| 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 |
#include <stdlib.h> |
| 25 |
#include <string.h> |
| 26 |
#include <errno.h> |
| 27 |
#include "amdgpu.h" |
| 28 |
#include "amdgpu_drm.h" |
| 29 |
#include "amdgpu_internal.h" |
| 30 |
#include "util_math.h" |
| 31 |
|
| 32 |
drm_public int amdgpu_va_range_query(amdgpu_device_handle dev, |
| 33 |
enum amdgpu_gpu_va_range type, |
| 34 |
uint64_t *start, uint64_t *end) |
| 35 |
{ |
| 36 |
if (type != amdgpu_gpu_va_range_general) |
| 37 |
return -EINVAL; |
| 38 |
|
| 39 |
*start = dev->dev_info.virtual_address_offset; |
| 40 |
*end = dev->dev_info.virtual_address_max; |
| 41 |
return 0; |
| 42 |
} |
| 43 |
|
| 44 |
drm_private void amdgpu_vamgr_init(struct amdgpu_bo_va_mgr *mgr, uint64_t start, |
| 45 |
uint64_t max, uint64_t alignment) |
| 46 |
{ |
| 47 |
struct amdgpu_bo_va_hole *n; |
| 48 |
|
| 49 |
mgr->va_max = max; |
| 50 |
mgr->va_alignment = alignment; |
| 51 |
|
| 52 |
list_inithead(&mgr->va_holes); |
| 53 |
pthread_mutex_init(&mgr->bo_va_mutex, NULL); |
| 54 |
pthread_mutex_lock(&mgr->bo_va_mutex); |
| 55 |
n = calloc(1, sizeof(struct amdgpu_bo_va_hole)); |
| 56 |
n->size = mgr->va_max - start; |
| 57 |
n->offset = start; |
| 58 |
list_add(&n->list, &mgr->va_holes); |
| 59 |
pthread_mutex_unlock(&mgr->bo_va_mutex); |
| 60 |
} |
| 61 |
|
| 62 |
drm_private void amdgpu_vamgr_deinit(struct amdgpu_bo_va_mgr *mgr) |
| 63 |
{ |
| 64 |
struct amdgpu_bo_va_hole *hole, *tmp; |
| 65 |
LIST_FOR_EACH_ENTRY_SAFE(hole, tmp, &mgr->va_holes, list) { |
| 66 |
list_del(&hole->list); |
| 67 |
free(hole); |
| 68 |
} |
| 69 |
pthread_mutex_destroy(&mgr->bo_va_mutex); |
| 70 |
} |
| 71 |
|
| 72 |
static drm_private int |
| 73 |
amdgpu_vamgr_subtract_hole(struct amdgpu_bo_va_hole *hole, uint64_t start_va, |
| 74 |
uint64_t end_va) |
| 75 |
{ |
| 76 |
if (start_va > hole->offset && end_va - hole->offset < hole->size) { |
| 77 |
struct amdgpu_bo_va_hole *n = calloc(1, sizeof(struct amdgpu_bo_va_hole)); |
| 78 |
if (!n) |
| 79 |
return -ENOMEM; |
| 80 |
|
| 81 |
n->size = start_va - hole->offset; |
| 82 |
n->offset = hole->offset; |
| 83 |
list_add(&n->list, &hole->list); |
| 84 |
|
| 85 |
hole->size -= (end_va - hole->offset); |
| 86 |
hole->offset = end_va; |
| 87 |
} else if (start_va > hole->offset) { |
| 88 |
hole->size = start_va - hole->offset; |
| 89 |
} else if (end_va - hole->offset < hole->size) { |
| 90 |
hole->size -= (end_va - hole->offset); |
| 91 |
hole->offset = end_va; |
| 92 |
} else { |
| 93 |
list_del(&hole->list); |
| 94 |
free(hole); |
| 95 |
} |
| 96 |
|
| 97 |
return 0; |
| 98 |
} |
| 99 |
|
| 100 |
static drm_private int |
| 101 |
amdgpu_vamgr_find_va(struct amdgpu_bo_va_mgr *mgr, uint64_t size, |
