Loading...
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#include <linux/module.h>
23#include <linux/kernel.h>
24#include <linux/init.h>
25#include <linux/miscdevice.h>
26#include <linux/fs.h>
27#include <linux/mm.h>
28#include <linux/mman.h>
29#include <linux/mmu_notifier.h>
30#include <linux/types.h>
31#include <linux/uaccess.h>
32#include <linux/sched.h>
33#include <linux/spinlock.h>
34#include <linux/slab.h>
35#include <linux/highmem.h>
36
37#include <xen/xen.h>
38#include <xen/grant_table.h>
39#include <xen/balloon.h>
40#include <xen/gntdev.h>
41#include <xen/events.h>
42#include <asm/xen/hypervisor.h>
43#include <asm/xen/hypercall.h>
44#include <asm/xen/page.h>
45
46MODULE_LICENSE("GPL");
47MODULE_AUTHOR("Derek G. Murray <Derek.Murray@cl.cam.ac.uk>, "
48 "Gerd Hoffmann <kraxel@redhat.com>");
49MODULE_DESCRIPTION("User-space granted page access driver");
50
51static int limit = 1024*1024;
52module_param(limit, int, 0644);
53MODULE_PARM_DESC(limit, "Maximum number of grants that may be mapped by "
54 "the gntdev device");
55
56static atomic_t pages_mapped = ATOMIC_INIT(0);
57
58static int use_ptemod;
59
60struct gntdev_priv {
61 struct list_head maps;
62 /* lock protects maps from concurrent changes */
63 spinlock_t lock;
64 struct mm_struct *mm;
65 struct mmu_notifier mn;
66};
67
68struct unmap_notify {
69 int flags;
70 /* Address relative to the start of the grant_map */
71 int addr;
72 int event;
73};
74
75struct grant_map {
76 struct list_head next;
77 struct vm_area_struct *vma;
78 int index;
79 int count;
80 int flags;
81 atomic_t users;
82 struct unmap_notify notify;
83 struct ioctl_gntdev_grant_ref *grants;
84 struct gnttab_map_grant_ref *map_ops;
85 struct gnttab_unmap_grant_ref *unmap_ops;
86 struct page **pages;
87};
88
89static int unmap_grant_pages(struct grant_map *map, int offset, int pages);
90
91/* ------------------------------------------------------------------ */
92
93static void gntdev_print_maps(struct gntdev_priv *priv,
94 char *text, int text_index)
95{
96#ifdef DEBUG
97 struct grant_map *map;
98
99 pr_debug("%s: maps list (priv %p)\n", __func__, priv);
100 list_for_each_entry(map, &priv->maps, next)
101 pr_debug(" index %2d, count %2d %s\n",
102 map->index, map->count,
103 map->index == text_index && text ? text : "");
104#endif
105}
106
107static struct grant_map *gntdev_alloc_map(struct gntdev_priv *priv, int count)
108{
109 struct grant_map *add;
110 int i;
111
112 add = kzalloc(sizeof(struct grant_map), GFP_KERNEL);
113 if (NULL == add)
114 return NULL;
115
116 add->grants = kzalloc(sizeof(add->grants[0]) * count, GFP_KERNEL);
117 add->map_ops = kzalloc(sizeof(add->map_ops[0]) * count, GFP_KERNEL);
118 add->unmap_ops = kzalloc(sizeof(add->unmap_ops[0]) * count, GFP_KERNEL);
119 add->pages = kzalloc(sizeof(add->pages[0]) * count, GFP_KERNEL);
120 if (NULL == add->grants ||
121 NULL == add->map_ops ||
122 NULL == add->unmap_ops ||
123 NULL == add->pages)
124 goto err;
125
126 if (alloc_xenballooned_pages(count, add->pages))
127 goto err;
128
129 for (i = 0; i < count; i++) {
130 add->map_ops[i].handle = -1;
131 add->unmap_ops[i].handle = -1;
132 }
133
134 add->index = 0;
135 add->count = count;
136 atomic_set(&add->users, 1);
137
138 return add;
139
140err:
141 kfree(add->pages);
142 kfree(add->grants);
143 kfree(add->map_ops);
144 kfree(add->unmap_ops);
145 kfree(add);
146 return NULL;
147}
148
149static void gntdev_add_map(struct gntdev_priv *priv, struct grant_map *add)
150{
151 struct grant_map *map;
152
153 list_for_each_entry(map, &priv->maps, next) {
154 if (add->index + add->count < map->index) {
155 list_add_tail(&add->next, &map->next);
156 goto done;
157 }
158 add->index = map->index + map->count;
159 }
160 list_add_tail(&add->next, &priv->maps);
161
162done:
163 gntdev_print_maps(priv, "[new]", add->index);
164}
165
166static struct grant_map *gntdev_find_map_index(struct gntdev_priv *priv,
167 int index, int count)
168{
169 struct grant_map *map;
170
171 list_for_each_entry(map, &priv->maps, next) {
172 if (map->index != index)
173 continue;
174 if (count && map->count != count)
175 continue;
176 return map;
177 }
178 return NULL;
179}
180
181static void gntdev_put_map(struct grant_map *map)
182{
183 if (!map)
184 return;
185
186 if (!atomic_dec_and_test(&map->users))
187 return;
188
189 atomic_sub(map->count, &pages_mapped);
190
191 if (map->notify.flags & UNMAP_NOTIFY_SEND_EVENT) {
192 notify_remote_via_evtchn(map->notify.event);
193 }
194
195 if (map->pages) {
196 if (!use_ptemod)
197 unmap_grant_pages(map, 0, map->count);
198
199 free_xenballooned_pages(map->count, map->pages);
200 }
201 kfree(map->pages);
202 kfree(map->grants);
203 kfree(map->map_ops);
204 kfree(map->unmap_ops);
205 kfree(map);
206}
207
208/* ------------------------------------------------------------------ */
209
210static int find_grant_ptes(pte_t *pte, pgtable_t token,
211 unsigned long addr, void *data)
212{
213 struct grant_map *map = data;
214 unsigned int pgnr = (addr - map->vma->vm_start) >> PAGE_SHIFT;
215 int flags = map->flags | GNTMAP_application_map | GNTMAP_contains_pte;
216 u64 pte_maddr;
217
218 BUG_ON(pgnr >= map->count);
219 pte_maddr = arbitrary_virt_to_machine(pte).maddr;
220
221 gnttab_set_map_op(&map->map_ops[pgnr], pte_maddr, flags,
222 map->grants[pgnr].ref,
223 map->grants[pgnr].domid);
224 gnttab_set_unmap_op(&map->unmap_ops[pgnr], pte_maddr, flags,
225 -1 /* handle */);
226 return 0;
227}
228
229static int map_grant_pages(struct grant_map *map)
230{
231 int i, err = 0;
232
233 if (!use_ptemod) {
234 /* Note: it could already be mapped */
235 if (map->map_ops[0].handle != -1)
236 return 0;
237 for (i = 0; i < map->count; i++) {
238 unsigned long addr = (unsigned long)
239 pfn_to_kaddr(page_to_pfn(map->pages[i]));
240 gnttab_set_map_op(&map->map_ops[i], addr, map->flags,
241 map->grants[i].ref,
242 map->grants[i].domid);
243 gnttab_set_unmap_op(&map->unmap_ops[i], addr,
244 map->flags, -1 /* handle */);
245 }
246 }
247
248 pr_debug("map %d+%d\n", map->index, map->count);
249 err = gnttab_map_refs(map->map_ops, map->pages, map->count);
250 if (err)
251 return err;
252
253 for (i = 0; i < map->count; i++) {
254 if (map->map_ops[i].status)
255 err = -EINVAL;
256 else {
257 BUG_ON(map->map_ops[i].handle == -1);
258 map->unmap_ops[i].handle = map->map_ops[i].handle;
259 pr_debug("map handle=%d\n", map->map_ops[i].handle);
260 }
261 }
262 return err;
263}
264
265static int __unmap_grant_pages(struct grant_map *map, int offset, int pages)
266{
267 int i, err = 0;
268
269 if (map->notify.flags & UNMAP_NOTIFY_CLEAR_BYTE) {
270 int pgno = (map->notify.addr >> PAGE_SHIFT);
271 if (pgno >= offset && pgno < offset + pages && use_ptemod) {
272 void __user *tmp = (void __user *)
273 map->vma->vm_start + map->notify.addr;
274 err = copy_to_user(tmp, &err, 1);
275 if (err)
276 return -EFAULT;
277 map->notify.flags &= ~UNMAP_NOTIFY_CLEAR_BYTE;
278 } else if (pgno >= offset && pgno < offset + pages) {
279 uint8_t *tmp = kmap(map->pages[pgno]);
280 tmp[map->notify.addr & (PAGE_SIZE-1)] = 0;
281 kunmap(map->pages[pgno]);
282 map->notify.flags &= ~UNMAP_NOTIFY_CLEAR_BYTE;
283 }
284 }
285
286 err = gnttab_unmap_refs(map->unmap_ops + offset, map->pages + offset, pages);
287 if (err)
288 return err;
289
290 for (i = 0; i < pages; i++) {
291 if (map->unmap_ops[offset+i].status)
292 err = -EINVAL;
293 pr_debug("unmap handle=%d st=%d\n",
294 map->unmap_ops[offset+i].handle,
295 map->unmap_ops[offset+i].status);
296 map->unmap_ops[offset+i].handle = -1;
297 }
298 return err;
299}
300
301static int unmap_grant_pages(struct grant_map *map, int offset, int pages)
302{
303 int range, err = 0;
304
305 pr_debug("unmap %d+%d [%d+%d]\n", map->index, map->count, offset, pages);
306
307 /* It is possible the requested range will have a "hole" where we
308 * already unmapped some of the grants. Only unmap valid ranges.
