drivers/xen/gntdev.c

   1 /******************************************************************************
   2  * gntdev.c
   3  *
   4  * Device for accessing (in user-space) pages that have been granted by other
   5  * domains.
   6  *
   7  * Copyright (c) 2006-2007, D G Murray.
   8  *           (c) 2009 Gerd Hoffmann <kraxel@redhat.com>
   9  *
  10  * This program is distributed in the hope that it will be useful,
  11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  * GNU General Public License for more details.
  14  *
  15  * You should have received a copy of the GNU General Public License
  16  * along with this program; if not, write to the Free Software
  17  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  18  */
  19
  20 #undef DEBUG
  21
  22 #define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt
  23
  24 #include <linux/module.h>
  25 #include <linux/kernel.h>
  26 #include <linux/init.h>
  27 #include <linux/miscdevice.h>
  28 #include <linux/fs.h>
  29 #include <linux/mm.h>
  30 #include <linux/mman.h>
  31 #include <linux/mmu_notifier.h>
  32 #include <linux/types.h>
  33 #include <linux/uaccess.h>
  34 #include <linux/sched.h>
  35 #include <linux/sched/mm.h>
  36 #include <linux/spinlock.h>
  37 #include <linux/slab.h>
  38 #include <linux/highmem.h>
  39 #include <linux/refcount.h>
  40
  41 #include <xen/xen.h>
  42 #include <xen/grant_table.h>
  43 #include <xen/balloon.h>
  44 #include <xen/gntdev.h>
  45 #include <xen/events.h>
  46 #include <xen/page.h>
  47 #include <asm/xen/hypervisor.h>
  48 #include <asm/xen/hypercall.h>
  49
  50 MODULE_LICENSE("GPL");
  51 MODULE_AUTHOR("Derek G. Murray <Derek.Murray@cl.cam.ac.uk>, "
  52               "Gerd Hoffmann <kraxel@redhat.com>");
  53 MODULE_DESCRIPTION("User-space granted page access driver");
  54
  55 static int limit = 1024*1024;
  56 module_param(limit, int, 0644);
  57 MODULE_PARM_DESC(limit, "Maximum number of grants that may be mapped by "
  58                 "the gntdev device");
  59
  60 static atomic_t pages_mapped = ATOMIC_INIT(0);
  61
  62 static int use_ptemod;
  63 #define populate_freeable_maps use_ptemod
  64
  65 struct gntdev_priv {
  66         /* maps with visible offsets in the file descriptor */
  67         struct list_head maps;
  68         /* maps that are not visible; will be freed on munmap.
  69          * Only populated if populate_freeable_maps == 1 */
  70         struct list_head freeable_maps;
  71         /* lock protects maps and freeable_maps */
  72         struct mutex lock;
  73         struct mm_struct *mm;
  74         struct mmu_notifier mn;
  75 };
  76
  77 struct unmap_notify {
  78         int flags;
  79         /* Address relative to the start of the grant_map */
  80         int addr;
  81         int event;
  82 };
  83
  84 struct grant_map {
  85         struct list_head next;
  86         struct vm_area_struct *vma;
  87         int index;
  88         int count;
  89         int flags;
  90         refcount_t users;
  91         struct unmap_notify notify;
  92         struct ioctl_gntdev_grant_ref *grants;
  93         struct gnttab_map_grant_ref   *map_ops;
  94         struct gnttab_unmap_grant_ref *unmap_ops;
  95         struct gnttab_map_grant_ref   *kmap_ops;
  96         struct gnttab_unmap_grant_ref *kunmap_ops;
  97         struct page **pages;
  98         unsigned long pages_vm_start;
  99 };
 100
 101 static int unmap_grant_pages(struct grant_map *map, int offset, int pages);
 102
 103 /* ------------------------------------------------------------------ */
 104
 105 static void gntdev_print_maps(struct gntdev_priv *priv,
 106                               char *text, int text_index)
 107 {
 108 #ifdef DEBUG
 109         struct grant_map *map;
 110
 111         pr_debug("%s: maps list (priv %p)\n", __func__, priv);
 112         list_for_each_entry(map, &priv->maps, next)
 113                 pr_debug("  index %2d, count %2d %s\n",
 114                        map->index, map->count,
 115                        map->index == text_index && text ? text : "");
 116 #endif
 117 }
 118
 119 static void gntdev_free_map(struct grant_map *map)
 120 {
 121         if (map == NULL)
 122                 return;
 123
 124         if (map->pages)
 125                 gnttab_free_pages(map->count, map->pages);
 126         kfree(map->pages);
 127         kfree(map->grants);
 128         kfree(map->map_ops);
 129         kfree(map->unmap_ops);
 130         kfree(map->kmap_ops);
 131         kfree(map->kunmap_ops);
 132         kfree(map);
 133 }
 134
 135 static struct grant_map *gntdev_alloc_map(struct gntdev_priv *priv, int count)
 136 {
 137         struct grant_map *add;
 138         int i;
 139
 140         add = kzalloc(sizeof(struct grant_map), GFP_KERNEL);
 141         if (NULL == add)
 142                 return NULL;
 143
 144         add->grants    = kcalloc(count, sizeof(add->grants[0]), GFP_KERNEL);
 145         add->map_ops   = kcalloc(count, sizeof(add->map_ops[0]), GFP_KERNEL);
 146         add->unmap_ops = kcalloc(count, sizeof(add->unmap_ops[0]), GFP_KERNEL);
 147         add->kmap_ops  = kcalloc(count, sizeof(add->kmap_ops[0]), GFP_KERNEL);
 148         add->kunmap_ops = kcalloc(count, sizeof(add->kunmap_ops[0]), GFP_KERNEL);
 149         add->pages     = kcalloc(count, sizeof(add->pages[0]), GFP_KERNEL);
