xref: /openbmc/linux/fs/fuse/virtio_fs.c (revision c4c3c32d)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * virtio-fs: Virtio Filesystem
4  * Copyright (C) 2018 Red Hat, Inc.
5  */
6 
7 #include <linux/fs.h>
8 #include <linux/dax.h>
9 #include <linux/pci.h>
10 #include <linux/pfn_t.h>
11 #include <linux/memremap.h>
12 #include <linux/module.h>
13 #include <linux/virtio.h>
14 #include <linux/virtio_fs.h>
15 #include <linux/delay.h>
16 #include <linux/fs_context.h>
17 #include <linux/fs_parser.h>
18 #include <linux/highmem.h>
19 #include <linux/uio.h>
20 #include "fuse_i.h"
21 
22 /* Used to help calculate the FUSE connection's max_pages limit for a request's
23  * size. Parts of the struct fuse_req are sliced into scattergather lists in
24  * addition to the pages used, so this can help account for that overhead.
25  */
26 #define FUSE_HEADER_OVERHEAD    4
27 
28 /* List of virtio-fs device instances and a lock for the list. Also provides
29  * mutual exclusion in device removal and mounting path
30  */
31 static DEFINE_MUTEX(virtio_fs_mutex);
32 static LIST_HEAD(virtio_fs_instances);
33 
34 enum {
35 	VQ_HIPRIO,
36 	VQ_REQUEST
37 };
38 
39 #define VQ_NAME_LEN	24
40 
41 /* Per-virtqueue state */
42 struct virtio_fs_vq {
43 	spinlock_t lock;
44 	struct virtqueue *vq;     /* protected by ->lock */
45 	struct work_struct done_work;
46 	struct list_head queued_reqs;
47 	struct list_head end_reqs;	/* End these requests */
48 	struct delayed_work dispatch_work;
49 	struct fuse_dev *fud;
50 	bool connected;
51 	long in_flight;
52 	struct completion in_flight_zero; /* No inflight requests */
53 	char name[VQ_NAME_LEN];
54 } ____cacheline_aligned_in_smp;
55 
56 /* A virtio-fs device instance */
57 struct virtio_fs {
58 	struct kref refcount;
59 	struct list_head list;    /* on virtio_fs_instances */
60 	char *tag;
61 	struct virtio_fs_vq *vqs;
62 	unsigned int nvqs;               /* number of virtqueues */
63 	unsigned int num_request_queues; /* number of request queues */
64 	struct dax_device *dax_dev;
65 
66 	/* DAX memory window where file contents are mapped */
67 	void *window_kaddr;
68 	phys_addr_t window_phys_addr;
69 	size_t window_len;
70 };
71 
72 struct virtio_fs_forget_req {
73 	struct fuse_in_header ih;
74 	struct fuse_forget_in arg;
75 };
76 
77 struct virtio_fs_forget {
78 	/* This request can be temporarily queued on virt queue */
79 	struct list_head list;
80 	struct virtio_fs_forget_req req;
81 };
82 
83 struct virtio_fs_req_work {
84 	struct fuse_req *req;
85 	struct virtio_fs_vq *fsvq;
86 	struct work_struct done_work;
87 };
88 
89 static int virtio_fs_enqueue_req(struct virtio_fs_vq *fsvq,
90 				 struct fuse_req *req, bool in_flight);
91 
92 static const struct constant_table dax_param_enums[] = {
93 	{"always",	FUSE_DAX_ALWAYS },
94 	{"never",	FUSE_DAX_NEVER },
95 	{"inode",	FUSE_DAX_INODE_USER },
96 	{}
97 };
98 
99 enum {
100 	OPT_DAX,
101 	OPT_DAX_ENUM,
102 };
103 
104 static const struct fs_parameter_spec virtio_fs_parameters[] = {
105 	fsparam_flag("dax", OPT_DAX),
106 	fsparam_enum("dax", OPT_DAX_ENUM, dax_param_enums),
107 	{}
108 };
109 
110 static int virtio_fs_parse_param(struct fs_context *fsc,
111 				 struct fs_parameter *param)
112 {
113 	struct fs_parse_result result;
114 	struct fuse_fs_context *ctx = fsc->fs_private;
115 	int opt;
116 
117 	opt = fs_parse(fsc, virtio_fs_parameters, param, &result);
118 	if (opt < 0)
119 		return opt;
120 
121 	switch (opt) {
122 	case OPT_DAX:
123 		ctx->dax_mode = FUSE_DAX_ALWAYS;
124 		break;
125 	case OPT_DAX_ENUM:
126 		ctx->dax_mode = result.uint_32;
127 		break;
128 	default:
129 		return -EINVAL;
130 	}
131 
132 	return 0;
133 }
134 
135 static void virtio_fs_free_fsc(struct fs_context *fsc)
136 {
137 	struct fuse_fs_context *ctx = fsc->fs_private;
138 
139 	kfree(ctx);
140 }
141 
142 static inline struct virtio_fs_vq *vq_to_fsvq(struct virtqueue *vq)
143 {
144 	struct virtio_fs *fs = vq->vdev->priv;
145 
146 	return &fs->vqs[vq->index];
147 }
148 
149 /* Should be called with fsvq->lock held. */
150 static inline void inc_in_flight_req(struct virtio_fs_vq *fsvq)
151 {
152 	fsvq->in_flight++;
153 }
154 
155 /* Should be called with fsvq->lock held. */
156 static inline void dec_in_flight_req(struct virtio_fs_vq *fsvq)
157 {
158 	WARN_ON(fsvq->in_flight <= 0);
159 	fsvq->in_flight--;
160 	if (!fsvq->in_flight)
161 		complete(&fsvq->in_flight_zero);
162 }
163 
164 static void release_virtio_fs_obj(struct kref *ref)
165 {
166 	struct virtio_fs *vfs = container_of(ref, struct virtio_fs, refcount);
167 
168 	kfree(vfs->vqs);
169 	kfree(vfs);
170 }
171 
172 /* Make sure virtiofs_mutex is held */
173 static void virtio_fs_put(struct virtio_fs *fs)
174 {
175 	kref_put(&fs->refcount, release_virtio_fs_obj);
176 }
177 
178 static void virtio_fs_fiq_release(struct fuse_iqueue *fiq)
179 {
180 	struct virtio_fs *vfs = fiq->priv;
181 
182 	mutex_lock(&virtio_fs_mutex);
183 	virtio_fs_put(vfs);
184 	mutex_unlock(&virtio_fs_mutex);
185 }
186 
187 static void virtio_fs_drain_queue(struct virtio_fs_vq *fsvq)
188 {
189 	WARN_ON(fsvq->in_flight < 0);
190 
191 	/* Wait for in flight requests to finish.*/
192 	spin_lock(&fsvq->lock);
193 	if (fsvq->in_flight) {
194 		/* We are holding virtio_fs_mutex. There should not be any
195 		 * waiters waiting for completion.
196 		 */
197 		reinit_completion(&fsvq->in_flight_zero);
198 		spin_unlock(&fsvq->lock);
199 		wait_for_completion(&fsvq->in_flight_zero);
200 	} else {
201 		spin_unlock(&fsvq->lock);
202 	}
203 
204 	flush_work(&fsvq->done_work);
205 	flush_delayed_work(&fsvq->dispatch_work);
206 }
207 
208 static void virtio_fs_drain_all_queues_locked(struct virtio_fs *fs)
209 {
210 	struct virtio_fs_vq *fsvq;
211 	int i;
212 
213 	for (i = 0; i < fs->nvqs; i++) {
214 		fsvq = &fs->vqs[i];
215 		virtio_fs_drain_queue(fsvq);
216 	}
217 }
218 
219 static void virtio_fs_drain_all_queues(struct virtio_fs *fs)
220 {
221 	/* Provides mutual exclusion between ->remove and ->kill_sb
222 	 * paths. We don't want both of these draining queue at the
223 	 * same time. Current completion logic reinits completion
224 	 * and that means there should not be any other thread
225 	 * doing reinit or waiting for completion already.