309 */
310 while (pages && !err) {
311 while (pages && map->unmap_ops[offset].handle == -1) {
312 offset++;
313 pages--;
314 }
315 range = 0;
316 while (range < pages) {
317 if (map->unmap_ops[offset+range].handle == -1) {
318 range--;
319 break;
320 }
321 range++;
322 }
323 err = __unmap_grant_pages(map, offset, range);
324 offset += range;
325 pages -= range;
326 }
327
328 return err;
329}
330
331/* ------------------------------------------------------------------ */
332
333static void gntdev_vma_open(struct vm_area_struct *vma)
334{
335 struct grant_map *map = vma->vm_private_data;
336
337 pr_debug("gntdev_vma_open %p\n", vma);
338 atomic_inc(&map->users);
339}
340
341static void gntdev_vma_close(struct vm_area_struct *vma)
342{
343 struct grant_map *map = vma->vm_private_data;
344
345 pr_debug("gntdev_vma_close %p\n", vma);
346 map->vma = NULL;
347 vma->vm_private_data = NULL;
348 gntdev_put_map(map);
349}
350
351static struct vm_operations_struct gntdev_vmops = {
352 .open = gntdev_vma_open,
353 .close = gntdev_vma_close,
354};
355
356/* ------------------------------------------------------------------ */
357
358static void mn_invl_range_start(struct mmu_notifier *mn,
359 struct mm_struct *mm,
360 unsigned long start, unsigned long end)
361{
362 struct gntdev_priv *priv = container_of(mn, struct gntdev_priv, mn);
363 struct grant_map *map;
364 unsigned long mstart, mend;
365 int err;
366
367 spin_lock(&priv->lock);
368 list_for_each_entry(map, &priv->maps, next) {
369 if (!map->vma)
370 continue;
371 if (map->vma->vm_start >= end)
372 continue;
373 if (map->vma->vm_end <= start)
374 continue;
375 mstart = max(start, map->vma->vm_start);
376 mend = min(end, map->vma->vm_end);
377 pr_debug("map %d+%d (%lx %lx), range %lx %lx, mrange %lx %lx\n",
378 map->index, map->count,
379 map->vma->vm_start, map->vma->vm_end,
380 start, end, mstart, mend);
381 err = unmap_grant_pages(map,
382 (mstart - map->vma->vm_start) >> PAGE_SHIFT,
383 (mend - mstart) >> PAGE_SHIFT);
384 WARN_ON(err);
385 }
386 spin_unlock(&priv->lock);
387}
388
389static void mn_invl_page(struct mmu_notifier *mn,
390 struct mm_struct *mm,
391 unsigned long address)
392{
393 mn_invl_range_start(mn, mm, address, address + PAGE_SIZE);
394}
395
396static void mn_release(struct mmu_notifier *mn,
397 struct mm_struct *mm)
398{
399 struct gntdev_priv *priv = container_of(mn, struct gntdev_priv, mn);
400 struct grant_map *map;
401 int err;
402
403 spin_lock(&priv->lock);
404 list_for_each_entry(map, &priv->maps, next) {
405 if (!map->vma)
406 continue;
407 pr_debug("map %d+%d (%lx %lx)\n",
408 map->index, map->count,
409 map->vma->vm_start, map->vma->vm_end);
410 err = unmap_grant_pages(map, /* offset */ 0, map->count);
411 WARN_ON(err);
412 }
413 spin_unlock(&priv->lock);
414}
415
416struct mmu_notifier_ops gntdev_mmu_ops = {
417 .release = mn_release,
418 .invalidate_page = mn_invl_page,
419 .invalidate_range_start = mn_invl_range_start,
420};
421
422/* ------------------------------------------------------------------ */
423
424static int gntdev_open(struct inode *inode, struct file *flip)
425{
426 struct gntdev_priv *priv;
427 int ret = 0;
428
429 priv = kzalloc(sizeof(*priv), GFP_KERNEL);
430 if (!priv)
431 return -ENOMEM;
432
433 INIT_LIST_HEAD(&priv->maps);
434 spin_lock_init(&priv->lock);
435
436 if (use_ptemod) {
437 priv->mm = get_task_mm(current);
438 if (!priv->mm) {
439 kfree(priv);
440 return -ENOMEM;
441 }
442 priv->mn.ops = &gntdev_mmu_ops;
443 ret = mmu_notifier_register(&priv->mn, priv->mm);
444 mmput(priv->mm);
445 }
446
447 if (ret) {
448 kfree(priv);
449 return ret;
450 }
451
452 flip->private_data = priv;
453 pr_debug("priv %p\n", priv);
454
455 return 0;
456}
457
458static int gntdev_release(struct inode *inode, struct file *flip)
459{
460 struct gntdev_priv *priv = flip->private_data;
461 struct grant_map *map;
462
463 pr_debug("priv %p\n", priv);
464
465 spin_lock(&priv->lock);
466 while (!list_empty(&priv->maps)) {
467 map = list_entry(priv->maps.next, struct grant_map, next);
468 list_del(&map->next);
469 gntdev_put_map(map);
470 }
471 spin_unlock(&priv->lock);
472
473 if (use_ptemod)
474 mmu_notifier_unregister(&priv->mn, priv->mm);
475 kfree(priv);
476 return 0;
477}
478
479static long gntdev_ioctl_map_grant_ref(struct gntdev_priv *priv,
480 struct ioctl_gntdev_map_grant_ref __user *u)
481{
482 struct ioctl_gntdev_map_grant_ref op;
483 struct grant_map *map;
484 int err;
485
486 if (copy_from_user(&op, u, sizeof(op)) != 0)
487 return -EFAULT;
488 pr_debug("priv %p, add %d\n", priv, op.count);
489 if (unlikely(op.count <= 0))
490 return -EINVAL;
491
492 err = -ENOMEM;
493 map = gntdev_alloc_map(priv, op.count);
494 if (!map)
495 return err;
496
497 if (unlikely(atomic_add_return(op.count, &pages_mapped) > limit)) {
498 pr_debug("can't map: over limit\n");
499 gntdev_put_map(map);
500 return err;
501 }
502
503 if (copy_from_user(map->grants, &u->refs,
504 sizeof(map->grants[0]) * op.count) != 0) {
505 gntdev_put_map(map);
506 return err;
507 }
508
509 spin_lock(&priv->lock);
510 gntdev_add_map(priv, map);
511 op.index = map->index << PAGE_SHIFT;
512 spin_unlock(&priv->lock);
513
514 if (copy_to_user(u, &op, sizeof(op)) != 0)
515 return -EFAULT;
516
517 return 0;
518}
519
520static long gntdev_ioctl_unmap_grant_ref(struct gntdev_priv *priv,
521 struct ioctl_gntdev_unmap_grant_ref __user *u)
522{
523 struct ioctl_gntdev_unmap_grant_ref op;
524 struct grant_map *map;
525 int err = -ENOENT;
526
527 if (copy_from_user(&op, u, sizeof(op)) != 0)
528 return -EFAULT;
529 pr_debug("priv %p, del %d+%d\n", priv, (int)op.index, (int)op.count);
530
531 spin_lock(&priv->lock);
532 map = gntdev_find_map_index(priv, op.index >> PAGE_SHIFT, op.count);
533 if (map) {
