xref: /linux-6.15/fs/fuse/dev.c (revision a669c2df)
1 /*
2   FUSE: Filesystem in Userspace
3   Copyright (C) 2001-2008  Miklos Szeredi <[email protected]>
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/init.h>
12 #include <linux/module.h>
13 #include <linux/poll.h>
14 #include <linux/sched/signal.h>
15 #include <linux/uio.h>
16 #include <linux/miscdevice.h>
17 #include <linux/pagemap.h>
18 #include <linux/file.h>
19 #include <linux/slab.h>
20 #include <linux/pipe_fs_i.h>
21 #include <linux/swap.h>
22 #include <linux/splice.h>
23 #include <linux/sched.h>
24 
25 #define CREATE_TRACE_POINTS
26 #include "fuse_trace.h"
27 
28 MODULE_ALIAS_MISCDEV(FUSE_MINOR);
29 MODULE_ALIAS("devname:fuse");
30 
31 /* Ordinary requests have even IDs, while interrupts IDs are odd */
32 #define FUSE_INT_REQ_BIT (1ULL << 0)
33 #define FUSE_REQ_ID_STEP (1ULL << 1)
34 
35 static struct kmem_cache *fuse_req_cachep;
36 
37 static void end_requests(struct list_head *head);
38 
39 static struct fuse_dev *fuse_get_dev(struct file *file)
40 {
41 	/*
42 	 * Lockless access is OK, because file->private data is set
43 	 * once during mount and is valid until the file is released.
44 	 */
45 	return READ_ONCE(file->private_data);
46 }
47 
48 static void fuse_request_init(struct fuse_mount *fm, struct fuse_req *req)
49 {
50 	INIT_LIST_HEAD(&req->list);
51 	INIT_LIST_HEAD(&req->intr_entry);
52 	init_waitqueue_head(&req->waitq);
53 	refcount_set(&req->count, 1);
54 	__set_bit(FR_PENDING, &req->flags);
55 	req->fm = fm;
56 }
57 
58 static struct fuse_req *fuse_request_alloc(struct fuse_mount *fm, gfp_t flags)
59 {
60 	struct fuse_req *req = kmem_cache_zalloc(fuse_req_cachep, flags);
61 	if (req)
62 		fuse_request_init(fm, req);
63 
64 	return req;
65 }
66 
67 static void fuse_request_free(struct fuse_req *req)
68 {
69 	kmem_cache_free(fuse_req_cachep, req);
70 }
71 
72 static void __fuse_get_request(struct fuse_req *req)
73 {
74 	refcount_inc(&req->count);
75 }
76 
77 /* Must be called with > 1 refcount */
78 static void __fuse_put_request(struct fuse_req *req)
79 {
80 	refcount_dec(&req->count);
81 }
82 
83 void fuse_set_initialized(struct fuse_conn *fc)
84 {
85 	/* Make sure stores before this are seen on another CPU */
86 	smp_wmb();
87 	fc->initialized = 1;
88 }
89 
90 static bool fuse_block_alloc(struct fuse_conn *fc, bool for_background)
91 {
92 	return !fc->initialized || (for_background && fc->blocked);
93 }
94 
95 static void fuse_drop_waiting(struct fuse_conn *fc)
96 {
97 	/*
98 	 * lockess check of fc->connected is okay, because atomic_dec_and_test()
99 	 * provides a memory barrier matched with the one in fuse_wait_aborted()
100 	 * to ensure no wake-up is missed.
101 	 */
102 	if (atomic_dec_and_test(&fc->num_waiting) &&
103 	    !READ_ONCE(fc->connected)) {
104 		/* wake up aborters */
105 		wake_up_all(&fc->blocked_waitq);
106 	}
107 }
108 
109 static void fuse_put_request(struct fuse_req *req);
110 
111 static struct fuse_req *fuse_get_req(struct mnt_idmap *idmap,
112 				     struct fuse_mount *fm,
113 				     bool for_background)
114 {
115 	struct fuse_conn *fc = fm->fc;
116 	struct fuse_req *req;
117 	bool no_idmap = !fm->sb || (fm->sb->s_iflags & SB_I_NOIDMAP);
118 	kuid_t fsuid;
119 	kgid_t fsgid;
120 	int err;
121 
122 	atomic_inc(&fc->num_waiting);
123 
124 	if (fuse_block_alloc(fc, for_background)) {
125 		err = -EINTR;
126 		if (wait_event_killable_exclusive(fc->blocked_waitq,
127 				!fuse_block_alloc(fc, for_background)))
128 			goto out;
129 	}
130 	/* Matches smp_wmb() in fuse_set_initialized() */
131 	smp_rmb();
132 
133 	err = -ENOTCONN;
134 	if (!fc->connected)
135 		goto out;
136 
137 	err = -ECONNREFUSED;
138 	if (fc->conn_error)
139 		goto out;
140 
141 	req = fuse_request_alloc(fm, GFP_KERNEL);
142 	err = -ENOMEM;
143 	if (!req) {
144 		if (for_background)
145 			wake_up(&fc->blocked_waitq);
146 		goto out;
147 	}
148 
149 	req->in.h.pid = pid_nr_ns(task_pid(current), fc->pid_ns);
150 
151 	__set_bit(FR_WAITING, &req->flags);
152 	if (for_background)
153 		__set_bit(FR_BACKGROUND, &req->flags);
154 
155 	/*
156 	 * Keep the old behavior when idmappings support was not
157 	 * declared by a FUSE server.
158 	 *
159 	 * For those FUSE servers who support idmapped mounts,
160 	 * we send UID/GID only along with "inode creation"
161 	 * fuse requests, otherwise idmap == &invalid_mnt_idmap and
162 	 * req->in.h.{u,g}id will be equal to FUSE_INVALID_UIDGID.
163 	 */
164 	fsuid = no_idmap ? current_fsuid() : mapped_fsuid(idmap, fc->user_ns);
165 	fsgid = no_idmap ? current_fsgid() : mapped_fsgid(idmap, fc->user_ns);
166 	req->in.h.uid = from_kuid(fc->user_ns, fsuid);
167 	req->in.h.gid = from_kgid(fc->user_ns, fsgid);
168 
169 	if (no_idmap && unlikely(req->in.h.uid == ((uid_t)-1) ||
170 				 req->in.h.gid == ((gid_t)-1))) {
171 		fuse_put_request(req);
172 		return ERR_PTR(-EOVERFLOW);
173 	}
174 
175 	return req;
176 
177  out:
178 	fuse_drop_waiting(fc);
179 	return ERR_PTR(err);
180 }
181 
182 static void fuse_put_request(struct fuse_req *req)
183 {
184 	struct fuse_conn *fc = req->fm->fc;
185 
186 	if (refcount_dec_and_test(&req->count)) {
187 		if (test_bit(FR_BACKGROUND, &req->flags)) {
188 			/*
189 			 * We get here in the unlikely case that a background
190 			 * request was allocated but not sent
191 			 */
192 			spin_lock(&fc->bg_lock);
193 			if (!fc->blocked)
194 				wake_up(&fc->blocked_waitq);
195 			spin_unlock(&fc->bg_lock);
196 		}
197 
198 		if (test_bit(FR_WAITING, &req->flags)) {
199 			__clear_bit(FR_WAITING, &req->flags);
200 			fuse_drop_waiting(fc);
201 		}
202 
203 		fuse_request_free(req);
204 	}
205 }
206 
207 unsigned int fuse_len_args(unsigned int numargs, struct fuse_arg *args)
208 {
209 	unsigned nbytes = 0;
210 	unsigned i;
211 
212 	for (i = 0; i < numargs; i++)
213 		nbytes += args[i].size;
214 
215 	return nbytes;
216 }
217 EXPORT_SYMBOL_GPL(fuse_len_args);
218 
219 static u64 fuse_get_unique_locked(struct fuse_iqueue *fiq)
220 {
221 	fiq->reqctr += FUSE_REQ_ID_STEP;
222 	return fiq->reqctr;
223 }
224 
225 u64 fuse_get_unique(struct fuse_iqueue *fiq)
226 {
227 	u64 ret;
228 
229 	spin_lock(&fiq->lock);
230 	ret = fuse_get_unique_locked(fiq);
231 	spin_unlock(&fiq->lock);
232 
233 	return ret;
234 }
235 EXPORT_SYMBOL_GPL(fuse_get_unique);
236 
237 static unsigned int fuse_req_hash(u64 unique)
238 {
239 	return hash_long(unique & ~FUSE_INT_REQ_BIT, FUSE_PQ_HASH_BITS);
240 }
241 
242 /*
243  * A new request is available, wake fiq->waitq
244  */
245 static void fuse_dev_wake_and_unlock(struct fuse_iqueue *fiq)
246 __releases(fiq->lock)
247 {
248 	wake_up(&fiq->waitq);
249 	kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
250 	spin_unlock(&fiq->lock);
251 }
252 
253 static void fuse_dev_queue_forget(struct fuse_iqueue *fiq, struct fuse_forget_link *forget)
254 {
255 	spin_lock(&fiq->lock);
256 	if (fiq->connected) {
257 		fiq->forget_list_tail->next = forget;
258 		fiq->forget_list_tail = forget;
259 		fuse_dev_wake_and_unlock(fiq);
260 	} else {
261 		kfree(forget);
262 		spin_unlock(&fiq->lock);
263 	}
264 }
265 
266 static void fuse_dev_queue_interrupt(struct fuse_iqueue *fiq, struct fuse_req *req)
267 {
268 	spin_lock(&fiq->lock);
269 	if (list_empty(&req->intr_entry)) {
270 		list_add_tail(&req->intr_entry, &fiq->interrupts);
271 		/*
272 		 * Pairs with smp_mb() implied by test_and_set_bit()
273 		 * from fuse_request_end().
