sys/dev/drm2/ttm/ttm_bo_vm.c

   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
  38 #include <sys/cdefs.h>
  39 __FBSDID("$FreeBSD$");
  40
  41 #include "opt_vm.h"
  42
  43 #include <dev/drm2/drmP.h>
  44 #include <dev/drm2/ttm/ttm_module.h>
  45 #include <dev/drm2/ttm/ttm_bo_driver.h>
  46 #include <dev/drm2/ttm/ttm_placement.h>
  47
  48 #include <vm/vm.h>
  49 #include <vm/vm_page.h>
  50 #include <vm/vm_pageout.h>
  51
  52 #define TTM_BO_VM_NUM_PREFAULT 16
  53
  54 RB_GENERATE(ttm_bo_device_buffer_objects, ttm_buffer_object, vm_rb,
  55     ttm_bo_cmp_rb_tree_items);
  56
  57 int
  58 ttm_bo_cmp_rb_tree_items(struct ttm_buffer_object *a,
  59     struct ttm_buffer_object *b)
  60 {
  61
  62         if (a->vm_node->start < b->vm_node->start) {
  63                 return (-1);
  64         } else if (a->vm_node->start > b->vm_node->start) {
  65                 return (1);
  66         } else {
  67                 return (0);
  68         }
  69 }
  70
  71 static struct ttm_buffer_object *ttm_bo_vm_lookup_rb(struct ttm_bo_device *bdev,
  72                                                      unsigned long page_start,
  73                                                      unsigned long num_pages)
  74 {
  75         unsigned long cur_offset;
  76         struct ttm_buffer_object *bo;
  77         struct ttm_buffer_object *best_bo = NULL;
  78
  79         bo = RB_ROOT(&bdev->addr_space_rb);
  80         while (bo != NULL) {
  81                 cur_offset = bo->vm_node->start;
  82                 if (page_start >= cur_offset) {
  83                         best_bo = bo;
  84                         if (page_start == cur_offset)
  85                                 break;
  86                         bo = RB_RIGHT(bo, vm_rb);
  87                 } else
  88                         bo = RB_LEFT(bo, vm_rb);
  89         }
  90
  91         if (unlikely(best_bo == NULL))
  92                 return NULL;
  93
  94         if (unlikely((best_bo->vm_node->start + best_bo->num_pages) <
  95                      (page_start + num_pages)))
  96                 return NULL;
  97
  98         return best_bo;
  99 }
 100
 101 static int
 102 ttm_bo_vm_fault(vm_object_t vm_obj, vm_ooffset_t offset,
 103     int prot, vm_page_t *mres)
 104 {
 105
 106         struct ttm_buffer_object *bo = vm_obj->handle;
 107         struct ttm_bo_device *bdev = bo->bdev;
 108         struct ttm_tt *ttm = NULL;
 109         vm_page_t m, m1;
 110         int ret;
 111         int retval = VM_PAGER_OK;
 112         struct ttm_mem_type_manager *man =
 113                 &bdev->man[bo->mem.mem_type];
 114
 115         vm_object_pip_add(vm_obj, 1);
 116         if (*mres != NULL) {
 117                 (void)vm_page_remove(*mres);
 118         }
 119 retry:
 120         VM_OBJECT_WUNLOCK(vm_obj);
 121         m = NULL;
 122
 123 reserve:
 124         ret = ttm_bo_reserve(bo, false, false, false, 0);
 125         if (unlikely(ret != 0)) {
 126                 if (ret == -EBUSY) {
 127                         kern_yield(PRI_USER);
 128                         goto reserve;
 129                 }
 130         }
 131
 132         if (bdev->driver->fault_reserve_notify) {
 133                 ret = bdev->driver->fault_reserve_notify(bo);
 134                 switch (ret) {
 135                 case 0:
 136                         break;
 137                 case -EBUSY:
 138                 case -ERESTARTSYS:
 139                 case -EINTR:
 140                         kern_yield(PRI_USER);
 141                         goto reserve;
 142                 default:
 143                         retval = VM_PAGER_ERROR;
 144                         goto out_unlock;
 145                 }
 146         }
 147
 148         /*
 149          * Wait for buffer data in transit, due to a pipelined
 150          * move.
