xref: /openbmc/linux/drivers/xen/gntdev.c (revision 8bd1369b)
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 #include <linux/refcount.h>
40 
41 #include <xen/xen.h>
42 #include <xen/grant_table.h>
43 #include <xen/balloon.h>
44 #include <xen/gntdev.h>
45 #include <xen/events.h>
46 #include <xen/page.h>
47 #include <asm/xen/hypervisor.h>
48 #include <asm/xen/hypercall.h>
49 
50 MODULE_LICENSE("GPL");
51 MODULE_AUTHOR("Derek G. Murray <Derek.Murray@cl.cam.ac.uk>, "
52 	      "Gerd Hoffmann <kraxel@redhat.com>");
53 MODULE_DESCRIPTION("User-space granted page access driver");
54 
55 static int limit = 1024*1024;
56 module_param(limit, int, 0644);
57 MODULE_PARM_DESC(limit, "Maximum number of grants that may be mapped by "
58 		"the gntdev device");
59 
60 static atomic_t pages_mapped = ATOMIC_INIT(0);
61 
62 static int use_ptemod;
63 #define populate_freeable_maps use_ptemod
64 
65 struct gntdev_priv {
66 	/* maps with visible offsets in the file descriptor */
67 	struct list_head maps;
68 	/* maps that are not visible; will be freed on munmap.
69 	 * Only populated if populate_freeable_maps == 1 */
70 	struct list_head freeable_maps;
71 	/* lock protects maps and freeable_maps */
72 	struct mutex lock;
73 	struct mm_struct *mm;
74 	struct mmu_notifier mn;
75 };
76 
77 struct unmap_notify {
78 	int flags;
79 	/* Address relative to the start of the grant_map */
80 	int addr;
81 	int event;
82 };
83 
84 struct grant_map {
85 	struct list_head next;
86 	struct vm_area_struct *vma;
87 	int index;
88 	int count;
89 	int flags;
90 	refcount_t users;
91 	struct unmap_notify notify;
92 	struct ioctl_gntdev_grant_ref *grants;
93 	struct gnttab_map_grant_ref   *map_ops;
94 	struct gnttab_unmap_grant_ref *unmap_ops;
95 	struct gnttab_map_grant_ref   *kmap_ops;
96 	struct gnttab_unmap_grant_ref *kunmap_ops;
97 	struct page **pages;
98 	unsigned long pages_vm_start;
99 };
100 
101 static int unmap_grant_pages(struct grant_map *map, int offset, int pages);
102 
103 /* ------------------------------------------------------------------ */
104 
105 static void gntdev_print_maps(struct gntdev_priv *priv,
106 			      char *text, int text_index)
107 {
108 #ifdef DEBUG
109 	struct grant_map *map;
110 
111 	pr_debug("%s: maps list (priv %p)\n", __func__, priv);
112 	list_for_each_entry(map, &priv->maps, next)
113 		pr_debug("  index %2d, count %2d %s\n",
114 		       map->index, map->count,
115 		       map->index == text_index && text ? text : "");
116 #endif
117 }
118 
119 static void gntdev_free_map(struct grant_map *map)
120 {
121 	if (map == NULL)
122 		return;
123 
124 	if (map->pages)
125 		gnttab_free_pages(map->count, map->pages);
126 	kfree(map->pages);
127 	kfree(map->grants);
128 	kfree(map->map_ops);
129 	kfree(map->unmap_ops);
130 	kfree(map->kmap_ops);
131 	kfree(map->kunmap_ops);
132 	kfree(map);
133 }
134 
135 static struct grant_map *gntdev_alloc_map(struct gntdev_priv *priv, int count)
136 {
137 	struct grant_map *add;
138 	int i;
139 
140 	add = kzalloc(sizeof(struct grant_map), GFP_KERNEL);
141 	if (NULL == add)
142 		return NULL;
143 
144 	add->grants    = kcalloc(count, sizeof(add->grants[0]), GFP_KERNEL);
145 	add->map_ops   = kcalloc(count, sizeof(add->map_ops[0]), GFP_KERNEL);
146 	add->unmap_ops = kcalloc(count, sizeof(add->unmap_ops[0]), GFP_KERNEL);
147 	add->kmap_ops  = kcalloc(count, sizeof(add->kmap_ops[0]), GFP_KERNEL);
148 	add->kunmap_ops = kcalloc(count, sizeof(add->kunmap_ops[0]), GFP_KERNEL);
149 	add->pages     = kcalloc(count, sizeof(add->pages[0]), GFP_KERNEL);
150 	if (NULL == add->grants    ||
151 	    NULL == add->map_ops   ||
152 	    NULL == add->unmap_ops ||
153 	    NULL == add->kmap_ops  ||
154 	    NULL == add->kunmap_ops ||
155 	    NULL == add->pages)
156 		goto err;
157 
158 	if (gnttab_alloc_pages(count, add->pages))
159 		goto err;
160 
161 	for (i = 0; i < count; i++) {
