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