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 #include <vm/vm_page2.h>
  53
  54 RB_GENERATE(ttm_bo_device_buffer_objects, ttm_buffer_object, vm_rb,
  55     ttm_bo_cmp_rb_tree_items);
  56
  57
  58 #define TTM_BO_VM_NUM_PREFAULT 16
  59
  60 int
  61 ttm_bo_cmp_rb_tree_items(struct ttm_buffer_object *a,
  62     struct ttm_buffer_object *b)
  63 {
  64     if (a->vm_node->start < b->vm_node->start) {
  65         return (-1);
  66     } else if (a->vm_node->start > b->vm_node->start) {
  67         return (1);
  68     } else {
  69         return (0);
  70     }
  71 }
  72
  73
  74 static struct ttm_buffer_object *ttm_bo_vm_lookup_rb(struct ttm_bo_device *bdev,
  75                                                      unsigned long page_start,
  76                                                      unsigned long num_pages)
  77 {
  78         unsigned long cur_offset;
  79         struct ttm_buffer_object *bo;
  80         struct ttm_buffer_object *best_bo = NULL;
  81
  82         bo = RB_ROOT(&bdev->addr_space_rb);
  83         while (bo != NULL) {
  84                 cur_offset = bo->vm_node->start;
  85                 if (page_start >= cur_offset) {
  86                         best_bo = bo;
  87                         if (page_start == cur_offset)
  88                                 break;
  89                         bo = RB_RIGHT(bo, vm_rb);
  90                 } else
  91                         bo = RB_LEFT(bo, vm_rb);
  92         }
  93
  94         if (unlikely(best_bo == NULL))
  95                 return NULL;
  96
  97         if (unlikely((best_bo->vm_node->start + best_bo->num_pages) <
  98                      (page_start + num_pages)))
  99                 return NULL;
 100
 101         return best_bo;
 102 }
 103
 104 static int
 105 ttm_bo_vm_fault(vm_object_t vm_obj, vm_ooffset_t offset,
 106     int prot, vm_page_t *mres)
 107 {
 108         struct ttm_buffer_object *bo = vm_obj->handle;
 109         struct ttm_bo_device *bdev = bo->bdev;
 110         struct ttm_tt *ttm = NULL;
 111         vm_page_t m, m1, oldm;
 112         int ret;
 113         int retval = VM_PAGER_OK;
 114         struct ttm_mem_type_manager *man =
 115                 &bdev->man[bo->mem.mem_type];
 116
 117         vm_object_pip_add(vm_obj, 1);
 118         oldm = *mres;
 119         if (oldm != NULL) {
 120                 vm_page_remove(oldm);
 121                 *mres = NULL;
 122         } else
 123                 oldm = NULL;
 124 retry:
 125         VM_OBJECT_UNLOCK(vm_obj);
 126         m = NULL;
 127
 128 reserve:
 129         ret = ttm_bo_reserve(bo, false, false, false, 0);
 130         if (unlikely(ret != 0)) {
 131                 if (ret == -EBUSY) {
 132                         lwkt_yield();
 133                         goto reserve;
 134                 }
 135         }
 136
 137         if (bdev->driver->fault_reserve_notify) {
 138                 ret = bdev->driver->fault_reserve_notify(bo);
 139                 switch (ret) {
 140                 case 0:
 141                         break;
 142                 case -EBUSY:
 143                 case -ERESTARTSYS:
 144                 case -EINTR:
 145                         lwkt_yield();
 146                         goto reserve;
 147                 default:
 148                         retval = VM_PAGER_ERROR;
 149                         goto out_unlock;
 150                 }
 151         }
 152
 153         /*
 154          * Wait for buffer data in transit, due to a pipelined
 155          * move.
 156          */
 157
 158         lockmgr(&bdev->fence_lock, LK_EXCLUSIVE);
 159         if (test_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags)) {
 160                 /*
 161                  * Here, the behavior differs between Linux and FreeBSD.
 162                  *
 163                  * On Linux, the wait is interruptible (3rd argument to
 164                  * ttm_bo_wait). There must be some mechanism to resume
 165                  * page fault handling, once the signal is processed.
 166                  *
 167                  * On FreeBSD, the wait is uninteruptible. This is not a
 168                  * problem as we can't end up with an unkillable process
 169                  * here, because the wait will eventually time out.
