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