534 list_del(&map->next);
535 gntdev_put_map(map);
536 err = 0;
537 }
538 spin_unlock(&priv->lock);
539 return err;
540}
541
542static long gntdev_ioctl_get_offset_for_vaddr(struct gntdev_priv *priv,
543 struct ioctl_gntdev_get_offset_for_vaddr __user *u)
544{
545 struct ioctl_gntdev_get_offset_for_vaddr op;
546 struct vm_area_struct *vma;
547 struct grant_map *map;
548
549 if (copy_from_user(&op, u, sizeof(op)) != 0)
550 return -EFAULT;
551 pr_debug("priv %p, offset for vaddr %lx\n", priv, (unsigned long)op.vaddr);
552
553 vma = find_vma(current->mm, op.vaddr);
554 if (!vma || vma->vm_ops != &gntdev_vmops)
555 return -EINVAL;
556
557 map = vma->vm_private_data;
558 if (!map)
559 return -EINVAL;
560
561 op.offset = map->index << PAGE_SHIFT;
562 op.count = map->count;
563
564 if (copy_to_user(u, &op, sizeof(op)) != 0)
565 return -EFAULT;
566 return 0;
567}
568
569static long gntdev_ioctl_notify(struct gntdev_priv *priv, void __user *u)
570{
571 struct ioctl_gntdev_unmap_notify op;
572 struct grant_map *map;
573 int rc;
574
575 if (copy_from_user(&op, u, sizeof(op)))
576 return -EFAULT;
577
578 if (op.action & ~(UNMAP_NOTIFY_CLEAR_BYTE|UNMAP_NOTIFY_SEND_EVENT))
579 return -EINVAL;
580
581 spin_lock(&priv->lock);
582
583 list_for_each_entry(map, &priv->maps, next) {
584 uint64_t begin = map->index << PAGE_SHIFT;
585 uint64_t end = (map->index + map->count) << PAGE_SHIFT;
586 if (op.index >= begin && op.index < end)
587 goto found;
588 }
589 rc = -ENOENT;
590 goto unlock_out;
591
592 found:
593 if ((op.action & UNMAP_NOTIFY_CLEAR_BYTE) &&
594 (map->flags & GNTMAP_readonly)) {
595 rc = -EINVAL;
596 goto unlock_out;
597 }
598
599 map->notify.flags = op.action;
600 map->notify.addr = op.index - (map->index << PAGE_SHIFT);
601 map->notify.event = op.event_channel_port;
602 rc = 0;
603 unlock_out:
604 spin_unlock(&priv->lock);
605 return rc;
606}
607
608static long gntdev_ioctl(struct file *flip,
609 unsigned int cmd, unsigned long arg)
610{
611 struct gntdev_priv *priv = flip->private_data;
612 void __user *ptr = (void __user *)arg;
613
614 switch (cmd) {
615 case IOCTL_GNTDEV_MAP_GRANT_REF:
616 return gntdev_ioctl_map_grant_ref(priv, ptr);
617
618 case IOCTL_GNTDEV_UNMAP_GRANT_REF:
619 return gntdev_ioctl_unmap_grant_ref(priv, ptr);
620
621 case IOCTL_GNTDEV_GET_OFFSET_FOR_VADDR:
622 return gntdev_ioctl_get_offset_for_vaddr(priv, ptr);
623
624 case IOCTL_GNTDEV_SET_UNMAP_NOTIFY:
625 return gntdev_ioctl_notify(priv, ptr);
626
627 default:
628 pr_debug("priv %p, unknown cmd %x\n", priv, cmd);
629 return -ENOIOCTLCMD;
630 }
631
632 return 0;
633}
634
635static int gntdev_mmap(struct file *flip, struct vm_area_struct *vma)
636{
637 struct gntdev_priv *priv = flip->private_data;
638 int index = vma->vm_pgoff;
639 int count = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
640 struct grant_map *map;
641 int i, err = -EINVAL;
642
643 if ((vma->vm_flags & VM_WRITE) && !(vma->vm_flags & VM_SHARED))
644 return -EINVAL;
645
646 pr_debug("map %d+%d at %lx (pgoff %lx)\n",
647 index, count, vma->vm_start, vma->vm_pgoff);
648
649 spin_lock(&priv->lock);
650 map = gntdev_find_map_index(priv, index, count);
651 if (!map)
652 goto unlock_out;
653 if (use_ptemod && map->vma)
654 goto unlock_out;
655 if (use_ptemod && priv->mm != vma->vm_mm) {
656 printk(KERN_WARNING "Huh? Other mm?\n");
657 goto unlock_out;
658 }
659
660 atomic_inc(&map->users);
661
662 vma->vm_ops = &gntdev_vmops;
663
664 vma->vm_flags |= VM_RESERVED|VM_DONTEXPAND;
665
666 if (use_ptemod)
667 vma->vm_flags |= VM_DONTCOPY|VM_PFNMAP;
668
669 vma->vm_private_data = map;
670
671 if (use_ptemod)
672 map->vma = vma;
673
674 if (map->flags) {
675 if ((vma->vm_flags & VM_WRITE) &&
676 (map->flags & GNTMAP_readonly))
677 goto out_unlock_put;
678 } else {
679 map->flags = GNTMAP_host_map;
680 if (!(vma->vm_flags & VM_WRITE))
681 map->flags |= GNTMAP_readonly;
682 }
683
684 spin_unlock(&priv->lock);
685
686 if (use_ptemod) {
687 err = apply_to_page_range(vma->vm_mm, vma->vm_start,
688 vma->vm_end - vma->vm_start,
689 find_grant_ptes, map);
690 if (err) {
691 printk(KERN_WARNING "find_grant_ptes() failure.\n");
692 goto out_put_map;
693 }
694 }
695
696 err = map_grant_pages(map);
697 if (err)
698 goto out_put_map;
699
700 if (!use_ptemod) {
701 for (i = 0; i < count; i++) {
702 err = vm_insert_page(vma, vma->vm_start + i*PAGE_SIZE,
703 map->pages[i]);
704 if (err)
705 goto out_put_map;
706 }
707 }
708
709 return 0;
710
711unlock_out:
712 spin_unlock(&priv->lock);
713 return err;
714
715out_unlock_put:
716 spin_unlock(&priv->lock);
717out_put_map:
718 if (use_ptemod)
719 map->vma = NULL;
720 gntdev_put_map(map);
721 return err;
722}
723
724static const struct file_operations gntdev_fops = {
725 .owner = THIS_MODULE,
726 .open = gntdev_open,
727 .release = gntdev_release,
728 .mmap = gntdev_mmap,
729 .unlocked_ioctl = gntdev_ioctl
730};
731
732static struct miscdevice gntdev_miscdev = {
733 .minor = MISC_DYNAMIC_MINOR,
734 .name = "xen/gntdev",
735 .fops = &gntdev_fops,
736};
737
738/* ------------------------------------------------------------------ */
739
740static int __init gntdev_init(void)
741{
742 int err;
743
744 if (!xen_domain())
745 return -ENODEV;
746
747 use_ptemod = xen_pv_domain();
748
749 err = misc_register(&gntdev_miscdev);
750 if (err != 0) {
751 printk(KERN_ERR "Could not register gntdev device\n");
752 return err;
753 }
754 return 0;
755}
756
757static void __exit gntdev_exit(void)
758{
759 misc_deregister(&gntdev_miscdev);
760}
761
762module_init(gntdev_init);
763module_exit(gntdev_exit);
764
765/* ------------------------------------------------------------------ */
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 * (c) 2018 Oleksandr Andrushchenko, EPAM Systems Inc.
