1 /******************************************************************************
4 * Device for accessing (in user-space) pages that have been granted by other
7 * Copyright (c) 2006-2007, D G Murray.
8 * (c) 2009 Gerd Hoffmann <kraxel@redhat.com>
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 #include <linux/module.h>
23 #include <linux/kernel.h>
24 #include <linux/init.h>
25 #include <linux/miscdevice.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>
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>
46 MODULE_LICENSE("GPL");
47 MODULE_AUTHOR("Derek G. Murray <Derek.Murray@cl.cam.ac.uk>, "
48 "Gerd Hoffmann <kraxel@redhat.com>");
49 MODULE_DESCRIPTION("User-space granted page access driver");
51 static int limit = 1024*1024;
52 module_param(limit, int, 0644);
53 MODULE_PARM_DESC(limit, "Maximum number of grants that may be mapped by "
56 static atomic_t pages_mapped = ATOMIC_INIT(0);
58 static int use_ptemod;
61 struct list_head maps;
62 /* lock protects maps from concurrent changes */
65 struct mmu_notifier mn;
70 /* Address relative to the start of the grant_map */
76 struct list_head next;
77 struct vm_area_struct *vma;
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 gnttab_map_grant_ref *kmap_ops;
90 static int unmap_grant_pages(struct grant_map *map, int offset, int pages);
92 /* ------------------------------------------------------------------ */
94 static void gntdev_print_maps(struct gntdev_priv *priv,
95 char *text, int text_index)
98 struct grant_map *map;
100 pr_debug("%s: maps list (priv %p)\n", __func__, priv);
101 list_for_each_entry(map, &priv->maps, next)
102 pr_debug(" index %2d, count %2d %s\n",
103 map->index, map->count,
104 map->index == text_index && text ? text : "");
108 static void gntdev_free_map(struct grant_map *map)
114 free_xenballooned_pages(map->count, map->pages);
118 kfree(map->unmap_ops);
119 kfree(map->kmap_ops);
123 static struct grant_map *gntdev_alloc_map(struct gntdev_priv *priv, int count)
125 struct grant_map *add;
128 add = kzalloc(sizeof(struct grant_map), GFP_KERNEL);
132 add->grants = kcalloc(count, sizeof(add->grants[0]), GFP_KERNEL);
133 add->map_ops = kcalloc(count, sizeof(add->map_ops[0]), GFP_KERNEL);
134 add->unmap_ops = kcalloc(count, sizeof(add->unmap_ops[0]), GFP_KERNEL);
135 add->kmap_ops = kcalloc(count, sizeof(add->kmap_ops[0]), GFP_KERNEL);
136 add->pages = kcalloc(count, sizeof(add->pages[0]), GFP_KERNEL);
137 if (NULL == add->grants ||
138 NULL == add->map_ops ||
139 NULL == add->unmap_ops ||
140 NULL == add->kmap_ops ||
144 if (alloc_xenballooned_pages(count, add->pages, false /* lowmem */))
147 for (i = 0; i < count; i++) {
148 add->map_ops[i].handle = -1;
149 add->unmap_ops[i].handle = -1;
150 add->kmap_ops[i].handle = -1;
155 atomic_set(&add->users, 1);
160 gntdev_free_map(add);
164 static void gntdev_add_map(struct gntdev_priv *priv, struct grant_map *add)
166 struct grant_map *map;
168 list_for_each_entry(map, &priv->maps, next) {
169 if (add->index + add->count < map->index) {
170 list_add_tail(&add->next, &map->next);
173 add->index = map->index + map->count;
175 list_add_tail(&add->next, &priv->maps);
178 gntdev_print_maps(priv, "[new]", add->index);
181 static struct grant_map *gntdev_find_map_index(struct gntdev_priv *priv,
182 int index, int count)
184 struct grant_map *map;
186 list_for_each_entry(map, &priv->maps, next) {
187 if (map->index != index)
189 if (count && map->count != count)
196 static void gntdev_put_map(struct grant_map *map)
201 if (!atomic_dec_and_test(&map->users))
