 */
1831 	if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
1832 		struct iovec *iov = iov_page;
1833 
1834 		iov->iov_base = (void __user *)arg;
1835 		iov->iov_len = _IOC_SIZE(cmd);
1836 
1837 		if (_IOC_DIR(cmd) & _IOC_WRITE) {
1838 			in_iov = iov;
1839 			in_iovs = 1;
1840 		}
1841 
1842 		if (_IOC_DIR(cmd) & _IOC_READ) {
1843 			out_iov = iov;
1844 			out_iovs = 1;
1845 		}
1846 	}
1847 
1848  retry:
1849 	inarg.in_size = in_size = iov_length(in_iov, in_iovs);
1850 	inarg.out_size = out_size = iov_length(out_iov, out_iovs);
1851 
1852 	/*
1853 	 * Out data can be used either for actual out data or iovs,
1854 	 * make sure there always is at least one page.
1855 	 */
1856 	out_size = max_t(size_t, out_size, PAGE_SIZE);
1857 	max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
1858 
1859 	/* make sure there are enough buffer pages and init request with them */
1860 	err = -ENOMEM;
1861 	if (max_pages > FUSE_MAX_PAGES_PER_REQ)
1862 		goto out;
1863 	while (num_pages < max_pages) {
1864 		pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
1865 		if (!pages[num_pages])
1866 			goto out;
1867 		num_pages++;
1868 	}
1869 
1870 	req = fuse_get_req(fc);
1871 	if (IS_ERR(req)) {
1872 		err = PTR_ERR(req);
1873 		req = NULL;
1874 		goto out;
1875 	}
1876 	memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
1877 	req->num_pages = num_pages;
1878 
1879 	/* okay, let's send it to the client */
1880 	req->in.h.opcode = FUSE_IOCTL;
1881 	req->in.h.nodeid = ff->nodeid;
1882 	req->in.numargs = 1;
1883 	req->in.args[0].size = sizeof(inarg);
1884 	req->in.args[0].value = &inarg;
1885 	if (in_size) {
1886 		req->in.numargs++;
1887 		req->in.args[1].size = in_size;
1888 		req->in.argpages = 1;
1889 
1890 		err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
1891 					   false);
1892 		if (err)
1893 			goto out;
1894 	}
1895 
1896 	req->out.numargs = 2;
1897 	req->out.args[0].size = sizeof(outarg);
1898 	req->out.args[0].value = &outarg;
1899 	req->out.args[1].size = out_size;
1900 	req->out.argpages = 1;
1901 	req->out.argvar = 1;
1902 
1903 	fuse_request_send(fc, req);
1904 	err = req->out.h.error;
1905 	transferred = req->out.args[1].size;
1906 	fuse_put_request(fc, req);
1907 	req = NULL;
1908 	if (err)
1909 		goto out;
1910 
1911 	/* did it ask for retry? */
1912 	if (outarg.flags & FUSE_IOCTL_RETRY) {
1913 		void *vaddr;
1914 
1915 		/* no retry if in restricted mode */
1916 		err = -EIO;
1917 		if (!(flags & FUSE_IOCTL_UNRESTRICTED))
1918 			goto out;
1919 
1920 		in_iovs = outarg.in_iovs;
1921 		out_iovs = outarg.out_iovs;
1922 
1923 		/*
1924 		 * Make sure things are in boundary, separate checks
1925 		 * are to protect against overflow.
1926 		 */
1927 		err = -ENOMEM;
1928 		if (in_iovs > FUSE_IOCTL_MAX_IOV ||
1929 		    out_iovs > FUSE_IOCTL_MAX_IOV ||
1930 		    in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
1931 			goto out;
1932 
1933 		vaddr = kmap_atomic(pages[0]);
1934 		err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
1935 					    transferred, in_iovs + out_iovs,
1936 					    (flags & FUSE_IOCTL_COMPAT) != 0);
1937 		kunmap_atomic(vaddr);
1938 		if (err)
1939 			goto out;
1940 
1941 		in_iov = iov_page;
1942 		out_iov = in_iov + in_iovs;
1943 
1944 		err = fuse_verify_ioctl_iov(in_iov, in_iovs);
1945 		if (err)
1946 			goto out;
1947 
1948 		err = fuse_verify_ioctl_iov(out_iov, out_iovs);
1949 		if (err)
1950 			goto out;
1951 
1952 		goto retry;
1953 	}
1954 
1955 	err = -EIO;
1956 	if (transferred > inarg.out_size)
1957 		goto out;
1958 
1959 	err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
1960  out:
1961 	if (req)
1962 		fuse_put_request(fc, req);
1963 	free_page((unsigned long) iov_page);
1964 	while (num_pages)
1965 		__free_page(pages[--num_pages]);
1966 	kfree(pages);
1967 
1968 	return err ? err : outarg.result;
1969 }
1970 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
1971 
1972 long fuse_ioctl_common(struct file *file, unsigned int cmd,
1973 		       unsigned long arg, unsigned int flags)
1974 {
1975 	struct inode *inode = file->f_dentry->d_inode;
1976 	struct fuse_conn *fc = get_fuse_conn(inode);
1977 
1978 	if (!fuse_allow_task(fc, current))
1979 		return -EACCES;
1980 
1981 	if (is_bad_inode(inode))
1982 		return -EIO;
1983 
1984 	return fuse_do_ioctl(file, cmd, arg, flags);
1985 }
1986 
1987 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
1988 			    unsigned long arg)
1989 {
1990 	return fuse_ioctl_common(file, cmd, arg, 0);
1991 }
1992 
1993 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
1994 				   unsigned long arg)
1995 {
1996 	return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
1997 }
1998 
1999 /*
2000  * All files which have been polled are linked to RB tree
2001  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2002  * find the matching one.
