xref: /openbmc/linux/drivers/xen/gntdev.c (revision 94cdda6b)
1 /******************************************************************************
2  * gntdev.c
3  *
4  * Device for accessing (in user-space) pages that have been granted by other
5  * domains.
6  *
7  * Copyright (c) 2006-2007, D G Murray.
8  *           (c) 2009 Gerd Hoffmann <kraxel@redhat.com>
9  *
10  * This program is distributed in the hope that it will be useful,
11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
13  * GNU General Public License for more details.
14  *
15  * You should have received a copy of the GNU General Public License
16  * along with this program; if not, write to the Free Software
17  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
18  */
19 
20 #undef DEBUG
21 
22 #define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt
23 
24 #include <linux/module.h>
25 #include <linux/kernel.h>
26 #include <linux/init.h>
27 #include <linux/miscdevice.h>
28 #include <linux/fs.h>
29 #include <linux/mm.h>
30 #include <linux/mman.h>
31 #include <linux/mmu_notifier.h>
32 #include <linux/types.h>
33 #include <linux/uaccess.h>
34 #include <linux/sched.h>
35 #include <linux/spinlock.h>
36 #include <linux/slab.h>
37 #include <linux/highmem.h>
38 
39 #include <xen/xen.h>
40 #include <xen/grant_table.h>
41 #include <xen/balloon.h>
42 #include <xen/gntdev.h>
43 #include <xen/events.h>
44 #include <asm/xen/hypervisor.h>
45 #include <asm/xen/hypercall.h>
46 #include <asm/xen/page.h>
47 
48 MODULE_LICENSE("GPL");
49 MODULE_AUTHOR("Derek G. Murray <Derek.Murray@cl.cam.ac.uk>, "
50 	      "Gerd Hoffmann <kraxel@redhat.com>");
51 MODULE_DESCRIPTION("User-space granted page access driver");
52 
53 static int limit = 1024*1024;
54 module_param(limit, int, 0644);
55 MODULE_PARM_DESC(limit, "Maximum number of grants that may be mapped by "
56 		"the gntdev device");
57 
58 static atomic_t pages_mapped = ATOMIC_INIT(0);
59 
60 static int use_ptemod;
61 #define populate_freeable_maps use_ptemod
62 
63 struct gntdev_priv {
64 	/* maps with visible offsets in the file descriptor */
65 	struct list_head maps;
66 	/* maps that are not visible; will be freed on munmap.
67 	 * Only populated if populate_freeable_maps == 1 */
68 	struct list_head freeable_maps;
69 	/* lock protects maps and freeable_maps */
70 	struct mutex lock;
71 	struct mm_struct *mm;
72 	struct mmu_notifier mn;
73 };
74 
75 struct unmap_notify {
76 	int flags;
77 	/* Address relative to the start of the grant_map */
78 	int addr;
79 	int event;
80 };
81 
82 struct grant_map {
83 	struct list_head next;
84 	struct vm_area_struct *vma;
85 	int index;
86 	int count;
87 	int flags;
88 	atomic_t users;
89 	struct unmap_notify notify;
90 	struct ioctl_gntdev_grant_ref *grants;
91 	struct gnttab_map_grant_ref   *map_ops;
92 	struct gnttab_unmap_grant_ref *unmap_ops;
93 	struct gnttab_map_grant_ref   *kmap_ops;
94 	struct gnttab_unmap_grant_ref *kunmap_ops;
95 	struct page **pages;
96 	unsigned long pages_vm_start;
97 };
98 
99 static int unmap_grant_pages(struct grant_map *map, int offset, int pages);
100 
101 /* ------------------------------------------------------------------ */
102 
103 static void gntdev_print_maps(struct gntdev_priv *priv,
104 			      char *text, int text_index)
105 {
106 #ifdef DEBUG
107 	struct grant_map *map;
108 
109 	pr_debug("%s: maps list (priv %p)\n", __func__, priv);
110 	list_for_each_entry(map, &priv->maps, next)
111 		pr_debug("  index %2d, count %2d %s\n",
112 		       map->index, map->count,
113 		       map->index == text_index && text ? text : "");
114 #endif
115 }
116 
117 static void gntdev_free_map(struct grant_map *map)
118 {
119 	if (map == NULL)
120 		return;
121 
122 	if (map->pages)