204 atomic_sub(map->count, &pages_mapped);
206 if (map->notify.flags & UNMAP_NOTIFY_SEND_EVENT) {
207 notify_remote_via_evtchn(map->notify.event);
208 evtchn_put(map->notify.event);
211 if (map->pages && !use_ptemod)
212 unmap_grant_pages(map, 0, map->count);
213 gntdev_free_map(map);
216 /* ------------------------------------------------------------------ */
218 static int find_grant_ptes(pte_t *pte, pgtable_t token,
219 unsigned long addr, void *data)
221 struct grant_map *map = data;
222 unsigned int pgnr = (addr - map->vma->vm_start) >> PAGE_SHIFT;
223 int flags = map->flags | GNTMAP_application_map | GNTMAP_contains_pte;
226 BUG_ON(pgnr >= map->count);
227 pte_maddr = arbitrary_virt_to_machine(pte).maddr;
229 gnttab_set_map_op(&map->map_ops[pgnr], pte_maddr, flags,
230 map->grants[pgnr].ref,
231 map->grants[pgnr].domid);
232 gnttab_set_unmap_op(&map->unmap_ops[pgnr], pte_maddr, flags,
237 static int map_grant_pages(struct grant_map *map)
242 /* Note: it could already be mapped */
243 if (map->map_ops[0].handle != -1)
245 for (i = 0; i < map->count; i++) {
246 unsigned long addr = (unsigned long)
247 pfn_to_kaddr(page_to_pfn(map->pages[i]));
248 gnttab_set_map_op(&map->map_ops[i], addr, map->flags,
250 map->grants[i].domid);
251 gnttab_set_unmap_op(&map->unmap_ops[i], addr,
252 map->flags, -1 /* handle */);
256 * Setup the map_ops corresponding to the pte entries pointing
257 * to the kernel linear addresses of the struct pages.
258 * These ptes are completely different from the user ptes dealt
259 * with find_grant_ptes.
261 for (i = 0; i < map->count; i++) {
263 unsigned long address = (unsigned long)
264 pfn_to_kaddr(page_to_pfn(map->pages[i]));
267 BUG_ON(PageHighMem(map->pages[i]));
269 ptep = lookup_address(address, &level);
270 pte_maddr = arbitrary_virt_to_machine(ptep).maddr;
271 gnttab_set_map_op(&map->kmap_ops[i], pte_maddr,
276 map->grants[i].domid);
280 pr_debug("map %d+%d\n", map->index, map->count);
281 err = gnttab_map_refs(map->map_ops, use_ptemod ? map->kmap_ops : NULL,
282 map->pages, map->count);
286 for (i = 0; i < map->count; i++) {
287 if (map->map_ops[i].status)
290 BUG_ON(map->map_ops[i].handle == -1);
291 map->unmap_ops[i].handle = map->map_ops[i].handle;
292 pr_debug("map handle=%d\n", map->map_ops[i].handle);
298 static int __unmap_grant_pages(struct grant_map *map, int offset, int pages)
302 if (map->notify.flags & UNMAP_NOTIFY_CLEAR_BYTE) {
303 int pgno = (map->notify.addr >> PAGE_SHIFT);
304 if (pgno >= offset && pgno < offset + pages && use_ptemod) {
305 void __user *tmp = (void __user *)
306 map->vma->vm_start + map->notify.addr;
307 err = copy_to_user(tmp, &err, 1);
310 map->notify.flags &= ~UNMAP_NOTIFY_CLEAR_BYTE;
311 } else if (pgno >= offset && pgno < offset + pages) {
312 uint8_t *tmp = kmap(map->pages[pgno]);
313 tmp[map->notify.addr & (PAGE_SIZE-1)] = 0;
314 kunmap(map->pages[pgno]);
315 map->notify.flags &= ~UNMAP_NOTIFY_CLEAR_BYTE;
319 err = gnttab_unmap_refs(map->unmap_ops + offset,
320 use_ptemod ? map->kmap_ops + offset : NULL, map->pages + offset,
325 for (i = 0; i < pages; i++) {
326 if (map->unmap_ops[offset+i].status)
328 pr_debug("unmap handle=%d st=%d\n",
329 map->unmap_ops[offset+i].handle,
330 map->unmap_ops[offset+i].status);
331 map->unmap_ops[offset+i].handle = -1;
336 static int unmap_grant_pages(struct grant_map *map, int offset, int pages)
340 pr_debug("unmap %d+%d [%d+%d]\n", map->index, map->count, offset, pages);
342 /* It is possible the requested range will have a "hole" where we
343 * already unmapped some of the grants. Only unmap valid ranges.