226 	 */
227 	mutex_lock(&virtio_fs_mutex);
228 	virtio_fs_drain_all_queues_locked(fs);
229 	mutex_unlock(&virtio_fs_mutex);
230 }
231 
232 static void virtio_fs_start_all_queues(struct virtio_fs *fs)
233 {
234 	struct virtio_fs_vq *fsvq;
235 	int i;
236 
237 	for (i = 0; i < fs->nvqs; i++) {
238 		fsvq = &fs->vqs[i];
239 		spin_lock(&fsvq->lock);
240 		fsvq->connected = true;
241 		spin_unlock(&fsvq->lock);
242 	}
243 }
244 
245 /* Add a new instance to the list or return -EEXIST if tag name exists*/
246 static int virtio_fs_add_instance(struct virtio_fs *fs)
247 {
248 	struct virtio_fs *fs2;
249 	bool duplicate = false;
250 
251 	mutex_lock(&virtio_fs_mutex);
252 
253 	list_for_each_entry(fs2, &virtio_fs_instances, list) {
254 		if (strcmp(fs->tag, fs2->tag) == 0)
255 			duplicate = true;
256 	}
257 
258 	if (!duplicate)
259 		list_add_tail(&fs->list, &virtio_fs_instances);
260 
261 	mutex_unlock(&virtio_fs_mutex);
262 
263 	if (duplicate)
264 		return -EEXIST;
265 	return 0;
266 }
267 
268 /* Return the virtio_fs with a given tag, or NULL */
269 static struct virtio_fs *virtio_fs_find_instance(const char *tag)
270 {
271 	struct virtio_fs *fs;
272 
273 	mutex_lock(&virtio_fs_mutex);
274 
275 	list_for_each_entry(fs, &virtio_fs_instances, list) {
276 		if (strcmp(fs->tag, tag) == 0) {
277 			kref_get(&fs->refcount);
278 			goto found;
279 		}
280 	}
281 
282 	fs = NULL; /* not found */
283 
284 found:
285 	mutex_unlock(&virtio_fs_mutex);
286 
287 	return fs;
288 }
289 
290 static void virtio_fs_free_devs(struct virtio_fs *fs)
291 {
292 	unsigned int i;
293 
294 	for (i = 0; i < fs->nvqs; i++) {
295 		struct virtio_fs_vq *fsvq = &fs->vqs[i];
296 
297 		if (!fsvq->fud)
298 			continue;
299 
300 		fuse_dev_free(fsvq->fud);
301 		fsvq->fud = NULL;
302 	}
303 }
304 
305 /* Read filesystem name from virtio config into fs->tag (must kfree()). */
306 static int virtio_fs_read_tag(struct virtio_device *vdev, struct virtio_fs *fs)
307 {
308 	char tag_buf[sizeof_field(struct virtio_fs_config, tag)];
309 	char *end;
310 	size_t len;
311 
312 	virtio_cread_bytes(vdev, offsetof(struct virtio_fs_config, tag),
313 			   &tag_buf, sizeof(tag_buf));
314 	end = memchr(tag_buf, '\0', sizeof(tag_buf));
315 	if (end == tag_buf)
316 		return -EINVAL; /* empty tag */
317 	if (!end)
318 		end = &tag_buf[sizeof(tag_buf)];
319 
320 	len = end - tag_buf;
321 	fs->tag = devm_kmalloc(&vdev->dev, len + 1, GFP_KERNEL);
322 	if (!fs->tag)
323 		return -ENOMEM;
324 	memcpy(fs->tag, tag_buf, len);
325 	fs->tag[len] = '\0';
326 	return 0;
327 }
328 
329 /* Work function for hiprio completion */
330 static void virtio_fs_hiprio_done_work(struct work_struct *work)
331 {
332 	struct virtio_fs_vq *fsvq = container_of(work, struct virtio_fs_vq,
333 						 done_work);
334 	struct virtqueue *vq = fsvq->vq;
335 
336 	/* Free completed FUSE_FORGET requests */
337 	spin_lock(&fsvq->lock);
338 	do {
339 		unsigned int len;
340 		void *req;
341 
342 		virtqueue_disable_cb(vq);
343 
344 		while ((req = virtqueue_get_buf(vq, &len)) != NULL) {
345 			kfree(req);
346 			dec_in_flight_req(fsvq);
347 		}
348 	} while (!virtqueue_enable_cb(vq) && likely(!virtqueue_is_broken(vq)));
349 	spin_unlock(&fsvq->lock);
350 }
351 
352 static void virtio_fs_request_dispatch_work(struct work_struct *work)
353 {
354 	struct fuse_req *req;
355 	struct virtio_fs_vq *fsvq = container_of(work, struct virtio_fs_vq,
356 						 dispatch_work.work);
357 	int ret;
358 
359 	pr_debug("virtio-fs: worker %s called.\n", __func__);
360 	while (1) {
361 		spin_lock(&fsvq->lock);
362 		req = list_first_entry_or_null(&fsvq->end_reqs, struct fuse_req,
363 					       list);
364 		if (!req) {
365 			spin_unlock(&fsvq->lock);
366 			break;
367 		}
368 
369 		list_del_init(&req->list);
370 		spin_unlock(&fsvq->lock);
371 		fuse_request_end(req);
372 	}
373 
374 	/* Dispatch pending requests */
375 	while (1) {
376 		spin_lock(&fsvq->lock);
377 		req = list_first_entry_or_null(&fsvq->queued_reqs,
378 					       struct fuse_req, list);
379 		if (!req) {
380 			spin_unlock(&fsvq->lock);
381 			return;
382 		}
383 		list_del_init(&req->list);
384 		spin_unlock(&fsvq->lock);
385 
386 		ret = virtio_fs_enqueue_req(fsvq, req, true);
387 		if (ret < 0) {
388 			if (ret == -ENOMEM || ret == -ENOSPC) {
389 				spin_lock(&fsvq->lock);
390 				list_add_tail(&req->list, &fsvq->queued_reqs);
391 				schedule_delayed_work(&fsvq->dispatch_work,
392 						      msecs_to_jiffies(1));
393 				spin_unlock(&fsvq->lock);
394 				return;
395 			}
396 			req->out.h.error = ret;
397 			spin_lock(&fsvq->lock);
398 			dec_in_flight_req(fsvq);
399 			spin_unlock(&fsvq->lock);
400 			pr_err("virtio-fs: virtio_fs_enqueue_req() failed %d\n",
401 			       ret);
402 			fuse_request_end(req);
403 		}
404 	}
405 }
406 
407 /*
408  * Returns 1 if queue is full and sender should wait a bit before sending
409  * next request, 0 otherwise.
