xref: /openbmc/linux/drivers/xen/gntdev.c (revision 77a87824)
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 <xen/page.h>
45 #include <asm/xen/hypervisor.h>
46 #include <asm/xen/hypercall.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 static int __unmap_grant_pages(struct grant_map *map, int offset, int pages)
331 {
332 	int i, err = 0;
333 	struct gntab_unmap_queue_data unmap_data;
334 
335 	if (map->notify.flags & UNMAP_NOTIFY_CLEAR_BYTE) {
336 		int pgno = (map->notify.addr >> PAGE_SHIFT);
337 		if (pgno >= offset && pgno < offset + pages) {
338 			/* No need for kmap, pages are in lowmem */
339 			uint8_t *tmp = pfn_to_kaddr(page_to_pfn(map->pages[pgno]));
340 			tmp[map->notify.addr & (PAGE_SIZE-1)] = 0;
341 			map->notify.flags &= ~UNMAP_NOTIFY_CLEAR_BYTE;
342 		}
343 	}
344 
345 	unmap_data.unmap_ops = map->unmap_ops + offset;
346 	unmap_data.kunmap_ops = use_ptemod ? map->kunmap_ops + offset : NULL;
347 	unmap_data.pages = map->pages + offset;
348 	unmap_data.count = pages;
349 
350 	err = gnttab_unmap_refs_sync(&unmap_data);
351 	if (err)
352 		return err;
353 
354 	for (i = 0; i < pages; i++) {
355 		if (map->unmap_ops[offset+i].status)
356 			err = -EINVAL;
357 		pr_debug("unmap handle=%d st=%d\n",
358 			map->unmap_ops[offset+i].handle,
359 			map->unmap_ops[offset+i].status);
360 		map->unmap_ops[offset+i].handle = -1;
361 	}
362 	return err;
363 }
364 
365 static int unmap_grant_pages(struct grant_map *map, int offset, int pages)
366 {
367 	int range, err = 0;
368 
369 	pr_debug("unmap %d+%d [%d+%d]\n", map->index, map->count, offset, pages);
370 
371 	/* It is possible the requested range will have a "hole" where we
372 	 * already unmapped some of the grants. Only unmap valid ranges.
373 	 */
374 	while (pages && !err) {
375 		while (pages && map->unmap_ops[offset].handle == -1) {
376 			offset++;
377 			pages--;
378 		}
379 		range = 0;
380 		while (range < pages) {
381 			if (map->unmap_ops[offset+range].handle == -1) {
382 				range--;
383 				break;
384 			}
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 	atomic_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_invl_page(struct mmu_notifier *mn,
486 			 struct mm_struct *mm,
487 			 unsigned long address)
488 {
489 	mn_invl_range_start(mn, mm, address, address + PAGE_SIZE);
490 }
491 
492 static void mn_release(struct mmu_notifier *mn,
493 		       struct mm_struct *mm)
494 {
495 	struct gntdev_priv *priv = container_of(mn, struct gntdev_priv, mn);
496 	struct grant_map *map;
497 	int err;
498 
499 	mutex_lock(&priv->lock);
500 	list_for_each_entry(map, &priv->maps, next) {
501 		if (!map->vma)
502 			continue;
503 		pr_debug("map %d+%d (%lx %lx)\n",
504 				map->index, map->count,
505 				map->vma->vm_start, map->vma->vm_end);
506 		err = unmap_grant_pages(map, /* offset */ 0, map->count);
507 		WARN_ON(err);
508 	}
509 	list_for_each_entry(map, &priv->freeable_maps, next) {
510 		if (!map->vma)
511 			continue;
512 		pr_debug("map %d+%d (%lx %lx)\n",
513 				map->index, map->count,
514 				map->vma->vm_start, map->vma->vm_end);
515 		err = unmap_grant_pages(map, /* offset */ 0, map->count);
516 		WARN_ON(err);
517 	}
518 	mutex_unlock(&priv->lock);
519 }
520 
521 static const struct mmu_notifier_ops gntdev_mmu_ops = {
522 	.release                = mn_release,
523 	.invalidate_page        = mn_invl_page,
524 	.invalidate_range_start = mn_invl_range_start,
525 };
526 
527 /* ------------------------------------------------------------------ */
528 
529 static int gntdev_open(struct inode *inode, struct file *flip)
530 {
531 	struct gntdev_priv *priv;
532 	int ret = 0;
533 
534 	priv = kzalloc(sizeof(*priv), GFP_KERNEL);
