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/module.h> 26 #include <linux/kernel.h> 27 #include <linux/init.h> 28 #include <linux/miscdevice.h> 29 #include <linux/fs.h> 30 #include <linux/uaccess.h> 31 #include <linux/sched.h> 32 #include <linux/sched/mm.h> 33 #include <linux/spinlock.h> 34 #include <linux/slab.h> 35 #include <linux/highmem.h> 36 #include <linux/refcount.h> 37 #ifdef CONFIG_XEN_GRANT_DMA_ALLOC 38 #include <linux/of_device.h> 39 #endif 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 #include "gntdev-common.h" 51 #ifdef CONFIG_XEN_GNTDEV_DMABUF 52 #include "gntdev-dmabuf.h" 53 #endif 54 55 MODULE_LICENSE("GPL"); 56 MODULE_AUTHOR("Derek G. Murray <Derek.Murray@cl.cam.ac.uk>, " 57 "Gerd Hoffmann <kraxel@redhat.com>"); 58 MODULE_DESCRIPTION("User-space granted page access driver"); 59 60 static int limit = 1024*1024; 61 module_param(limit, int, 0644); 62 MODULE_PARM_DESC(limit, "Maximum number of grants that may be mapped by " 63 "the gntdev device"); 64 65 static atomic_t pages_mapped = ATOMIC_INIT(0); 66 67 static int use_ptemod; 68 #define populate_freeable_maps use_ptemod 69 70 static int unmap_grant_pages(struct gntdev_grant_map *map, 71 int offset, int pages); 72 73 static struct miscdevice gntdev_miscdev; 74 75 /* ------------------------------------------------------------------ */ 76 77 bool gntdev_account_mapped_pages(int count) 78 { 79 return atomic_add_return(count, &pages_mapped) > limit; 80 } 81 82 static void gntdev_print_maps(struct gntdev_priv *priv, 83 char *text, int text_index) 84 { 85 #ifdef DEBUG 86 struct gntdev_grant_map *map; 87 88 pr_debug("%s: maps list (priv %p)\n", __func__, priv); 89 list_for_each_entry(map, &priv->maps, next) 90 pr_debug(" index %2d, count %2d %s\n", 91 map->index, map->count, 92 map->index == text_index && text ? text : ""); 93 #endif 94 } 95 96 static void gntdev_free_map(struct gntdev_grant_map *map) 97 { 98 if (map == NULL) 99 return; 100 101 #ifdef CONFIG_XEN_GRANT_DMA_ALLOC 102 if (map->dma_vaddr) { 103 struct gnttab_dma_alloc_args args; 104 105 args.dev = map->dma_dev; 106 args.coherent = !!(map->dma_flags & GNTDEV_DMA_FLAG_COHERENT); 107 args.nr_pages = map->count; 108 args.pages = map->pages; 109 args.frames = map->frames; 110 args.vaddr = map->dma_vaddr; 111 args.dev_bus_addr = map->dma_bus_addr; 112 113 gnttab_dma_free_pages(&args); 114 } else 115 #endif 116 if (map->pages) 117 gnttab_free_pages(map->count, map->pages); 118 119 #ifdef CONFIG_XEN_GRANT_DMA_ALLOC 120 kfree(map->frames); 121 #endif 122 kfree(map->pages); 123 kfree(map->grants); 124 kfree(map->map_ops); 125 kfree(map->unmap_ops); 126 kfree(map->kmap_ops); 127 kfree(map->kunmap_ops); 128 kfree(map); 129 } 130 131 struct gntdev_grant_map *gntdev_alloc_map(struct gntdev_priv *priv, int count, 132 int dma_flags) 133 { 134 struct gntdev_grant_map *add; 135 int i; 136 137 add = kzalloc(sizeof(*add), GFP_KERNEL); 138 if (NULL == add) 139 return NULL; 140 141 add->grants = kcalloc(count, sizeof(add->grants[0]), GFP_KERNEL); 142 add->map_ops = kcalloc(count, sizeof(add->map_ops[0]), GFP_KERNEL); 143 add->unmap_ops = kcalloc(count, sizeof(add->unmap_ops[0]), GFP_KERNEL); 144 add->kmap_ops = kcalloc(count, sizeof(add->kmap_ops[0]), GFP_KERNEL); 145 add->kunmap_ops = kcalloc(count, sizeof(add->kunmap_ops[0]), GFP_KERNEL); 146 add->pages = kcalloc(count, sizeof(add->pages[0]), GFP_KERNEL); 147 if (NULL == add->grants || 148 NULL == add->map_ops || 149 NULL == add->unmap_ops || 150 NULL == add->kmap_ops || 151 NULL == add->kunmap_ops || 152 NULL == add->pages) 153 goto err; 154 155 #ifdef CONFIG_XEN_GRANT_DMA_ALLOC 156 add->dma_flags = dma_flags; 157 158 /* 159 * Check if this mapping is requested to be backed 160 * by a DMA buffer. 161 */ 162 if (dma_flags & (GNTDEV_DMA_FLAG_WC | GNTDEV_DMA_FLAG_COHERENT)) { 163 struct gnttab_dma_alloc_args args; 164 165 add->frames = kcalloc(count, sizeof(add->frames[0]), 166 GFP_KERNEL); 167 if (!add->frames) 168 goto err; 169 170 /* Remember the device, so we can free DMA memory. */ 171 add->dma_dev = priv->dma_dev; 172 173 args.dev = priv->dma_dev; 174 args.coherent = !!