 150         if (NULL == add->grants    ||
 151             NULL == add->map_ops   ||
 152             NULL == add->unmap_ops ||
 153             NULL == add->kmap_ops  ||
 154             NULL == add->kunmap_ops ||
 155             NULL == add->pages)
 156                 goto err;
 157
 158         if (gnttab_alloc_pages(count, add->pages))
 159                 goto err;
 160
 161         for (i = 0; i < count; i++) {
 162                 add->map_ops[i].handle = -1;
 163                 add->unmap_ops[i].handle = -1;
 164                 add->kmap_ops[i].handle = -1;
 165                 add->kunmap_ops[i].handle = -1;
 166         }
 167
 168         add->index = 0;
 169         add->count = count;
 170         refcount_set(&add->users, 1);
 171
 172         return add;
 173
 174 err:
 175         gntdev_free_map(add);
 176         return NULL;
 177 }
 178
 179 static void gntdev_add_map(struct gntdev_priv *priv, struct grant_map *add)
 180 {
 181         struct grant_map *map;
 182
 183         list_for_each_entry(map, &priv->maps, next) {
 184                 if (add->index + add->count < map->index) {
 185                         list_add_tail(&add->next, &map->next);
 186                         goto done;
 187                 }
 188                 add->index = map->index + map->count;
 189         }
 190         list_add_tail(&add->next, &priv->maps);
 191
 192 done:
 193         gntdev_print_maps(priv, "[new]", add->index);
 194 }
 195
 196 static struct grant_map *gntdev_find_map_index(struct gntdev_priv *priv,
 197                 int index, int count)
 198 {
 199         struct grant_map *map;
 200
 201         list_for_each_entry(map, &priv->maps, next) {
 202                 if (map->index != index)
 203                         continue;
 204                 if (count && map->count != count)
 205                         continue;
 206                 return map;
 207         }
 208         return NULL;
 209 }
 210
 211 static void gntdev_put_map(struct gntdev_priv *priv, struct grant_map *map)
 212 {
 213         if (!map)
 214                 return;
 215
 216         if (!refcount_dec_and_test(&map->users))
 217                 return;
 218
 219         atomic_sub(map->count, &pages_mapped);
 220
 221         if (map->notify.flags & UNMAP_NOTIFY_SEND_EVENT) {
 222                 notify_remote_via_evtchn(map->notify.event);
 223                 evtchn_put(map->notify.event);
 224         }
 225
 226         if (populate_freeable_maps && priv) {
 227                 mutex_lock(&priv->lock);
 228                 list_del(&map->next);
 229                 mutex_unlock(&priv->lock);
 230         }
 231
 232         if (map->pages && !use_ptemod)
 233                 unmap_grant_pages(map, 0, map->count);
 234         gntdev_free_map(map);
 235 }
 236
 237 /* ------------------------------------------------------------------ */
 238
 239 static int find_grant_ptes(pte_t *pte, pgtable_t token,
 240                 unsigned long addr, void *data)
 241 {
 242         struct grant_map *map = data;
 243         unsigned int pgnr = (addr - map->vma->vm_start) >> PAGE_SHIFT;
 244         int flags = map->flags | GNTMAP_application_map | GNTMAP_contains_pte;
 245         u64 pte_maddr;
 246
 247         BUG_ON(pgnr >= map->count);
 248         pte_maddr = arbitrary_virt_to_machine(pte).maddr;
 249
 250         /*
 251          * Set the PTE as special to force get_user_pages_fast() fall
 252          * back to the slow path.  If this is not supported as part of
 253          * the grant map, it will be done afterwards.
 254          */
 255         if (xen_feature(XENFEAT_gnttab_map_avail_bits))
 256                 flags |= (1 << _GNTMAP_guest_avail0);
 257
 258         gnttab_set_map_op(&map->map_ops[pgnr], pte_maddr, flags,
 259                           map->grants[pgnr].ref,
 260                           map->grants[pgnr].domid);
 261         gnttab_set_unmap_op(&map->unmap_ops[pgnr], pte_maddr, flags,
 262                             -1 /* handle */);
 263         return 0;
 264 }
 265
 266 #ifdef CONFIG_X86
 267 static int set_grant_ptes_as_special(pte_t *pte, pgtable_t token,
 268                                      unsigned long addr, void *data)
 269 {
 270         set_pte_at(current->mm, addr, pte, pte_mkspecial(*pte));
 271         return 0;
 272 }
 273 #endif
 274
 275 static int map_grant_pages(struct grant_map *map)
 276 {
 277         int i, err = 0;
 278
 279         if (!use_ptemod) {
 280                 /* Note: it could already be mapped */
 281                 if (map->map_ops[0].handle != -1)
 282                         return 0;
 283                 for (i = 0; i < map->count; i++) {
 284                         unsigned long addr = (unsigned long)
 285                                 pfn_to_kaddr(page_to_pfn(map->pages[i]));
 286                         gnttab_set_map_op(&map->map_ops[i], addr, map->flags,
 287                                 map->grants[i].ref,
 288                                 map->grants[i].domid);
 289                         gnttab_set_unmap_op(&map->unmap_ops[i], addr,
 290                                 map->flags, -1 /* handle */);
 291                 }
 292         } else {
 293                 /*
 294                  * Setup the map_ops corresponding to the pte entries pointing
 295                  * to the kernel linear addresses of the struct pages.