535 	if (!priv)
536 		return -ENOMEM;
537 
538 	INIT_LIST_HEAD(&priv->maps);
539 	INIT_LIST_HEAD(&priv->freeable_maps);
540 	mutex_init(&priv->lock);
541 
542 	if (use_ptemod) {
543 		priv->mm = get_task_mm(current);
544 		if (!priv->mm) {
545 			kfree(priv);
546 			return -ENOMEM;
547 		}
548 		priv->mn.ops = &gntdev_mmu_ops;
549 		ret = mmu_notifier_register(&priv->mn, priv->mm);
550 		mmput(priv->mm);
551 	}
552 
553 	if (ret) {
554 		kfree(priv);
555 		return ret;
556 	}
557 
558 	flip->private_data = priv;
559 	pr_debug("priv %p\n", priv);
560 
561 	return 0;
562 }
563 
564 static int gntdev_release(struct inode *inode, struct file *flip)
565 {
566 	struct gntdev_priv *priv = flip->private_data;
567 	struct grant_map *map;
568 
569 	pr_debug("priv %p\n", priv);
570 
571 	mutex_lock(&priv->lock);
572 	while (!list_empty(&priv->maps)) {
573 		map = list_entry(priv->maps.next, struct grant_map, next);
574 		list_del(&map->next);
575 		gntdev_put_map(NULL /* already removed */, map);
576 	}
577 	WARN_ON(!list_empty(&priv->freeable_maps));
578 	mutex_unlock(&priv->lock);
579 
580 	if (use_ptemod)
581 		mmu_notifier_unregister(&priv->mn, priv->mm);
582 	kfree(priv);
583 	return 0;
584 }
585 
586 static long gntdev_ioctl_map_grant_ref(struct gntdev_priv *priv,
587 				       struct ioctl_gntdev_map_grant_ref __user *u)
588 {
589 	struct ioctl_gntdev_map_grant_ref op;
590 	struct grant_map *map;
591 	int err;
592 
593 	if (copy_from_user(&op, u, sizeof(op)) != 0)
594 		return -EFAULT;
595 	pr_debug("priv %p, add %d\n", priv, op.count);
596 	if (unlikely(op.count <= 0))
597 		return -EINVAL;
598 
599 	err = -ENOMEM;
600 	map = gntdev_alloc_map(priv, op.count);
601 	if (!map)
602 		return err;
603 
604 	if (unlikely(atomic_add_return(op.count, &pages_mapped) > limit)) {
605 		pr_debug("can't map: over limit\n");
606 		gntdev_put_map(NULL, map);
607 		return err;
608 	}
609 
610 	if (copy_from_user(map->grants, &u->refs,
611 			   sizeof(map->grants[0]) * op.count) != 0) {
612 		gntdev_put_map(NULL, map);
613 		return -EFAULT;
614 	}
615 
616 	mutex_lock(&priv->lock);
617 	gntdev_add_map(priv, map);
618 	op.index = map->index << PAGE_SHIFT;
619 	mutex_unlock(&priv->lock);
620 
621 	if (copy_to_user(u, &op, sizeof(op)) != 0)
622 		return -EFAULT;
623 
624 	return 0;
625 }
626 
627 static long gntdev_ioctl_unmap_grant_ref(struct gntdev_priv *priv,
628 					 struct ioctl_gntdev_unmap_grant_ref __user *u)
629 {
630 	struct ioctl_gntdev_unmap_grant_ref op;
631 	struct grant_map *map;
632 	int err = -ENOENT;
633 
634 	if (copy_from_user(&op, u, sizeof(op)) != 0)
635 		return -EFAULT;
636 	pr_debug("priv %p, del %d+%d\n", priv, (int)op.index, (int)op.count);
637 
638 	mutex_lock(&priv->lock);
639 	map = gntdev_find_map_index(priv, op.index >> PAGE_SHIFT, op.count);
640 	if (map) {
641 		list_del(&map->next);
642 		if (populate_freeable_maps)
643 			list_add_tail(&map->next, &priv->freeable_maps);
644 		err = 0;
645 	}
646 	mutex_unlock(&priv->lock);
647 	if (map)
648 		gntdev_put_map(priv, map);
649 	return err;
650 }
651 
652 static long gntdev_ioctl_get_offset_for_vaddr(struct gntdev_priv *priv,
653 					      struct ioctl_gntdev_get_offset_for_vaddr __user *u)
654 {
655 	struct ioctl_gntdev_get_offset_for_vaddr op;
656 	struct vm_area_struct *vma;
657 	struct grant_map *map;
658 	int rv = -EINVAL;
659 
660 	if (copy_from_user(&op, u, sizeof(op)) != 0)
661 		return -EFAULT;
662 	pr_debug("priv %p, offset for vaddr %lx\n", priv, (unsigned long)op.vaddr);
663 
664 	down_read(&current->mm->mmap_sem);
665 	vma = find_vma(current->mm, op.vaddr);
666 	if (!vma || vma->vm_ops != &gntdev_vmops)
