| 1 |
/* |
| 2 |
* Copyright © 2014 Broadcom |
| 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 (including the next |
| 12 |
* paragraph) shall be included in all copies or substantial portions of the |
| 13 |
* Software. |
| 14 |
* |
| 15 |
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 16 |
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 17 |
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL |
| 18 |
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 19 |
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING |
| 20 |
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS |
| 21 |
* IN THE SOFTWARE. |
| 22 |
*/ |
| 23 |
|
| 24 |
/** |
| 25 |
* Command list validator for VC4. |
| 26 |
* |
| 27 |
* The VC4 has no IOMMU between it and system memory. So, a user with |
| 28 |
* access to execute command lists could escalate privilege by |
| 29 |
* overwriting system memory (drawing to it as a framebuffer) or |
| 30 |
* reading system memory it shouldn't (reading it as a texture, or |
| 31 |
* uniform data, or vertex data). |
| 32 |
* |
| 33 |
* This validates command lists to ensure that all accesses are within |
| 34 |
* the bounds of the GEM objects referenced. It explicitly whitelists |
| 35 |
* packets, and looks at the offsets in any address fields to make |
| 36 |
* sure they're constrained within the BOs they reference. |
| 37 |
* |
| 38 |
* Note that because of the validation that's happening anyway, this |
| 39 |
* is where GEM relocation processing happens. |
| 40 |
*/ |
| 41 |
|
| 42 |
#include "vc4_drv.h" |
| 43 |
#include "vc4_packet.h" |
| 44 |
|
| 45 |
#define VALIDATE_ARGS \ |
| 46 |
struct vc4_exec_info *exec, \ |
| 47 |
void *validated, \ |
| 48 |
void *untrusted |
| 49 |
|
| 50 |
/** Return the width in pixels of a 64-byte microtile. */ |
| 51 |
static uint32_t |
| 52 |
utile_width(int cpp) |
| 53 |
{ |
| 54 |
switch (cpp) { |
| 55 |
case 1: |
| 56 |
case 2: |
| 57 |
return 8; |
| 58 |
case 4: |
| 59 |
return 4; |
| 60 |
case 8: |
| 61 |
return 2; |
| 62 |
default: |
| 63 |
DRM_ERROR("unknown cpp: %d\n", cpp); |
| 64 |
return 1; |
| 65 |
} |
| 66 |
} |
| 67 |
|
| 68 |
/** Return the height in pixels of a 64-byte microtile. */ |
| 69 |
static uint32_t |
| 70 |
utile_height(int cpp) |
| 71 |
{ |
| 72 |
switch (cpp) { |
| 73 |
case 1: |
| 74 |
return 8; |
| 75 |
case 2: |
| 76 |
case 4: |
| 77 |
case 8: |
| 78 |
return 4; |
| 79 |
default: |
| 80 |
DRM_ERROR("unknown cpp: %d\n", cpp); |
| 81 |
return 1; |
| 82 |
} |
| 83 |
} |
| 84 |
|
| 85 |
/** |
| 86 |
* The texture unit decides what tiling format a particular miplevel is using |
| 87 |
* this function, so we lay out our miptrees accordingly. |
| 88 |
*/ |
| 89 |
static bool |
| 90 |
size_is_lt(uint32_t width, uint32_t height, int cpp) |
| 91 |
{ |
| 92 |
return (width <= 4 * utile_width(cpp) || |
| 93 |
height <= 4 * utile_height(cpp)); |
| 94 |
} |
| 95 |
|
| 96 |
struct drm_gem_cma_object * |
| 97 |
vc4_use_bo(struct vc4_exec_info *exec, uint32_t hindex) |
| 98 |
{ |
| 99 |
struct drm_gem_cma_object *obj; |
| 100 |
struct drm_vc4_bo *bo; |
| 101 |
|
| 102 |
if (hindex >= exec->bo_count) { |
| 103 |
DRM_ERROR("BO index %d greater than BO count %d\n", |
| 104 |
hindex, exec->bo_count); |
| 105 |
return NULL; |
| 106 |
} |
| 107 |
obj = exec->bo[hindex]; |
| 108 |
bo = to_vc4_bo(&obj->base); |
| 109 |
|
| 110 |
if (bo->validated_shader) { |
| 111 |
DRM_ERROR("Trying to use shader BO as something other than " |
| 112 |
"a shader\n"); |
| 113 |
return NULL; |
| 114 |
} |
| 115 |
|
| 116 |
return obj; |
| 117 |
} |
| 118 |
|
| 119 |
static struct drm_gem_cma_object * |
| 120 |
vc4_use_handle(struct vc4_exec_info *exec, uint32_t gem_handles_packet_index) |
| 121 |
{ |
| 122 |
return vc4_use_bo(exec, exec->bo_index[gem_handles_packet_index]); |
| 123 |
} |
| 124 |
|
| 125 |
static bool |
| 126 |
validate_bin_pos(struct vc4_exec_info *exec, void *untrusted, uint32_t pos) |
| 127 |
{ |
| 128 |
/* Note that the untrusted pointer passed to these functions is |
| 129 |
* incremented past the packet byte. |
| 130 |
*/ |