274 		 */
275 		smp_mb();
276 		if (test_bit(FR_FINISHED, &req->flags)) {
277 			list_del_init(&req->intr_entry);
278 			spin_unlock(&fiq->lock);
279 		} else  {
280 			fuse_dev_wake_and_unlock(fiq);
281 		}
282 	} else {
283 		spin_unlock(&fiq->lock);
284 	}
285 }
286 
287 static void fuse_dev_queue_req(struct fuse_iqueue *fiq, struct fuse_req *req)
288 {
289 	spin_lock(&fiq->lock);
290 	if (fiq->connected) {
291 		if (req->in.h.opcode != FUSE_NOTIFY_REPLY)
292 			req->in.h.unique = fuse_get_unique_locked(fiq);
293 		list_add_tail(&req->list, &fiq->pending);
294 		fuse_dev_wake_and_unlock(fiq);
295 	} else {
296 		spin_unlock(&fiq->lock);
297 		req->out.h.error = -ENOTCONN;
298 		clear_bit(FR_PENDING, &req->flags);
299 		fuse_request_end(req);
300 	}
301 }
302 
303 const struct fuse_iqueue_ops fuse_dev_fiq_ops = {
304 	.send_forget	= fuse_dev_queue_forget,
305 	.send_interrupt	= fuse_dev_queue_interrupt,
306 	.send_req	= fuse_dev_queue_req,
307 };
308 EXPORT_SYMBOL_GPL(fuse_dev_fiq_ops);
309 
310 static void fuse_send_one(struct fuse_iqueue *fiq, struct fuse_req *req)
311 {
312 	req->in.h.len = sizeof(struct fuse_in_header) +
313 		fuse_len_args(req->args->in_numargs,
314 			      (struct fuse_arg *) req->args->in_args);
315 	trace_fuse_request_send(req);
316 	fiq->ops->send_req(fiq, req);
317 }
318 
319 void fuse_queue_forget(struct fuse_conn *fc, struct fuse_forget_link *forget,
320 		       u64 nodeid, u64 nlookup)
321 {
322 	struct fuse_iqueue *fiq = &fc->iq;
323 
324 	forget->forget_one.nodeid = nodeid;
325 	forget->forget_one.nlookup = nlookup;
326 
327 	fiq->ops->send_forget(fiq, forget);
328 }
329 
330 static void flush_bg_queue(struct fuse_conn *fc)
331 {
332 	struct fuse_iqueue *fiq = &fc->iq;
333 
334 	while (fc->active_background < fc->max_background &&
335 	       !list_empty(&fc->bg_queue)) {
336 		struct fuse_req *req;
337 
338 		req = list_first_entry(&fc->bg_queue, struct fuse_req, list);
339 		list_del(&req->list);
340 		fc->active_background++;
341 		fuse_send_one(fiq, req);
342 	}
343 }
344 
345 /*
346  * This function is called when a request is finished.  Either a reply
347  * has arrived or it was aborted (and not yet sent) or some error
348  * occurred during communication with userspace, or the device file
349  * was closed.  The requester thread is woken up (if still waiting),
350  * the 'end' callback is called if given, else the reference to the
351  * request is released
352  */
353 void fuse_request_end(struct fuse_req *req)
354 {
355 	struct fuse_mount *fm = req->fm;
356 	struct fuse_conn *fc = fm->fc;
357 	struct fuse_iqueue *fiq = &fc->iq;
358 
359 	if (test_and_set_bit(FR_FINISHED, &req->flags))
360 		goto put_request;
361 
362 	trace_fuse_request_end(req);
363 	/*
364 	 * test_and_set_bit() implies smp_mb() between bit
365 	 * changing and below FR_INTERRUPTED check. Pairs with
366 	 * smp_mb() from queue_interrupt().
367 	 */
368 	if (test_bit(FR_INTERRUPTED, &req->flags)) {
369 		spin_lock(&fiq->lock);
370 		list_del_init(&req->intr_entry);
371 		spin_unlock(&fiq->lock);
372 	}
373 	WARN_ON(test_bit(FR_PENDING, &req->flags));
374 	WARN_ON(test_bit(FR_SENT, &req->flags));
375 	if (test_bit(FR_BACKGROUND, &req->flags)) {
376 		spin_lock(&fc->bg_lock);
377 		clear_bit(FR_BACKGROUND, &req->flags);
378 		if (fc->num_background == fc->max_background) {
379 			fc->blocked = 0;
380 			wake_up(&fc->blocked_waitq);
381 		} else if (!fc->blocked) {
382 			/*
383 			 * Wake up next waiter, if any.  It's okay to use
384 			 * waitqueue_active(), as we've already synced up
385 			 * fc->blocked with waiters with the wake_up() call
386 			 * above.
387 			 */
388 			if (waitqueue_active(&fc->blocked_waitq))
389 				wake_up(&fc->blocked_waitq);
390 		}
391 
392 		fc->num_background--;
393 		fc->active_background--;
394 		flush_bg_queue(fc);
395 		spin_unlock(&fc->bg_lock);
396 	} else {
397 		/* Wake up waiter sleeping in request_wait_answer() */
398 		wake_up(&req->waitq);
399 	}
400 
401 	if (test_bit(FR_ASYNC, &req->flags))
402 		req->args->end(fm, req->args, req->out.h.error);
403 put_request:
404 	fuse_put_request(req);
405 }
406 EXPORT_SYMBOL_GPL(fuse_request_end);
407 
408 static int queue_interrupt(struct fuse_req *req)
409 {
410 	struct fuse_iqueue *fiq = &req->fm->fc->iq;
411 
412 	/* Check for we've sent request to interrupt this req */
413 	if (unlikely(!test_bit(FR_INTERRUPTED, &req->flags)))
414 		return -EINVAL;
415 
416 	fiq->ops->send_interrupt(fiq, req);
417 
418 	return 0;
419 }
420 
421 static void request_wait_answer(struct fuse_req *req)
422 {
423 	struct fuse_conn *fc = req->fm->fc;
424 	struct fuse_iqueue *fiq = &fc->iq;
425 	int err;
426 
427 	if (!fc->no_interrupt) {
428 		/* Any signal may interrupt this */
429 		err = wait_event_interruptible(req->waitq,
430 					test_bit(FR_FINISHED, &req->flags));
431 		if (!err)
432 			return;
433 
434 		set_bit(FR_INTERRUPTED, &req->flags);
435 		/* matches barrier in fuse_dev_do_read() */
436 		smp_mb__after_atomic();
437 		if (test_bit(FR_SENT, &req->flags))
438 			queue_interrupt(req);
439 	}
440 
441 	if (!test_bit(FR_FORCE, &req->flags)) {
442 		/* Only fatal signals may interrupt this */
443 		err = wait_event_killable(req->waitq,
444 					test_bit(FR_FINISHED, &req->flags));
445 		if (!err)
446 			return;
447 
448 		spin_lock(&fiq->lock);
449 		/* Request is not yet in userspace, bail out */
450 		if (test_bit(FR_PENDING, &req->flags)) {
451 			list_del(&req->list);
452 			spin_unlock(&fiq->lock);
453 			__fuse_put_request(req);
454 			req->out.h.error = -EINTR;
455 			return;
456 		}
457 		spin_unlock(&fiq->lock);
458 	}
459 
460 	/*
461 	 * Either request is already in userspace, or it was forced.
462 	 * Wait it out.
463 	 */
464 	wait_event(req->waitq, test_bit(FR_FINISHED, &req->flags));
465 }
466 
467 static void __fuse_request_send(struct fuse_req *req)
468 {
469 	struct fuse_iqueue *fiq = &req->fm->fc->iq;
470 
471 	BUG_ON(test_bit(FR_BACKGROUND, &req->flags));
472 
473 	/* acquire extra reference, since request is still needed after
474 	   fuse_request_end() */
475 	__fuse_get_request(req);
476 	fuse_send_one(fiq, req);
477 
478 	request_wait_answer(req);
479 	/* Pairs with smp_wmb() in fuse_request_end() */
480 	smp_rmb();
481 }
482 
483 static void fuse_adjust_compat(struct fuse_conn *fc, struct fuse_args *args)
484 {
485 	if (fc->minor < 4 && args->opcode == FUSE_STATFS)
486 		args->out_args[0].size = FUSE_COMPAT_STATFS_SIZE;
487 
488 	if (fc->minor < 9) {
489 		switch (args->opcode) {
490 		case FUSE_LOOKUP:
491 		case FUSE_CREATE:
492 		case FUSE_MKNOD:
493 		case FUSE_MKDIR:
494 		case FUSE_SYMLINK:
495 		case FUSE_LINK:
496 			args->out_args[0].size = FUSE_COMPAT_ENTRY_OUT_SIZE;
497 			break;
498 		case FUSE_GETATTR:
499 		case FUSE_SETATTR:
500 			args->out_args[0].size = FUSE_COMPAT_ATTR_OUT_SIZE;
501 			break;
502 		}
503 	}
504 	if (fc->minor < 12) {
505 		switch (args->opcode) {
506 		case FUSE_CREATE:
507 			args->in_args[0].size = sizeof(struct fuse_open_in);
508 			break;
509 		case FUSE_MKNOD:
510 			args->in_args[0].size = FUSE_COMPAT_MKNOD_IN_SIZE;
511 			break;
512 		}
513 	}
514 }
515 
516 static void fuse_force_creds(struct fuse_req *req)
517 {
518 	struct fuse_conn *fc = req->fm->fc;
519 
520 	if (!req->fm->sb || req->fm->sb->s_iflags & SB_I_NOIDMAP) {
521 		req->in.h.uid = from_kuid_munged(fc->user_ns, current_fsuid());
522 		req->in.h.gid = from_kgid_munged(fc->user_ns, current_fsgid());
523 	} else {
524 		req->in.h.uid = FUSE_INVALID_UIDGID;
525 		req->in.h.gid = FUSE_INVALID_UIDGID;
526 	}
527 
528 	req->in.h.pid = pid_nr_ns(task_pid(current), fc->pid_ns);
529 }
530 
531 static void fuse_args_to_req(struct fuse_req *req, struct fuse_args *args)
532 {
533 	req->in.h.opcode = args->opcode;
534 	req->in.h.nodeid = args->nodeid;
535 	req->args = args;
536 	if (args->is_ext)
537 		req->in.h.total_extlen = args->in_args[args->ext_idx].size / 8;
538 	if (args->end)
539 		__set_bit(FR_ASYNC, &req->flags);
540 }
541 
542 ssize_t __fuse_simple_request(struct mnt_idmap *idmap,
543 			      struct fuse_mount *fm,
544 			      struct fuse_args *args)
545 {
546 	struct fuse_conn *fc = fm->fc;
547 	struct fuse_req *req;
548 	ssize_t ret;
549 
550 	if (args->force) {
551 		atomic_inc(&fc->num_waiting);
552 		req = fuse_request_alloc(fm, GFP_KERNEL | __GFP_NOFAIL);
553 
554 		if (!args->nocreds)
555 			fuse_force_creds(req);
556 
557 		__set_bit(FR_WAITING, &req->flags);
558 		__set_bit(FR_FORCE, &req->flags);
559 	} else {
560 		WARN_ON(args->nocreds);
561 		req = fuse_get_req(idmap, fm, false);
562 		if (IS_ERR(req))
563 			return PTR_ERR(req);
564 	}
565 
566 	/* Needs to be done after fuse_get_req() so that fc->minor is valid */
567 	fuse_adjust_compat(fc, args);
568 	fuse_args_to_req(req, args);
569 
570 	if (!args->noreply)
571 		__set_bit(FR_ISREPLY, &req->flags);
572 	__fuse_request_send(req);
573 	ret = req->out.h.error;
574 	if (!ret && args->out_argvar) {
575 		BUG_ON(args->out_numargs == 0);
576 		ret = args->out_args[args->out_numargs - 1].size;
577 	}
578 	fuse_put_request(req);
579 
580 	return ret;
581 }
582 
583 static bool fuse_request_queue_background(struct fuse_req *req)
584 {
585 	struct fuse_mount *fm = req->fm;
586 	struct fuse_conn *fc = fm->fc;
587 	bool queued = false;
588 
589 	WARN_ON(!test_bit(FR_BACKGROUND, &req->flags));
590 	if (!test_bit(FR_WAITING, &req->flags)) {
591 		__set_bit(FR_WAITING, &req->flags);
592 		atomic_inc(&fc->num_waiting);
593 	}
594 	__set_bit(FR_ISREPLY, &req->flags);
595 	spin_lock(&fc->bg_lock);
596 	if (likely(fc->connected)) {
597 		fc->num_background++;
598 		if (fc->num_background == fc->max_background)
599 			fc->blocked = 1;
600 		list_add_tail(&req->list, &fc->bg_queue);
601 		flush_bg_queue(fc);
602 		queued = true;
603 	}
604 	spin_unlock(&fc->bg_lock);
605 
606 	return queued;
607 }
608 
609 int fuse_simple_background(struct fuse_mount *fm, struct fuse_args *args,
610 			    gfp_t gfp_flags)
611 {
612 	struct fuse_req *req;
613 
614 	if (args->force) {
615 		WARN_ON(!args->nocreds);
616 		req = fuse_request_alloc(fm, gfp_flags);
617 		if (!req)
618 			return -ENOMEM;
619 		__set_bit(FR_BACKGROUND, &req->flags);
620 	} else {
621 		WARN_ON(args->nocreds);
622 		req = fuse_get_req(&invalid_mnt_idmap, fm, true);
623 		if (IS_ERR(req))
624 			return PTR_ERR(req);
625 	}
626 
627 	fuse_args_to_req(req, args);
628 
629 	if (!fuse_request_queue_background(req)) {
630 		fuse_put_request(req);
631 		return -ENOTCONN;
632 	}
633 
634 	return 0;
635 }
636 EXPORT_SYMBOL_GPL(fuse_simple_background);
637 
638 static int fuse_simple_notify_reply(struct fuse_mount *fm,
639 				    struct fuse_args *args, u64 unique)
640 {
641 	struct fuse_req *req;
642 	struct fuse_iqueue *fiq = &fm->fc->iq;
643 
644 	req = fuse_get_req(&invalid_mnt_idmap, fm, false);
645 	if (IS_ERR(req))
646 		return PTR_ERR(req);
647 
648 	__clear_bit(FR_ISREPLY, &req->flags);
649 	req->in.h.unique = unique;
650 
651 	fuse_args_to_req(req, args);
652 
653 	fuse_send_one(fiq, req);
654 
655 	return 0;
656 }
657 
658 /*
659  * Lock the request.  Up to the next unlock_request() there mustn't be
660  * anything that could cause a page-fault.  If the request was already
661  * aborted bail out.
