1 /****************************************************************************** 2 * gntdev.c 3 * 4 * Device for accessing (in user-space) pages that have been granted by other 5 * domains. 6 * 7 * Copyright (c) 2006-2007, D G Murray. 8 * (c) 2009 Gerd Hoffmann <kraxel@redhat.com> 9 * (c) 2018 Oleksandr Andrushchenko, EPAM Systems Inc. 10 * 11 * This program is distributed in the hope that it will be useful, 12 * but WITHOUT ANY WARRANTY; without even the implied warranty of 13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 14 * GNU General Public License for more details. 15 * 16 * You should have received a copy of the GNU General Public License 17 * along with this program; if not, write to the Free Software 18 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA 19 */ 20 21 #undef DEBUG 22 23 #define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt 24 25 #include <linux/dma-mapping.h> 26 #include <linux/module.h> 27 #include <linux/kernel.h> 28 #include <linux/init.h> 29 #include <linux/miscdevice.h> 30 #include <linux/fs.h> 31 #include <linux/uaccess.h> 32 #include <linux/sched.h> 33 #include <linux/sched/mm.h> 34 #include <linux/spinlock.h> 35 #include <linux/slab.h> 36 #include <linux/highmem.h> 37 #include <linux/refcount.h> 38 #include <linux/workqueue.h> 39 40 #include <xen/xen.h> 41 #include <xen/grant_table.h> 42 #include <xen/balloon.h> 43 #include <xen/gntdev.h> 44 #include <xen/events.h> 45 #include <xen/page.h> 46 #include <asm/xen/hypervisor.h> 47 #include <asm/xen/hypercall.h> 48 49 #include "gntdev-common.h" 50 #ifdef CONFIG_XEN_GNTDEV_DMABUF 51 #include "gntdev-dmabuf.h" 52 #endif 53 54 MODULE_LICENSE("GPL"); 55 MODULE_AUTHOR("Derek G. Murray <Derek.Murray@cl.cam.ac.uk>, " 56 "Gerd Hoffmann <kraxel@redhat.com>"); 57 MODULE_DESCRIPTION("User-space granted page access driver"); 58 59 static unsigned int limit = 64*1024; 60 module_param(limit, uint, 0644); 61 MODULE_PARM_DESC(limit, 62 "Maximum number of grants that may be mapped by one mapping request"); 63 64 /* True in PV mode, false otherwise */ 65 static int use_ptemod; 66 67 static void unmap_grant_pages(struct gntdev_grant_map *map, 68 int offset, int pages); 69 70 static struct miscdevice gntdev_miscdev; 71 72 /* ------------------------------------------------------------------ */ 73 74 bool gntdev_test_page_count(unsigned int count) 75 { 76 return !count || count > limit; 77 } 78 79 static void gntdev_print_maps(struct gntdev_priv *priv, 80 char *text, int text_index) 81 { 82 #ifdef DEBUG 83 struct gntdev_grant_map *map; 84 85 pr_debug("%s: maps list (priv %p)\n", __func__, priv); 86 list_for_each_entry(map, &priv->maps, next) 87 pr_debug(" index %2d, count %2d %s\n", 88 map->index, map->count, 89 map->index == text_index && text ? text : ""); 90 #endif 91 } 92 93 static void gntdev_free_map(struct gntdev_grant_map *map) 94 { 95 if (map == NULL) 96 return; 97 98 #ifdef CONFIG_XEN_GRANT_DMA_ALLOC 99 if (map->dma_vaddr) { 100 struct gnttab_dma_alloc_args args; 101 102 args.dev = map->dma_dev; 103 args.coherent = !!(map->dma_flags & GNTDEV_DMA_FLAG_COHERENT); 104 args.nr_pages = map->count; 105 args.pages = map->pages; 106 args.frames = map->frames; 107 args.vaddr = map->dma_vaddr; 108 args.dev_bus_addr = map->dma_bus_addr; 109 110 gnttab_dma_free_pages(&args); 111 } else 112 #endif 113 if (map->pages) 114 gnttab_free_pages(map->count, map->pages); 115 116 #ifdef CONFIG_XEN_GRANT_DMA_ALLOC 117 kvfree(map->frames); 118 #endif 119 kvfree(map->pages); 120 kvfree(map->grants); 121 kvfree(map->map_ops); 122 kvfree(map->unmap_ops); 123 kvfree(map->kmap_ops); 124 kvfree(map->kunmap_ops); 125 kvfree(map->being_removed); 126 kfree(map); 127 } 128 129 struct gntdev_grant_map *gntdev_alloc_map(struct gntdev_priv *priv, int count, 130 int dma_flags) 131 { 132 struct gntdev_grant_map *add; 133 int i; 134 135 add = kzalloc(sizeof(*add), GFP_KERNEL); 136 if (NULL == add) 137 return NULL; 138 139 add->grants = kvmalloc_array(count, sizeof(add->grants[0]), 140 GFP_KERNEL); 141 add->map_ops = kvmalloc_array(count, sizeof(add->map_ops[0]), 142 GFP_KERNEL); 143 add->unmap_ops = kvmalloc_array(count, sizeof(add->unmap_ops[0]), 144 GFP_KERNEL); 145 add->pages = kvcalloc(count, sizeof(add->pages[0]), GFP_KERNEL); 146 add->being_removed = 147 kvcalloc(count, sizeof(add->being_removed[0]), GFP_KERNEL); 148 if (NULL == add->grants || 149 NULL == add->map_ops || 150 NULL == add->unmap_ops || 151 NULL == add->pages || 152 NULL == add->being_removed) 153 goto err; 154 if (use_ptemod) { 155 add->kmap_ops = kvmalloc_array(count, sizeof(add->kmap_ops[0]), 156 GFP_KERNEL); 157 add->kunmap_ops = kvmalloc_array(count, sizeof(add->kunmap_ops[0]), 158 GFP_KERNEL); 159 if (NULL == add->kmap_ops || NULL == add->kunmap_ops) 160 goto err; 161 } 162 163 #ifdef CONFIG_XEN_GRANT_DMA_ALLOC 164 add->dma_flags = dma_flags; 165 166 /* 167 * Check if this mapping is requested to be backed 168 * by a DMA buffer. 169 */ 170 if (dma_flags & (GNTDEV_DMA_FLAG_WC | GNTDEV_DMA_FLAG_COHERENT)) { 171 struct gnttab_dma_alloc_args args; 172 173 add->frames = kvcalloc(count, sizeof(add->frames[0]), 174 GFP_KERNEL); 175 if (!add->frames) 176 goto err; 177 178 /* Remember the device, so we can free DMA memory. */ 179 add->dma_dev = priv->dma_dev; 180 181 args.dev = priv->dma_dev; 182 args.coherent = !!(dma_flags & GNTDEV_DMA_FLAG_COHERENT); 183 args.nr_pages = count; 184 args.pages = add->pages; 185 args.frames = add->frames; 186 187 if (gnttab_dma_alloc_pages(&args)) 188 goto err; 189 190 add->dma_vaddr = args.vaddr; 191 add->dma_bus_addr = args.dev_bus_addr; 192 } else 193 #endif 194 if (gnttab_alloc_pages(count, add->pages)) 195 goto err; 196 197 for (i = 0; i < count; i++) { 198 add->grants[i].domid = DOMID_INVALID; 199 add->grants[i].ref = INVALID_GRANT_REF; 200 add->map_ops[i].handle = INVALID_GRANT_HANDLE; 201 add->unmap_ops[i].handle = INVALID_GRANT_HANDLE; 202 if (use_ptemod) { 203 add->kmap_ops[i].handle = INVALID_GRANT_HANDLE; 204 add->kunmap_ops[i].handle = INVALID_GRANT_HANDLE; 205 } 206 } 207 208 add->index = 0; 209 add->count = count; 210 refcount_set(&add->users, 1); 211 212 return add; 213 214 err: 215 gntdev_free_map(add); 216 return NULL; 217 } 218 219 void gntdev_add_map(struct gntdev_priv *priv, struct gntdev_grant_map *add) 220 { 221 struct gntdev_grant_map *map; 222 223 list_for_each_entry(map, &priv->maps, next) { 224 if (add->index + add->count < map->index) { 225 list_add_tail(&add->next, &map->next); 226 goto done; 227 } 228 add->index = map->index + map->count; 229 } 230 list_add_tail(&add->next, &priv->maps); 231 232 done: 233 gntdev_print_maps(priv, "[new]", add->index); 234 } 235 236 static struct gntdev_grant_map *gntdev_find_map_index(struct gntdev_priv *priv, 237 int index, int count) 238 { 239 struct gntdev_grant_map *map; 240 241 list_for_each_entry(map, &priv->maps, next) { 242 if (map->index != index) 243 continue; 244 if (count && map->count != count) 245 continue; 246 return map; 247 } 248 return NULL; 249 } 250 251 void gntdev_put_map(struct gntdev_priv *priv, struct gntdev_grant_map *map) 252 { 253 if (!map) 254 return; 255 256 if (!refcount_dec_and_test(&map->users)) 257 return; 258 259 if (map->pages && !use_ptemod) { 260 /* 261 * Increment the reference count. This ensures that the 262 * subsequent call to unmap_grant_pages() will not wind up 263 * re-entering itself. It *can* wind up calling 264 * gntdev_put_map() recursively, but such calls will be with a 265 * reference count greater than 1, so they will return before 266 * this code is reached. The recursion depth is thus limited to 267 * 1. Do NOT use refcount_inc() here, as it will detect that 268 * the reference count is zero and WARN(). 269 */ 270 refcount_set(&map->users, 1); 271 272 /* 273 * Unmap the grants. This may or may not be asynchronous, so it 274 * is possible that the reference count is 1 on return, but it 275 * could also be greater than 1. 