xref: /linux/drivers/xen/gntdev.c (revision a3a4a816b4b194c45d0217e8b9e08b2639802cda)
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/sched/mm.h>
36 #include <linux/spinlock.h>
37 #include <linux/slab.h>
38 #include <linux/highmem.h>
39 
40 #include <xen/xen.h>
41 #include <xen/grant_table.h>
42 #include <xen/balloon.h>
43 #include <xen/gntdev.h>
44 #include <xen/events.h>
45 #include <xen/page.h>
46 #include <asm/xen/hypervisor.h>
47 #include <asm/xen/hypercall.h>
48 
49 MODULE_LICENSE("GPL");
50 MODULE_AUTHOR("Derek G. Murray <Derek.Murray@cl.cam.ac.uk>, "
51 	      "Gerd Hoffmann <kraxel@redhat.com>");
52 MODULE_DESCRIPTION("User-space granted page access driver");
53 
54 static int limit = 1024*1024;
55 module_param(limit, int, 0644);
56 MODULE_PARM_DESC(limit, "Maximum number of grants that may be mapped by "
57 		"the gntdev device");
58 
59 static atomic_t pages_mapped = ATOMIC_INIT(0);
60 
61 static int use_ptemod;
62 #define populate_freeable_maps use_ptemod
63 
64 struct gntdev_priv {
65 	/* maps with visible offsets in the file descriptor */
66 	struct list_head maps;
67 	/* maps that are not visible; will be freed on munmap.
68 	 * Only populated if populate_freeable_maps == 1 */
69 	struct list_head freeable_maps;
70 	/* lock protects maps and freeable_maps */
71 	struct mutex lock;
72 	struct mm_struct *mm;
73 	struct mmu_notifier mn;
74 };
75 
76 struct unmap_notify {
77 	int flags;
78 	/* Address relative to the start of the grant_map */
79 	int addr;
80 	int event;
81 };
82 
83 struct grant_map {
84 	struct list_head next;
85 	struct vm_area_struct *vma;
86 	int index;
87 	int count;
88 	int flags;
89 	atomic_t users;
90 	struct unmap_notify notify;
91 	struct ioctl_gntdev_grant_ref *grants;
92 	struct gnttab_map_grant_ref   *map_ops;
93 	struct gnttab_unmap_grant_ref *unmap_ops;
94 	struct gnttab_map_grant_ref   *kmap_ops;
95 	struct gnttab_unmap_grant_ref *kunmap_ops;
96 	struct page **pages;
97 	unsigned long pages_vm_start;
98 };
99 
100 static int unmap_grant_pages(struct grant_map *map, int offset, int pages);
101 
102 /* ------------------------------------------------------------------ */
103 
104 static void gntdev_print_maps(struct gntdev_priv *priv,
105 			      char *text, int text_index)
106 {
107 #ifdef DEBUG
108 	struct grant_map *map;
109 
110 	pr_debug("%s: maps list (priv %p)\n", __func__, priv);
111 	list_for_each_entry(map, &priv->maps, next)
112 		pr_debug("  index %2d, count %2d %s\n",
113 		       map->index, map->count,
114 		       map->index == text_index && text ? text : "");
115 #endif
116 }
117 
118 static void gntdev_free_map(struct grant_map *map)
119 {
120 	if (map == NULL)
121 		return;
122 
123 	if (map->pages)
124 		gnttab_free_pages(map->count, map->pages);
125 	kfree(map->pages);
126 	kfree(map->grants);
127 	kfree(map->map_ops);
128 	kfree(map->unmap_ops);
129 	kfree(map->kmap_ops);
130 	kfree(map->kunmap_ops);
131 	kfree(map);
132 }
133 
134 static struct grant_map *gntdev_alloc_map(struct gntdev_priv *priv, int count)
135 {
136 	struct grant_map *add;
137 	int i;
138 
139 	add = kzalloc(sizeof(struct grant_map), GFP_KERNEL);
140 	if (NULL == add)
141 		return NULL;
142 
143 	add->grants    = kcalloc(count, sizeof(add->grants[0]), GFP_KERNEL);
144 	add->map_ops   = kcalloc(count, sizeof(add->map_ops[0]), GFP_KERNEL);
145 	add->unmap_ops = kcalloc(count, sizeof(add->unmap_ops[0]), GFP_KERNEL);
146 	add->kmap_ops  = kcalloc(count, sizeof(add->kmap_ops[0]), GFP_KERNEL);
147 	add->kunmap_ops = kcalloc(count, sizeof(add->kunmap_ops[0]), GFP_KERNEL);
148 	add->pages     = kcalloc(count, sizeof(add->pages[0]), GFP_KERNEL);
149 	if (NULL == add->grants    ||
150 	    NULL == add->map_ops   ||
151 	    NULL == add->unmap_ops ||
152 	    NULL == add->kmap_ops  ||
153 	    NULL == add->kunmap_ops ||
154 	    NULL == add->pages)
155 		goto err;
156 
157 	if (gnttab_alloc_pages(count, add->pages))
158 		goto err;
159 