(dma_flags & GNTDEV_DMA_FLAG_COHERENT); 175 args.nr_pages = count; 176 args.pages = add->pages; 177 args.frames = add->frames; 178 179 if (gnttab_dma_alloc_pages(&args)) 180 goto err; 181 182 add->dma_vaddr = args.vaddr; 183 add->dma_bus_addr = args.dev_bus_addr; 184 } else 185 #endif 186 if (gnttab_alloc_pages(count, add->pages)) 187 goto err; 188 189 for (i = 0; i < count; i++) { 190 add->map_ops[i].handle = -1; 191 add->unmap_ops[i].handle = -1; 192 add->kmap_ops[i].handle = -1; 193 add->kunmap_ops[i].handle = -1; 194 } 195 196 add->index = 0; 197 add->count = count; 198 refcount_set(&add->users, 1); 199 200 return add; 201 202 err: 203 gntdev_free_map(add); 204 return NULL; 205 } 206 207 void gntdev_add_map(struct gntdev_priv *priv, struct gntdev_grant_map *add) 208 { 209 struct gntdev_grant_map *map; 210 211 list_for_each_entry(map, &priv->maps, next) { 212 if (add->index + add->count < map->index) { 213 list_add_tail(&add->next, &map->next); 214 goto done; 215 } 216 add->index = map->index + map->count; 217 } 218 list_add_tail(&add->next, &priv->maps); 219 220 done: 221 gntdev_print_maps(priv, "[new]", add->index); 222 } 223 224 static struct gntdev_grant_map *gntdev_find_map_index(struct gntdev_priv *priv, 225 int index, int count) 226 { 227 struct gntdev_grant_map *map; 228 229 list_for_each_entry(map, &priv->maps, next) { 230 if (map->index != index) 231 continue; 232 if (count && map->count != count) 233 continue; 234 return map; 235 } 236 return NULL; 237 } 238 239 void gntdev_put_map(struct gntdev_priv *priv, struct gntdev_grant_map *map) 240 { 241 if (!map) 242 return; 243 244 if (!refcount_dec_and_test(&map->users)) 245 return; 246 247 atomic_sub(map->count, &pages_mapped); 248 249 if (map->notify.flags & UNMAP_NOTIFY_SEND_EVENT) { 250 notify_remote_via_evtchn(map->notify.event); 251 evtchn_put(map->notify.event); 252 } 253 254 if (populate_freeable_maps && priv) { 255 mutex_lock(&priv->lock); 256 list_del(&map->next); 257 mutex_unlock(&priv->lock); 258 } 259 260 if (map->pages && !use_ptemod) 261 unmap_grant_pages(map, 0, map->count); 262 gntdev_free_map(map); 263 } 264 265 /* ------------------------------------------------------------------ */ 266 267 static int find_grant_ptes(pte_t *pte, pgtable_t token, 268 unsigned long addr, void *data) 269 { 270 struct gntdev_grant_map *map = data; 271 unsigned int pgnr = (addr - map->vma->vm_start) >> PAGE_SHIFT; 272 int flags = map->flags | GNTMAP_application_map | GNTMAP_contains_pte; 273 u64 pte_maddr; 274 275 BUG_ON(pgnr >= map->count); 276 pte_maddr = arbitrary_virt_to_machine(pte).maddr; 277 278 /* 279 * Set the PTE as special to force get_user_pages_fast() fall 280 * back to the slow path. If this is not supported as part of 281 * the grant map, it will be done afterwards. 282 */ 283 if (xen_feature(XENFEAT_gnttab_map_avail_bits)) 284 flags |= (1 << _GNTMAP_guest_avail0); 285 286 gnttab_set_map_op(&map->map_ops[pgnr], pte_maddr, flags, 287 map->grants[pgnr].ref, 288 map->grants[pgnr].domid); 289 gnttab_set_unmap_op(&map->unmap_ops[pgnr], pte_maddr, flags, 290 -1 /* handle */); 291 return 0; 292 } 293 294 #ifdef CONFIG_X86 295 static int set_grant_ptes_as_special(pte_t *pte, pgtable_t token, 296 unsigned long addr, void *data) 297 { 298 set_pte_at(current->mm, addr, pte, pte_mkspecial(*pte)); 299 return 0; 300 } 301 #endif 302 303 int gntdev_map_grant_pages(struct gntdev_grant_map *map) 304 { 305 int i, err = 0; 306 307 if (!use_ptemod) { 308 /* Note: it could already be mapped */ 309 if (map->map_ops[0].handle != -1) 310 return 0; 311 for (i = 0; i < map->count; i++) { 312 unsigned long addr = (unsigned long) 313 pfn_to_kaddr(page_to_pfn(map->pages[i])); 314 gnttab_set_map_op(&map->map_ops[i], addr, map->flags, 315 map->grants[i].ref, 316 map->grants[i].domid); 317 gnttab_set_unmap_op(&map->unmap_ops[i], addr, 318 map->flags, -1 /* handle */); 319 } 320 } else { 321 /* 322 * Setup the map_ops corresponding to the pte entries pointing 323 * to the kernel linear addresses of the struct pages. 