| 131 |
return (untrusted - 1 == exec->bin_u + pos); |
| 132 |
} |
| 133 |
|
| 134 |
static uint32_t |
| 135 |
gl_shader_rec_size(uint32_t pointer_bits) |
| 136 |
{ |
| 137 |
uint32_t attribute_count = pointer_bits & 7; |
| 138 |
bool extended = pointer_bits & 8; |
| 139 |
|
| 140 |
if (attribute_count == 0) |
| 141 |
attribute_count = 8; |
| 142 |
|
| 143 |
if (extended) |
| 144 |
return 100 + attribute_count * 4; |
| 145 |
else |
| 146 |
return 36 + attribute_count * 8; |
| 147 |
} |
| 148 |
|
| 149 |
bool |
| 150 |
vc4_check_tex_size(struct vc4_exec_info *exec, struct drm_gem_cma_object *fbo, |
| 151 |
uint32_t offset, uint8_t tiling_format, |
| 152 |
uint32_t width, uint32_t height, uint8_t cpp) |
| 153 |
{ |
| 154 |
uint32_t aligned_width, aligned_height, stride, size; |
| 155 |
uint32_t utile_w = utile_width(cpp); |
| 156 |
uint32_t utile_h = utile_height(cpp); |
| 157 |
|
| 158 |
/* The shaded vertex format stores signed 12.4 fixed point |
| 159 |
* (-2048,2047) offsets from the viewport center, so we should |
| 160 |
* never have a render target larger than 4096. The texture |
| 161 |
* unit can only sample from 2048x2048, so it's even more |
| 162 |
* restricted. This lets us avoid worrying about overflow in |
| 163 |
* our math. |
| 164 |
*/ |
| 165 |
if (width > 4096 || height > 4096) { |
| 166 |
DRM_ERROR("Surface dimensions (%d,%d) too large", width, height); |
| 167 |
return false; |
| 168 |
} |
| 169 |
|
| 170 |
switch (tiling_format) { |
| 171 |
case VC4_TILING_FORMAT_LINEAR: |
| 172 |
aligned_width = round_up(width, utile_w); |
| 173 |
aligned_height = height; |
| 174 |
break; |
| 175 |
case VC4_TILING_FORMAT_T: |
| 176 |
aligned_width = round_up(width, utile_w * 8); |
| 177 |
aligned_height = round_up(height, utile_h * 8); |
| 178 |
break; |
| 179 |
case VC4_TILING_FORMAT_LT: |
| 180 |
aligned_width = round_up(width, utile_w); |
| 181 |
aligned_height = round_up(height, utile_h); |
| 182 |
break; |
| 183 |
default: |
| 184 |
DRM_ERROR("buffer tiling %d unsupported\n", tiling_format); |
| 185 |
return false; |
| 186 |
} |
| 187 |
|
| 188 |
stride = aligned_width * cpp; |
| 189 |
size = stride * aligned_height; |
| 190 |
|
| 191 |
if (size + offset < size || |
| 192 |
size + offset > fbo->base.size) { |
| 193 |
DRM_ERROR("Overflow in %dx%d (%dx%d) fbo size (%d + %d > %zd)\n", |
| 194 |
width, height, |
| 195 |
aligned_width, aligned_height, |
| 196 |
size, offset, fbo->base.size); |
| 197 |
return false; |
| 198 |
} |
| 199 |
|
| 200 |
return true; |
| 201 |
} |
| 202 |
|
| 203 |
static int |
| 204 |
validate_flush(VALIDATE_ARGS) |
| 205 |
{ |
| 206 |
if (!validate_bin_pos(exec, untrusted, exec->args->bin_cl_size - 1)) { |
| 207 |
DRM_ERROR("Bin CL must end with VC4_PACKET_FLUSH\n"); |
| 208 |
return -EINVAL; |
| 209 |
} |
| 210 |
exec->found_flush = true; |
| 211 |
|
| 212 |
return 0; |
| 213 |
} |
| 214 |
|
| 215 |
static int |
| 216 |
validate_start_tile_binning(VALIDATE_ARGS) |
| 217 |
{ |
| 218 |
if (exec->found_start_tile_binning_packet) { |
| 219 |
DRM_ERROR("Duplicate VC4_PACKET_START_TILE_BINNING\n"); |
| 220 |
return -EINVAL; |
| 221 |
} |
| 222 |
exec->found_start_tile_binning_packet = true; |
| 223 |
|
| 224 |
if (!exec->found_tile_binning_mode_config_packet) { |
| 225 |
DRM_ERROR("missing VC4_PACKET_TILE_BINNING_MODE_CONFIG\n"); |
| 226 |
return -EINVAL; |
| 227 |
} |
| 228 |
|
| 229 |
return 0; |
| 230 |
} |
| 231 |
|
| 232 |
static int |
| 233 |
validate_increment_semaphore(VALIDATE_ARGS) |
| 234 |
{ |
| 235 |
if (!validate_bin_pos(exec, untrusted, exec->args->bin_cl_size - 2)) { |
| 236 |
DRM_ERROR("Bin CL must end with " |
| 237 |
"VC4_PACKET_INCREMENT_SEMAPHORE\n"); |
| 238 |
return -EINVAL; |
| 239 |
} |
| 240 |
exec->found_increment_semaphore_packet = true; |
| 241 |
|
| 242 |
return 0; |
| 243 |
} |
| 244 |
|
| 245 |
static int |
| 246 |
validate_indexed_prim_list(VALIDATE_ARGS) |
| 247 |
{ |
| 248 |
struct drm_gem_cma_object *ib; |
| 249 |
uint32_t length = *(uint32_t *)(untrusted + 1); |
| 250 |
uint32_t offset = *(uint32_t *)(untrusted + 5); |
| 251 |