662  */
663 static int lock_request(struct fuse_req *req)
664 {
665 	int err = 0;
666 	if (req) {
667 		spin_lock(&req->waitq.lock);
668 		if (test_bit(FR_ABORTED, &req->flags))
669 			err = -ENOENT;
670 		else
671 			set_bit(FR_LOCKED, &req->flags);
672 		spin_unlock(&req->waitq.lock);
673 	}
674 	return err;
675 }
676 
677 /*
678  * Unlock request.  If it was aborted while locked, caller is responsible
679  * for unlocking and ending the request.
680  */
681 static int unlock_request(struct fuse_req *req)
682 {
683 	int err = 0;
684 	if (req) {
685 		spin_lock(&req->waitq.lock);
686 		if (test_bit(FR_ABORTED, &req->flags))
687 			err = -ENOENT;
688 		else
689 			clear_bit(FR_LOCKED, &req->flags);
690 		spin_unlock(&req->waitq.lock);
691 	}
692 	return err;
693 }
694 
695 struct fuse_copy_state {
696 	int write;
697 	struct fuse_req *req;
698 	struct iov_iter *iter;
699 	struct pipe_buffer *pipebufs;
700 	struct pipe_buffer *currbuf;
701 	struct pipe_inode_info *pipe;
702 	unsigned long nr_segs;
703 	struct page *pg;
704 	unsigned len;
705 	unsigned offset;
706 	unsigned move_pages:1;
707 };
708 
709 static void fuse_copy_init(struct fuse_copy_state *cs, int write,
710 			   struct iov_iter *iter)
711 {
712 	memset(cs, 0, sizeof(*cs));
713 	cs->write = write;
714 	cs->iter = iter;
715 }
716 
717 /* Unmap and put previous page of userspace buffer */
718 static void fuse_copy_finish(struct fuse_copy_state *cs)
719 {
720 	if (cs->currbuf) {
721 		struct pipe_buffer *buf = cs->currbuf;
722 
723 		if (cs->write)
724 			buf->len = PAGE_SIZE - cs->len;
725 		cs->currbuf = NULL;
726 	} else if (cs->pg) {
727 		if (cs->write) {
728 			flush_dcache_page(cs->pg);
729 			set_page_dirty_lock(cs->pg);
730 		}
731 		put_page(cs->pg);
732 	}
733 	cs->pg = NULL;
734 }
735 
736 /*
737  * Get another pagefull of userspace buffer, and map it to kernel
738  * address space, and lock request
739  */
740 static int fuse_copy_fill(struct fuse_copy_state *cs)
741 {
742 	struct page *page;
743 	int err;
744 
745 	err = unlock_request(cs->req);
746 	if (err)
747 		return err;
748 
749 	fuse_copy_finish(cs);
750 	if (cs->pipebufs) {
751 		struct pipe_buffer *buf = cs->pipebufs;
752 
753 		if (!cs->write) {
754 			err = pipe_buf_confirm(cs->pipe, buf);
755 			if (err)
756 				return err;
757 
758 			BUG_ON(!cs->nr_segs);
759 			cs->currbuf = buf;
760 			cs->pg = buf->page;
761 			cs->offset = buf->offset;
762 			cs->len = buf->len;
763 			cs->pipebufs++;
764 			cs->nr_segs--;
765 		} else {
766 			if (cs->nr_segs >= cs->pipe->max_usage)
767 				return -EIO;
768 
769 			page = alloc_page(GFP_HIGHUSER);
770 			if (!page)
771 				return -ENOMEM;
772 
773 			buf->page = page;
774 			buf->offset = 0;
775 			buf->len = 0;
776 
777 			cs->currbuf = buf;
778 			cs->pg = page;
779 			cs->offset = 0;
780 			cs->len = PAGE_SIZE;
781 			cs->pipebufs++;
782 			cs->nr_segs++;
783 		}
784 	} else {
785 		size_t off;
786 		err = iov_iter_get_pages2(cs->iter, &page, PAGE_SIZE, 1, &off);
787 		if (err < 0)
788 			return err;
789 		BUG_ON(!err);
790 		cs->len = err;
791 		cs->offset = off;
792 		cs->pg = page;
793 	}
794 
795 	return lock_request(cs->req);
796 }
797 
798 /* Do as much copy to/from userspace buffer as we can */
799 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
800 {
801 	unsigned ncpy = min(*size, cs->len);
802 	if (val) {
803 		void *pgaddr = kmap_local_page(cs->pg);
804 		void *buf = pgaddr + cs->offset;
805 
806 		if (cs->write)
807 			memcpy(buf, *val, ncpy);
808 		else
809 			memcpy(*val, buf, ncpy);
810 
811 		kunmap_local(pgaddr);
812 		*val += ncpy;
813 	}
814 	*size -= ncpy;
815 	cs->len -= ncpy;
816 	cs->offset += ncpy;
817 	return ncpy;
818 }
819 
820 static int fuse_check_folio(struct folio *folio)
821 {
822 	if (folio_mapped(folio) ||
823 	    folio->mapping != NULL ||
824 	    (folio->flags & PAGE_FLAGS_CHECK_AT_PREP &
825 	     ~(1 << PG_locked |
826 	       1 << PG_referenced |
827 	       1 << PG_lru |
828 	       1 << PG_active |
829 	       1 << PG_workingset |
830 	       1 << PG_reclaim |
831 	       1 << PG_waiters |
832 	       LRU_GEN_MASK | LRU_REFS_MASK))) {
833 		dump_page(&folio->page, "fuse: trying to steal weird page");
834 		return 1;
835 	}
836 	return 0;
837 }
838 
839 static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
840 {
841 	int err;
842 	struct folio *oldfolio = page_folio(*pagep);
843 	struct folio *newfolio;
844 	struct pipe_buffer *buf = cs->pipebufs;
845 
846 	folio_get(oldfolio);
847 	err = unlock_request(cs->req);
848 	if (err)
849 		goto out_put_old;
850 
851 	fuse_copy_finish(cs);
852 
853 	err = pipe_buf_confirm(cs->pipe, buf);
854 	if (err)
855 		goto out_put_old;
856 
857 	BUG_ON(!cs->nr_segs);
858 	cs->currbuf = buf;
859 	cs->len = buf->len;
860 	cs->pipebufs++;
861 	cs->nr_segs--;
862 
863 	if (cs->len != PAGE_SIZE)
864 		goto out_fallback;
865 
866 	if (!pipe_buf_try_steal(cs->pipe, buf))
867 		goto out_fallback;
868 
869 	newfolio = page_folio(buf->page);
870 
871 	folio_clear_uptodate(newfolio);
872 	folio_clear_mappedtodisk(newfolio);
873 
874 	if (fuse_check_folio(newfolio) != 0)
875 		goto out_fallback_unlock;
876 
877 	/*
878 	 * This is a new and locked page, it shouldn't be mapped or
879 	 * have any special flags on it
880 	 */
881 	if (WARN_ON(folio_mapped(oldfolio)))
882 		goto out_fallback_unlock;
883 	if (WARN_ON(folio_has_private(oldfolio)))
884 		goto out_fallback_unlock;
885 	if (WARN_ON(folio_test_dirty(oldfolio) ||
886 				folio_test_writeback(oldfolio)))
887 		goto out_fallback_unlock;
888 	if (WARN_ON(folio_test_mlocked(oldfolio)))
889 		goto out_fallback_unlock;
890 
891 	replace_page_cache_folio(oldfolio, newfolio);
892 
893 	folio_get(newfolio);
894 
895 	if (!(buf->flags & PIPE_BUF_FLAG_LRU))
896 		folio_add_lru(newfolio);
897 
898 	/*
899 	 * Release while we have extra ref on stolen page.  Otherwise
900 	 * anon_pipe_buf_release() might think the page can be reused.
901 	 */
902 	pipe_buf_release(cs->pipe, buf);
903 
904 	err = 0;
905 	spin_lock(&cs->req->waitq.lock);
906 	if (test_bit(FR_ABORTED, &cs->req->flags))
907 		err = -ENOENT;
908 	else
909 		*pagep = &newfolio->page;
910 	spin_unlock(&cs->req->waitq.lock);
911 
912 	if (err) {
913 		folio_unlock(newfolio);
914 		folio_put(newfolio);
915 		goto out_put_old;
916 	}
917 
918 	folio_unlock(oldfolio);
919 	/* Drop ref for ap->pages[] array */
920 	folio_put(oldfolio);
921 	cs->len = 0;
922 
923 	err = 0;
924 out_put_old:
925 	/* Drop ref obtained in this function */
926 	folio_put(oldfolio);
927 	return err;
928 
929 out_fallback_unlock:
930 	folio_unlock(newfolio);
931 out_fallback:
932 	cs->pg = buf->page;
933 	cs->offset = buf->offset;
934 
935 	err = lock_request(cs->req);
936 	if (!err)
937 		err = 1;
938 
939 	goto out_put_old;
940 }
941 
942 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
943 			 unsigned offset, unsigned count)
944 {
945 	struct pipe_buffer *buf;
946 	int err;
947 
948 	if (cs->nr_segs >= cs->pipe->max_usage)
949 		return -EIO;
950 
951 	get_page(page);
952 	err = unlock_request(cs->req);
953 	if (err) {
954 		put_page(page);
955 		return err;
956 	}
957 
958 	fuse_copy_finish(cs);
959 
960 	buf = cs->pipebufs;
961 	buf->page = page;
962 	buf->offset = offset;
963 	buf->len = count;
964 
965 	cs->pipebufs++;
966 	cs->nr_segs++;
967 	cs->len = 0;
968 
969 	return 0;
970 }
971 
972 /*
973  * Copy a page in the request to/from the userspace buffer.  Must be
974  * done atomically
975  */
976 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
977 			  unsigned offset, unsigned count, int zeroing)
978 {
979 	int err;
980 	struct page *page = *pagep;
981 
982 	if (page && zeroing && count < PAGE_SIZE)
983 		clear_highpage(page);
984 
985 	while (count) {
986 		if (cs->write && cs->pipebufs && page) {
987 			/*
988 			 * Can't control lifetime of pipe buffers, so always
989 			 * copy user pages.