324 * These ptes are completely different from the user ptes dealt 325 * with find_grant_ptes. 326 */ 327 for (i = 0; i < map->count; i++) { 328 unsigned long address = (unsigned long) 329 pfn_to_kaddr(page_to_pfn(map->pages[i])); 330 BUG_ON(PageHighMem(map->pages[i])); 331 332 gnttab_set_map_op(&map->kmap_ops[i], address, 333 map->flags | GNTMAP_host_map, 334 map->grants[i].ref, 335 map->grants[i].domid); 336 gnttab_set_unmap_op(&map->kunmap_ops[i], address, 337 map->flags | GNTMAP_host_map, -1); 338 } 339 } 340 341 pr_debug("map %d+%d\n", map->index, map->count); 342 err = gnttab_map_refs(map->map_ops, use_ptemod ? map->kmap_ops : NULL, 343 map->pages, map->count); 344 if (err) 345 return err; 346 347 for (i = 0; i < map->count; i++) { 348 if (map->map_ops[i].status) { 349 err = -EINVAL; 350 continue; 351 } 352 353 map->unmap_ops[i].handle = map->map_ops[i].handle; 354 if (use_ptemod) 355 map->kunmap_ops[i].handle = map->kmap_ops[i].handle; 356 #ifdef CONFIG_XEN_GRANT_DMA_ALLOC 357 else if (map->dma_vaddr) { 358 unsigned long bfn; 359 360 bfn = pfn_to_bfn(page_to_pfn(map->pages[i])); 361 map->unmap_ops[i].dev_bus_addr = __pfn_to_phys(bfn); 362 } 363 #endif 364 } 365 return err; 366 } 367 368 static int __unmap_grant_pages(struct gntdev_grant_map *map, int offset, 369 int pages) 370 { 371 int i, err = 0; 372 struct gntab_unmap_queue_data unmap_data; 373 374 if (map->notify.flags & UNMAP_NOTIFY_CLEAR_BYTE) { 375 int pgno = (map->notify.addr >> PAGE_SHIFT); 376 if (pgno >= offset && pgno < offset + pages) { 377 /* No need for kmap, pages are in lowmem */ 378 uint8_t *tmp = pfn_to_kaddr(page_to_pfn(map->pages[pgno])); 379 tmp[map->notify.addr & (PAGE_SIZE-1)] = 0; 380 map->notify.flags &= ~UNMAP_NOTIFY_CLEAR_BYTE; 381 } 382 } 383 384 unmap_data.unmap_ops = map->unmap_ops + offset; 385 unmap_data.kunmap_ops = use_ptemod ? map->kunmap_ops + offset : NULL; 386 unmap_data.pages = map->pages + offset; 387 unmap_data.count = pages; 388 389 err = gnttab_unmap_refs_sync(&unmap_data); 390 if (err) 391 return err; 392 393 for (i = 0; i < pages; i++) { 394 if (map->unmap_ops[offset+i].status) 395 err = -EINVAL; 396 pr_debug("unmap handle=%d st=%d\n", 397 map->unmap_ops[offset+i].handle, 398 map->unmap_ops[offset+i].status); 399 map->unmap_ops[offset+i].handle = -1; 400 } 401 return err; 402 } 403 404 static int unmap_grant_pages(struct gntdev_grant_map *map, int offset, 405 int pages) 406 { 407 int range, err = 0; 408 409 pr_debug("unmap %d+%d [%d+%d]\n", map->index, map->count, offset, pages); 410 411 /* It is possible the requested range will have a "hole" where we 412 * already unmapped some of the grants. Only unmap valid ranges. 413 */ 414 while (pages && !err) { 415 while (pages && map->unmap_ops[offset].handle == -1) { 416 offset++; 417 pages--; 418 } 419 range = 0; 420 while (range < pages) { 421 if (map->unmap_ops[offset+range].handle == -1) 422 break; 423 range++; 424 } 425 err = __unmap_grant_pages(map, offset, range); 426 offset += range; 427 pages -= range; 428 } 429 430 return err; 431 } 432 433 /* ------------------------------------------------------------------ */ 434 435 static void gntdev_vma_open(struct vm_area_struct *vma) 436 { 437 struct gntdev_grant_map *map = vma->vm_private_data; 438 439 pr_debug("gntdev_vma_open %p\n", vma); 440 refcount_inc(&map->users); 441 } 442 443 static void gntdev_vma_close(struct vm_area_struct *vma) 444 { 445 struct gntdev_grant_map *map = vma->vm_private_data; 446 struct file *file = vma->vm_file; 447 struct gntdev_priv *priv = file->private_data; 448 449 pr_debug("gntdev_vma_close %p\n", vma); 450 if (use_ptemod) { 451 /* It is possible that an mmu notifier could be running 452 * concurrently, so take priv->lock to ensure that the vma won't 453 * vanishing during the unmap_grant_pages call, since we will 454 * spin here until that completes. Such a concurrent call will 455 * not do any unmapping, since that has been done prior to 456 * closing the vma, but it may still iterate the unmap_ops