1 /* 2 * Copyright © 2013 Intel Corporation 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 * Authors: 24 * Brad Volkin <bradley.d.volkin@intel.com> 25 * 26 */ 27 28 #include "gt/intel_engine.h" 29 30 #include "i915_drv.h" 31 #include "i915_memcpy.h" 32 33 /** 34 * DOC: batch buffer command parser 35 * 36 * Motivation: 37 * Certain OpenGL features (e.g. transform feedback, performance monitoring) 38 * require userspace code to submit batches containing commands such as 39 * MI_LOAD_REGISTER_IMM to access various registers. Unfortunately, some 40 * generations of the hardware will noop these commands in "unsecure" batches 41 * (which includes all userspace batches submitted via i915) even though the 42 * commands may be safe and represent the intended programming model of the 43 * device. 44 * 45 * The software command parser is similar in operation to the command parsing 46 * done in hardware for unsecure batches. However, the software parser allows 47 * some operations that would be noop'd by hardware, if the parser determines 48 * the operation is safe, and submits the batch as "secure" to prevent hardware 49 * parsing. 50 * 51 * Threats: 52 * At a high level, the hardware (and software) checks attempt to prevent 53 * granting userspace undue privileges. There are three categories of privilege. 54 * 55 * First, commands which are explicitly defined as privileged or which should 56 * only be used by the kernel driver. The parser rejects such commands 57 * 58 * Second, commands which access registers. To support correct/enhanced 59 * userspace functionality, particularly certain OpenGL extensions, the parser 60 * provides a whitelist of registers which userspace may safely access 61 * 62 * Third, commands which access privileged memory (i.e. GGTT, HWS page, etc). 63 * The parser always rejects such commands. 64 * 65 * The majority of the problematic commands fall in the MI_* range, with only a 66 * few specific commands on each engine (e.g. PIPE_CONTROL and MI_FLUSH_DW). 67 * 68 * Implementation: 69 * Each engine maintains tables of commands and registers which the parser 70 * uses in scanning batch buffers submitted to that engine. 71 * 72 * Since the set of commands that the parser must check for is significantly 73 * smaller than the number of commands supported, the parser tables contain only 74 * those commands required by the parser. This generally works because command 75 * opcode ranges have standard command length encodings. So for commands that 76 * the parser does not need to check, it can easily skip them. This is 77 * implemented via a per-engine length decoding vfunc. 78 * 79 * Unfortunately, there are a number of commands that do not follow the standard 80 * length encoding for their opcode range, primarily amongst the MI_* commands. 81 * To handle this, the parser provides a way to define explicit "skip" entries 82 * in the per-engine command tables. 83 * 84 * Other command table entries map fairly directly to high level categories 85 * mentioned above: rejected, register whitelist. The parser implements a number 86 * of checks, including the privileged memory checks, via a general bitmasking 87 * mechanism. 88 */ 89 90 /* 91 * A command that requires special handling by the command parser. 92 */ 93 struct drm_i915_cmd_descriptor { 94 /* 95 * Flags describing how the command parser processes the command. 96 * 97 * CMD_DESC_FIXED: The command has a fixed length if this is set, 98 * a length mask if not set 99 * CMD_DESC_SKIP: The command is allowed but does not follow the 100 * standard length encoding for the opcode range in 101 * which it falls 102 * CMD_DESC_REJECT: The command is never allowed 103 * CMD_DESC_REGISTER: The command should be checked against the 104 * register whitelist for the appropriate ring 105 */ 106 u32 flags; 107 #define CMD_DESC_FIXED (1<<0) 108 #define CMD_DESC_SKIP (1<<1) 109 #define CMD_DESC_REJECT (1<<2) 110 #define CMD_DESC_REGISTER (1<<3) 111 #define CMD_DESC_BITMASK (1<<4) 112 113 /* 114 * The command's unique identification bits and the bitmask to get them. 115 * This isn't strictly the opcode field as defined in the spec and may 116 * also include type, subtype, and/or subop fields. 117 */ 118 struct { 119 u32 value; 120 u32 mask; 121 } cmd; 122 123 /* 124 * The command's length. The command is either fixed length (i.e. does 125 * not include a length field) or has a length field mask. The flag 126 * CMD_DESC_FIXED indicates a fixed length. Otherwise, the command has 127 * a length mask. All command entries in a command table must include 128 * length information. 