345 while (pages && !err) {
346 while (pages && map->unmap_ops[offset].handle == -1) {
351 while (range < pages) {
352 if (map->unmap_ops[offset+range].handle == -1) {
358 err = __unmap_grant_pages(map, offset, range);
366 /* ------------------------------------------------------------------ */
368 static void gntdev_vma_open(struct vm_area_struct *vma)
370 struct grant_map *map = vma->vm_private_data;
372 pr_debug("gntdev_vma_open %p\n", vma);
373 atomic_inc(&map->users);
376 static void gntdev_vma_close(struct vm_area_struct *vma)
378 struct grant_map *map = vma->vm_private_data;
380 pr_debug("gntdev_vma_close %p\n", vma);
382 struct file *file = vma->vm_file;
383 struct gntdev_priv *priv = file->private_data;
384 /* It is possible that an mmu notifier could be running
385 * concurrently, so take priv->lock to ensure that the vma won't
386 * vanishing during the unmap_grant_pages call, since we will
387 * spin here until that completes. Such a concurrent call will
388 * not do any unmapping, since that has been done prior to
389 * closing the vma, but it may still iterate the unmap_ops list.
391 spin_lock(&priv->lock);
393 spin_unlock(&priv->lock);
395 vma->vm_private_data = NULL;
399 static struct vm_operations_struct gntdev_vmops = {
400 .open = gntdev_vma_open,
401 .close = gntdev_vma_close,
404 /* ------------------------------------------------------------------ */
406 static void mn_invl_range_start(struct mmu_notifier *mn,
407 struct mm_struct *mm,
408 unsigned long start, unsigned long end)
410 struct gntdev_priv *priv = container_of(mn, struct gntdev_priv, mn);
411 struct grant_map *map;
412 unsigned long mstart, mend;
415 spin_lock(&priv->lock);
416 list_for_each_entry(map, &priv->maps, next) {
419 if (map->vma->vm_start >= end)
421 if (map->vma->vm_end <= start)
423 mstart = max(start, map->vma->vm_start);
424 mend = min(end, map->vma->vm_end);
425 pr_debug("map %d+%d (%lx %lx), range %lx %lx, mrange %lx %lx\n",
426 map->index, map->count,
427 map->vma->vm_start, map->vma->vm_end,
428 start, end, mstart, mend);
429 err = unmap_grant_pages(map,
430 (mstart - map->vma->vm_start) >> PAGE_SHIFT,
431 (mend - mstart) >> PAGE_SHIFT);
434 spin_unlock(&priv->lock);
437 static void mn_invl_page(struct mmu_notifier *mn,
438 struct mm_struct *mm,
439 unsigned long address)
441 mn_invl_range_start(mn, mm, address, address + PAGE_SIZE);
444 static void mn_release(struct mmu_notifier *mn,
445 struct mm_struct *mm)
447 struct gntdev_priv *priv = container_of(mn, struct gntdev_priv, mn);
448 struct grant_map *map;
451 spin_lock(&priv->lock);
452 list_for_each_entry(map, &priv->maps, next) {
455 pr_debug("map %d+%d (%lx %lx)\n",
456 map->index, map->count,
457 map->vma->vm_start, map->vma->vm_end);
458 err = unmap_grant_pages(map, /* offset */ 0, map->count);
461 spin_unlock(&priv->lock);
464 static struct mmu_notifier_ops gntdev_mmu_ops = {
465 .release = mn_release,
466 .invalidate_page = mn_invl_page,
467 .invalidate_range_start = mn_invl_range_start,
470 /* ------------------------------------------------------------------ */
472 static int gntdev_open(struct inode *inode, struct file *flip)
474 struct gntdev_priv *priv;
477 priv = kzalloc(sizeof(*priv), GFP_KERNEL);