667 		goto out_unlock;
668 
669 	map = vma->vm_private_data;
670 	if (!map)
671 		goto out_unlock;
672 
673 	op.offset = map->index << PAGE_SHIFT;
674 	op.count = map->count;
675 	rv = 0;
676 
677  out_unlock:
678 	up_read(&current->mm->mmap_sem);
679 
680 	if (rv == 0 && copy_to_user(u, &op, sizeof(op)) != 0)
681 		return -EFAULT;
682 	return rv;
683 }
684 
685 static long gntdev_ioctl_notify(struct gntdev_priv *priv, void __user *u)
686 {
687 	struct ioctl_gntdev_unmap_notify op;
688 	struct grant_map *map;
689 	int rc;
690 	int out_flags;
691 	unsigned int out_event;
692 
693 	if (copy_from_user(&op, u, sizeof(op)))
694 		return -EFAULT;
695 
696 	if (op.action & ~(UNMAP_NOTIFY_CLEAR_BYTE|UNMAP_NOTIFY_SEND_EVENT))
697 		return -EINVAL;
698 
699 	/* We need to grab a reference to the event channel we are going to use
700 	 * to send the notify before releasing the reference we may already have
701 	 * (if someone has called this ioctl twice). This is required so that
702 	 * it is possible to change the clear_byte part of the notification
703 	 * without disturbing the event channel part, which may now be the last
704 	 * reference to that event channel.
705 	 */
706 	if (op.action & UNMAP_NOTIFY_SEND_EVENT) {
707 		if (evtchn_get(op.event_channel_port))
708 			return -EINVAL;
709 	}
710 
711 	out_flags = op.action;
712 	out_event = op.event_channel_port;
713 
714 	mutex_lock(&priv->lock);
715 
716 	list_for_each_entry(map, &priv->maps, next) {
717 		uint64_t begin = map->index << PAGE_SHIFT;
718 		uint64_t end = (map->index + map->count) << PAGE_SHIFT;
719 		if (op.index >= begin && op.index < end)
720 			goto found;
721 	}
722 	rc = -ENOENT;
723 	goto unlock_out;
724 
725  found:
726 	if ((op.action & UNMAP_NOTIFY_CLEAR_BYTE) &&
727 			(map->flags & GNTMAP_readonly)) {
728 		rc = -EINVAL;
729 		goto unlock_out;
730 	}
731 
732 	out_flags = map->notify.flags;
733 	out_event = map->notify.event;
734 
735 	map->notify.flags = op.action;
736 	map->notify.addr = op.index - (map->index << PAGE_SHIFT);
737 	map->notify.event = op.event_channel_port;
738 
739 	rc = 0;
740 
741  unlock_out:
742 	mutex_unlock(&priv->lock);
743 
744 	/* Drop the reference to the event channel we did not save in the map */
745 	if (out_flags & UNMAP_NOTIFY_SEND_EVENT)
746 		evtchn_put(out_event);
747 
748 	return rc;
749 }
750 
751 #define GNTDEV_COPY_BATCH 16
752 
753 struct gntdev_copy_batch {
754 	struct gnttab_copy ops[GNTDEV_COPY_BATCH];
755 	struct page *pages[GNTDEV_COPY_BATCH];
756 	s16 __user *status[GNTDEV_COPY_BATCH];
757 	unsigned int nr_ops;
758 	unsigned int nr_pages;
759 };
760 
761 static int gntdev_get_page(struct gntdev_copy_batch *batch, void __user *virt,
762 			   bool writeable, unsigned long *gfn)
763 {
764 	unsigned long addr = (unsigned long)virt;
765 	struct page *page;
766 	unsigned long xen_pfn;
767 	int ret;
768 
769 	ret = get_user_pages_fast(addr, 1, writeable, &page);
770 	if (ret < 0)
771 		return ret;
772 
773 	batch->pages[batch->nr_pages++] = page;
774 
775 	xen_pfn = page_to_xen_pfn(page) + XEN_PFN_DOWN(addr & ~PAGE_MASK);
776 	*gfn = pfn_to_gfn(xen_pfn);
777 
778 	return 0;
779 }
780 
781 static void gntdev_put_pages(struct gntdev_copy_batch *batch)
782 {
783 	unsigned int i;
784 
785 	for (i = 0; i < batch->nr_pages; i++)
786 		put_page(batch->pages[i]);
787 	batch->nr_pages = 0;
788 }
789 
790 static int gntdev_copy(struct gntdev_copy_batch *batch)
791 {
792 	unsigned int i;
793 
794 	gnttab_batch_copy(batch->ops, batch->nr_ops);
795 	gntdev_put_pages(batch);
796 
797 	/*
798 	 * For each completed op, update the status if the op failed
799 	 * and all previous ops for the segment were successful.