| 102 |
uint64_t alignment, uint64_t base_required, |
| 103 |
bool search_from_top, uint64_t *va_out) |
| 104 |
{ |
| 105 |
struct amdgpu_bo_va_hole *hole, *n; |
| 106 |
uint64_t offset = 0; |
| 107 |
int ret; |
| 108 |
|
| 109 |
|
| 110 |
alignment = MAX2(alignment, mgr->va_alignment); |
| 111 |
size = ALIGN(size, mgr->va_alignment); |
| 112 |
|
| 113 |
if (base_required % alignment) |
| 114 |
return -EINVAL; |
| 115 |
|
| 116 |
pthread_mutex_lock(&mgr->bo_va_mutex); |
| 117 |
if (!search_from_top) { |
| 118 |
LIST_FOR_EACH_ENTRY_SAFE_REV(hole, n, &mgr->va_holes, list) { |
| 119 |
if (base_required) { |
| 120 |
if (hole->offset > base_required || |
| 121 |
(hole->offset + hole->size) < (base_required + size)) |
| 122 |
continue; |
| 123 |
offset = base_required; |
| 124 |
} else { |
| 125 |
uint64_t waste = hole->offset % alignment; |
| 126 |
waste = waste ? alignment - waste : 0; |
| 127 |
offset = hole->offset + waste; |
| 128 |
if (offset >= (hole->offset + hole->size) || |
| 129 |
size > (hole->offset + hole->size) - offset) { |
| 130 |
continue; |
| 131 |
} |
| 132 |
} |
| 133 |
ret = amdgpu_vamgr_subtract_hole(hole, offset, offset + size); |
| 134 |
pthread_mutex_unlock(&mgr->bo_va_mutex); |
| 135 |
*va_out = offset; |
| 136 |
return ret; |
| 137 |
} |
| 138 |
} else { |
| 139 |
LIST_FOR_EACH_ENTRY_SAFE(hole, n, &mgr->va_holes, list) { |
| 140 |
if (base_required) { |
| 141 |
if (hole->offset > base_required || |
| 142 |
(hole->offset + hole->size) < (base_required + size)) |
| 143 |
continue; |
| 144 |
offset = base_required; |
| 145 |
} else { |
| 146 |
if (size > hole->size) |
| 147 |
continue; |
| 148 |
|
| 149 |
offset = hole->offset + hole->size - size; |
| 150 |
offset -= offset % alignment; |
| 151 |
if (offset < hole->offset) { |
| 152 |
continue; |
| 153 |
} |
| 154 |
} |
| 155 |
|
| 156 |
ret = amdgpu_vamgr_subtract_hole(hole, offset, offset + size); |
| 157 |
pthread_mutex_unlock(&mgr->bo_va_mutex); |
| 158 |
*va_out = offset; |
| 159 |
return ret; |
| 160 |
} |
| 161 |
} |
| 162 |
|
| 163 |
pthread_mutex_unlock(&mgr->bo_va_mutex); |
| 164 |
return -ENOMEM; |
| 165 |
} |
| 166 |
|
| 167 |
static drm_private void |
| 168 |
amdgpu_vamgr_free_va(struct amdgpu_bo_va_mgr *mgr, uint64_t va, uint64_t size) |
| 169 |
{ |
| 170 |
struct amdgpu_bo_va_hole *hole, *next; |
| 171 |
|
| 172 |
if (va == AMDGPU_INVALID_VA_ADDRESS) |
| 173 |
return; |
| 174 |
|
| 175 |
size = ALIGN(size, mgr->va_alignment); |
| 176 |
|
| 177 |
pthread_mutex_lock(&mgr->bo_va_mutex); |
| 178 |
hole = container_of(&mgr->va_holes, hole, list); |
| 179 |
LIST_FOR_EACH_ENTRY(next, &mgr->va_holes, list) { |
| 180 |
if (next->offset < va) |
| 181 |
break; |