410  */
411 static int send_forget_request(struct virtio_fs_vq *fsvq,
412 			       struct virtio_fs_forget *forget,
413 			       bool in_flight)
414 {
415 	struct scatterlist sg;
416 	struct virtqueue *vq;
417 	int ret = 0;
418 	bool notify;
419 	struct virtio_fs_forget_req *req = &forget->req;
420 
421 	spin_lock(&fsvq->lock);
422 	if (!fsvq->connected) {
423 		if (in_flight)
424 			dec_in_flight_req(fsvq);
425 		kfree(forget);
426 		goto out;
427 	}
428 
429 	sg_init_one(&sg, req, sizeof(*req));
430 	vq = fsvq->vq;
431 	dev_dbg(&vq->vdev->dev, "%s\n", __func__);
432 
433 	ret = virtqueue_add_outbuf(vq, &sg, 1, forget, GFP_ATOMIC);
434 	if (ret < 0) {
435 		if (ret == -ENOMEM || ret == -ENOSPC) {
436 			pr_debug("virtio-fs: Could not queue FORGET: err=%d. Will try later\n",
437 				 ret);
438 			list_add_tail(&forget->list, &fsvq->queued_reqs);
439 			schedule_delayed_work(&fsvq->dispatch_work,
440 					      msecs_to_jiffies(1));
441 			if (!in_flight)
442 				inc_in_flight_req(fsvq);
443 			/* Queue is full */
444 			ret = 1;
445 		} else {
446 			pr_debug("virtio-fs: Could not queue FORGET: err=%d. Dropping it.\n",
447 				 ret);
448 			kfree(forget);
449 			if (in_flight)
450 				dec_in_flight_req(fsvq);
451 		}
452 		goto out;
453 	}
454 
455 	if (!in_flight)
456 		inc_in_flight_req(fsvq);
457 	notify = virtqueue_kick_prepare(vq);
458 	spin_unlock(&fsvq->lock);
459 
460 	if (notify)
461 		virtqueue_notify(vq);
462 	return ret;
463 out:
464 	spin_unlock(&fsvq->lock);
465 	return ret;
466 }
467 
468 static void virtio_fs_hiprio_dispatch_work(struct work_struct *work)
469 {
470 	struct virtio_fs_forget *forget;
471 	struct virtio_fs_vq *fsvq = container_of(work, struct virtio_fs_vq,
472 						 dispatch_work.work);
473 	pr_debug("virtio-fs: worker %s called.\n", __func__);
474 	while (1) {
475 		spin_lock(&fsvq->lock);
476 		forget = list_first_entry_or_null(&fsvq->queued_reqs,
477 					struct virtio_fs_forget, list);
478 		if (!forget) {
479 			spin_unlock(&fsvq->lock);
480 			return;
481 		}
482 
483 		list_del(&forget->list);
484 		spin_unlock(&fsvq->lock);
485 		if (send_forget_request(fsvq, forget, true))
486 			return;
487 	}
488 }
489 
490 /* Allocate and copy args into req->argbuf */
491 static int copy_args_to_argbuf(struct fuse_req *req)
492 {
493 	struct fuse_args *args = req->args;
494 	unsigned int offset = 0;
495 	unsigned int num_in;
496 	unsigned int num_out;
497 	unsigned int len;
498 	unsigned int i;
499 
500 	num_in = args->in_numargs - args->in_pages;
501 	num_out = args->out_numargs - args->out_pages;
502 	len = fuse_len_args(num_in, (struct fuse_arg *) args->in_args) +
503 	      fuse_len_args(num_out, args->out_args);
504 
505 	req->argbuf = kmalloc(len, GFP_ATOMIC);
506 	if (!req->argbuf)
507 		return -ENOMEM;
508 
509 	for (i = 0; i < num_in; i++) {
510 		memcpy(req->argbuf + offset,
511 		       args->in_args[i].value,
512 		       args->in_args[i].size);
513 		offset += args->in_args[i].size;
514 	}
515 
516 	return 0;
517 }
518 
519 /* Copy args out of and free req->argbuf */
520 static void copy_args_from_argbuf(struct fuse_args *args, struct fuse_req *req)
521 {
522 	unsigned int remaining;
523 	unsigned int offset;
524 	unsigned int num_in;
525 	unsigned int num_out;
526 	unsigned int i;
527 
528 	remaining = req->out.h.len - sizeof(req->out.h);
529 	num_in = args->in_numargs - args->in_pages;
530 	num_out = args->out_numargs - args->out_pages;
531 	offset = fuse_len_args(num_in, (struct fuse_arg *)args->in_args);
532 
533 	for (i = 0; i < num_out; i++) {
534 		unsigned int argsize = args->out_args[i].size;
535 
536 		if (args->out_argvar &&
537 		    i == args->out_numargs - 1 &&
538 		    argsize > remaining) {
539 			argsize = remaining;
540 		}
541 
542 		memcpy(args->out_args[i].value, req->argbuf + offset, argsize);
543 		offset += argsize;
544 
545 		if (i != args->out_numargs - 1)
546 			remaining -= argsize;
547 	}
548 
549 	/* Store the actual size of the variable-length arg */
550 	if (args->out_argvar)
551 		args->out_args[args->out_numargs - 1].size = remaining;
552 
553 	kfree(req->argbuf);
554 	req->argbuf = NULL;
555 }
556 
557 /* Work function for request completion */
558 static void virtio_fs_request_complete(struct fuse_req *req,
559 				       struct virtio_fs_vq *fsvq)
560 {
561 	struct fuse_pqueue *fpq = &fsvq->fud->pq;
562 	struct fuse_args *args;
563 	struct fuse_args_pages *ap;
564 	unsigned int len, i, thislen;
565 	struct page *page;
566 
567 	/*
568 	 * TODO verify that server properly follows FUSE protocol
569 	 * (oh.uniq, oh.len)
570 	 */
571 	args = req->args;
572 	copy_args_from_argbuf(args, req);
573 
574 	if (args->out_pages && args->page_zeroing) {
575 		len = args->out_args[args->out_numargs - 1].size;
576 		ap = container_of(args, typeof(*ap), args);
577 		for (i = 0; i < ap->num_pages; i++) {
578 			thislen = ap->descs[i].length;
579 			if (len < thislen) {
580 				WARN_ON(ap->descs[i].offset);
581 				page = ap->pages[i];
582 				zero_user_segment(page, len, thislen);
583 				len = 0;
584 			} else {
585 				len -= thislen;
586 			}
587 		}
588 	}
589 
590 	spin_lock(&fpq->lock);
591 	clear_bit(FR_SENT, &req->flags);
592 	spin_unlock(&fpq->lock);
593 
594 	fuse_request_end(req);
595 	spin_lock(&fsvq->lock);
596 	dec_in_flight_req(fsvq);
597 	spin_unlock(&fsvq->lock);
598 }
599 
600 static void virtio_fs_complete_req_work(struct work_struct *work)
601 {
602 	struct virtio_fs_req_work *w =
603 		container_of(work, typeof(*w), done_work);
604 
605 	virtio_fs_request_complete(w->req, w->fsvq);
606 	kfree(w);
607 }
608 
609 static void virtio_fs_requests_done_work(struct work_struct *work)
610 {
611 	struct virtio_fs_vq *fsvq = container_of(work, struct virtio_fs_vq,
612 						 done_work);
613 	struct fuse_pqueue *fpq = &fsvq->fud->pq;
614 	struct virtqueue *vq = fsvq->vq;
615 	struct fuse_req *req;
616 	struct fuse_req *next;
617 	unsigned int len;
618 	LIST_HEAD(reqs);
619 
620 	/* Collect completed requests off the virtqueue */
621 	spin_lock(&fsvq->lock);
622 	do {
623 		virtqueue_disable_cb(vq);
624 
625 		while ((req = virtqueue_get_buf(vq, &len)) != NULL) {
626 			spin_lock(&fpq->lock);
627 			list_move_tail(&req->list, &reqs);
628 			spin_unlock(&fpq->lock);
629 		}
630 	} while (!virtqueue_enable_cb(vq) && likely(!virtqueue_is_broken(vq)));