list. 457 */ 458 mutex_lock(&priv->lock); 459 map->vma = NULL; 460 mutex_unlock(&priv->lock); 461 } 462 vma->vm_private_data = NULL; 463 gntdev_put_map(priv, map); 464 } 465 466 static struct page *gntdev_vma_find_special_page(struct vm_area_struct *vma, 467 unsigned long addr) 468 { 469 struct gntdev_grant_map *map = vma->vm_private_data; 470 471 return map->pages[(addr - map->pages_vm_start) >> PAGE_SHIFT]; 472 } 473 474 static const struct vm_operations_struct gntdev_vmops = { 475 .open = gntdev_vma_open, 476 .close = gntdev_vma_close, 477 .find_special_page = gntdev_vma_find_special_page, 478 }; 479 480 /* ------------------------------------------------------------------ */ 481 482 static bool in_range(struct gntdev_grant_map *map, 483 unsigned long start, unsigned long end) 484 { 485 if (!map->vma) 486 return false; 487 if (map->vma->vm_start >= end) 488 return false; 489 if (map->vma->vm_end <= start) 490 return false; 491 492 return true; 493 } 494 495 static void unmap_if_in_range(struct gntdev_grant_map *map, 496 unsigned long start, unsigned long end) 497 { 498 unsigned long mstart, mend; 499 int err; 500 501 mstart = max(start, map->vma->vm_start); 502 mend = min(end, map->vma->vm_end); 503 pr_debug("map %d+%d (%lx %lx), range %lx %lx, mrange %lx %lx\n", 504 map->index, map->count, 505 map->vma->vm_start, map->vma->vm_end, 506 start, end, mstart, mend); 507 err = unmap_grant_pages(map, 508 (mstart - map->vma->vm_start) >> PAGE_SHIFT, 509 (mend - mstart) >> PAGE_SHIFT); 510 WARN_ON(err); 511 } 512 513 static int mn_invl_range_start(struct mmu_notifier *mn, 514 struct mm_struct *mm, 515 unsigned long start, unsigned long end, 516 bool blockable) 517 { 518 struct gntdev_priv *priv = container_of(mn, struct gntdev_priv, mn); 519 struct gntdev_grant_map *map; 520 int ret = 0; 521 522 /* TODO do we really need a mutex here? */ 523 if (blockable) 524 mutex_lock(&priv->lock); 525 else if (!mutex_trylock(&priv->lock)) 526 return -EAGAIN; 527 528 list_for_each_entry(map, &priv->maps, next) { 529 if (in_range(map, start, end)) { 530 ret = -EAGAIN; 531 goto out_unlock; 532 } 533 unmap_if_in_range(map, start, end); 534 } 535 list_for_each_entry(map, &priv->freeable_maps, next) { 536 if (in_range(map, start, end)) { 537 ret = -EAGAIN; 538 goto out_unlock; 539 } 540 unmap_if_in_range(map, start, end); 541 } 542 543 out_unlock: 544 mutex_unlock(&priv->lock); 545 546 return ret; 547 } 548 549 static void mn_release(struct mmu_notifier *mn, 550 struct mm_struct *mm) 551 { 552 struct gntdev_priv *priv = container_of(mn, struct gntdev_priv, mn); 553 struct gntdev_grant_map *map; 554 int err; 555 556 mutex_lock(&priv->lock); 557 list_for_each_entry(map, &priv->maps, next) { 558 if (!map->vma) 559 continue; 560 pr_debug("map %d+%d (%lx %lx)\n", 561 map->index, map->count, 562 map->vma->vm_start, map->vma->vm_end); 563 err = unmap_grant_pages(map, /* offset */ 0, map->count); 564 WARN_ON(err); 565 } 566 list_for_each_entry(map, &priv->freeable_maps, next) { 567 if (!map->vma) 568 continue; 569 pr_debug("map %d+%d (%lx %lx)\n", 570 map->index, map->count, 571 map->vma->vm_start, map->vma->vm_end); 572 err = unmap_grant_pages(map, /* offset */ 0, map->count); 573 WARN_ON(err); 574 } 575 mutex_unlock(&priv->lock); 576 } 577 578 static const struct mmu_notifier_ops gntdev_mmu_ops = { 579 .release = mn_release, 580 .invalidate_range_start = mn_invl_range_start, 581 }; 582 583 /* ------------------------------------------------------------------ */ 584 585 static int gntdev_open(struct inode *inode, struct file *flip) 586 { 587 struct gntdev_priv *priv; 588 int ret = 0; 589 590 priv = kzalloc(sizeof(*priv), GFP_KERNEL); 591 if (!priv) 592 return -ENOMEM; 593 594 INIT_LIST_HEAD(&priv->maps); 595 INIT_LIST_HEAD(&priv->freeable_maps); 596 mutex_init(&priv->lock); 597 598 #ifdef CONFIG_XEN_GNTDEV_DMABUF 599 priv->dmabuf_priv = gntdev_dmabuf_init(); 600 if (IS_ERR(priv->dmabuf_priv)) { 601 ret = PTR_ERR(priv->dmabuf_priv); 602 kfree(priv); 603 return ret; 604 } 605 #endif 606 607 if (use_ptemod) { 