 296                  * These ptes are completely different from the user ptes dealt
 297                  * with find_grant_ptes.
 298                  */
 299                 for (i = 0; i < map->count; i++) {
 300                         unsigned long address = (unsigned long)
 301                                 pfn_to_kaddr(page_to_pfn(map->pages[i]));
 302                         BUG_ON(PageHighMem(map->pages[i]));
 303
 304                         gnttab_set_map_op(&map->kmap_ops[i], address,
 305                                 map->flags | GNTMAP_host_map,
 306                                 map->grants[i].ref,
 307                                 map->grants[i].domid);
 308                         gnttab_set_unmap_op(&map->kunmap_ops[i], address,
 309                                 map->flags | GNTMAP_host_map, -1);
 310                 }
 311         }
 312
 313         pr_debug("map %d+%d\n", map->index, map->count);
 314         err = gnttab_map_refs(map->map_ops, use_ptemod ? map->kmap_ops : NULL,
 315                         map->pages, map->count);
 316         if (err)
 317                 return err;
 318
 319         for (i = 0; i < map->count; i++) {
 320                 if (map->map_ops[i].status) {
 321                         err = -EINVAL;
 322                         continue;
 323                 }
 324
 325                 map->unmap_ops[i].handle = map->map_ops[i].handle;
 326                 if (use_ptemod)
 327                         map->kunmap_ops[i].handle = map->kmap_ops[i].handle;
 328         }
 329         return err;
 330 }
 331
 332 static int __unmap_grant_pages(struct grant_map *map, int offset, int pages)
 333 {
 334         int i, err = 0;
 335         struct gntab_unmap_queue_data unmap_data;
 336
 337         if (map->notify.flags & UNMAP_NOTIFY_CLEAR_BYTE) {
 338                 int pgno = (map->notify.addr >> PAGE_SHIFT);
 339                 if (pgno >= offset && pgno < offset + pages) {
 340                         /* No need for kmap, pages are in lowmem */
 341                         uint8_t *tmp = pfn_to_kaddr(page_to_pfn(map->pages[pgno]));
 342                         tmp[map->notify.addr & (PAGE_SIZE-1)] = 0;
 343                         map->notify.flags &= ~UNMAP_NOTIFY_CLEAR_BYTE;
 344                 }
 345         }
 346
 347         unmap_data.unmap_ops = map->unmap_ops + offset;
 348         unmap_data.kunmap_ops = use_ptemod ? map->kunmap_ops + offset : NULL;
 349         unmap_data.pages = map->pages + offset;
 350         unmap_data.count = pages;
 351
 352         err = gnttab_unmap_refs_sync(&unmap_data);
 353         if (err)
 354                 return err;
 355
 356         for (i = 0; i < pages; i++) {
 357                 if (map->unmap_ops[offset+i].status)
 358                         err = -EINVAL;
 359                 pr_debug("unmap handle=%d st=%d\n",
 360                         map->unmap_ops[offset+i].handle,
 361                         map->unmap_ops[offset+i].status);
 362                 map->unmap_ops[offset+i].handle = -1;
 363         }
 364         return err;
 365 }
 366
 367 static int unmap_grant_pages(struct grant_map *map, int offset, int pages)
 368 {
 369         int range, err = 0;
 370
 371         pr_debug("unmap %d+%d [%d+%d]\n", map->index, map->count, offset, pages);
 372
 373         /* It is possible the requested range will have a "hole" where we
 374          * already unmapped some of the grants. Only unmap valid ranges.
 375          */
 376         while (pages && !err) {
 377                 while (pages && map->unmap_ops[offset].handle == -1) {
 378                         offset++;
 379                         pages--;
 380                 }
 381                 range = 0;
 382                 while (range < pages) {
 383                         if (map->unmap_ops[offset+range].handle == -1) {
 384                                 range--;
 385                                 break;
 386                         }
 387                         range++;
 388                 }
 389                 err = __unmap_grant_pages(map, offset, range);
 390                 offset += range;
 391                 pages -= range;
 392         }
 393
 394         return err;
 395 }
 396
 397 /* ------------------------------------------------------------------ */
 398
 399 static void gntdev_vma_open(struct vm_area_struct *vma)
 400 {
 401         struct grant_map *map = vma->vm_private_data;
 402
 403         pr_debug("gntdev_vma_open %p\n", vma);
 404         refcount_inc(&map->users);
 405 }
 406
 407 static void gntdev_vma_close(struct vm_area_struct *vma)
 408 {
 409         struct grant_map *map = vma->vm_private_data;
 410         struct file *file = vma->vm_file;
 411         struct gntdev_priv *priv = file->private_data;
 412
 413         pr_debug("gntdev_vma_close %p\n", vma);
 414         if (use_ptemod) {
 415                 /* It is possible that an mmu notifier could be running
 416                  * concurrently, so take priv->lock to ensure that the vma won't
 417                  * vanishing during the unmap_grant_pages call, since we will
 418                  * spin here until that completes. Such a concurrent call will
 419                  * not do any unmapping, since that has been done prior to
 420                  * closing the vma, but it may still iterate the unmap_ops list.