uint32_t max_index = *(uint32_t *)(untrusted + 9); |
| 252 |
uint32_t index_size = (*(uint8_t *)(untrusted + 0) >> 4) ? 2 : 1; |
| 253 |
struct vc4_shader_state *shader_state; |
| 254 |
|
| 255 |
/* Check overflow condition */ |
| 256 |
if (exec->shader_state_count == 0) { |
| 257 |
DRM_ERROR("shader state must precede primitives\n"); |
| 258 |
return -EINVAL; |
| 259 |
} |
| 260 |
shader_state = &exec->shader_state[exec->shader_state_count - 1]; |
| 261 |
|
| 262 |
if (max_index > shader_state->max_index) |
| 263 |
shader_state->max_index = max_index; |
| 264 |
|
| 265 |
ib = vc4_use_handle(exec, 0); |
| 266 |
if (!ib) |
| 267 |
return -EINVAL; |
| 268 |
|
| 269 |
if (offset > ib->base.size || |
| 270 |
(ib->base.size - offset) / index_size < length) { |
| 271 |
DRM_ERROR("IB access overflow (%d + %d*%d > %zd)\n", |
| 272 |
offset, length, index_size, ib->base.size); |
| 273 |
return -EINVAL; |
| 274 |
} |
| 275 |
|
| 276 |
*(uint32_t *)(validated + 5) = ib->paddr + offset; |
| 277 |
|
| 278 |
return 0; |
| 279 |
} |
| 280 |
|
| 281 |
static int |
| 282 |
validate_gl_array_primitive(VALIDATE_ARGS) |
| 283 |
{ |
| 284 |
uint32_t length = *(uint32_t *)(untrusted + 1); |
| 285 |
uint32_t base_index = *(uint32_t *)(untrusted + 5); |
| 286 |
uint32_t max_index; |
| 287 |
struct vc4_shader_state *shader_state; |
| 288 |
|
| 289 |
/* Check overflow condition */ |
| 290 |
if (exec->shader_state_count == 0) { |
| 291 |
DRM_ERROR("shader state must precede primitives\n"); |
| 292 |
return -EINVAL; |
| 293 |
} |
| 294 |
shader_state = &exec->shader_state[exec->shader_state_count - 1]; |
| 295 |
|
| 296 |
if (length + base_index < length) { |
| 297 |
DRM_ERROR("primitive vertex count overflow\n"); |
| 298 |
return -EINVAL; |
| 299 |
} |
| 300 |
max_index = length + base_index - 1; |
| 301 |
|
| 302 |
if (max_index > shader_state->max_index) |
| 303 |
shader_state->max_index = max_index; |
| 304 |
|
| 305 |
return 0; |
| 306 |
} |
| 307 |
|
| 308 |
static int |
| 309 |
validate_gl_shader_state(VALIDATE_ARGS) |
| 310 |
{ |
| 311 |
uint32_t i = exec->shader_state_count++; |
| 312 |
|
| 313 |
if (i >= exec->shader_state_size) { |
| 314 |
DRM_ERROR("More requests for shader states than declared\n"); |
| 315 |
return -EINVAL; |
| 316 |
} |
| 317 |
|
| 318 |
exec->shader_state[i].addr = *(uint32_t *)untrusted; |
| 319 |
exec->shader_state[i].max_index = 0; |
| 320 |
|
| 321 |
if (exec->shader_state[i].addr & ~0xf) { |
| 322 |
DRM_ERROR("high bits set in GL shader rec reference\n"); |
| 323 |
return -EINVAL; |
| 324 |
} |
| 325 |
|
| 326 |
*(uint32_t *)validated = (exec->shader_rec_p + |
| 327 |
exec->shader_state[i].addr); |
| 328 |
|
| 329 |
exec->shader_rec_p += |
| 330 |
roundup(gl_shader_rec_size(exec->shader_state[i].addr), 16); |
| 331 |
|
| 332 |
return 0; |
| 333 |
} |
| 334 |
|
| 335 |
static int |
| 336 |
validate_tile_binning_config(VALIDATE_ARGS) |
| 337 |
{ |
| 338 |
struct drm_device *dev = exec->exec_bo->base.dev; |
| 339 |
uint8_t flags; |
| 340 |
uint32_t tile_state_size, tile_alloc_size; |
| 341 |
uint32_t tile_count; |
| 342 |
|
| 343 |
if (exec->found_tile_binning_mode_config_packet) { |
| 344 |
DRM_ERROR("Duplicate VC4_PACKET_TILE_BINNING_MODE_CONFIG\n"); |
| 345 |
return -EINVAL; |
| 346 |
} |
| 347 |
exec->found_tile_binning_mode_config_packet = true; |
| 348 |
|
| 349 |
exec->bin_tiles_x = *(uint8_t *)(untrusted + 12); |
| 350 |
exec->bin_tiles_y = *(uint8_t *)(untrusted + 13); |
| 351 |
tile_count = exec->bin_tiles_x * exec->bin_tiles_y; |
| 352 |
flags = *(uint8_t *)(untrusted + 14); |
| 353 |
|
| 354 |
if (exec->bin_tiles_x == 0 || |
| 355 |
exec->bin_tiles_y == 0) { |
| 356 |
DRM_ERROR("Tile binning config of %dx%d too small\n", |
| 357 |
exec->bin_tiles_x, exec->bin_tiles_y); |
| 358 |
return -EINVAL; |
| 359 |
} |
| 360 |
|
| 361 |
if (flags & (VC4_BIN_CONFIG_DB_NON_MS | |
| 362 |
VC4_BIN_CONFIG_TILE_BUFFER_64BIT)) { |
| 363 |
DRM_ERROR("unsupported binning config flags 0x%02x\n", flags); |
| 364 |
return -EINVAL; |
| 365 |
} |
| 366 |
|
| 367 |
/* The tile state data array is 48 bytes per tile, and we put it at |