608 priv->mm = get_task_mm(current); 609 if (!priv->mm) { 610 kfree(priv); 611 return -ENOMEM; 612 } 613 priv->mn.ops = &gntdev_mmu_ops; 614 ret = mmu_notifier_register(&priv->mn, priv->mm); 615 mmput(priv->mm); 616 } 617 618 if (ret) { 619 kfree(priv); 620 return ret; 621 } 622 623 flip->private_data = priv; 624 #ifdef CONFIG_XEN_GRANT_DMA_ALLOC 625 priv->dma_dev = gntdev_miscdev.this_device; 626 627 /* 628 * The device is not spawn from a device tree, so arch_setup_dma_ops 629 * is not called, thus leaving the device with dummy DMA ops. 630 * Fix this by calling of_dma_configure() with a NULL node to set 631 * default DMA ops. 632 */ 633 of_dma_configure(priv->dma_dev, NULL, true); 634 #endif 635 pr_debug("priv %p\n", priv); 636 637 return 0; 638 } 639 640 static int gntdev_release(struct inode *inode, struct file *flip) 641 { 642 struct gntdev_priv *priv = flip->private_data; 643 struct gntdev_grant_map *map; 644 645 pr_debug("priv %p\n", priv); 646 647 mutex_lock(&priv->lock); 648 while (!list_empty(&priv->maps)) { 649 map = list_entry(priv->maps.next, 650 struct gntdev_grant_map, next); 651 list_del(&map->next); 652 gntdev_put_map(NULL /* already removed */, map); 653 } 654 WARN_ON(!list_empty(&priv->freeable_maps)); 655 mutex_unlock(&priv->lock); 656 657 #ifdef CONFIG_XEN_GNTDEV_DMABUF 658 gntdev_dmabuf_fini(priv->dmabuf_priv); 659 #endif 660 661 if (use_ptemod) 662 mmu_notifier_unregister(&priv->mn, priv->mm); 663 664 kfree(priv); 665 return 0; 666 } 667 668 static long gntdev_ioctl_map_grant_ref(struct gntdev_priv *priv, 669 struct ioctl_gntdev_map_grant_ref __user *u) 670 { 671 struct ioctl_gntdev_map_grant_ref op; 672 struct gntdev_grant_map *map; 673 int err; 674 675 if (copy_from_user(&op, u, sizeof(op)) != 0) 676 return -EFAULT; 677 pr_debug("priv %p, add %d\n", priv, op.count); 678 if (unlikely(op.count <= 0)) 679 return -EINVAL; 680 681 err = -ENOMEM; 682 map = gntdev_alloc_map(priv, op.count, 0 /* This is not a dma-buf. */); 683 if (!map) 684 return err; 685 686 if (unlikely(gntdev_account_mapped_pages(op.count))) { 687 pr_debug("can't map: over limit\n"); 688 gntdev_put_map(NULL, map); 689 return err; 690 } 691 692 if (copy_from_user(map->grants, &u->refs, 693 sizeof(map->grants[0]) * op.count) != 0) { 694 gntdev_put_map(NULL, map); 695 return -EFAULT; 696 } 697 698 mutex_lock(&priv->lock); 699 gntdev_add_map(priv, map); 700 op.index = map->index << PAGE_SHIFT; 701 mutex_unlock(&priv->lock); 702 703 if (copy_to_user(u, &op, sizeof(op)) != 0) 704 return -EFAULT; 705 706 return 0; 707 } 708 709 static long gntdev_ioctl_unmap_grant_ref(struct gntdev_priv *priv, 710 struct ioctl_gntdev_unmap_grant_ref __user *u) 711 { 712 struct ioctl_gntdev_unmap_grant_ref op; 713 struct gntdev_grant_map *map; 714 int err = -ENOENT; 715 716 if (copy_from_user(&op, u, sizeof(op)) != 0) 717 return -EFAULT; 718 pr_debug("priv %p, del %d+%d\n", priv, (int)op.index, (int)op.count); 719 720 mutex_lock(&priv->lock); 721 map = gntdev_find_map_index(priv, op.index >> PAGE_SHIFT, op.count); 722 if (map) { 723 list_del(&map->next); 724 if (populate_freeable_maps) 725 list_add_tail(&map->next, &priv->freeable_maps); 726 err = 0; 727 } 728 mutex_unlock(&priv->lock); 729 if (map) 730 gntdev_put_map(priv, map); 731 return err; 732 } 733 734 static long gntdev_ioctl_get_offset_for_vaddr(struct gntdev_priv *priv, 735 struct ioctl_gntdev_get_offset_for_vaddr __user *u) 736 { 737 struct ioctl_gntdev_get_offset_for_vaddr op; 738 struct vm_area_struct *vma; 739 struct gntdev_grant_map *map; 740 int rv = -EINVAL; 741 742 if (copy_from_user(&op, u, sizeof(op)) != 0) 743 return -EFAULT; 744 pr_debug("priv %p, offset for vaddr %lx\n", priv, (unsigned long)op.vaddr); 745 746 down_read(¤t->mm->mmap_sem); 747 vma = find_vma(current->mm, op.vaddr); 748 if (!vma || vma->vm_ops != &gntdev_vmops) 749 goto out_unlock; 750 751 map = vma->vm_private_data; 752 if (!map) 753 goto out_unlock; 754 755 op.offset = map->index << PAGE_SHIFT; 756 op.count = map->count; 757 rv = 0; 758 759 out_unlock: 760 