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