123 		gnttab_free_pages(map->count, map->pages);
124 	kfree(map->pages);
125 	kfree(map->grants);
126 	kfree(map->map_ops);
127 	kfree(map->unmap_ops);
128 	kfree(map->kmap_ops);
129 	kfree(map->kunmap_ops);
130 	kfree(map);
131 }
132 
133 static struct grant_map *gntdev_alloc_map(struct gntdev_priv *priv, int count)
134 {
135 	struct grant_map *add;
136 	int i;
137 
138 	add = kzalloc(sizeof(struct grant_map), GFP_KERNEL);
139 	if (NULL == add)
140 		return NULL;
141 
142 	add->grants    = kcalloc(count, sizeof(add->grants[0]), GFP_KERNEL);
143 	add->map_ops   = kcalloc(count, sizeof(add->map_ops[0]), GFP_KERNEL);
144 	add->unmap_ops = kcalloc(count, sizeof(add->unmap_ops[0]), GFP_KERNEL);
145 	add->kmap_ops  = kcalloc(count, sizeof(add->kmap_ops[0]), GFP_KERNEL);
146 	add->kunmap_ops = kcalloc(count, sizeof(add->kunmap_ops[0]), GFP_KERNEL);
147 	add->pages     = kcalloc(count, sizeof(add->pages[0]), GFP_KERNEL);
148 	if (NULL == add->grants    ||
149 	    NULL == add->map_ops   ||
150 	    NULL == add->unmap_ops ||
151 	    NULL == add->kmap_ops  ||
152 	    NULL == add->kunmap_ops ||
153 	    NULL == add->pages)
154 		goto err;
155 
156 	if (gnttab_alloc_pages(count, add->pages))
157 		goto err;
158 
159 	for (i = 0; i < count; i++) {
160 		add->map_ops[i].handle = -1;
161 		add->unmap_ops[i].handle = -1;
162 		add->kmap_ops[i].handle = -1;
163 		add->kunmap_ops[i].handle = -1;
164 	}
165 
166 	add->index = 0;
167 	add->count = count;
168 	atomic_set(&add->users, 1);
169 
170 	return add;
171 
172 err:
173 	gntdev_free_map(add);
174 	return NULL;
175 }
176 
177 static void gntdev_add_map(struct gntdev_priv *priv, struct grant_map *add)
178 {
179 	struct grant_map *map;
180 
181 	list_for_each_entry(map, &priv->maps, next) {
182 		if (add->index + add->count < map->index) {
183 			list_add_tail(&add->next, &map->next);
184 			goto done;
185 		}
186 		add->index = map->index + map->count;
187 	}
188 	list_add_tail(&add->next, &priv->maps);
189 
190 done:
191 	gntdev_print_maps(priv, "[new]", add->index);
192 }
193 
194 static struct grant_map *gntdev_find_map_index(struct gntdev_priv *priv,
195 		int index, int count)
196 {
197 	struct grant_map *map;
198 
199 	list_for_each_entry(map, &priv->maps, next) {
200 		if (map->index != index)
201 			continue;
202 		if (count && map->count != count)
203 			continue;
204 		return map;
205 	}
206 	return NULL;
207 }
208 
209 static void gntdev_put_map(struct gntdev_priv *priv, struct grant_map *map)
210 {
211 	if (!map)
212 		return;
213 
214 	if (!atomic_dec_and_test(&map->users))
215 		return;
216 
217 	atomic_sub(map->count, &pages_mapped);
218 
219 	if (map->notify.flags & UNMAP_NOTIFY_SEND_EVENT) {
220 		notify_remote_via_evtchn(map->notify.event);
221 		evtchn_put(map->notify.event);
222 	}
223 
224 	if (populate_freeable_maps && priv) {
225 		mutex_lock(&priv->lock);
226 		list_del(&map->next);
227 		mutex_unlock(&priv->lock);
228 	}
229 
230 	if (map->pages && !use_ptemod)
231 		unmap_grant_pages(map, 0, map->count);
232 	gntdev_free_map(map);
233 }
234 
235 /* ------------------------------------------------------------------ */
236 
237 static int find_grant_ptes(pte_t *pte, pgtable_t token,
238 		unsigned long addr, void *data)
239 {
240 	struct grant_map *map = data;
241 	unsigned int pgnr = (addr - map->vma->vm_start) >> PAGE_SHIFT;
242 	int flags = map->flags | GNTMAP_application_map | GNTMAP_contains_pte;
243 	u64 pte_maddr;
244 
245 	BUG_ON(pgnr >= map->count);
246 	pte_maddr = arbitrary_virt_to_machine(pte).maddr;
247 
248 	/*
249 	 * Set the PTE as special to force get_user_pages_fast() fall
250 	 * back to the slow path.  If this is not supported as part of
251 	 * the grant map, it will be done afterwards.