up_read(¤t->mm->mmap_sem); 761 762 if (rv == 0 && copy_to_user(u, &op, sizeof(op)) != 0) 763 return -EFAULT; 764 return rv; 765 } 766 767 static long gntdev_ioctl_notify(struct gntdev_priv *priv, void __user *u) 768 { 769 struct ioctl_gntdev_unmap_notify op; 770 struct gntdev_grant_map *map; 771 int rc; 772 int out_flags; 773 unsigned int out_event; 774 775 if (copy_from_user(&op, u, sizeof(op))) 776 return -EFAULT; 777 778 if (op.action & ~(UNMAP_NOTIFY_CLEAR_BYTE|UNMAP_NOTIFY_SEND_EVENT)) 779 return -EINVAL; 780 781 /* We need to grab a reference to the event channel we are going to use 782 * to send the notify before releasing the reference we may already have 783 * (if someone has called this ioctl twice). This is required so that 784 * it is possible to change the clear_byte part of the notification 785 * without disturbing the event channel part, which may now be the last 786 * reference to that event channel. 787 */ 788 if (op.action & UNMAP_NOTIFY_SEND_EVENT) { 789 if (evtchn_get(op.event_channel_port)) 790 return -EINVAL; 791 } 792 793 out_flags = op.action; 794 out_event = op.event_channel_port; 795 796 mutex_lock(&priv->lock); 797 798 list_for_each_entry(map, &priv->maps, next) { 799 uint64_t begin = map->index << PAGE_SHIFT; 800 uint64_t end = (map->index + map->count) << PAGE_SHIFT; 801 if (op.index >= begin && op.index < end) 802 goto found; 803 } 804 rc = -ENOENT; 805 goto unlock_out; 806 807 found: 808 if ((op.action & UNMAP_NOTIFY_CLEAR_BYTE) && 809 (map->flags & GNTMAP_readonly)) { 810 rc = -EINVAL; 811 goto unlock_out; 812 } 813 814 out_flags = map->notify.flags; 815 out_event = map->notify.event; 816 817 map->notify.flags = op.action; 818 map->notify.addr = op.index - (map->index << PAGE_SHIFT); 819 map->notify.event = op.event_channel_port; 820 821 rc = 0; 822 823 unlock_out: 824 mutex_unlock(&priv->lock); 825 826 /* Drop the reference to the event channel we did not save in the map */ 827 if (out_flags & UNMAP_NOTIFY_SEND_EVENT) 828 evtchn_put(out_event); 829 830 return rc; 831 } 832 833 #define GNTDEV_COPY_BATCH 16 834 835 struct gntdev_copy_batch { 836 struct gnttab_copy ops[GNTDEV_COPY_BATCH]; 837 struct page *pages[GNTDEV_COPY_BATCH]; 838 s16 __user *status[GNTDEV_COPY_BATCH]; 839 unsigned int nr_ops; 840 unsigned int nr_pages; 841 }; 842 843 static int gntdev_get_page(struct gntdev_copy_batch *batch, void __user *virt, 844 bool writeable, unsigned long *gfn) 845 { 846 unsigned long addr = (unsigned long)virt; 847 struct page *page; 848 unsigned long xen_pfn; 849 int ret; 850 851 ret = get_user_pages_fast(addr, 1, writeable, &page); 852 if (ret < 0) 853 return ret; 854 855 batch->pages[batch->nr_pages++] = page; 856 857 xen_pfn = page_to_xen_pfn(page) + XEN_PFN_DOWN(addr & ~PAGE_MASK); 858 *gfn = pfn_to_gfn(xen_pfn); 859 860 return 0; 861 } 862 863 static void gntdev_put_pages(struct gntdev_copy_batch *batch) 864 { 865 unsigned int i; 866 867 for (i = 0; i < batch->nr_pages; i++) 868 put_page(batch->pages[i]); 869 batch->nr_pages = 0; 870 } 871 872 static int gntdev_copy(struct gntdev_copy_batch *batch) 873 { 874 unsigned int i; 875 876 gnttab_batch_copy(batch->ops, batch->nr_ops); 877 gntdev_put_pages(batch); 878 879 /* 880 * For each completed op, update the status if the op failed 881 * and all previous ops for the segment were successful. 882 */ 883 for (i = 0; i < batch->nr_ops; i++) { 884 s16 status = batch->ops[i].status; 885 s16 old_status; 886 887 if (status == GNTST_okay) 888 continue; 889 890 if (__get_user(old_status, batch->status[i])) 891 return -EFAULT; 892 893 if (old_status != GNTST_okay) 894 continue; 895 896 if (__put_user(status, batch->status[i])) 897 return -EFAULT; 898 } 899 900 batch->nr_ops = 0; 901 return 0; 902 } 903 904 static int gntdev_grant_copy_seg(struct gntdev_copy_batch *batch, 905 struct gntdev_grant_copy_segment *seg, 906 s16 __user *status) 907 { 908 uint16_t copied = 0; 909 910 /* 911 * Disallow local -> local copies since there is only space in 912 * batch->pages for one page per-op and this would be a very 913 * expensive memcpy(). 