| 368 |
* the start of a BO containing both it and the tile alloc. |
| 369 |
*/ |
| 370 |
tile_state_size = 48 * tile_count; |
| 371 |
|
| 372 |
/* Since the tile alloc array will follow us, align. */ |
| 373 |
exec->tile_alloc_offset = roundup(tile_state_size, 4096); |
| 374 |
|
| 375 |
*(uint8_t *)(validated + 14) = |
| 376 |
((flags & ~(VC4_BIN_CONFIG_ALLOC_INIT_BLOCK_SIZE_MASK | |
| 377 |
VC4_BIN_CONFIG_ALLOC_BLOCK_SIZE_MASK)) | |
| 378 |
VC4_BIN_CONFIG_AUTO_INIT_TSDA | |
| 379 |
VC4_SET_FIELD(VC4_BIN_CONFIG_ALLOC_INIT_BLOCK_SIZE_32, |
| 380 |
VC4_BIN_CONFIG_ALLOC_INIT_BLOCK_SIZE) | |
| 381 |
VC4_SET_FIELD(VC4_BIN_CONFIG_ALLOC_BLOCK_SIZE_128, |
| 382 |
VC4_BIN_CONFIG_ALLOC_BLOCK_SIZE)); |
| 383 |
|
| 384 |
/* Initial block size. */ |
| 385 |
tile_alloc_size = 32 * tile_count; |
| 386 |
|
| 387 |
/* |
| 388 |
* The initial allocation gets rounded to the next 256 bytes before |
| 389 |
* the hardware starts fulfilling further allocations. |
| 390 |
*/ |
| 391 |
tile_alloc_size = roundup(tile_alloc_size, 256); |
| 392 |
|
| 393 |
/* Add space for the extra allocations. This is what gets used first, |
| 394 |
* before overflow memory. It must have at least 4096 bytes, but we |
| 395 |
* want to avoid overflow memory usage if possible. |
| 396 |
*/ |
| 397 |
tile_alloc_size += 1024 * 1024; |
| 398 |
|
| 399 |
exec->tile_bo = drm_gem_cma_create(dev, exec->tile_alloc_offset + |
| 400 |
tile_alloc_size); |
| 401 |
if (!exec->tile_bo) |
| 402 |
return -ENOMEM; |
| 403 |
list_addtail(&to_vc4_bo(&exec->tile_bo->base)->unref_head, |
| 404 |
&exec->unref_list); |
| 405 |
|
| 406 |
/* tile alloc address. */ |
| 407 |
*(uint32_t *)(validated + 0) = (exec->tile_bo->paddr + |
| 408 |
exec->tile_alloc_offset); |
| 409 |
/* tile alloc size. */ |
| 410 |
*(uint32_t *)(validated + 4) = tile_alloc_size; |
| 411 |
/* tile state address. */ |
| 412 |
*(uint32_t *)(validated + 8) = exec->tile_bo->paddr; |
| 413 |
|
| 414 |
return 0; |
| 415 |
} |
| 416 |
|
| 417 |
static int |
| 418 |
validate_gem_handles(VALIDATE_ARGS) |
| 419 |
{ |
| 420 |
memcpy(exec->bo_index, untrusted, sizeof(exec->bo_index)); |
| 421 |
return 0; |
| 422 |
} |
| 423 |
|
| 424 |
#define VC4_DEFINE_PACKET(packet, func) \ |
| 425 |
[packet] = { packet ## _SIZE, #packet, func } |
| 426 |
|
| 427 |
static const struct cmd_info { |
| 428 |
uint16_t len; |
| 429 |
const char *name; |
| 430 |
int (*func)(struct vc4_exec_info *exec, void *validated, |
| 431 |
void *untrusted); |
| 432 |
} cmd_info[] = { |
| 433 |
VC4_DEFINE_PACKET(VC4_PACKET_HALT, NULL), |
| 434 |
VC4_DEFINE_PACKET(VC4_PACKET_NOP, NULL), |
| 435 |
VC4_DEFINE_PACKET(VC4_PACKET_FLUSH, validate_flush), |
| 436 |
VC4_DEFINE_PACKET(VC4_PACKET_FLUSH_ALL, NULL), |
| 437 |
VC4_DEFINE_PACKET(VC4_PACKET_START_TILE_BINNING, |
| 438 |
validate_start_tile_binning), |
| 439 |
VC4_DEFINE_PACKET(VC4_PACKET_INCREMENT_SEMAPHORE, |
| 440 |
validate_increment_semaphore), |
| 441 |
|
| 442 |
VC4_DEFINE_PACKET(VC4_PACKET_GL_INDEXED_PRIMITIVE, |
| 443 |
validate_indexed_prim_list), |
| 444 |
VC4_DEFINE_PACKET(VC4_PACKET_GL_ARRAY_PRIMITIVE, |
| 445 |
validate_gl_array_primitive), |
| 446 |
|
| 447 |
VC4_DEFINE_PACKET(VC4_PACKET_PRIMITIVE_LIST_FORMAT, NULL), |
| 448 |
|
| 449 |
VC4_DEFINE_PACKET(VC4_PACKET_GL_SHADER_STATE, validate_gl_shader_state), |
| 450 |
|
| 451 |
VC4_DEFINE_PACKET(VC4_PACKET_CONFIGURATION_BITS, NULL), |
| 452 |
VC4_DEFINE_PACKET(VC4_PACKET_FLAT_SHADE_FLAGS, NULL), |
| 453 |
VC4_DEFINE_PACKET(VC4_PACKET_POINT_SIZE, NULL), |
| 454 |
VC4_DEFINE_PACKET(VC4_PACKET_LINE_WIDTH, NULL), |
| 455 |
VC4_DEFINE_PACKET(VC4_PACKET_RHT_X_BOUNDARY, NULL), |
| 456 |
VC4_DEFINE_PACKET(VC4_PACKET_DEPTH_OFFSET, NULL), |
| 457 |
VC4_DEFINE_PACKET(VC4_PACKET_CLIP_WINDOW, NULL), |
| 458 |
VC4_DEFINE_PACKET(VC4_PACKET_VIEWPORT_OFFSET, NULL), |
| 459 |
VC4_DEFINE_PACKET(VC4_PACKET_CLIPPER_XY_SCALING, NULL), |
| 460 |
/* Note: The docs say this was also 105, but it was 106 in the |
| 461 |
* initial userland code drop. |
| 462 |
*/ |
| 463 |
VC4_DEFINE_PACKET(VC4_PACKET_CLIPPER_Z_SCALING, NULL), |
| 464 |
|