162 		add->map_ops[i].handle = -1;
163 		add->unmap_ops[i].handle = -1;
164 		add->kmap_ops[i].handle = -1;
165 		add->kunmap_ops[i].handle = -1;
166 	}
167 
168 	add->index = 0;
169 	add->count = count;
170 	refcount_set(&add->users, 1);
171 
172 	return add;
173 
174 err:
175 	gntdev_free_map(add);
176 	return NULL;
177 }
178 
179 static void gntdev_add_map(struct gntdev_priv *priv, struct grant_map *add)
180 {
181 	struct grant_map *map;
182 
183 	list_for_each_entry(map, &priv->maps, next) {
184 		if (add->index + add->count < map->index) {
185 			list_add_tail(&add->next, &map->next);
186 			goto done;
187 		}
188 		add->index = map->index + map->count;
189 	}
190 	list_add_tail(&add->next, &priv->maps);
191 
192 done:
193 	gntdev_print_maps(priv, "[new]", add->index);
194 }
195 
196 static struct grant_map *gntdev_find_map_index(struct gntdev_priv *priv,
197 		int index, int count)
198 {
199 	struct grant_map *map;
200 
201 	list_for_each_entry(map, &priv->maps, next) {
202 		if (map->index != index)
203 			continue;
204 		if (count && map->count != count)
205 			continue;
206 		return map;
207 	}
208 	return NULL;
209 }
210 
211 static void gntdev_put_map(struct gntdev_priv *priv, struct grant_map *map)
212 {
213 	if (!map)
214 		return;
215 
216 	if (!refcount_dec_and_test(&map->users))
217 		return;
218 
219 	atomic_sub(map->count, &pages_mapped);
220 
221 	if (map->notify.flags & UNMAP_NOTIFY_SEND_EVENT) {
222 		notify_remote_via_evtchn(map->notify.event);
223 		evtchn_put(map->notify.event);
224 	}
225 
226 	if (populate_freeable_maps && priv) {
227 		mutex_lock(&priv->lock);
228 		list_del(&map->next);
229 		mutex_unlock(&priv->lock);
230 	}
231 
232 	if (map->pages && !use_ptemod)
233 		unmap_grant_pages(map, 0, map->count);
234 	gntdev_free_map(map);
235 }
236 
237 /* ------------------------------------------------------------------ */
238 
239 static int find_grant_ptes(pte_t *pte, pgtable_t token,
240 		unsigned long addr, void *data)
241 {
242 	struct grant_map *map = data;
243 	unsigned int pgnr = (addr - map->vma->vm_start) >> PAGE_SHIFT;
244 	int flags = map->flags | GNTMAP_application_map | GNTMAP_contains_pte;
245 	u64 pte_maddr;
246 
247 	BUG_ON(pgnr >= map->count);
248 	pte_maddr = arbitrary_virt_to_machine(pte).maddr;
249 
250 	/*
251 	 * Set the PTE as special to force get_user_pages_fast() fall
252 	 * back to the slow path.  If this is not supported as part of
253 	 * the grant map, it will be done afterwards.
254 	 */
255 	if (xen_feature(XENFEAT_gnttab_map_avail_bits))
256 		flags |= (1 << _GNTMAP_guest_avail0);
257 
258 	gnttab_set_map_op(&map->map_ops[pgnr], pte_maddr, flags,
259 			  map->grants[pgnr].ref,
260 			  map->grants[pgnr].domid);
261 	gnttab_set_unmap_op(&map->unmap_ops[pgnr], pte_maddr, flags,
262 			    -1 /* handle */);
263 	return 0;
264 }
265 
266 #ifdef CONFIG_X86
267 static int set_grant_ptes_as_special(pte_t *pte, pgtable_t token,
268 				     unsigned long addr, void *data)
269 {
270 	set_pte_at(current->mm, addr, pte, pte_mkspecial(*pte));
271 	return 0;
272 }
273 #endif
274 
275 static int map_grant_pages(struct grant_map *map)
276 {
277 	int i, err = 0;
278 
279 	if (!use_ptemod) {
280 		/* Note: it could already be mapped */
281 		if (map->map_ops[0].handle != -1)
282 			return 0;
283 		for (i = 0; i < map->count; i++) {
284 			unsigned long addr = (unsigned long)
285 				pfn_to_kaddr(page_to_pfn(map->pages[i]));
286 			gnttab_set_map_op(&map->map_ops[i], addr, map->flags,
287 				map->grants[i].ref,
288 				map->grants[i].domid);
289 			gnttab_set_unmap_op(&map->unmap_ops[i], addr,
290 				map->flags, -1 /* handle */);
291 		}
292 	} else {
293 		/*
294 		 * Setup the map_ops corresponding to the pte entries pointing
295 		 * to the kernel linear addresses of the struct pages.
296 		 * These ptes are completely different from the user ptes dealt
297 		 * with find_grant_ptes.