160 	for (i = 0; i < count; i++) {
161 		add->map_ops[i].handle = -1;
162 		add->unmap_ops[i].handle = -1;
163 		add->kmap_ops[i].handle = -1;
164 		add->kunmap_ops[i].handle = -1;
165 	}
166 
167 	add->index = 0;
168 	add->count = count;
169 	atomic_set(&add->users, 1);
170 
171 	return add;
172 
173 err:
174 	gntdev_free_map(add);
175 	return NULL;
176 }
177 
178 static void gntdev_add_map(struct gntdev_priv *priv, struct grant_map *add)
179 {
180 	struct grant_map *map;
181 
182 	list_for_each_entry(map, &priv->maps, next) {
183 		if (add->index + add->count < map->index) {
184 			list_add_tail(&add->next, &map->next);
185 			goto done;
186 		}
187 		add->index = map->index + map->count;
188 	}
189 	list_add_tail(&add->next, &priv->maps);
190 
191 done:
192 	gntdev_print_maps(priv, "[new]", add->index);
193 }
194 
195 static struct grant_map *gntdev_find_map_index(struct gntdev_priv *priv,
196 		int index, int count)
197 {
198 	struct grant_map *map;
199 
200 	list_for_each_entry(map, &priv->maps, next) {
201 		if (map->index != index)
202 			continue;
203 		if (count && map->count != count)
204 			continue;
205 		return map;
206 	}
207 	return NULL;
208 }
209 
210 static void gntdev_put_map(struct gntdev_priv *priv, struct grant_map *map)
211 {
212 	if (!map)
213 		return;
214 
215 	if (!atomic_dec_and_test(&map->users))
216 		return;
217 
218 	atomic_sub(map->count, &pages_mapped);
219 
220 	if (map->notify.flags & UNMAP_NOTIFY_SEND_EVENT) {
221 		notify_remote_via_evtchn(map->notify.event);
222 		evtchn_put(map->notify.event);
223 	}
224 
225 	if (populate_freeable_maps && priv) {
226 		mutex_lock(&priv->lock);
227 		list_del(&map->next);
228 		mutex_unlock(&priv->lock);
229 	}
230 
231 	if (map->pages && !use_ptemod)
232 		unmap_grant_pages(map, 0, map->count);
233 	gntdev_free_map(map);
234 }
235 
236 /* ------------------------------------------------------------------ */
237 
238 static int find_grant_ptes(pte_t *pte, pgtable_t token,
239 		unsigned long addr, void *data)
240 {
241 	struct grant_map *map = data;
242 	unsigned int pgnr = (addr - map->vma->vm_start) >> PAGE_SHIFT;
243 	int flags = map->flags | GNTMAP_application_map | GNTMAP_contains_pte;
244 	u64 pte_maddr;
245 
246 	BUG_ON(pgnr >= map->count);
247 	pte_maddr = arbitrary_virt_to_machine(pte).maddr;
248 
249 	/*
250 	 * Set the PTE as special to force get_user_pages_fast() fall
251 	 * back to the slow path.  If this is not supported as part of
252 	 * the grant map, it will be done afterwards.
253 	 */
254 	if (xen_feature(XENFEAT_gnttab_map_avail_bits))
255 		flags |= (1 << _GNTMAP_guest_avail0);
256 
257 	gnttab_set_map_op(&map->map_ops[pgnr], pte_maddr, flags,
258 			  map->grants[pgnr].ref,
259 			  map->grants[pgnr].domid);
260 	gnttab_set_unmap_op(&map->unmap_ops[pgnr], pte_maddr, flags,
261 			    -1 /* handle */);
262 	return 0;
263 }
264 
265 #ifdef CONFIG_X86
266 static int set_grant_ptes_as_special(pte_t *pte, pgtable_t token,
267 				     unsigned long addr, void *data)
268 {
269 	set_pte_at(current->mm, addr, pte, pte_mkspecial(*pte));
270 	return 0;
271 }
272 #endif
273 
274 static int map_grant_pages(struct grant_map *map)
275 {
276 	int i, err = 0;
277 
278 	if (!use_ptemod) {
279 		/* Note: it could already be mapped */
280 		if (map->map_ops[0].handle != -1)
281 			return 0;
282 		for (i = 0; i < map->count; i++) {
283 			unsigned long addr = (unsigned long)
284 				pfn_to_kaddr(page_to_pfn(map->pages[i]));
285 			gnttab_set_map_op(&map->map_ops[i], addr, map->flags,
286 				map->grants[i].ref,
287 				map->grants[i].domid);
288 			gnttab_set_unmap_op(&map->unmap_ops[i], addr,
289 				map->flags, -1 /* handle */);
290 		}
291 	} else {
292 		/*
293 		 * Setup the map_ops corresponding to the pte entries pointing
294 		 * to the kernel linear addresses of the struct pages.
295 		 * These ptes are completely different from the user ptes dealt
296 		 * with find_grant_ptes.
