1 /************************************************************************** 2 * 3 * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA 4 * All Rights Reserved. 5 * 6 * Permission is hereby granted, free of charge, to any person obtaining a 7 * copy of this software and associated documentation files (the 8 * "Software"), to deal in the Software without restriction, including 9 * without limitation the rights to use, copy, modify, merge, publish, 10 * distribute, sub license, and/or sell copies of the Software, and to 11 * permit persons to whom the Software is furnished to do so, subject to 12 * the following conditions: 13 * 14 * The above copyright notice and this permission notice (including the 15 * next paragraph) shall be included in all copies or substantial portions 16 * of the Software. 17 * 18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 19 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 20 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL 21 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, 22 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR 23 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE 24 * USE OR OTHER DEALINGS IN THE SOFTWARE. 25 * 26 **************************************************************************/ 27 /* 28 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com> 29 */ 30 /* 31 * Copyright (c) 2013 The FreeBSD Foundation 32 * All rights reserved. 33 * 34 * Portions of this software were developed by Konstantin Belousov 35 * <kib@FreeBSD.org> under sponsorship from the FreeBSD Foundation. 36 * 37 *$FreeBSD: head/sys/dev/drm2/ttm/ttm_bo_vm.c 253710 2013-07-27 16:44:37Z kib $ 38 */ 39 40 #include "opt_vm.h" 41 42 #define pr_fmt(fmt) "[TTM] " fmt 43 44 #include <drm/ttm/ttm_module.h> 45 #include <drm/ttm/ttm_bo_driver.h> 46 #include <drm/ttm/ttm_placement.h> 47 #include <vm/vm.h> 48 #include <vm/vm_page.h> 49 #include <linux/errno.h> 50 #include <linux/export.h> 51 #include <linux/rbtree.h> 52 53 #define TTM_BO_VM_NUM_PREFAULT 16 54 55 static struct ttm_buffer_object *ttm_bo_vm_lookup_rb(struct ttm_bo_device *bdev, 56 unsigned long page_start, 57 unsigned long num_pages) 58 { 59 struct rb_node *cur = bdev->addr_space_rb.rb_node; 60 unsigned long cur_offset; 61 struct ttm_buffer_object *bo; 62 struct ttm_buffer_object *best_bo = NULL; 63 64 while (likely(cur != NULL)) { 65 bo = rb_entry(cur, struct ttm_buffer_object, vm_rb); 66 cur_offset = bo->vm_node->start; 67 if (page_start >= cur_offset) { 68 cur = cur->rb_right; 69 best_bo = bo; 70 if (page_start == cur_offset) 71 break; 72 } else 73 cur = cur->rb_left; 74 } 75 76 if (unlikely(best_bo == NULL)) 77 return NULL; 78 79 if (unlikely((best_bo->vm_node->start + best_bo->num_pages) < 80 (page_start + num_pages))) 81 return NULL; 82 83 return best_bo; 84 } 85 86 static int 87 ttm_bo_vm_fault(vm_object_t vm_obj, vm_ooffset_t offset, 88 int prot, vm_page_t *mres) 89 { 90 struct ttm_buffer_object *bo = vm_obj->handle; 91 struct ttm_bo_device *bdev = bo->bdev; 92 struct ttm_tt *ttm = NULL; 93 vm_page_t m, m1, oldm; 94 int ret; 95 int retval = VM_PAGER_OK; 96 struct ttm_mem_type_manager *man = 97 &bdev->man[bo->mem.mem_type]; 98 99 vm_object_pip_add(vm_obj, 1); 100 oldm = *mres; 101 if (oldm != NULL) { 102 vm_page_remove(oldm); 103 *mres = NULL; 104 } else 105 oldm = NULL; 106 retry: 107 VM_OBJECT_WUNLOCK(vm_obj); 108 m = NULL; 109 110 reserve: 111 ret = ttm_bo_reserve(bo, false, false, false, 0); 112 if (unlikely(ret != 0)) { 113 if (ret == -EBUSY) { 114 lwkt_yield(); 115 goto reserve; 116 } 