| 465 |
VC4_DEFINE_PACKET(VC4_PACKET_TILE_BINNING_MODE_CONFIG, |
| 466 |
validate_tile_binning_config), |
| 467 |
|
| 468 |
VC4_DEFINE_PACKET(VC4_PACKET_GEM_HANDLES, validate_gem_handles), |
| 469 |
}; |
| 470 |
|
| 471 |
int |
| 472 |
vc4_validate_bin_cl(struct drm_device *dev, |
| 473 |
void *validated, |
| 474 |
void *unvalidated, |
| 475 |
struct vc4_exec_info *exec) |
| 476 |
{ |
| 477 |
uint32_t len = exec->args->bin_cl_size; |
| 478 |
uint32_t dst_offset = 0; |
| 479 |
uint32_t src_offset = 0; |
| 480 |
|
| 481 |
while (src_offset < len) { |
| 482 |
void *dst_pkt = validated + dst_offset; |
| 483 |
void *src_pkt = unvalidated + src_offset; |
| 484 |
u8 cmd = *(uint8_t *)src_pkt; |
| 485 |
const struct cmd_info *info; |
| 486 |
|
| 487 |
if (cmd >= ARRAY_SIZE(cmd_info)) { |
| 488 |
DRM_ERROR("0x%08x: packet %d out of bounds\n", |
| 489 |
src_offset, cmd); |
| 490 |
return -EINVAL; |
| 491 |
} |
| 492 |
|
| 493 |
info = &cmd_info[cmd]; |
| 494 |
if (!info->name) { |
| 495 |
DRM_ERROR("0x%08x: packet %d invalid\n", |
| 496 |
src_offset, cmd); |
| 497 |
return -EINVAL; |
| 498 |
} |
| 499 |
|
| 500 |
if (src_offset + info->len > len) { |
| 501 |
DRM_ERROR("0x%08x: packet %d (%s) length 0x%08x " |
| 502 |
"exceeds bounds (0x%08x)\n", |
| 503 |
src_offset, cmd, info->name, info->len, |
| 504 |
src_offset + len); |
| 505 |
return -EINVAL; |
| 506 |
} |
| 507 |
|
| 508 |
if (cmd != VC4_PACKET_GEM_HANDLES) |
| 509 |
memcpy(dst_pkt, src_pkt, info->len); |
| 510 |
|
| 511 |
if (info->func && info->func(exec, |
| 512 |
dst_pkt + 1, |
| 513 |
src_pkt + 1)) { |
| 514 |
DRM_ERROR("0x%08x: packet %d (%s) failed to validate\n", |
| 515 |
src_offset, cmd, info->name); |
| 516 |
return -EINVAL; |
| 517 |
} |
| 518 |
|
| 519 |
src_offset += info->len; |
| 520 |
/* GEM handle loading doesn't produce HW packets. */ |
| 521 |
if (cmd != VC4_PACKET_GEM_HANDLES) |
| 522 |
dst_offset += info->len; |
| 523 |
|
| 524 |
/* When the CL hits halt, it'll stop reading anything else. */ |
| 525 |
if (cmd == VC4_PACKET_HALT) |
| 526 |
break; |
| 527 |
} |
| 528 |
|
| 529 |
exec->ct0ea = exec->ct0ca + dst_offset; |
| 530 |
|
| 531 |
if (!exec->found_start_tile_binning_packet) { |
| 532 |
DRM_ERROR("Bin CL missing VC4_PACKET_START_TILE_BINNING\n"); |
| 533 |
return -EINVAL; |
| 534 |
} |
| 535 |
|
| 536 |
/* The bin CL must be ended with INCREMENT_SEMAPHORE and FLUSH. The |
| 537 |
* semaphore is used to trigger the render CL to start up, and the |
| 538 |
* FLUSH is what caps the bin lists with |
| 539 |
* VC4_PACKET_RETURN_FROM_SUB_LIST (so they jump back to the main |
| 540 |
* render CL when they get called to) and actually triggers the queued |
| 541 |
* semaphore increment. |
| 542 |
*/ |
| 543 |
if (!exec->found_increment_semaphore_packet || !exec->found_flush) { |
| 544 |
DRM_ERROR("Bin CL missing VC4_PACKET_INCREMENT_SEMAPHORE + " |
| 545 |
"VC4_PACKET_FLUSH\n"); |
| 546 |
return -EINVAL; |
| 547 |
} |
| 548 |
|
| 549 |
return 0; |
| 550 |
} |
| 551 |
|
| 552 |
static bool |
| 553 |
reloc_tex(struct vc4_exec_info *exec, |
| 554 |
void *uniform_data_u, |
| 555 |
struct vc4_texture_sample_info *sample, |
| 556 |
uint32_t texture_handle_index) |
| 557 |
|
| 558 |
{ |
| 559 |
struct drm_gem_cma_object *tex; |
| 560 |
uint32_t p0 = *(uint32_t *)(uniform_data_u + sample->p_offset[0]); |
| 561 |
uint32_t p1 = *(uint32_t *)(uniform_data_u + sample->p_offset[1]); |
| 562 |
uint32_t p2 = (sample->p_offset[2] != ~0 ? |
| 563 |
*(uint32_t *)(uniform_data_u + sample->p_offset[2]) : 0); |
| 564 |
uint32_t p3 = (sample->p_offset[3] != ~0 ? |
| 565 |
*(uint32_t *)(uniform_data_u + sample->p_offset[3]) : 0); |
| 566 |
uint32_t *validated_p0 = exec->uniforms_v + sample->p_offset[0]; |
| 567 |
uint32_t offset = p0 & VC4_TEX_P0_OFFSET_MASK; |
| 568 |
uint32_t miplevels = VC4_GET_FIELD(p0, VC4_TEX_P0_MIPLVLS); |
| 569 |
uint32_t width = VC4_GET_FIELD(p1, VC4_TEX_P1_WIDTH); |
| 570 |
uint32_t height = VC4_GET_FIELD(p1, VC4_TEX_P1_HEIGHT); |
| 571 |
uint32_t cpp, tiling_format, utile_w, utile_h; |
| 572 |
uint32_t i; |
| 573 |
uint32_t cube_map_stride = 0; |