2003  */
2004 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2005 					      struct rb_node **parent_out)
2006 {
2007 	struct rb_node **link = &fc->polled_files.rb_node;
2008 	struct rb_node *last = NULL;
2009 
2010 	while (*link) {
2011 		struct fuse_file *ff;
2012 
2013 		last = *link;
2014 		ff = rb_entry(last, struct fuse_file, polled_node);
2015 
2016 		if (kh < ff->kh)
2017 			link = &last->rb_left;
2018 		else if (kh > ff->kh)
2019 			link = &last->rb_right;
2020 		else
2021 			return link;
2022 	}
2023 
2024 	if (parent_out)
2025 		*parent_out = last;
2026 	return link;
2027 }
2028 
2029 /*
2030  * The file is about to be polled.  Make sure it's on the polled_files
2031  * RB tree.  Note that files once added to the polled_files tree are
2032  * not removed before the file is released.  This is because a file
2033  * polled once is likely to be polled again.
2034  */
2035 static void fuse_register_polled_file(struct fuse_conn *fc,
2036 				      struct fuse_file *ff)
2037 {
2038 	spin_lock(&fc->lock);
2039 	if (RB_EMPTY_NODE(&ff->polled_node)) {
2040 		struct rb_node **link, *parent;
2041 
2042 		link = fuse_find_polled_node(fc, ff->kh, &parent);
2043 		BUG_ON(*link);
2044 		rb_link_node(&ff->polled_node, parent, link);
2045 		rb_insert_color(&ff->polled_node, &fc->polled_files);
2046 	}
2047 	spin_unlock(&fc->lock);
2048 }
2049 
2050 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2051 {
2052 	struct fuse_file *ff = file->private_data;
2053 	struct fuse_conn *fc = ff->fc;
2054 	struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2055 	struct fuse_poll_out outarg;
2056 	struct fuse_req *req;
2057 	int err;
2058 
2059 	if (fc->no_poll)
2060 		return DEFAULT_POLLMASK;
2061 
2062 	poll_wait(file, &ff->poll_wait, wait);
2063 
2064 	/*
2065 	 * Ask for notification iff there's someone waiting for it.
2066 	 * The client may ignore the flag and always notify.
2067 	 */
2068 	if (waitqueue_active(&ff->poll_wait)) {
2069 		inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2070 		fuse_register_polled_file(fc, ff);
2071 	}
2072 
2073 	req = fuse_get_req(fc);
2074 	if (IS_ERR(req))
2075 		return POLLERR;
2076 
2077 	req->in.h.opcode = FUSE_POLL;
2078 	req->in.h.nodeid = ff->nodeid;
2079 	req->in.numargs = 1;
2080 	req->in.args[0].size = sizeof(inarg);
2081 	req->in.args[0].value = &inarg;
2082 	req->out.numargs = 1;
2083 	req->out.args[0].size = sizeof(outarg);
2084 	req->out.args[0].value = &outarg;
2085 	fuse_request_send(fc, req);
2086 	err = req->out.h.error;
2087 	fuse_put_request(fc, req);
2088 
2089 	if (!err)
2090 		return outarg.revents;
2091 	if (err == -ENOSYS) {
2092 		fc->no_poll = 1;
2093 		return DEFAULT_POLLMASK;
2094 	}
2095 	return POLLERR;
2096 }
2097 EXPORT_SYMBOL_GPL(fuse_file_poll);
2098 
2099 /*
2100  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2101  * wakes up the poll waiters.