252 	 */
253 	if (xen_feature(XENFEAT_gnttab_map_avail_bits))
254 		flags |= (1 << _GNTMAP_guest_avail0);
255 
256 	gnttab_set_map_op(&map->map_ops[pgnr], pte_maddr, flags,
257 			  map->grants[pgnr].ref,
258 			  map->grants[pgnr].domid);
259 	gnttab_set_unmap_op(&map->unmap_ops[pgnr], pte_maddr, flags,
260 			    -1 /* handle */);
261 	return 0;
262 }
263 
264 #ifdef CONFIG_X86
265 static int set_grant_ptes_as_special(pte_t *pte, pgtable_t token,
266 				     unsigned long addr, void *data)
267 {
268 	set_pte_at(current->mm, addr, pte, pte_mkspecial(*pte));
269 	return 0;
270 }
271 #endif
272 
273 static int map_grant_pages(struct grant_map *map)
274 {
275 	int i, err = 0;
276 
277 	if (!use_ptemod) {
278 		/* Note: it could already be mapped */
279 		if (map->map_ops[0].handle != -1)
280 			return 0;
281 		for (i = 0; i < map->count; i++) {
282 			unsigned long addr = (unsigned long)
283 				pfn_to_kaddr(page_to_pfn(map->pages[i]));
284 			gnttab_set_map_op(&map->map_ops[i], addr, map->flags,
285 				map->grants[i].ref,
286 				map->grants[i].domid);
287 			gnttab_set_unmap_op(&map->unmap_ops[i], addr,
288 				map->flags, -1 /* handle */);
289 		}
290 	} else {
291 		/*
292 		 * Setup the map_ops corresponding to the pte entries pointing
293 		 * to the kernel linear addresses of the struct pages.
294 		 * These ptes are completely different from the user ptes dealt
295 		 * with find_grant_ptes.
296 		 */
297 		for (i = 0; i < map->count; i++) {
298 			unsigned long address = (unsigned long)
299 				pfn_to_kaddr(page_to_pfn(map->pages[i]));
300 			BUG_ON(PageHighMem(map->pages[i]));
301 
302 			gnttab_set_map_op(&map->kmap_ops[i], address,
303 				map->flags | GNTMAP_host_map,
304 				map->grants[i].ref,
305 				map->grants[i].domid);
306 			gnttab_set_unmap_op(&map->kunmap_ops[i], address,
307 				map->flags | GNTMAP_host_map, -1);
308 		}
309 	}
310 
311 	pr_debug("map %d+%d\n", map->index, map->count);
312 	err = gnttab_map_refs(map->map_ops, use_ptemod ? map->kmap_ops : NULL,
313 			map->pages, map->count);
314 	if (err)
315 		return err;
316 
317 	for (i = 0; i < map->count; i++) {
318 		if (map->map_ops[i].status) {
319 			err = -EINVAL;
320 			continue;
321 		}
322 
323 		map->unmap_ops[i].handle = map->map_ops[i].handle;
324 		if (use_ptemod)
325 			map->kunmap_ops[i].handle = map->kmap_ops[i].handle;
326 	}
327 	return err;
328 }
329 
330 struct unmap_grant_pages_callback_data
331 {
332 	struct completion completion;
333 	int result;
334 };
335 
336 static void unmap_grant_callback(int result,
337 				 struct gntab_unmap_queue_data *data)
338 {
339 	struct unmap_grant_pages_callback_data* d = data->data;
340 
341 	d->result = result;
342 	complete(&d->completion);
343 }
344 
345 static int __unmap_grant_pages(struct grant_map *map, int offset, int pages)
346 {
347 	int i, err = 0;
348 	struct gntab_unmap_queue_data unmap_data;
349 	struct unmap_grant_pages_callback_data data;