276 */ 277 unmap_grant_pages(map, 0, map->count); 278 279 /* Check if the memory now needs to be freed */ 280 if (!refcount_dec_and_test(&map->users)) 281 return; 282 283 /* 284 * All pages have been returned to the hypervisor, so free the 285 * map. 286 */ 287 } 288 289 if (map->notify.flags & UNMAP_NOTIFY_SEND_EVENT) { 290 notify_remote_via_evtchn(map->notify.event); 291 evtchn_put(map->notify.event); 292 } 293 gntdev_free_map(map); 294 } 295 296 /* ------------------------------------------------------------------ */ 297 298 static int find_grant_ptes(pte_t *pte, unsigned long addr, void *data) 299 { 300 struct gntdev_grant_map *map = data; 301 unsigned int pgnr = (addr - map->vma->vm_start) >> PAGE_SHIFT; 302 int flags = map->flags | GNTMAP_application_map | GNTMAP_contains_pte | 303 (1 << _GNTMAP_guest_avail0); 304 u64 pte_maddr; 305 306 BUG_ON(pgnr >= map->count); 307 pte_maddr = arbitrary_virt_to_machine(pte).maddr; 308 309 gnttab_set_map_op(&map->map_ops[pgnr], pte_maddr, flags, 310 map->grants[pgnr].ref, 311 map->grants[pgnr].domid); 312 gnttab_set_unmap_op(&map->unmap_ops[pgnr], pte_maddr, flags, 313 INVALID_GRANT_HANDLE); 314 return 0; 315 } 316 317 int gntdev_map_grant_pages(struct gntdev_grant_map *map) 318 { 319 size_t alloced = 0; 320 int i, err = 0; 321 322 if (!use_ptemod) { 323 /* Note: it could already be mapped */ 324 if (map->map_ops[0].handle != INVALID_GRANT_HANDLE) 325 return 0; 326 for (i = 0; i < map->count; i++) { 327 unsigned long addr = (unsigned long) 328 pfn_to_kaddr(page_to_pfn(map->pages[i])); 329 gnttab_set_map_op(&map->map_ops[i], addr, map->flags, 330 map->grants[i].ref, 331 map->grants[i].domid); 332 gnttab_set_unmap_op(&map->unmap_ops[i], addr, 333 map->flags, INVALID_GRANT_HANDLE); 334 } 335 } else { 336 /* 337 * Setup the map_ops corresponding to the pte entries pointing 338 * to the kernel linear addresses of the struct pages. 339 * These ptes are completely different from the user ptes dealt 340 * with find_grant_ptes. 341 * Note that GNTMAP_device_map isn't needed here: The 342 * dev_bus_addr output field gets consumed only from ->map_ops, 343 * and by not requesting it when mapping we also avoid needing 344 * to mirror dev_bus_addr into ->unmap_ops (and holding an extra 345 * reference to the page in the hypervisor). 346 */ 347 unsigned int flags = (map->flags & ~GNTMAP_device_map) | 348 GNTMAP_host_map; 349 350 for (i = 0; i < map->count; i++) { 351 unsigned long address = (unsigned long) 352 pfn_to_kaddr(page_to_pfn(map->pages[i])); 353 BUG_ON(PageHighMem(map->pages[i])); 354 355 gnttab_set_map_op(&map->kmap_ops[i], address, flags, 356 map->grants[i].ref, 357 map->grants[i].domid); 358 gnttab_set_unmap_op(&map->kunmap_ops[i], address, 359 flags, INVALID_GRANT_HANDLE); 360 } 361 } 362 363 pr_debug("map %d+%d\n", map->index, map->count); 364 err = gnttab_map_refs(map->map_ops, map->kmap_ops, map->pages, 365 map->count); 366 367 for (i = 0; i < map->count; i++) { 368 if (map->map_ops[i].status == GNTST_okay) { 369 map->unmap_ops[i].handle = map->map_ops[i].handle; 370 if (!use_ptemod) 371 alloced++; 372 } else if (!err) 373 err = -EINVAL; 374 375 if (map->flags & GNTMAP_device_map) 376 map->unmap_ops[i].dev_bus_addr = map->map_ops[i].dev_bus_addr; 377 378 if (use_ptemod) { 379 if (map->kmap_ops[i].status == GNTST_okay) { 380 if (map->map_ops[i].status == GNTST_okay) 381 alloced++; 382 map->kunmap_ops[i].handle = map->kmap_ops[i].handle; 383 } else if (!err) 384 err = -EINVAL; 385 } 386 } 387 atomic_add(alloced, &map->live_grants); 388 return err; 389 } 390 391 static void __unmap_grant_pages_done(int result, 392 struct gntab_unmap_queue_data *data) 393 { 394 unsigned int i; 395 struct gntdev_grant_map *map = data->data; 396 unsigned int offset = data->unmap_ops - map->unmap_ops; 397 398 for (i = 0; i < data->count; i++) { 399 WARN_ON(map->unmap_ops[offset+i].status); 400 pr_debug("unmap handle=%d st=%d\n", 401 map->unmap_ops[offset+i].handle, 402 map->unmap_ops[offset+i].status); 403 map->unmap_ops[offset+i].handle = INVALID_GRANT_HANDLE; 404 if (use_ptemod) { 405 WARN_ON(map->kunmap_ops[offset+i].status); 406 pr_debug("kunmap handle=%u st=%d\n", 407 map->kunmap_ops[offset+i].handle, 408 map->kunmap_ops[offset+i].status); 409 map->kunmap_ops[offset+i].handle = INVALID_GRANT_HANDLE; 410 } 411 } 412 /* 413 * Decrease the live-grant counter. This must happen after the loop to 414 * prevent premature reuse of the grants by gnttab_mmap(). 