 421                  */
 422                 mutex_lock(&priv->lock);
 423                 map->vma = NULL;
 424                 mutex_unlock(&priv->lock);
 425         }
 426         vma->vm_private_data = NULL;
 427         gntdev_put_map(priv, map);
 428 }
 429
 430 static struct page *gntdev_vma_find_special_page(struct vm_area_struct *vma,
 431                                                  unsigned long addr)
 432 {
 433         struct grant_map *map = vma->vm_private_data;
 434
 435         return map->pages[(addr - map->pages_vm_start) >> PAGE_SHIFT];
 436 }
 437
 438 static const struct vm_operations_struct gntdev_vmops = {
 439         .open = gntdev_vma_open,
 440         .close = gntdev_vma_close,
 441         .find_special_page = gntdev_vma_find_special_page,
 442 };
 443
 444 /* ------------------------------------------------------------------ */
 445
 446 static void unmap_if_in_range(struct grant_map *map,
 447                               unsigned long start, unsigned long end)
 448 {
 449         unsigned long mstart, mend;
 450         int err;
 451
 452         if (!map->vma)
 453                 return;
 454         if (map->vma->vm_start >= end)
 455                 return;
 456         if (map->vma->vm_end <= start)
 457                 return;
 458         mstart = max(start, map->vma->vm_start);
 459         mend   = min(end,   map->vma->vm_end);
 460         pr_debug("map %d+%d (%lx %lx), range %lx %lx, mrange %lx %lx\n",
 461                         map->index, map->count,
 462                         map->vma->vm_start, map->vma->vm_end,
 463                         start, end, mstart, mend);
 464         err = unmap_grant_pages(map,
 465                                 (mstart - map->vma->vm_start) >> PAGE_SHIFT,
 466                                 (mend - mstart) >> PAGE_SHIFT);
 467         WARN_ON(err);
 468 }
 469
 470 static void mn_invl_range_start(struct mmu_notifier *mn,
 471                                 struct mm_struct *mm,
 472                                 unsigned long start, unsigned long end)
 473 {
 474         struct gntdev_priv *priv = container_of(mn, struct gntdev_priv, mn);
 475         struct grant_map *map;
 476
 477         mutex_lock(&priv->lock);
 478         list_for_each_entry(map, &priv->maps, next) {
 479                 unmap_if_in_range(map, start, end);
 480         }
 481         list_for_each_entry(map, &priv->freeable_maps, next) {
 482                 unmap_if_in_range(map, start, end);
 483         }
 484         mutex_unlock(&priv->lock);
 485 }
 486
 487 static void mn_invl_page(struct mmu_notifier *mn,
 488                          struct mm_struct *mm,
 489                          unsigned long address)
 490 {
 491         mn_invl_range_start(mn, mm, address, address + PAGE_SIZE);
 492 }
 493
 494 static void mn_release(struct mmu_notifier *mn,
 495                        struct mm_struct *mm)
 496 {
 497         struct gntdev_priv *priv = container_of(mn, struct gntdev_priv, mn);
 498         struct grant_map *map;
 499         int err;
 500
 501         mutex_lock(&priv->lock);
 502         list_for_each_entry(map, &priv->maps, next) {
 503                 if (!map->vma)
 504                         continue;
 505                 pr_debug("map %d+%d (%lx %lx)\n",
 506                                 map->index, map->count,
 507                                 map->vma->vm_start, map->vma->vm_end);
 508                 err = unmap_grant_pages(map, /* offset */ 0, map->count);
 509                 WARN_ON(err);
 510         }
 511         list_for_each_entry(map, &priv->freeable_maps, next) {
 512                 if (!map->vma)
 513                         continue;
 514                 pr_debug("map %d+%d (%lx %lx)\n",
 515                                 map->index, map->count,
 516                                 map->vma->vm_start, map->vma->vm_end);
 517                 err = unmap_grant_pages(map, /* offset */ 0, map->count);
 518                 WARN_ON(err);
 519         }
 520         mutex_unlock(&priv->lock);
 521 }
 522
 523 static const struct mmu_notifier_ops gntdev_mmu_ops = {
 524         .release                = mn_release,
 525         .invalidate_page        = mn_invl_page,
 526         .invalidate_range_start = mn_invl_range_start,
 527 };
 528
 529 /* ------------------------------------------------------------------ */
 530
 531 static int gntdev_open(struct inode *inode, struct file *flip)
 532 {
 533         struct gntdev_priv *priv;
 534         int ret = 0;
 535
 536         priv = kzalloc(sizeof(*priv), GFP_KERNEL);
 537         if (!priv)
 538                 return -ENOMEM;
 539
 540         INIT_LIST_HEAD(&priv->maps);
 541         INIT_LIST_HEAD(&priv->freeable_maps);
 542         mutex_init(&priv->lock);
 543
 544         if (use_ptemod) {
 545                 priv->mm = get_task_mm(current);
 546                 if (!priv->mm) {
 547                         kfree(priv);
 548                         return -ENOMEM;
 549                 }
 550                 priv->mn.ops = &gntdev_mmu_ops;
 551                 ret = mmu_notifier_register(&priv->mn, priv->mm);
 552                 mmput(priv->mm);
 553         }
 554
 555         if (ret) {
 556                 kfree(priv);
 557                 return ret;
 558         }
 559
 560         flip->private_data = priv;