990 			 */
991 			if (cs->req->args->user_pages) {
992 				err = fuse_copy_fill(cs);
993 				if (err)
994 					return err;
995 			} else {
996 				return fuse_ref_page(cs, page, offset, count);
997 			}
998 		} else if (!cs->len) {
999 			if (cs->move_pages && page &&
1000 			    offset == 0 && count == PAGE_SIZE) {
1001 				err = fuse_try_move_page(cs, pagep);
1002 				if (err <= 0)
1003 					return err;
1004 			} else {
1005 				err = fuse_copy_fill(cs);
1006 				if (err)
1007 					return err;
1008 			}
1009 		}
1010 		if (page) {
1011 			void *mapaddr = kmap_local_page(page);
1012 			void *buf = mapaddr + offset;
1013 			offset += fuse_copy_do(cs, &buf, &count);
1014 			kunmap_local(mapaddr);
1015 		} else
1016 			offset += fuse_copy_do(cs, NULL, &count);
1017 	}
1018 	if (page && !cs->write)
1019 		flush_dcache_page(page);
1020 	return 0;
1021 }
1022 
1023 /* Copy pages in the request to/from userspace buffer */
1024 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
1025 			   int zeroing)
1026 {
1027 	unsigned i;
1028 	struct fuse_req *req = cs->req;
1029 	struct fuse_args_pages *ap = container_of(req->args, typeof(*ap), args);
1030 
1031 	if (ap->uses_folios) {
1032 		for (i = 0; i < ap->num_folios && (nbytes || zeroing); i++) {
1033 			int err;
1034 			unsigned int offset = ap->folio_descs[i].offset;
1035 			unsigned int count = min(nbytes, ap->folio_descs[i].length);
1036 			struct page *orig, *pagep;
1037 
1038 			orig = pagep = &ap->folios[i]->page;
1039 
1040 			err = fuse_copy_page(cs, &pagep, offset, count, zeroing);
1041 			if (err)
1042 				return err;
1043 
1044 			nbytes -= count;
1045 
1046 			/*
1047 			 *  fuse_copy_page may have moved a page from a pipe
1048 			 *  instead of copying into our given page, so update
1049 			 *  the folios if it was replaced.
1050 			 */
1051 			if (pagep != orig)
1052 				ap->folios[i] = page_folio(pagep);
1053 		}
1054 	} else {
1055 		for (i = 0; i < ap->num_pages && (nbytes || zeroing); i++) {
1056 			int err;
1057 			unsigned int offset = ap->descs[i].offset;
1058 			unsigned int count = min(nbytes, ap->descs[i].length);
1059 
1060 			err = fuse_copy_page(cs, &ap->pages[i], offset, count, zeroing);
1061 			if (err)
1062 				return err;
1063 
1064 			nbytes -= count;
1065 		}
1066 	}
1067 	return 0;
1068 }
1069 
1070 /* Copy a single argument in the request to/from userspace buffer */
1071 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
1072 {
1073 	while (size) {
1074 		if (!cs->len) {
1075 			int err = fuse_copy_fill(cs);
1076 			if (err)
1077 				return err;
1078 		}
1079 		fuse_copy_do(cs, &val, &size);
1080 	}
1081 	return 0;
1082 }
1083 
1084 /* Copy request arguments to/from userspace buffer */
1085 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
1086 			  unsigned argpages, struct fuse_arg *args,
1087 			  int zeroing)
1088 {
1089 	int err = 0;
1090 	unsigned i;
1091 
1092 	for (i = 0; !err && i < numargs; i++)  {
1093 		struct fuse_arg *arg = &args[i];
1094 		if (i == numargs - 1 && argpages)
1095 			err = fuse_copy_pages(cs, arg->size, zeroing);
1096 		else
1097 			err = fuse_copy_one(cs, arg->value, arg->size);
1098 	}
1099 	return err;
1100 }
1101 
1102 static int forget_pending(struct fuse_iqueue *fiq)
1103 {
1104 	return fiq->forget_list_head.next != NULL;
1105 }
1106 
1107 static int request_pending(struct fuse_iqueue *fiq)
1108 {
1109 	return !list_empty(&fiq->pending) || !list_empty(&fiq->interrupts) ||
1110 		forget_pending(fiq);
1111 }
1112 
1113 /*
1114  * Transfer an interrupt request to userspace
1115  *
1116  * Unlike other requests this is assembled on demand, without a need
1117  * to allocate a separate fuse_req structure.
1118  *
1119  * Called with fiq->lock held, releases it
1120  */
1121 static int fuse_read_interrupt(struct fuse_iqueue *fiq,
1122 			       struct fuse_copy_state *cs,
1123 			       size_t nbytes, struct fuse_req *req)
1124 __releases(fiq->lock)
1125 {
1126 	struct fuse_in_header ih;
1127 	struct fuse_interrupt_in arg;
1128 	unsigned reqsize = sizeof(ih) + sizeof(arg);
1129 	int err;
1130 
1131 	list_del_init(&req->intr_entry);
1132 	memset(&ih, 0, sizeof(ih));
1133 	memset(&arg, 0, sizeof(arg));
1134 	ih.len = reqsize;
1135 	ih.opcode = FUSE_INTERRUPT;
1136 	ih.unique = (req->in.h.unique | FUSE_INT_REQ_BIT);
1137 	arg.unique = req->in.h.unique;
1138 
1139 	spin_unlock(&fiq->lock);
1140 	if (nbytes < reqsize)
1141 		return -EINVAL;
1142 
1143 	err = fuse_copy_one(cs, &ih, sizeof(ih));
1144 	if (!err)
1145 		err = fuse_copy_one(cs, &arg, sizeof(arg));
1146 	fuse_copy_finish(cs);
1147 
1148 	return err ? err : reqsize;
1149 }
1150 
1151 static struct fuse_forget_link *fuse_dequeue_forget(struct fuse_iqueue *fiq,
1152 						    unsigned int max,
1153 						    unsigned int *countp)
1154 {
1155 	struct fuse_forget_link *head = fiq->forget_list_head.next;
1156 	struct fuse_forget_link **newhead = &head;
1157 	unsigned count;
1158 
1159 	for (count = 0; *newhead != NULL && count < max; count++)
1160 		newhead = &(*newhead)->next;
1161 
1162 	fiq->forget_list_head.next = *newhead;
1163 	*newhead = NULL;
1164 	if (fiq->forget_list_head.next == NULL)
1165 		fiq->forget_list_tail = &fiq->forget_list_head;
1166 
1167 	if (countp != NULL)
1168 		*countp = count;
1169 
1170 	return head;
1171 }
1172 
1173 static int fuse_read_single_forget(struct fuse_iqueue *fiq,
1174 				   struct fuse_copy_state *cs,
1175 				   size_t nbytes)
1176 __releases(fiq->lock)
1177 {
1178 	int err;
1179 	struct fuse_forget_link *forget = fuse_dequeue_forget(fiq, 1, NULL);
1180 	struct fuse_forget_in arg = {
1181 		.nlookup = forget->forget_one.nlookup,
1182 	};
1183 	struct fuse_in_header ih = {
1184 		.opcode = FUSE_FORGET,
1185 		.nodeid = forget->forget_one.nodeid,
1186 		.unique = fuse_get_unique_locked(fiq),
1187 		.len = sizeof(ih) + sizeof(arg),
1188 	};
1189 
1190 	spin_unlock(&fiq->lock);
1191 	kfree(forget);
1192 	if (nbytes < ih.len)
1193 		return -EINVAL;
1194 
1195 	err = fuse_copy_one(cs, &ih, sizeof(ih));
1196 	if (!err)
1197 		err = fuse_copy_one(cs, &arg, sizeof(arg));
1198 	fuse_copy_finish(cs);
1199 
1200 	if (err)
1201 		return err;
1202 
1203 	return ih.len;
1204 }
1205 
1206 static int fuse_read_batch_forget(struct fuse_iqueue *fiq,
1207 				   struct fuse_copy_state *cs, size_t nbytes)
1208 __releases(fiq->lock)
1209 {
1210 	int err;
1211 	unsigned max_forgets;
1212 	unsigned count;
1213 	struct fuse_forget_link *head;
1214 	struct fuse_batch_forget_in arg = { .count = 0 };
1215 	struct fuse_in_header ih = {
1216 		.opcode = FUSE_BATCH_FORGET,
1217 		.unique = fuse_get_unique_locked(fiq),
1218 		.len = sizeof(ih) + sizeof(arg),
1219 	};
1220 
1221 	if (nbytes < ih.len) {
1222 		spin_unlock(&fiq->lock);
1223 		return -EINVAL;
1224 	}
1225 
1226 	max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1227 	head = fuse_dequeue_forget(fiq, max_forgets, &count);
1228 	spin_unlock(&fiq->lock);
1229 
1230 	arg.count = count;
1231 	ih.len += count * sizeof(struct fuse_forget_one);
1232 	err = fuse_copy_one(cs, &ih, sizeof(ih));
1233 	if (!err)
1234 		err = fuse_copy_one(cs, &arg, sizeof(arg));
1235 
1236 	while (head) {
1237 		struct fuse_forget_link *forget = head;
1238 
1239 		if (!err) {
1240 			err = fuse_copy_one(cs, &forget->forget_one,
1241 					    sizeof(forget->forget_one));
1242 		}
1243 		head = forget->next;
1244 		kfree(forget);
1245 	}
1246 
1247 	fuse_copy_finish(cs);
1248 
1249 	if (err)
1250 		return err;
1251 
1252 	return ih.len;
1253 }
1254 
1255 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_iqueue *fiq,
1256 			    struct fuse_copy_state *cs,
1257 			    size_t nbytes)
1258 __releases(fiq->lock)
1259 {
1260 	if (fc->minor < 16 || fiq->forget_list_head.next->next == NULL)
1261 		return fuse_read_single_forget(fiq, cs, nbytes);
1262 	else
1263 		return fuse_read_batch_forget(fiq, cs, nbytes);
1264 }
1265 
1266 /*
1267  * Read a single request into the userspace filesystem's buffer.  This
1268  * function waits until a request is available, then removes it from
1269  * the pending list and copies request data to userspace buffer.  If
1270  * no reply is needed (FORGET) or request has been aborted or there
1271  * was an error during the copying then it's finished by calling
1272  * fuse_request_end().  Otherwise add it to the processing list, and set
1273  * the 'sent' flag.