10 *
11 * This program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
15 *
16 * You should have received a copy of the GNU General Public License
17 * along with this program; if not, write to the Free Software
18 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
19 */
20
21#undef DEBUG
22
23#define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt
24
25#include <linux/dma-mapping.h>
26#include <linux/module.h>
27#include <linux/kernel.h>
28#include <linux/init.h>
29#include <linux/miscdevice.h>
30#include <linux/fs.h>
31#include <linux/uaccess.h>
32#include <linux/sched.h>
33#include <linux/sched/mm.h>
34#include <linux/spinlock.h>
35#include <linux/slab.h>
36#include <linux/highmem.h>
37#include <linux/refcount.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#include "gntdev-common.h"
49#ifdef CONFIG_XEN_GNTDEV_DMABUF
50#include "gntdev-dmabuf.h"
51#endif
52
53MODULE_LICENSE("GPL");
54MODULE_AUTHOR("Derek G. Murray <Derek.Murray@cl.cam.ac.uk>, "
55 "Gerd Hoffmann <kraxel@redhat.com>");
56MODULE_DESCRIPTION("User-space granted page access driver");
57
58static unsigned int limit = 64*1024;
59module_param(limit, uint, 0644);
60MODULE_PARM_DESC(limit,
61 "Maximum number of grants that may be mapped by one mapping request");
62
63static int use_ptemod;
64
65static int unmap_grant_pages(struct gntdev_grant_map *map,
66 int offset, int pages);
67
68static struct miscdevice gntdev_miscdev;
69
70/* ------------------------------------------------------------------ */
71
72bool gntdev_test_page_count(unsigned int count)
73{
74 return !count || count > limit;
75}
76
77static void gntdev_print_maps(struct gntdev_priv *priv,
78 char *text, int text_index)
79{
80#ifdef DEBUG
81 struct gntdev_grant_map *map;
82
83 pr_debug("%s: maps list (priv %p)\n", __func__, priv);
84 list_for_each_entry(map, &priv->maps, next)
85 pr_debug(" index %2d, count %2d %s\n",
86 map->index, map->count,
87 map->index == text_index && text ? text : "");
88#endif
89}
90
91static void gntdev_free_map(struct gntdev_grant_map *map)
92{
93 if (map == NULL)
94 return;
95
96#ifdef CONFIG_XEN_GRANT_DMA_ALLOC
97 if (map->dma_vaddr) {
98 struct gnttab_dma_alloc_args args;
99
100 args.dev = map->dma_dev;
101 args.coherent = !!(map->dma_flags & GNTDEV_DMA_FLAG_COHERENT);
102 args.nr_pages = map->count;
103 args.pages = map->pages;
104 args.frames = map->frames;
105 args.vaddr = map->dma_vaddr;
106 args.dev_bus_addr = map->dma_bus_addr;
107
108 gnttab_dma_free_pages(&args);
109 } else
110#endif
111 if (map->pages)
112 gnttab_free_pages(map->count, map->pages);
113
114#ifdef CONFIG_XEN_GRANT_DMA_ALLOC
115 kvfree(map->frames);
116#endif
117 kvfree(map->pages);
118 kvfree(map->grants);
119 kvfree(map->map_ops);
120 kvfree(map->unmap_ops);
121 kvfree(map->kmap_ops);
122 kvfree(map->kunmap_ops);
123 kfree(map);
124}
125
126struct gntdev_grant_map *gntdev_alloc_map(struct gntdev_priv *priv, int count,
127 int dma_flags)
128{
129 struct gntdev_grant_map *add;
130 int i;
131
132 add = kzalloc(sizeof(*add), GFP_KERNEL);
133 if (NULL == add)
134 return NULL;
135
136 add->grants = kvmalloc_array(count, sizeof(add->grants[0]),
137 GFP_KERNEL);
138 add->map_ops = kvmalloc_array(count, sizeof(add->map_ops[0]),
139 GFP_KERNEL);
140 add->unmap_ops = kvmalloc_array(count, sizeof(add->unmap_ops[0]),
141 GFP_KERNEL);
142 add->pages = kvcalloc(count, sizeof(add->pages[0]), GFP_KERNEL);
143 if (NULL == add->grants ||
144 NULL == add->map_ops ||
145 NULL == add->unmap_ops ||
146 NULL == add->pages)
147 goto err;
148 if (use_ptemod) {
149 add->kmap_ops = kvmalloc_array(count, sizeof(add->kmap_ops[0]),
150 GFP_KERNEL);
151 add->kunmap_ops = kvmalloc_array(count, sizeof(add->kunmap_ops[0]),
152 GFP_KERNEL);
153 if (NULL == add->kmap_ops || NULL == add->kunmap_ops)
154 goto err;
155 }
156
157#ifdef CONFIG_XEN_GRANT_DMA_ALLOC
158 add->dma_flags = dma_flags;
159
160 /*
161 * Check if this mapping is requested to be backed
162 * by a DMA buffer.
163 */
164 if (dma_flags & (GNTDEV_DMA_FLAG_WC | GNTDEV_DMA_FLAG_COHERENT)) {
165 struct gnttab_dma_alloc_args args;
166
167 add->frames = kvcalloc(count, sizeof(add->frames[0]),
168 GFP_KERNEL);
169 if (!add->frames)
170 goto err;
171
172 /* Remember the device, so we can free DMA memory. */
173 add->dma_dev = priv->dma_dev;
174
175 args.dev = priv->dma_dev;
176 args.coherent = !!(dma_flags & GNTDEV_DMA_FLAG_COHERENT);
177 args.nr_pages = count;
178 args.pages = add->pages;
179 args.frames = add->frames;
180
181 if (gnttab_dma_alloc_pages(&args))
182 goto err;
183
184 add->dma_vaddr = args.vaddr;
185 add->dma_bus_addr = args.dev_bus_addr;
186 } else
187#endif
188 if (gnttab_alloc_pages(count, add->pages))
189 goto err;
190
191 for (i = 0; i < count; i++) {
192 add->grants[i].domid = DOMID_INVALID;
193 add->grants[i].ref = INVALID_GRANT_REF;
194 add->map_ops[i].handle = INVALID_GRANT_HANDLE;
195 add->unmap_ops[i].handle = INVALID_GRANT_HANDLE;
196 if (use_ptemod) {
197 add->kmap_ops[i].handle = INVALID_GRANT_HANDLE;
198 add->kunmap_ops[i].handle = INVALID_GRANT_HANDLE;
199 }
200 }
201
202 add->index = 0;
203 add->count = count;
204 refcount_set(&add->users, 1);
205
206 return add;
207
208err:
209 gntdev_free_map(add);
210 return NULL;
211}
212
213void gntdev_add_map(struct gntdev_priv *priv, struct gntdev_grant_map *add)
214{
215 struct gntdev_grant_map *map;
216
217 list_for_each_entry(map, &priv->maps, next) {
218 if (add->index + add->count < map->index) {
219 list_add_tail(&add->next, &map->next);
220 goto done;
221 }
222 add->index = map->index + map->count;
223 }
224 list_add_tail(&add->next, &priv->maps);
225
226done:
227 gntdev_print_maps(priv, "[new]", add->index);
228}
229
230static struct gntdev_grant_map *gntdev_find_map_index(struct gntdev_priv *priv,
231 int index, int count)
232{
233 struct gntdev_grant_map *map;
234
235 list_for_each_entry(map, &priv->maps, next) {
236 if (map->index != index)
237 continue;
238 if (count && map->count != count)
239 continue;
240 return map;
241 }
242 return NULL;
243}
244
245void gntdev_put_map(struct gntdev_priv *priv, struct gntdev_grant_map *map)
246{
247 if (!map)
248 return;
249
250 if (!refcount_dec_and_test(&map->users))
251 return;
252
253 if (map->notify.flags & UNMAP_NOTIFY_SEND_EVENT) {
254 notify_remote_via_evtchn(map->notify.event);
255 evtchn_put(map->notify.event);
256 }
257
258 if (map->pages && !use_ptemod)
259 unmap_grant_pages(map, 0, map->count);
260 gntdev_free_map(map);
261}
262
263/* ------------------------------------------------------------------ */
264
265static int find_grant_ptes(pte_t *pte, unsigned long addr, void *data)
266{
267 struct gntdev_grant_map *map = data;
268 unsigned int pgnr = (addr - map->vma->vm_start) >> PAGE_SHIFT;
269 int flags = map->flags | GNTMAP_application_map | GNTMAP_contains_pte;
270 u64 pte_maddr;
271
272 BUG_ON(pgnr >= map->count);
273 pte_maddr = arbitrary_virt_to_machine(pte).maddr;
274
275 /*
276 * Set the PTE as special to force get_user_pages_fast() fall
277 * back to the slow path. If this is not supported as part of
278 * the grant map, it will be done afterwards.