800 	 */
801 	for (i = 0; i < batch->nr_ops; i++) {
802 		s16 status = batch->ops[i].status;
803 		s16 old_status;
804 
805 		if (status == GNTST_okay)
806 			continue;
807 
808 		if (__get_user(old_status, batch->status[i]))
809 			return -EFAULT;
810 
811 		if (old_status != GNTST_okay)
812 			continue;
813 
814 		if (__put_user(status, batch->status[i]))
815 			return -EFAULT;
816 	}
817 
818 	batch->nr_ops = 0;
819 	return 0;
820 }
821 
822 static int gntdev_grant_copy_seg(struct gntdev_copy_batch *batch,
823 				 struct gntdev_grant_copy_segment *seg,
824 				 s16 __user *status)
825 {
826 	uint16_t copied = 0;
827 
828 	/*
829 	 * Disallow local -> local copies since there is only space in
830 	 * batch->pages for one page per-op and this would be a very
831 	 * expensive memcpy().
832 	 */
833 	if (!(seg->flags & (GNTCOPY_source_gref | GNTCOPY_dest_gref)))
834 		return -EINVAL;
835 
836 	/* Can't cross page if source/dest is a grant ref. */
837 	if (seg->flags & GNTCOPY_source_gref) {
838 		if (seg->source.foreign.offset + seg->len > XEN_PAGE_SIZE)
839 			return -EINVAL;
840 	}
841 	if (seg->flags & GNTCOPY_dest_gref) {
842 		if (seg->dest.foreign.offset + seg->len > XEN_PAGE_SIZE)
843 			return -EINVAL;
844 	}
845 
846 	if (put_user(GNTST_okay, status))
847 		return -EFAULT;
848 
849 	while (copied < seg->len) {
850 		struct gnttab_copy *op;
851 		void __user *virt;
852 		size_t len, off;
853 		unsigned long gfn;
854 		int ret;
855 
856 		if (batch->nr_ops >= GNTDEV_COPY_BATCH) {
857 			ret = gntdev_copy(batch);
858 			if (ret < 0)
859 				return ret;
860 		}
861 
862 		len = seg->len - copied;
863 
864 		op = &batch->ops[batch->nr_ops];
865 		op->flags = 0;
866 
867 		if (seg->flags & GNTCOPY_source_gref) {
868 			op->source.u.ref = seg->source.foreign.ref;
869 			op->source.domid = seg->source.foreign.domid;
870 			op->source.offset = seg->source.foreign.offset + copied;
871 			op->flags |= GNTCOPY_source_gref;
872 		} else {
873 			virt = seg->source.virt + copied;
874 			off = (unsigned long)virt & ~XEN_PAGE_MASK;
875 			len = min(len, (size_t)XEN_PAGE_SIZE - off);
876 
877 			ret = gntdev_get_page(batch, virt, false, &gfn);
878 			if (ret < 0)
879 				return ret;
880 
881 			op->source.u.gmfn = gfn;
882 			op->source.domid = DOMID_SELF;
883 			op->source.offset = off;
884 		}
885 
886 		if (seg->flags & GNTCOPY_dest_gref) {
887 			op->dest.u.ref = seg->dest.foreign.ref;
888 			op->dest.domid = seg->dest.foreign.domid;
889 			op->dest.offset = seg->dest.foreign.offset + copied;
890 			op->flags |= GNTCOPY_dest_gref;
891 		} else {
892 			virt = seg->dest.virt + copied;
893 			off = (unsigned long)virt & ~XEN_PAGE_MASK;
894 			len = min(len, (size_t)XEN_PAGE_SIZE - off);
895 
896 			ret = gntdev_get_page(batch, virt, true, &gfn);
897 			if (ret < 0)
898 				return ret;
899 
900 			op->dest.u.gmfn = gfn;
901 			op->dest.domid = DOMID_SELF;