| 182 |
hole = next; |
| 183 |
} |
| 184 |
|
| 185 |
if (&hole->list != &mgr->va_holes) { |
| 186 |
/* Grow upper hole if it's adjacent */ |
| 187 |
if (hole->offset == (va + size)) { |
| 188 |
hole->offset = va; |
| 189 |
hole->size += size; |
| 190 |
/* Merge lower hole if it's adjacent */ |
| 191 |
if (next != hole && |
| 192 |
&next->list != &mgr->va_holes && |
| 193 |
(next->offset + next->size) == va) { |
| 194 |
next->size += hole->size; |
| 195 |
list_del(&hole->list); |
| 196 |
free(hole); |
| 197 |
} |
| 198 |
goto out; |
| 199 |
} |
| 200 |
} |
| 201 |
|
| 202 |
/* Grow lower hole if it's adjacent */ |
| 203 |
if (next != hole && &next->list != &mgr->va_holes && |
| 204 |
(next->offset + next->size) == va) { |
| 205 |
next->size += size; |
| 206 |
goto out; |
| 207 |
} |
| 208 |
|
| 209 |
/* FIXME on allocation failure we just lose virtual address space |
| 210 |
* maybe print a warning |
| 211 |
*/ |
| 212 |
next = calloc(1, sizeof(struct amdgpu_bo_va_hole)); |
| 213 |
if (next) { |
| 214 |
next->size = size; |
| 215 |
next->offset = va; |
| 216 |
list_add(&next->list, &hole->list); |
| 217 |
} |
| 218 |
|
| 219 |
out: |
| 220 |
pthread_mutex_unlock(&mgr->bo_va_mutex); |
| 221 |
} |
| 222 |
|
| 223 |
drm_public int amdgpu_va_range_alloc(amdgpu_device_handle dev, |
| 224 |
enum amdgpu_gpu_va_range va_range_type, |
| 225 |
uint64_t size, |
| 226 |
uint64_t va_base_alignment, |
| 227 |
uint64_t va_base_required, |
| 228 |
uint64_t *va_base_allocated, |
| 229 |
amdgpu_va_handle *va_range_handle, |
| 230 |
uint64_t flags) |
| 231 |
{ |
| 232 |
return amdgpu_va_range_alloc2(&dev->va_mgr, va_range_type, size, |
| 233 |
va_base_alignment, va_base_required, |
| 234 |
va_base_allocated, va_range_handle, |
| 235 |
flags); |
| 236 |
} |
| 237 |
|
| 238 |
drm_public int amdgpu_va_range_alloc2(amdgpu_va_manager_handle va_mgr, |
| 239 |
enum amdgpu_gpu_va_range va_range_type, |
| 240 |
uint64_t size, |
| 241 |
uint64_t va_base_alignment, |
| 242 |
uint64_t va_base_required, |
| 243 |
uint64_t *va_base_allocated, |
| 244 |
amdgpu_va_handle *va_range_handle, |
| 245 |
uint64_t flags) |
| 246 |
{ |
| 247 |
struct amdgpu_bo_va_mgr *vamgr; |
| 248 |
bool search_from_top = !!(flags & AMDGPU_VA_RANGE_REPLAYABLE); |
| 249 |
int ret; |
| 250 |
|
| 251 |
/* Clear the flag when the high VA manager is not initialized */ |
| 252 |
if (flags & AMDGPU_VA_RANGE_HIGH && !va_mgr->vamgr_high_32.va_max) |
| 253 |
flags &= ~AMDGPU_VA_RANGE_HIGH; |
| 254 |
|
| 255 |
if (flags & AMDGPU_VA_RANGE_HIGH) { |
| 256 |
if (flags & AMDGPU_VA_RANGE_32_BIT) |
| 257 |
vamgr = &va_mgr->vamgr_high_32; |
| 258 |
else |
| 259 |
vamgr = &va_mgr->vamgr_high; |
| 260 |
} else { |
| 261 |