279 */
280 if (xen_feature(XENFEAT_gnttab_map_avail_bits))
281 flags |= (1 << _GNTMAP_guest_avail0);
282
283 gnttab_set_map_op(&map->map_ops[pgnr], pte_maddr, flags,
284 map->grants[pgnr].ref,
285 map->grants[pgnr].domid);
286 gnttab_set_unmap_op(&map->unmap_ops[pgnr], pte_maddr, flags,
287 INVALID_GRANT_HANDLE);
288 return 0;
289}
290
291#ifdef CONFIG_X86
292static int set_grant_ptes_as_special(pte_t *pte, unsigned long addr, void *data)
293{
294 set_pte_at(current->mm, addr, pte, pte_mkspecial(*pte));
295 return 0;
296}
297#endif
298
299int gntdev_map_grant_pages(struct gntdev_grant_map *map)
300{
301 int i, err = 0;
302
303 if (!use_ptemod) {
304 /* Note: it could already be mapped */
305 if (map->map_ops[0].handle != INVALID_GRANT_HANDLE)
306 return 0;
307 for (i = 0; i < map->count; i++) {
308 unsigned long addr = (unsigned long)
309 pfn_to_kaddr(page_to_pfn(map->pages[i]));
310 gnttab_set_map_op(&map->map_ops[i], addr, map->flags,
311 map->grants[i].ref,
312 map->grants[i].domid);
313 gnttab_set_unmap_op(&map->unmap_ops[i], addr,
314 map->flags, INVALID_GRANT_HANDLE);
315 }
316 } else {
317 /*
318 * Setup the map_ops corresponding to the pte entries pointing
319 * to the kernel linear addresses of the struct pages.
320 * These ptes are completely different from the user ptes dealt
321 * with find_grant_ptes.
322 * Note that GNTMAP_device_map isn't needed here: The
323 * dev_bus_addr output field gets consumed only from ->map_ops,
324 * and by not requesting it when mapping we also avoid needing
325 * to mirror dev_bus_addr into ->unmap_ops (and holding an extra
326 * reference to the page in the hypervisor).
327 */
328 unsigned int flags = (map->flags & ~GNTMAP_device_map) |
329 GNTMAP_host_map;
330
331 for (i = 0; i < map->count; i++) {
332 unsigned long address = (unsigned long)
333 pfn_to_kaddr(page_to_pfn(map->pages[i]));
334 BUG_ON(PageHighMem(map->pages[i]));
335
336 gnttab_set_map_op(&map->kmap_ops[i], address, flags,
337 map->grants[i].ref,
338 map->grants[i].domid);
339 gnttab_set_unmap_op(&map->kunmap_ops[i], address,
340 flags, INVALID_GRANT_HANDLE);
341 }
342 }
343
344 pr_debug("map %d+%d\n", map->index, map->count);
345 err = gnttab_map_refs(map->map_ops, map->kmap_ops, map->pages,
346 map->count);
347
348 for (i = 0; i < map->count; i++) {
349 if (map->map_ops[i].status == GNTST_okay)
350 map->unmap_ops[i].handle = map->map_ops[i].handle;
351 else if (!err)
352 err = -EINVAL;
353
354 if (map->flags & GNTMAP_device_map)
355 map->unmap_ops[i].dev_bus_addr = map->map_ops[i].dev_bus_addr;
356
357 if (use_ptemod) {
358 if (map->kmap_ops[i].status == GNTST_okay)
359 map->kunmap_ops[i].handle = map->kmap_ops[i].handle;
360 else if (!err)
361 err = -EINVAL;
362 }
363 }
364 return err;
365}
366
367static int __unmap_grant_pages(struct gntdev_grant_map *map, int offset,
368 int pages)
369{
370 int i, err = 0;
371 struct gntab_unmap_queue_data unmap_data;
372
373 if (map->notify.flags & UNMAP_NOTIFY_CLEAR_BYTE) {
374 int pgno = (map->notify.addr >> PAGE_SHIFT);
375 if (pgno >= offset && pgno < offset + pages) {
376 /* No need for kmap, pages are in lowmem */
377 uint8_t *tmp = pfn_to_kaddr(page_to_pfn(map->pages[pgno]));
378 tmp[map->notify.addr & (PAGE_SIZE-1)] = 0;
379 map->notify.flags &= ~UNMAP_NOTIFY_CLEAR_BYTE;
380 }
381 }
382
383 unmap_data.unmap_ops = map->unmap_ops + offset;
384 unmap_data.kunmap_ops = use_ptemod ? map->kunmap_ops + offset : NULL;
385 unmap_data.pages = map->pages + offset;
386 unmap_data.count = pages;
387
388 err = gnttab_unmap_refs_sync(&unmap_data);
389 if (err)
390 return err;
391
392 for (i = 0; i < pages; i++) {
393 if (map->unmap_ops[offset+i].status)
394 err = -EINVAL;
395 pr_debug("unmap handle=%d st=%d\n",
396 map->unmap_ops[offset+i].handle,
397 map->unmap_ops[offset+i].status);
398 map->unmap_ops[offset+i].handle = INVALID_GRANT_HANDLE;
399 if (use_ptemod) {
400 if (map->kunmap_ops[offset+i].status)
401 err = -EINVAL;
402 pr_debug("kunmap handle=%u st=%d\n",
403 map->kunmap_ops[offset+i].handle,
404 map->kunmap_ops[offset+i].status);
405 map->kunmap_ops[offset+i].handle = INVALID_GRANT_HANDLE;
406 }
407 }
408 return err;
409}
410
411static int unmap_grant_pages(struct gntdev_grant_map *map, int offset,
412 int pages)
413{
414 int range, err = 0;
415
416 pr_debug("unmap %d+%d [%d+%d]\n", map->index, map->count, offset, pages);
417
418 /* It is possible the requested range will have a "hole" where we
419 * already unmapped some of the grants. Only unmap valid ranges.
