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         if (m1 == NULL) {
 241                 if (vm_page_insert(m, vm_obj, OFF_TO_IDX(offset))) {
 242                         vm_page_xunbusy(m);
 243                         VM_OBJECT_WUNLOCK(vm_obj);
 244                         vm_wait(vm_obj);
 245                         VM_OBJECT_WLOCK(vm_obj);
 246                         ttm_mem_io_unlock(man);
 247                         ttm_bo_unreserve(bo);
 248                         goto retry;
 249                 }
 250         } else {
 251                 KASSERT(m == m1,
 252                     ("inconsistent insert bo %p m %p m1 %p offset %jx",
 253                     bo, m, m1, (uintmax_t)offset));
 254         }
 255         vm_page_valid(m);
 256         if (*mres != NULL) {
 257                 KASSERT(*mres != m, ("losing %p %p", *mres, m));
 258                 vm_page_free(*mres);
 259         }
 260         *mres = m;
 261
 262 out_io_unlock1:
 263         ttm_mem_io_unlock(man);
 264 out_unlock1:
 265         ttm_bo_unreserve(bo);
 266         vm_object_pip_wakeup(vm_obj);
 267         return (retval);
 268
 269 out_io_unlock:
 270         VM_OBJECT_WLOCK(vm_obj);
 271         goto out_io_unlock1;
 272
 273 out_unlock:
 274         VM_OBJECT_WLOCK(vm_obj);
 275         goto out_unlock1;
 276 }
 277
 278 static int
 279 ttm_bo_vm_ctor(void *handle, vm_ooffset_t size, vm_prot_t prot,
 280     vm_ooffset_t foff, struct ucred *cred, u_short *color)
 281 {
 282
 283         /*
 284          * On Linux, a reference to the buffer object is acquired here.
 285          * The reason is that this function is not called when the
 286          * mmap() is initialized, but only when a process forks for
 287          * instance. Therefore on Linux, the reference on the bo is
 288          * acquired either in ttm_bo_mmap() or ttm_bo_vm_open(). It's
 289          * then released in ttm_bo_vm_close().
 290          *
 291          * Here, this function is called during mmap() initialization.
 292          * Thus, the reference acquired in ttm_bo_mmap_single() is
 293          * sufficient.
 294          */
 295
 296         *color = 0;
 297         return (0);
 298 }
 299
 300 static void
 301 ttm_bo_vm_dtor(void *handle)
 302 {
 303         struct ttm_buffer_object *bo = handle;
 304
 305         ttm_bo_unref(&bo);
 306 }
 307
 308 static struct cdev_pager_ops ttm_pager_ops = {
 309         .cdev_pg_fault = ttm_bo_vm_fault,
 310         .cdev_pg_ctor = ttm_bo_vm_ctor,
 311         .cdev_pg_dtor = ttm_bo_vm_dtor
 312 };
 313
 314 int
 315 ttm_bo_mmap_single(struct ttm_bo_device *bdev, vm_ooffset_t *offset, vm_size_t size,
 316     struct vm_object **obj_res, int nprot)
 317 {
 318         struct ttm_bo_driver *driver;
 319         struct ttm_buffer_object *bo;
 320         struct vm_object *vm_obj;
 321         int ret;
 322
 323         rw_wlock(&bdev->vm_lock);
 324         bo = ttm_bo_vm_lookup_rb(bdev, OFF_TO_IDX(*offset), OFF_TO_IDX(size));
 325         if (likely(bo != NULL))
 326                 refcount_acquire(&bo->kref);
 327         rw_wunlock(&bdev->vm_lock);
 328
 329         if (unlikely(bo == NULL)) {
 330                 printf("[TTM] Could not find buffer object to map\n");
 331                 return (-EINVAL);
 332         }
 333
 334         driver = bo->bdev->driver;
 335         if (unlikely(!driver->verify_access)) {
 336                 ret = -EPERM;
 337                 goto out_unref;
 338         }
 339         ret = driver->verify_access(bo);
 340         if (unlikely(ret != 0))
 341                 goto out_unref;
 342
 343         vm_obj = cdev_pager_allocate(bo, OBJT_MGTDEVICE, &ttm_pager_ops,
 344             size, nprot, 0, curthread->td_ucred);
 345         if (vm_obj == NULL) {
 346                 ret = -EINVAL;
 347                 goto out_unref;
 348         }
 349         /*
 350          * Note: We're transferring the bo reference to vm_obj->handle here.
