1 /******************************************************************************
4 * Device for accessing (in user-space) pages that have been granted by other
7 * Copyright (c) 2006-2007, D G Murray.
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.
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
22 #define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt
24 #include <linux/module.h>
25 #include <linux/kernel.h>
26 #include <linux/init.h>
27 #include <linux/miscdevice.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>
40 #include <xen/grant_table.h>
41 #include <xen/balloon.h>
42 #include <xen/gntdev.h>
43 #include <xen/events.h>
45 #include <asm/xen/hypervisor.h>
46 #include <asm/xen/hypercall.h>
48 MODULE_LICENSE("GPL");
51 MODULE_DESCRIPTION("User-space granted page access driver");
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 "
58 static atomic_t pages_mapped = ATOMIC_INIT(0);
60 static int use_ptemod;
61 #define populate_freeable_maps use_ptemod
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 */
72 struct mmu_notifier mn;
77 /* Address relative to the start of the grant_map */
83 struct list_head next;
84 struct vm_area_struct *vma;
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;
96 unsigned long pages_vm_start;
99 static int unmap_grant_pages(struct grant_map *map, int offset, int pages);
101 /* ------------------------------------------------------------------ */
103 static void gntdev_print_maps(struct gntdev_priv *priv,
104 char *text, int text_index)
107 struct grant_map *map;
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 : "");
117 static void gntdev_free_map(struct grant_map *map)
123 gnttab_free_pages(map->count, map->pages);
127 kfree(map->unmap_ops);
128 kfree(map->kmap_ops);
129 kfree(map->kunmap_ops);
133 static struct grant_map *gntdev_alloc_map(struct gntdev_priv *priv, int count)
135 struct grant_map *add;
138 add = kzalloc(sizeof(struct grant_map), GFP_KERNEL);
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 ||
156 if (gnttab_alloc_pages(count, add->pages))
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;
168 atomic_set(&add->users, 1);
173 gntdev_free_map(add);
177 static void gntdev_add_map(struct gntdev_priv *priv, struct grant_map *add)
179 struct grant_map *map;
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);
186 add->index = map->index + map->count;
188 list_add_tail(&add->next, &priv->maps);
191 gntdev_print_maps(priv, "[new]", add->index);
194 static struct grant_map *gntdev_find_map_index(struct gntdev_priv *priv,
195 int index, int count)
197 struct grant_map *map;
199 list_for_each_entry(map, &priv->maps, next) {
200 if (map->index != index)
202 if (count && map->count != count)
209 static void gntdev_put_map(struct gntdev_priv *priv, struct grant_map *map)
214 if (!atomic_dec_and_test(&map->users))
217 atomic_sub(map->count, &pages_mapped);
219 if (map->notify.flags & UNMAP_NOTIFY_SEND_EVENT) {
220 notify_remote_via_evtchn(map->notify.event);
221 evtchn_put(map->notify.event);
224 if (populate_freeable_maps && priv) {
225 mutex_lock(&priv->lock);
226 list_del(&map->next);
227 mutex_unlock(&priv->lock);
230 if (map->pages && !use_ptemod)
231 unmap_grant_pages(map, 0, map->count);
232 gntdev_free_map(map);
235 /* ------------------------------------------------------------------ */
237 static int find_grant_ptes(pte_t *pte, pgtable_t token,
238 unsigned long addr, void *data)
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;
245 BUG_ON(pgnr >= map->count);
246 pte_maddr = arbitrary_virt_to_machine(pte).maddr;
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.
253 if (xen_feature(XENFEAT_gnttab_map_avail_bits))
254 flags |= (1 << _GNTMAP_guest_avail0);
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,
265 static int set_grant_ptes_as_special(pte_t *pte, pgtable_t token,
266 unsigned long addr, void *data)
268 set_pte_at(current->mm, addr, pte, pte_mkspecial(*pte));
273 static int map_grant_pages(struct grant_map *map)
278 /* Note: it could already be mapped */
279 if (map->map_ops[0].handle != -1)
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,
286 map->grants[i].domid);
287 gnttab_set_unmap_op(&map->unmap_ops[i], addr,
288 map->flags, -1 /* handle */);
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.