297 		 */
298 		for (i = 0; i < map->count; i++) {
299 			unsigned long address = (unsigned long)
300 				pfn_to_kaddr(page_to_pfn(map->pages[i]));
301 			BUG_ON(PageHighMem(map->pages[i]));
302 
303 			gnttab_set_map_op(&map->kmap_ops[i], address,
304 				map->flags | GNTMAP_host_map,
305 				map->grants[i].ref,
306 				map->grants[i].domid);
307 			gnttab_set_unmap_op(&map->kunmap_ops[i], address,
308 				map->flags | GNTMAP_host_map, -1);
309 		}
310 	}
311 
312 	pr_debug("map %d+%d\n", map->index, map->count);
313 	err = gnttab_map_refs(map->map_ops, use_ptemod ? map->kmap_ops : NULL,
314 			map->pages, map->count);
315 	if (err)
316 		return err;
317 
318 	for (i = 0; i < map->count; i++) {
319 		if (map->map_ops[i].status) {
320 			err = -EINVAL;
321 			continue;
322 		}
323 
324 		map->unmap_ops[i].handle = map->map_ops[i].handle;
325 		if (use_ptemod)
326 			map->kunmap_ops[i].handle = map->kmap_ops[i].handle;
327 	}
328 	return err;
329 }
330 
331 static int __unmap_grant_pages(struct grant_map *map, int offset, int pages)
332 {
333 	int i, err = 0;
334 	struct gntab_unmap_queue_data unmap_data;
335 
336 	if (map->notify.flags & UNMAP_NOTIFY_CLEAR_BYTE) {
337 		int pgno = (map->notify.addr >> PAGE_SHIFT);
338 		if (pgno >= offset && pgno < offset + pages) {
339 			/* No need for kmap, pages are in lowmem */
340 			uint8_t *tmp = pfn_to_kaddr(page_to_pfn(map->pages[pgno]));
341 			tmp[map->notify.addr & (PAGE_SIZE-1)] = 0;
342 			map->notify.flags &= ~UNMAP_NOTIFY_CLEAR_BYTE;
343 		}
344 	}
345 
346 	unmap_data.unmap_ops = map->unmap_ops + offset;
347 	unmap_data.kunmap_ops = use_ptemod ? map->kunmap_ops + offset : NULL;
348 	unmap_data.pages = map->pages + offset;
349 	unmap_data.count = pages;
350 
351 	err = gnttab_unmap_refs_sync(&unmap_data);
352 	if (err)
353 		return err;
354 
355 	for (i = 0; i < pages; i++) {
356 		if (map->unmap_ops[offset+i].status)
357 			err = -EINVAL;
358 		pr_debug("unmap handle=%d st=%d\n",
359 			map->unmap_ops[offset+i].handle,
360 			map->unmap_ops[offset+i].status);
361 		map->unmap_ops[offset+i].handle = -1;
362 	}
363 	return err;
364 }
365 
366 static int unmap_grant_pages(struct grant_map *map, int offset, int pages)
367 {
368 	int range, err = 0;
369 
370 	pr_debug("unmap %d+%d [%d+%d]\n", map->index, map->count, offset, pages);
371 
372 	/* It is possible the requested range will have a "hole" where we
373 	 * already unmapped some of the grants. Only unmap valid ranges.
374 	 */
375 	while (pages && !err) {
376 		while (pages && map->unmap_ops[offset].handle == -1) {
377 			offset++;
378 			pages--;
379 		}
380 		range = 0;
381 		while (range < pages) {
382 			if (map->unmap_ops[offset+range].handle == -1) {
383 				range--;
384 				break;
385 			}
386 			range++;
387 		}
388 		err = __unmap_grant_pages(map, offset, range);
389 		offset += range;
390 		pages -= range;
391 	}
392 
393 	return err;
394 }
395 
396 /* ------------------------------------------------------------------ */
397 
398 static void gntdev_vma_open(struct vm_area_struct *vma)
399 {
400 	struct grant_map *map = vma->vm_private_data;
401 
402 	pr_debug("gntdev_vma_open %p\n", vma);
403 	atomic_inc(&map->users);
404 }
405 
406 static void gntdev_vma_close(struct vm_area_struct *vma)
407 {
408 	struct grant_map *map = vma->vm_private_data;
409 	struct file *file = vma->vm_file;
410 	struct gntdev_priv *priv = file->private_data;
411 
412 	pr_debug("gntdev_vma_close %p\n", vma);
413 	if (use_ptemod) {
414 		/* It is possible that an mmu notifier could be running
415 		 * concurrently, so take priv->lock to ensure that the vma won't
416 		 * vanishing during the unmap_grant_pages call, since we will
417 		 * spin here until that completes. Such a concurrent call will
418 		 * not do any unmapping, since that has been done prior to
419 		 * closing the vma, but it may still iterate the unmap_ops list.