631 	spin_unlock(&fsvq->lock);
632 
633 	/* End requests */
634 	list_for_each_entry_safe(req, next, &reqs, list) {
635 		list_del_init(&req->list);
636 
637 		/* blocking async request completes in a worker context */
638 		if (req->args->may_block) {
639 			struct virtio_fs_req_work *w;
640 
641 			w = kzalloc(sizeof(*w), GFP_NOFS | __GFP_NOFAIL);
642 			INIT_WORK(&w->done_work, virtio_fs_complete_req_work);
643 			w->fsvq = fsvq;
644 			w->req = req;
645 			schedule_work(&w->done_work);
646 		} else {
647 			virtio_fs_request_complete(req, fsvq);
648 		}
649 	}
650 }
651 
652 /* Virtqueue interrupt handler */
653 static void virtio_fs_vq_done(struct virtqueue *vq)
654 {
655 	struct virtio_fs_vq *fsvq = vq_to_fsvq(vq);
656 
657 	dev_dbg(&vq->vdev->dev, "%s %s\n", __func__, fsvq->name);
658 
659 	schedule_work(&fsvq->done_work);
660 }
661 
662 static void virtio_fs_init_vq(struct virtio_fs_vq *fsvq, char *name,
663 			      int vq_type)
664 {
665 	strscpy(fsvq->name, name, VQ_NAME_LEN);
666 	spin_lock_init(&fsvq->lock);
667 	INIT_LIST_HEAD(&fsvq->queued_reqs);
668 	INIT_LIST_HEAD(&fsvq->end_reqs);
669 	init_completion(&fsvq->in_flight_zero);
670 
671 	if (vq_type == VQ_REQUEST) {
672 		INIT_WORK(&fsvq->done_work, virtio_fs_requests_done_work);
673 		INIT_DELAYED_WORK(&fsvq->dispatch_work,
674 				  virtio_fs_request_dispatch_work);
675 	} else {
676 		INIT_WORK(&fsvq->done_work, virtio_fs_hiprio_done_work);
677 		INIT_DELAYED_WORK(&fsvq->dispatch_work,
678 				  virtio_fs_hiprio_dispatch_work);
679 	}
680 }
681 
682 /* Initialize virtqueues */
683 static int virtio_fs_setup_vqs(struct virtio_device *vdev,
684 			       struct virtio_fs *fs)
685 {
686 	struct virtqueue **vqs;
687 	vq_callback_t **callbacks;
688 	const char **names;
689 	unsigned int i;
690 	int ret = 0;
691 
692 	virtio_cread_le(vdev, struct virtio_fs_config, num_request_queues,
693 			&fs->num_request_queues);
694 	if (fs->num_request_queues == 0)
695 		return -EINVAL;
696 
697 	fs->nvqs = VQ_REQUEST + fs->num_request_queues;
698 	fs->vqs = kcalloc(fs->nvqs, sizeof(fs->vqs[VQ_HIPRIO]), GFP_KERNEL);
699 	if (!fs->vqs)
700 		return -ENOMEM;
701 
702 	vqs = kmalloc_array(fs->nvqs, sizeof(vqs[VQ_HIPRIO]), GFP_KERNEL);
703 	callbacks = kmalloc_array(fs->nvqs, sizeof(callbacks[VQ_HIPRIO]),
704 					GFP_KERNEL);
705 	names = kmalloc_array(fs->nvqs, sizeof(names[VQ_HIPRIO]), GFP_KERNEL);
706 	if (!vqs || !callbacks || !names) {
707 		ret = -ENOMEM;
708 		goto out;
709 	}
710 
711 	/* Initialize the hiprio/forget request virtqueue */
712 	callbacks[VQ_HIPRIO] = virtio_fs_vq_done;
713 	virtio_fs_init_vq(&fs->vqs[VQ_HIPRIO], "hiprio", VQ_HIPRIO);
714 	names[VQ_HIPRIO] = fs->vqs[VQ_HIPRIO].name;
715 
716 	/* Initialize the requests virtqueues */
717 	for (i = VQ_REQUEST; i < fs->nvqs; i++) {
718 		char vq_name[VQ_NAME_LEN];
719 
720 		snprintf(vq_name, VQ_NAME_LEN, "requests.%u", i - VQ_REQUEST);
721 		virtio_fs_init_vq(&fs->vqs[i], vq_name, VQ_REQUEST);
722 		callbacks[i] = virtio_fs_vq_done;
723 		names[i] = fs->vqs[i].name;
724 	}
725 
726 	ret = virtio_find_vqs(vdev, fs->nvqs, vqs, callbacks, names, NULL);
727 	if (ret < 0)
728 		goto out;
729 
730 	for (i = 0; i < fs->nvqs; i++)
731 		fs->vqs[i].vq = vqs[i];
732 
733 	virtio_fs_start_all_queues(fs);
734 out:
735 	kfree(names);
736 	kfree(callbacks);
737 	kfree(vqs);
738 	if (ret)
739 		kfree(fs->vqs);
740 	return ret;
741 }
742 
743 /* Free virtqueues (device must already be reset) */
744 static void virtio_fs_cleanup_vqs(struct virtio_device *vdev)
745 {
746 	vdev->config->del_vqs(vdev);
747 }
748 
749 /* Map a window offset to a page frame number.  The window offset will have
750  * been produced by .iomap_begin(), which maps a file offset to a window
751  * offset.
752  */
753 static long virtio_fs_direct_access(struct dax_device *dax_dev, pgoff_t pgoff,
754 				    long nr_pages, enum dax_access_mode mode,
755 				    void **kaddr, pfn_t *pfn)
756 {
757 	struct virtio_fs *fs = dax_get_private(dax_dev);
758 	phys_addr_t offset = PFN_PHYS(pgoff);
759 	size_t max_nr_pages = fs->window_len / PAGE_SIZE - pgoff;
760 
761 	if (kaddr)
762 		*kaddr = fs->window_kaddr + offset;
763 	if (pfn)
764 		*pfn = phys_to_pfn_t(fs->window_phys_addr + offset,
765 					PFN_DEV | PFN_MAP);
766 	return nr_pages > max_nr_pages ? max_nr_pages : nr_pages;
767 }
768 
769 static int virtio_fs_zero_page_range(struct dax_device *dax_dev,
770 				     pgoff_t pgoff, size_t nr_pages)
771 {
772 	long rc;
773 	void *kaddr;
774 
775 	rc = dax_direct_access(dax_dev, pgoff, nr_pages, DAX_ACCESS, &kaddr,
776 			       NULL);
777 	if (rc < 0)
778 		return dax_mem2blk_err(rc);
779 
780 	memset(kaddr, 0, nr_pages << PAGE_SHIFT);
781 	dax_flush(dax_dev, kaddr, nr_pages << PAGE_SHIFT);
782 	return 0;
783 }
784 
785 static const struct dax_operations virtio_fs_dax_ops = {
786 	.direct_access = virtio_fs_direct_access,
787 	.zero_page_range = virtio_fs_zero_page_range,
788 };
789 
790 static void virtio_fs_cleanup_dax(void *data)
791 {
792 	struct dax_device *dax_dev = data;
793 
794 	kill_dax(dax_dev);
795 	put_dax(dax_dev);
796 }
797 
798 static int virtio_fs_setup_dax(struct virtio_device *vdev, struct virtio_fs *fs)
799 {
800 	struct virtio_shm_region cache_reg;
801 	struct dev_pagemap *pgmap;
802 	bool have_cache;
803 
804 	if (!IS_ENABLED(CONFIG_FUSE_DAX))
805 		return 0;
806 
807 	/* Get cache region */
808 	have_cache = virtio_get_shm_region(vdev, &cache_reg,
809 					   (u8)VIRTIO_FS_SHMCAP_ID_CACHE);
810 	if (!have_cache) {
811 		dev_notice(&vdev->dev, "%s: No cache capability\n", __func__);
812 		return 0;
813 	}
814 
815 	if (!devm_request_mem_region(&vdev->dev, cache_reg.addr, cache_reg.len,
816 				     dev_name(&vdev->dev))) {
817 		dev_warn(&vdev->dev, "could not reserve region addr=0x%llx len=0x%llx\n",
818 			 cache_reg.addr, cache_reg.len);
819 		return -EBUSY;
820 	}
821 
822 	dev_notice(&vdev->dev, "Cache len: 0x%llx @ 0x%llx\n", cache_reg.len,
823 		   cache_reg.addr);
824 
825 	pgmap = devm_kzalloc(&vdev->dev, sizeof(*pgmap), GFP_KERNEL);
826 	if (!pgmap)
827 		return -ENOMEM;
828 
829 	pgmap->type = MEMORY_DEVICE_FS_DAX;
830 
831 	/* Ideally we would directly use the PCI BAR resource but
832 	 * devm_memremap_pages() wants its own copy in pgmap.  So
833 	 * initialize a struct resource from scratch (only the start
834 	 * and end fields will be used).