420 */
421 while (pages && !err) {
422 while (pages &&
423 map->unmap_ops[offset].handle == INVALID_GRANT_HANDLE) {
424 offset++;
425 pages--;
426 }
427 range = 0;
428 while (range < pages) {
429 if (map->unmap_ops[offset + range].handle ==
430 INVALID_GRANT_HANDLE)
431 break;
432 range++;
433 }
434 err = __unmap_grant_pages(map, offset, range);
435 offset += range;
436 pages -= range;
437 }
438
439 return err;
440}
441
442/* ------------------------------------------------------------------ */
443
444static void gntdev_vma_open(struct vm_area_struct *vma)
445{
446 struct gntdev_grant_map *map = vma->vm_private_data;
447
448 pr_debug("gntdev_vma_open %p\n", vma);
449 refcount_inc(&map->users);
450}
451
452static void gntdev_vma_close(struct vm_area_struct *vma)
453{
454 struct gntdev_grant_map *map = vma->vm_private_data;
455 struct file *file = vma->vm_file;
456 struct gntdev_priv *priv = file->private_data;
457
458 pr_debug("gntdev_vma_close %p\n", vma);
459 if (use_ptemod) {
460 WARN_ON(map->vma != vma);
461 mmu_interval_notifier_remove(&map->notifier);
462 map->vma = NULL;
463 }
464 vma->vm_private_data = NULL;
465 gntdev_put_map(priv, map);
466}
467
468static struct page *gntdev_vma_find_special_page(struct vm_area_struct *vma,
469 unsigned long addr)
470{
471 struct gntdev_grant_map *map = vma->vm_private_data;
472
473 return map->pages[(addr - map->pages_vm_start) >> PAGE_SHIFT];
474}
475
476static const struct vm_operations_struct gntdev_vmops = {
477 .open = gntdev_vma_open,
478 .close = gntdev_vma_close,
479 .find_special_page = gntdev_vma_find_special_page,
480};
481
482/* ------------------------------------------------------------------ */
483
484static bool gntdev_invalidate(struct mmu_interval_notifier *mn,
485 const struct mmu_notifier_range *range,
486 unsigned long cur_seq)
487{
488 struct gntdev_grant_map *map =
489 container_of(mn, struct gntdev_grant_map, notifier);
490 unsigned long mstart, mend;
491 int err;
492
493 if (!mmu_notifier_range_blockable(range))
494 return false;
495
496 /*
497 * If the VMA is split or otherwise changed the notifier is not
498 * updated, but we don't want to process VA's outside the modified
499 * VMA. FIXME: It would be much more understandable to just prevent
500 * modifying the VMA in the first place.
501 */
502 if (map->vma->vm_start >= range->end ||
503 map->vma->vm_end <= range->start)
504 return true;
505
506 mstart = max(range->start, map->vma->vm_start);
507 mend = min(range->end, map->vma->vm_end);
508 pr_debug("map %d+%d (%lx %lx), range %lx %lx, mrange %lx %lx\n",
509 map->index, map->count,
510 map->vma->vm_start, map->vma->vm_end,
511 range->start, range->end, mstart, mend);
512 err = unmap_grant_pages(map,
513 (mstart - map->vma->vm_start) >> PAGE_SHIFT,
514 (mend - mstart) >> PAGE_SHIFT);
515 WARN_ON(err);
516
517 return true;
518}
519
520static const struct mmu_interval_notifier_ops gntdev_mmu_ops = {
521 .invalidate = gntdev_invalidate,
522};
523
524/* ------------------------------------------------------------------ */
525
526static int gntdev_open(struct inode *inode, struct file *flip)
527{
528 struct gntdev_priv *priv;
529
530 priv = kzalloc(sizeof(*priv), GFP_KERNEL);
531 if (!priv)
532 return -ENOMEM;
533
534 INIT_LIST_HEAD(&priv->maps);
535 mutex_init(&priv->lock);
536
537#ifdef CONFIG_XEN_GNTDEV_DMABUF
538 priv->dmabuf_priv = gntdev_dmabuf_init(flip);
539 if (IS_ERR(priv->dmabuf_priv)) {
540 int ret = PTR_ERR(priv->dmabuf_priv);
541
542 kfree(priv);
543 return ret;
544 }
545#endif
546
547 flip->private_data = priv;
548#ifdef CONFIG_XEN_GRANT_DMA_ALLOC
549 priv->dma_dev = gntdev_miscdev.this_device;
550 dma_coerce_mask_and_coherent(priv->dma_dev, DMA_BIT_MASK(64));
551#endif
552 pr_debug("priv %p\n", priv);
553
554 return 0;
555}
556
557static int gntdev_release(struct inode *inode, struct file *flip)
558{
559 struct gntdev_priv *priv = flip->private_data;
560 struct gntdev_grant_map *map;
561
562 pr_debug("priv %p\n", priv);
563
564 mutex_lock(&priv->lock);
565 while (!list_empty(&priv->maps)) {
566 map = list_entry(priv->maps.next,
567 struct gntdev_grant_map, next);
568 list_del(&map->next);
569 gntdev_put_map(NULL /* already removed */, map);
570 }
571 mutex_unlock(&priv->lock);
572
573#ifdef CONFIG_XEN_GNTDEV_DMABUF
574 gntdev_dmabuf_fini(priv->dmabuf_priv);
575#endif
576
577 kfree(priv);
578 return 0;
579}
580
581static long gntdev_ioctl_map_grant_ref(struct gntdev_priv *priv,
582 struct ioctl_gntdev_map_grant_ref __user *u)
583{
584 struct ioctl_gntdev_map_grant_ref op;
585 struct gntdev_grant_map *map;
586 int err;
587
588 if (copy_from_user(&op, u, sizeof(op)) != 0)
589 return -EFAULT;
590 pr_debug("priv %p, add %d\n", priv, op.count);
591 if (unlikely(gntdev_test_page_count(op.count)))
592 return -EINVAL;
593
594 err = -ENOMEM;
595 map = gntdev_alloc_map(priv, op.count, 0 /* This is not a dma-buf. */);
596 if (!map)
597 return err;
598
599 if (copy_from_user(map->grants, &u->refs,
600 sizeof(map->grants[0]) * op.count) != 0) {
601 gntdev_put_map(NULL, map);
602 return -EFAULT;
603 }
604
605 mutex_lock(&priv->lock);
606 gntdev_add_map(priv, map);
607 op.index = map->index << PAGE_SHIFT;
608 mutex_unlock(&priv->lock);
609
610 if (copy_to_user(u, &op, sizeof(op)) != 0)
611 return -EFAULT;
612
613 return 0;
614}
615
616static long gntdev_ioctl_unmap_grant_ref(struct gntdev_priv *priv,
617 struct ioctl_gntdev_unmap_grant_ref __user *u)
618{
619 struct ioctl_gntdev_unmap_grant_ref op;
620 struct gntdev_grant_map *map;
621 int err = -ENOENT;
622
623 if (copy_from_user(&op, u, sizeof(op)) != 0)
624 return -EFAULT;
625 pr_debug("priv %p, del %d+%d\n", priv, (int)op.index, (int)op.count);
626
627 mutex_lock(&priv->lock);
628 map = gntdev_find_map_index(priv, op.index >> PAGE_SHIFT, op.count);
629 if (map) {
630 list_del(&map->next);
631 err = 0;
632 }
633 mutex_unlock(&priv->lock);
634 if (map)
635 gntdev_put_map(priv, map);
636 return err;
637}
638
639static long gntdev_ioctl_get_offset_for_vaddr(struct gntdev_priv *priv,
640 struct ioctl_gntdev_get_offset_for_vaddr __user *u)
641{
642 struct ioctl_gntdev_get_offset_for_vaddr op;
643 struct vm_area_struct *vma;
644 struct gntdev_grant_map *map;
645 int rv = -EINVAL;
646
647 if (copy_from_user(&op, u, sizeof(op)) != 0)
648 return -EFAULT;
649 pr_debug("priv %p, offset for vaddr %lx\n", priv, (unsigned long)op.vaddr);
650
651 mmap_read_lock(current->mm);
652 vma = find_vma(current->mm, op.vaddr);
653 if (!vma || vma->vm_ops != &gntdev_vmops)
654 goto out_unlock;
655
656 map = vma->vm_private_data;
657 if (!map)
658 goto out_unlock;
659
660 op.offset = map->index << PAGE_SHIFT;
661 op.count = map->count;
662 rv = 0;
663
664 out_unlock:
665 mmap_read_unlock(current->mm);
666
667 if (rv == 0 && copy_to_user(u, &op, sizeof(op)) != 0)
668 return -EFAULT;
669 return rv;
670}
671
672static long gntdev_ioctl_notify(struct gntdev_priv *priv, void __user *u)
673{
674 struct ioctl_gntdev_unmap_notify op;
675 struct gntdev_grant_map *map;
676 int rc;
677 int out_flags;
678 evtchn_port_t out_event;
679
680 if (copy_from_user(&op, u, sizeof(op)))
681 return -EFAULT;
682
683 if (op.action & ~(UNMAP_NOTIFY_CLEAR_BYTE|UNMAP_NOTIFY_SEND_EVENT))
684 return -EINVAL;
685
686 /* We need to grab a reference to the event channel we are going to use
687 * to send the notify before releasing the reference we may already have
688 * (if someone has called this ioctl twice). This is required so that
689 * it is possible to change the clear_byte part of the notification
690 * without disturbing the event channel part, which may now be the last
691 * reference to that event channel.