415 */ 416 atomic_sub(data->count, &map->live_grants); 417 418 /* Release reference taken by __unmap_grant_pages */ 419 gntdev_put_map(NULL, map); 420 } 421 422 static void __unmap_grant_pages(struct gntdev_grant_map *map, int offset, 423 int pages) 424 { 425 if (map->notify.flags & UNMAP_NOTIFY_CLEAR_BYTE) { 426 int pgno = (map->notify.addr >> PAGE_SHIFT); 427 428 if (pgno >= offset && pgno < offset + pages) { 429 /* No need for kmap, pages are in lowmem */ 430 uint8_t *tmp = pfn_to_kaddr(page_to_pfn(map->pages[pgno])); 431 432 tmp[map->notify.addr & (PAGE_SIZE-1)] = 0; 433 map->notify.flags &= ~UNMAP_NOTIFY_CLEAR_BYTE; 434 } 435 } 436 437 map->unmap_data.unmap_ops = map->unmap_ops + offset; 438 map->unmap_data.kunmap_ops = use_ptemod ? map->kunmap_ops + offset : NULL; 439 map->unmap_data.pages = map->pages + offset; 440 map->unmap_data.count = pages; 441 map->unmap_data.done = __unmap_grant_pages_done; 442 map->unmap_data.data = map; 443 refcount_inc(&map->users); /* to keep map alive during async call below */ 444 445 gnttab_unmap_refs_async(&map->unmap_data); 446 } 447 448 static void unmap_grant_pages(struct gntdev_grant_map *map, int offset, 449 int pages) 450 { 451 int range; 452 453 if (atomic_read(&map->live_grants) == 0) 454 return; /* Nothing to do */ 455 456 pr_debug("unmap %d+%d [%d+%d]\n", map->index, map->count, offset, pages); 457 458 /* It is possible the requested range will have a "hole" where we 459 * already unmapped some of the grants. Only unmap valid ranges. 460 */ 461 while (pages) { 462 while (pages && map->being_removed[offset]) { 463 offset++; 464 pages--; 465 } 466 range = 0; 467 while (range < pages) { 468 if (map->being_removed[offset + range]) 469 break; 470 map->being_removed[offset + range] = true; 471 range++; 472 } 473 if (range) 474 __unmap_grant_pages(map, offset, range); 475 offset += range; 476 pages -= range; 477 } 478 } 479 480 /* ------------------------------------------------------------------ */ 481 482 static void gntdev_vma_open(struct vm_area_struct *vma) 483 { 484 struct gntdev_grant_map *map = vma->vm_private_data; 485 486 pr_debug("gntdev_vma_open %p\n", vma); 487 refcount_inc(&map->users); 488 } 489 490 static void gntdev_vma_close(struct vm_area_struct *vma) 491 { 492 struct gntdev_grant_map *map = vma->vm_private_data; 493 struct file *file = vma->vm_file; 494 struct gntdev_priv *priv = file->private_data; 495 496 pr_debug("gntdev_vma_close %p\n", vma); 497 if (use_ptemod) { 498 WARN_ON(map->vma != vma); 499 mmu_interval_notifier_remove(&map->notifier); 500 map->vma = NULL; 501 } 502 vma->vm_private_data = NULL; 503 gntdev_put_map(priv, map); 504 } 505 506 static struct page *gntdev_vma_find_special_page(struct vm_area_struct *vma, 507 unsigned long addr) 508 { 509 struct gntdev_grant_map *map = vma->vm_private_data; 510 511 return map->pages[(addr - map->pages_vm_start) >> PAGE_SHIFT]; 512 } 513 514 static const struct vm_operations_struct gntdev_vmops = { 515 .open = gntdev_vma_open, 516 .close = gntdev_vma_close, 517 .find_special_page = gntdev_vma_find_special_page, 518 }; 519 520 /* ------------------------------------------------------------------ */ 521 522 static bool gntdev_invalidate(struct mmu_interval_notifier *mn, 523 const struct mmu_notifier_range *range, 524 unsigned long cur_seq) 525 { 526 struct gntdev_grant_map *map = 527 container_of(mn, struct gntdev_grant_map, notifier); 528 unsigned long mstart, mend; 529 530 if (!mmu_notifier_range_blockable(range)) 531 return false; 532 533 /* 534 * If the VMA is split or otherwise changed the notifier is not 535 * updated, but we don't want to process VA's outside the modified 536 * VMA. FIXME: It would be much more understandable to just prevent 537 * modifying the VMA in the first place. 