1274  */
1275 static ssize_t fuse_dev_do_read(struct fuse_dev *fud, struct file *file,
1276 				struct fuse_copy_state *cs, size_t nbytes)
1277 {
1278 	ssize_t err;
1279 	struct fuse_conn *fc = fud->fc;
1280 	struct fuse_iqueue *fiq = &fc->iq;
1281 	struct fuse_pqueue *fpq = &fud->pq;
1282 	struct fuse_req *req;
1283 	struct fuse_args *args;
1284 	unsigned reqsize;
1285 	unsigned int hash;
1286 
1287 	/*
1288 	 * Require sane minimum read buffer - that has capacity for fixed part
1289 	 * of any request header + negotiated max_write room for data.
1290 	 *
1291 	 * Historically libfuse reserves 4K for fixed header room, but e.g.
1292 	 * GlusterFS reserves only 80 bytes
1293 	 *
1294 	 *	= `sizeof(fuse_in_header) + sizeof(fuse_write_in)`
1295 	 *
1296 	 * which is the absolute minimum any sane filesystem should be using
1297 	 * for header room.
1298 	 */
1299 	if (nbytes < max_t(size_t, FUSE_MIN_READ_BUFFER,
1300 			   sizeof(struct fuse_in_header) +
1301 			   sizeof(struct fuse_write_in) +
1302 			   fc->max_write))
1303 		return -EINVAL;
1304 
1305  restart:
1306 	for (;;) {
1307 		spin_lock(&fiq->lock);
1308 		if (!fiq->connected || request_pending(fiq))
1309 			break;
1310 		spin_unlock(&fiq->lock);
1311 
1312 		if (file->f_flags & O_NONBLOCK)
1313 			return -EAGAIN;
1314 		err = wait_event_interruptible_exclusive(fiq->waitq,
1315 				!fiq->connected || request_pending(fiq));
1316 		if (err)
1317 			return err;
1318 	}
1319 
1320 	if (!fiq->connected) {
1321 		err = fc->aborted ? -ECONNABORTED : -ENODEV;
1322 		goto err_unlock;
1323 	}
1324 
1325 	if (!list_empty(&fiq->interrupts)) {
1326 		req = list_entry(fiq->interrupts.next, struct fuse_req,
1327 				 intr_entry);
1328 		return fuse_read_interrupt(fiq, cs, nbytes, req);
1329 	}
1330 
1331 	if (forget_pending(fiq)) {
1332 		if (list_empty(&fiq->pending) || fiq->forget_batch-- > 0)
1333 			return fuse_read_forget(fc, fiq, cs, nbytes);
1334 
1335 		if (fiq->forget_batch <= -8)
1336 			fiq->forget_batch = 16;
1337 	}
1338 
1339 	req = list_entry(fiq->pending.next, struct fuse_req, list);
1340 	clear_bit(FR_PENDING, &req->flags);
1341 	list_del_init(&req->list);
1342 	spin_unlock(&fiq->lock);
1343 
1344 	args = req->args;
1345 	reqsize = req->in.h.len;
1346 
1347 	/* If request is too large, reply with an error and restart the read */
1348 	if (nbytes < reqsize) {
1349 		req->out.h.error = -EIO;
1350 		/* SETXATTR is special, since it may contain too large data */
1351 		if (args->opcode == FUSE_SETXATTR)
1352 			req->out.h.error = -E2BIG;
1353 		fuse_request_end(req);
1354 		goto restart;
1355 	}
1356 	spin_lock(&fpq->lock);
1357 	/*
1358 	 *  Must not put request on fpq->io queue after having been shut down by
1359 	 *  fuse_abort_conn()
1360 	 */
1361 	if (!fpq->connected) {
1362 		req->out.h.error = err = -ECONNABORTED;
1363 		goto out_end;
1364 
1365 	}
1366 	list_add(&req->list, &fpq->io);
1367 	spin_unlock(&fpq->lock);
1368 	cs->req = req;
1369 	err = fuse_copy_one(cs, &req->in.h, sizeof(req->in.h));
1370 	if (!err)
1371 		err = fuse_copy_args(cs, args->in_numargs, args->in_pages,
1372 				     (struct fuse_arg *) args->in_args, 0);
1373 	fuse_copy_finish(cs);
1374 	spin_lock(&fpq->lock);
1375 	clear_bit(FR_LOCKED, &req->flags);
1376 	if (!fpq->connected) {
1377 		err = fc->aborted ? -ECONNABORTED : -ENODEV;
1378 		goto out_end;
1379 	}
1380 	if (err) {
1381 		req->out.h.error = -EIO;
1382 		goto out_end;
1383 	}
1384 	if (!test_bit(FR_ISREPLY, &req->flags)) {
1385 		err = reqsize;
1386 		goto out_end;
1387 	}
1388 	hash = fuse_req_hash(req->in.h.unique);
1389 	list_move_tail(&req->list, &fpq->processing[hash]);
1390 	__fuse_get_request(req);
1391 	set_bit(FR_SENT, &req->flags);
1392 	spin_unlock(&fpq->lock);
1393 	/* matches barrier in request_wait_answer() */
1394 	smp_mb__after_atomic();
1395 	if (test_bit(FR_INTERRUPTED, &req->flags))
1396 		queue_interrupt(req);
1397 	fuse_put_request(req);
1398 
1399 	return reqsize;
1400 
1401 out_end:
1402 	if (!test_bit(FR_PRIVATE, &req->flags))
1403 		list_del_init(&req->list);
1404 	spin_unlock(&fpq->lock);
1405 	fuse_request_end(req);
1406 	return err;
1407 
1408  err_unlock:
1409 	spin_unlock(&fiq->lock);
1410 	return err;
1411 }
1412 
1413 static int fuse_dev_open(struct inode *inode, struct file *file)
1414 {
1415 	/*
1416 	 * The fuse device's file's private_data is used to hold
1417 	 * the fuse_conn(ection) when it is mounted, and is used to
1418 	 * keep track of whether the file has been mounted already.
1419 	 */
1420 	file->private_data = NULL;
1421 	return 0;
1422 }
1423 
1424 static ssize_t fuse_dev_read(struct kiocb *iocb, struct iov_iter *to)
1425 {
1426 	struct fuse_copy_state cs;
1427 	struct file *file = iocb->ki_filp;
1428 	struct fuse_dev *fud = fuse_get_dev(file);
1429 
1430 	if (!fud)
1431 		return -EPERM;
1432 
1433 	if (!user_backed_iter(to))
1434 		return -EINVAL;
1435 
1436 	fuse_copy_init(&cs, 1, to);
1437 
1438 	return fuse_dev_do_read(fud, file, &cs, iov_iter_count(to));
1439 }
1440 
1441 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1442 				    struct pipe_inode_info *pipe,
1443 				    size_t len, unsigned int flags)
1444 {
1445 	int total, ret;
1446 	int page_nr = 0;
1447 	struct pipe_buffer *bufs;
1448 	struct fuse_copy_state cs;
1449 	struct fuse_dev *fud = fuse_get_dev(in);
1450 
1451 	if (!fud)
1452 		return -EPERM;
1453 
1454 	bufs = kvmalloc_array(pipe->max_usage, sizeof(struct pipe_buffer),
1455 			      GFP_KERNEL);
1456 	if (!bufs)
1457 		return -ENOMEM;
1458 
1459 	fuse_copy_init(&cs, 1, NULL);
1460 	cs.pipebufs = bufs;
1461 	cs.pipe = pipe;
1462 	ret = fuse_dev_do_read(fud, in, &cs, len);
1463 	if (ret < 0)
1464 		goto out;
1465 
1466 	if (pipe_occupancy(pipe->head, pipe->tail) + cs.nr_segs > pipe->max_usage) {
1467 		ret = -EIO;
1468 		goto out;
1469 	}
1470 
1471 	for (ret = total = 0; page_nr < cs.nr_segs; total += ret) {
1472 		/*
1473 		 * Need to be careful about this.  Having buf->ops in module
1474 		 * code can Oops if the buffer persists after module unload.
1475 		 */
1476 		bufs[page_nr].ops = &nosteal_pipe_buf_ops;
1477 		bufs[page_nr].flags = 0;
1478 		ret = add_to_pipe(pipe, &bufs[page_nr++]);
1479 		if (unlikely(ret < 0))
1480 			break;
1481 	}
1482 	if (total)
1483 		ret = total;
1484 out:
1485 	for (; page_nr < cs.nr_segs; page_nr++)
1486 		put_page(bufs[page_nr].page);
1487 
1488 	kvfree(bufs);
1489 	return ret;
1490 }
1491 
1492 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1493 			    struct fuse_copy_state *cs)
1494 {
1495 	struct fuse_notify_poll_wakeup_out outarg;
1496 	int err = -EINVAL;
1497 
1498 	if (size != sizeof(outarg))
1499 		goto err;
1500 
1501 	err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1502 	if (err)
1503 		goto err;
1504 
1505 	fuse_copy_finish(cs);
1506 	return fuse_notify_poll_wakeup(fc, &outarg);
1507 
1508 err:
1509 	fuse_copy_finish(cs);
1510 	return err;
1511 }
1512 
1513 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1514 				   struct fuse_copy_state *cs)
1515 {
1516 	struct fuse_notify_inval_inode_out outarg;
1517 	int err = -EINVAL;
1518 
1519 	if (size != sizeof(outarg))
1520 		goto err;
1521 
1522 	err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1523 	if (err)
1524 		goto err;
1525 	fuse_copy_finish(cs);
1526 
1527 	down_read(&fc->killsb);
1528 	err = fuse_reverse_inval_inode(fc, outarg.ino,
1529 				       outarg.off, outarg.len);
1530 	up_read(&fc->killsb);
1531 	return err;
1532 
1533 err:
1534 	fuse_copy_finish(cs);
1535 	return err;
1536 }
1537 
1538 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1539 				   struct fuse_copy_state *cs)
1540 {
1541 	struct fuse_notify_inval_entry_out outarg;
1542 	int err = -ENOMEM;
1543 	char *buf;
1544 	struct qstr name;
1545 
1546 	buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1547 	if (!buf)
1548 		goto err;
1549 
1550 	err = -EINVAL;
1551 	if (size < sizeof(outarg))
1552 		goto err;
1553 
1554 	err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1555 	if (err)
1556 		goto err;
1557 
1558 	err = -ENAMETOOLONG;
1559 	if (outarg.namelen > FUSE_NAME_MAX)
1560 		goto err;
1561 
1562 	err = -EINVAL;
1563 	if (size != sizeof(outarg) + outarg.namelen + 1)
