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/errno.h>
  50 #include <linux/export.h>
  51
  52 RB_GENERATE(ttm_bo_device_buffer_objects, ttm_buffer_object, vm_rb,
  53     ttm_bo_cmp_rb_tree_items);
  54
  55
  56 #define TTM_BO_VM_NUM_PREFAULT 16
  57
  58 int
  59 ttm_bo_cmp_rb_tree_items(struct ttm_buffer_object *a,
  60     struct ttm_buffer_object *b)
  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
  72 static struct ttm_buffer_object *ttm_bo_vm_lookup_rb(struct ttm_bo_device *bdev,
  73                                                      unsigned long page_start,
  74                                                      unsigned long num_pages)
  75 {
  76         unsigned long cur_offset;
  77         struct ttm_buffer_object *bo;
  78         struct ttm_buffer_object *best_bo = NULL;
  79
  80         bo = RB_ROOT(&bdev->addr_space_rb);
  81         while (bo != NULL) {
  82                 cur_offset = bo->vm_node->start;
  83                 if (page_start >= cur_offset) {
  84                         best_bo = bo;
  85                         if (page_start == cur_offset)
  86                                 break;
  87                         bo = RB_RIGHT(bo, vm_rb);
  88                 } else
  89                         bo = RB_LEFT(bo, vm_rb);
  90         }
  91
  92         if (unlikely(best_bo == NULL))
  93                 return NULL;
  94
  95         if (unlikely((best_bo->vm_node->start + best_bo->num_pages) <
  96                      (page_start + num_pages)))
  97                 return NULL;
  98
  99         return best_bo;
 100 }
 101
 102 static int
 103 ttm_bo_vm_fault(vm_object_t vm_obj, vm_ooffset_t offset,
 104     int prot, vm_page_t *mres)
 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, oldm;
 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         oldm = *mres;
 117         if (oldm != NULL) {
 118                 vm_page_remove(oldm);
 119                 *mres = NULL;
 120         } else
 121                 oldm = NULL;
 122 retry:
 123         VM_OBJECT_UNLOCK(vm_obj);
 124         m = NULL;
 125
 126 reserve:
 127         ret = ttm_bo_reserve(bo, false, false, false, 0);
 128         if (unlikely(ret != 0)) {
 129                 if (ret == -EBUSY) {
 130                         lwkt_yield();
 131                         goto reserve;
 132                 }
 133         }
 134
 135         if (bdev->driver->fault_reserve_notify) {
 136                 ret = bdev->driver->fault_reserve_notify(bo);
 137                 switch (ret) {
 138                 case 0:
 139                         break;
 140                 case -EBUSY:
 141                 case -ERESTARTSYS:
 142                 case -EINTR:
 143                         lwkt_yield();
 144                         goto reserve;
 145                 default:
 146                         retval = VM_PAGER_ERROR;
 147                         goto out_unlock;
 148                 }
 149         }
 150
 151         /*
 152          * Wait for buffer data in transit, due to a pipelined
 153          * move.
 154          */
 155
 156         lockmgr(&bdev->fence_lock, LK_EXCLUSIVE);
 157         if (test_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags)) {
 158                 /*
 159                  * Here, the behavior differs between Linux and FreeBSD.
 160                  *
 161                  * On Linux, the wait is interruptible (3rd argument to
 162                  * ttm_bo_wait). There must be some mechanism to resume
 163                  * page fault handling, once the signal is processed.
 164                  *
 165                  * On FreeBSD, the wait is uninteruptible. This is not a
 166                  * problem as we can't end up with an unkillable process
 167                  * here, because the wait will eventually time out.
 168                  *
 169                  * An example of this situation is the Xorg process
 170                  * which uses SIGALRM internally. The signal could
 171                  * interrupt the wait, causing the page fault to fail
 172                  * and the process to receive SIGSEGV.