| 574 |
enum vc4_texture_data_type type; |
| 575 |
|
| 576 |
tex = vc4_use_bo(exec, texture_handle_index); |
| 577 |
if (!tex) |
| 578 |
return false; |
| 579 |
|
| 580 |
if (sample->is_direct) { |
| 581 |
uint32_t remaining_size = tex->base.size - p0; |
| 582 |
|
| 583 |
if (p0 > tex->base.size - 4) { |
| 584 |
DRM_ERROR("UBO offset greater than UBO size\n"); |
| 585 |
goto fail; |
| 586 |
} |
| 587 |
if (p1 > remaining_size - 4) { |
| 588 |
DRM_ERROR("UBO clamp would allow reads " |
| 589 |
"outside of UBO\n"); |
| 590 |
goto fail; |
| 591 |
} |
| 592 |
*validated_p0 = tex->paddr + p0; |
| 593 |
return true; |
| 594 |
} |
| 595 |
|
| 596 |
if (width == 0) |
| 597 |
width = 2048; |
| 598 |
if (height == 0) |
| 599 |
height = 2048; |
| 600 |
|
| 601 |
if (p0 & VC4_TEX_P0_CMMODE_MASK) { |
| 602 |
if (VC4_GET_FIELD(p2, VC4_TEX_P2_PTYPE) == |
| 603 |
VC4_TEX_P2_PTYPE_CUBE_MAP_STRIDE) |
| 604 |
cube_map_stride = p2 & VC4_TEX_P2_CMST_MASK; |
| 605 |
if (VC4_GET_FIELD(p3, VC4_TEX_P2_PTYPE) == |
| 606 |
VC4_TEX_P2_PTYPE_CUBE_MAP_STRIDE) { |
| 607 |
if (cube_map_stride) { |
| 608 |
DRM_ERROR("Cube map stride set twice\n"); |
| 609 |
goto fail; |
| 610 |
} |
| 611 |
|
| 612 |
cube_map_stride = p3 & VC4_TEX_P2_CMST_MASK; |
| 613 |
} |
| 614 |
if (!cube_map_stride) { |
| 615 |
DRM_ERROR("Cube map stride not set\n"); |
| 616 |
goto fail; |
| 617 |
} |
| 618 |
} |
| 619 |
|
| 620 |
type = (VC4_GET_FIELD(p0, VC4_TEX_P0_TYPE) | |
| 621 |
(VC4_GET_FIELD(p1, VC4_TEX_P1_TYPE4) << 4)); |
| 622 |
|
| 623 |
switch (type) { |
| 624 |
case VC4_TEXTURE_TYPE_RGBA8888: |
| 625 |
case VC4_TEXTURE_TYPE_RGBX8888: |
| 626 |
case VC4_TEXTURE_TYPE_RGBA32R: |
| 627 |
cpp = 4; |
| 628 |
break; |
| 629 |
case VC4_TEXTURE_TYPE_RGBA4444: |
| 630 |
case VC4_TEXTURE_TYPE_RGBA5551: |
| 631 |
case VC4_TEXTURE_TYPE_RGB565: |
| 632 |
case VC4_TEXTURE_TYPE_LUMALPHA: |
| 633 |
case VC4_TEXTURE_TYPE_S16F: |
| 634 |
case VC4_TEXTURE_TYPE_S16: |
| 635 |
cpp = 2; |
| 636 |
break; |
| 637 |
case VC4_TEXTURE_TYPE_LUMINANCE: |
| 638 |
case VC4_TEXTURE_TYPE_ALPHA: |
| 639 |
case VC4_TEXTURE_TYPE_S8: |
| 640 |
cpp = 1; |
| 641 |
break; |
| 642 |
case VC4_TEXTURE_TYPE_ETC1: |
| 643 |
/* ETC1 is arranged as 64-bit blocks, where each block is 4x4 |
| 644 |
* pixels. |
| 645 |
*/ |
| 646 |
cpp = 8; |
| 647 |
width = (width + 3) >> 2; |
| 648 |
height = (height + 3) >> 2; |
| 649 |
break; |
| 650 |
case VC4_TEXTURE_TYPE_BW1: |
| 651 |
case VC4_TEXTURE_TYPE_A4: |
| 652 |
case VC4_TEXTURE_TYPE_A1: |
| 653 |
case VC4_TEXTURE_TYPE_RGBA64: |
| 654 |
case VC4_TEXTURE_TYPE_YUV422R: |
| 655 |
default: |
| 656 |
DRM_ERROR("Texture format %d unsupported\n", type); |
| 657 |
goto fail; |
| 658 |
} |
| 659 |
utile_w = utile_width(cpp); |
| 660 |
utile_h = utile_height(cpp); |
| 661 |
|
| 662 |
if (type == VC4_TEXTURE_TYPE_RGBA32R) { |
| 663 |
tiling_format = VC4_TILING_FORMAT_LINEAR; |
| 664 |
} else { |
| 665 |
if (size_is_lt(width, height, cpp)) |
| 666 |
tiling_format = VC4_TILING_FORMAT_LT; |
| 667 |
else |
| 668 |
tiling_format = VC4_TILING_FORMAT_T; |
| 669 |
} |
| 670 |
|
| 671 |
if (!vc4_check_tex_size(exec, tex, offset + cube_map_stride * 5, |
| 672 |
tiling_format, width, height, cpp)) { |
| 673 |
goto fail; |
| 674 |
} |
| 675 |
|
| 676 |
/* The mipmap levels are stored before the base of the texture. Make |
| 677 |
* sure there is actually space in the BO. |
| 678 |
*/ |
| 679 |
for (i = 1; i <= miplevels; i++) { |
| 680 |
uint32_t level_width = max(width >> i, 1u); |
| 681 |
uint32_t level_height = max(height >> i, 1u); |
| 682 |
uint32_t aligned_width, aligned_height; |
| 683 |
uint32_t level_size; |
| 684 |
|
| 685 |
/* Once the levels get small enough, they drop from T to LT. */ |
| 686 |
if (tiling_format == VC4_TILING_FORMAT_T && |
| 687 |
size_is_lt(level_width, level_height, cpp)) { |
| 688 |
tiling_format = VC4_TILING_FORMAT_LT; |
| 689 |
} |
| 690 |
|
| 691 |
switch (tiling_format) { |
| 692 |
case VC4_TILING_FORMAT_T: |
| 693 |
aligned_width = round_up(level_width, utile_w * 8); |
| 694 |
aligned_height = round_up(level_height, utile_h * 8); |