298 		 */
299 		for (i = 0; i < map->count; i++) {
300 			unsigned long address = (unsigned long)
301 				pfn_to_kaddr(page_to_pfn(map->pages[i]));
302 			BUG_ON(PageHighMem(map->pages[i]));
303 
304 			gnttab_set_map_op(&map->kmap_ops[i], address,
305 				map->flags | GNTMAP_host_map,
306 				map->grants[i].ref,
307 				map->grants[i].domid);
308 			gnttab_set_unmap_op(&map->kunmap_ops[i], address,
309 				map->flags | GNTMAP_host_map, -1);
310 		}
311 	}
312 
313 	pr_debug("map %d+%d\n", map->index, map->count);
314 	err = gnttab_map_refs(map->map_ops, use_ptemod ? map->kmap_ops : NULL,
315 			map->pages, map->count);
316 	if (err)
317 		return err;
318 
319 	for (i = 0; i < map->count; i++) {
320 		if (map->map_ops[i].status) {
321 			err = -EINVAL;
322 			continue;
323 		}
324 
325 		map->unmap_ops[i].handle = map->map_ops[i].handle;
326 		if (use_ptemod)
327 			map->kunmap_ops[i].handle = map->kmap_ops[i].handle;
328 	}
329 	return err;
330 }
331 
332 static int __unmap_grant_pages(struct grant_map *map, int offset, int pages)
333 {
334 	int i, err = 0;
335 	struct gntab_unmap_queue_data unmap_data;
336 
337 	if (map->notify.flags & UNMAP_NOTIFY_CLEAR_BYTE) {
338 		int pgno = (map->notify.addr >> PAGE_SHIFT);
339 		if (pgno >= offset && pgno < offset + pages) {
340 			/* No need for kmap, pages are in lowmem */
341 			uint8_t *tmp = pfn_to_kaddr(page_to_pfn(map->pages[pgno]));
342 			tmp[map->notify.addr & (PAGE_SIZE-1)] = 0;
343 			map->notify.flags &= ~UNMAP_NOTIFY_CLEAR_BYTE;
344 		}
345 	}
346 
347 	unmap_data.unmap_ops = map->unmap_ops + offset;
348 	unmap_data.kunmap_ops = use_ptemod ? map->kunmap_ops + offset : NULL;
349 	unmap_data.pages = map->pages + offset;
350 	unmap_data.count = pages;
351 
352 	err = gnttab_unmap_refs_sync(&unmap_data);
353 	if (err)
354 		return err;
355 
356 	for (i = 0; i < pages; i++) {
357 		if (map->unmap_ops[offset+i].status)
358 			err = -EINVAL;
359 		pr_debug("unmap handle=%d st=%d\n",
360 			map->unmap_ops[offset+i].handle,
361 			map->unmap_ops[offset+i].status);
362 		map->unmap_ops[offset+i].handle = -1;
363 	}
364 	return err;
365 }
366 
367 static int unmap_grant_pages(struct grant_map *map, int offset, int pages)
368 {
369 	int range, err = 0;
370 
371 	pr_debug("unmap %d+%d [%d+%d]\n", map->index, map->count, offset, pages);
372 
373 	/* It is possible the requested range will have a "hole" where we
374 	 * already unmapped some of the grants. Only unmap valid ranges.
375 	 */
376 	while (pages && !err) {
377 		while (pages && map->unmap_ops[offset].handle == -1) {
378 			offset++;
379 			pages--;
380 		}
381 		range = 0;
382 		while (range < pages) {
383 			if (map->unmap_ops[offset+range].handle == -1)
384 				break;
385 			range++;
386 		}
387 		err = __unmap_grant_pages(map, offset, range);
388 		offset += range;
389 		pages -= range;
390 	}
391 
392 	return err;
393 }
394 
395 /* ------------------------------------------------------------------ */
396 
397 static void gntdev_vma_open(struct vm_area_struct *vma)
398 {
399 	struct grant_map *map = vma->vm_private_data;
400 
401 	pr_debug("gntdev_vma_open %p\n", vma);
402 	refcount_inc(&map->users);
403 }
404 
405 static void gntdev_vma_close(struct vm_area_struct *vma)
406 {
407 	struct grant_map *map = vma->vm_private_data;
408 	struct file *file = vma->vm_file;
409 	struct gntdev_priv *priv = file->private_data;
410 
411 	pr_debug("gntdev_vma_close %p\n", vma);
412 	if (use_ptemod) {
413 		/* It is possible that an mmu notifier could be running
414 		 * concurrently, so take priv->lock to ensure that the vma won't
415 		 * vanishing during the unmap_grant_pages call, since we will
416 		 * spin here until that completes. Such a concurrent call will
417 		 * not do any unmapping, since that has been done prior to
418 		 * closing the vma, but it may still iterate the unmap_ops list.