117 } 118 119 if (bdev->driver->fault_reserve_notify) { 120 ret = bdev->driver->fault_reserve_notify(bo); 121 switch (ret) { 122 case 0: 123 break; 124 case -EBUSY: 125 case -ERESTARTSYS: 126 case -EINTR: 127 lwkt_yield(); 128 goto reserve; 129 default: 130 retval = VM_PAGER_ERROR; 131 goto out_unlock; 132 } 133 } 134 135 /* 136 * Wait for buffer data in transit, due to a pipelined 137 * move. 138 */ 139 140 lockmgr(&bdev->fence_lock, LK_EXCLUSIVE); 141 if (test_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags)) { 142 /* 143 * Here, the behavior differs between Linux and FreeBSD. 144 * 145 * On Linux, the wait is interruptible (3rd argument to 146 * ttm_bo_wait). There must be some mechanism to resume 147 * page fault handling, once the signal is processed. 148 * 149 * On FreeBSD, the wait is uninteruptible. This is not a 150 * problem as we can't end up with an unkillable process 151 * here, because the wait will eventually time out. 152 * 153 * An example of this situation is the Xorg process 154 * which uses SIGALRM internally. The signal could 155 * interrupt the wait, causing the page fault to fail 156 * and the process to receive SIGSEGV. 157 */ 158 ret = ttm_bo_wait(bo, false, false, false); 159 lockmgr(&bdev->fence_lock, LK_RELEASE); 160 if (unlikely(ret != 0)) { 161 retval = VM_PAGER_ERROR; 162 goto out_unlock; 163 } 164 } else 165 lockmgr(&bdev->fence_lock, LK_RELEASE); 166 167 ret = ttm_mem_io_lock(man, true); 168 if (unlikely(ret != 0)) { 169 retval = VM_PAGER_ERROR; 170 goto out_unlock; 171 } 172 ret = ttm_mem_io_reserve_vm(bo); 173 if (unlikely(ret != 0)) { 174 retval = VM_PAGER_ERROR; 175 goto out_io_unlock; 176 } 177 178 /* 179 * Strictly, we're not allowed to modify vma->vm_page_prot here, 180 * since the mmap_sem is only held in read mode. However, we 181 * modify only the caching bits of vma->vm_page_prot and 182 * consider those bits protected by 183 * the bo->mutex, as we should be the only writers. 184 * There shouldn't really be any readers of these bits except 185 * within vm_insert_mixed()? fork? 186 * 187 * TODO: Add a list of vmas to the bo, and change the 188 * vma->vm_page_prot when the object changes caching policy, with 189 * the correct locks held. 190 */ 191 if (!bo->mem.bus.is_iomem) { 192 /* Allocate all page at once, most common usage */ 193 ttm = bo->ttm; 194 if (ttm->bdev->driver->ttm_tt_populate(ttm)) { 195 retval = VM_PAGER_ERROR; 196 goto out_io_unlock; 197 } 198 } 199 200 if (bo->mem.bus.is_iomem) { 201 m = vm_phys_fictitious_to_vm_page(bo->mem.bus.base + 202 bo->mem.bus.offset + offset); 203 pmap_page_set_memattr(m, ttm_io_prot(bo->mem.placement)); 204 } else { 205 ttm = bo->ttm; 206 m = ttm->pages[OFF_TO_IDX(offset)]; 207 if (unlikely(!m)) { 208 retval = VM_PAGER_ERROR; 209 goto out_io_unlock; 210 } 211 pmap_page_set_memattr(m, 212 (bo->mem.placement & TTM_PL_FLAG_CACHED) ? 213 VM_MEMATTR_WRITE_BACK : ttm_io_prot(bo->mem.placement)); 214 } 215 216 VM_OBJECT_WLOCK(vm_obj); 217 if ((m->flags & PG_BUSY) != 0) { 218 #if 0 219 vm_page_sleep(m, "ttmpbs"); 220 #endif 221 ttm_mem_io_unlock(man); 222 ttm_bo_unreserve(bo); 223 goto retry; 224 } 225 m->valid = VM_PAGE_BITS_ALL; 226 *mres = m; 227 m1 = vm_page_lookup(vm_obj, OFF_TO_IDX(offset)); 228 if (m1 == NULL) { 229 vm_page_insert(m, vm_obj, OFF_TO_IDX(offset)); 230 } else { 231 KASSERT(m == m1, 232 ("inconsistent insert bo %p m %p m1 %p offset %jx", 233 bo, m, m1, (uintmax_t)offset)); 234 } 235 vm_page_busy_try(m, FALSE); 236 237 if (oldm != NULL) { 238 vm_page_free(oldm); 239 } 240 241 out_io_unlock1: 242 