902 			op->dest.offset = off;
903 		}
904 
905 		op->len = len;
906 		copied += len;
907 
908 		batch->status[batch->nr_ops] = status;
909 		batch->nr_ops++;
910 	}
911 
912 	return 0;
913 }
914 
915 static long gntdev_ioctl_grant_copy(struct gntdev_priv *priv, void __user *u)
916 {
917 	struct ioctl_gntdev_grant_copy copy;
918 	struct gntdev_copy_batch batch;
919 	unsigned int i;
920 	int ret = 0;
921 
922 	if (copy_from_user(&copy, u, sizeof(copy)))
923 		return -EFAULT;
924 
925 	batch.nr_ops = 0;
926 	batch.nr_pages = 0;
927 
928 	for (i = 0; i < copy.count; i++) {
929 		struct gntdev_grant_copy_segment seg;
930 
931 		if (copy_from_user(&seg, &copy.segments[i], sizeof(seg))) {
932 			ret = -EFAULT;
933 			goto out;
934 		}
935 
936 		ret = gntdev_grant_copy_seg(&batch, &seg, &copy.segments[i].status);
937 		if (ret < 0)
938 			goto out;
939 
940 		cond_resched();
941 	}
942 	if (batch.nr_ops)
943 		ret = gntdev_copy(&batch);
944 	return ret;
945 
946   out:
947 	gntdev_put_pages(&batch);
948 	return ret;
949 }
950 
951 static long gntdev_ioctl(struct file *flip,
952 			 unsigned int cmd, unsigned long arg)
953 {
954 	struct gntdev_priv *priv = flip->private_data;
955 	void __user *ptr = (void __user *)arg;
956 
957 	switch (cmd) {
958 	case IOCTL_GNTDEV_MAP_GRANT_REF:
959 		return gntdev_ioctl_map_grant_ref(priv, ptr);
960 
961 	case IOCTL_GNTDEV_UNMAP_GRANT_REF:
962 		return gntdev_ioctl_unmap_grant_ref(priv, ptr);
963 
964 	case IOCTL_GNTDEV_GET_OFFSET_FOR_VADDR:
965 		return gntdev_ioctl_get_offset_for_vaddr(priv, ptr);
966 
967 	case IOCTL_GNTDEV_SET_UNMAP_NOTIFY:
968 		return gntdev_ioctl_notify(priv, ptr);
969 
970 	case IOCTL_GNTDEV_GRANT_COPY:
971 		return gntdev_ioctl_grant_copy(priv, ptr);
972 
973 	default:
974 		pr_debug("priv %p, unknown cmd %x\n", priv, cmd);
975 		return -ENOIOCTLCMD;
976 	}
977 
978 	return 0;
979 }
980 
981 static int gntdev_mmap(struct file *flip, struct vm_area_struct *vma)
982 {
983 	struct gntdev_priv *priv = flip->private_data;
984 	int index = vma->vm_pgoff;
985 	int count = vma_pages(vma);
986 	struct grant_map *map;
987 	int i, err = -EINVAL;
988 
989 	if ((vma->vm_flags & VM_WRITE) && !(vma->vm_flags & VM_SHARED))
990 		return -EINVAL;
991 
992 	pr_debug("map %d+%d at %lx (pgoff %lx)\n",
993 			index, count, vma->vm_start, vma->vm_pgoff);
994 
995 	mutex_lock(&priv->lock);
996 	map = gntdev_find_map_index(priv, index, count);
997 	if (!map)
998 		goto unlock_out;
999 	if (use_ptemod && map->vma)
1000 		goto unlock_out;
1001 	if (use_ptemod && priv->mm != vma->vm_mm) {
1002 		pr_warn("Huh? Other mm?\n");
1003 		goto unlock_out;
1004 	}
1005 
1006 	atomic_inc(&map->users);
1007 
1008 	vma->vm_ops = &gntdev_vmops;
1009 
1010 	vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP | VM_IO;
1011 
1012 	if (use_ptemod)
1013 		vma->vm_flags |= VM_DONTCOPY;
1014 
1015 	vma->vm_private_data = map;
1016 
1017 	if (use_ptemod)
1018 		map->vma = vma;