481 INIT_LIST_HEAD(&priv->maps);
482 spin_lock_init(&priv->lock);
485 priv->mm = get_task_mm(current);
490 priv->mn.ops = &gntdev_mmu_ops;
491 ret = mmu_notifier_register(&priv->mn, priv->mm);
500 flip->private_data = priv;
501 pr_debug("priv %p\n", priv);
506 static int gntdev_release(struct inode *inode, struct file *flip)
508 struct gntdev_priv *priv = flip->private_data;
509 struct grant_map *map;
511 pr_debug("priv %p\n", priv);
513 while (!list_empty(&priv->maps)) {
514 map = list_entry(priv->maps.next, struct grant_map, next);
515 list_del(&map->next);
520 mmu_notifier_unregister(&priv->mn, priv->mm);
525 static long gntdev_ioctl_map_grant_ref(struct gntdev_priv *priv,
526 struct ioctl_gntdev_map_grant_ref __user *u)
528 struct ioctl_gntdev_map_grant_ref op;
529 struct grant_map *map;
532 if (copy_from_user(&op, u, sizeof(op)) != 0)
534 pr_debug("priv %p, add %d\n", priv, op.count);
535 if (unlikely(op.count <= 0))
539 map = gntdev_alloc_map(priv, op.count);
543 if (unlikely(atomic_add_return(op.count, &pages_mapped) > limit)) {
544 pr_debug("can't map: over limit\n");
549 if (copy_from_user(map->grants, &u->refs,
550 sizeof(map->grants[0]) * op.count) != 0) {
555 spin_lock(&priv->lock);
556 gntdev_add_map(priv, map);
557 op.index = map->index << PAGE_SHIFT;
558 spin_unlock(&priv->lock);
560 if (copy_to_user(u, &op, sizeof(op)) != 0)
566 static long gntdev_ioctl_unmap_grant_ref(struct gntdev_priv *priv,
567 struct ioctl_gntdev_unmap_grant_ref __user *u)
569 struct ioctl_gntdev_unmap_grant_ref op;
570 struct grant_map *map;
573 if (copy_from_user(&op, u, sizeof(op)) != 0)
575 pr_debug("priv %p, del %d+%d\n", priv, (int)op.index, (int)op.count);
577 spin_lock(&priv->lock);
578 map = gntdev_find_map_index(priv, op.index >> PAGE_SHIFT, op.count);
580 list_del(&map->next);
583 spin_unlock(&priv->lock);
589 static long gntdev_ioctl_get_offset_for_vaddr(struct gntdev_priv *priv,
590 struct ioctl_gntdev_get_offset_for_vaddr __user *u)
592 struct ioctl_gntdev_get_offset_for_vaddr op;
593 struct vm_area_struct *vma;
594 struct grant_map *map;
597 if (copy_from_user(&op, u, sizeof(op)) != 0)
599 pr_debug("priv %p, offset for vaddr %lx\n", priv, (unsigned long)op.vaddr);
601 down_read(¤t->mm->mmap_sem);
602 vma = find_vma(current->mm, op.vaddr);
603 if (!vma || vma->vm_ops != &gntdev_vmops)
606 map = vma->vm_private_data;
610 op.offset = map->index << PAGE_SHIFT;
611 op.count = map->count;
615 up_read(¤t->mm->mmap_sem);
617 if (rv == 0 && copy_to_user(u, &op, sizeof(op)) != 0)
622 static long gntdev_ioctl_notify(struct gntdev_priv *priv, void __user *u)
624 struct ioctl_gntdev_unmap_notify op;
625 struct grant_map *map;
628 unsigned int out_event;
630 if (copy_from_user(&op, u, sizeof(op)))
633 if (op.action & ~(UNMAP_NOTIFY_CLEAR_BYTE|UNMAP_NOTIFY_SEND_EVENT))
636 /* We need to grab a reference to the event channel we are going to use
637 * to send the notify before releasing the reference we may already have
638 * (if someone has called this ioctl twice). This is required so that
639 * it is possible to change the clear_byte part of the notification
640 * without disturbing the event channel part, which may now be the last
641 * reference to that event channel.