 561         pr_debug("priv %p\n", priv);
 562
 563         return 0;
 564 }
 565
 566 static int gntdev_release(struct inode *inode, struct file *flip)
 567 {
 568         struct gntdev_priv *priv = flip->private_data;
 569         struct grant_map *map;
 570
 571         pr_debug("priv %p\n", priv);
 572
 573         mutex_lock(&priv->lock);
 574         while (!list_empty(&priv->maps)) {
 575                 map = list_entry(priv->maps.next, struct grant_map, next);
 576                 list_del(&map->next);
 577                 gntdev_put_map(NULL /* already removed */, map);
 578         }
 579         WARN_ON(!list_empty(&priv->freeable_maps));
 580         mutex_unlock(&priv->lock);
 581
 582         if (use_ptemod)
 583                 mmu_notifier_unregister(&priv->mn, priv->mm);
 584         kfree(priv);
 585         return 0;
 586 }
 587
 588 static long gntdev_ioctl_map_grant_ref(struct gntdev_priv *priv,
 589                                        struct ioctl_gntdev_map_grant_ref __user *u)
 590 {
 591         struct ioctl_gntdev_map_grant_ref op;
 592         struct grant_map *map;
 593         int err;
 594
 595         if (copy_from_user(&op, u, sizeof(op)) != 0)
 596                 return -EFAULT;
 597         pr_debug("priv %p, add %d\n", priv, op.count);
 598         if (unlikely(op.count <= 0))
 599                 return -EINVAL;
 600
 601         err = -ENOMEM;
 602         map = gntdev_alloc_map(priv, op.count);
 603         if (!map)
 604                 return err;
 605
 606         if (unlikely(atomic_add_return(op.count, &pages_mapped) > limit)) {
 607                 pr_debug("can't map: over limit\n");
 608                 gntdev_put_map(NULL, map);
 609                 return err;
 610         }
 611
 612         if (copy_from_user(map->grants, &u->refs,
 613                            sizeof(map->grants[0]) * op.count) != 0) {
 614                 gntdev_put_map(NULL, map);
 615                 return -EFAULT;
 616         }
 617
 618         mutex_lock(&priv->lock);
 619         gntdev_add_map(priv, map);
 620         op.index = map->index << PAGE_SHIFT;
 621         mutex_unlock(&priv->lock);
 622
 623         if (copy_to_user(u, &op, sizeof(op)) != 0)
 624                 return -EFAULT;
 625
 626         return 0;
 627 }
 628
 629 static long gntdev_ioctl_unmap_grant_ref(struct gntdev_priv *priv,
 630                                          struct ioctl_gntdev_unmap_grant_ref __user *u)
 631 {
 632         struct ioctl_gntdev_unmap_grant_ref op;
 633         struct grant_map *map;
 634         int err = -ENOENT;
 635
 636         if (copy_from_user(&op, u, sizeof(op)) != 0)
 637                 return -EFAULT;
 638         pr_debug("priv %p, del %d+%d\n", priv, (int)op.index, (int)op.count);
 639
 640         mutex_lock(&priv->lock);
 641         map = gntdev_find_map_index(priv, op.index >> PAGE_SHIFT, op.count);
 642         if (map) {
 643                 list_del(&map->next);
 644                 if (populate_freeable_maps)
 645                         list_add_tail(&map->next, &priv->freeable_maps);
 646                 err = 0;
 647         }
 648         mutex_unlock(&priv->lock);
 649         if (map)
 650                 gntdev_put_map(priv, map);
 651         return err;
 652 }
 653
 654 static long gntdev_ioctl_get_offset_for_vaddr(struct gntdev_priv *priv,
 655                                               struct ioctl_gntdev_get_offset_for_vaddr __user *u)
 656 {
 657         struct ioctl_gntdev_get_offset_for_vaddr op;
 658         struct vm_area_struct *vma;
 659         struct grant_map *map;
 660         int rv = -EINVAL;
 661
 662         if (copy_from_user(&op, u, sizeof(op)) != 0)
 663                 return -EFAULT;
 664         pr_debug("priv %p, offset for vaddr %lx\n", priv, (unsigned long)op.vaddr);
 665
 666         down_read(&current->mm->mmap_sem);
 667         vma = find_vma(current->mm, op.vaddr);
 668         if (!vma || vma->vm_ops != &gntdev_vmops)
 669                 goto out_unlock;
 670
 671         map = vma->vm_private_data;
 672         if (!map)
 673                 goto out_unlock;
 674
 675         op.offset = map->index << PAGE_SHIFT;
 676         op.count = map->count;
 677         rv = 0;
 678
 679  out_unlock:
 680         up_read(&current->mm->mmap_sem);
 681
 682         if (rv == 0 && copy_to_user(u, &op, sizeof(op)) != 0)
 683                 return -EFAULT;
 684         return rv;
 685 }
 686
 687 static long gntdev_ioctl_notify(struct gntdev_priv *priv, void __user *u)
 688 {
 689         struct ioctl_gntdev_unmap_notify op;
 690         struct grant_map *map;
 691         int rc;
 692         int out_flags;
 693         unsigned int out_event;
 694
 695         if (copy_from_user(&op, u, sizeof(op)))
 696                 return -EFAULT;
 697
 698         if (op.action & ~(UNMAP_NOTIFY_CLEAR_BYTE|UNMAP_NOTIFY_SEND_EVENT))
 699                 return -EINVAL;
 700
 701         /* We need to grab a reference to the event channel we are going to use
 702          * to send the notify before releasing the reference we may already have
 703          * (if someone has called this ioctl twice). This is required so that
 704          * it is possible to change the clear_byte part of the notification
 705          * without disturbing the event channel part, which may now be the last
 706          * reference to that event channel.