| 695 |
break; |
| 696 |
case VC4_TILING_FORMAT_LT: |
| 697 |
aligned_width = round_up(level_width, utile_w); |
| 698 |
aligned_height = round_up(level_height, utile_h); |
| 699 |
break; |
| 700 |
default: |
| 701 |
aligned_width = round_up(level_width, utile_w); |
| 702 |
aligned_height = level_height; |
| 703 |
break; |
| 704 |
} |
| 705 |
|
| 706 |
level_size = aligned_width * cpp * aligned_height; |
| 707 |
|
| 708 |
if (offset < level_size) { |
| 709 |
DRM_ERROR("Level %d (%dx%d -> %dx%d) size %db " |
| 710 |
"overflowed buffer bounds (offset %d)\n", |
| 711 |
i, level_width, level_height, |
| 712 |
aligned_width, aligned_height, |
| 713 |
level_size, offset); |
| 714 |
goto fail; |
| 715 |
} |
| 716 |
|
| 717 |
offset -= level_size; |
| 718 |
} |
| 719 |
|
| 720 |
*validated_p0 = tex->paddr + p0; |
| 721 |
|
| 722 |
return true; |
| 723 |
fail: |
| 724 |
DRM_INFO("Texture p0 at %d: 0x%08x\n", sample->p_offset[0], p0); |
| 725 |
DRM_INFO("Texture p1 at %d: 0x%08x\n", sample->p_offset[1], p1); |
| 726 |
DRM_INFO("Texture p2 at %d: 0x%08x\n", sample->p_offset[2], p2); |
| 727 |
DRM_INFO("Texture p3 at %d: 0x%08x\n", sample->p_offset[3], p3); |
| 728 |
return false; |
| 729 |
} |
| 730 |
|
| 731 |
static int |
| 732 |
validate_gl_shader_rec(struct drm_device *dev, |
| 733 |
struct vc4_exec_info *exec, |
| 734 |
struct vc4_shader_state *state) |
| 735 |
{ |
| 736 |
uint32_t *src_handles; |
| 737 |
void *pkt_u, *pkt_v; |
| 738 |
static const uint32_t shader_reloc_offsets[] = { |
| 739 |
4, /* fs */ |
| 740 |
16, /* vs */ |
| 741 |
28, /* cs */ |
| 742 |
}; |
| 743 |
uint32_t shader_reloc_count = ARRAY_SIZE(shader_reloc_offsets); |
| 744 |
struct drm_gem_cma_object *bo[shader_reloc_count + 8]; |
| 745 |
uint32_t nr_attributes, nr_relocs, packet_size; |
| 746 |
int i; |
| 747 |
|
| 748 |
nr_attributes = state->addr & 0x7; |
| 749 |
if (nr_attributes == 0) |
| 750 |
nr_attributes = 8; |
| 751 |
packet_size = gl_shader_rec_size(state->addr); |
| 752 |
|
| 753 |
nr_relocs = ARRAY_SIZE(shader_reloc_offsets) + nr_attributes; |
| 754 |
if (nr_relocs * 4 > exec->shader_rec_size) { |
| 755 |
DRM_ERROR("overflowed shader recs reading %d handles " |
| 756 |
"from %d bytes left\n", |
| 757 |
nr_relocs, exec->shader_rec_size); |
| 758 |
return -EINVAL; |
| 759 |
} |
| 760 |
src_handles = exec->shader_rec_u; |
| 761 |
exec->shader_rec_u += nr_relocs * 4; |
| 762 |
exec->shader_rec_size -= nr_relocs * 4; |
| 763 |
|
| 764 |
if (packet_size > exec->shader_rec_size) { |
| 765 |
DRM_ERROR("overflowed shader recs copying %db packet " |
| 766 |
"from %d bytes left\n", |
| 767 |
packet_size, exec->shader_rec_size); |
| 768 |
return -EINVAL; |
| 769 |
} |
| 770 |
pkt_u = exec->shader_rec_u; |
| 771 |
pkt_v = exec->shader_rec_v; |
| 772 |
memcpy(pkt_v, pkt_u, packet_size); |
| 773 |
exec->shader_rec_u += packet_size; |
| 774 |
/* Shader recs have to be aligned to 16 bytes (due to the attribute |
| 775 |
* flags being in the low bytes), so round the next validated shader |
| 776 |
* rec address up. This should be safe, since we've got so many |
| 777 |
* relocations in a shader rec packet. |
| 778 |
*/ |
| 779 |
BUG_ON(roundup(packet_size, 16) - packet_size > nr_relocs * 4); |
| 780 |
exec->shader_rec_v += roundup(packet_size, 16); |
| 781 |
exec->shader_rec_size -= packet_size; |
| 782 |
|
| 783 |
for (i = 0; i < shader_reloc_count; i++) { |
| 784 |
if (src_handles[i] > exec->bo_count) { |
| 785 |
DRM_ERROR("Shader handle %d too big\n", src_handles[i]); |
| 786 |
return -EINVAL; |
| 787 |
} |
| 788 |
|
| 789 |
bo[i] = exec->bo[src_handles[i]]; |
| 790 |
if (!bo[i]) |
| 791 |
return -EINVAL; |
| 792 |
} |
| 793 |
for (i = shader_reloc_count; i < nr_relocs; i++) { |
| 794 |
bo[i] = vc4_use_bo(exec, src_handles[i]); |
| 795 |
if (!bo[i]) |
| 796 |
return -EINVAL; |
| 797 |
} |
| 798 |
|
| 799 |
if (((*(uint16_t *)pkt_u & VC4_SHADER_FLAG_FS_SINGLE_THREAD) == 0) != |
| 800 |
to_vc4_bo(&bo[0]->base)->validated_shader->is_threaded) { |
| 801 |
DRM_ERROR("Thread mode of CL and FS do not match\n"); |
| 802 |
return -EINVAL; |