914 */ 915 if (!(seg->flags & (GNTCOPY_source_gref | GNTCOPY_dest_gref))) 916 return -EINVAL; 917 918 /* Can't cross page if source/dest is a grant ref. */ 919 if (seg->flags & GNTCOPY_source_gref) { 920 if (seg->source.foreign.offset + seg->len > XEN_PAGE_SIZE) 921 return -EINVAL; 922 } 923 if (seg->flags & GNTCOPY_dest_gref) { 924 if (seg->dest.foreign.offset + seg->len > XEN_PAGE_SIZE) 925 return -EINVAL; 926 } 927 928 if (put_user(GNTST_okay, status)) 929 return -EFAULT; 930 931 while (copied < seg->len) { 932 struct gnttab_copy *op; 933 void __user *virt; 934 size_t len, off; 935 unsigned long gfn; 936 int ret; 937 938 if (batch->nr_ops >= GNTDEV_COPY_BATCH) { 939 ret = gntdev_copy(batch); 940 if (ret < 0) 941 return ret; 942 } 943 944 len = seg->len - copied; 945 946 op = &batch->ops[batch->nr_ops]; 947 op->flags = 0; 948 949 if (seg->flags & GNTCOPY_source_gref) { 950 op->source.u.ref = seg->source.foreign.ref; 951 op->source.domid = seg->source.foreign.domid; 952 op->source.offset = seg->source.foreign.offset + copied; 953 op->flags |= GNTCOPY_source_gref; 954 } else { 955 virt = seg->source.virt + copied; 956 off = (unsigned long)virt & ~XEN_PAGE_MASK; 957 len = min(len, (size_t)XEN_PAGE_SIZE - off); 958 959 ret = gntdev_get_page(batch, virt, false, &gfn); 960 if (ret < 0) 961 return ret; 962 963 op->source.u.gmfn = gfn; 964 op->source.domid = DOMID_SELF; 965 op->source.offset = off; 966 } 967 968 if (seg->flags & GNTCOPY_dest_gref) { 969 op->dest.u.ref = seg->dest.foreign.ref; 970 op->dest.domid = seg->dest.foreign.domid; 971 op->dest.offset = seg->dest.foreign.offset + copied; 972 op->flags |= GNTCOPY_dest_gref; 973 } else { 974 virt = seg->dest.virt + copied; 975 off = (unsigned long)virt & ~XEN_PAGE_MASK; 976 len = min(len, (size_t)XEN_PAGE_SIZE - off); 977 978 ret = gntdev_get_page(batch, virt, true, &gfn); 979 if (ret < 0) 980 return ret; 981 982 op->dest.u.gmfn = gfn; 983 op->dest.domid = DOMID_SELF; 984 op->dest.offset = off; 985 } 986 987 op->len = len; 988 copied += len; 989 990 batch->status[batch->nr_ops] = status; 991 batch->nr_ops++; 992 } 993 994 return 0; 995 } 996 997 static long gntdev_ioctl_grant_copy(struct gntdev_priv *priv, void __user *u) 998 { 999 struct ioctl_gntdev_grant_copy copy; 1000 struct gntdev_copy_batch batch; 1001 unsigned int i; 1002 int ret = 0; 1003 1004 if (copy_from_user(©, u, sizeof(copy))) 1005 return -EFAULT; 1006 1007 batch.nr_ops = 0; 1008 batch.nr_pages = 0; 1009 1010 for (i = 0; i < copy.count; i++) { 1011 struct gntdev_grant_copy_segment seg; 1012 1013 if (copy_from_user(&seg, ©.segments[i], sizeof(seg))) { 1014 ret = -EFAULT; 1015 goto out; 1016 } 1017 1018 ret = gntdev_grant_copy_seg(&batch, &seg, ©.segments[i].status); 1019 if (ret < 0) 1020 goto out; 1021 1022 cond_resched(); 1023 } 1024 if (batch.nr_ops) 1025 ret = gntdev_copy(&batch); 1026 return ret; 1027 1028 out: 1029 gntdev_put_pages(&batch); 1030 return ret; 1031 } 1032 1033 static long gntdev_ioctl(struct file *flip, 1034 unsigned int cmd, unsigned long arg) 1035 { 1036 struct gntdev_priv *priv = flip->private_data; 1037 void __user *ptr = (void __user *)arg; 1038 1039 switch (cmd) { 1040 case IOCTL_GNTDEV_MAP_GRANT_REF: 1041 return gntdev_ioctl_map_grant_ref(priv, ptr); 1042 1043 case IOCTL_GNTDEV_UNMAP_GRANT_REF: 1044 return gntdev_ioctl_unmap_grant_ref(priv, ptr); 1045 1046 case IOCTL_GNTDEV_GET_OFFSET_FOR_VADDR: 1047 return gntdev_ioctl_get_offset_for_vaddr(priv, ptr); 1048 1049 case IOCTL_GNTDEV_SET_UNMAP_NOTIFY: 1050 return gntdev_ioctl_notify(priv, ptr); 1051 1052 case IOCTL_GNTDEV_GRANT_COPY: 1053 return gntdev_ioctl_grant_copy(priv, ptr); 1054 1055 #ifdef CONFIG_XEN_GNTDEV_DMABUF 1056 case IOCTL_GNTDEV_DMABUF_EXP_FROM_REFS: 1057 return gntdev_ioctl_dmabuf_exp_from_refs(priv, use_ptemod, ptr); 1058 1059 case IOCTL_GNTDEV_DMABUF_EXP_WAIT_RELEASED: 1060 return gntdev_ioctl_dmabuf_exp_wait_released(priv, ptr); 1061 1062 case IOCTL_GNTDEV_DMABUF_IMP_TO_REFS: 1063 return gntdev_ioctl_dmabuf_imp_to_refs(priv, ptr); 1064 1065 case IOCTL_GNTDEV_DMABUF_IMP_RELEASE: 1066 return gntdev_ioctl_dmabuf_imp_release(priv, ptr); 1067 #endif 1068 1069 default: 1070 pr_debug("priv %p, unknown cmd %x\n", priv, cmd); 1071 return -ENOIOCTLCMD; 1072 } 1073 1074 return 0; 1075 } 1076 1077 static int gntdev_mmap(struct file *flip, struct vm_area_struct *vma) 1078 { 1079 struct gntdev_priv *priv = flip->private_data; 1080 int index = vma->vm_pgoff; 1081 int count = vma_pages(vma); 1082 struct gntdev_grant_map *map; 1083 int i, err = -EINVAL; 1084 1085 if ((vma->vm_flags & VM_WRITE) && !