 707          */
 708         if (op.action & UNMAP_NOTIFY_SEND_EVENT) {
 709                 if (evtchn_get(op.event_channel_port))
 710                         return -EINVAL;
 711         }
 712
 713         out_flags = op.action;
 714         out_event = op.event_channel_port;
 715
 716         mutex_lock(&priv->lock);
 717
 718         list_for_each_entry(map, &priv->maps, next) {
 719                 uint64_t begin = map->index << PAGE_SHIFT;
 720                 uint64_t end = (map->index + map->count) << PAGE_SHIFT;
 721                 if (op.index >= begin && op.index < end)
 722                         goto found;
 723         }
 724         rc = -ENOENT;
 725         goto unlock_out;
 726
 727  found:
 728         if ((op.action & UNMAP_NOTIFY_CLEAR_BYTE) &&
 729                         (map->flags & GNTMAP_readonly)) {
 730                 rc = -EINVAL;
 731                 goto unlock_out;
 732         }
 733
 734         out_flags = map->notify.flags;
 735         out_event = map->notify.event;
 736
 737         map->notify.flags = op.action;
 738         map->notify.addr = op.index - (map->index << PAGE_SHIFT);
 739         map->notify.event = op.event_channel_port;
 740
 741         rc = 0;
 742
 743  unlock_out:
 744         mutex_unlock(&priv->lock);
 745
 746         /* Drop the reference to the event channel we did not save in the map */
 747         if (out_flags & UNMAP_NOTIFY_SEND_EVENT)
 748                 evtchn_put(out_event);
 749
 750         return rc;
 751 }
 752
 753 #define GNTDEV_COPY_BATCH 16
 754
 755 struct gntdev_copy_batch {
 756         struct gnttab_copy ops[GNTDEV_COPY_BATCH];
 757         struct page *pages[GNTDEV_COPY_BATCH];
 758         s16 __user *status[GNTDEV_COPY_BATCH];
 759         unsigned int nr_ops;
 760         unsigned int nr_pages;
 761 };
 762
 763 static int gntdev_get_page(struct gntdev_copy_batch *batch, void __user *virt,
 764                            bool writeable, unsigned long *gfn)
 765 {
 766         unsigned long addr = (unsigned long)virt;
 767         struct page *page;
 768         unsigned long xen_pfn;
 769         int ret;
 770
 771         ret = get_user_pages_fast(addr, 1, writeable, &page);
 772         if (ret < 0)
 773                 return ret;
 774
 775         batch->pages[batch->nr_pages++] = page;
 776
 777         xen_pfn = page_to_xen_pfn(page) + XEN_PFN_DOWN(addr & ~PAGE_MASK);
 778         *gfn = pfn_to_gfn(xen_pfn);
 779
 780         return 0;
 781 }
 782
 783 static void gntdev_put_pages(struct gntdev_copy_batch *batch)
 784 {
 785         unsigned int i;
 786
 787         for (i = 0; i < batch->nr_pages; i++)
 788                 put_page(batch->pages[i]);
 789         batch->nr_pages = 0;
 790 }
 791
 792 static int gntdev_copy(struct gntdev_copy_batch *batch)
 793 {
 794         unsigned int i;
 795
 796         gnttab_batch_copy(batch->ops, batch->nr_ops);
 797         gntdev_put_pages(batch);
 798
 799         /*
 800          * For each completed op, update the status if the op failed
 801          * and all previous ops for the segment were successful.
 802          */
 803         for (i = 0; i < batch->nr_ops; i++) {
 804                 s16 status = batch->ops[i].status;
 805                 s16 old_status;
 806
 807                 if (status == GNTST_okay)
 808                         continue;
 809
 810                 if (__get_user(old_status, batch->status[i]))
 811                         return -EFAULT;
 812
 813                 if (old_status != GNTST_okay)
 814                         continue;
 815
 816                 if (__put_user(status, batch->status[i]))
 817                         return -EFAULT;
 818         }
 819
 820         batch->nr_ops = 0;
 821         return 0;
 822 }
 823
 824 static int gntdev_grant_copy_seg(struct gntdev_copy_batch *batch,
 825                                  struct gntdev_grant_copy_segment *seg,
 826                                  s16 __user *status)
 827 {
 828         uint16_t copied = 0;
 829
 830         /*
 831          * Disallow local -> local copies since there is only space in
 832          * batch->pages for one page per-op and this would be a very
 833          * expensive memcpy().