419 		 */
420 		mutex_lock(&priv->lock);
421 		map->vma = NULL;
422 		mutex_unlock(&priv->lock);
423 	}
424 	vma->vm_private_data = NULL;
425 	gntdev_put_map(priv, map);
426 }
427 
428 static struct page *gntdev_vma_find_special_page(struct vm_area_struct *vma,
429 						 unsigned long addr)
430 {
431 	struct grant_map *map = vma->vm_private_data;
432 
433 	return map->pages[(addr - map->pages_vm_start) >> PAGE_SHIFT];
434 }
435 
436 static const struct vm_operations_struct gntdev_vmops = {
437 	.open = gntdev_vma_open,
438 	.close = gntdev_vma_close,
439 	.find_special_page = gntdev_vma_find_special_page,
440 };
441 
442 /* ------------------------------------------------------------------ */
443 
444 static void unmap_if_in_range(struct grant_map *map,
445 			      unsigned long start, unsigned long end)
446 {
447 	unsigned long mstart, mend;
448 	int err;
449 
450 	if (!map->vma)
451 		return;
452 	if (map->vma->vm_start >= end)
453 		return;
454 	if (map->vma->vm_end <= start)
455 		return;
456 	mstart = max(start, map->vma->vm_start);
457 	mend   = min(end,   map->vma->vm_end);
458 	pr_debug("map %d+%d (%lx %lx), range %lx %lx, mrange %lx %lx\n",
459 			map->index, map->count,
460 			map->vma->vm_start, map->vma->vm_end,
461 			start, end, mstart, mend);
462 	err = unmap_grant_pages(map,
463 				(mstart - map->vma->vm_start) >> PAGE_SHIFT,
464 				(mend - mstart) >> PAGE_SHIFT);
465 	WARN_ON(err);
466 }
467 
468 static void mn_invl_range_start(struct mmu_notifier *mn,
469 				struct mm_struct *mm,
470 				unsigned long start, unsigned long end)
471 {
472 	struct gntdev_priv *priv = container_of(mn, struct gntdev_priv, mn);
473 	struct grant_map *map;
474 
475 	mutex_lock(&priv->lock);
476 	list_for_each_entry(map, &priv->maps, next) {
477 		unmap_if_in_range(map, start, end);
478 	}
479 	list_for_each_entry(map, &priv->freeable_maps, next) {
480 		unmap_if_in_range(map, start, end);
481 	}
482 	mutex_unlock(&priv->lock);
483 }
484 
485 static void mn_release(struct mmu_notifier *mn,
486 		       struct mm_struct *mm)
487 {
488 	struct gntdev_priv *priv = container_of(mn, struct gntdev_priv, mn);
489 	struct grant_map *map;
490 	int err;
491 
492 	mutex_lock(&priv->lock);
493 	list_for_each_entry(map, &priv->maps, next) {
494 		if (!map->vma)
495 			continue;
496 		pr_debug("map %d+%d (%lx %lx)\n",
497 				map->index, map->count,
498 				map->vma->vm_start, map->vma->vm_end);
499 		err = unmap_grant_pages(map, /* offset */ 0, map->count);
500 		WARN_ON(err);
501 	}
502 	list_for_each_entry(map, &priv->freeable_maps, next) {
503 		if (!map->vma)
504 			continue;
505 		pr_debug("map %d+%d (%lx %lx)\n",
506 				map->index, map->count,
507 				map->vma->vm_start, map->vma->vm_end);
508 		err = unmap_grant_pages(map, /* offset */ 0, map->count);
509 		WARN_ON(err);
510 	}
511 	mutex_unlock(&priv->lock);
512 }
513 
514 static const struct mmu_notifier_ops gntdev_mmu_ops = {
515 	.release                = mn_release,
516 	.invalidate_range_start = mn_invl_range_start,
517 };
518 
519 /* ------------------------------------------------------------------ */
520 
521 static int gntdev_open(struct inode *inode, struct file *flip)
522 {
523 	struct gntdev_priv *priv;
524 	int ret = 0;
525 
526 	priv = kzalloc(sizeof(*priv), GFP_KERNEL);
527 	if (!priv)
528 		return -ENOMEM;
529 
530 	INIT_LIST_HEAD(&priv->maps);
531 	INIT_LIST_HEAD(&priv->freeable_maps);
532 	mutex_init(&priv->lock);
533 
534 	if (use_ptemod) {
535 		priv->mm = get_task_mm(current);
536 		if (!priv->mm) {
537 			kfree(priv);
538 			return -ENOMEM;
539 		}
540 		priv->mn.ops = &gntdev_mmu_ops;
541 		ret = mmu_notifier_register(&priv->mn, priv->mm);
542 		mmput(priv->mm);
543 	}
544 
545 	if (ret) {
546 		kfree(priv);
547 		return ret;
548 	}
549 
550 	flip->private_data = priv;
551 	pr_debug("priv %p\n", priv);
552 
553 	return 0;
554 }
555 
556 static int gntdev_release(struct inode *inode, struct file *flip)
557 {
558 	struct gntdev_priv *priv = flip->private_data;
559 	struct grant_map *map;
560 
561 	pr_debug("priv %p\n", priv);