538 */ 539 if (map->vma->vm_start >= range->end || 540 map->vma->vm_end <= range->start) 541 return true; 542 543 mstart = max(range->start, map->vma->vm_start); 544 mend = min(range->end, map->vma->vm_end); 545 pr_debug("map %d+%d (%lx %lx), range %lx %lx, mrange %lx %lx\n", 546 map->index, map->count, 547 map->vma->vm_start, map->vma->vm_end, 548 range->start, range->end, mstart, mend); 549 unmap_grant_pages(map, 550 (mstart - map->vma->vm_start) >> PAGE_SHIFT, 551 (mend - mstart) >> PAGE_SHIFT); 552 553 return true; 554 } 555 556 static const struct mmu_interval_notifier_ops gntdev_mmu_ops = { 557 .invalidate = gntdev_invalidate, 558 }; 559 560 /* ------------------------------------------------------------------ */ 561 562 static int gntdev_open(struct inode *inode, struct file *flip) 563 { 564 struct gntdev_priv *priv; 565 566 priv = kzalloc(sizeof(*priv), GFP_KERNEL); 567 if (!priv) 568 return -ENOMEM; 569 570 INIT_LIST_HEAD(&priv->maps); 571 mutex_init(&priv->lock); 572 573 #ifdef CONFIG_XEN_GNTDEV_DMABUF 574 priv->dmabuf_priv = gntdev_dmabuf_init(flip); 575 if (IS_ERR(priv->dmabuf_priv)) { 576 int ret = PTR_ERR(priv->dmabuf_priv); 577 578 kfree(priv); 579 return ret; 580 } 581 #endif 582 583 flip->private_data = priv; 584 #ifdef CONFIG_XEN_GRANT_DMA_ALLOC 585 priv->dma_dev = gntdev_miscdev.this_device; 586 dma_coerce_mask_and_coherent(priv->dma_dev, DMA_BIT_MASK(64)); 587 #endif 588 pr_debug("priv %p\n", priv); 589 590 return 0; 591 } 592 593 static int gntdev_release(struct inode *inode, struct file *flip) 594 { 595 struct gntdev_priv *priv = flip->private_data; 596 struct gntdev_grant_map *map; 597 598 pr_debug("priv %p\n", priv); 599 600 mutex_lock(&priv->lock); 601 while (!list_empty(&priv->maps)) { 602 map = list_entry(priv->maps.next, 603 struct gntdev_grant_map, next); 604 list_del(&map->next); 605 gntdev_put_map(NULL /* already removed */, map); 606 } 607 mutex_unlock(&priv->lock); 608 609 #ifdef CONFIG_XEN_GNTDEV_DMABUF 610 gntdev_dmabuf_fini(priv->dmabuf_priv); 611 #endif 612 613 kfree(priv); 614 return 0; 615 } 616 617 static long gntdev_ioctl_map_grant_ref(struct gntdev_priv *priv, 618 struct ioctl_gntdev_map_grant_ref __user *u) 619 { 620 struct ioctl_gntdev_map_grant_ref op; 621 struct gntdev_grant_map *map; 622 int err; 623 624 if (copy_from_user(&op, u, sizeof(op)) != 0) 625 return -EFAULT; 626 pr_debug("priv %p, add %d\n", priv, op.count); 627 if (unlikely(gntdev_test_page_count(op.count))) 628 return -EINVAL; 629 630 err = -ENOMEM; 631 map = gntdev_alloc_map(priv, op.count, 0 /* This is not a dma-buf. */); 632 if (!map) 633 return err; 634 635 if (copy_from_user(map->grants, &u->refs, 636 sizeof(map->grants[0]) * op.count) != 0) { 637 gntdev_put_map(NULL, map); 638 return -EFAULT; 639 } 640 641 mutex_lock(&priv->lock); 642 gntdev_add_map(priv, map); 643 op.index = map->index << PAGE_SHIFT; 644 mutex_unlock(&priv->lock); 645 646 if (copy_to_user(u, &op, sizeof(op)) != 0) 647 return -EFAULT; 648 649 return 0; 650 } 651 652 static long gntdev_ioctl_unmap_grant_ref(struct gntdev_priv *priv, 653 struct ioctl_gntdev_unmap_grant_ref __user *u) 654 { 655 struct ioctl_gntdev_unmap_grant_ref op; 656 struct gntdev_grant_map *map; 657 int err = -ENOENT; 658 659 if (copy_from_user(&op, u, sizeof(op)) != 0) 660 return -EFAULT; 661 pr_debug("priv %p, del %d+%d\n", priv, (int)op.index, (int)op.count); 662 663 mutex_lock(&priv->lock); 664 map = gntdev_find_map_index(priv, op.index >> PAGE_SHIFT, op.count); 665 if (map) { 666 list_del(&map->next); 667 err = 0; 668 } 669 mutex_unlock(&priv->lock); 670 if (map) 671 gntdev_put_map(priv, map); 672 return err; 673 } 674 675 static long gntdev_ioctl_get_offset_for_vaddr(struct gntdev_priv *priv, 676 struct ioctl_gntdev_get_offset_for_vaddr __user *u) 677 { 678 struct ioctl_gntdev_get_offset_for_vaddr op; 679 struct vm_area_struct *vma; 680 struct gntdev_grant_map *map; 681 int rv = -EINVAL; 682 683 if (copy_from_user(&op, u, sizeof(op)) != 0) 684 return -EFAULT; 685 pr_debug("priv %p, offset for vaddr %lx\n", priv, (unsigned long)op.vaddr); 686 687 mmap_read_lock(current->mm); 688 vma = find_vma(current->mm, op.vaddr); 689 if (!vma || vma->vm_ops != &gntdev_vmops) 690 goto out_unlock; 691 692 map = vma->vm_private_data; 693 if (!map) 694 goto out_unlock; 695 696 op.offset = map->index << PAGE_SHIFT; 697 op.count = map->count; 698 rv = 0; 699 700 out_unlock: 701 