835 	 */
836 	pgmap->range = (struct range) {
837 		.start = (phys_addr_t) cache_reg.addr,
838 		.end = (phys_addr_t) cache_reg.addr + cache_reg.len - 1,
839 	};
840 	pgmap->nr_range = 1;
841 
842 	fs->window_kaddr = devm_memremap_pages(&vdev->dev, pgmap);
843 	if (IS_ERR(fs->window_kaddr))
844 		return PTR_ERR(fs->window_kaddr);
845 
846 	fs->window_phys_addr = (phys_addr_t) cache_reg.addr;
847 	fs->window_len = (phys_addr_t) cache_reg.len;
848 
849 	dev_dbg(&vdev->dev, "%s: window kaddr 0x%px phys_addr 0x%llx len 0x%llx\n",
850 		__func__, fs->window_kaddr, cache_reg.addr, cache_reg.len);
851 
852 	fs->dax_dev = alloc_dax(fs, &virtio_fs_dax_ops);
853 	if (IS_ERR(fs->dax_dev))
854 		return PTR_ERR(fs->dax_dev);
855 
856 	return devm_add_action_or_reset(&vdev->dev, virtio_fs_cleanup_dax,
857 					fs->dax_dev);
858 }
859 
860 static int virtio_fs_probe(struct virtio_device *vdev)
861 {
862 	struct virtio_fs *fs;
863 	int ret;
864 
865 	fs = kzalloc(sizeof(*fs), GFP_KERNEL);
866 	if (!fs)
867 		return -ENOMEM;
868 	kref_init(&fs->refcount);
869 	vdev->priv = fs;
870 
871 	ret = virtio_fs_read_tag(vdev, fs);
872 	if (ret < 0)
873 		goto out;
874 
875 	ret = virtio_fs_setup_vqs(vdev, fs);
876 	if (ret < 0)
877 		goto out;
878 
879 	/* TODO vq affinity */
880 
881 	ret = virtio_fs_setup_dax(vdev, fs);
882 	if (ret < 0)
883 		goto out_vqs;
884 
885 	/* Bring the device online in case the filesystem is mounted and
886 	 * requests need to be sent before we return.
887 	 */
888 	virtio_device_ready(vdev);
889 
890 	ret = virtio_fs_add_instance(fs);
891 	if (ret < 0)
892 		goto out_vqs;
893 
894 	return 0;
895 
896 out_vqs:
897 	virtio_reset_device(vdev);
898 	virtio_fs_cleanup_vqs(vdev);
899 	kfree(fs->vqs);
900 
901 out:
902 	vdev->priv = NULL;
903 	kfree(fs);
904 	return ret;
905 }
906 
907 static void virtio_fs_stop_all_queues(struct virtio_fs *fs)
908 {
909 	struct virtio_fs_vq *fsvq;
910 	int i;
911 
912 	for (i = 0; i < fs->nvqs; i++) {
913 		fsvq = &fs->vqs[i];
914 		spin_lock(&fsvq->lock);
915 		fsvq->connected = false;
916 		spin_unlock(&fsvq->lock);
917 	}
918 }
919 
920 static void virtio_fs_remove(struct virtio_device *vdev)
921 {
922 	struct virtio_fs *fs = vdev->priv;
923 
924 	mutex_lock(&virtio_fs_mutex);
925 	/* This device is going away. No one should get new reference */
926 	list_del_init(&fs->list);
927 	virtio_fs_stop_all_queues(fs);
928 	virtio_fs_drain_all_queues_locked(fs);
929 	virtio_reset_device(vdev);
930 	virtio_fs_cleanup_vqs(vdev);
931 
932 	vdev->priv = NULL;
933 	/* Put device reference on virtio_fs object */
934 	virtio_fs_put(fs);
935 	mutex_unlock(&virtio_fs_mutex);
936 }
937 
938 #ifdef CONFIG_PM_SLEEP
939 static int virtio_fs_freeze(struct virtio_device *vdev)
940 {
941 	/* TODO need to save state here */
942 	pr_warn("virtio-fs: suspend/resume not yet supported\n");
943 	return -EOPNOTSUPP;
944 }
945 
946 static int virtio_fs_restore(struct virtio_device *vdev)
947 {
948 	 /* TODO need to restore state here */
949 	return 0;
950 }
951 #endif /* CONFIG_PM_SLEEP */
952 
953 static const struct virtio_device_id id_table[] = {
954 	{ VIRTIO_ID_FS, VIRTIO_DEV_ANY_ID },
955 	{},
956 };
957 
958 static const unsigned int feature_table[] = {};
959 
960 static struct virtio_driver virtio_fs_driver = {
961 	.driver.name		= KBUILD_MODNAME,
962 	.driver.owner		= THIS_MODULE,
963 	.id_table		= id_table,
964 	.feature_table		= feature_table,
965 	.feature_table_size	= ARRAY_SIZE(feature_table),
966 	.probe			= virtio_fs_probe,
967 	.remove			= virtio_fs_remove,
968 #ifdef CONFIG_PM_SLEEP
969 	.freeze			= virtio_fs_freeze,
970 	.restore		= virtio_fs_restore,
971 #endif
972 };
973 
974 static void virtio_fs_wake_forget_and_unlock(struct fuse_iqueue *fiq)
975 __releases(fiq->lock)
976 {
977 	struct fuse_forget_link *link;
978 	struct virtio_fs_forget *forget;
979 	struct virtio_fs_forget_req *req;
980 	struct virtio_fs *fs;
981 	struct virtio_fs_vq *fsvq;
982 	u64 unique;
983 
984 	link = fuse_dequeue_forget(fiq, 1, NULL);
985 	unique = fuse_get_unique(fiq);
986 
987 	fs = fiq->priv;
988 	fsvq = &fs->vqs[VQ_HIPRIO];
989 	spin_unlock(&fiq->lock);
990 
991 	/* Allocate a buffer for the request */
992 	forget = kmalloc(sizeof(*forget), GFP_NOFS | __GFP_NOFAIL);
993 	req = &forget->req;
994 
995 	req->ih = (struct fuse_in_header){
996 		.opcode = FUSE_FORGET,
997 		.nodeid = link->forget_one.nodeid,
998 		.unique = unique,
999 		.len = sizeof(*req),
1000 	};
1001 	req->arg = (struct fuse_forget_in){
1002 		.nlookup = link->forget_one.nlookup,
1003 	};
1004 
1005 	send_forget_request(fsvq, forget, false);
1006 	kfree(link);
1007 }
1008 
1009 static void virtio_fs_wake_interrupt_and_unlock(struct fuse_iqueue *fiq)
1010 __releases(fiq->lock)
1011 {
1012 	/*
1013 	 * TODO interrupts.