 834          */
 835         if (!(seg->flags & (GNTCOPY_source_gref | GNTCOPY_dest_gref)))
 836                 return -EINVAL;
 837
 838         /* Can't cross page if source/dest is a grant ref. */
 839         if (seg->flags & GNTCOPY_source_gref) {
 840                 if (seg->source.foreign.offset + seg->len > XEN_PAGE_SIZE)
 841                         return -EINVAL;
 842         }
 843         if (seg->flags & GNTCOPY_dest_gref) {
 844                 if (seg->dest.foreign.offset + seg->len > XEN_PAGE_SIZE)
 845                         return -EINVAL;
 846         }
 847
 848         if (put_user(GNTST_okay, status))
 849                 return -EFAULT;
 850
 851         while (copied < seg->len) {
 852                 struct gnttab_copy *op;
 853                 void __user *virt;
 854                 size_t len, off;
 855                 unsigned long gfn;
 856                 int ret;
 857
 858                 if (batch->nr_ops >= GNTDEV_COPY_BATCH) {
 859                         ret = gntdev_copy(batch);
 860                         if (ret < 0)
 861                                 return ret;
 862                 }
 863
 864                 len = seg->len - copied;
 865
 866                 op = &batch->ops[batch->nr_ops];
 867                 op->flags = 0;
 868
 869                 if (seg->flags & GNTCOPY_source_gref) {
 870                         op->source.u.ref = seg->source.foreign.ref;
 871                         op->source.domid = seg->source.foreign.domid;
 872                         op->source.offset = seg->source.foreign.offset + copied;
 873                         op->flags |= GNTCOPY_source_gref;
 874                 } else {
 875                         virt = seg->source.virt + copied;
 876                         off = (unsigned long)virt & ~XEN_PAGE_MASK;
 877                         len = min(len, (size_t)XEN_PAGE_SIZE - off);
 878
 879                         ret = gntdev_get_page(batch, virt, false, &gfn);
 880                         if (ret < 0)
 881                                 return ret;
 882
 883                         op->source.u.gmfn = gfn;
 884                         op->source.domid = DOMID_SELF;
 885                         op->source.offset = off;
 886                 }
 887
 888                 if (seg->flags & GNTCOPY_dest_gref) {
 889                         op->dest.u.ref = seg->dest.foreign.ref;
 890                         op->dest.domid = seg->dest.foreign.domid;
 891                         op->dest.offset = seg->dest.foreign.offset + copied;
 892                         op->flags |= GNTCOPY_dest_gref;
 893                 } else {
 894                         virt = seg->dest.virt + copied;
 895                         off = (unsigned long)virt & ~XEN_PAGE_MASK;
 896                         len = min(len, (size_t)XEN_PAGE_SIZE - off);
 897
 898                         ret = gntdev_get_page(batch, virt, true, &gfn);
 899                         if (ret < 0)
 900                                 return ret;
 901
 902                         op->dest.u.gmfn = gfn;
 903                         op->dest.domid = DOMID_SELF;
 904                         op->dest.offset = off;
 905                 }
 906
 907                 op->len = len;
 908                 copied += len;
 909
 910                 batch->status[batch->nr_ops] = status;
 911                 batch->nr_ops++;
 912         }
 913
 914         return 0;
 915 }
 916
 917 static long gntdev_ioctl_grant_copy(struct gntdev_priv *priv, void __user *u)
 918 {
 919         struct ioctl_gntdev_grant_copy copy;
 920         struct gntdev_copy_batch batch;
 921         unsigned int i;
 922         int ret = 0;
 923
 924         if (copy_from_user(&copy, u, sizeof(copy)))
 925                 return -EFAULT;
 926
 927         batch.nr_ops = 0;
 928         batch.nr_pages = 0;
 929
 930         for (i = 0; i < copy.count; i++) {
 931                 struct gntdev_grant_copy_segment seg;
 932
 933                 if (copy_from_user(&seg, &copy.segments[i], sizeof(seg))) {
 934                         ret = -EFAULT;
 935                         goto out;
 936                 }
 937
 938                 ret = gntdev_grant_copy_seg(&batch, &seg, &copy.segments[i].status);
 939                 if (ret < 0)
 940                         goto out;
 941
 942                 cond_resched();
 943         }
 944         if (batch.nr_ops)
 945                 ret = gntdev_copy(&batch);
 946         return ret;
 947
 948   out:
 949         gntdev_put_pages(&batch);
 950         return ret;
 951 }
 952
 953 static long gntdev_ioctl(struct file *flip,
 954                          unsigned int cmd, unsigned long arg)
 955 {
 956         struct gntdev_priv *priv = flip->private_data;
 957         void __user *ptr = (void __user *)arg;
 958
 959         switch (cmd) {
 960         case IOCTL_GNTDEV_MAP_GRANT_REF:
 961                 return gntdev_ioctl_map_grant_ref(priv, ptr);
 962
 963         case IOCTL_GNTDEV_UNMAP_GRANT_REF:
 964                 return gntdev_ioctl_unmap_grant_ref(priv, ptr);