350 
351 	init_completion(&data.completion);
352 	unmap_data.data = &data;
353 	unmap_data.done= &unmap_grant_callback;
354 
355 	if (map->notify.flags & UNMAP_NOTIFY_CLEAR_BYTE) {
356 		int pgno = (map->notify.addr >> PAGE_SHIFT);
357 		if (pgno >= offset && pgno < offset + pages) {
358 			/* No need for kmap, pages are in lowmem */
359 			uint8_t *tmp = pfn_to_kaddr(page_to_pfn(map->pages[pgno]));
360 			tmp[map->notify.addr & (PAGE_SIZE-1)] = 0;
361 			map->notify.flags &= ~UNMAP_NOTIFY_CLEAR_BYTE;
362 		}
363 	}
364 
365 	unmap_data.unmap_ops = map->unmap_ops + offset;
366 	unmap_data.kunmap_ops = use_ptemod ? map->kunmap_ops + offset : NULL;
367 	unmap_data.pages = map->pages + offset;
368 	unmap_data.count = pages;
369 
370 	gnttab_unmap_refs_async(&unmap_data);
371 
372 	wait_for_completion(&data.completion);
373 	if (data.result)
374 		return data.result;
375 
376 	for (i = 0; i < pages; i++) {
377 		if (map->unmap_ops[offset+i].status)
378 			err = -EINVAL;
379 		pr_debug("unmap handle=%d st=%d\n",
380 			map->unmap_ops[offset+i].handle,
381 			map->unmap_ops[offset+i].status);
382 		map->unmap_ops[offset+i].handle = -1;
383 	}
384 	return err;
385 }
386 
387 static int unmap_grant_pages(struct grant_map *map, int offset, int pages)
388 {
389 	int range, err = 0;
390 
391 	pr_debug("unmap %d+%d [%d+%d]\n", map->index, map->count, offset, pages);
392 
393 	/* It is possible the requested range will have a "hole" where we
394 	 * already unmapped some of the grants. Only unmap valid ranges.
395 	 */
396 	while (pages && !err) {
397 		while (pages && map->unmap_ops[offset].handle == -1) {
398 			offset++;
399 			pages--;
400 		}
401 		range = 0;
402 		while (range < pages) {
403 			if (map->unmap_ops[offset+range].handle == -1) {
404 				range--;
405 				break;
406 			}
407 			range++;
408 		}
409 		err = __unmap_grant_pages(map, offset, range);
410 		offset += range;
411 		pages -= range;
412 	}
413 
414 	return err;
415 }
416 
417 /* ------------------------------------------------------------------ */
418 
419 static void gntdev_vma_open(struct vm_area_struct *vma)
420 {
421 	struct grant_map *map = vma->vm_private_data;
422 
423 	pr_debug("gntdev_vma_open %p\n", vma);
424 	atomic_inc(&map->users);
425 }
426 
427 static void gntdev_vma_close(struct vm_area_struct *vma)
428 {
429 	struct grant_map *map = vma->vm_private_data;
430 	struct file *file = vma->vm_file;
431 	struct gntdev_priv *priv = file->private_data;
432 
433 	pr_debug("gntdev_vma_close %p\n", vma);
434 	if (use_ptemod) {
435 		/* It is possible that an mmu notifier could be running
436 		 * concurrently, so take priv->lock to ensure that the vma won't
437 		 * vanishing during the unmap_grant_pages call, since we will
438 		 * spin here until that completes. Such a concurrent call will
439 		 * not do any unmapping, since that has been done prior to
440 		 * closing the vma, but it may still iterate the unmap_ops list.