562 
563 	mutex_lock(&priv->lock);
564 	while (!list_empty(&priv->maps)) {
565 		map = list_entry(priv->maps.next, struct grant_map, next);
566 		list_del(&map->next);
567 		gntdev_put_map(NULL /* already removed */, map);
568 	}
569 	WARN_ON(!list_empty(&priv->freeable_maps));
570 	mutex_unlock(&priv->lock);
571 
572 	if (use_ptemod)
573 		mmu_notifier_unregister(&priv->mn, priv->mm);
574 	kfree(priv);
575 	return 0;
576 }
577 
578 static long gntdev_ioctl_map_grant_ref(struct gntdev_priv *priv,
579 				       struct ioctl_gntdev_map_grant_ref __user *u)
580 {
581 	struct ioctl_gntdev_map_grant_ref op;
582 	struct grant_map *map;
583 	int err;
584 
585 	if (copy_from_user(&op, u, sizeof(op)) != 0)
586 		return -EFAULT;
587 	pr_debug("priv %p, add %d\n", priv, op.count);
588 	if (unlikely(op.count <= 0))
589 		return -EINVAL;
590 
591 	err = -ENOMEM;
592 	map = gntdev_alloc_map(priv, op.count);
593 	if (!map)
594 		return err;
595 
596 	if (unlikely(atomic_add_return(op.count, &pages_mapped) > limit)) {
597 		pr_debug("can't map: over limit\n");
598 		gntdev_put_map(NULL, map);
599 		return err;
600 	}
601 
602 	if (copy_from_user(map->grants, &u->refs,
603 			   sizeof(map->grants[0]) * op.count) != 0) {
604 		gntdev_put_map(NULL, map);
605 		return -EFAULT;
606 	}
607 
608 	mutex_lock(&priv->lock);
609 	gntdev_add_map(priv, map);
610 	op.index = map->index << PAGE_SHIFT;
611 	mutex_unlock(&priv->lock);
612 
613 	if (copy_to_user(u, &op, sizeof(op)) != 0)
614 		return -EFAULT;
615 
616 	return 0;
617 }
618 
619 static long gntdev_ioctl_unmap_grant_ref(struct gntdev_priv *priv,
620 					 struct ioctl_gntdev_unmap_grant_ref __user *u)
621 {
622 	struct ioctl_gntdev_unmap_grant_ref op;
623 	struct grant_map *map;
624 	int err = -ENOENT;
625 
626 	if (copy_from_user(&op, u, sizeof(op)) != 0)
627 		return -EFAULT;
628 	pr_debug("priv %p, del %d+%d\n", priv, (int)op.index, (int)op.count);
629 
630 	mutex_lock(&priv->lock);
631 	map = gntdev_find_map_index(priv, op.index >> PAGE_SHIFT, op.count);
632 	if (map) {
633 		list_del(&map->next);
634 		if (populate_freeable_maps)
635 			list_add_tail(&map->next, &priv->freeable_maps);
636 		err = 0;
637 	}
638 	mutex_unlock(&priv->lock);
639 	if (map)
640 		gntdev_put_map(priv, map);
641 	return err;
642 }
643 
644 static long gntdev_ioctl_get_offset_for_vaddr(struct gntdev_priv *priv,
645 					      struct ioctl_gntdev_get_offset_for_vaddr __user *u)
646 {
647 	struct ioctl_gntdev_get_offset_for_vaddr op;
648 	struct vm_area_struct *vma;
649 	struct grant_map *map;
650 	int rv = -EINVAL;
651 
652 	if (copy_from_user(&op, u, sizeof(op)) != 0)
653 		return -EFAULT;
654 	pr_debug("priv %p, offset for vaddr %lx\n", priv, (unsigned long)op.vaddr);
655 
656 	down_read(&current->mm->mmap_sem);
657 	vma = find_vma(current->mm, op.vaddr);
658 	if (!vma || vma->vm_ops != &gntdev_vmops)
659 		goto out_unlock;
660 
661 	map = vma->vm_private_data;
662 	if (!map)
663 		goto out_unlock;
664 
665 	op.offset = map->index << PAGE_SHIFT;
666 	op.count = map->count;
667 	rv = 0;
668 
669  out_unlock:
670 	up_read(&current->mm->mmap_sem);
671 
672 	if (rv == 0 && copy_to_user(u, &op, sizeof(op)) != 0)
673 		return -EFAULT;
674 	return rv;
675 }
676 
677 static long gntdev_ioctl_notify(struct gntdev_priv *priv, void __user *u)
678 {
679 	struct ioctl_gntdev_unmap_notify op;
680 	struct grant_map *map;
681 	int rc;
682 	int out_flags;
683 	unsigned int out_event;
684 
685 	if (copy_from_user(&op, u, sizeof(op)))
686 		return -EFAULT;
687 
688 	if (op.action & ~(UNMAP_NOTIFY_CLEAR_BYTE|UNMAP_NOTIFY_SEND_EVENT))
689 		return -EINVAL;
690 
691 	/* We need to grab a reference to the event channel we are going to use
692 	 * to send the notify before releasing the reference we may already have
693 	 * (if someone has called this ioctl twice). This is required so that
694 	 * it is possible to change the clear_byte part of the notification
695 	 * without disturbing the event channel part, which may now be the last
696 	 * reference to that event channel.