| 803 |
} |
| 804 |
|
| 805 |
if (to_vc4_bo(&bo[1]->base)->validated_shader->is_threaded || |
| 806 |
to_vc4_bo(&bo[2]->base)->validated_shader->is_threaded) { |
| 807 |
DRM_ERROR("cs and vs cannot be threaded\n"); |
| 808 |
return -EINVAL; |
| 809 |
} |
| 810 |
|
| 811 |
for (i = 0; i < shader_reloc_count; i++) { |
| 812 |
struct vc4_validated_shader_info *validated_shader; |
| 813 |
uint32_t o = shader_reloc_offsets[i]; |
| 814 |
uint32_t src_offset = *(uint32_t *)(pkt_u + o); |
| 815 |
uint32_t *texture_handles_u; |
| 816 |
void *uniform_data_u; |
| 817 |
uint32_t tex, uni; |
| 818 |
|
| 819 |
*(uint32_t *)(pkt_v + o) = bo[i]->paddr + src_offset; |
| 820 |
|
| 821 |
if (src_offset != 0) { |
| 822 |
DRM_ERROR("Shaders must be at offset 0 of " |
| 823 |
"the BO.\n"); |
| 824 |
return -EINVAL; |
| 825 |
} |
| 826 |
|
| 827 |
validated_shader = to_vc4_bo(&bo[i]->base)->validated_shader; |
| 828 |
if (!validated_shader) |
| 829 |
return -EINVAL; |
| 830 |
|
| 831 |
if (validated_shader->uniforms_src_size > |
| 832 |
exec->uniforms_size) { |
| 833 |
DRM_ERROR("Uniforms src buffer overflow\n"); |
| 834 |
return -EINVAL; |
| 835 |
} |
| 836 |
|
| 837 |
texture_handles_u = exec->uniforms_u; |
| 838 |
uniform_data_u = (texture_handles_u + |
| 839 |
validated_shader->num_texture_samples); |
| 840 |
|
| 841 |
memcpy(exec->uniforms_v, uniform_data_u, |
| 842 |
validated_shader->uniforms_size); |
| 843 |
|
| 844 |
for (tex = 0; |
| 845 |
tex < validated_shader->num_texture_samples; |
| 846 |
tex++) { |
| 847 |
if (!reloc_tex(exec, |
| 848 |
uniform_data_u, |
| 849 |
&validated_shader->texture_samples[tex], |
| 850 |
texture_handles_u[tex])) { |
| 851 |
return -EINVAL; |
| 852 |
} |
| 853 |
} |
| 854 |
|
| 855 |
/* Fill in the uniform slots that need this shader's |
| 856 |
* start-of-uniforms address (used for resetting the uniform |
| 857 |
* stream in the presence of control flow). |
| 858 |
*/ |
| 859 |
for (uni = 0; |
| 860 |
uni < validated_shader->num_uniform_addr_offsets; |
| 861 |
uni++) { |
| 862 |
uint32_t o = validated_shader->uniform_addr_offsets[uni]; |
| 863 |
((uint32_t *)exec->uniforms_v)[o] = exec->uniforms_p; |
| 864 |
} |
| 865 |
|
| 866 |
*(uint32_t *)(pkt_v + o + 4) = exec->uniforms_p; |
| 867 |
|
| 868 |
exec->uniforms_u += validated_shader->uniforms_src_size; |
| 869 |
exec->uniforms_v += validated_shader->uniforms_size; |
| 870 |
exec->uniforms_p += validated_shader->uniforms_size; |
| 871 |
} |
| 872 |
|
| 873 |
for (i = 0; i < nr_attributes; i++) { |
| 874 |
struct drm_gem_cma_object *vbo = |
| 875 |
bo[ARRAY_SIZE(shader_reloc_offsets) + i]; |
| 876 |
uint32_t o = 36 + i * 8; |
| 877 |
uint32_t offset = *(uint32_t *)(pkt_u + o + 0); |
| 878 |
uint32_t attr_size = *(uint8_t *)(pkt_u + o + 4) + 1; |
| 879 |
uint32_t stride = *(uint8_t *)(pkt_u + o + 5); |
| 880 |
uint32_t max_index; |
| 881 |
|
| 882 |
if (state->addr & 0x8) |
| 883 |
stride |= (*(uint32_t *)(pkt_u + 100 + i * 4)) & ~0xff; |
| 884 |
|
| 885 |
if (vbo->base.size < offset || |
| 886 |
vbo->base.size - offset < attr_size) { |
| 887 |
DRM_ERROR("BO offset overflow (%d + %d > %zd)\n", |
| 888 |
offset, attr_size, vbo->base.size); |
| 889 |
return -EINVAL; |
| 890 |
} |
| 891 |
|
| 892 |
if (stride != 0) { |
| 893 |
max_index = ((vbo->base.size - offset - attr_size) / |
| 894 |
stride); |
| 895 |
if (state->max_index > max_index) { |
| 896 |
DRM_ERROR("primitives use index %d out of " |
| 897 |
"supplied %d\n", |
| 898 |
state->max_index, max_index); |
| 899 |
return -EINVAL; |
| 900 |
} |
| 901 |
} |
| 902 |
|
| 903 |
*(uint32_t *)(pkt_v + o) = vbo->paddr + offset; |
| 904 |
} |
| 905 |
|
| 906 |
return 0; |
| 907 |
} |
| 908 |
|
| 909 |
int |
| 910 |
vc4_validate_shader_recs(struct drm_device *dev, |
| 911 |
struct vc4_exec_info *exec) |
| 912 |
{ |
| 913 |
uint32_t i; |
| 914 |
int ret = 0; |
| 915 |
|
| 916 |
for (i = 0; i < exec->shader_state_count; i++) { |
| 917 |
ret = validate_gl_shader_rec(dev, exec, &exec->shader_state[i]); |
| 918 |
if (ret) |
| 919 |
return ret; |
| 920 |
} |
| 921 |
|
| 922 |
return ret; |
| 923 |
} |