1014 	 *
1015 	 * Normal fs operations on a local filesystems aren't interruptible.
1016 	 * Exceptions are blocking lock operations; for example fcntl(F_SETLKW)
1017 	 * with shared lock between host and guest.
1018 	 */
1019 	spin_unlock(&fiq->lock);
1020 }
1021 
1022 /* Count number of scatter-gather elements required */
1023 static unsigned int sg_count_fuse_pages(struct fuse_page_desc *page_descs,
1024 				       unsigned int num_pages,
1025 				       unsigned int total_len)
1026 {
1027 	unsigned int i;
1028 	unsigned int this_len;
1029 
1030 	for (i = 0; i < num_pages && total_len; i++) {
1031 		this_len =  min(page_descs[i].length, total_len);
1032 		total_len -= this_len;
1033 	}
1034 
1035 	return i;
1036 }
1037 
1038 /* Return the number of scatter-gather list elements required */
1039 static unsigned int sg_count_fuse_req(struct fuse_req *req)
1040 {
1041 	struct fuse_args *args = req->args;
1042 	struct fuse_args_pages *ap = container_of(args, typeof(*ap), args);
1043 	unsigned int size, total_sgs = 1 /* fuse_in_header */;
1044 
1045 	if (args->in_numargs - args->in_pages)
1046 		total_sgs += 1;
1047 
1048 	if (args->in_pages) {
1049 		size = args->in_args[args->in_numargs - 1].size;
1050 		total_sgs += sg_count_fuse_pages(ap->descs, ap->num_pages,
1051 						 size);
1052 	}
1053 
1054 	if (!test_bit(FR_ISREPLY, &req->flags))
1055 		return total_sgs;
1056 
1057 	total_sgs += 1 /* fuse_out_header */;
1058 
1059 	if (args->out_numargs - args->out_pages)
1060 		total_sgs += 1;
1061 
1062 	if (args->out_pages) {
1063 		size = args->out_args[args->out_numargs - 1].size;
1064 		total_sgs += sg_count_fuse_pages(ap->descs, ap->num_pages,
1065 						 size);
1066 	}
1067 
1068 	return total_sgs;
1069 }
1070 
1071 /* Add pages to scatter-gather list and return number of elements used */
1072 static unsigned int sg_init_fuse_pages(struct scatterlist *sg,
1073 				       struct page **pages,
1074 				       struct fuse_page_desc *page_descs,
1075 				       unsigned int num_pages,
1076 				       unsigned int total_len)
1077 {
1078 	unsigned int i;
1079 	unsigned int this_len;
1080 
1081 	for (i = 0; i < num_pages && total_len; i++) {
1082 		sg_init_table(&sg[i], 1);
1083 		this_len =  min(page_descs[i].length, total_len);
1084 		sg_set_page(&sg[i], pages[i], this_len, page_descs[i].offset);
1085 		total_len -= this_len;
1086 	}
1087 
1088 	return i;
1089 }
1090 
1091 /* Add args to scatter-gather list and return number of elements used */
1092 static unsigned int sg_init_fuse_args(struct scatterlist *sg,
1093 				      struct fuse_req *req,
1094 				      struct fuse_arg *args,
1095 				      unsigned int numargs,
1096 				      bool argpages,
1097 				      void *argbuf,
1098 				      unsigned int *len_used)
1099 {
1100 	struct fuse_args_pages *ap = container_of(req->args, typeof(*ap), args);
1101 	unsigned int total_sgs = 0;
1102 	unsigned int len;
1103 
1104 	len = fuse_len_args(numargs - argpages, args);
1105 	if (len)
1106 		sg_init_one(&sg[total_sgs++], argbuf, len);
1107 
1108 	if (argpages)
1109 		total_sgs += sg_init_fuse_pages(&sg[total_sgs],
1110 						ap->pages, ap->descs,
1111 						ap->num_pages,
1112 						args[numargs - 1].size);
1113 
1114 	if (len_used)
1115 		*len_used = len;
1116 
1117 	return total_sgs;
1118 }
1119 
1120 /* Add a request to a virtqueue and kick the device */
1121 static int virtio_fs_enqueue_req(struct virtio_fs_vq *fsvq,
1122 				 struct fuse_req *req, bool in_flight)
1123 {
1124 	/* requests need at least 4 elements */
1125 	struct scatterlist *stack_sgs[6];
1126 	struct scatterlist stack_sg[ARRAY_SIZE(stack_sgs)];
1127 	struct scatterlist **sgs = stack_sgs;
1128 	struct scatterlist *sg = stack_sg;
1129 	struct virtqueue *vq;
1130 	struct fuse_args *args = req->args;
1131 	unsigned int argbuf_used = 0;
1132 	unsigned int out_sgs = 0;
1133 	unsigned int in_sgs = 0;
1134 	unsigned int total_sgs;
1135 	unsigned int i;
1136 	int ret;
1137 	bool notify;
1138 	struct fuse_pqueue *fpq;
1139 
1140 	/* Does the sglist fit on the stack? */
1141 	total_sgs = sg_count_fuse_req(req);
1142 	if (total_sgs > ARRAY_SIZE(stack_sgs)) {
1143 		sgs = kmalloc_array(total_sgs, sizeof(sgs[0]), GFP_ATOMIC);
1144 		sg = kmalloc_array(total_sgs, sizeof(sg[0]), GFP_ATOMIC);
1145 		if (!sgs || !sg) {
1146 			ret = -ENOMEM;
1147 			goto out;
1148 		}
1149 	}
1150 
1151 	/* Use a bounce buffer since stack args cannot be mapped */
1152 	ret = copy_args_to_argbuf(req);
1153 	if (ret < 0)
1154 		goto out;
1155 
1156 	/* Request elements */
1157 	sg_init_one(&sg[out_sgs++], &req->in.h, sizeof(req->in.h));
1158 	out_sgs += sg_init_fuse_args(&sg[out_sgs], req,
1159 				     (struct fuse_arg *)args->in_args,
1160 				     args->in_numargs, args->in_pages,
1161 				     req->argbuf, &argbuf_used);
1162 
1163 	/* Reply elements */
1164 	if (test_bit(FR_ISREPLY, &req->flags)) {
1165 		sg_init_one(&sg[out_sgs + in_sgs++],
1166 			    &req->out.h, sizeof(req->out.h));
1167 		in_sgs += sg_init_fuse_args(&sg[out_sgs + in_sgs], req,
1168 					    args->out_args, args->out_numargs,
1169 					    args->out_pages,
1170 					    req->argbuf + argbuf_used, NULL);
1171 	}
1172 
1173 	WARN_ON(out_sgs + in_sgs != total_sgs);
1174 
1175 	for (i = 0; i < total_sgs; i++)
1176 		sgs[i] = &sg[i];
1177 
1178 	spin_lock(&fsvq->lock);
1179 
1180 	if (!fsvq->connected) {
1181 		spin_unlock(&fsvq->lock);
1182 		ret = -ENOTCONN;
1183 		goto out;
1184 	}
1185 
1186 	vq = fsvq->vq;
1187 	ret = virtqueue_add_sgs(vq, sgs, out_sgs, in_sgs, req, GFP_ATOMIC);
1188 	if (ret < 0) {
1189 		spin_unlock(&fsvq->lock);