441 		 */
442 		mutex_lock(&priv->lock);
443 		map->vma = NULL;
444 		mutex_unlock(&priv->lock);
445 	}
446 	vma->vm_private_data = NULL;
447 	gntdev_put_map(priv, map);
448 }
449 
450 static struct page *gntdev_vma_find_special_page(struct vm_area_struct *vma,
451 						 unsigned long addr)
452 {
453 	struct grant_map *map = vma->vm_private_data;
454 
455 	return map->pages[(addr - map->pages_vm_start) >> PAGE_SHIFT];
456 }
457 
458 static struct vm_operations_struct gntdev_vmops = {
459 	.open = gntdev_vma_open,
460 	.close = gntdev_vma_close,
461 	.find_special_page = gntdev_vma_find_special_page,
462 };
463 
464 /* ------------------------------------------------------------------ */
465 
466 static void unmap_if_in_range(struct grant_map *map,
467 			      unsigned long start, unsigned long end)
468 {
469 	unsigned long mstart, mend;
470 	int err;
471 
472 	if (!map->vma)
473 		return;
474 	if (map->vma->vm_start >= end)
475 		return;
476 	if (map->vma->vm_end <= start)
477 		return;
478 	mstart = max(start, map->vma->vm_start);
479 	mend   = min(end,   map->vma->vm_end);
480 	pr_debug("map %d+%d (%lx %lx), range %lx %lx, mrange %lx %lx\n",
481 			map->index, map->count,
482 			map->vma->vm_start, map->vma->vm_end,
483 			start, end, mstart, mend);
484 	err = unmap_grant_pages(map,
485 				(mstart - map->vma->vm_start) >> PAGE_SHIFT,
486 				(mend - mstart) >> PAGE_SHIFT);
487 	WARN_ON(err);
488 }
489 
490 static void mn_invl_range_start(struct mmu_notifier *mn,
491 				struct mm_struct *mm,
492 				unsigned long start, unsigned long end)
493 {
494 	struct gntdev_priv *priv = container_of(mn, struct gntdev_priv, mn);
495 	struct grant_map *map;
496 
497 	mutex_lock(&priv->lock);
498 	list_for_each_entry(map, &priv->maps, next) {
499 		unmap_if_in_range(map, start, end);
500 	}
501 	list_for_each_entry(map, &priv->freeable_maps, next) {
502 		unmap_if_in_range(map, start, end);
503 	}
504 	mutex_unlock(&priv->lock);
505 }
506 
507 static void mn_invl_page(struct mmu_notifier *mn,
508 			 struct mm_struct *mm,
509 			 unsigned long address)
510 {
511 	mn_invl_range_start(mn, mm, address, address + PAGE_SIZE);
512 }
513 
514 static void mn_release(struct mmu_notifier *mn,
515 		       struct mm_struct *mm)
516 {
517 	struct gntdev_priv *priv = container_of(mn, struct gntdev_priv, mn);
518 	struct grant_map *map;
519 	int err;
520 
521 	mutex_lock(&priv->lock);
522 	list_for_each_entry(map, &priv->maps, next) {
523 		if (!map->vma)
524 			continue;
525 		pr_debug("map %d+%d (%lx %lx)\n",
526 				map->index, map->count,
527 				map->vma->vm_start, map->vma->vm_end);
528 		err = unmap_grant_pages(map, /* offset */ 0, map->count);
529 		WARN_ON(err);
530 	}
531 	list_for_each_entry(map, &priv->freeable_maps, next) {
532 		if (!map->vma)
533 			continue;
534 		pr_debug("map %d+%d (%lx %lx)\n",
535 				map->index, map->count,
536 				map->vma->vm_start, map->vma->vm_end);
537 		err = unmap_grant_pages(map, /* offset */ 0, map->count);
538 		WARN_ON(err);
539 	}
540 	mutex_unlock(&priv->lock);
541 }
542 
543 static struct mmu_notifier_ops gntdev_mmu_ops = {
544 	.release                = mn_release,
545 	.invalidate_page        = mn_invl_page,
546 	.invalidate_range_start = mn_invl_range_start,
547 };
548 
549 /* ------------------------------------------------------------------ */
550 
551 static int gntdev_open(struct inode *inode, struct file *flip)
552 {
553 	struct gntdev_priv *priv;
554 	int ret = 0;
555 
556 	priv = kzalloc(sizeof(*priv), GFP_KERNEL);
557 	if (!priv)
558 		return -ENOMEM;
559 
560 	INIT_LIST_HEAD(&priv->maps);
561 	INIT_LIST_HEAD(&priv->freeable_maps);
562 	mutex_init(&priv->lock);
563 
564 	if (use_ptemod) {
565 		priv->mm = get_task_mm(current);
566 		if (!priv->mm) {
567 			kfree(priv);
568 			return -ENOMEM;
569 		}
570 		priv->mn.ops = &gntdev_mmu_ops;
571 		ret = mmu_notifier_register(&priv->mn, priv->mm);
572 		mmput(priv->mm);
573 	}
574 
575 	if (ret) {
576 		kfree(priv);
577 		return ret;
578 	}
579 
580 	flip->private_data = priv;
581 	pr_debug("priv %p\n", priv);
582 
583 	return 0;
584 }
585 
586 static int gntdev_release(struct inode *inode, struct file *flip)