697 	 */
698 	if (op.action & UNMAP_NOTIFY_SEND_EVENT) {
699 		if (evtchn_get(op.event_channel_port))
700 			return -EINVAL;
701 	}
702 
703 	out_flags = op.action;
704 	out_event = op.event_channel_port;
705 
706 	mutex_lock(&priv->lock);
707 
708 	list_for_each_entry(map, &priv->maps, next) {
709 		uint64_t begin = map->index << PAGE_SHIFT;
710 		uint64_t end = (map->index + map->count) << PAGE_SHIFT;
711 		if (op.index >= begin && op.index < end)
712 			goto found;
713 	}
714 	rc = -ENOENT;
715 	goto unlock_out;
716 
717  found:
718 	if ((op.action & UNMAP_NOTIFY_CLEAR_BYTE) &&
719 			(map->flags & GNTMAP_readonly)) {
720 		rc = -EINVAL;
721 		goto unlock_out;
722 	}
723 
724 	out_flags = map->notify.flags;
725 	out_event = map->notify.event;
726 
727 	map->notify.flags = op.action;
728 	map->notify.addr = op.index - (map->index << PAGE_SHIFT);
729 	map->notify.event = op.event_channel_port;
730 
731 	rc = 0;
732 
733  unlock_out:
734 	mutex_unlock(&priv->lock);
735 
736 	/* Drop the reference to the event channel we did not save in the map */
737 	if (out_flags & UNMAP_NOTIFY_SEND_EVENT)
738 		evtchn_put(out_event);
739 
740 	return rc;
741 }
742 
743 #define GNTDEV_COPY_BATCH 16
744 
745 struct gntdev_copy_batch {
746 	struct gnttab_copy ops[GNTDEV_COPY_BATCH];
747 	struct page *pages[GNTDEV_COPY_BATCH];
748 	s16 __user *status[GNTDEV_COPY_BATCH];
749 	unsigned int nr_ops;
750 	unsigned int nr_pages;
751 };
752 
753 static int gntdev_get_page(struct gntdev_copy_batch *batch, void __user *virt,
754 			   bool writeable, unsigned long *gfn)
755 {
756 	unsigned long addr = (unsigned long)virt;
757 	struct page *page;
758 	unsigned long xen_pfn;
759 	int ret;
760 
761 	ret = get_user_pages_fast(addr, 1, writeable, &page);
762 	if (ret < 0)
763 		return ret;
764 
765 	batch->pages[batch->nr_pages++] = page;
766 
767 	xen_pfn = page_to_xen_pfn(page) + XEN_PFN_DOWN(addr & ~PAGE_MASK);
768 	*gfn = pfn_to_gfn(xen_pfn);
769 
770 	return 0;
771 }
772 
773 static void gntdev_put_pages(struct gntdev_copy_batch *batch)
774 {
775 	unsigned int i;
776 
777 	for (i = 0; i < batch->nr_pages; i++)
778 		put_page(batch->pages[i]);
779 	batch->nr_pages = 0;
780 }
781 
782 static int gntdev_copy(struct gntdev_copy_batch *batch)
783 {
784 	unsigned int i;
785 
786 	gnttab_batch_copy(batch->ops, batch->nr_ops);
787 	gntdev_put_pages(batch);
788 
789 	/*
790 	 * For each completed op, update the status if the op failed
791 	 * and all previous ops for the segment were successful.
792 	 */
793 	for (i = 0; i < batch->nr_ops; i++) {
794 		s16 status = batch->ops[i].status;
795 		s16 old_status;
796 
797 		if (status == GNTST_okay)
798 			continue;
799 
800 		if (__get_user(old_status, batch->status[i]))
801 			return -EFAULT;
802 
803 		if (old_status != GNTST_okay)
804 			continue;
805 
806 		if (__put_user(status, batch->status[i]))
807 			return -EFAULT;
808 	}
809 
810 	batch->nr_ops = 0;
811 	return 0;
812 }
813 
814 static int gntdev_grant_copy_seg(struct gntdev_copy_batch *batch,
815 				 struct gntdev_grant_copy_segment *seg,
816 				 s16 __user *status)
817 {
818 	uint16_t copied = 0;
819 
820 	/*
821 	 * Disallow local -> local copies since there is only space in
822 	 * batch->pages for one page per-op and this would be a very
823 	 * expensive memcpy().
