amdgpu/amdgpu_vamgr.c

   1 /*
   2  * Copyright 2014 Advanced Micro Devices, Inc.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE SOFTWARE.
  21  *
  22  */
  23
  24 #ifdef HAVE_CONFIG_H
  25 #include "config.h"
  26 #endif
  27
  28 #include <stdlib.h>
  29 #include <string.h>
  30 #include <errno.h>
  31 #include "amdgpu.h"
  32 #include "amdgpu_drm.h"
  33 #include "amdgpu_internal.h"
  34 #include "util_math.h"
  35
  36 static struct amdgpu_bo_va_mgr vamgr = {{0}};
  37
  38 int amdgpu_va_range_query(amdgpu_device_handle dev,
  39                           enum amdgpu_gpu_va_range type, uint64_t *start, uint64_t *end)
  40 {
  41         if (type == amdgpu_gpu_va_range_general) {
  42                 *start = dev->dev_info.virtual_address_offset;
  43                 *end = dev->dev_info.virtual_address_max;
  44                 return 0;
  45         }
  46         return -EINVAL;
  47 }
  48
  49 static void amdgpu_vamgr_init(struct amdgpu_bo_va_mgr *mgr, uint64_t start,
  50                               uint64_t max, uint64_t alignment)
  51 {
  52         mgr->va_offset = start;
  53         mgr->va_max = max;
  54         mgr->va_alignment = alignment;
  55
  56         list_inithead(&mgr->va_holes);
  57         pthread_mutex_init(&mgr->bo_va_mutex, NULL);
  58 }
  59
  60 static void amdgpu_vamgr_deinit(struct amdgpu_bo_va_mgr *mgr)
  61 {
  62         struct amdgpu_bo_va_hole *hole;
  63         LIST_FOR_EACH_ENTRY(hole, &mgr->va_holes, list) {
  64                 list_del(&hole->list);
  65                 free(hole);
  66         }
  67         pthread_mutex_destroy(&mgr->bo_va_mutex);
  68 }
  69
  70 drm_private struct amdgpu_bo_va_mgr *
  71 amdgpu_vamgr_get_global(struct amdgpu_device *dev)
  72 {
  73         int ref;
  74         ref = atomic_inc_return(&vamgr.refcount);
  75
  76         if (ref == 1)
  77                 amdgpu_vamgr_init(&vamgr, dev->dev_info.virtual_address_offset,
  78                                   dev->dev_info.virtual_address_max,
  79                                   dev->dev_info.virtual_address_alignment);
  80         return &vamgr;
  81 }
  82
  83 drm_private void
  84 amdgpu_vamgr_reference(struct amdgpu_bo_va_mgr **dst,
  85                        struct amdgpu_bo_va_mgr *src)
  86 {
  87         if (update_references(&(*dst)->refcount, NULL))
  88                 amdgpu_vamgr_deinit(*dst);
  89         *dst = src;
  90 }
  91
  92 drm_private uint64_t
  93 amdgpu_vamgr_find_va(struct amdgpu_bo_va_mgr *mgr, uint64_t size,
  94                      uint64_t alignment, uint64_t base_required)
  95 {
  96         struct amdgpu_bo_va_hole *hole, *n;
  97         uint64_t offset = 0, waste = 0;
  98
  99         alignment = MAX2(alignment, mgr->va_alignment);
 100         size = ALIGN(size, mgr->va_alignment);
 101
 102         if (base_required % alignment)
 103                 return AMDGPU_INVALID_VA_ADDRESS;
 104
 105         pthread_mutex_lock(&mgr->bo_va_mutex);
 106         /* TODO: using more appropriate way to track the holes */
 107         /* first look for a hole */
 108         LIST_FOR_EACH_ENTRY_SAFE(hole, n, &vamgr.va_holes, list) {
 109                 if (base_required) {
 110                         if(hole->offset > base_required ||
 111                                 (hole->offset + hole->size) < (base_required + size))
 112                                 continue;
 113                         waste = base_required - hole->offset;
 114                         offset = base_required;
 115                 } else {
 116                         offset = hole->offset;
 117                         waste = offset % alignment;
 118                         waste = waste ? alignment - waste : 0;
 119                         offset += waste;
 120                         if (offset >= (hole->offset + hole->size)) {
 121                                 continue;
 122                         }
 123                 }
 124                 if (!waste && hole->size == size) {
 125                         offset = hole->offset;
 126                         list_del(&hole->list);
 127                         free(hole);
 128                         pthread_mutex_unlock(&mgr->bo_va_mutex);
 129                         return offset;
 130                 }
 131                 if ((hole->size - waste) > size) {
 132                         if (waste) {
 133                                 n = calloc(1, sizeof(struct amdgpu_bo_va_hole));
 134                                 n->size = waste;
 135                                 n->offset = hole->offset;
 136                                 list_add(&n->list, &hole->list);
 137                         }
 138                         hole->size -= (size + waste);
 139                         hole->offset += size + waste;
 140                         pthread_mutex_unlock(&mgr->bo_va_mutex);
 141                         return offset;
 142                 }
 143                 if ((hole->size - waste) == size) {
 144                         hole->size = waste;
 145                         pthread_mutex_unlock(&mgr->bo_va_mutex);
 146                         return offset;
 147                 }
 148         }
 149
 150         if (base_required) {
 151                 if (base_required < mgr->va_offset)
 152                         return AMDGPU_INVALID_VA_ADDRESS;
 153                 offset = mgr->va_offset;
 154                 waste = base_required - mgr->va_offset;
 155         } else {
 156                 offset = mgr->va_offset;
 157                 waste = offset % alignment;
 158                 waste = waste ? alignment - waste : 0;