587 {
588 	struct gntdev_priv *priv = flip->private_data;
589 	struct grant_map *map;
590 
591 	pr_debug("priv %p\n", priv);
592 
593 	while (!list_empty(&priv->maps)) {
594 		map = list_entry(priv->maps.next, struct grant_map, next);
595 		list_del(&map->next);
596 		gntdev_put_map(NULL /* already removed */, map);
597 	}
598 	WARN_ON(!list_empty(&priv->freeable_maps));
599 
600 	if (use_ptemod)
601 		mmu_notifier_unregister(&priv->mn, priv->mm);
602 	kfree(priv);
603 	return 0;
604 }
605 
606 static long gntdev_ioctl_map_grant_ref(struct gntdev_priv *priv,
607 				       struct ioctl_gntdev_map_grant_ref __user *u)
608 {
609 	struct ioctl_gntdev_map_grant_ref op;
610 	struct grant_map *map;
611 	int err;
612 
613 	if (copy_from_user(&op, u, sizeof(op)) != 0)
614 		return -EFAULT;
615 	pr_debug("priv %p, add %d\n", priv, op.count);
616 	if (unlikely(op.count <= 0))
617 		return -EINVAL;
618 
619 	err = -ENOMEM;
620 	map = gntdev_alloc_map(priv, op.count);
621 	if (!map)
622 		return err;
623 
624 	if (unlikely(atomic_add_return(op.count, &pages_mapped) > limit)) {
625 		pr_debug("can't map: over limit\n");
626 		gntdev_put_map(NULL, map);
627 		return err;
628 	}
629 
630 	if (copy_from_user(map->grants, &u->refs,
631 			   sizeof(map->grants[0]) * op.count) != 0) {
632 		gntdev_put_map(NULL, map);
633 		return -EFAULT;
634 	}
635 
636 	mutex_lock(&priv->lock);
637 	gntdev_add_map(priv, map);
638 	op.index = map->index << PAGE_SHIFT;
639 	mutex_unlock(&priv->lock);
640 
641 	if (copy_to_user(u, &op, sizeof(op)) != 0)
642 		return -EFAULT;
643 
644 	return 0;
645 }
646 
647 static long gntdev_ioctl_unmap_grant_ref(struct gntdev_priv *priv,
648 					 struct ioctl_gntdev_unmap_grant_ref __user *u)
649 {
650 	struct ioctl_gntdev_unmap_grant_ref op;
651 	struct grant_map *map;
652 	int err = -ENOENT;
653 
654 	if (copy_from_user(&op, u, sizeof(op)) != 0)
655 		return -EFAULT;
656 	pr_debug("priv %p, del %d+%d\n", priv, (int)op.index, (int)op.count);
657 
658 	mutex_lock(&priv->lock);
659 	map = gntdev_find_map_index(priv, op.index >> PAGE_SHIFT, op.count);
660 	if (map) {
661 		list_del(&map->next);
662 		if (populate_freeable_maps)
663 			list_add_tail(&map->next, &priv->freeable_maps);
664 		err = 0;
665 	}
666 	mutex_unlock(&priv->lock);
667 	if (map)
668 		gntdev_put_map(priv, map);
669 	return err;
670 }
671 
672 static long gntdev_ioctl_get_offset_for_vaddr(struct gntdev_priv *priv,
673 					      struct ioctl_gntdev_get_offset_for_vaddr __user *u)
674 {
675 	struct ioctl_gntdev_get_offset_for_vaddr op;
676 	struct vm_area_struct *vma;
677 	struct grant_map *map;
678 	int rv = -EINVAL;
679 
680 	if (copy_from_user(&op, u, sizeof(op)) != 0)
681 		return -EFAULT;
682 	pr_debug("priv %p, offset for vaddr %lx\n", priv, (unsigned long)op.vaddr);
683 
684 	down_read(&current->mm->mmap_sem);
685 	vma = find_vma(current->mm, op.vaddr);
686 	if (!vma || vma->vm_ops != &gntdev_vmops)
687 		goto out_unlock;
688 
689 	map = vma->vm_private_data;
690 	if (!map)
691 		goto out_unlock;
692 
693 	op.offset = map->index << PAGE_SHIFT;
694 	op.count = map->count;
695 	rv = 0;
696 
697  out_unlock:
698 	up_read(&current->mm->mmap_sem);
699 
700 	if (rv == 0 && copy_to_user(u, &op, sizeof(op)) != 0)
701 		return -EFAULT;
702 	return rv;
703 }
704 
705 static long gntdev_ioctl_notify(struct gntdev_priv *priv, void __user *u)
706 {
707 	struct ioctl_gntdev_unmap_notify op;
708 	struct grant_map *map;
709 	int rc;
710 	int out_flags;
711 	unsigned int out_event;
712 
713 	if (copy_from_user(&op, u, sizeof(op)))
714 		return -EFAULT;
715 
716 	if (op.action & ~(UNMAP_NOTIFY_CLEAR_BYTE|UNMAP_NOTIFY_SEND_EVENT))
717 		return -EINVAL;
718 
719 	/* We need to grab a reference to the event channel we are going to use
720 	 * to send the notify before releasing the reference we may already have
721 	 * (if someone has called this ioctl twice). This is required so that
722 	 * it is possible to change the clear_byte part of the notification
723 	 * without disturbing the event channel part, which may now be the last
724 	 * reference to that event channel.