420 		 */
421 		mutex_lock(&priv->lock);
422 		map->vma = NULL;
423 		mutex_unlock(&priv->lock);
424 	}
425 	vma->vm_private_data = NULL;
426 	gntdev_put_map(priv, map);
427 }
428 
429 static struct page *gntdev_vma_find_special_page(struct vm_area_struct *vma,
430 						 unsigned long addr)
431 {
432 	struct grant_map *map = vma->vm_private_data;
433 
434 	return map->pages[(addr - map->pages_vm_start) >> PAGE_SHIFT];
435 }
436 
437 static const struct vm_operations_struct gntdev_vmops = {
438 	.open = gntdev_vma_open,
439 	.close = gntdev_vma_close,
440 	.find_special_page = gntdev_vma_find_special_page,
441 };
442 
443 /* ------------------------------------------------------------------ */
444 
445 static void unmap_if_in_range(struct grant_map *map,
446 			      unsigned long start, unsigned long end)
447 {
448 	unsigned long mstart, mend;
449 	int err;
450 
451 	if (!map->vma)
452 		return;
453 	if (map->vma->vm_start >= end)
454 		return;
455 	if (map->vma->vm_end <= start)
456 		return;
457 	mstart = max(start, map->vma->vm_start);
458 	mend   = min(end,   map->vma->vm_end);
459 	pr_debug("map %d+%d (%lx %lx), range %lx %lx, mrange %lx %lx\n",
460 			map->index, map->count,
461 			map->vma->vm_start, map->vma->vm_end,
462 			start, end, mstart, mend);
463 	err = unmap_grant_pages(map,
464 				(mstart - map->vma->vm_start) >> PAGE_SHIFT,
465 				(mend - mstart) >> PAGE_SHIFT);
466 	WARN_ON(err);
467 }
468 
469 static void mn_invl_range_start(struct mmu_notifier *mn,
470 				struct mm_struct *mm,
471 				unsigned long start, unsigned long end)
472 {
473 	struct gntdev_priv *priv = container_of(mn, struct gntdev_priv, mn);
474 	struct grant_map *map;
475 
476 	mutex_lock(&priv->lock);
477 	list_for_each_entry(map, &priv->maps, next) {
478 		unmap_if_in_range(map, start, end);
479 	}
480 	list_for_each_entry(map, &priv->freeable_maps, next) {
481 		unmap_if_in_range(map, start, end);
482 	}
483 	mutex_unlock(&priv->lock);
484 }
485 
486 static void mn_invl_page(struct mmu_notifier *mn,
487 			 struct mm_struct *mm,
488 			 unsigned long address)
489 {
490 	mn_invl_range_start(mn, mm, address, address + PAGE_SIZE);
491 }
492 
493 static void mn_release(struct mmu_notifier *mn,
494 		       struct mm_struct *mm)
495 {
496 	struct gntdev_priv *priv = container_of(mn, struct gntdev_priv, mn);
497 	struct grant_map *map;
498 	int err;
499 
500 	mutex_lock(&priv->lock);
501 	list_for_each_entry(map, &priv->maps, next) {
502 		if (!map->vma)
503 			continue;
504 		pr_debug("map %d+%d (%lx %lx)\n",
505 				map->index, map->count,
506 				map->vma->vm_start, map->vma->vm_end);
507 		err = unmap_grant_pages(map, /* offset */ 0, map->count);
508 		WARN_ON(err);
509 	}
510 	list_for_each_entry(map, &priv->freeable_maps, next) {
511 		if (!map->vma)
512 			continue;
513 		pr_debug("map %d+%d (%lx %lx)\n",
514 				map->index, map->count,
515 				map->vma->vm_start, map->vma->vm_end);
516 		err = unmap_grant_pages(map, /* offset */ 0, map->count);
517 		WARN_ON(err);
518 	}
519 	mutex_unlock(&priv->lock);
520 }
521 
522 static const struct mmu_notifier_ops gntdev_mmu_ops = {
523 	.release                = mn_release,
524 	.invalidate_page        = mn_invl_page,
525 	.invalidate_range_start = mn_invl_range_start,
526 };
527 
528 /* ------------------------------------------------------------------ */
529 
530 static int gntdev_open(struct inode *inode, struct file *flip)
531 {
532 	struct gntdev_priv *priv;
533 	int ret = 0;
534 
535 	priv = kzalloc(sizeof(*priv), GFP_KERNEL);
536 	if (!priv)
537 		return -ENOMEM;
538 
539 	INIT_LIST_HEAD(&priv->maps);
540 	INIT_LIST_HEAD(&priv->freeable_maps);
541 	mutex_init(&priv->lock);
542 
543 	if (use_ptemod) {
544 		priv->mm = get_task_mm(current);
545 		if (!priv->mm) {
546 			kfree(priv);
547 			return -ENOMEM;
548 		}
549 		priv->mn.ops = &gntdev_mmu_ops;
550 		ret = mmu_notifier_register(&priv->mn, priv->mm);
551 		mmput(priv->mm);
552 	}
553 
554 	if (ret) {
555 		kfree(priv);
556 		return ret;
557 	}
558 
559 	flip->private_data = priv;
560 	pr_debug("priv %p\n", priv);
561 
562 	return 0;
563 }
564 
565 static int gntdev_release(struct inode *inode, struct file *flip)