 351          */
 352         *offset = 0;
 353         *obj_res = vm_obj;
 354         return 0;
 355 out_unref:
 356         ttm_bo_unref(&bo);
 357         return ret;
 358 }
 359
 360 void
 361 ttm_bo_release_mmap(struct ttm_buffer_object *bo)
 362 {
 363         vm_object_t vm_obj;
 364         vm_page_t m;
 365         int i;
 366
 367         vm_obj = cdev_pager_lookup(bo);
 368         if (vm_obj == NULL)
 369                 return;
 370
 371         VM_OBJECT_WLOCK(vm_obj);
 372 retry:
 373         for (i = 0; i < bo->num_pages; i++) {
 374                 m = vm_page_lookup(vm_obj, i);
 375                 if (m == NULL)
 376                         continue;
 377                 if (vm_page_busy_acquire(m, VM_ALLOC_WAITFAIL) == 0)
 378                         goto retry;
 379                 cdev_pager_free_page(vm_obj, m);
 380         }
 381         VM_OBJECT_WUNLOCK(vm_obj);
 382
 383         vm_object_deallocate(vm_obj);
 384 }
 385
 386 #if 0
 387 int ttm_fbdev_mmap(struct vm_area_struct *vma, struct ttm_buffer_object *bo)
 388 {
 389         if (vma->vm_pgoff != 0)
 390                 return -EACCES;
 391
 392         vma->vm_ops = &ttm_bo_vm_ops;
 393         vma->vm_private_data = ttm_bo_reference(bo);
 394         vma->vm_flags |= VM_IO | VM_MIXEDMAP | VM_DONTEXPAND;
 395         return 0;
 396 }
 397
 398 ssize_t ttm_bo_io(struct ttm_bo_device *bdev, struct file *filp,
 399                   const char __user *wbuf, char __user *rbuf, size_t count,
 400                   loff_t *f_pos, bool write)
 401 {
 402         struct ttm_buffer_object *bo;
 403         struct ttm_bo_driver *driver;
 404         struct ttm_bo_kmap_obj map;
 405         unsigned long dev_offset = (*f_pos >> PAGE_SHIFT);
 406         unsigned long kmap_offset;
 407         unsigned long kmap_end;
 408         unsigned long kmap_num;
 409         size_t io_size;
 410         unsigned int page_offset;
 411         char *virtual;
 412         int ret;
 413         bool no_wait = false;
 414         bool dummy;
 415
 416         read_lock(&bdev->vm_lock);
 417         bo = ttm_bo_vm_lookup_rb(bdev, dev_offset, 1);
 418         if (likely(bo != NULL))
 419                 ttm_bo_reference(bo);
 420         read_unlock(&bdev->vm_lock);
 421
 422         if (unlikely(bo == NULL))
 423                 return -EFAULT;
 424
 425         driver = bo->bdev->driver;
 426         if (unlikely(!driver->verify_access)) {
 427                 ret = -EPERM;
 428                 goto out_unref;
 429         }
 430
 431         ret = driver->verify_access(bo, filp);
 432         if (unlikely(ret != 0))
 433                 goto out_unref;
 434
 435         kmap_offset = dev_offset - bo->vm_node->start;
 436         if (unlikely(kmap_offset >= bo->num_pages)) {
 437                 ret = -EFBIG;
 438                 goto out_unref;
 439         }
 440
 441         page_offset = *f_pos & ~PAGE_MASK;
 442         io_size = bo->num_pages - kmap_offset;
 443         io_size = (io_size << PAGE_SHIFT) - page_offset;
 444         if (count < io_size)
 445                 io_size = count;
 446
 447         kmap_end = (*f_pos + count - 1) >> PAGE_SHIFT;
 448         kmap_num = kmap_end - kmap_offset + 1;