 173                  */
 174                 ret = ttm_bo_wait(bo, false, false, false);
 175                 lockmgr(&bdev->fence_lock, LK_RELEASE);
 176                 if (unlikely(ret != 0)) {
 177                         retval = VM_PAGER_ERROR;
 178                         goto out_unlock;
 179                 }
 180         } else
 181                 lockmgr(&bdev->fence_lock, LK_RELEASE);
 182
 183         ret = ttm_mem_io_lock(man, true);
 184         if (unlikely(ret != 0)) {
 185                 retval = VM_PAGER_ERROR;
 186                 goto out_unlock;
 187         }
 188         ret = ttm_mem_io_reserve_vm(bo);
 189         if (unlikely(ret != 0)) {
 190                 retval = VM_PAGER_ERROR;
 191                 goto out_io_unlock;
 192         }
 193
 194         /*
 195          * Strictly, we're not allowed to modify vma->vm_page_prot here,
 196          * since the mmap_sem is only held in read mode. However, we
 197          * modify only the caching bits of vma->vm_page_prot and
 198          * consider those bits protected by
 199          * the bo->mutex, as we should be the only writers.
 200          * There shouldn't really be any readers of these bits except
 201          * within vm_insert_mixed()? fork?
 202          *
 203          * TODO: Add a list of vmas to the bo, and change the
 204          * vma->vm_page_prot when the object changes caching policy, with
 205          * the correct locks held.
 206          */
 207         if (!bo->mem.bus.is_iomem) {
 208                 /* Allocate all page at once, most common usage */
 209                 ttm = bo->ttm;
 210                 if (ttm->bdev->driver->ttm_tt_populate(ttm)) {
 211                         retval = VM_PAGER_ERROR;
 212                         goto out_io_unlock;
 213                 }
 214         }
 215
 216         if (bo->mem.bus.is_iomem) {
 217                 m = vm_phys_fictitious_to_vm_page(bo->mem.bus.base +
 218                     bo->mem.bus.offset + offset);
 219                 pmap_page_set_memattr(m, ttm_io_prot(bo->mem.placement));
 220         } else {
 221                 ttm = bo->ttm;
 222                 m = (struct vm_page *)ttm->pages[OFF_TO_IDX(offset)];
 223                 if (unlikely(!m)) {
 224                         retval = VM_PAGER_ERROR;
 225                         goto out_io_unlock;
 226                 }
 227                 pmap_page_set_memattr(m,
 228                     (bo->mem.placement & TTM_PL_FLAG_CACHED) ?
 229                     VM_MEMATTR_WRITE_BACK : ttm_io_prot(bo->mem.placement));
 230         }
 231
 232         VM_OBJECT_LOCK(vm_obj);
 233         if ((m->busy_count & PBUSY_LOCKED) != 0) {
 234 #if 0
 235                 vm_page_sleep(m, "ttmpbs");
 236 #endif
 237                 ttm_mem_io_unlock(man);
 238                 ttm_bo_unreserve(bo);
 239                 goto retry;
 240         }
 241         m->valid = VM_PAGE_BITS_ALL;
 242         *mres = m;
 243         m1 = vm_page_lookup(vm_obj, OFF_TO_IDX(offset));
 244         if (m1 == NULL) {
 245                 vm_page_insert(m, vm_obj, OFF_TO_IDX(offset));
 246         } else {
 247                 KASSERT(m == m1,
 248                     ("inconsistent insert bo %p m %p m1 %p offset %jx",
 249                     bo, m, m1, (uintmax_t)offset));
 250         }
 251         vm_page_busy_try(m, FALSE);
 252
 253         if (oldm != NULL) {
 254                 vm_page_free(oldm);
 255         }
 256
 257 out_io_unlock1:
 258         ttm_mem_io_unlock(man);
 259 out_unlock1:
 260         ttm_bo_unreserve(bo);
 261         vm_object_pip_wakeup(vm_obj);
 262         return (retval);
 263
 264 out_io_unlock:
 265         VM_OBJECT_LOCK(vm_obj);
 266         goto out_io_unlock1;
 267
 268 out_unlock:
 269         VM_OBJECT_LOCK(vm_obj);
 270         goto out_unlock1;
 271 }
 272
 273 static int
 274 ttm_bo_vm_ctor(void *handle, vm_ooffset_t size, vm_prot_t prot,
 275     vm_ooffset_t foff, struct ucred *cred, u_short *color)
 276 {
 277
 278         /*
 279          * On Linux, a reference to the buffer object is acquired here.
 280          * The reason is that this function is not called when the
 281          * mmap() is initialized, but only when a process forks for
 282          * instance. Therefore on Linux, the reference on the bo is
 283          * acquired either in ttm_bo_mmap() or ttm_bo_vm_open(). It's
 284          * then released in ttm_bo_vm_close().