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]));
302 gnttab_set_map_op(&map->kmap_ops[i], address,
303 map->flags | GNTMAP_host_map,
305 map->grants[i].domid);
306 gnttab_set_unmap_op(&map->kunmap_ops[i], address,
307 map->flags | GNTMAP_host_map, -1);
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);
317 for (i = 0; i < map->count; i++) {
318 if (map->map_ops[i].status) {
323 map->unmap_ops[i].handle = map->map_ops[i].handle;
325 map->kunmap_ops[i].handle = map->kmap_ops[i].handle;
330 static int __unmap_grant_pages(struct grant_map *map, int offset, int pages)
333 struct gntab_unmap_queue_data unmap_data;
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;
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;
350 err = gnttab_unmap_refs_sync(&unmap_data);
354 for (i = 0; i < pages; i++) {
355 if (map->unmap_ops[offset+i].status)
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;
365 static int unmap_grant_pages(struct grant_map *map, int offset, int pages)
369 pr_debug("unmap %d+%d [%d+%d]\n", map->index, map->count, offset, pages);
371 /* It is possible the requested range will have a "hole" where we
372 * already unmapped some of the grants. Only unmap valid ranges.
374 while (pages && !err) {
375 while (pages && map->unmap_ops[offset].handle == -1) {
380 while (range < pages) {
381 if (map->unmap_ops[offset+range].handle == -1) {
387 err = __unmap_grant_pages(map, offset, range);
395 /* ------------------------------------------------------------------ */
397 static void gntdev_vma_open(struct vm_area_struct *vma)
399 struct grant_map *map = vma->vm_private_data;
401 pr_debug("gntdev_vma_open %p\n", vma);
402 atomic_inc(&map->users);
405 static void gntdev_vma_close(struct vm_area_struct *vma)
407 struct grant_map *map = vma->vm_private_data;
408 struct file *file = vma->vm_file;
409 struct gntdev_priv *priv = file->private_data;
411 pr_debug("gntdev_vma_close %p\n", vma);
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.
420 mutex_lock(&priv->lock);
422 mutex_unlock(&priv->lock);
424 vma->vm_private_data = NULL;
425 gntdev_put_map(priv, map);
428 static struct page *gntdev_vma_find_special_page(struct vm_area_struct *vma,
431 struct grant_map *map = vma->vm_private_data;
433 return map->pages[(addr - map->pages_vm_start) >> PAGE_SHIFT];
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,
442 /* ------------------------------------------------------------------ */
444 static void unmap_if_in_range(struct grant_map *map,
445 unsigned long start, unsigned long end)
447 unsigned long mstart, mend;
452 if (map->vma->vm_start >= end)
454 if (map->vma->vm_end <= start)
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);
468 static void mn_invl_range_start(struct mmu_notifier *mn,
469 struct mm_struct *mm,
470 unsigned long start, unsigned long end)
472 struct gntdev_priv *priv = container_of(mn, struct gntdev_priv, mn);
473 struct grant_map *map;
475 mutex_lock(&priv->lock);
476 list_for_each_entry(map, &priv->maps, next) {
477 unmap_if_in_range(map, start, end);
479 list_for_each_entry(map, &priv->freeable_maps, next) {
480 unmap_if_in_range(map, start, end);
482 mutex_unlock(&priv->lock);
485 static void mn_invl_page(struct mmu_notifier *mn,
486 struct mm_struct *mm,
487 unsigned long address)
489 mn_invl_range_start(mn, mm, address, address + PAGE_SIZE);
492 static void mn_release(struct mmu_notifier *mn,
493 struct mm_struct *mm)
495 struct gntdev_priv *priv = container_of(mn, struct gntdev_priv, mn);
496 struct grant_map *map;
499 mutex_lock(&priv->lock);