 965
 966         case IOCTL_GNTDEV_GET_OFFSET_FOR_VADDR:
 967                 return gntdev_ioctl_get_offset_for_vaddr(priv, ptr);
 968
 969         case IOCTL_GNTDEV_SET_UNMAP_NOTIFY:
 970                 return gntdev_ioctl_notify(priv, ptr);
 971
 972         case IOCTL_GNTDEV_GRANT_COPY:
 973                 return gntdev_ioctl_grant_copy(priv, ptr);
 974
 975         default:
 976                 pr_debug("priv %p, unknown cmd %x\n", priv, cmd);
 977                 return -ENOIOCTLCMD;
 978         }
 979
 980         return 0;
 981 }
 982
 983 static 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 grant_map *map;
 989         int i, 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         if (use_ptemod && priv->mm != vma->vm_mm) {
1004                 pr_warn("Huh? Other mm?\n");
1005                 goto unlock_out;
1006         }
1007
1008         refcount_inc(&map->users);
1009
1010         vma->vm_ops = &gntdev_vmops;
1011
1012         vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP | VM_MIXEDMAP;
1013
1014         if (use_ptemod)
1015                 vma->vm_flags |= VM_DONTCOPY;
1016
1017         vma->vm_private_data = map;
1018
1019         if (use_ptemod)
1020                 map->vma = vma;
1021
1022         if (map->flags) {
1023                 if ((vma->vm_flags & VM_WRITE) &&
1024                                 (map->flags & GNTMAP_readonly))
1025                         goto out_unlock_put;
1026         } else {
1027                 map->flags = GNTMAP_host_map;
1028                 if (!(vma->vm_flags & VM_WRITE))
1029                         map->flags |= GNTMAP_readonly;
1030         }
1031
1032         mutex_unlock(&priv->lock);
1033
1034         if (use_ptemod) {
1035                 err = apply_to_page_range(vma->vm_mm, vma->vm_start,
1036                                           vma->vm_end - vma->vm_start,
1037                                           find_grant_ptes, map);
1038                 if (err) {
1039                         pr_warn("find_grant_ptes() failure.\n");
1040                         goto out_put_map;
1041                 }
1042         }
1043
1044         err = map_grant_pages(map);
1045         if (err)
1046                 goto out_put_map;
1047
1048         if (!use_ptemod) {
1049                 for (i = 0; i < count; i++) {
1050                         err = vm_insert_page(vma, vma->vm_start + i*PAGE_SIZE,
1051                                 map->pages[i]);
1052                         if (err)
1053                                 goto out_put_map;
1054                 }
1055         } else {
1056 #ifdef CONFIG_X86
1057                 /*
1058                  * If the PTEs were not made special by the grant map
1059                  * hypercall, do so here.
1060                  *
1061                  * This is racy since the mapping is already visible
1062                  * to userspace but userspace should be well-behaved
1063                  * enough to not touch it until the mmap() call
1064                  * returns.
1065                  */
1066                 if (!xen_feature(XENFEAT_gnttab_map_avail_bits)) {
1067                         apply_to_page_range(vma->vm_mm, vma->vm_start,
1068                                             vma->vm_end - vma->vm_start,
1069                                             set_grant_ptes_as_special, NULL);
1070                 }
1071 #endif
1072                 map->pages_vm_start = vma->vm_start;
1073         }
1074
1075         return 0;
1076
1077 unlock_out:
1078         mutex_unlock(&priv->lock);
1079         return err;
1080
1081 out_unlock_put:
1082         mutex_unlock(&priv->lock);
1083 out_put_map:
1084         if (use_ptemod)
1085                 map->vma = NULL;
1086         gntdev_put_map(priv, map);
1087         return err;
1088 }
1089
1090 static const struct file_operations gntdev_fops = {
1091         .owner = THIS_MODULE,
1092         .open = gntdev_open,
1093         .release = gntdev_release,
1094         .mmap = gntdev_mmap,
1095         .unlocked_ioctl = gntdev_ioctl
1096 };
1097
1098 static struct miscdevice gntdev_miscdev = {
1099         .minor        = MISC_DYNAMIC_MINOR,
1100         .name         = "xen/gntdev",
1101         .fops         = &gntdev_fops,
1102 };
1103
1104 /* ------------------------------------------------------------------ */
1105
1106 static int __init gntdev_init(void)
1107 {
1108         int err;
1109
1110         if (!xen_domain())
1111                 return -ENODEV;
1112
1113         use_ptemod = !xen_feature(XENFEAT_auto_translated_physmap);
1114
1115         err = misc_register(&gntdev_miscdev);
1116         if (err != 0) {
1117                 pr_err("Could not register gntdev device\n");
1118                 return err;
1119         }
1120         return 0;
1121 }
1122
1123 static void __exit gntdev_exit(void)
1124 {
1125         misc_deregister(&gntdev_miscdev);
1126 }
1127
1128 module_init(gntdev_init);
1129 module_exit(gntdev_exit);
1130
1131 /* ------------------------------------------------------------------ */