725 	 */
726 	if (op.action & UNMAP_NOTIFY_SEND_EVENT) {
727 		if (evtchn_get(op.event_channel_port))
728 			return -EINVAL;
729 	}
730 
731 	out_flags = op.action;
732 	out_event = op.event_channel_port;
733 
734 	mutex_lock(&priv->lock);
735 
736 	list_for_each_entry(map, &priv->maps, next) {
737 		uint64_t begin = map->index << PAGE_SHIFT;
738 		uint64_t end = (map->index + map->count) << PAGE_SHIFT;
739 		if (op.index >= begin && op.index < end)
740 			goto found;
741 	}
742 	rc = -ENOENT;
743 	goto unlock_out;
744 
745  found:
746 	if ((op.action & UNMAP_NOTIFY_CLEAR_BYTE) &&
747 			(map->flags & GNTMAP_readonly)) {
748 		rc = -EINVAL;
749 		goto unlock_out;
750 	}
751 
752 	out_flags = map->notify.flags;
753 	out_event = map->notify.event;
754 
755 	map->notify.flags = op.action;
756 	map->notify.addr = op.index - (map->index << PAGE_SHIFT);
757 	map->notify.event = op.event_channel_port;
758 
759 	rc = 0;
760 
761  unlock_out:
762 	mutex_unlock(&priv->lock);
763 
764 	/* Drop the reference to the event channel we did not save in the map */
765 	if (out_flags & UNMAP_NOTIFY_SEND_EVENT)
766 		evtchn_put(out_event);
767 
768 	return rc;
769 }
770 
771 static long gntdev_ioctl(struct file *flip,
772 			 unsigned int cmd, unsigned long arg)
773 {
774 	struct gntdev_priv *priv = flip->private_data;
775 	void __user *ptr = (void __user *)arg;
776 
777 	switch (cmd) {
778 	case IOCTL_GNTDEV_MAP_GRANT_REF:
779 		return gntdev_ioctl_map_grant_ref(priv, ptr);
780 
781 	case IOCTL_GNTDEV_UNMAP_GRANT_REF:
782 		return gntdev_ioctl_unmap_grant_ref(priv, ptr);
783 
784 	case IOCTL_GNTDEV_GET_OFFSET_FOR_VADDR:
785 		return gntdev_ioctl_get_offset_for_vaddr(priv, ptr);
786 
787 	case IOCTL_GNTDEV_SET_UNMAP_NOTIFY:
788 		return gntdev_ioctl_notify(priv, ptr);
789 
790 	default:
791 		pr_debug("priv %p, unknown cmd %x\n", priv, cmd);
792 		return -ENOIOCTLCMD;
793 	}
794 
795 	return 0;
796 }
797 
798 static int gntdev_mmap(struct file *flip, struct vm_area_struct *vma)
799 {
800 	struct gntdev_priv *priv = flip->private_data;
801 	int index = vma->vm_pgoff;
802 	int count = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
803 	struct grant_map *map;
804 	int i, err = -EINVAL;
805 
806 	if ((vma->vm_flags & VM_WRITE) && !(vma->vm_flags & VM_SHARED))
807 		return -EINVAL;
808 
809 	pr_debug("map %d+%d at %lx (pgoff %lx)\n",
810 			index, count, vma->vm_start, vma->vm_pgoff);
811 
812 	mutex_lock(&priv->lock);
813 	map = gntdev_find_map_index(priv, index, count);
814 	if (!map)
815 		goto unlock_out;
816 	if (use_ptemod && map->vma)
817 		goto unlock_out;
818 	if (use_ptemod && priv->mm != vma->vm_mm) {
819 		pr_warn("Huh? Other mm?\n");
820 		goto unlock_out;
821 	}
822 
823 	atomic_inc(&map->users);
824 
825 	vma->vm_ops = &gntdev_vmops;
826 
827 	vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
828 
829 	if (use_ptemod)
830 		vma->vm_flags |= VM_DONTCOPY;