566 {
567 	struct gntdev_priv *priv = flip->private_data;
568 	struct grant_map *map;
569 
570 	pr_debug("priv %p\n", priv);
571 
572 	mutex_lock(&priv->lock);
573 	while (!list_empty(&priv->maps)) {
574 		map = list_entry(priv->maps.next, struct grant_map, next);
575 		list_del(&map->next);
576 		gntdev_put_map(NULL /* already removed */, map);
577 	}
578 	WARN_ON(!list_empty(&priv->freeable_maps));
579 	mutex_unlock(&priv->lock);
580 
581 	if (use_ptemod)
582 		mmu_notifier_unregister(&priv->mn, priv->mm);
583 	kfree(priv);
584 	return 0;
585 }
586 
587 static long gntdev_ioctl_map_grant_ref(struct gntdev_priv *priv,
588 				       struct ioctl_gntdev_map_grant_ref __user *u)
589 {
590 	struct ioctl_gntdev_map_grant_ref op;
591 	struct grant_map *map;
592 	int err;
593 
594 	if (copy_from_user(&op, u, sizeof(op)) != 0)
595 		return -EFAULT;
596 	pr_debug("priv %p, add %d\n", priv, op.count);
597 	if (unlikely(op.count <= 0))
598 		return -EINVAL;
599 
600 	err = -ENOMEM;
601 	map = gntdev_alloc_map(priv, op.count);
602 	if (!map)
603 		return err;
604 
605 	if (unlikely(atomic_add_return(op.count, &pages_mapped) > limit)) {
606 		pr_debug("can't map: over limit\n");
607 		gntdev_put_map(NULL, map);
608 		return err;
609 	}
610 
611 	if (copy_from_user(map->grants, &u->refs,
612 			   sizeof(map->grants[0]) * op.count) != 0) {
613 		gntdev_put_map(NULL, map);
614 		return -EFAULT;
615 	}
616 
617 	mutex_lock(&priv->lock);
618 	gntdev_add_map(priv, map);
619 	op.index = map->index << PAGE_SHIFT;
620 	mutex_unlock(&priv->lock);
621 
622 	if (copy_to_user(u, &op, sizeof(op)) != 0)
623 		return -EFAULT;
624 
625 	return 0;
626 }
627 
628 static long gntdev_ioctl_unmap_grant_ref(struct gntdev_priv *priv,
629 					 struct ioctl_gntdev_unmap_grant_ref __user *u)
630 {
631 	struct ioctl_gntdev_unmap_grant_ref op;
632 	struct grant_map *map;
633 	int err = -ENOENT;
634 
635 	if (copy_from_user(&op, u, sizeof(op)) != 0)
636 		return -EFAULT;
637 	pr_debug("priv %p, del %d+%d\n", priv, (int)op.index, (int)op.count);
638 
639 	mutex_lock(&priv->lock);
640 	map = gntdev_find_map_index(priv, op.index >> PAGE_SHIFT, op.count);
641 	if (map) {
642 		list_del(&map->next);
643 		if (populate_freeable_maps)
644 			list_add_tail(&map->next, &priv->freeable_maps);
645 		err = 0;
646 	}
647 	mutex_unlock(&priv->lock);
648 	if (map)
649 		gntdev_put_map(priv, map);
650 	return err;
651 }
652 
653 static long gntdev_ioctl_get_offset_for_vaddr(struct gntdev_priv *priv,
654 					      struct ioctl_gntdev_get_offset_for_vaddr __user *u)
655 {
656 	struct ioctl_gntdev_get_offset_for_vaddr op;
657 	struct vm_area_struct *vma;
658 	struct grant_map *map;
659 	int rv = -EINVAL;
660 
661 	if (copy_from_user(&op, u, sizeof(op)) != 0)
662 		return -EFAULT;
663 	pr_debug("priv %p, offset for vaddr %lx\n", priv, (unsigned long)op.vaddr);
664 
665 	down_read(&current->mm->mmap_sem);
666 	vma = find_vma(current->mm, op.vaddr);
667 	if (!vma || vma->vm_ops != &gntdev_vmops)
668 		goto out_unlock;
669 
670 	map = vma->vm_private_data;
671 	if (!map)
672 		goto out_unlock;
673 
674 	op.offset = map->index << PAGE_SHIFT;
675 	op.count = map->count;
676 	rv = 0;
677 
678  out_unlock:
679 	up_read(&current->mm->mmap_sem);
680 
681 	if (rv == 0 && copy_to_user(u, &op, sizeof(op)) != 0)
682 		return -EFAULT;
683 	return rv;
684 }
685 
686 static long gntdev_ioctl_notify(struct gntdev_priv *priv, void __user *u)
687 {
688 	struct ioctl_gntdev_unmap_notify op;
689 	struct grant_map *map;
690 	int rc;
691 	int out_flags;
692 	unsigned int out_event;
693 
694 	if (copy_from_user(&op, u, sizeof(op)))
695 		return -EFAULT;
696 
697 	if (op.action & ~(UNMAP_NOTIFY_CLEAR_BYTE|UNMAP_NOTIFY_SEND_EVENT))
698 		return -EINVAL;
699 
700 	/* We need to grab a reference to the event channel we are going to use
701 	 * to send the notify before releasing the reference we may already have
702 	 * (if someone has called this ioctl twice). This is required so that
703 	 * it is possible to change the clear_byte part of the notification
704 	 * without disturbing the event channel part, which may now be the last
705 	 * reference to that event channel.