ttm_mem_io_unlock(man); 243 out_unlock1: 244 ttm_bo_unreserve(bo); 245 vm_object_pip_wakeup(vm_obj); 246 return (retval); 247 248 out_io_unlock: 249 VM_OBJECT_WLOCK(vm_obj); 250 goto out_io_unlock1; 251 252 out_unlock: 253 VM_OBJECT_WLOCK(vm_obj); 254 goto out_unlock1; 255 } 256 257 static int 258 ttm_bo_vm_ctor(void *handle, vm_ooffset_t size, vm_prot_t prot, 259 vm_ooffset_t foff, struct ucred *cred, u_short *color) 260 { 261 262 /* 263 * On Linux, a reference to the buffer object is acquired here. 264 * The reason is that this function is not called when the 265 * mmap() is initialized, but only when a process forks for 266 * instance. Therefore on Linux, the reference on the bo is 267 * acquired either in ttm_bo_mmap() or ttm_bo_vm_open(). It's 268 * then released in ttm_bo_vm_close(). 269 * 270 * Here, this function is called during mmap() intialization. 271 * Thus, the reference acquired in ttm_bo_mmap_single() is 272 * sufficient. 273 */ 274 *color = 0; 275 return (0); 276 } 277 278 static void 279 ttm_bo_vm_dtor(void *handle) 280 { 281 struct ttm_buffer_object *bo = handle; 282 283 ttm_bo_unref(&bo); 284 } 285 286 static struct cdev_pager_ops ttm_pager_ops = { 287 .cdev_pg_fault = ttm_bo_vm_fault, 288 .cdev_pg_ctor = ttm_bo_vm_ctor, 289 .cdev_pg_dtor = ttm_bo_vm_dtor 290 }; 291 292 int 293 ttm_bo_mmap_single(struct ttm_bo_device *bdev, vm_ooffset_t *offset, vm_size_t size, 294 struct vm_object **obj_res, int nprot) 295 { 296 struct ttm_bo_driver *driver; 297 struct ttm_buffer_object *bo; 298 struct vm_object *vm_obj; 299 int ret; 300 301 *obj_res = NULL; 302 303 lockmgr(&bdev->vm_lock, LK_EXCLUSIVE); 304 bo = ttm_bo_vm_lookup_rb(bdev, OFF_TO_IDX(*offset), OFF_TO_IDX(size)); 305 if (likely(bo != NULL)) 306 kref_get(&bo->kref); 307 lockmgr(&bdev->vm_lock, LK_RELEASE); 308 309 if (unlikely(bo == NULL)) { 310 kprintf("[TTM] Could not find buffer object to map\n"); 311 return (EINVAL); 312 } 313 314 driver = bo->bdev->driver; 315 if (unlikely(!driver->verify_access)) { 316 ret = EPERM; 317 goto out_unref; 318 } 319 ret = -driver->verify_access(bo); 320 if (unlikely(ret != 0)) 321 goto out_unref; 322 323 vm_obj = cdev_pager_allocate(bo, OBJT_MGTDEVICE, &ttm_pager_ops, 324 size, nprot, 0, curthread->td_ucred); 325 326 if (vm_obj == NULL) { 327 ret = EINVAL; 328 goto out_unref; 329 } 330 /* 331 * Note: We're transferring the bo reference to vm_obj->handle here. 332 */ 333 *offset = 0; 334 *obj_res = vm_obj; 335 return 0; 336 out_unref: 337 ttm_bo_unref(&bo); 338 return ret; 339 } 340 EXPORT_SYMBOL(ttm_bo_mmap); 341 342 void 343 ttm_bo_release_mmap(struct ttm_buffer_object *bo) 344 { 345 vm_object_t vm_obj; 346 vm_page_t m; 347 int i; 348 349 vm_obj = cdev_pager_lookup(bo); 350 if (vm_obj == NULL) 351 return; 352 353 VM_OBJECT_WLOCK(vm_obj); 354 for (i = 0; i < bo->num_pages; i++) { 355 m = vm_page_lookup_busy_wait(vm_obj, i, TRUE, "ttm_unm"); 356 if (m == NULL) 357 continue; 358 cdev_pager_free_page(vm_obj, m); 359 } 360 VM_OBJECT_WUNLOCK(vm_obj); 361 362 vm_object_deallocate(vm_obj); 363 } 364 365 #if 0 366 int ttm_fbdev_mmap(struct vm_area_struct *vma, struct ttm_buffer_object *bo) 367 { 368 if (vma->vm_pgoff != 0) 369 return -EACCES; 370 371 vma->vm_ops = &ttm_bo_vm_ops; 372 vma->vm_private_data = ttm_bo_reference(bo); 373 vma->vm_flags |= VM_IO | VM_MIXEDMAP | VM_DONTEXPAND; 374 return 0; 375 } 376 EXPORT_SYMBOL(ttm_fbdev_mmap); 377 378 379 ssize_t ttm_bo_io(struct ttm_bo_device *bdev, struct file *filp, 380 const char __user *wbuf, char __user *rbuf, size_t count, 381 loff_t *f_pos, bool write) 382 { 383 