1190 		goto out;
1191 	}
1192 
1193 	/* Request successfully sent. */
1194 	fpq = &fsvq->fud->pq;
1195 	spin_lock(&fpq->lock);
1196 	list_add_tail(&req->list, fpq->processing);
1197 	spin_unlock(&fpq->lock);
1198 	set_bit(FR_SENT, &req->flags);
1199 	/* matches barrier in request_wait_answer() */
1200 	smp_mb__after_atomic();
1201 
1202 	if (!in_flight)
1203 		inc_in_flight_req(fsvq);
1204 	notify = virtqueue_kick_prepare(vq);
1205 
1206 	spin_unlock(&fsvq->lock);
1207 
1208 	if (notify)
1209 		virtqueue_notify(vq);
1210 
1211 out:
1212 	if (ret < 0 && req->argbuf) {
1213 		kfree(req->argbuf);
1214 		req->argbuf = NULL;
1215 	}
1216 	if (sgs != stack_sgs) {
1217 		kfree(sgs);
1218 		kfree(sg);
1219 	}
1220 
1221 	return ret;
1222 }
1223 
1224 static void virtio_fs_wake_pending_and_unlock(struct fuse_iqueue *fiq)
1225 __releases(fiq->lock)
1226 {
1227 	unsigned int queue_id = VQ_REQUEST; /* TODO multiqueue */
1228 	struct virtio_fs *fs;
1229 	struct fuse_req *req;
1230 	struct virtio_fs_vq *fsvq;
1231 	int ret;
1232 
1233 	WARN_ON(list_empty(&fiq->pending));
1234 	req = list_last_entry(&fiq->pending, struct fuse_req, list);
1235 	clear_bit(FR_PENDING, &req->flags);
1236 	list_del_init(&req->list);
1237 	WARN_ON(!list_empty(&fiq->pending));
1238 	spin_unlock(&fiq->lock);
1239 
1240 	fs = fiq->priv;
1241 
1242 	pr_debug("%s: opcode %u unique %#llx nodeid %#llx in.len %u out.len %u\n",
1243 		  __func__, req->in.h.opcode, req->in.h.unique,
1244 		 req->in.h.nodeid, req->in.h.len,
1245 		 fuse_len_args(req->args->out_numargs, req->args->out_args));
1246 
1247 	fsvq = &fs->vqs[queue_id];
1248 	ret = virtio_fs_enqueue_req(fsvq, req, false);
1249 	if (ret < 0) {
1250 		if (ret == -ENOMEM || ret == -ENOSPC) {
1251 			/*
1252 			 * Virtqueue full. Retry submission from worker
1253 			 * context as we might be holding fc->bg_lock.
1254 			 */
1255 			spin_lock(&fsvq->lock);
1256 			list_add_tail(&req->list, &fsvq->queued_reqs);
1257 			inc_in_flight_req(fsvq);
1258 			schedule_delayed_work(&fsvq->dispatch_work,
1259 						msecs_to_jiffies(1));
1260 			spin_unlock(&fsvq->lock);
1261 			return;
1262 		}
1263 		req->out.h.error = ret;
1264 		pr_err("virtio-fs: virtio_fs_enqueue_req() failed %d\n", ret);
1265 
1266 		/* Can't end request in submission context. Use a worker */
1267 		spin_lock(&fsvq->lock);
1268 		list_add_tail(&req->list, &fsvq->end_reqs);
1269 		schedule_delayed_work(&fsvq->dispatch_work, 0);
1270 		spin_unlock(&fsvq->lock);
1271 		return;
1272 	}
1273 }
1274 
1275 static const struct fuse_iqueue_ops virtio_fs_fiq_ops = {
1276 	.wake_forget_and_unlock		= virtio_fs_wake_forget_and_unlock,
1277 	.wake_interrupt_and_unlock	= virtio_fs_wake_interrupt_and_unlock,
1278 	.wake_pending_and_unlock	= virtio_fs_wake_pending_and_unlock,
1279 	.release			= virtio_fs_fiq_release,
1280 };
1281 
1282 static inline void virtio_fs_ctx_set_defaults(struct fuse_fs_context *ctx)
1283 {
1284 	ctx->rootmode = S_IFDIR;
1285 	ctx->default_permissions = 1;
1286 	ctx->allow_other = 1;
1287 	ctx->max_read = UINT_MAX;
1288 	ctx->blksize = 512;
1289 	ctx->destroy = true;
1290 	ctx->no_control = true;
1291 	ctx->no_force_umount = true;
1292 }
1293 
1294 static int virtio_fs_fill_super(struct super_block *sb, struct fs_context *fsc)
1295 {
1296 	struct fuse_mount *fm = get_fuse_mount_super(sb);
1297 	struct fuse_conn *fc = fm->fc;
1298 	struct virtio_fs *fs = fc->iq.priv;
1299 	struct fuse_fs_context *ctx = fsc->fs_private;
1300 	unsigned int i;
1301 	int err;
1302 
1303 	virtio_fs_ctx_set_defaults(ctx);
1304 	mutex_lock(&virtio_fs_mutex);
1305 
1306 	/* After holding mutex, make sure virtiofs device is still there.
1307 	 * Though we are holding a reference to it, drive ->remove might
1308 	 * still have cleaned up virtual queues. In that case bail out.
1309 	 */
1310 	err = -EINVAL;
1311 	if (list_empty(&fs->list)) {
1312 		pr_info("virtio-fs: tag <%s> not found\n", fs->tag);
1313 		goto err;
1314 	}
1315 
1316 	err = -ENOMEM;
1317 	/* Allocate fuse_dev for hiprio and notification queues */
1318 	for (i = 0; i < fs->nvqs; i++) {
1319 		struct virtio_fs_vq *fsvq = &fs->vqs[i];
1320 
1321 		fsvq->fud = fuse_dev_alloc();
1322 		if (!fsvq->fud)
1323 			goto err_free_fuse_devs;
1324 	}
1325 
1326 	/* virtiofs allocates and installs its own fuse devices */
1327 	ctx->fudptr = NULL;
1328 	if (ctx->dax_mode != FUSE_DAX_NEVER) {
1329 		if (ctx->dax_mode == FUSE_DAX_ALWAYS && !fs->dax_dev) {
1330 			err = -EINVAL;
1331 			pr_err("virtio-fs: dax can't be enabled as filesystem"
1332 			       " device does not support it.\n");
1333 			goto err_free_fuse_devs;
1334 		}
1335 		ctx->dax_dev = fs->dax_dev;
1336 	}
1337 	err = fuse_fill_super_common(sb, ctx);
1338 	if (err < 0)
1339 		goto err_free_fuse_devs;
1340 
1341 	for (i = 0; i < fs->nvqs; i++) {
1342 		struct virtio_fs_vq *fsvq = &fs->vqs[i];
1343 
1344 		fuse_dev_install(fsvq->fud, fc);
1345 	}
1346 
1347 	/* Previous unmount will stop all queues. Start these again */
1348 	virtio_fs_start_all_queues(fs);
1349 	fuse_send_init(fm);
1350 	mutex_unlock(&virtio_fs_mutex);
1351 	return 0;
1352 
1353 err_free_fuse_devs:
1354 	virtio_fs_free_devs(fs);
1355 err:
1356 	mutex_unlock(&virtio_fs_mutex);
1357 	return err;
1358 }
1359 
1360 static void virtio_fs_conn_destroy(struct fuse_mount *fm)
1361 {
1362 	struct fuse_conn *fc = fm->fc;
1363 	struct virtio_fs *vfs = fc->iq.priv;
1364 	struct virtio_fs_vq *fsvq = &vfs->vqs[VQ_HIPRIO];
1365 
1366 	/* Stop dax worker. Soon evict_inodes() will be called which
1367 	 * will free all memory ranges belonging to all inodes.