 159         }
 160
 161         if (offset + waste + size > mgr->va_max) {
 162                 pthread_mutex_unlock(&mgr->bo_va_mutex);
 163                 return AMDGPU_INVALID_VA_ADDRESS;
 164         }
 165
 166         if (waste) {
 167                 n = calloc(1, sizeof(struct amdgpu_bo_va_hole));
 168                 n->size = waste;
 169                 n->offset = offset;
 170                 list_add(&n->list, &mgr->va_holes);
 171         }
 172
 173         offset += waste;
 174         mgr->va_offset += size + waste;
 175         pthread_mutex_unlock(&mgr->bo_va_mutex);
 176         return offset;
 177 }
 178
 179 drm_private void
 180 amdgpu_vamgr_free_va(struct amdgpu_bo_va_mgr *mgr, uint64_t va, uint64_t size)
 181 {
 182         struct amdgpu_bo_va_hole *hole;
 183
 184         if (va == AMDGPU_INVALID_VA_ADDRESS)
 185                 return;
 186
 187         size = ALIGN(size, mgr->va_alignment);
 188
 189         pthread_mutex_lock(&mgr->bo_va_mutex);
 190         if ((va + size) == mgr->va_offset) {
 191                 mgr->va_offset = va;
 192                 /* Delete uppermost hole if it reaches the new top */
 193                 if (!LIST_IS_EMPTY(&mgr->va_holes)) {
 194                         hole = container_of(mgr->va_holes.next, hole, list);
 195                         if ((hole->offset + hole->size) == va) {
 196                                 mgr->va_offset = hole->offset;
 197                                 list_del(&hole->list);
 198                                 free(hole);
 199                         }
 200                 }
 201         } else {
 202                 struct amdgpu_bo_va_hole *next;
 203
 204                 hole = container_of(&mgr->va_holes, hole, list);
 205                 LIST_FOR_EACH_ENTRY(next, &mgr->va_holes, list) {
 206                         if (next->offset < va)
 207                                 break;
 208                         hole = next;
 209                 }
 210
 211                 if (&hole->list != &mgr->va_holes) {
 212                         /* Grow upper hole if it's adjacent */
 213                         if (hole->offset == (va + size)) {
 214                                 hole->offset = va;
 215                                 hole->size += size;
 216                                 /* Merge lower hole if it's adjacent */
 217                                 if (next != hole
 218                                                 && &next->list != &mgr->va_holes
 219                                                 && (next->offset + next->size) == va) {
 220                                         next->size += hole->size;
 221                                         list_del(&hole->list);
 222                                         free(hole);
 223                                 }
 224                                 goto out;
 225                         }
 226                 }
 227
 228                 /* Grow lower hole if it's adjacent */
 229                 if (next != hole && &next->list != &mgr->va_holes &&
 230                                 (next->offset + next->size) == va) {
 231                         next->size += size;
 232                         goto out;
 233                 }
 234
 235                 /* FIXME on allocation failure we just lose virtual address space
 236                  * maybe print a warning
 237                  */
 238                 next = calloc(1, sizeof(struct amdgpu_bo_va_hole));
 239                 if (next) {
 240                         next->size = size;
 241                         next->offset = va;
 242                         list_add(&next->list, &hole->list);
 243                 }
 244         }
 245 out:
 246         pthread_mutex_unlock(&mgr->bo_va_mutex);
 247 }
 248
 249 int amdgpu_va_range_alloc(amdgpu_device_handle dev,
 250                           enum amdgpu_gpu_va_range va_range_type,
 251                           uint64_t size,
 252                           uint64_t va_base_alignment,
 253                           uint64_t va_base_required,
 254                           uint64_t *va_base_allocated,
 255                           amdgpu_va_handle *va_range_handle,
 256                           uint64_t flags)
 257 {
 258         va_base_alignment = MAX2(va_base_alignment, dev->vamgr->va_alignment);
 259         size = ALIGN(size, vamgr.va_alignment);
 260
 261         *va_base_allocated = amdgpu_vamgr_find_va(dev->vamgr, size,
 262                                         va_base_alignment, va_base_required);
 263
 264         if (*va_base_allocated != AMDGPU_INVALID_VA_ADDRESS) {
 265                 struct amdgpu_va* va;
 266                 va = calloc(1, sizeof(struct amdgpu_va));
 267                 if(!va){
 268                         amdgpu_vamgr_free_va(dev->vamgr, *va_base_allocated, size);
 269                         return -ENOMEM;
 270                 }
 271                 va->dev = dev;
 272                 va->address = *va_base_allocated;
 273                 va->size = size;
 274                 va->range = va_range_type;
 275                 *va_range_handle = va;
 276         } else {
 277                 return -EINVAL;
 278         }
 279
 280         return 0;
 281 }
 282
 283 int amdgpu_va_range_free(amdgpu_va_handle va_range_handle)
 284 {
 285         if(!va_range_handle || !va_range_handle->address)
 286                 return 0;
 287         amdgpu_vamgr_free_va(va_range_handle->dev->vamgr, va_range_handle->address,
 288                         va_range_handle->size);
 289         free(va_range_handle);
 290         return 0;
 291 }