 170                  *
 171                  * An example of this situation is the Xorg process
 172                  * which uses SIGALRM internally. The signal could
 173                  * interrupt the wait, causing the page fault to fail
 174                  * and the process to receive SIGSEGV.
 175                  */
 176                 ret = ttm_bo_wait(bo, false, false, false);
 177                 lockmgr(&bdev->fence_lock, LK_RELEASE);
 178                 if (unlikely(ret != 0)) {
 179                         retval = VM_PAGER_ERROR;
 180                         goto out_unlock;
 181                 }
 182         } else
 183                 lockmgr(&bdev->fence_lock, LK_RELEASE);
 184
 185         ret = ttm_mem_io_lock(man, true);
 186         if (unlikely(ret != 0)) {
 187                 retval = VM_PAGER_ERROR;
 188                 goto out_unlock;
 189         }
 190         ret = ttm_mem_io_reserve_vm(bo);
 191         if (unlikely(ret != 0)) {
 192                 retval = VM_PAGER_ERROR;
 193                 goto out_io_unlock;
 194         }
 195
 196         /*
 197          * Strictly, we're not allowed to modify vma->vm_page_prot here,
 198          * since the mmap_sem is only held in read mode. However, we
 199          * modify only the caching bits of vma->vm_page_prot and
 200          * consider those bits protected by
 201          * the bo->mutex, as we should be the only writers.
 202          * There shouldn't really be any readers of these bits except
 203          * within vm_insert_mixed()? fork?
 204          *
 205          * TODO: Add a list of vmas to the bo, and change the
 206          * vma->vm_page_prot when the object changes caching policy, with
 207          * the correct locks held.
 208          */
 209         if (!bo->mem.bus.is_iomem) {
 210                 /* Allocate all page at once, most common usage */
 211                 ttm = bo->ttm;
 212                 if (ttm->bdev->driver->ttm_tt_populate(ttm)) {
 213                         retval = VM_PAGER_ERROR;
 214                         goto out_io_unlock;
 215                 }
 216         }
 217
 218         if (bo->mem.bus.is_iomem) {
 219                 m = vm_phys_fictitious_to_vm_page(bo->mem.bus.base +
 220                     bo->mem.bus.offset + offset);
 221                 pmap_page_set_memattr(m, ttm_io_prot(bo->mem.placement));
 222         } else {
 223                 ttm = bo->ttm;
 224                 m = (struct vm_page *)ttm->pages[OFF_TO_IDX(offset)];
 225                 if (unlikely(!m)) {
 226                         retval = VM_PAGER_ERROR;
 227                         goto out_io_unlock;
 228                 }
 229                 pmap_page_set_memattr(m,
 230                     (bo->mem.placement & TTM_PL_FLAG_CACHED) ?
 231                     VM_MEMATTR_WRITE_BACK : ttm_io_prot(bo->mem.placement));
 232         }
 233
 234         VM_OBJECT_LOCK(vm_obj);
 235         if ((m->busy_count & PBUSY_LOCKED) != 0) {
 236 #if 0
 237                 vm_page_sleep(m, "ttmpbs");
 238 #endif
 239                 ttm_mem_io_unlock(man);
 240                 ttm_bo_unreserve(bo);
 241                 goto retry;
 242         }
 243         m->valid = VM_PAGE_BITS_ALL;
 244         *mres = m;
 245         m1 = vm_page_lookup(vm_obj, OFF_TO_IDX(offset));
 246         if (m1 == NULL) {
 247                 vm_page_insert(m, vm_obj, OFF_TO_IDX(offset));
 248         } else {
 249                 KASSERT(m == m1,
 250                     ("inconsistent insert bo %p m %p m1 %p offset %jx",
 251                     bo, m, m1, (uintmax_t)offset));
 252         }
 253         vm_page_busy_try(m, FALSE);
 254
 255         if (oldm != NULL) {
 256                 vm_page_free(oldm);
 257         }
 258
 259 out_io_unlock1:
 260         ttm_mem_io_unlock(man);
 261 out_unlock1:
 262         ttm_bo_unreserve(bo);
 263         vm_object_pip_wakeup(vm_obj);
 264         return (retval);
 265
 266 out_io_unlock:
 267         VM_OBJECT_LOCK(vm_obj);
 268         goto out_io_unlock1;
 269
 270 out_unlock:
 271         VM_OBJECT_LOCK(vm_obj);
 272         goto out_unlock1;
 273 }
 274
 275 static int
 276 ttm_bo_vm_ctor(void *handle, vm_ooffset_t size, vm_prot_t prot,
 277     vm_ooffset_t foff, struct ucred *cred, u_short *color)
 278 {
 279
 280         /*
 281          * On Linux, a reference to the buffer object is acquired here.