mmap_read_unlock(current->mm); 702 703 if (rv == 0 && copy_to_user(u, &op, sizeof(op)) != 0) 704 return -EFAULT; 705 return rv; 706 } 707 708 static long gntdev_ioctl_notify(struct gntdev_priv *priv, void __user *u) 709 { 710 struct ioctl_gntdev_unmap_notify op; 711 struct gntdev_grant_map *map; 712 int rc; 713 int out_flags; 714 evtchn_port_t out_event; 715 716 if (copy_from_user(&op, u, sizeof(op))) 717 return -EFAULT; 718 719 if (op.action & ~(UNMAP_NOTIFY_CLEAR_BYTE|UNMAP_NOTIFY_SEND_EVENT)) 720 return -EINVAL; 721 722 /* We need to grab a reference to the event channel we are going to use 723 * to send the notify before releasing the reference we may already have 724 * (if someone has called this ioctl twice). This is required so that 725 * it is possible to change the clear_byte part of the notification 726 * without disturbing the event channel part, which may now be the last 727 * reference to that event channel. 728 */ 729 if (op.action & UNMAP_NOTIFY_SEND_EVENT) { 730 if (evtchn_get(op.event_channel_port)) 731 return -EINVAL; 732 } 733 734 out_flags = op.action; 735 out_event = op.event_channel_port; 736 737 mutex_lock(&priv->lock); 738 739 list_for_each_entry(map, &priv->maps, next) { 740 uint64_t begin = map->index << PAGE_SHIFT; 741 uint64_t end = (map->index + map->count) << PAGE_SHIFT; 742 if (op.index >= begin && op.index < end) 743 goto found; 744 } 745 rc = -ENOENT; 746 goto unlock_out; 747 748 found: 749 if ((op.action & UNMAP_NOTIFY_CLEAR_BYTE) && 750 (map->flags & GNTMAP_readonly)) { 751 rc = -EINVAL; 752 goto unlock_out; 753 } 754 755 out_flags = map->notify.flags; 756 out_event = map->notify.event; 757 758 map->notify.flags = op.action; 759 map->notify.addr = op.index - (map->index << PAGE_SHIFT); 760 map->notify.event = op.event_channel_port; 761 762 rc = 0; 763 764 unlock_out: 765 mutex_unlock(&priv->lock); 766 767 /* Drop the reference to the event channel we did not save in the map */ 768 if (out_flags & UNMAP_NOTIFY_SEND_EVENT) 769 evtchn_put(out_event); 770 771 return rc; 772 } 773 774 #define GNTDEV_COPY_BATCH 16 775 776 struct gntdev_copy_batch { 777 struct gnttab_copy ops[GNTDEV_COPY_BATCH]; 778 struct page *pages[GNTDEV_COPY_BATCH]; 779 s16 __user *status[GNTDEV_COPY_BATCH]; 780 unsigned int nr_ops; 781 unsigned int nr_pages; 782 bool writeable; 783 }; 784 785 static int gntdev_get_page(struct gntdev_copy_batch *batch, void __user *virt, 786 unsigned long *gfn) 787 { 788 unsigned long addr = (unsigned long)virt; 789 struct page *page; 790 unsigned long xen_pfn; 791 int ret; 792 793 ret = pin_user_pages_fast(addr, 1, batch->writeable ? FOLL_WRITE : 0, &page); 794 if (ret < 0) 795 return ret; 796 797 batch->pages[batch->nr_pages++] = page; 798 799 xen_pfn = page_to_xen_pfn(page) + XEN_PFN_DOWN(addr & ~PAGE_MASK); 800 *gfn = pfn_to_gfn(xen_pfn); 801 802 return 0; 803 } 804 805 static void gntdev_put_pages(struct gntdev_copy_batch *batch) 806 { 807 unpin_user_pages_dirty_lock(batch->pages, batch->nr_pages, batch->writeable); 808 batch->nr_pages = 0; 809 batch->writeable = false; 810 } 811 812 static int gntdev_copy(struct gntdev_copy_batch *batch) 813 { 814 unsigned int i; 815 816 gnttab_batch_copy(batch->ops, batch->nr_ops); 817 gntdev_put_pages(batch); 818 819 /* 820 * For each completed op, update the status if the op failed 821 * and all previous ops for the segment were successful. 822 */ 823 for (i = 0; i < batch->nr_ops; i++) { 824 s16 status = batch->ops[i].status; 825 s16 old_status; 826 827 if (status == GNTST_okay) 828 continue; 829 830 if (__get_user(old_status, batch->status[i])) 831 return -EFAULT; 832 833 if (old_status != GNTST_okay) 834 continue; 835 836 if (__put_user(status, batch->status[i])) 837 return -EFAULT; 838 } 839 840 batch->nr_ops = 0; 841 return 0; 842 } 843 844 static int gntdev_grant_copy_seg(struct gntdev_copy_batch *batch, 845 struct gntdev_grant_copy_segment *seg, 846 s16 __user *status) 847 { 848 uint16_t copied = 0; 849 850 /* 851 * Disallow local -> local copies since there is only space in 852 * batch->pages for one page per-op and this would be a very 853 * expensive memcpy(). 