struct ttm_buffer_object *bo; 384 struct ttm_bo_driver *driver; 385 struct ttm_bo_kmap_obj map; 386 unsigned long dev_offset = (*f_pos >> PAGE_SHIFT); 387 unsigned long kmap_offset; 388 unsigned long kmap_end; 389 unsigned long kmap_num; 390 size_t io_size; 391 unsigned int page_offset; 392 char *virtual; 393 int ret; 394 bool no_wait = false; 395 bool dummy; 396 397 lockmgr(&bdev->vm_lock, LK_EXCLUSIVE); 398 bo = ttm_bo_vm_lookup_rb(bdev, dev_offset, 1); 399 if (likely(bo != NULL)) 400 ttm_bo_reference(bo); 401 lockmgr(&bdev->vm_lock, LK_RELEASE); 402 403 if (unlikely(bo == NULL)) 404 return -EFAULT; 405 406 driver = bo->bdev->driver; 407 if (unlikely(!driver->verify_access)) { 408 ret = -EPERM; 409 goto out_unref; 410 } 411 412 ret = driver->verify_access(bo, filp); 413 if (unlikely(ret != 0)) 414 goto out_unref; 415 416 kmap_offset = dev_offset - bo->vm_node->start; 417 if (unlikely(kmap_offset >= bo->num_pages)) { 418 ret = -EFBIG; 419 goto out_unref; 420 } 421 422 page_offset = *f_pos & ~PAGE_MASK; 423 io_size = bo->num_pages - kmap_offset; 424 io_size = (io_size << PAGE_SHIFT) - page_offset; 425 if (count < io_size) 426 io_size = count; 427 428 kmap_end = (*f_pos + count - 1) >> PAGE_SHIFT; 429 kmap_num = kmap_end - kmap_offset + 1; 430 431 ret = ttm_bo_reserve(bo, true, no_wait, false, 0); 432 433 switch (ret) { 434 case 0: 435 break; 436 case -EBUSY: 437 ret = -EAGAIN; 438 goto out_unref; 439 default: 440 goto out_unref; 441 } 442 443 ret = ttm_bo_kmap(bo, kmap_offset, kmap_num, &map); 444 if (unlikely(ret != 0)) { 445 ttm_bo_unreserve(bo); 446 goto out_unref; 447 } 448 449 virtual = ttm_kmap_obj_virtual(&map, &dummy); 450 virtual += page_offset; 451 452 if (write) 453 ret = copy_from_user(virtual, wbuf, io_size); 454 else 455 ret = copy_to_user(rbuf, virtual, io_size); 456 457 ttm_bo_kunmap(&map); 458 ttm_bo_unreserve(bo); 459 ttm_bo_unref(&bo); 460 461 if (unlikely(ret != 0)) 462 return -EFBIG; 463 464 *f_pos += io_size; 465 466 return io_size; 467 out_unref: 468 ttm_bo_unref(&bo); 469 return ret; 470 } 471 472 ssize_t ttm_bo_fbdev_io(struct ttm_buffer_object *bo, const char __user *wbuf, 473 char __user *rbuf, size_t count, loff_t *f_pos, 474 bool write) 475 { 476 struct ttm_bo_kmap_obj map; 477 unsigned long kmap_offset; 478 unsigned long kmap_end; 479 unsigned long kmap_num; 480 size_t io_size; 481 unsigned int page_offset; 482 char *virtual; 483 int ret; 484 bool no_wait = false; 485 bool dummy; 486 487 kmap_offset = (*f_pos >> PAGE_SHIFT); 488 if (unlikely(kmap_offset >= bo->num_pages)) 489 return -EFBIG; 490 491 page_offset = *f_pos & ~PAGE_MASK; 492 io_size = bo->num_pages - kmap_offset; 493 io_size = (io_size << PAGE_SHIFT) - page_offset; 494 if (count < io_size) 495 io_size = count; 496 497 kmap_end = (*f_pos + count - 1) >> PAGE_SHIFT; 498 kmap_num = kmap_end - kmap_offset + 1; 499 500 ret = ttm_bo_reserve(bo, true, no_wait, false, 0); 501 502 switch (ret) { 503 case 0: 504 break; 505 case -EBUSY: 506 return -EAGAIN; 507 default: 508 return ret; 509 } 510 511 ret = ttm_bo_kmap(bo, kmap_offset, kmap_num, &map); 512 if (unlikely(ret != 0)) { 513 ttm_bo_unreserve(bo); 514 return ret; 515 } 516 517 virtual = ttm_kmap_obj_virtual(&map, &dummy); 518 virtual += page_offset; 519 520 if (write) 521 ret = copy_from_user(virtual, wbuf, io_size); 522 else 523 ret = copy_to_user(rbuf, virtual, io_size); 524 525 ttm_bo_kunmap(&map); 526 ttm_bo_unreserve(bo); 527 ttm_bo_unref(&bo); 528 529 if (unlikely(ret != 0)) 530 return ret; 531 532 *f_pos += io_size; 533 534 return io_size; 535 } 536 #endif 537