 282          * The reason is that this function is not called when the
 283          * mmap() is initialized, but only when a process forks for
 284          * instance. Therefore on Linux, the reference on the bo is
 285          * acquired either in ttm_bo_mmap() or ttm_bo_vm_open(). It's
 286          * then released in ttm_bo_vm_close().
 287          *
 288          * Here, this function is called during mmap() intialization.
 289          * Thus, the reference acquired in ttm_bo_mmap_single() is
 290          * sufficient.
 291          */
 292         *color = 0;
 293         return (0);
 294 }
 295
 296 static void
 297 ttm_bo_vm_dtor(void *handle)
 298 {
 299         struct ttm_buffer_object *bo = handle;
 300
 301         ttm_bo_unref(&bo);
 302 }
 303
 304 static struct cdev_pager_ops ttm_pager_ops = {
 305         .cdev_pg_fault = ttm_bo_vm_fault,
 306         .cdev_pg_ctor = ttm_bo_vm_ctor,
 307         .cdev_pg_dtor = ttm_bo_vm_dtor
 308 };
 309
 310 int
 311 ttm_bo_mmap_single(struct ttm_bo_device *bdev, vm_ooffset_t *offset, vm_size_t size,
 312     struct vm_object **obj_res, int nprot)
 313 {
 314         struct ttm_bo_driver *driver;
 315         struct ttm_buffer_object *bo;
 316         struct vm_object *vm_obj;
 317         int ret;
 318
 319         *obj_res = NULL;
 320
 321         lockmgr(&bdev->vm_lock, LK_EXCLUSIVE);
 322         bo = ttm_bo_vm_lookup_rb(bdev, OFF_TO_IDX(*offset), OFF_TO_IDX(size));
 323         if (likely(bo != NULL))
 324                 kref_get(&bo->kref);
 325         lockmgr(&bdev->vm_lock, LK_RELEASE);
 326
 327         if (unlikely(bo == NULL)) {
 328                 pr_err("Could not find buffer object to map\n");
 329                 return (EINVAL);
 330         }
 331
 332         driver = bo->bdev->driver;
 333         if (unlikely(!driver->verify_access)) {
 334                 ret = EPERM;
 335                 goto out_unref;
 336         }
 337         ret = -driver->verify_access(bo);
 338         if (unlikely(ret != 0))
 339                 goto out_unref;
 340
 341         vm_obj = cdev_pager_allocate(bo, OBJT_MGTDEVICE, &ttm_pager_ops,
 342             size, nprot, 0, curthread->td_ucred);
 343
 344         if (vm_obj == NULL) {
 345                 ret = EINVAL;
 346                 goto out_unref;
 347         }
 348         /*
 349          * Note: We're transferring the bo reference to vm_obj->handle here.
 350          */
 351         *offset = 0;
 352         *obj_res = vm_obj;
 353         return 0;
 354 out_unref:
 355         ttm_bo_unref(&bo);
 356         return ret;
 357 }
 358 EXPORT_SYMBOL(ttm_bo_mmap);
 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_LOCK(vm_obj);
 372         for (i = 0; i < bo->num_pages; i++) {
 373                 m = vm_page_lookup_busy_wait(vm_obj, i, TRUE, "ttm_unm");
 374                 if (m == NULL)
 375                         continue;
 376                 cdev_pager_free_page(vm_obj, m);
 377         }
 378         VM_OBJECT_UNLOCK(vm_obj);
 379
 380         vm_object_deallocate(vm_obj);
 381 }
 382
 383 #if 0
 384 int ttm_fbdev_mmap(struct vm_area_struct *vma, struct ttm_buffer_object *bo)
 385 {
 386         if (vma->vm_pgoff != 0)
 387                 return -EACCES;
 388
 389         vma->vm_ops = &ttm_bo_vm_ops;
 390         vma->vm_private_data = ttm_bo_reference(bo);
 391         vma->vm_flags |= VM_IO | VM_MIXEDMAP | VM_DONTEXPAND;
 392         return 0;
 393 }
 394 EXPORT_SYMBOL(ttm_fbdev_mmap);
 395 #endif