 449
 450         ret = ttm_bo_reserve(bo, true, no_wait, false, 0);
 451
 452         switch (ret) {
 453         case 0:
 454                 break;
 455         case -EBUSY:
 456                 ret = -EAGAIN;
 457                 goto out_unref;
 458         default:
 459                 goto out_unref;
 460         }
 461
 462         ret = ttm_bo_kmap(bo, kmap_offset, kmap_num, &map);
 463         if (unlikely(ret != 0)) {
 464                 ttm_bo_unreserve(bo);
 465                 goto out_unref;
 466         }
 467
 468         virtual = ttm_kmap_obj_virtual(&map, &dummy);
 469         virtual += page_offset;
 470
 471         if (write)
 472                 ret = copy_from_user(virtual, wbuf, io_size);
 473         else
 474                 ret = copy_to_user(rbuf, virtual, io_size);
 475
 476         ttm_bo_kunmap(&map);
 477         ttm_bo_unreserve(bo);
 478         ttm_bo_unref(&bo);
 479
 480         if (unlikely(ret != 0))
 481                 return -EFBIG;
 482
 483         *f_pos += io_size;
 484
 485         return io_size;
 486 out_unref:
 487         ttm_bo_unref(&bo);
 488         return ret;
 489 }
 490
 491 ssize_t ttm_bo_fbdev_io(struct ttm_buffer_object *bo, const char __user *wbuf,
 492                         char __user *rbuf, size_t count, loff_t *f_pos,
 493                         bool write)
 494 {
 495         struct ttm_bo_kmap_obj map;
 496         unsigned long kmap_offset;
 497         unsigned long kmap_end;
 498         unsigned long kmap_num;
 499         size_t io_size;
 500         unsigned int page_offset;
 501         char *virtual;
 502         int ret;
 503         bool no_wait = false;
 504         bool dummy;
 505
 506         kmap_offset = (*f_pos >> PAGE_SHIFT);
 507         if (unlikely(kmap_offset >= bo->num_pages))
 508                 return -EFBIG;
 509
 510         page_offset = *f_pos & ~PAGE_MASK;
 511         io_size = bo->num_pages - kmap_offset;
 512         io_size = (io_size << PAGE_SHIFT) - page_offset;
 513         if (count < io_size)
 514                 io_size = count;
 515
 516         kmap_end = (*f_pos + count - 1) >> PAGE_SHIFT;
 517         kmap_num = kmap_end - kmap_offset + 1;
 518
 519         ret = ttm_bo_reserve(bo, true, no_wait, false, 0);
 520
 521         switch (ret) {
 522         case 0:
 523                 break;
 524         case -EBUSY:
 525                 return -EAGAIN;
 526         default:
 527                 return ret;
 528         }
 529
 530         ret = ttm_bo_kmap(bo, kmap_offset, kmap_num, &map);
 531         if (unlikely(ret != 0)) {
 532                 ttm_bo_unreserve(bo);
 533                 return ret;
 534         }
 535
 536         virtual = ttm_kmap_obj_virtual(&map, &dummy);
 537         virtual += page_offset;
 538
 539         if (write)
 540                 ret = copy_from_user(virtual, wbuf, io_size);
 541         else
 542                 ret = copy_to_user(rbuf, virtual, io_size);
 543
 544         ttm_bo_kunmap(&map);
 545         ttm_bo_unreserve(bo);
 546         ttm_bo_unref(&bo);
 547
 548         if (unlikely(ret != 0))
 549                 return ret;
 550
 551         *f_pos += io_size;
 552
 553         return io_size;
 554 }
 555 #endif