706 	 */
707 	if (op.action & UNMAP_NOTIFY_SEND_EVENT) {
708 		if (evtchn_get(op.event_channel_port))
709 			return -EINVAL;
710 	}
711 
712 	out_flags = op.action;
713 	out_event = op.event_channel_port;
714 
715 	mutex_lock(&priv->lock);
716 
717 	list_for_each_entry(map, &priv->maps, next) {
718 		uint64_t begin = map->index << PAGE_SHIFT;
719 		uint64_t end = (map->index + map->count) << PAGE_SHIFT;
720 		if (op.index >= begin && op.index < end)
721 			goto found;
722 	}
723 	rc = -ENOENT;
724 	goto unlock_out;
725 
726  found:
727 	if ((op.action & UNMAP_NOTIFY_CLEAR_BYTE) &&
728 			(map->flags & GNTMAP_readonly)) {
729 		rc = -EINVAL;
730 		goto unlock_out;
731 	}
732 
733 	out_flags = map->notify.flags;
734 	out_event = map->notify.event;
735 
736 	map->notify.flags = op.action;
737 	map->notify.addr = op.index - (map->index << PAGE_SHIFT);
738 	map->notify.event = op.event_channel_port;
739 
740 	rc = 0;
741 
742  unlock_out:
743 	mutex_unlock(&priv->lock);
744 
745 	/* Drop the reference to the event channel we did not save in the map */
746 	if (out_flags & UNMAP_NOTIFY_SEND_EVENT)
747 		evtchn_put(out_event);
748 
749 	return rc;
750 }
751 
752 #define GNTDEV_COPY_BATCH 16
753 
754 struct gntdev_copy_batch {
755 	struct gnttab_copy ops[GNTDEV_COPY_BATCH];
756 	struct page *pages[GNTDEV_COPY_BATCH];
757 	s16 __user *status[GNTDEV_COPY_BATCH];
758 	unsigned int nr_ops;
759 	unsigned int nr_pages;
760 };
761 
762 static int gntdev_get_page(struct gntdev_copy_batch *batch, void __user *virt,
763 			   bool writeable, unsigned long *gfn)
764 {
765 	unsigned long addr = (unsigned long)virt;
766 	struct page *page;
767 	unsigned long xen_pfn;
768 	int ret;
769 
770 	ret = get_user_pages_fast(addr, 1, writeable, &page);
771 	if (ret < 0)
772 		return ret;
773 
774 	batch->pages[batch->nr_pages++] = page;
775 
776 	xen_pfn = page_to_xen_pfn(page) + XEN_PFN_DOWN(addr & ~PAGE_MASK);
777 	*gfn = pfn_to_gfn(xen_pfn);
778 
779 	return 0;
780 }
781 
782 static void gntdev_put_pages(struct gntdev_copy_batch *batch)
783 {
784 	unsigned int i;
785 
786 	for (i = 0; i < batch->nr_pages; i++)
787 		put_page(batch->pages[i]);
788 	batch->nr_pages = 0;
789 }
790 
791 static int gntdev_copy(struct gntdev_copy_batch *batch)
792 {
793 	unsigned int i;
794 
795 	gnttab_batch_copy(batch->ops, batch->nr_ops);
796 	gntdev_put_pages(batch);
797 
798 	/*
799 	 * For each completed op, update the status if the op failed
800 	 * and all previous ops for the segment were successful.
801 	 */
802 	for (i = 0; i < batch->nr_ops; i++) {
803 		s16 status = batch->ops[i].status;
804 		s16 old_status;
805 
806 		if (status == GNTST_okay)
807 			continue;
808 
809 		if (__get_user(old_status, batch->status[i]))
810 			return -EFAULT;
811 
812 		if (old_status != GNTST_okay)
813 			continue;
814 
815 		if (__put_user(status, batch->status[i]))
816 			return -EFAULT;
817 	}
818 
819 	batch->nr_ops = 0;
820 	return 0;
821 }
822 
823 static int gntdev_grant_copy_seg(struct gntdev_copy_batch *batch,
824 				 struct gntdev_grant_copy_segment *seg,
825 				 s16 __user *status)
826 {
827 	uint16_t copied = 0;
828 
829 	/*
830 	 * Disallow local -> local copies since there is only space in
831 	 * batch->pages for one page per-op and this would be a very
832 	 * expensive memcpy().