2102  */
2103 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2104 			    struct fuse_notify_poll_wakeup_out *outarg)
2105 {
2106 	u64 kh = outarg->kh;
2107 	struct rb_node **link;
2108 
2109 	spin_lock(&fc->lock);
2110 
2111 	link = fuse_find_polled_node(fc, kh, NULL);
2112 	if (*link) {
2113 		struct fuse_file *ff;
2114 
2115 		ff = rb_entry(*link, struct fuse_file, polled_node);
2116 		wake_up_interruptible_sync(&ff->poll_wait);
2117 	}
2118 
2119 	spin_unlock(&fc->lock);
2120 	return 0;
2121 }
2122 
2123 static ssize_t fuse_loop_dio(struct file *filp, const struct iovec *iov,
2124 			     unsigned long nr_segs, loff_t *ppos, int rw)
2125 {
2126 	const struct iovec *vector = iov;
2127 	ssize_t ret = 0;
2128 
2129 	while (nr_segs > 0) {
2130 		void __user *base;
2131 		size_t len;
2132 		ssize_t nr;
2133 
2134 		base = vector->iov_base;
2135 		len = vector->iov_len;
2136 		vector++;
2137 		nr_segs--;
2138 
2139 		if (rw == WRITE)
2140 			nr = __fuse_direct_write(filp, base, len, ppos);
2141 		else
2142 			nr = fuse_direct_read(filp, base, len, ppos);
2143 
2144 		if (nr < 0) {
2145 			if (!ret)
2146 				ret = nr;
2147 			break;
2148 		}
2149 		ret += nr;
2150 		if (nr != len)
2151 			break;
2152 	}
2153 
2154 	return ret;
2155 }
2156 
2157 
2158 static ssize_t
2159 fuse_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2160 			loff_t offset, unsigned long nr_segs)
2161 {
2162 	ssize_t ret = 0;
2163 	struct file *file = NULL;
2164 	loff_t pos = 0;
2165 
2166 	file = iocb->ki_filp;
2167 	pos = offset;
2168 
2169 	ret = fuse_loop_dio(file, iov, nr_segs, &pos, rw);
2170 
2171 	return ret;
2172 }
2173 
2174 static const struct file_operations fuse_file_operations = {
2175 	.llseek		= fuse_file_llseek,
2176 	.read		= do_sync_read,
2177 	.aio_read	= fuse_file_aio_read,
2178 	.write		= do_sync_write,
2179 	.aio_write	= fuse_file_aio_write,
2180 	.mmap		= fuse_file_mmap,
2181 	.open		= fuse_open,
2182 	.flush		= fuse_flush,
2183 	.release	= fuse_release,
2184 	.fsync		= fuse_fsync,
2185 	.lock		= fuse_file_lock,
2186 	.flock		= fuse_file_flock,
2187 	.splice_read	= generic_file_splice_read,
2188 	.unlocked_ioctl	= fuse_file_ioctl,
2189 	.compat_ioctl	= fuse_file_compat_ioctl,
2190 	.poll		= fuse_file_poll,
2191 };
2192 
2193 static const struct file_operations fuse_direct_io_file_operations = {
2194 	.llseek		= fuse_file_llseek,
2195 	.read		= fuse_direct_read,
2196 	.write		= fuse_direct_write,
2197 	.mmap		= fuse_direct_mmap,
2198 	.open		= fuse_open,
2199 	.flush		= fuse_flush,
2200 	.release	= fuse_release,
2201 	.fsync		= fuse_fsync,
2202 	.lock		= fuse_file_lock,
2203 	.flock		= fuse_file_flock,
2204 	.unlocked_ioctl	= fuse_file_ioctl,
2205 	.compat_ioctl	= fuse_file_compat_ioctl,
2206 	.poll		= fuse_file_poll,
2207 	/* no splice_read */
2208 };
2209 
2210 static const struct address_space_operations fuse_file_aops  = {
2211 	.readpage	= fuse_readpage,
2212 	.writepage	= fuse_writepage,
2213 	.launder_page	= fuse_launder_page,
2214 	.readpages	= fuse_readpages,
2215 	.set_page_dirty	= __set_page_dirty_nobuffers,
2216 	.bmap		= fuse_bmap,
2217 	.direct_IO	= fuse_direct_IO,
2218 };
2219 
2220 void fuse_init_file_inode(struct inode *inode)
2221 {
2222 	inode->i_fop = &fuse_file_operations;
2223 	inode->i_data.a_ops = &fuse_file_aops;
2224 }
2225