1564 		goto err;
1565 
1566 	name.name = buf;
1567 	name.len = outarg.namelen;
1568 	err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1569 	if (err)
1570 		goto err;
1571 	fuse_copy_finish(cs);
1572 	buf[outarg.namelen] = 0;
1573 
1574 	down_read(&fc->killsb);
1575 	err = fuse_reverse_inval_entry(fc, outarg.parent, 0, &name, outarg.flags);
1576 	up_read(&fc->killsb);
1577 	kfree(buf);
1578 	return err;
1579 
1580 err:
1581 	kfree(buf);
1582 	fuse_copy_finish(cs);
1583 	return err;
1584 }
1585 
1586 static int fuse_notify_delete(struct fuse_conn *fc, unsigned int size,
1587 			      struct fuse_copy_state *cs)
1588 {
1589 	struct fuse_notify_delete_out outarg;
1590 	int err = -ENOMEM;
1591 	char *buf;
1592 	struct qstr name;
1593 
1594 	buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1595 	if (!buf)
1596 		goto err;
1597 
1598 	err = -EINVAL;
1599 	if (size < sizeof(outarg))
1600 		goto err;
1601 
1602 	err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1603 	if (err)
1604 		goto err;
1605 
1606 	err = -ENAMETOOLONG;
1607 	if (outarg.namelen > FUSE_NAME_MAX)
1608 		goto err;
1609 
1610 	err = -EINVAL;
1611 	if (size != sizeof(outarg) + outarg.namelen + 1)
1612 		goto err;
1613 
1614 	name.name = buf;
1615 	name.len = outarg.namelen;
1616 	err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1617 	if (err)
1618 		goto err;
1619 	fuse_copy_finish(cs);
1620 	buf[outarg.namelen] = 0;
1621 
1622 	down_read(&fc->killsb);
1623 	err = fuse_reverse_inval_entry(fc, outarg.parent, outarg.child, &name, 0);
1624 	up_read(&fc->killsb);
1625 	kfree(buf);
1626 	return err;
1627 
1628 err:
1629 	kfree(buf);
1630 	fuse_copy_finish(cs);
1631 	return err;
1632 }
1633 
1634 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1635 			     struct fuse_copy_state *cs)
1636 {
1637 	struct fuse_notify_store_out outarg;
1638 	struct inode *inode;
1639 	struct address_space *mapping;
1640 	u64 nodeid;
1641 	int err;
1642 	pgoff_t index;
1643 	unsigned int offset;
1644 	unsigned int num;
1645 	loff_t file_size;
1646 	loff_t end;
1647 
1648 	err = -EINVAL;
1649 	if (size < sizeof(outarg))
1650 		goto out_finish;
1651 
1652 	err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1653 	if (err)
1654 		goto out_finish;
1655 
1656 	err = -EINVAL;
1657 	if (size - sizeof(outarg) != outarg.size)
1658 		goto out_finish;
1659 
1660 	nodeid = outarg.nodeid;
1661 
1662 	down_read(&fc->killsb);
1663 
1664 	err = -ENOENT;
1665 	inode = fuse_ilookup(fc, nodeid,  NULL);
1666 	if (!inode)
1667 		goto out_up_killsb;
1668 
1669 	mapping = inode->i_mapping;
1670 	index = outarg.offset >> PAGE_SHIFT;
1671 	offset = outarg.offset & ~PAGE_MASK;
1672 	file_size = i_size_read(inode);
1673 	end = outarg.offset + outarg.size;
1674 	if (end > file_size) {
1675 		file_size = end;
1676 		fuse_write_update_attr(inode, file_size, outarg.size);
1677 	}
1678 
1679 	num = outarg.size;
1680 	while (num) {
1681 		struct folio *folio;
1682 		struct page *page;
1683 		unsigned int this_num;
1684 
1685 		folio = filemap_grab_folio(mapping, index);
1686 		err = PTR_ERR(folio);
1687 		if (IS_ERR(folio))
1688 			goto out_iput;
1689 
1690 		page = &folio->page;
1691 		this_num = min_t(unsigned, num, folio_size(folio) - offset);
1692 		err = fuse_copy_page(cs, &page, offset, this_num, 0);
1693 		if (!folio_test_uptodate(folio) && !err && offset == 0 &&
1694 		    (this_num == folio_size(folio) || file_size == end)) {
1695 			folio_zero_segment(folio, this_num, folio_size(folio));
1696 			folio_mark_uptodate(folio);
1697 		}
1698 		folio_unlock(folio);
1699 		folio_put(folio);
1700 
1701 		if (err)
1702 			goto out_iput;
1703 
1704 		num -= this_num;
1705 		offset = 0;
1706 		index++;
1707 	}
1708 
1709 	err = 0;
1710 
1711 out_iput:
1712 	iput(inode);
1713 out_up_killsb:
1714 	up_read(&fc->killsb);
1715 out_finish:
1716 	fuse_copy_finish(cs);
1717 	return err;
1718 }
1719 
1720 struct fuse_retrieve_args {
1721 	struct fuse_args_pages ap;
1722 	struct fuse_notify_retrieve_in inarg;
1723 };
1724 
1725 static void fuse_retrieve_end(struct fuse_mount *fm, struct fuse_args *args,
1726 			      int error)
1727 {
1728 	struct fuse_retrieve_args *ra =
1729 		container_of(args, typeof(*ra), ap.args);
1730 
1731 	release_pages(ra->ap.pages, ra->ap.num_pages);
1732 	kfree(ra);
1733 }
1734 
1735 static int fuse_retrieve(struct fuse_mount *fm, struct inode *inode,
1736 			 struct fuse_notify_retrieve_out *outarg)
1737 {
1738 	int err;
1739 	struct address_space *mapping = inode->i_mapping;
1740 	pgoff_t index;
1741 	loff_t file_size;
1742 	unsigned int num;
1743 	unsigned int offset;
1744 	size_t total_len = 0;
1745 	unsigned int num_pages;
1746 	struct fuse_conn *fc = fm->fc;
1747 	struct fuse_retrieve_args *ra;
1748 	size_t args_size = sizeof(*ra);
1749 	struct fuse_args_pages *ap;
1750 	struct fuse_args *args;
1751 
1752 	offset = outarg->offset & ~PAGE_MASK;
1753 	file_size = i_size_read(inode);
1754 
1755 	num = min(outarg->size, fc->max_write);
1756 	if (outarg->offset > file_size)
1757 		num = 0;
1758 	else if (outarg->offset + num > file_size)
1759 		num = file_size - outarg->offset;
1760 
1761 	num_pages = (num + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1762 	num_pages = min(num_pages, fc->max_pages);
1763 
1764 	args_size += num_pages * (sizeof(ap->pages[0]) + sizeof(ap->descs[0]));
1765 
1766 	ra = kzalloc(args_size, GFP_KERNEL);
1767 	if (!ra)
1768 		return -ENOMEM;
1769 
1770 	ap = &ra->ap;
1771 	ap->pages = (void *) (ra + 1);
1772 	ap->descs = (void *) (ap->pages + num_pages);
1773 
1774 	args = &ap->args;
1775 	args->nodeid = outarg->nodeid;
1776 	args->opcode = FUSE_NOTIFY_REPLY;
1777 	args->in_numargs = 2;
1778 	args->in_pages = true;
1779 	args->end = fuse_retrieve_end;
1780 
1781 	index = outarg->offset >> PAGE_SHIFT;
1782 
1783 	while (num && ap->num_pages < num_pages) {
1784 		struct folio *folio;
1785 		unsigned int this_num;
1786 
1787 		folio = filemap_get_folio(mapping, index);
1788 		if (IS_ERR(folio))
1789 			break;
1790 
1791 		this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1792 		ap->pages[ap->num_pages] = &folio->page;
1793 		ap->descs[ap->num_pages].offset = offset;
1794 		ap->descs[ap->num_pages].length = this_num;
1795 		ap->num_pages++;
1796 
1797 		offset = 0;
1798 		num -= this_num;
1799 		total_len += this_num;
1800 		index++;
1801 	}
1802 	ra->inarg.offset = outarg->offset;
1803 	ra->inarg.size = total_len;
1804 	args->in_args[0].size = sizeof(ra->inarg);
1805 	args->in_args[0].value = &ra->inarg;
1806 	args->in_args[1].size = total_len;
1807 
1808 	err = fuse_simple_notify_reply(fm, args, outarg->notify_unique);
1809 	if (err)
1810 		fuse_retrieve_end(fm, args, err);
1811 
1812 	return err;
1813 }
1814 
1815 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1816 				struct fuse_copy_state *cs)
1817 {
1818 	struct fuse_notify_retrieve_out outarg;
1819 	struct fuse_mount *fm;
1820 	struct inode *inode;
1821 	u64 nodeid;
1822 	int err;
1823 
1824 	err = -EINVAL;
1825 	if (size != sizeof(outarg))
1826 		goto copy_finish;
1827 
1828 	err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1829 	if (err)
1830 		goto copy_finish;
1831 
1832 	fuse_copy_finish(cs);
1833 
1834 	down_read(&fc->killsb);
1835 	err = -ENOENT;
1836 	nodeid = outarg.nodeid;
1837 
1838 	inode = fuse_ilookup(fc, nodeid, &fm);
1839 	if (inode) {
1840 		err = fuse_retrieve(fm, inode, &outarg);
1841 		iput(inode);
1842 	}
1843 	up_read(&fc->killsb);
1844 
1845 	return err;
1846 
1847 copy_finish:
1848 	fuse_copy_finish(cs);
1849 	return err;
1850 }
1851 
1852 /*
1853  * Resending all processing queue requests.
1854  *
1855  * During a FUSE daemon panics and failover, it is possible for some inflight
1856  * requests to be lost and never returned. As a result, applications awaiting
1857  * replies would become stuck forever. To address this, we can use notification
1858  * to trigger resending of these pending requests to the FUSE daemon, ensuring
1859  * they are properly processed again.
1860  *
1861  * Please note that this strategy is applicable only to idempotent requests or
1862  * if the FUSE daemon takes careful measures to avoid processing duplicated
1863  * non-idempotent requests.