692 */
693 if (op.action & UNMAP_NOTIFY_SEND_EVENT) {
694 if (evtchn_get(op.event_channel_port))
695 return -EINVAL;
696 }
697
698 out_flags = op.action;
699 out_event = op.event_channel_port;
700
701 mutex_lock(&priv->lock);
702
703 list_for_each_entry(map, &priv->maps, next) {
704 uint64_t begin = map->index << PAGE_SHIFT;
705 uint64_t end = (map->index + map->count) << PAGE_SHIFT;
706 if (op.index >= begin && op.index < end)
707 goto found;
708 }
709 rc = -ENOENT;
710 goto unlock_out;
711
712 found:
713 if ((op.action & UNMAP_NOTIFY_CLEAR_BYTE) &&
714 (map->flags & GNTMAP_readonly)) {
715 rc = -EINVAL;
716 goto unlock_out;
717 }
718
719 out_flags = map->notify.flags;
720 out_event = map->notify.event;
721
722 map->notify.flags = op.action;
723 map->notify.addr = op.index - (map->index << PAGE_SHIFT);
724 map->notify.event = op.event_channel_port;
725
726 rc = 0;
727
728 unlock_out:
729 mutex_unlock(&priv->lock);
730
731 /* Drop the reference to the event channel we did not save in the map */
732 if (out_flags & UNMAP_NOTIFY_SEND_EVENT)
733 evtchn_put(out_event);
734
735 return rc;
736}
737
738#define GNTDEV_COPY_BATCH 16
739
740struct gntdev_copy_batch {
741 struct gnttab_copy ops[GNTDEV_COPY_BATCH];
742 struct page *pages[GNTDEV_COPY_BATCH];
743 s16 __user *status[GNTDEV_COPY_BATCH];
744 unsigned int nr_ops;
745 unsigned int nr_pages;
746 bool writeable;
747};
748
749static int gntdev_get_page(struct gntdev_copy_batch *batch, void __user *virt,
750 unsigned long *gfn)
751{
752 unsigned long addr = (unsigned long)virt;
753 struct page *page;
754 unsigned long xen_pfn;
755 int ret;
756
757 ret = pin_user_pages_fast(addr, 1, batch->writeable ? FOLL_WRITE : 0, &page);
758 if (ret < 0)
759 return ret;
760
761 batch->pages[batch->nr_pages++] = page;
762
763 xen_pfn = page_to_xen_pfn(page) + XEN_PFN_DOWN(addr & ~PAGE_MASK);
764 *gfn = pfn_to_gfn(xen_pfn);
765
766 return 0;
767}
768
769static void gntdev_put_pages(struct gntdev_copy_batch *batch)
770{
771 unpin_user_pages_dirty_lock(batch->pages, batch->nr_pages, batch->writeable);
772 batch->nr_pages = 0;
773 batch->writeable = false;
774}
775
776static int gntdev_copy(struct gntdev_copy_batch *batch)
777{
778 unsigned int i;
779
780 gnttab_batch_copy(batch->ops, batch->nr_ops);
781 gntdev_put_pages(batch);
782
783 /*
784 * For each completed op, update the status if the op failed
785 * and all previous ops for the segment were successful.
786 */
787 for (i = 0; i < batch->nr_ops; i++) {
788 s16 status = batch->ops[i].status;
789 s16 old_status;
790
791 if (status == GNTST_okay)
792 continue;
793
794 if (__get_user(old_status, batch->status[i]))
795 return -EFAULT;
796
797 if (old_status != GNTST_okay)
798 continue;
799
800 if (__put_user(status, batch->status[i]))
801 return -EFAULT;
802 }
803
804 batch->nr_ops = 0;
805 return 0;
806}
807
808static int gntdev_grant_copy_seg(struct gntdev_copy_batch *batch,
809 struct gntdev_grant_copy_segment *seg,
810 s16 __user *status)
811{
812 uint16_t copied = 0;
813
814 /*
815 * Disallow local -> local copies since there is only space in
816 * batch->pages for one page per-op and this would be a very
817 * expensive memcpy().
818 */
819 if (!(seg->flags & (GNTCOPY_source_gref | GNTCOPY_dest_gref)))
820 return -EINVAL;
821
822 /* Can't cross page if source/dest is a grant ref. */
823 if (seg->flags & GNTCOPY_source_gref) {
824 if (seg->source.foreign.offset + seg->len > XEN_PAGE_SIZE)
825 return -EINVAL;
826 }
827 if (seg->flags & GNTCOPY_dest_gref) {
828 if (seg->dest.foreign.offset + seg->len > XEN_PAGE_SIZE)
829 return -EINVAL;
830 }
831
832 if (put_user(GNTST_okay, status))
833 return -EFAULT;
834
835 while (copied < seg->len) {
836 struct gnttab_copy *op;
837 void __user *virt;
838 size_t len, off;
839 unsigned long gfn;
840 int ret;
841
842 if (batch->nr_ops >= GNTDEV_COPY_BATCH) {
843 ret = gntdev_copy(batch);
844 if (ret < 0)
845 return ret;
846 }
847
848 len = seg->len - copied;
849
850 op = &batch->ops[batch->nr_ops];
851 op->flags = 0;
852
853 if (seg->flags & GNTCOPY_source_gref) {
854 op->source.u.ref = seg->source.foreign.ref;
855 op->source.domid = seg->source.foreign.domid;
856 op->source.offset = seg->source.foreign.offset + copied;
857 op->flags |= GNTCOPY_source_gref;
858 } else {
859 virt = seg->source.virt + copied;
860 off = (unsigned long)virt & ~XEN_PAGE_MASK;
861 len = min(len, (size_t)XEN_PAGE_SIZE - off);
862 batch->writeable = false;
863
864 ret = gntdev_get_page(batch, virt, &gfn);
865 if (ret < 0)
866 return ret;
867
868 op->source.u.gmfn = gfn;
869 op->source.domid = DOMID_SELF;
870 op->source.offset = off;
871 }
872
873 if (seg->flags & GNTCOPY_dest_gref) {
874 op->dest.u.ref = seg->dest.foreign.ref;
875 op->dest.domid = seg->dest.foreign.domid;
876 op->dest.offset = seg->dest.foreign.offset + copied;
877 op->flags |= GNTCOPY_dest_gref;
878 } else {
879 virt = seg->dest.virt + copied;
880 off = (unsigned long)virt & ~XEN_PAGE_MASK;
881 len = min(len, (size_t)XEN_PAGE_SIZE - off);
882 batch->writeable = true;
883
884 ret = gntdev_get_page(batch, virt, &gfn);
885 if (ret < 0)
886 return ret;
887
888 op->dest.u.gmfn = gfn;
889 op->dest.domid = DOMID_SELF;
890 op->dest.offset = off;
891 }
892
893 op->len = len;
894 copied += len;
895
896 batch->status[batch->nr_ops] = status;
897 batch->nr_ops++;
898 }
899
900 return 0;
901}
902
903static long gntdev_ioctl_grant_copy(struct gntdev_priv *priv, void __user *u)
904{
905 struct ioctl_gntdev_grant_copy copy;
906 struct gntdev_copy_batch batch;
907 unsigned int i;
908 int ret = 0;
909
910 if (copy_from_user(©, u, sizeof(copy)))
911 return -EFAULT;
912
913 batch.nr_ops = 0;
914 batch.nr_pages = 0;
915
916 for (i = 0; i < copy.count; i++) {
917 struct gntdev_grant_copy_segment seg;
918
919 if (copy_from_user(&seg, ©.segments[i], sizeof(seg))) {
920 ret = -EFAULT;
921 goto out;
922 }
923