if (flags & AMDGPU_VA_RANGE_32_BIT) |
| 262 |
vamgr = &va_mgr->vamgr_32; |
| 263 |
else |
| 264 |
vamgr = &va_mgr->vamgr_low; |
| 265 |
} |
| 266 |
|
| 267 |
va_base_alignment = MAX2(va_base_alignment, vamgr->va_alignment); |
| 268 |
size = ALIGN(size, vamgr->va_alignment); |
| 269 |
|
| 270 |
ret = amdgpu_vamgr_find_va(vamgr, size, |
| 271 |
va_base_alignment, va_base_required, |
| 272 |
search_from_top, va_base_allocated); |
| 273 |
|
| 274 |
if (!(flags & AMDGPU_VA_RANGE_32_BIT) && ret) { |
| 275 |
/* fallback to 32bit address */ |
| 276 |
if (flags & AMDGPU_VA_RANGE_HIGH) |
| 277 |
vamgr = &va_mgr->vamgr_high_32; |
| 278 |
else |
| 279 |
vamgr = &va_mgr->vamgr_32; |
| 280 |
ret = amdgpu_vamgr_find_va(vamgr, size, |
| 281 |
va_base_alignment, va_base_required, |
| 282 |
search_from_top, va_base_allocated); |
| 283 |
} |
| 284 |
|
| 285 |
if (!ret) { |
| 286 |
struct amdgpu_va* va; |
| 287 |
va = calloc(1, sizeof(struct amdgpu_va)); |
| 288 |
if(!va){ |
| 289 |
amdgpu_vamgr_free_va(vamgr, *va_base_allocated, size); |
| 290 |
return -ENOMEM; |
| 291 |
} |
| 292 |
va->address = *va_base_allocated; |
| 293 |
va->size = size; |
| 294 |
va->range = va_range_type; |
| 295 |
va->vamgr = vamgr; |
| 296 |
*va_range_handle = va; |
| 297 |
} |
| 298 |
|
| 299 |
return ret; |
| 300 |
} |
| 301 |
|
| 302 |
drm_public int amdgpu_va_range_free(amdgpu_va_handle va_range_handle) |
| 303 |
{ |
| 304 |
if(!va_range_handle || !va_range_handle->address) |
| 305 |
return 0; |
| 306 |
|
| 307 |
amdgpu_vamgr_free_va(va_range_handle->vamgr, |
| 308 |
va_range_handle->address, |
| 309 |
va_range_handle->size); |
| 310 |
free(va_range_handle); |
| 311 |
return 0; |
| 312 |
} |
| 313 |
|
| 314 |
drm_public uint64_t amdgpu_va_get_start_addr(amdgpu_va_handle va_handle) |
| 315 |
{ |
| 316 |
return va_handle->address; |
| 317 |
} |
| 318 |
|
| 319 |
drm_public amdgpu_va_manager_handle amdgpu_va_manager_alloc(void) |
| 320 |
{ |
| 321 |
amdgpu_va_manager_handle r = calloc(1, sizeof(struct amdgpu_va_manager)); |
| 322 |
return r; |
| 323 |
} |
| 324 |
|
| 325 |
drm_public void amdgpu_va_manager_init(struct amdgpu_va_manager *va_mgr, |
| 326 |
uint64_t low_va_offset, uint64_t low_va_max, |
| 327 |
uint64_t high_va_offset, uint64_t high_va_max, |
| 328 |
uint32_t virtual_address_alignment) |
| 329 |
{ |
| 330 |
amdgpu_va_manager_init2(va_mgr, low_va_offset, low_va_max, |
| 331 |
high_va_offset, high_va_max, |
| 332 |
virtual_address_alignment, 0); |
| 333 |
} |
| 334 |
|
| 335 |
drm_public void amdgpu_va_manager_init2(struct amdgpu_va_manager *va_mgr, |
| 336 |
uint64_t low_va_offset, uint64_t low_va_max, |
| 337 |
uint64_t high_va_offset, uint64_t high_va_max, |
| 338 |