824 	 */
825 	if (!(seg->flags & (GNTCOPY_source_gref | GNTCOPY_dest_gref)))
826 		return -EINVAL;
827 
828 	/* Can't cross page if source/dest is a grant ref. */
829 	if (seg->flags & GNTCOPY_source_gref) {
830 		if (seg->source.foreign.offset + seg->len > XEN_PAGE_SIZE)
831 			return -EINVAL;
832 	}
833 	if (seg->flags & GNTCOPY_dest_gref) {
834 		if (seg->dest.foreign.offset + seg->len > XEN_PAGE_SIZE)
835 			return -EINVAL;
836 	}
837 
838 	if (put_user(GNTST_okay, status))
839 		return -EFAULT;
840 
841 	while (copied < seg->len) {
842 		struct gnttab_copy *op;
843 		void __user *virt;
844 		size_t len, off;
845 		unsigned long gfn;
846 		int ret;
847 
848 		if (batch->nr_ops >= GNTDEV_COPY_BATCH) {
849 			ret = gntdev_copy(batch);
850 			if (ret < 0)
851 				return ret;
852 		}
853 
854 		len = seg->len - copied;
855 
856 		op = &batch->ops[batch->nr_ops];
857 		op->flags = 0;
858 
859 		if (seg->flags & GNTCOPY_source_gref) {
860 			op->source.u.ref = seg->source.foreign.ref;
861 			op->source.domid = seg->source.foreign.domid;
862 			op->source.offset = seg->source.foreign.offset + copied;
863 			op->flags |= GNTCOPY_source_gref;
864 		} else {
865 			virt = seg->source.virt + copied;
866 			off = (unsigned long)virt & ~XEN_PAGE_MASK;
867 			len = min(len, (size_t)XEN_PAGE_SIZE - off);
868 
869 			ret = gntdev_get_page(batch, virt, false, &gfn);
870 			if (ret < 0)
871 				return ret;
872 
873 			op->source.u.gmfn = gfn;
874 			op->source.domid = DOMID_SELF;
875 			op->source.offset = off;
876 		}
877 
878 		if (seg->flags & GNTCOPY_dest_gref) {
879 			op->dest.u.ref = seg->dest.foreign.ref;
880 			op->dest.domid = seg->dest.foreign.domid;
881 			op->dest.offset = seg->dest.foreign.offset + copied;
882 			op->flags |= GNTCOPY_dest_gref;
883 		} else {
884 			virt = seg->dest.virt + copied;
885 			off = (unsigned long)virt & ~XEN_PAGE_MASK;
886 			len = min(len, (size_t)XEN_PAGE_SIZE - off);
887 
888 			ret = gntdev_get_page(batch, virt, true, &gfn);
889 			if (ret < 0)
890 				return ret;
891 
892 			op->dest.u.gmfn = gfn;
893 			op->dest.domid = DOMID_SELF;
894 			op->dest.offset = off;
895 		}
896 
897 		op->len = len;
898 		copied += len;
899 
900 		batch->status[batch->nr_ops] = status;
901 		batch->nr_ops++;
902 	}
903 
904 	return 0;
905 }
906 
907 static long gntdev_ioctl_grant_copy(struct gntdev_priv *priv, void __user *u)
908 {
909 	struct ioctl_gntdev_grant_copy copy;
910 	struct gntdev_copy_batch batch;
911 	unsigned int i;
912 	int ret = 0;
913 
914 	if (copy_from_user(&copy, u, sizeof(copy)))
915 		return -EFAULT;
916 
917 	batch.nr_ops = 0;
918 	batch.nr_pages = 0;
919 
920 	for (i = 0; i < copy.count; i++) {
921 		struct gntdev_grant_copy_segment seg;
922 
923 		if (copy_from_user(&seg, &copy.segments[i], sizeof(seg))) {
924 			ret = -EFAULT;
925 			goto out;
926 		}
927 
928 		ret = gntdev_grant_copy_seg(&batch, &seg, &copy.segments[i].status);
929 		if (ret < 0)
930 			goto out;
931 
932 		cond_resched();
933 	}
934 	if (batch.nr_ops)
935 		ret = gntdev_copy(&batch);
936 	return ret;
937 
938   out:
939 	gntdev_put_pages(&batch);
940 	return ret;
941 }
942 
943 static long gntdev_ioctl(struct file *flip,
944 			 unsigned int cmd, unsigned long arg)
945 {
946 	struct gntdev_priv *priv = flip->private_data;
947 	void __user *ptr = (void __user *)arg;
948 
949 	switch (cmd) {
950 	case IOCTL_GNTDEV_MAP_GRANT_REF:
951 		return gntdev_ioctl_map_grant_ref(priv, ptr);
952 
953 	case IOCTL_GNTDEV_UNMAP_GRANT_REF:
954 		return gntdev_ioctl_unmap_grant_ref(priv, ptr);
955 
956 	case IOCTL_GNTDEV_GET_OFFSET_FOR_VADDR:
957 		return gntdev_ioctl_get_offset_for_vaddr(priv, ptr);
958 
959 	case IOCTL_GNTDEV_SET_UNMAP_NOTIFY:
960 		return gntdev_ioctl_notify(priv, ptr);
961 
962 	case IOCTL_GNTDEV_GRANT_COPY:
963 		return gntdev_ioctl_grant_copy(priv, ptr);
964 
965 	default:
966 		pr_debug("priv %p, unknown cmd %x\n", priv, cmd);
967 		return -ENOIOCTLCMD;
968 	}
969 
970 	return 0;
971 }
972 