1864  */
1865 static void fuse_resend(struct fuse_conn *fc)
1866 {
1867 	struct fuse_dev *fud;
1868 	struct fuse_req *req, *next;
1869 	struct fuse_iqueue *fiq = &fc->iq;
1870 	LIST_HEAD(to_queue);
1871 	unsigned int i;
1872 
1873 	spin_lock(&fc->lock);
1874 	if (!fc->connected) {
1875 		spin_unlock(&fc->lock);
1876 		return;
1877 	}
1878 
1879 	list_for_each_entry(fud, &fc->devices, entry) {
1880 		struct fuse_pqueue *fpq = &fud->pq;
1881 
1882 		spin_lock(&fpq->lock);
1883 		for (i = 0; i < FUSE_PQ_HASH_SIZE; i++)
1884 			list_splice_tail_init(&fpq->processing[i], &to_queue);
1885 		spin_unlock(&fpq->lock);
1886 	}
1887 	spin_unlock(&fc->lock);
1888 
1889 	list_for_each_entry_safe(req, next, &to_queue, list) {
1890 		set_bit(FR_PENDING, &req->flags);
1891 		clear_bit(FR_SENT, &req->flags);
1892 		/* mark the request as resend request */
1893 		req->in.h.unique |= FUSE_UNIQUE_RESEND;
1894 	}
1895 
1896 	spin_lock(&fiq->lock);
1897 	if (!fiq->connected) {
1898 		spin_unlock(&fiq->lock);
1899 		list_for_each_entry(req, &to_queue, list)
1900 			clear_bit(FR_PENDING, &req->flags);
1901 		end_requests(&to_queue);
1902 		return;
1903 	}
1904 	/* iq and pq requests are both oldest to newest */
1905 	list_splice(&to_queue, &fiq->pending);
1906 	fuse_dev_wake_and_unlock(fiq);
1907 }
1908 
1909 static int fuse_notify_resend(struct fuse_conn *fc)
1910 {
1911 	fuse_resend(fc);
1912 	return 0;
1913 }
1914 
1915 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1916 		       unsigned int size, struct fuse_copy_state *cs)
1917 {
1918 	/* Don't try to move pages (yet) */
1919 	cs->move_pages = 0;
1920 
1921 	switch (code) {
1922 	case FUSE_NOTIFY_POLL:
1923 		return fuse_notify_poll(fc, size, cs);
1924 
1925 	case FUSE_NOTIFY_INVAL_INODE:
1926 		return fuse_notify_inval_inode(fc, size, cs);
1927 
1928 	case FUSE_NOTIFY_INVAL_ENTRY:
1929 		return fuse_notify_inval_entry(fc, size, cs);
1930 
1931 	case FUSE_NOTIFY_STORE:
1932 		return fuse_notify_store(fc, size, cs);
1933 
1934 	case FUSE_NOTIFY_RETRIEVE:
1935 		return fuse_notify_retrieve(fc, size, cs);
1936 
1937 	case FUSE_NOTIFY_DELETE:
1938 		return fuse_notify_delete(fc, size, cs);
1939 
1940 	case FUSE_NOTIFY_RESEND:
1941 		return fuse_notify_resend(fc);
1942 
1943 	default:
1944 		fuse_copy_finish(cs);
1945 		return -EINVAL;
1946 	}
1947 }
1948 
1949 /* Look up request on processing list by unique ID */
1950 static struct fuse_req *request_find(struct fuse_pqueue *fpq, u64 unique)
1951 {
1952 	unsigned int hash = fuse_req_hash(unique);
1953 	struct fuse_req *req;
1954 
1955 	list_for_each_entry(req, &fpq->processing[hash], list) {
1956 		if (req->in.h.unique == unique)
1957 			return req;
1958 	}
1959 	return NULL;
1960 }
1961 
1962 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_args *args,
1963 			 unsigned nbytes)
1964 {
1965 	unsigned reqsize = sizeof(struct fuse_out_header);
1966 
1967 	reqsize += fuse_len_args(args->out_numargs, args->out_args);
1968 
1969 	if (reqsize < nbytes || (reqsize > nbytes && !args->out_argvar))
1970 		return -EINVAL;
1971 	else if (reqsize > nbytes) {
1972 		struct fuse_arg *lastarg = &args->out_args[args->out_numargs-1];
1973 		unsigned diffsize = reqsize - nbytes;
1974 
1975 		if (diffsize > lastarg->size)
1976 			return -EINVAL;
1977 		lastarg->size -= diffsize;
1978 	}
1979 	return fuse_copy_args(cs, args->out_numargs, args->out_pages,
1980 			      args->out_args, args->page_zeroing);
1981 }
1982 
1983 /*
1984  * Write a single reply to a request.  First the header is copied from
1985  * the write buffer.  The request is then searched on the processing
1986  * list by the unique ID found in the header.  If found, then remove
1987  * it from the list and copy the rest of the buffer to the request.
1988  * The request is finished by calling fuse_request_end().
1989  */
1990 static ssize_t fuse_dev_do_write(struct fuse_dev *fud,
1991 				 struct fuse_copy_state *cs, size_t nbytes)
1992 {
1993 	int err;
1994 	struct fuse_conn *fc = fud->fc;
1995 	struct fuse_pqueue *fpq = &fud->pq;
1996 	struct fuse_req *req;
1997 	struct fuse_out_header oh;
1998 
1999 	err = -EINVAL;
2000 	if (nbytes < sizeof(struct fuse_out_header))
2001 		goto out;
2002 
2003 	err = fuse_copy_one(cs, &oh, sizeof(oh));
2004 	if (err)
2005 		goto copy_finish;
2006 
2007 	err = -EINVAL;
2008 	if (oh.len != nbytes)
2009 		goto copy_finish;
2010 
2011 	/*
2012 	 * Zero oh.unique indicates unsolicited notification message
2013 	 * and error contains notification code.
2014 	 */
2015 	if (!oh.unique) {
2016 		err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
2017 		goto out;
2018 	}
2019 
2020 	err = -EINVAL;
2021 	if (oh.error <= -512 || oh.error > 0)
2022 		goto copy_finish;
2023 
2024 	spin_lock(&fpq->lock);
2025 	req = NULL;
2026 	if (fpq->connected)
2027 		req = request_find(fpq, oh.unique & ~FUSE_INT_REQ_BIT);
2028 
2029 	err = -ENOENT;
2030 	if (!req) {
2031 		spin_unlock(&fpq->lock);
2032 		goto copy_finish;
2033 	}
2034 
2035 	/* Is it an interrupt reply ID? */
2036 	if (oh.unique & FUSE_INT_REQ_BIT) {
2037 		__fuse_get_request(req);
2038 		spin_unlock(&fpq->lock);
2039 
2040 		err = 0;
2041 		if (nbytes != sizeof(struct fuse_out_header))
2042 			err = -EINVAL;
2043 		else if (oh.error == -ENOSYS)
2044 			fc->no_interrupt = 1;
2045 		else if (oh.error == -EAGAIN)
2046 			err = queue_interrupt(req);
2047 
2048 		fuse_put_request(req);
2049 
2050 		goto copy_finish;
2051 	}
2052 
2053 	clear_bit(FR_SENT, &req->flags);
2054 	list_move(&req->list, &fpq->io);
2055 	req->out.h = oh;
2056 	set_bit(FR_LOCKED, &req->flags);
2057 	spin_unlock(&fpq->lock);
2058 	cs->req = req;
2059 	if (!req->args->page_replace)
2060 		cs->move_pages = 0;
2061 
2062 	if (oh.error)
2063 		err = nbytes != sizeof(oh) ? -EINVAL : 0;
2064 	else
2065 		err = copy_out_args(cs, req->args, nbytes);
2066 	fuse_copy_finish(cs);
2067 
2068 	spin_lock(&fpq->lock);
2069 	clear_bit(FR_LOCKED, &req->flags);
2070 	if (!fpq->connected)
2071 		err = -ENOENT;
2072 	else if (err)
2073 		req->out.h.error = -EIO;
2074 	if (!test_bit(FR_PRIVATE, &req->flags))
2075 		list_del_init(&req->list);
2076 	spin_unlock(&fpq->lock);
2077 
2078 	fuse_request_end(req);
2079 out:
2080 	return err ? err : nbytes;
2081 
2082 copy_finish:
2083 	fuse_copy_finish(cs);
2084 	goto out;
2085 }
2086 
2087 static ssize_t fuse_dev_write(struct kiocb *iocb, struct iov_iter *from)
2088 {
2089 	struct fuse_copy_state cs;
2090 	struct fuse_dev *fud = fuse_get_dev(iocb->ki_filp);
2091 
2092 	if (!fud)
2093 		return -EPERM;
2094 
2095 	if (!user_backed_iter(from))
2096 		return -EINVAL;
2097 
2098 	fuse_copy_init(&cs, 0, from);
2099 
2100 	return fuse_dev_do_write(fud, &cs, iov_iter_count(from));
2101 }
2102 
2103 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
2104 				     struct file *out, loff_t *ppos,
2105 				     size_t len, unsigned int flags)
2106 {
2107 	unsigned int head, tail, mask, count;
2108 	unsigned nbuf;
2109 	unsigned idx;
2110 	struct pipe_buffer *bufs;
2111 	struct fuse_copy_state cs;
2112 	struct fuse_dev *fud;
2113 	size_t rem;
2114 	ssize_t ret;
2115 
2116 	fud = fuse_get_dev(out);
2117 	if (!fud)
2118 		return -EPERM;
2119 
2120 	pipe_lock(pipe);
2121 
2122 	head = pipe->head;
2123 	tail = pipe->tail;
2124 	mask = pipe->ring_size - 1;
2125 	count = head - tail;
2126 
2127 	bufs = kvmalloc_array(count, sizeof(struct pipe_buffer), GFP_KERNEL);
2128 	if (!bufs) {
2129 		pipe_unlock(pipe);
2130 		return -ENOMEM;
2131 	}
2132 
2133 	nbuf = 0;
2134 	rem = 0;
2135 	for (idx = tail; idx != head && rem < len; idx++)
2136 		rem += pipe->bufs[idx & mask].len;
2137 
2138 	ret = -EINVAL;
2139 	if (rem < len)
2140 		goto out_free;
2141 
2142 	rem = len;
2143 	while (rem) {
2144 		struct pipe_buffer *ibuf;
2145 		struct pipe_buffer *obuf;
2146 
2147 		if (WARN_ON(nbuf >= count || tail == head))
2148 			goto out_free;
2149 
2150 		ibuf = &pipe->bufs[tail & mask];
2151 		obuf = &bufs[nbuf];
2152 
2153 		if (rem >= ibuf->len) {
2154 			*obuf = *ibuf;
2155 			ibuf->ops = NULL;
2156 			tail++;
2157 			pipe->tail = tail;
2158 		} else {
2159 			if (!pipe_buf_get(pipe, ibuf))
2160 				goto out_free;
2161 
2162 			*obuf = *ibuf;
2163 			obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2164 			obuf->len = rem;
2165 			ibuf->offset += obuf->len;
2166 			ibuf->len -= obuf->len;
2167 		}
2168 		nbuf++;
2169 		rem -= obuf->len;
2170 	}
2171 	pipe_unlock(pipe);
2172 
2173 	fuse_copy_init(&cs, 0, NULL);
2174 	cs.pipebufs = bufs;
2175 	cs.nr_segs = nbuf;
2176 	cs.pipe = pipe;
2177 
2178 	if (flags & SPLICE_F_MOVE)
2179 		cs.move_pages = 1;
2180 
2181 	ret = fuse_dev_do_write(fud, &cs, len);
2182 
2183 	pipe_lock(pipe);
2184 out_free:
2185 	for (idx = 0; idx < nbuf; idx++) {
2186 		struct pipe_buffer *buf = &bufs[idx];
2187 
2188 		if (buf->ops)
2189 			pipe_buf_release(pipe, buf);
2190 	}
2191 	pipe_unlock(pipe);
2192 
2193 	kvfree(bufs);
2194 	return ret;
2195 }
2196 
2197 static __poll_t fuse_dev_poll(struct file *file, poll_table *wait)
2198 {
2199 	__poll_t mask = EPOLLOUT | EPOLLWRNORM;
2200 	struct fuse_iqueue *fiq;
2201 	struct fuse_dev *fud = fuse_get_dev(file);
2202 
2203 	if (!fud)
2204 		return EPOLLERR;
2205 
2206 	fiq = &fud->fc->iq;
2207 	poll_wait(file, &fiq->waitq, wait);
2208 
2209 	spin_lock(&fiq->lock);
2210 	if (!fiq->connected)
2211 		mask = EPOLLERR;
2212 	else if (request_pending(fiq))
2213 		mask |= EPOLLIN | EPOLLRDNORM;
2214 	spin_unlock(&fiq->lock);
2215 
2216 	return mask;
2217 }
2218 
2219 /* Abort all requests on the given list (pending or processing) */
2220 static void end_requests(struct list_head *head)
2221 {
2222 	while (!list_empty(head)) {
2223 		struct fuse_req *req;
2224 		req = list_entry(head->next, struct fuse_req, list);
2225 		req->out.h.error = -ECONNABORTED;
2226 		clear_bit(FR_SENT, &req->flags);
2227 		list_del_init(&req->list);
2228 		fuse_request_end(req);
2229 	}
2230 }
2231 
2232 static void end_polls(struct fuse_conn *fc)
2233 {
2234 	struct rb_node *p;
2235 
2236 	p = rb_first(&fc->polled_files);
2237 
2238 	while (p) {
2239 		struct fuse_file *ff;
2240 		ff = rb_entry(p, struct fuse_file, polled_node);
2241 		wake_up_interruptible_all(&ff->poll_wait);
2242 
2243 		p = rb_next(p);
2244 	}
2245 }
2246 
2247 /*
2248  * Abort all requests.