854 */ 855 if (!(seg->flags & (GNTCOPY_source_gref | GNTCOPY_dest_gref))) 856 return -EINVAL; 857 858 /* Can't cross page if source/dest is a grant ref. */ 859 if (seg->flags & GNTCOPY_source_gref) { 860 if (seg->source.foreign.offset + seg->len > XEN_PAGE_SIZE) 861 return -EINVAL; 862 } 863 if (seg->flags & GNTCOPY_dest_gref) { 864 if (seg->dest.foreign.offset + seg->len > XEN_PAGE_SIZE) 865 return -EINVAL; 866 } 867 868 if (put_user(GNTST_okay, status)) 869 return -EFAULT; 870 871 while (copied < seg->len) { 872 struct gnttab_copy *op; 873 void __user *virt; 874 size_t len, off; 875 unsigned long gfn; 876 int ret; 877 878 if (batch->nr_ops >= GNTDEV_COPY_BATCH) { 879 ret = gntdev_copy(batch); 880 if (ret < 0) 881 return ret; 882 } 883 884 len = seg->len - copied; 885 886 op = &batch->ops[batch->nr_ops]; 887 op->flags = 0; 888 889 if (seg->flags & GNTCOPY_source_gref) { 890 op->source.u.ref = seg->source.foreign.ref; 891 op->source.domid = seg->source.foreign.domid; 892 op->source.offset = seg->source.foreign.offset + copied; 893 op->flags |= GNTCOPY_source_gref; 894 } else { 895 virt = seg->source.virt + copied; 896 off = (unsigned long)virt & ~XEN_PAGE_MASK; 897 len = min(len, (size_t)XEN_PAGE_SIZE - off); 898 batch->writeable = false; 899 900 ret = gntdev_get_page(batch, virt, &gfn); 901 if (ret < 0) 902 return ret; 903 904 op->source.u.gmfn = gfn; 905 op->source.domid = DOMID_SELF; 906 op->source.offset = off; 907 } 908 909 if (seg->flags & GNTCOPY_dest_gref) { 910 op->dest.u.ref = seg->dest.foreign.ref; 911 op->dest.domid = seg->dest.foreign.domid; 912 op->dest.offset = seg->dest.foreign.offset + copied; 913 op->flags |= GNTCOPY_dest_gref; 914 } else { 915 virt = seg->dest.virt + copied; 916 off = (unsigned long)virt & ~XEN_PAGE_MASK; 917 len = min(len, (size_t)XEN_PAGE_SIZE - off); 918 batch->writeable = true; 919 920 ret = gntdev_get_page(batch, virt, &gfn); 921 if (ret < 0) 922 return ret; 923 924 op->dest.u.gmfn = gfn; 925 op->dest.domid = DOMID_SELF; 926 op->dest.offset = off; 927 } 928 929 op->len = len; 930 copied += len; 931 932 batch->status[batch->nr_ops] = status; 933 batch->nr_ops++; 934 } 935 936 return 0; 937 } 938 939 static long gntdev_ioctl_grant_copy(struct gntdev_priv *priv, void __user *u) 940 { 941 struct ioctl_gntdev_grant_copy copy; 942 struct gntdev_copy_batch batch; 943 unsigned int i; 944 int ret = 0; 945 946 if (copy_from_user(©, u, sizeof(copy))) 947 return -EFAULT; 948 949 batch.nr_ops = 0; 950 batch.nr_pages = 0; 951 952 for (i = 0; i < copy.count; i++) { 953 struct gntdev_grant_copy_segment seg; 954 955 if (copy_from_user(&seg, ©.segments[i], sizeof(seg))) { 956 ret = -EFAULT; 957 goto out; 958 } 959 960 ret = gntdev_grant_copy_seg(&batch, &seg, ©.segments[i].status); 961 if (ret < 0) 962 goto out; 963 964 cond_resched(); 965 } 966 if (batch.nr_ops) 967 ret = gntdev_copy(&batch); 968 return ret; 969 970 out: 971 gntdev_put_pages(&batch); 972 return ret; 973 } 974 975 static long gntdev_ioctl(struct file *flip, 976 unsigned int cmd, unsigned long arg) 977 { 978 struct gntdev_priv *priv = flip->private_data; 979 void __user *ptr = (void __user *)arg; 980 981 switch (cmd) { 982 case IOCTL_GNTDEV_MAP_GRANT_REF: 983 return gntdev_ioctl_map_grant_ref(priv, ptr); 984 985 case IOCTL_GNTDEV_UNMAP_GRANT_REF: 986 return gntdev_ioctl_unmap_grant_ref(priv, ptr); 987 988 case IOCTL_GNTDEV_GET_OFFSET_FOR_VADDR: 989 return gntdev_ioctl_get_offset_for_vaddr(priv, ptr); 990 991 case IOCTL_GNTDEV_SET_UNMAP_NOTIFY: 992 return gntdev_ioctl_notify(priv, ptr); 993 994 case IOCTL_GNTDEV_GRANT_COPY: 995 return gntdev_ioctl_grant_copy(priv, ptr); 996 997 #ifdef CONFIG_XEN_GNTDEV_DMABUF 998 case IOCTL_GNTDEV_DMABUF_EXP_FROM_REFS: 999 return gntdev_ioctl_dmabuf_exp_from_refs(priv, use_ptemod, ptr); 1000 1001 case IOCTL_GNTDEV_DMABUF_EXP_WAIT_RELEASED: 1002 return gntdev_ioctl_dmabuf_exp_wait_released(priv, ptr); 1003 1004 case IOCTL_GNTDEV_DMABUF_IMP_TO_REFS: 1005 return gntdev_ioctl_dmabuf_imp_to_refs(priv, ptr); 1006 1007 case IOCTL_GNTDEV_DMABUF_IMP_RELEASE: 1008 return gntdev_ioctl_dmabuf_imp_release(priv, ptr); 1009 #endif 1010 1011 default: 1012 pr_debug("priv %p, unknown cmd %x\n", priv, cmd); 1013 return -ENOIOCTLCMD; 1014 } 1015 1016 return 0; 1017 } 1018 1019 static int gntdev_mmap(struct file *flip, struct vm_area_struct *vma) 1020 { 1021 struct gntdev_priv *priv = flip->private_data; 1022 int index = vma->vm_pgoff; 1023 int count = vma_pages(vma); 1024 struct gntdev_grant_map *map; 1025 int err = -EINVAL; 1026 1027 if ((vma->vm_flags & VM_WRITE) && !