 285          *
 286          * Here, this function is called during mmap() intialization.
 287          * Thus, the reference acquired in ttm_bo_mmap_single() is
 288          * sufficient.
 289          */
 290         *color = 0;
 291         return (0);
 292 }
 293
 294 static void
 295 ttm_bo_vm_dtor(void *handle)
 296 {
 297         struct ttm_buffer_object *bo = handle;
 298
 299         ttm_bo_unref(&bo);
 300 }
 301
 302 static struct cdev_pager_ops ttm_pager_ops = {
 303         .cdev_pg_fault = ttm_bo_vm_fault,
 304         .cdev_pg_ctor = ttm_bo_vm_ctor,
 305         .cdev_pg_dtor = ttm_bo_vm_dtor
 306 };
 307
 308 int
 309 ttm_bo_mmap_single(struct ttm_bo_device *bdev, vm_ooffset_t *offset, vm_size_t size,
 310     struct vm_object **obj_res, int nprot)
 311 {
 312         struct ttm_bo_driver *driver;
 313         struct ttm_buffer_object *bo;
 314         struct vm_object *vm_obj;
 315         int ret;
 316
 317         *obj_res = NULL;
 318
 319         lockmgr(&bdev->vm_lock, LK_EXCLUSIVE);
 320         bo = ttm_bo_vm_lookup_rb(bdev, OFF_TO_IDX(*offset), OFF_TO_IDX(size));
 321         if (likely(bo != NULL))
 322                 kref_get(&bo->kref);
 323         lockmgr(&bdev->vm_lock, LK_RELEASE);
 324
 325         if (unlikely(bo == NULL)) {
 326                 pr_err("Could not find buffer object to map\n");
 327                 return (EINVAL);
 328         }
 329
 330         driver = bo->bdev->driver;
 331         if (unlikely(!driver->verify_access)) {
 332                 ret = EPERM;
 333                 goto out_unref;
 334         }
 335         ret = -driver->verify_access(bo);
 336         if (unlikely(ret != 0))
 337                 goto out_unref;
 338
 339         vm_obj = cdev_pager_allocate(bo, OBJT_MGTDEVICE, &ttm_pager_ops,
 340             size, nprot, 0, curthread->td_ucred);
 341
 342         if (vm_obj == NULL) {
 343                 ret = EINVAL;
 344                 goto out_unref;
 345         }
 346         /*
 347          * Note: We're transferring the bo reference to vm_obj->handle here.
 348          */
 349         *offset = 0;
 350         *obj_res = vm_obj;
 351         return 0;
 352 out_unref:
 353         ttm_bo_unref(&bo);
 354         return ret;
 355 }
 356 EXPORT_SYMBOL(ttm_bo_mmap);
 357
 358 void
 359 ttm_bo_release_mmap(struct ttm_buffer_object *bo)
 360 {
 361         vm_object_t vm_obj;
 362         vm_page_t m;
 363         int i;
 364
 365         vm_obj = cdev_pager_lookup(bo);
 366         if (vm_obj == NULL)
 367                 return;
 368
 369         VM_OBJECT_LOCK(vm_obj);
 370         for (i = 0; i < bo->num_pages; i++) {
 371                 m = vm_page_lookup_busy_wait(vm_obj, i, TRUE, "ttm_unm");
 372                 if (m == NULL)
 373                         continue;
 374                 cdev_pager_free_page(vm_obj, m);
 375         }
 376         VM_OBJECT_UNLOCK(vm_obj);
 377
 378         vm_object_deallocate(vm_obj);
 379 }
 380
 381 #if 0
 382 int ttm_fbdev_mmap(struct vm_area_struct *vma, struct ttm_buffer_object *bo)
 383 {
 384         if (vma->vm_pgoff != 0)
 385                 return -EACCES;
 386
 387         vma->vm_ops = &ttm_bo_vm_ops;
 388         vma->vm_private_data = ttm_bo_reference(bo);
 389         vma->vm_flags |= VM_IO | VM_MIXEDMAP | VM_DONTEXPAND;
 390         return 0;
 391 }
 392 EXPORT_SYMBOL(ttm_fbdev_mmap);
 393 #endif