 151          */
 152
 153         mtx_lock(&bdev->fence_lock);
 154         if (test_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags)) {
 155                 /*
 156                  * Here, the behavior differs between Linux and FreeBSD.
 157                  *
 158                  * On Linux, the wait is interruptible (3rd argument to
 159                  * ttm_bo_wait). There must be some mechanism to resume
 160                  * page fault handling, once the signal is processed.
 161                  *
 162                  * On FreeBSD, the wait is uninteruptible. This is not a
 163                  * problem as we can't end up with an unkillable process
 164                  * here, because the wait will eventually time out.
 165                  *
 166                  * An example of this situation is the Xorg process
 167                  * which uses SIGALRM internally. The signal could
 168                  * interrupt the wait, causing the page fault to fail
 169                  * and the process to receive SIGSEGV.
 170                  */
 171                 ret = ttm_bo_wait(bo, false, false, false);
 172                 mtx_unlock(&bdev->fence_lock);
 173                 if (unlikely(ret != 0)) {
 174                         retval = VM_PAGER_ERROR;
 175                         goto out_unlock;
 176                 }
 177         } else
 178                 mtx_unlock(&bdev->fence_lock);
 179
 180         ret = ttm_mem_io_lock(man, true);
 181         if (unlikely(ret != 0)) {
 182                 retval = VM_PAGER_ERROR;
 183                 goto out_unlock;
 184         }
 185         ret = ttm_mem_io_reserve_vm(bo);
 186         if (unlikely(ret != 0)) {
 187                 retval = VM_PAGER_ERROR;
 188                 goto out_io_unlock;
 189         }
 190
 191         /*
 192          * Strictly, we're not allowed to modify vma->vm_page_prot here,
 193          * since the mmap_sem is only held in read mode. However, we
 194          * modify only the caching bits of vma->vm_page_prot and
 195          * consider those bits protected by
 196          * the bo->mutex, as we should be the only writers.
 197          * There shouldn't really be any readers of these bits except
 198          * within vm_insert_mixed()? fork?
 199          *
 200          * TODO: Add a list of vmas to the bo, and change the
 201          * vma->vm_page_prot when the object changes caching policy, with
 202          * the correct locks held.
 203          */
 204         if (!bo->mem.bus.is_iomem) {
 205                 /* Allocate all page at once, most common usage */
 206                 ttm = bo->ttm;
 207                 if (ttm->bdev->driver->ttm_tt_populate(ttm)) {
 208                         retval = VM_PAGER_ERROR;
 209                         goto out_io_unlock;
 210                 }
 211         }
 212
 213         if (bo->mem.bus.is_iomem) {
 214                 m = PHYS_TO_VM_PAGE(bo->mem.bus.base + bo->mem.bus.offset +
 215                     offset);
 216                 KASSERT((m->flags & PG_FICTITIOUS) != 0,
 217                     ("physical address %#jx not fictitious",
 218                     (uintmax_t)(bo->mem.bus.base + bo->mem.bus.offset
 219                     + offset)));
 220                 pmap_page_set_memattr(m, ttm_io_prot(bo->mem.placement));
 221         } else {
 222                 ttm = bo->ttm;
 223                 m = ttm->pages[OFF_TO_IDX(offset)];
 224                 if (unlikely(!m)) {
 225                         retval = VM_PAGER_ERROR;
 226                         goto out_io_unlock;
 227                 }
 228                 pmap_page_set_memattr(m,
 229                     (bo->mem.placement & TTM_PL_FLAG_CACHED) ?