831 
832 	vma->vm_private_data = map;
833 
834 	if (use_ptemod)
835 		map->vma = vma;
836 
837 	if (map->flags) {
838 		if ((vma->vm_flags & VM_WRITE) &&
839 				(map->flags & GNTMAP_readonly))
840 			goto out_unlock_put;
841 	} else {
842 		map->flags = GNTMAP_host_map;
843 		if (!(vma->vm_flags & VM_WRITE))
844 			map->flags |= GNTMAP_readonly;
845 	}
846 
847 	mutex_unlock(&priv->lock);
848 
849 	if (use_ptemod) {
850 		err = apply_to_page_range(vma->vm_mm, vma->vm_start,
851 					  vma->vm_end - vma->vm_start,
852 					  find_grant_ptes, map);
853 		if (err) {
854 			pr_warn("find_grant_ptes() failure.\n");
855 			goto out_put_map;
856 		}
857 	}
858 
859 	err = map_grant_pages(map);
860 	if (err)
861 		goto out_put_map;
862 
863 	if (!use_ptemod) {
864 		for (i = 0; i < count; i++) {
865 			err = vm_insert_page(vma, vma->vm_start + i*PAGE_SIZE,
866 				map->pages[i]);
867 			if (err)
868 				goto out_put_map;
869 		}
870 	} else {
871 #ifdef CONFIG_X86
872 		/*
873 		 * If the PTEs were not made special by the grant map
874 		 * hypercall, do so here.
875 		 *
876 		 * This is racy since the mapping is already visible
877 		 * to userspace but userspace should be well-behaved
878 		 * enough to not touch it until the mmap() call
879 		 * returns.
880 		 */
881 		if (!xen_feature(XENFEAT_gnttab_map_avail_bits)) {
882 			apply_to_page_range(vma->vm_mm, vma->vm_start,
883 					    vma->vm_end - vma->vm_start,
884 					    set_grant_ptes_as_special, NULL);
885 		}
886 #endif
887 		map->pages_vm_start = vma->vm_start;
888 	}
889 
890 	return 0;
891 
892 unlock_out:
893 	mutex_unlock(&priv->lock);
894 	return err;
895 
896 out_unlock_put:
897 	mutex_unlock(&priv->lock);
898 out_put_map:
899 	if (use_ptemod)
900 		map->vma = NULL;
901 	gntdev_put_map(priv, map);
902 	return err;
903 }
904 
905 static const struct file_operations gntdev_fops = {
906 	.owner = THIS_MODULE,
907 	.open = gntdev_open,
908 	.release = gntdev_release,
909 	.mmap = gntdev_mmap,
910 	.unlocked_ioctl = gntdev_ioctl
911 };
912 
913 static struct miscdevice gntdev_miscdev = {
914 	.minor        = MISC_DYNAMIC_MINOR,
915 	.name         = "xen/gntdev",
916 	.fops         = &gntdev_fops,
917 };
918 
919 /* ------------------------------------------------------------------ */
920 
921 static int __init gntdev_init(void)
922 {
923 	int err;
924 
925 	if (!xen_domain())
926 		return -ENODEV;
927 
928 	use_ptemod = !xen_feature(XENFEAT_auto_translated_physmap);
929 
930 	err = misc_register(&gntdev_miscdev);
931 	if (err != 0) {
932 		pr_err("Could not register gntdev device\n");
933 		return err;
934 	}
935 	return 0;
936 }
937 
938 static void __exit gntdev_exit(void)
939 {
940 	misc_deregister(&gntdev_miscdev);
941 }
942 
943 module_init(gntdev_init);
944 module_exit(gntdev_exit);
945 
946 /* ------------------------------------------------------------------ */
947