(vma->vm_flags & VM_SHARED)) 1086 return -EINVAL; 1087 1088 pr_debug("map %d+%d at %lx (pgoff %lx)\n", 1089 index, count, vma->vm_start, vma->vm_pgoff); 1090 1091 mutex_lock(&priv->lock); 1092 map = gntdev_find_map_index(priv, index, count); 1093 if (!map) 1094 goto unlock_out; 1095 if (use_ptemod && map->vma) 1096 goto unlock_out; 1097 if (use_ptemod && priv->mm != vma->vm_mm) { 1098 pr_warn("Huh? Other mm?\n"); 1099 goto unlock_out; 1100 } 1101 1102 refcount_inc(&map->users); 1103 1104 vma->vm_ops = &gntdev_vmops; 1105 1106 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP | VM_MIXEDMAP; 1107 1108 if (use_ptemod) 1109 vma->vm_flags |= VM_DONTCOPY; 1110 1111 vma->vm_private_data = map; 1112 1113 if (use_ptemod) 1114 map->vma = vma; 1115 1116 if (map->flags) { 1117 if ((vma->vm_flags & VM_WRITE) && 1118 (map->flags & GNTMAP_readonly)) 1119 goto out_unlock_put; 1120 } else { 1121 map->flags = GNTMAP_host_map; 1122 if (!(vma->vm_flags & VM_WRITE)) 1123 map->flags |= GNTMAP_readonly; 1124 } 1125 1126 mutex_unlock(&priv->lock); 1127 1128 if (use_ptemod) { 1129 map->pages_vm_start = vma->vm_start; 1130 err = apply_to_page_range(vma->vm_mm, vma->vm_start, 1131 vma->vm_end - vma->vm_start, 1132 find_grant_ptes, map); 1133 if (err) { 1134 pr_warn("find_grant_ptes() failure.\n"); 1135 goto out_put_map; 1136 } 1137 } 1138 1139 err = gntdev_map_grant_pages(map); 1140 if (err) 1141 goto out_put_map; 1142 1143 if (!use_ptemod) { 1144 for (i = 0; i < count; i++) { 1145 err = vm_insert_page(vma, vma->vm_start + i*PAGE_SIZE, 1146 map->pages[i]); 1147 if (err) 1148 goto out_put_map; 1149 } 1150 } else { 1151 #ifdef CONFIG_X86 1152 /* 1153 * If the PTEs were not made special by the grant map 1154 * hypercall, do so here. 1155 * 1156 * This is racy since the mapping is already visible 1157 * to userspace but userspace should be well-behaved 1158 * enough to not touch it until the mmap() call 1159 * returns. 1160 */ 1161 if (!xen_feature(XENFEAT_gnttab_map_avail_bits)) { 1162 apply_to_page_range(vma->vm_mm, vma->vm_start, 1163 vma->vm_end - vma->vm_start, 1164 set_grant_ptes_as_special, NULL); 1165 } 1166 #endif 1167 } 1168 1169 return 0; 1170 1171 unlock_out: 1172 mutex_unlock(&priv->lock); 1173 return err; 1174 1175 out_unlock_put: 1176 mutex_unlock(&priv->lock); 1177 out_put_map: 1178 if (use_ptemod) { 1179 map->vma = NULL; 1180 unmap_grant_pages(map, 0, map->count); 1181 } 1182 gntdev_put_map(priv, map); 1183 return err; 1184 } 1185 1186 static const struct file_operations gntdev_fops = { 1187 .owner = THIS_MODULE, 1188 .open = gntdev_open, 1189 .release = gntdev_release, 1190 .mmap = gntdev_mmap, 1191 .unlocked_ioctl = gntdev_ioctl 1192 }; 1193 1194 static struct miscdevice gntdev_miscdev = { 1195 .minor = MISC_DYNAMIC_MINOR, 1196 .name = "xen/gntdev", 1197 .fops = &gntdev_fops, 1198 }; 1199 1200 /* ------------------------------------------------------------------ */ 1201 1202 static int __init gntdev_init(void) 1203 { 1204 int err; 1205 1206 if (!xen_domain()) 1207 return -ENODEV; 1208 1209 use_ptemod = !xen_feature(XENFEAT_auto_translated_physmap); 1210 1211 err = misc_register(&gntdev_miscdev); 1212 if (err != 0) { 1213 pr_err("Could not register gntdev device\n"); 1214 return err; 1215 } 1216 return 0; 1217 } 1218 1219 static void __exit gntdev_exit(void) 1220 { 1221 misc_deregister(&gntdev_miscdev); 1222 } 1223 1224 module_init(gntdev_init); 1225 module_exit(gntdev_exit); 1226 1227 /* ------------------------------------------------------------------ */ 1228