2249  *
2250  * Emergency exit in case of a malicious or accidental deadlock, or just a hung
2251  * filesystem.
2252  *
2253  * The same effect is usually achievable through killing the filesystem daemon
2254  * and all users of the filesystem.  The exception is the combination of an
2255  * asynchronous request and the tricky deadlock (see
2256  * Documentation/filesystems/fuse.rst).
2257  *
2258  * Aborting requests under I/O goes as follows: 1: Separate out unlocked
2259  * requests, they should be finished off immediately.  Locked requests will be
2260  * finished after unlock; see unlock_request(). 2: Finish off the unlocked
2261  * requests.  It is possible that some request will finish before we can.  This
2262  * is OK, the request will in that case be removed from the list before we touch
2263  * it.
2264  */
2265 void fuse_abort_conn(struct fuse_conn *fc)
2266 {
2267 	struct fuse_iqueue *fiq = &fc->iq;
2268 
2269 	spin_lock(&fc->lock);
2270 	if (fc->connected) {
2271 		struct fuse_dev *fud;
2272 		struct fuse_req *req, *next;
2273 		LIST_HEAD(to_end);
2274 		unsigned int i;
2275 
2276 		/* Background queuing checks fc->connected under bg_lock */
2277 		spin_lock(&fc->bg_lock);
2278 		fc->connected = 0;
2279 		spin_unlock(&fc->bg_lock);
2280 
2281 		fuse_set_initialized(fc);
2282 		list_for_each_entry(fud, &fc->devices, entry) {
2283 			struct fuse_pqueue *fpq = &fud->pq;
2284 
2285 			spin_lock(&fpq->lock);
2286 			fpq->connected = 0;
2287 			list_for_each_entry_safe(req, next, &fpq->io, list) {
2288 				req->out.h.error = -ECONNABORTED;
2289 				spin_lock(&req->waitq.lock);
2290 				set_bit(FR_ABORTED, &req->flags);
2291 				if (!test_bit(FR_LOCKED, &req->flags)) {
2292 					set_bit(FR_PRIVATE, &req->flags);
2293 					__fuse_get_request(req);
2294 					list_move(&req->list, &to_end);
2295 				}
2296 				spin_unlock(&req->waitq.lock);
2297 			}
2298 			for (i = 0; i < FUSE_PQ_HASH_SIZE; i++)
2299 				list_splice_tail_init(&fpq->processing[i],
2300 						      &to_end);
2301 			spin_unlock(&fpq->lock);
2302 		}
2303 		spin_lock(&fc->bg_lock);
2304 		fc->blocked = 0;
2305 		fc->max_background = UINT_MAX;
2306 		flush_bg_queue(fc);
2307 		spin_unlock(&fc->bg_lock);
2308 
2309 		spin_lock(&fiq->lock);
2310 		fiq->connected = 0;
2311 		list_for_each_entry(req, &fiq->pending, list)
2312 			clear_bit(FR_PENDING, &req->flags);
2313 		list_splice_tail_init(&fiq->pending, &to_end);
2314 		while (forget_pending(fiq))
2315 			kfree(fuse_dequeue_forget(fiq, 1, NULL));
2316 		wake_up_all(&fiq->waitq);
2317 		spin_unlock(&fiq->lock);
2318 		kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
2319 		end_polls(fc);
2320 		wake_up_all(&fc->blocked_waitq);
2321 		spin_unlock(&fc->lock);
2322 
2323 		end_requests(&to_end);
2324 	} else {
2325 		spin_unlock(&fc->lock);
2326 	}
2327 }
2328 EXPORT_SYMBOL_GPL(fuse_abort_conn);
2329 
2330 void fuse_wait_aborted(struct fuse_conn *fc)
2331 {
2332 	/* matches implicit memory barrier in fuse_drop_waiting() */
2333 	smp_mb();
2334 	wait_event(fc->blocked_waitq, atomic_read(&fc->num_waiting) == 0);
2335 }
2336 
2337 int fuse_dev_release(struct inode *inode, struct file *file)
2338 {
2339 	struct fuse_dev *fud = fuse_get_dev(file);
2340 
2341 	if (fud) {
2342 		struct fuse_conn *fc = fud->fc;
2343 		struct fuse_pqueue *fpq = &fud->pq;
2344 		LIST_HEAD(to_end);
2345 		unsigned int i;
2346 
2347 		spin_lock(&fpq->lock);
2348 		WARN_ON(!list_empty(&fpq->io));
2349 		for (i = 0; i < FUSE_PQ_HASH_SIZE; i++)
2350 			list_splice_init(&fpq->processing[i], &to_end);
2351 		spin_unlock(&fpq->lock);
2352 
2353 		end_requests(&to_end);
2354 
2355 		/* Are we the last open device? */
2356 		if (atomic_dec_and_test(&fc->dev_count)) {
2357 			WARN_ON(fc->iq.fasync != NULL);
2358 			fuse_abort_conn(fc);
2359 		}
2360 		fuse_dev_free(fud);
2361 	}
2362 	return 0;
2363 }
2364 EXPORT_SYMBOL_GPL(fuse_dev_release);
2365 
2366 static int fuse_dev_fasync(int fd, struct file *file, int on)
2367 {
2368 	struct fuse_dev *fud = fuse_get_dev(file);
2369 
2370 	if (!fud)
2371 		return -EPERM;
2372 
2373 	/* No locking - fasync_helper does its own locking */
2374 	return fasync_helper(fd, file, on, &fud->fc->iq.fasync);
2375 }
2376 
2377 static int fuse_device_clone(struct fuse_conn *fc, struct file *new)
2378 {
2379 	struct fuse_dev *fud;
2380 
2381 	if (new->private_data)
2382 		return -EINVAL;
2383 
2384 	fud = fuse_dev_alloc_install(fc);
2385 	if (!fud)
2386 		return -ENOMEM;
2387 
2388 	new->private_data = fud;
2389 	atomic_inc(&fc->dev_count);
2390 
2391 	return 0;
2392 }
2393 
2394 static long fuse_dev_ioctl_clone(struct file *file, __u32 __user *argp)
2395 {
2396 	int res;
2397 	int oldfd;
2398 	struct fuse_dev *fud = NULL;
2399 	struct fd f;
2400 
2401 	if (get_user(oldfd, argp))
2402 		return -EFAULT;
2403 
2404 	f = fdget(oldfd);
2405 	if (!fd_file(f))
2406 		return -EINVAL;
2407 
2408 	/*
2409 	 * Check against file->f_op because CUSE
2410 	 * uses the same ioctl handler.
2411 	 */
2412 	if (fd_file(f)->f_op == file->f_op)
2413 		fud = fuse_get_dev(fd_file(f));
2414 
2415 	res = -EINVAL;
2416 	if (fud) {
2417 		mutex_lock(&fuse_mutex);
2418 		res = fuse_device_clone(fud->fc, file);
2419 		mutex_unlock(&fuse_mutex);
2420 	}
2421 
2422 	fdput(f);
2423 	return res;
2424 }
2425 
2426 static long fuse_dev_ioctl_backing_open(struct file *file,
2427 					struct fuse_backing_map __user *argp)
2428 {
2429 	struct fuse_dev *fud = fuse_get_dev(file);
2430 	struct fuse_backing_map map;
2431 
2432 	if (!fud)
2433 		return -EPERM;
2434 
2435 	if (!IS_ENABLED(CONFIG_FUSE_PASSTHROUGH))
2436 		return -EOPNOTSUPP;
2437 
2438 	if (copy_from_user(&map, argp, sizeof(map)))
2439 		return -EFAULT;
2440 
2441 	return fuse_backing_open(fud->fc, &map);
2442 }
2443 
2444 static long fuse_dev_ioctl_backing_close(struct file *file, __u32 __user *argp)
2445 {
2446 	struct fuse_dev *fud = fuse_get_dev(file);
2447 	int backing_id;
2448 
2449 	if (!fud)
2450 		return -EPERM;
2451 
2452 	if (!IS_ENABLED(CONFIG_FUSE_PASSTHROUGH))
2453 		return -EOPNOTSUPP;
2454 
2455 	if (get_user(backing_id, argp))
2456 		return -EFAULT;
2457 
2458 	return fuse_backing_close(fud->fc, backing_id);
2459 }
2460 
2461 static long fuse_dev_ioctl(struct file *file, unsigned int cmd,
2462 			   unsigned long arg)
2463 {
2464 	void __user *argp = (void __user *)arg;
2465 
2466 	switch (cmd) {
2467 	case FUSE_DEV_IOC_CLONE:
2468 		return fuse_dev_ioctl_clone(file, argp);
2469 
2470 	case FUSE_DEV_IOC_BACKING_OPEN:
2471 		return fuse_dev_ioctl_backing_open(file, argp);
2472 
2473 	case FUSE_DEV_IOC_BACKING_CLOSE:
2474 		return fuse_dev_ioctl_backing_close(file, argp);
2475 
2476 	default:
2477 		return -ENOTTY;
2478 	}
2479 }
2480 
2481 const struct file_operations fuse_dev_operations = {
2482 	.owner		= THIS_MODULE,
2483 	.open		= fuse_dev_open,
2484 	.read_iter	= fuse_dev_read,
2485 	.splice_read	= fuse_dev_splice_read,
2486 	.write_iter	= fuse_dev_write,
2487 	.splice_write	= fuse_dev_splice_write,
2488 	.poll		= fuse_dev_poll,
2489 	.release	= fuse_dev_release,
2490 	.fasync		= fuse_dev_fasync,
2491 	.unlocked_ioctl = fuse_dev_ioctl,
2492 	.compat_ioctl   = compat_ptr_ioctl,
2493 };
2494 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2495 
2496 static struct miscdevice fuse_miscdevice = {
2497 	.minor = FUSE_MINOR,
2498 	.name  = "fuse",
2499 	.fops = &fuse_dev_operations,
2500 };
2501 
2502 int __init fuse_dev_init(void)
2503 {
2504 	int err = -ENOMEM;
2505 	fuse_req_cachep = kmem_cache_create("fuse_request",
2506 					    sizeof(struct fuse_req),
2507 					    0, 0, NULL);
2508 	if (!fuse_req_cachep)
2509 		goto out;
2510 
2511 	err = misc_register(&fuse_miscdevice);
2512 	if (err)
2513 		goto out_cache_clean;
2514 
2515 	return 0;
2516 
2517  out_cache_clean:
2518 	kmem_cache_destroy(fuse_req_cachep);
2519  out:
2520 	return err;
2521 }
2522 
2523 void fuse_dev_cleanup(void)
2524 {
2525 	misc_deregister(&fuse_miscdevice);
2526 	kmem_cache_destroy(fuse_req_cachep);
2527 }
2528