500 list_for_each_entry(map, &priv->maps, next) {
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);
509 list_for_each_entry(map, &priv->freeable_maps, next) {
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);
518 mutex_unlock(&priv->lock);
521 static 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,
527 /* ------------------------------------------------------------------ */
529 static int gntdev_open(struct inode *inode, struct file *flip)
531 struct gntdev_priv *priv;
534 priv = kzalloc(sizeof(*priv), GFP_KERNEL);
538 INIT_LIST_HEAD(&priv->maps);
539 INIT_LIST_HEAD(&priv->freeable_maps);
540 mutex_init(&priv->lock);
543 priv->mm = get_task_mm(current);
548 priv->mn.ops = &gntdev_mmu_ops;
549 ret = mmu_notifier_register(&priv->mn, priv->mm);
558 flip->private_data = priv;
559 pr_debug("priv %p\n", priv);
564 static int gntdev_release(struct inode *inode, struct file *flip)
566 struct gntdev_priv *priv = flip->private_data;
567 struct grant_map *map;
569 pr_debug("priv %p\n", priv);
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);
577 WARN_ON(!list_empty(&priv->freeable_maps));
578 mutex_unlock(&priv->lock);
581 mmu_notifier_unregister(&priv->mn, priv->mm);
586 static long gntdev_ioctl_map_grant_ref(struct gntdev_priv *priv,
587 struct ioctl_gntdev_map_grant_ref __user *u)
589 struct ioctl_gntdev_map_grant_ref op;
590 struct grant_map *map;
593 if (copy_from_user(&op, u, sizeof(op)) != 0)
595 pr_debug("priv %p, add %d\n", priv, op.count);
596 if (unlikely(op.count <= 0))
600 map = gntdev_alloc_map(priv, op.count);
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);
610 if (copy_from_user(map->grants, &u->refs,
611 sizeof(map->grants[0]) * op.count) != 0) {
612 gntdev_put_map(NULL, map);
616 mutex_lock(&priv->lock);
617 gntdev_add_map(priv, map);
618 op.index = map->index << PAGE_SHIFT;
619 mutex_unlock(&priv->lock);
621 if (copy_to_user(u, &op, sizeof(op)) != 0)
627 static long gntdev_ioctl_unmap_grant_ref(struct gntdev_priv *priv,
628 struct ioctl_gntdev_unmap_grant_ref __user *u)
630 struct ioctl_gntdev_unmap_grant_ref op;
631 struct grant_map *map;
634 if (copy_from_user(&op, u, sizeof(op)) != 0)
636 pr_debug("priv %p, del %d+%d\n", priv, (int)op.index, (int)op.count);
638 mutex_lock(&priv->lock);
639 map = gntdev_find_map_index(priv, op.index >> PAGE_SHIFT, op.count);
641 list_del(&map->next);
642 if (populate_freeable_maps)
643 list_add_tail(&map->next, &priv->freeable_maps);
646 mutex_unlock(&priv->lock);
648 gntdev_put_map(priv, map);
652 static long gntdev_ioctl_get_offset_for_vaddr(struct gntdev_priv *priv,
653 struct ioctl_gntdev_get_offset_for_vaddr __user *u)
655 struct ioctl_gntdev_get_offset_for_vaddr op;
656 struct vm_area_struct *vma;
657 struct grant_map *map;
660 if (copy_from_user(&op, u, sizeof(op)) != 0)
662 pr_debug("priv %p, offset for vaddr %lx\n", priv, (unsigned long)op.vaddr);
664 down_read(¤t->mm->mmap_sem);
665 vma = find_vma(current->mm, op.vaddr);
666 if (!vma || vma->vm_ops != &gntdev_vmops)
669 map = vma->vm_private_data;
673 op.offset = map->index << PAGE_SHIFT;
674 op.count = map->count;
678 up_read(¤t->mm->mmap_sem);
680 if (rv == 0 && copy_to_user(u, &op, sizeof(op)) != 0)
685 static long gntdev_ioctl_notify(struct gntdev_priv *priv, void __user *u)
687 struct ioctl_gntdev_unmap_notify op;
688 struct grant_map *map;
691 unsigned int out_event;
693 if (copy_from_user(&op, u, sizeof(op)))
696 if (op.action & ~(UNMAP_NOTIFY_CLEAR_BYTE|UNMAP_NOTIFY_SEND_EVENT))
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.