uint32_t virtual_address_alignment, |
| 339 |
uint32_t flags) |
| 340 |
{ |
| 341 |
uint64_t start, max; |
| 342 |
|
| 343 |
va_mgr->address_prt_wa_control_bit = ~0; |
| 344 |
|
| 345 |
start = low_va_offset; |
| 346 |
max = MIN2(low_va_max, 0x100000000ULL); |
| 347 |
amdgpu_vamgr_init(&va_mgr->vamgr_32, start, max, |
| 348 |
virtual_address_alignment); |
| 349 |
|
| 350 |
start = max; |
| 351 |
if ((flags & AMDGPU_VA_MGR_RESERVE_HALF_VA_FOR_PRT) && !high_va_max) { |
| 352 |
/* Reserve the half VA range for PRT by splitting it in two |
| 353 |
* equal halves where one bit controls whether it's the LOW or |
| 354 |
* HIGH half. |
| 355 |
*/ |
| 356 |
va_mgr->address_prt_wa_control_bit = util_last_bit64(low_va_offset ^ low_va_max) - 1; |
| 357 |
max = low_va_max ^ (1ull << va_mgr->address_prt_wa_control_bit); |
| 358 |
} else { |
| 359 |
max = MAX2(low_va_max, 0x100000000ULL); |
| 360 |
} |
| 361 |
|
| 362 |
amdgpu_vamgr_init(&va_mgr->vamgr_low, start, max, |
| 363 |
virtual_address_alignment); |
| 364 |
|
| 365 |
start = high_va_offset; |
| 366 |
max = MIN2(high_va_max, (start & ~0xffffffffULL) + 0x100000000ULL); |
| 367 |
amdgpu_vamgr_init(&va_mgr->vamgr_high_32, start, max, |
| 368 |
virtual_address_alignment); |
| 369 |
|
| 370 |
start = max; |
| 371 |
if ((flags & AMDGPU_VA_MGR_RESERVE_HALF_VA_FOR_PRT) && high_va_max) { |
| 372 |
/* Reserve the half VA range for PRT by splitting it in two |
| 373 |
* equal halves where one bit controls whether it's the LOW or |
| 374 |
* HIGH half. |
| 375 |
*/ |
| 376 |
va_mgr->address_prt_wa_control_bit = util_last_bit64(high_va_offset ^ high_va_max) - 1; |
| 377 |
max = high_va_max ^ (1ull << va_mgr->address_prt_wa_control_bit); |
| 378 |
} else { |
| 379 |
max = MAX2(high_va_max, (start & ~0xffffffffULL) + 0x100000000ULL); |
| 380 |
} |
| 381 |
|
| 382 |
amdgpu_vamgr_init(&va_mgr->vamgr_high, start, max, |
| 383 |
virtual_address_alignment); |
| 384 |
} |
| 385 |
|
| 386 |
drm_public void amdgpu_va_manager_deinit(struct amdgpu_va_manager *va_mgr) |
| 387 |
{ |
| 388 |
amdgpu_vamgr_deinit(&va_mgr->vamgr_32); |
| 389 |
amdgpu_vamgr_deinit(&va_mgr->vamgr_low); |
| 390 |
amdgpu_vamgr_deinit(&va_mgr->vamgr_high_32); |
| 391 |
amdgpu_vamgr_deinit(&va_mgr->vamgr_high); |
| 392 |
} |
| 393 |
|
| 394 |
drm_public int amdgpu_va_manager_query_sw_info(struct amdgpu_va_manager *va_mgr, |
| 395 |
enum amdgpu_va_manager_sw_info info, |
| 396 |
void *value) |
| 397 |
{ |
| 398 |
uint32_t *val32 = (uint32_t*)value; |
| 399 |
|
| 400 |
switch (info) { |
| 401 |
case amdgpu_va_manager_sw_info_address_prt_wa_control_bit: |
| 402 |
*val32 = va_mgr->address_prt_wa_control_bit; |
| 403 |
return 0; |
| 404 |
} |
| 405 |
return -EINVAL; |
| 406 |
} |