973 static int gntdev_mmap(struct file *flip, struct vm_area_struct *vma)
974 {
975 	struct gntdev_priv *priv = flip->private_data;
976 	int index = vma->vm_pgoff;
977 	int count = vma_pages(vma);
978 	struct grant_map *map;
979 	int i, err = -EINVAL;
980 
981 	if ((vma->vm_flags & VM_WRITE) && !(vma->vm_flags & VM_SHARED))
982 		return -EINVAL;
983 
984 	pr_debug("map %d+%d at %lx (pgoff %lx)\n",
985 			index, count, vma->vm_start, vma->vm_pgoff);
986 
987 	mutex_lock(&priv->lock);
988 	map = gntdev_find_map_index(priv, index, count);
989 	if (!map)
990 		goto unlock_out;
991 	if (use_ptemod && map->vma)
992 		goto unlock_out;
993 	if (use_ptemod && priv->mm != vma->vm_mm) {
994 		pr_warn("Huh? Other mm?\n");
995 		goto unlock_out;
996 	}
997 
998 	refcount_inc(&map->users);
999 
1000 	vma->vm_ops = &gntdev_vmops;
1001 
1002 	vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP | VM_MIXEDMAP;
1003 
1004 	if (use_ptemod)
1005 		vma->vm_flags |= VM_DONTCOPY;
1006 
1007 	vma->vm_private_data = map;
1008 
1009 	if (use_ptemod)
1010 		map->vma = vma;
1011 
1012 	if (map->flags) {
1013 		if ((vma->vm_flags & VM_WRITE) &&
1014 				(map->flags & GNTMAP_readonly))
1015 			goto out_unlock_put;
1016 	} else {
1017 		map->flags = GNTMAP_host_map;
1018 		if (!(vma->vm_flags & VM_WRITE))
1019 			map->flags |= GNTMAP_readonly;
1020 	}
1021 
1022 	mutex_unlock(&priv->lock);
1023 
1024 	if (use_ptemod) {
1025 		map->pages_vm_start = vma->vm_start;
1026 		err = apply_to_page_range(vma->vm_mm, vma->vm_start,
1027 					  vma->vm_end - vma->vm_start,
1028 					  find_grant_ptes, map);
1029 		if (err) {
1030 			pr_warn("find_grant_ptes() failure.\n");
1031 			goto out_put_map;
1032 		}
1033 	}
1034 
1035 	err = map_grant_pages(map);
1036 	if (err)
1037 		goto out_put_map;
1038 
1039 	if (!use_ptemod) {
1040 		for (i = 0; i < count; i++) {
1041 			err = vm_insert_page(vma, vma->vm_start + i*PAGE_SIZE,
1042 				map->pages[i]);
1043 			if (err)
1044 				goto out_put_map;
1045 		}
1046 	} else {
1047 #ifdef CONFIG_X86
1048 		/*
1049 		 * If the PTEs were not made special by the grant map
1050 		 * hypercall, do so here.
1051 		 *
1052 		 * This is racy since the mapping is already visible
1053 		 * to userspace but userspace should be well-behaved
1054 		 * enough to not touch it until the mmap() call
1055 		 * returns.
1056 		 */
1057 		if (!xen_feature(XENFEAT_gnttab_map_avail_bits)) {
1058 			apply_to_page_range(vma->vm_mm, vma->vm_start,
1059 					    vma->vm_end - vma->vm_start,
1060 					    set_grant_ptes_as_special, NULL);
1061 		}
1062 #endif
1063 	}
1064 
1065 	return 0;
1066 
1067 unlock_out:
1068 	mutex_unlock(&priv->lock);
1069 	return err;
1070 
1071 out_unlock_put:
1072 	mutex_unlock(&priv->lock);
1073 out_put_map:
1074 	if (use_ptemod) {
1075 		map->vma = NULL;
1076 		unmap_grant_pages(map, 0, map->count);
1077 	}
1078 	gntdev_put_map(priv, map);
1079 	return err;
1080 }
1081 
1082 static const struct file_operations gntdev_fops = {
1083 	.owner = THIS_MODULE,
1084 	.open = gntdev_open,
1085 	.release = gntdev_release,
1086 	.mmap = gntdev_mmap,
1087 	.unlocked_ioctl = gntdev_ioctl
1088 };
1089 
1090 static struct miscdevice gntdev_miscdev = {
1091 	.minor        = MISC_DYNAMIC_MINOR,
1092 	.name         = "xen/gntdev",
1093 	.fops         = &gntdev_fops,
1094 };
1095 
1096 /* ------------------------------------------------------------------ */
1097 
1098 static int __init gntdev_init(void)
1099 {
1100 	int err;
1101 
1102 	if (!xen_domain())
1103 		return -ENODEV;
1104 
1105 	use_ptemod = !xen_feature(XENFEAT_auto_translated_physmap);
1106 
1107 	err = misc_register(&gntdev_miscdev);
1108 	if (err != 0) {
1109 		pr_err("Could not register gntdev device\n");
1110 		return err;
1111 	}
1112 	return 0;
1113 }
1114 
1115 static void __exit gntdev_exit(void)
1116 {
1117 	misc_deregister(&gntdev_miscdev);
1118 }
1119 
1120 module_init(gntdev_init);
1121 module_exit(gntdev_exit);
1122 
1123 /* ------------------------------------------------------------------ */
1124