833 	 */
834 	if (!(seg->flags & (GNTCOPY_source_gref | GNTCOPY_dest_gref)))
835 		return -EINVAL;
836 
837 	/* Can't cross page if source/dest is a grant ref. */
838 	if (seg->flags & GNTCOPY_source_gref) {
839 		if (seg->source.foreign.offset + seg->len > XEN_PAGE_SIZE)
840 			return -EINVAL;
841 	}
842 	if (seg->flags & GNTCOPY_dest_gref) {
843 		if (seg->dest.foreign.offset + seg->len > XEN_PAGE_SIZE)
844 			return -EINVAL;
845 	}
846 
847 	if (put_user(GNTST_okay, status))
848 		return -EFAULT;
849 
850 	while (copied < seg->len) {
851 		struct gnttab_copy *op;
852 		void __user *virt;
853 		size_t len, off;
854 		unsigned long gfn;
855 		int ret;
856 
857 		if (batch->nr_ops >= GNTDEV_COPY_BATCH) {
858 			ret = gntdev_copy(batch);
859 			if (ret < 0)
860 				return ret;
861 		}
862 
863 		len = seg->len - copied;
864 
865 		op = &batch->ops[batch->nr_ops];
866 		op->flags = 0;
867 
868 		if (seg->flags & GNTCOPY_source_gref) {
869 			op->source.u.ref = seg->source.foreign.ref;
870 			op->source.domid = seg->source.foreign.domid;
871 			op->source.offset = seg->source.foreign.offset + copied;
872 			op->flags |= GNTCOPY_source_gref;
873 		} else {
874 			virt = seg->source.virt + copied;
875 			off = (unsigned long)virt & ~XEN_PAGE_MASK;
876 			len = min(len, (size_t)XEN_PAGE_SIZE - off);
877 
878 			ret = gntdev_get_page(batch, virt, false, &gfn);
879 			if (ret < 0)
880 				return ret;
881 
882 			op->source.u.gmfn = gfn;
883 			op->source.domid = DOMID_SELF;
884 			op->source.offset = off;
885 		}
886 
887 		if (seg->flags & GNTCOPY_dest_gref) {
888 			op->dest.u.ref = seg->dest.foreign.ref;
889 			op->dest.domid = seg->dest.foreign.domid;
890 			op->dest.offset = seg->dest.foreign.offset + copied;
891 			op->flags |= GNTCOPY_dest_gref;
892 		} else {
893 			virt = seg->dest.virt + copied;
894 			off = (unsigned long)virt & ~XEN_PAGE_MASK;
895 			len = min(len, (size_t)XEN_PAGE_SIZE - off);
896 
897 			ret = gntdev_get_page(batch, virt, true, &gfn);
898 			if (ret < 0)
899 				return ret;
900 
901 			op->dest.u.gmfn = gfn;
902 			op->dest.domid = DOMID_SELF;
903 			op->dest.offset = off;
904 		}
905 
906 		op->len = len;
907 		copied += len;
908 
909 		batch->status[batch->nr_ops] = status;
910 		batch->nr_ops++;
911 	}
912 
913 	return 0;
914 }
915 
916 static long gntdev_ioctl_grant_copy(struct gntdev_priv *priv, void __user *u)
917 {
918 	struct ioctl_gntdev_grant_copy copy;
919 	struct gntdev_copy_batch batch;
920 	unsigned int i;
921 	int ret = 0;
922 
923 	if (copy_from_user(&copy, u, sizeof(copy)))
924 		return -EFAULT;
925 
926 	batch.nr_ops = 0;
927 	batch.nr_pages = 0;
928 
929 	for (i = 0; i < copy.count; i++) {
930 		struct gntdev_grant_copy_segment seg;
931 
932 		if (copy_from_user(&seg, &copy.segments[i], sizeof(seg))) {
933 			ret = -EFAULT;
934 			goto out;
935 		}
936 
937 		ret = gntdev_grant_copy_seg(&batch, &seg, &copy.segments[i].status);
938 		if (ret < 0)
939 			goto out;
940 
941 		cond_resched();
942 	}
943 	if (batch.nr_ops)
944 		ret = gntdev_copy(&batch);
945 	return ret;
946 
947   out:
948 	gntdev_put_pages(&batch);
949 	return ret;
950 }
951 
952 static long gntdev_ioctl(struct file *flip,
953 			 unsigned int cmd, unsigned long arg)
954 {
955 	struct gntdev_priv *priv = flip->private_data;
956 	void __user *ptr = (void __user *)arg;
957 
958 	switch (cmd) {
959 	case IOCTL_GNTDEV_MAP_GRANT_REF:
960 		return gntdev_ioctl_map_grant_ref(priv, ptr);
961 
962 	case IOCTL_GNTDEV_UNMAP_GRANT_REF:
963 		return gntdev_ioctl_unmap_grant_ref(priv, ptr);
964 
965 	case IOCTL_GNTDEV_GET_OFFSET_FOR_VADDR:
966 		return gntdev_ioctl_get_offset_for_vaddr(priv, ptr);
967 
968 	case IOCTL_GNTDEV_SET_UNMAP_NOTIFY:
969 		return gntdev_ioctl_notify(priv, ptr);
970 
971 	case IOCTL_GNTDEV_GRANT_COPY:
972 		return gntdev_ioctl_grant_copy(priv, ptr);
973 
974 	default:
975 		pr_debug("priv %p, unknown cmd %x\n", priv, cmd);
976 		return -ENOIOCTLCMD;
977 	}
978 
979 	return 0;
980 }
981 