643 if (op.action & UNMAP_NOTIFY_SEND_EVENT) {
644 if (evtchn_get(op.event_channel_port))
648 out_flags = op.action;
649 out_event = op.event_channel_port;
651 spin_lock(&priv->lock);
653 list_for_each_entry(map, &priv->maps, next) {
654 uint64_t begin = map->index << PAGE_SHIFT;
655 uint64_t end = (map->index + map->count) << PAGE_SHIFT;
656 if (op.index >= begin && op.index < end)
663 if ((op.action & UNMAP_NOTIFY_CLEAR_BYTE) &&
664 (map->flags & GNTMAP_readonly)) {
669 out_flags = map->notify.flags;
670 out_event = map->notify.event;
672 map->notify.flags = op.action;
673 map->notify.addr = op.index - (map->index << PAGE_SHIFT);
674 map->notify.event = op.event_channel_port;
679 spin_unlock(&priv->lock);
681 /* Drop the reference to the event channel we did not save in the map */
682 if (out_flags & UNMAP_NOTIFY_SEND_EVENT)
683 evtchn_put(out_event);
688 static long gntdev_ioctl(struct file *flip,
689 unsigned int cmd, unsigned long arg)
691 struct gntdev_priv *priv = flip->private_data;
692 void __user *ptr = (void __user *)arg;
695 case IOCTL_GNTDEV_MAP_GRANT_REF:
696 return gntdev_ioctl_map_grant_ref(priv, ptr);
698 case IOCTL_GNTDEV_UNMAP_GRANT_REF:
699 return gntdev_ioctl_unmap_grant_ref(priv, ptr);
701 case IOCTL_GNTDEV_GET_OFFSET_FOR_VADDR:
702 return gntdev_ioctl_get_offset_for_vaddr(priv, ptr);
704 case IOCTL_GNTDEV_SET_UNMAP_NOTIFY:
705 return gntdev_ioctl_notify(priv, ptr);
708 pr_debug("priv %p, unknown cmd %x\n", priv, cmd);
715 static int gntdev_mmap(struct file *flip, struct vm_area_struct *vma)
717 struct gntdev_priv *priv = flip->private_data;
718 int index = vma->vm_pgoff;
719 int count = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
720 struct grant_map *map;
721 int i, err = -EINVAL;
723 if ((vma->vm_flags & VM_WRITE) && !(vma->vm_flags & VM_SHARED))
726 pr_debug("map %d+%d at %lx (pgoff %lx)\n",
727 index, count, vma->vm_start, vma->vm_pgoff);
729 spin_lock(&priv->lock);
730 map = gntdev_find_map_index(priv, index, count);
733 if (use_ptemod && map->vma)
735 if (use_ptemod && priv->mm != vma->vm_mm) {
736 printk(KERN_WARNING "Huh? Other mm?\n");
740 atomic_inc(&map->users);
742 vma->vm_ops = &gntdev_vmops;
744 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
747 vma->vm_flags |= VM_DONTCOPY;
749 vma->vm_private_data = map;
755 if ((vma->vm_flags & VM_WRITE) &&
756 (map->flags & GNTMAP_readonly))
759 map->flags = GNTMAP_host_map;
760 if (!(vma->vm_flags & VM_WRITE))
761 map->flags |= GNTMAP_readonly;
764 spin_unlock(&priv->lock);
767 err = apply_to_page_range(vma->vm_mm, vma->vm_start,
768 vma->vm_end - vma->vm_start,
769 find_grant_ptes, map);
771 printk(KERN_WARNING "find_grant_ptes() failure.\n");
776 err = map_grant_pages(map);
781 for (i = 0; i < count; i++) {
782 err = vm_insert_page(vma, vma->vm_start + i*PAGE_SIZE,
792 spin_unlock(&priv->lock);
796 spin_unlock(&priv->lock);
804 static const struct file_operations gntdev_fops = {
805 .owner = THIS_MODULE,
807 .release = gntdev_release,
809 .unlocked_ioctl = gntdev_ioctl
812 static struct miscdevice gntdev_miscdev = {
813 .minor = MISC_DYNAMIC_MINOR,
814 .name = "xen/gntdev",
815 .fops = &gntdev_fops,
818 /* ------------------------------------------------------------------ */
820 static int __init gntdev_init(void)
827 use_ptemod = xen_pv_domain();
829 err = misc_register(&gntdev_miscdev);
831 printk(KERN_ERR "Could not register gntdev device\n");
837 static void __exit gntdev_exit(void)
839 misc_deregister(&gntdev_miscdev);
842 module_init(gntdev_init);
843 module_exit(gntdev_exit);
845 /* ------------------------------------------------------------------ */