(vma->vm_flags & VM_SHARED)) 1028 return -EINVAL; 1029 1030 pr_debug("map %d+%d at %lx (pgoff %lx)\n", 1031 index, count, vma->vm_start, vma->vm_pgoff); 1032 1033 mutex_lock(&priv->lock); 1034 map = gntdev_find_map_index(priv, index, count); 1035 if (!map) 1036 goto unlock_out; 1037 if (use_ptemod && map->vma) 1038 goto unlock_out; 1039 if (atomic_read(&map->live_grants)) { 1040 err = -EAGAIN; 1041 goto unlock_out; 1042 } 1043 refcount_inc(&map->users); 1044 1045 vma->vm_ops = &gntdev_vmops; 1046 1047 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP | VM_MIXEDMAP; 1048 1049 if (use_ptemod) 1050 vma->vm_flags |= VM_DONTCOPY; 1051 1052 vma->vm_private_data = map; 1053 if (map->flags) { 1054 if ((vma->vm_flags & VM_WRITE) && 1055 (map->flags & GNTMAP_readonly)) 1056 goto out_unlock_put; 1057 } else { 1058 map->flags = GNTMAP_host_map; 1059 if (!(vma->vm_flags & VM_WRITE)) 1060 map->flags |= GNTMAP_readonly; 1061 } 1062 1063 if (use_ptemod) { 1064 map->vma = vma; 1065 err = mmu_interval_notifier_insert_locked( 1066 &map->notifier, vma->vm_mm, vma->vm_start, 1067 vma->vm_end - vma->vm_start, &gntdev_mmu_ops); 1068 if (err) { 1069 map->vma = NULL; 1070 goto out_unlock_put; 1071 } 1072 } 1073 mutex_unlock(&priv->lock); 1074 1075 if (use_ptemod) { 1076 /* 1077 * gntdev takes the address of the PTE in find_grant_ptes() and 1078 * passes it to the hypervisor in gntdev_map_grant_pages(). The 1079 * purpose of the notifier is to prevent the hypervisor pointer 1080 * to the PTE from going stale. 1081 * 1082 * Since this vma's mappings can't be touched without the 1083 * mmap_lock, and we are holding it now, there is no need for 1084 * the notifier_range locking pattern. 1085 */ 1086 mmu_interval_read_begin(&map->notifier); 1087 1088 map->pages_vm_start = vma->vm_start; 1089 err = apply_to_page_range(vma->vm_mm, vma->vm_start, 1090 vma->vm_end - vma->vm_start, 1091 find_grant_ptes, map); 1092 if (err) { 1093 pr_warn("find_grant_ptes() failure.\n"); 1094 goto out_put_map; 1095 } 1096 } 1097 1098 err = gntdev_map_grant_pages(map); 1099 if (err) 1100 goto out_put_map; 1101 1102 if (!use_ptemod) { 1103 err = vm_map_pages_zero(vma, map->pages, map->count); 1104 if (err) 1105 goto out_put_map; 1106 } 1107 1108 return 0; 1109 1110 unlock_out: 1111 mutex_unlock(&priv->lock); 1112 return err; 1113 1114 out_unlock_put: 1115 mutex_unlock(&priv->lock); 1116 out_put_map: 1117 if (use_ptemod) { 1118 unmap_grant_pages(map, 0, map->count); 1119 if (map->vma) { 1120 mmu_interval_notifier_remove(&map->notifier); 1121 map->vma = NULL; 1122 } 1123 } 1124 gntdev_put_map(priv, map); 1125 return err; 1126 } 1127 1128 static const struct file_operations gntdev_fops = { 1129 .owner = THIS_MODULE, 1130 .open = gntdev_open, 1131 .release = gntdev_release, 1132 .mmap = gntdev_mmap, 1133 .unlocked_ioctl = gntdev_ioctl 1134 }; 1135 1136 static struct miscdevice gntdev_miscdev = { 1137 .minor = MISC_DYNAMIC_MINOR, 1138 .name = "xen/gntdev", 1139 .fops = &gntdev_fops, 1140 }; 1141 1142 /* ------------------------------------------------------------------ */ 1143 1144 static int __init gntdev_init(void) 1145 { 1146 int err; 1147 1148 if (!xen_domain()) 1149 return -ENODEV; 1150 1151 use_ptemod = !xen_feature(XENFEAT_auto_translated_physmap); 1152 1153 err = misc_register(&gntdev_miscdev); 1154 if (err != 0) { 1155 pr_err("Could not register gntdev device\n"); 1156 return err; 1157 } 1158 return 0; 1159 } 1160 1161 static void __exit gntdev_exit(void) 1162 { 1163 misc_deregister(&gntdev_miscdev); 1164 } 1165 1166 module_init(gntdev_init); 1167 module_exit(gntdev_exit); 1168 1169 /* ------------------------------------------------------------------ */ 1170