982 static int gntdev_mmap(struct file *flip, struct vm_area_struct *vma)
983 {
984 	struct gntdev_priv *priv = flip->private_data;
985 	int index = vma->vm_pgoff;
986 	int count = vma_pages(vma);
987 	struct grant_map *map;
988 	int i, err = -EINVAL;
989 
990 	if ((vma->vm_flags & VM_WRITE) && !(vma->vm_flags & VM_SHARED))
991 		return -EINVAL;
992 
993 	pr_debug("map %d+%d at %lx (pgoff %lx)\n",
994 			index, count, vma->vm_start, vma->vm_pgoff);
995 
996 	mutex_lock(&priv->lock);
997 	map = gntdev_find_map_index(priv, index, count);
998 	if (!map)
999 		goto unlock_out;
1000 	if (use_ptemod && map->vma)
1001 		goto unlock_out;
1002 	if (use_ptemod && priv->mm != vma->vm_mm) {
1003 		pr_warn("Huh? Other mm?\n");
1004 		goto unlock_out;
1005 	}
1006 
1007 	atomic_inc(&map->users);
1008 
1009 	vma->vm_ops = &gntdev_vmops;
1010 
1011 	vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP | VM_MIXEDMAP;
1012 
1013 	if (use_ptemod)
1014 		vma->vm_flags |= VM_DONTCOPY;
1015 
1016 	vma->vm_private_data = map;
1017 
1018 	if (use_ptemod)
1019 		map->vma = vma;
1020 
1021 	if (map->flags) {
1022 		if ((vma->vm_flags & VM_WRITE) &&
1023 				(map->flags & GNTMAP_readonly))
1024 			goto out_unlock_put;
1025 	} else {
1026 		map->flags = GNTMAP_host_map;
1027 		if (!(vma->vm_flags & VM_WRITE))
1028 			map->flags |= GNTMAP_readonly;
1029 	}
1030 
1031 	mutex_unlock(&priv->lock);
1032 
1033 	if (use_ptemod) {
1034 		err = apply_to_page_range(vma->vm_mm, vma->vm_start,
1035 					  vma->vm_end - vma->vm_start,
1036 					  find_grant_ptes, map);
1037 		if (err) {
1038 			pr_warn("find_grant_ptes() failure.\n");
1039 			goto out_put_map;
1040 		}
1041 	}
1042 
1043 	err = map_grant_pages(map);
1044 	if (err)
1045 		goto out_put_map;
1046 
1047 	if (!use_ptemod) {
1048 		for (i = 0; i < count; i++) {
1049 			err = vm_insert_page(vma, vma->vm_start + i*PAGE_SIZE,
1050 				map->pages[i]);
1051 			if (err)
1052 				goto out_put_map;
1053 		}
1054 	} else {
1055 #ifdef CONFIG_X86
1056 		/*
1057 		 * If the PTEs were not made special by the grant map
1058 		 * hypercall, do so here.
1059 		 *
1060 		 * This is racy since the mapping is already visible
1061 		 * to userspace but userspace should be well-behaved
1062 		 * enough to not touch it until the mmap() call
1063 		 * returns.
1064 		 */
1065 		if (!xen_feature(XENFEAT_gnttab_map_avail_bits)) {
1066 			apply_to_page_range(vma->vm_mm, vma->vm_start,
1067 					    vma->vm_end - vma->vm_start,
1068 					    set_grant_ptes_as_special, NULL);
1069 		}
1070 #endif
1071 		map->pages_vm_start = vma->vm_start;
1072 	}
1073 
1074 	return 0;
1075 
1076 unlock_out:
1077 	mutex_unlock(&priv->lock);
1078 	return err;
1079 
1080 out_unlock_put:
1081 	mutex_unlock(&priv->lock);
1082 out_put_map:
1083 	if (use_ptemod)
1084 		map->vma = NULL;
1085 	gntdev_put_map(priv, map);
1086 	return err;
1087 }
1088 
1089 static const struct file_operations gntdev_fops = {
1090 	.owner = THIS_MODULE,
1091 	.open = gntdev_open,
1092 	.release = gntdev_release,
1093 	.mmap = gntdev_mmap,
1094 	.unlocked_ioctl = gntdev_ioctl
1095 };
1096 
1097 static struct miscdevice gntdev_miscdev = {
1098 	.minor        = MISC_DYNAMIC_MINOR,
1099 	.name         = "xen/gntdev",
1100 	.fops         = &gntdev_fops,
1101 };
1102 
1103 /* ------------------------------------------------------------------ */
1104 
1105 static int __init gntdev_init(void)
1106 {
1107 	int err;
1108 
1109 	if (!xen_domain())
1110 		return -ENODEV;
1111 
1112 	use_ptemod = !xen_feature(XENFEAT_auto_translated_physmap);
1113 
1114 	err = misc_register(&gntdev_miscdev);
1115 	if (err != 0) {
1116 		pr_err("Could not register gntdev device\n");
1117 		return err;
1118 	}
1119 	return 0;
1120 }
1121 
1122 static void __exit gntdev_exit(void)
1123 {
1124 	misc_deregister(&gntdev_miscdev);
1125 }
1126 
1127 module_init(gntdev_init);
1128 module_exit(gntdev_exit);
1129 
1130 /* ------------------------------------------------------------------ */
1131