924 ret = gntdev_grant_copy_seg(&batch, &seg, ©.segments[i].status);
925 if (ret < 0)
926 goto out;
927
928 cond_resched();
929 }
930 if (batch.nr_ops)
931 ret = gntdev_copy(&batch);
932 return ret;
933
934 out:
935 gntdev_put_pages(&batch);
936 return ret;
937}
938
939static long gntdev_ioctl(struct file *flip,
940 unsigned int cmd, unsigned long arg)
941{
942 struct gntdev_priv *priv = flip->private_data;
943 void __user *ptr = (void __user *)arg;
944
945 switch (cmd) {
946 case IOCTL_GNTDEV_MAP_GRANT_REF:
947 return gntdev_ioctl_map_grant_ref(priv, ptr);
948
949 case IOCTL_GNTDEV_UNMAP_GRANT_REF:
950 return gntdev_ioctl_unmap_grant_ref(priv, ptr);
951
952 case IOCTL_GNTDEV_GET_OFFSET_FOR_VADDR:
953 return gntdev_ioctl_get_offset_for_vaddr(priv, ptr);
954
955 case IOCTL_GNTDEV_SET_UNMAP_NOTIFY:
956 return gntdev_ioctl_notify(priv, ptr);
957
958 case IOCTL_GNTDEV_GRANT_COPY:
959 return gntdev_ioctl_grant_copy(priv, ptr);
960
961#ifdef CONFIG_XEN_GNTDEV_DMABUF
962 case IOCTL_GNTDEV_DMABUF_EXP_FROM_REFS:
963 return gntdev_ioctl_dmabuf_exp_from_refs(priv, use_ptemod, ptr);
964
965 case IOCTL_GNTDEV_DMABUF_EXP_WAIT_RELEASED:
966 return gntdev_ioctl_dmabuf_exp_wait_released(priv, ptr);
967
968 case IOCTL_GNTDEV_DMABUF_IMP_TO_REFS:
969 return gntdev_ioctl_dmabuf_imp_to_refs(priv, ptr);
970
971 case IOCTL_GNTDEV_DMABUF_IMP_RELEASE:
972 return gntdev_ioctl_dmabuf_imp_release(priv, ptr);
973#endif
974
975 default:
976 pr_debug("priv %p, unknown cmd %x\n", priv, cmd);
977 return -ENOIOCTLCMD;
978 }
979
980 return 0;
981}
982
983static int gntdev_mmap(struct file *flip, struct vm_area_struct *vma)
984{
985 struct gntdev_priv *priv = flip->private_data;
986 int index = vma->vm_pgoff;
987 int count = vma_pages(vma);
988 struct gntdev_grant_map *map;
989 int err = -EINVAL;
990
991 if ((vma->vm_flags & VM_WRITE) && !(vma->vm_flags & VM_SHARED))
992 return -EINVAL;
993
994 pr_debug("map %d+%d at %lx (pgoff %lx)\n",
995 index, count, vma->vm_start, vma->vm_pgoff);
996
997 mutex_lock(&priv->lock);
998 map = gntdev_find_map_index(priv, index, count);
999 if (!map)
1000 goto unlock_out;
1001 if (use_ptemod && map->vma)
1002 goto unlock_out;
1003 refcount_inc(&map->users);
1004
1005 vma->vm_ops = &gntdev_vmops;
1006
1007 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP | VM_MIXEDMAP;
1008
1009 if (use_ptemod)
1010 vma->vm_flags |= VM_DONTCOPY;
1011
1012 vma->vm_private_data = map;
1013 if (map->flags) {
1014 if ((vma->vm_flags & VM_WRITE) &&
1015 (map->flags & GNTMAP_readonly))
1016 goto out_unlock_put;
1017 } else {
1018 map->flags = GNTMAP_host_map;
1019 if (!(vma->vm_flags & VM_WRITE))
1020 map->flags |= GNTMAP_readonly;
1021 }
1022
1023 if (use_ptemod) {
1024 map->vma = vma;
1025 err = mmu_interval_notifier_insert_locked(
1026 &map->notifier, vma->vm_mm, vma->vm_start,
1027 vma->vm_end - vma->vm_start, &gntdev_mmu_ops);
1028 if (err) {
1029 map->vma = NULL;
1030 goto out_unlock_put;
1031 }
1032 }
1033 mutex_unlock(&priv->lock);
1034
1035 if (use_ptemod) {
1036 /*
1037 * gntdev takes the address of the PTE in find_grant_ptes() and
1038 * passes it to the hypervisor in gntdev_map_grant_pages(). The
1039 * purpose of the notifier is to prevent the hypervisor pointer
1040 * to the PTE from going stale.
1041 *
1042 * Since this vma's mappings can't be touched without the
1043 * mmap_lock, and we are holding it now, there is no need for
1044 * the notifier_range locking pattern.
1045 */
1046 mmu_interval_read_begin(&map->notifier);
1047
1048 map->pages_vm_start = vma->vm_start;
1049 err = apply_to_page_range(vma->vm_mm, vma->vm_start,
1050 vma->vm_end - vma->vm_start,
1051 find_grant_ptes, map);
1052 if (err) {
1053 pr_warn("find_grant_ptes() failure.\n");
1054 goto out_put_map;
1055 }
1056 }
1057
1058 err = gntdev_map_grant_pages(map);
1059 if (err)
1060 goto out_put_map;
1061
1062 if (!use_ptemod) {
1063 err = vm_map_pages_zero(vma, map->pages, map->count);
1064 if (err)
1065 goto out_put_map;
1066 } else {
1067#ifdef CONFIG_X86
1068 /*
1069 * If the PTEs were not made special by the grant map
1070 * hypercall, do so here.
1071 *
1072 * This is racy since the mapping is already visible
1073 * to userspace but userspace should be well-behaved
1074 * enough to not touch it until the mmap() call
1075 * returns.
1076 */
1077 if (!xen_feature(XENFEAT_gnttab_map_avail_bits)) {
1078 apply_to_page_range(vma->vm_mm, vma->vm_start,
1079 vma->vm_end - vma->vm_start,
1080 set_grant_ptes_as_special, NULL);
1081 }
1082#endif
1083 }
1084
1085 return 0;
1086
1087unlock_out:
1088 mutex_unlock(&priv->lock);
1089 return err;
1090
1091out_unlock_put:
1092 mutex_unlock(&priv->lock);
1093out_put_map:
1094 if (use_ptemod) {
1095 unmap_grant_pages(map, 0, map->count);
1096 if (map->vma) {
1097 mmu_interval_notifier_remove(&map->notifier);
1098 map->vma = NULL;
1099 }
1100 }
1101 gntdev_put_map(priv, map);
1102 return err;
1103}
1104
1105static const struct file_operations gntdev_fops = {
1106 .owner = THIS_MODULE,
1107 .open = gntdev_open,
1108 .release = gntdev_release,
1109 .mmap = gntdev_mmap,
1110 .unlocked_ioctl = gntdev_ioctl
1111};
1112
1113static struct miscdevice gntdev_miscdev = {
1114 .minor = MISC_DYNAMIC_MINOR,
1115 .name = "xen/gntdev",
1116 .fops = &gntdev_fops,
1117};
1118
1119/* ------------------------------------------------------------------ */
1120
1121static int __init gntdev_init(void)
1122{
1123 int err;
1124
1125 if (!xen_domain())
1126 return -ENODEV;
1127
1128 use_ptemod = !xen_feature(XENFEAT_auto_translated_physmap);
1129
1130 err = misc_register(&gntdev_miscdev);
1131 if (err != 0) {
1132 pr_err("Could not register gntdev device\n");
1133 return err;
1134 }
1135 return 0;
1136}
1137
1138static void __exit gntdev_exit(void)
1139{
1140 misc_deregister(&gntdev_miscdev);
1141}
1142
1143module_init(gntdev_init);
1144module_exit(gntdev_exit);
1145
1146/* ------------------------------------------------------------------ */