1368 	 */
1369 	if (IS_ENABLED(CONFIG_FUSE_DAX))
1370 		fuse_dax_cancel_work(fc);
1371 
1372 	/* Stop forget queue. Soon destroy will be sent */
1373 	spin_lock(&fsvq->lock);
1374 	fsvq->connected = false;
1375 	spin_unlock(&fsvq->lock);
1376 	virtio_fs_drain_all_queues(vfs);
1377 
1378 	fuse_conn_destroy(fm);
1379 
1380 	/* fuse_conn_destroy() must have sent destroy. Stop all queues
1381 	 * and drain one more time and free fuse devices. Freeing fuse
1382 	 * devices will drop their reference on fuse_conn and that in
1383 	 * turn will drop its reference on virtio_fs object.
1384 	 */
1385 	virtio_fs_stop_all_queues(vfs);
1386 	virtio_fs_drain_all_queues(vfs);
1387 	virtio_fs_free_devs(vfs);
1388 }
1389 
1390 static void virtio_kill_sb(struct super_block *sb)
1391 {
1392 	struct fuse_mount *fm = get_fuse_mount_super(sb);
1393 	bool last;
1394 
1395 	/* If mount failed, we can still be called without any fc */
1396 	if (sb->s_root) {
1397 		last = fuse_mount_remove(fm);
1398 		if (last)
1399 			virtio_fs_conn_destroy(fm);
1400 	}
1401 	kill_anon_super(sb);
1402 	fuse_mount_destroy(fm);
1403 }
1404 
1405 static int virtio_fs_test_super(struct super_block *sb,
1406 				struct fs_context *fsc)
1407 {
1408 	struct fuse_mount *fsc_fm = fsc->s_fs_info;
1409 	struct fuse_mount *sb_fm = get_fuse_mount_super(sb);
1410 
1411 	return fsc_fm->fc->iq.priv == sb_fm->fc->iq.priv;
1412 }
1413 
1414 static int virtio_fs_get_tree(struct fs_context *fsc)
1415 {
1416 	struct virtio_fs *fs;
1417 	struct super_block *sb;
1418 	struct fuse_conn *fc = NULL;
1419 	struct fuse_mount *fm;
1420 	unsigned int virtqueue_size;
1421 	int err = -EIO;
1422 
1423 	/* This gets a reference on virtio_fs object. This ptr gets installed
1424 	 * in fc->iq->priv. Once fuse_conn is going away, it calls ->put()
1425 	 * to drop the reference to this object.
1426 	 */
1427 	fs = virtio_fs_find_instance(fsc->source);
1428 	if (!fs) {
1429 		pr_info("virtio-fs: tag <%s> not found\n", fsc->source);
1430 		return -EINVAL;
1431 	}
1432 
1433 	virtqueue_size = virtqueue_get_vring_size(fs->vqs[VQ_REQUEST].vq);
1434 	if (WARN_ON(virtqueue_size <= FUSE_HEADER_OVERHEAD))
1435 		goto out_err;
1436 
1437 	err = -ENOMEM;
1438 	fc = kzalloc(sizeof(struct fuse_conn), GFP_KERNEL);
1439 	if (!fc)
1440 		goto out_err;
1441 
1442 	fm = kzalloc(sizeof(struct fuse_mount), GFP_KERNEL);
1443 	if (!fm)
1444 		goto out_err;
1445 
1446 	fuse_conn_init(fc, fm, fsc->user_ns, &virtio_fs_fiq_ops, fs);
1447 	fc->release = fuse_free_conn;
1448 	fc->delete_stale = true;
1449 	fc->auto_submounts = true;
1450 	fc->sync_fs = true;
1451 
1452 	/* Tell FUSE to split requests that exceed the virtqueue's size */
1453 	fc->max_pages_limit = min_t(unsigned int, fc->max_pages_limit,
1454 				    virtqueue_size - FUSE_HEADER_OVERHEAD);
1455 
1456 	fsc->s_fs_info = fm;
1457 	sb = sget_fc(fsc, virtio_fs_test_super, set_anon_super_fc);
1458 	if (fsc->s_fs_info)
1459 		fuse_mount_destroy(fm);
1460 	if (IS_ERR(sb))
1461 		return PTR_ERR(sb);
1462 
1463 	if (!sb->s_root) {
1464 		err = virtio_fs_fill_super(sb, fsc);
1465 		if (err) {
1466 			deactivate_locked_super(sb);
1467 			return err;
1468 		}
1469 
1470 		sb->s_flags |= SB_ACTIVE;
1471 	}
1472 
1473 	WARN_ON(fsc->root);
1474 	fsc->root = dget(sb->s_root);
1475 	return 0;
1476 
1477 out_err:
1478 	kfree(fc);
1479 	mutex_lock(&virtio_fs_mutex);
1480 	virtio_fs_put(fs);
1481 	mutex_unlock(&virtio_fs_mutex);
1482 	return err;
1483 }
1484 
1485 static const struct fs_context_operations virtio_fs_context_ops = {
1486 	.free		= virtio_fs_free_fsc,
1487 	.parse_param	= virtio_fs_parse_param,
1488 	.get_tree	= virtio_fs_get_tree,
1489 };
1490 
1491 static int virtio_fs_init_fs_context(struct fs_context *fsc)
1492 {
1493 	struct fuse_fs_context *ctx;
1494 
1495 	if (fsc->purpose == FS_CONTEXT_FOR_SUBMOUNT)
1496 		return fuse_init_fs_context_submount(fsc);
1497 
1498 	ctx = kzalloc(sizeof(struct fuse_fs_context), GFP_KERNEL);
1499 	if (!ctx)
1500 		return -ENOMEM;
1501 	fsc->fs_private = ctx;
1502 	fsc->ops = &virtio_fs_context_ops;
1503 	return 0;
1504 }
1505 
1506 static struct file_system_type virtio_fs_type = {
1507 	.owner		= THIS_MODULE,
1508 	.name		= "virtiofs",
1509 	.init_fs_context = virtio_fs_init_fs_context,
1510 	.kill_sb	= virtio_kill_sb,
1511 };
1512 
1513 static int __init virtio_fs_init(void)
1514 {
1515 	int ret;
1516 
1517 	ret = register_virtio_driver(&virtio_fs_driver);
1518 	if (ret < 0)
1519 		return ret;
1520 
1521 	ret = register_filesystem(&virtio_fs_type);
1522 	if (ret < 0) {
1523 		unregister_virtio_driver(&virtio_fs_driver);
1524 		return ret;
1525 	}
1526 
1527 	return 0;
1528 }
1529 module_init(virtio_fs_init);
1530 
1531 static void __exit virtio_fs_exit(void)
1532 {
1533 	unregister_filesystem(&virtio_fs_type);
1534 	unregister_virtio_driver(&virtio_fs_driver);
1535 }
1536 module_exit(virtio_fs_exit);
1537 
1538 MODULE_AUTHOR("Stefan Hajnoczi <stefanha@redhat.com>");
1539 MODULE_DESCRIPTION("Virtio Filesystem");
1540 MODULE_LICENSE("GPL");
1541 MODULE_ALIAS_FS(KBUILD_MODNAME);
1542 MODULE_DEVICE_TABLE(virtio, id_table);
1543