706 if (op.action & UNMAP_NOTIFY_SEND_EVENT) {
707 if (evtchn_get(op.event_channel_port))
711 out_flags = op.action;
712 out_event = op.event_channel_port;
714 mutex_lock(&priv->lock);
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)
726 if ((op.action & UNMAP_NOTIFY_CLEAR_BYTE) &&
727 (map->flags & GNTMAP_readonly)) {
732 out_flags = map->notify.flags;
733 out_event = map->notify.event;
735 map->notify.flags = op.action;
736 map->notify.addr = op.index - (map->index << PAGE_SHIFT);
737 map->notify.event = op.event_channel_port;
742 mutex_unlock(&priv->lock);
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);
751 static long gntdev_ioctl(struct file *flip,
752 unsigned int cmd, unsigned long arg)
754 struct gntdev_priv *priv = flip->private_data;
755 void __user *ptr = (void __user *)arg;
758 case IOCTL_GNTDEV_MAP_GRANT_REF:
759 return gntdev_ioctl_map_grant_ref(priv, ptr);
761 case IOCTL_GNTDEV_UNMAP_GRANT_REF:
762 return gntdev_ioctl_unmap_grant_ref(priv, ptr);
764 case IOCTL_GNTDEV_GET_OFFSET_FOR_VADDR:
765 return gntdev_ioctl_get_offset_for_vaddr(priv, ptr);
767 case IOCTL_GNTDEV_SET_UNMAP_NOTIFY:
768 return gntdev_ioctl_notify(priv, ptr);
771 pr_debug("priv %p, unknown cmd %x\n", priv, cmd);
778 static int gntdev_mmap(struct file *flip, struct vm_area_struct *vma)
780 struct gntdev_priv *priv = flip->private_data;
781 int index = vma->vm_pgoff;
782 int count = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
783 struct grant_map *map;
784 int i, err = -EINVAL;
786 if ((vma->vm_flags & VM_WRITE) && !(vma->vm_flags & VM_SHARED))
789 pr_debug("map %d+%d at %lx (pgoff %lx)\n",
790 index, count, vma->vm_start, vma->vm_pgoff);
792 mutex_lock(&priv->lock);
793 map = gntdev_find_map_index(priv, index, count);
796 if (use_ptemod && map->vma)
798 if (use_ptemod && priv->mm != vma->vm_mm) {
799 pr_warn("Huh? Other mm?\n");
803 atomic_inc(&map->users);
805 vma->vm_ops = &gntdev_vmops;
807 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
810 vma->vm_flags |= VM_DONTCOPY;
812 vma->vm_private_data = map;
818 if ((vma->vm_flags & VM_WRITE) &&
819 (map->flags & GNTMAP_readonly))
822 map->flags = GNTMAP_host_map;
823 if (!(vma->vm_flags & VM_WRITE))
824 map->flags |= GNTMAP_readonly;
827 mutex_unlock(&priv->lock);
830 err = apply_to_page_range(vma->vm_mm, vma->vm_start,
831 vma->vm_end - vma->vm_start,
832 find_grant_ptes, map);
834 pr_warn("find_grant_ptes() failure.\n");
839 err = map_grant_pages(map);
844 for (i = 0; i < count; i++) {
845 err = vm_insert_page(vma, vma->vm_start + i*PAGE_SIZE,
853 * If the PTEs were not made special by the grant map
854 * hypercall, do so here.
856 * This is racy since the mapping is already visible
857 * to userspace but userspace should be well-behaved
858 * enough to not touch it until the mmap() call
861 if (!xen_feature(XENFEAT_gnttab_map_avail_bits)) {
862 apply_to_page_range(vma->vm_mm, vma->vm_start,
863 vma->vm_end - vma->vm_start,
864 set_grant_ptes_as_special, NULL);
867 map->pages_vm_start = vma->vm_start;
873 mutex_unlock(&priv->lock);
877 mutex_unlock(&priv->lock);
881 gntdev_put_map(priv, map);
885 static const struct file_operations gntdev_fops = {
886 .owner = THIS_MODULE,
888 .release = gntdev_release,
890 .unlocked_ioctl = gntdev_ioctl
893 static struct miscdevice gntdev_miscdev = {
894 .minor = MISC_DYNAMIC_MINOR,
895 .name = "xen/gntdev",
896 .fops = &gntdev_fops,
899 /* ------------------------------------------------------------------ */
901 static int __init gntdev_init(void)
908 use_ptemod = !xen_feature(XENFEAT_auto_translated_physmap);
910 err = misc_register(&gntdev_miscdev);
912 pr_err("Could not register gntdev device\n");
918 static void __exit gntdev_exit(void)
920 misc_deregister(&gntdev_miscdev);
923 module_init(gntdev_init);
924 module_exit(gntdev_exit);
926 /* ------------------------------------------------------------------ */