 230                     VM_MEMATTR_WRITE_BACK : ttm_io_prot(bo->mem.placement));
 231         }
 232
 233         VM_OBJECT_WLOCK(vm_obj);
 234         if (vm_page_busy_acquire(m, VM_ALLOC_WAITFAIL) == 0) {
 235                 ttm_mem_io_unlock(man);
 236                 ttm_bo_unreserve(bo);
 237                 goto retry;
 238         }
 239         m1 = vm_page_lookup(vm_obj, OFF_TO_IDX(offset));
 240         /* XXX This looks like it should just be vm_page_replace? */
 241         if (m1 == NULL) {
 242                 if (vm_page_insert(m, vm_obj, OFF_TO_IDX(offset))) {
 243                         vm_page_xunbusy(m);
 244                         VM_OBJECT_WUNLOCK(vm_obj);
 245                         vm_wait(vm_obj);
 246                         VM_OBJECT_WLOCK(vm_obj);
 247                         ttm_mem_io_unlock(man);
 248                         ttm_bo_unreserve(bo);
 249                         goto retry;
 250                 }
 251         } else {
 252                 KASSERT(m == m1,
 253                     ("inconsistent insert bo %p m %p m1 %p offset %jx",
 254                     bo, m, m1, (uintmax_t)offset));
 255         }
 256         vm_page_valid(m);
 257         if (*mres != NULL) {
 258                 KASSERT(*mres != m, ("losing %p %p", *mres, m));
 259                 vm_page_xunbusy(*mres);
 260                 vm_page_free(*mres);
 261         }
 262         *mres = m;
 263
 264 out_io_unlock1:
 265         ttm_mem_io_unlock(man);
 266 out_unlock1:
 267         ttm_bo_unreserve(bo);
 268         vm_object_pip_wakeup(vm_obj);
 269         return (retval);
 270
 271 out_io_unlock:
 272         VM_OBJECT_WLOCK(vm_obj);
 273         goto out_io_unlock1;
 274
 275 out_unlock:
 276         VM_OBJECT_WLOCK(vm_obj);
 277         goto out_unlock1;
 278 }
 279
 280 static int
 281 ttm_bo_vm_ctor(void *handle, vm_ooffset_t size, vm_prot_t prot,
 282     vm_ooffset_t foff, struct ucred *cred, u_short *color)
 283 {
 284
 285         /*
 286          * On Linux, a reference to the buffer object is acquired here.
 287          * The reason is that this function is not called when the
 288          * mmap() is initialized, but only when a process forks for
 289          * instance. Therefore on Linux, the reference on the bo is
 290          * acquired either in ttm_bo_mmap() or ttm_bo_vm_open(). It's
 291          * then released in ttm_bo_vm_close().
 292          *
 293          * Here, this function is called during mmap() initialization.
 294          * Thus, the reference acquired in ttm_bo_mmap_single() is
 295          * sufficient.
 296          */
 297
 298         *color = 0;
 299         return (0);
 300 }
 301
 302 static void
 303 ttm_bo_vm_dtor(void *handle)
 304 {
 305         struct ttm_buffer_object *bo = handle;
 306
 307         ttm_bo_unref(&bo);
 308 }
 309
 310 static struct cdev_pager_ops ttm_pager_ops = {
 311         .cdev_pg_fault = ttm_bo_vm_fault,
 312         .cdev_pg_ctor = ttm_bo_vm_ctor,
 313         .cdev_pg_dtor = ttm_bo_vm_dtor
 314 };
 315
 316 int
 317 ttm_bo_mmap_single(struct ttm_bo_device *bdev, vm_ooffset_t *offset, vm_size_t size,
 318     struct vm_object **obj_res, int nprot)
 319 {
 320         struct ttm_bo_driver *driver;
 321         struct ttm_buffer_object *bo;
 322         struct vm_object *vm_obj;
 323         int ret;
 324
 325         rw_wlock(&bdev->vm_lock);
 326         bo = ttm_bo_vm_lookup_rb(bdev, OFF_TO_IDX(*offset), OFF_TO_IDX(size));
 327         if (likely(bo != NULL))
 328                 refcount_acquire(&bo->kref);
 329         rw_wunlock(&bdev->vm_lock);
 330
 331         if (unlikely(bo == NULL)) {
 332                 printf("[TTM] Could not find buffer object to map\n");
 333                 return (-EINVAL);
 334         }
 335
 336         driver = bo->bdev->driver;
 337         if (unlikely(!driver->verify_access)) {
 338                 ret = -EPERM;
 339                 goto out_unref;
 340         }
 341         ret = driver->verify_access(bo);
 342         if (unlikely(ret != 0))
 343                 goto out_unref;
 344
 345         vm_obj = cdev_pager_allocate(bo, OBJT_MGTDEVICE, &ttm_pager_ops,
 346             size, nprot, 0, curthread->td_ucred);
 347         if (vm_obj == NULL) {
 348                 ret = -EINVAL;
 349                 goto out_unref;
 350         }
 351         /*
 352          * Note: We're transferring the bo reference to vm_obj->handle here.