1019 
1020 	if (map->flags) {
1021 		if ((vma->vm_flags & VM_WRITE) &&
1022 				(map->flags & GNTMAP_readonly))
1023 			goto out_unlock_put;
1024 	} else {
1025 		map->flags = GNTMAP_host_map;
1026 		if (!(vma->vm_flags & VM_WRITE))
1027 			map->flags |= GNTMAP_readonly;
1028 	}
1029 
1030 	mutex_unlock(&priv->lock);
1031 
1032 	if (use_ptemod) {
1033 		err = apply_to_page_range(vma->vm_mm, vma->vm_start,
1034 					  vma->vm_end - vma->vm_start,
1035 					  find_grant_ptes, map);
1036 		if (err) {
1037 			pr_warn("find_grant_ptes() failure.\n");
1038 			goto out_put_map;
1039 		}
1040 	}
1041 
1042 	err = map_grant_pages(map);
1043 	if (err)
1044 		goto out_put_map;
1045 
1046 	if (!use_ptemod) {
1047 		for (i = 0; i < count; i++) {
1048 			err = vm_insert_page(vma, vma->vm_start + i*PAGE_SIZE,
1049 				map->pages[i]);
1050 			if (err)
1051 				goto out_put_map;
1052 		}
1053 	} else {
1054 #ifdef CONFIG_X86
1055 		/*
1056 		 * If the PTEs were not made special by the grant map
1057 		 * hypercall, do so here.
1058 		 *
1059 		 * This is racy since the mapping is already visible
1060 		 * to userspace but userspace should be well-behaved
1061 		 * enough to not touch it until the mmap() call
1062 		 * returns.
1063 		 */
1064 		if (!xen_feature(XENFEAT_gnttab_map_avail_bits)) {
1065 			apply_to_page_range(vma->vm_mm, vma->vm_start,
1066 					    vma->vm_end - vma->vm_start,
1067 					    set_grant_ptes_as_special, NULL);
1068 		}
1069 #endif
1070 		map->pages_vm_start = vma->vm_start;
1071 	}
1072 
1073 	return 0;
1074 
1075 unlock_out:
1076 	mutex_unlock(&priv->lock);
1077 	return err;
1078 
1079 out_unlock_put:
1080 	mutex_unlock(&priv->lock);
1081 out_put_map:
1082 	if (use_ptemod)
1083 		map->vma = NULL;
1084 	gntdev_put_map(priv, map);
1085 	return err;
1086 }
1087 
1088 static const struct file_operations gntdev_fops = {
1089 	.owner = THIS_MODULE,
1090 	.open = gntdev_open,
1091 	.release = gntdev_release,
1092 	.mmap = gntdev_mmap,
1093 	.unlocked_ioctl = gntdev_ioctl
1094 };
1095 
1096 static struct miscdevice gntdev_miscdev = {
1097 	.minor        = MISC_DYNAMIC_MINOR,
1098 	.name         = "xen/gntdev",
1099 	.fops         = &gntdev_fops,
1100 };
1101 
1102 /* ------------------------------------------------------------------ */
1103 
1104 static int __init gntdev_init(void)
1105 {
1106 	int err;
1107 
1108 	if (!xen_domain())
1109 		return -ENODEV;
1110 
1111 	use_ptemod = !xen_feature(XENFEAT_auto_translated_physmap);
1112 
1113 	err = misc_register(&gntdev_miscdev);
1114 	if (err != 0) {
1115 		pr_err("Could not register gntdev device\n");
1116 		return err;
1117 	}
1118 	return 0;
1119 }
1120 
1121 static void __exit gntdev_exit(void)
1122 {
1123 	misc_deregister(&gntdev_miscdev);
1124 }
1125 
1126 module_init(gntdev_init);
1127 module_exit(gntdev_exit);
1128 
1129 /* ------------------------------------------------------------------ */
1130