 353          */
 354         *offset = 0;
 355         *obj_res = vm_obj;
 356         return 0;
 357 out_unref:
 358         ttm_bo_unref(&bo);
 359         return ret;
 360 }
 361
 362 void
 363 ttm_bo_release_mmap(struct ttm_buffer_object *bo)
 364 {
 365         vm_object_t vm_obj;
 366         vm_page_t m;
 367         int i;
 368
 369         vm_obj = cdev_pager_lookup(bo);
 370         if (vm_obj == NULL)
 371                 return;
 372
 373         VM_OBJECT_WLOCK(vm_obj);
 374 retry:
 375         for (i = 0; i < bo->num_pages; i++) {
 376                 m = vm_page_lookup(vm_obj, i);
 377                 if (m == NULL)
 378                         continue;
 379                 if (vm_page_busy_acquire(m, VM_ALLOC_WAITFAIL) == 0)
 380                         goto retry;
 381                 cdev_pager_free_page(vm_obj, m);
 382         }
 383         VM_OBJECT_WUNLOCK(vm_obj);
 384
 385         vm_object_deallocate(vm_obj);
 386 }
 387
 388 #if 0
 389 int ttm_fbdev_mmap(struct vm_area_struct *vma, struct ttm_buffer_object *bo)
 390 {
 391         if (vma->vm_pgoff != 0)
 392                 return -EACCES;
 393
 394         vma->vm_ops = &ttm_bo_vm_ops;
 395         vma->vm_private_data = ttm_bo_reference(bo);
 396         vma->vm_flags |= VM_IO | VM_MIXEDMAP | VM_DONTEXPAND;
 397         return 0;
 398 }
 399
 400 ssize_t ttm_bo_io(struct ttm_bo_device *bdev, struct file *filp,
 401                   const char __user *wbuf, char __user *rbuf, size_t count,
 402                   loff_t *f_pos, bool write)
 403 {
 404         struct ttm_buffer_object *bo;
 405         struct ttm_bo_driver *driver;
 406         struct ttm_bo_kmap_obj map;
 407         unsigned long dev_offset = (*f_pos >> PAGE_SHIFT);
 408         unsigned long kmap_offset;
 409         unsigned long kmap_end;
 410         unsigned long kmap_num;
 411         size_t io_size;
 412         unsigned int page_offset;
 413         char *virtual;
 414         int ret;
 415         bool no_wait = false;
 416         bool dummy;
 417
 418         read_lock(&bdev->vm_lock);
 419         bo = ttm_bo_vm_lookup_rb(bdev, dev_offset, 1);
 420         if (likely(bo != NULL))
 421                 ttm_bo_reference(bo);
 422         read_unlock(&bdev->vm_lock);
 423
 424         if (unlikely(bo == NULL))
 425                 return -EFAULT;
 426
 427         driver = bo->bdev->driver;
 428         if (unlikely(!driver->verify_access)) {
 429                 ret = -EPERM;
 430                 goto out_unref;
 431         }
 432
 433         ret = driver->verify_access(bo, filp);
 434         if (unlikely(ret != 0))
 435                 goto out_unref;
 436
 437         kmap_offset = dev_offset - bo->vm_node->start;
 438         if (unlikely(kmap_offset >= bo->num_pages)) {
 439                 ret = -EFBIG;
 440                 goto out_unref;
 441         }
 442
 443         page_offset = *f_pos & ~PAGE_MASK;
 444         io_size = bo->num_pages - kmap_offset;
 445         io_size = (io_size << PAGE_SHIFT) - page_offset;
 446         if (count < io_size)
 447                 io_size = count;
 448
 449         kmap_end = (*f_pos + count - 1) >> PAGE_SHIFT;
 450         kmap_num = kmap_end - kmap_offset + 1;
 451
 452         ret = ttm_bo_reserve(bo, true, no_wait, false, 0);
 453
 454         switch (ret) {
 455         case 0:
 456                 break;
 457         case -EBUSY:
 458                 ret = -EAGAIN;
 459                 goto out_unref;
 460         default:
 461                 goto out_unref;
 462         }
 463
 464         ret = ttm_bo_kmap(bo, kmap_offset, kmap_num, &map);
 465         if (unlikely(ret != 0)) {
 466                 ttm_bo_unreserve(bo);
 467                 goto out_unref;
 468         }
 469
 470         virtual = ttm_kmap_obj_virtual(&map, &dummy);
 471         virtual += page_offset;
 472
 473         if (write)
 474                 ret = copy_from_user(virtual, wbuf, io_size);
 475         else
 476                 ret = copy_to_user(rbuf, virtual, io_size);
 477
 478         ttm_bo_kunmap(&map);
 479         ttm_bo_unreserve(bo);
 480         ttm_bo_unref(&bo);
 481
 482         if (unlikely(ret != 0))
 483                 return -EFBIG;
 484
 485         *f_pos += io_size;
 486
 487         return io_size;
 488 out_unref:
 489         ttm_bo_unref(&bo);
 490         return ret;
 491 }
 492
 493 ssize_t ttm_bo_fbdev_io(struct ttm_buffer_object *bo, const char __user *wbuf,
 494                         char __user *rbuf, size_t count, loff_t *f_pos,
 495                         bool write)
 496 {
 497         struct ttm_bo_kmap_obj map;
 498         unsigned long kmap_offset;
 499         unsigned long kmap_end;
 500         unsigned long kmap_num;
 501         size_t io_size;
 502         unsigned int page_offset;
 503         char *virtual;
 504         int ret;
 505         bool no_wait = false;
 506         bool dummy;
 507
 508         kmap_offset = (*f_pos >> PAGE_SHIFT);
 509         if (unlikely(kmap_offset >= bo->num_pages))
 510                 return -EFBIG;
 511
 512         page_offset = *f_pos & ~PAGE_MASK;
 513         io_size = bo->num_pages - kmap_offset;
 514         io_size = (io_size << PAGE_SHIFT) - page_offset;
 515         if (count < io_size)
 516                 io_size = count;
 517
 518         kmap_end = (*f_pos + count - 1) >> PAGE_SHIFT;
 519         kmap_num = kmap_end - kmap_offset + 1;
 520
 521         ret = ttm_bo_reserve(bo, true, no_wait, false, 0);
 522
 523         switch (ret) {
 524         case 0:
 525                 break;
 526         case -EBUSY:
 527                 return -EAGAIN;
 528         default:
 529                 return ret;
 530         }
 531
 532         ret = ttm_bo_kmap(bo, kmap_offset, kmap_num, &map);
 533         if (unlikely(ret != 0)) {
 534                 ttm_bo_unreserve(bo);
 535                 return ret;
 536         }
 537
 538         virtual = ttm_kmap_obj_virtual(&map, &dummy);
 539         virtual += page_offset;
 540
 541         if (write)
 542                 ret = copy_from_user(virtual, wbuf, io_size);
 543         else
 544                 ret = copy_to_user(rbuf, virtual, io_size);
 545
 546         ttm_bo_kunmap(&map);
 547         ttm_bo_unreserve(bo);
 548         ttm_bo_unref(&bo);
 549
 550         if (unlikely(ret != 0))
 551                 return ret;
 552
 553         *f_pos += io_size;
 554
 555         return io_size;
 556 }
 557 #endif