129 */ 130 union { 131 u32 fixed; 132 u32 mask; 133 } length; 134 135 /* 136 * Describes where to find a register address in the command to check 137 * against the ring's register whitelist. Only valid if flags has the 138 * CMD_DESC_REGISTER bit set. 139 * 140 * A non-zero step value implies that the command may access multiple 141 * registers in sequence (e.g. LRI), in that case step gives the 142 * distance in dwords between individual offset fields. 143 */ 144 struct { 145 u32 offset; 146 u32 mask; 147 u32 step; 148 } reg; 149 150 #define MAX_CMD_DESC_BITMASKS 3 151 /* 152 * Describes command checks where a particular dword is masked and 153 * compared against an expected value. If the command does not match 154 * the expected value, the parser rejects it. Only valid if flags has 155 * the CMD_DESC_BITMASK bit set. Only entries where mask is non-zero 156 * are valid. 157 * 158 * If the check specifies a non-zero condition_mask then the parser 159 * only performs the check when the bits specified by condition_mask 160 * are non-zero. 161 */ 162 struct { 163 u32 offset; 164 u32 mask; 165 u32 expected; 166 u32 condition_offset; 167 u32 condition_mask; 168 } bits[MAX_CMD_DESC_BITMASKS]; 169 }; 170 171 /* 172 * A table of commands requiring special handling by the command parser. 173 * 174 * Each engine has an array of tables. Each table consists of an array of 175 * command descriptors, which must be sorted with command opcodes in 176 * ascending order. 177 */ 178 struct drm_i915_cmd_table { 179 const struct drm_i915_cmd_descriptor *table; 180 int count; 181 }; 182 183 #define STD_MI_OPCODE_SHIFT (32 - 9) 184 #define STD_3D_OPCODE_SHIFT (32 - 16) 185 #define STD_2D_OPCODE_SHIFT (32 - 10) 186 #define STD_MFX_OPCODE_SHIFT (32 - 16) 187 #define MIN_OPCODE_SHIFT 16 188 189 #define CMD(op, opm, f, lm, fl, ...) \ 190 { \ 191 .flags = (fl) | ((f) ? CMD_DESC_FIXED : 0), \ 192 .cmd = { (op & ~0u << (opm)), ~0u << (opm) }, \ 193 .length = { (lm) }, \ 194 __VA_ARGS__ \ 195 } 196 197 /* Convenience macros to compress the tables */ 198 #define SMI STD_MI_OPCODE_SHIFT 199 #define S3D STD_3D_OPCODE_SHIFT 200 #define S2D STD_2D_OPCODE_SHIFT 201 #define SMFX STD_MFX_OPCODE_SHIFT 202 #define F true 203 #define S CMD_DESC_SKIP 204 #define R CMD_DESC_REJECT 205 #define W CMD_DESC_REGISTER 206 #define B CMD_DESC_BITMASK 207 208 /* Command Mask Fixed Len Action 209 ---------------------------------------------------------- */ 210 static const struct drm_i915_cmd_descriptor gen7_common_cmds[] = { 211 CMD( MI_NOOP, SMI, F, 1, S ), 212 CMD( MI_USER_INTERRUPT, SMI, F, 1, R ), 213 CMD( MI_WAIT_FOR_EVENT, SMI, F, 1, R ), 214 CMD( MI_ARB_CHECK, SMI, F, 1, S ), 215 CMD( MI_REPORT_HEAD, SMI, F, 1, S ), 216 CMD( MI_SUSPEND_FLUSH, SMI, F, 1, S ), 217 CMD( MI_SEMAPHORE_MBOX, SMI, !F, 0xFF, R ), 218 CMD( MI_STORE_DWORD_INDEX, SMI, !F, 0xFF, R ), 219 CMD( MI_LOAD_REGISTER_IMM(1), SMI, !F, 0xFF, W, 220 .reg = { .offset = 1, .mask = 0x007FFFFC, .step = 2 } ), 221 CMD( MI_STORE_REGISTER_MEM, SMI, F, 3, W | B, 222 .reg = { .offset = 1, .mask = 0x007FFFFC }, 223 .bits = {{ 224 .offset = 0, 225 .mask = MI_GLOBAL_GTT, 226 .expected = 0, 227 }}, ), 228 CMD( MI_LOAD_REGISTER_MEM, SMI, F, 3, W | B, 229 .reg = { .offset = 1, .mask = 0x007FFFFC }, 230 .bits = {{ 231 .offset = 0, 232 .mask = MI_GLOBAL_GTT, 233 .expected = 0, 234 }}, ), 235 /* 236 * MI_BATCH_BUFFER_START requires some special handling. It's not 237 * really a 'skip' action but it doesn't seem like it's worth adding 238 * a new action. See intel_engine_cmd_parser(). 239 */ 240 CMD( MI_BATCH_BUFFER_START, SMI, !F, 0xFF, S ), 241 }; 242 243 static const struct drm_i915_cmd_descriptor gen7_render_cmds[] = { 244 CMD( MI_FLUSH, SMI, F, 1, S ), 245 CMD( MI_ARB_ON_OFF, SMI, F, 1, R ), 246 CMD( MI_PREDICATE, SMI, F, 1, S ), 247 CMD( MI_TOPOLOGY_FILTER, SMI, F, 1, S ), 248 CMD( MI_SET_APPID, SMI, F, 1, S ), 249 CMD( MI_DISPLAY_FLIP, SMI, !F, 0xFF, R ), 250 CMD( MI_SET_CONTEXT, SMI, !F, 0xFF, R ), 251 CMD( MI_URB_CLEAR, SMI, !F, 0xFF, S ), 252 CMD( MI_STORE_DWORD_IMM, SMI, !F, 0x3F, B, 253 .bits = {{ 254 .offset = 0, 255 .mask = MI_GLOBAL_GTT, 256 .expected = 0, 257 }}, ), 258 CMD( MI_UPDATE_GTT, SMI, !F, 0xFF, R ), 259 CMD( MI_CLFLUSH, SMI, !F, 0x3FF, B, 260 .bits = {{ 261 .offset = 0, 262 .mask = MI_GLOBAL_GTT, 263 .expected = 0, 264 }}, ), 265 CMD( MI_REPORT_PERF_COUNT, SMI, !F, 0x3F, B, 266 .bits = {{ 267 .offset = 1, 268 .mask = MI_REPORT_PERF_COUNT_GGTT, 269 .expected = 0, 270 }}, ), 271 CMD( MI_CONDITIONAL_BATCH_BUFFER_END, SMI, !F, 0xFF, B, 272 .bits = {{ 273 .offset = 0, 274 .mask = MI_GLOBAL_GTT, 275 .expected = 0, 276 }}, ), 277 CMD( GFX_OP_3DSTATE_VF_STATISTICS, S3D, F, 1, S ), 278 CMD( PIPELINE_SELECT, S3D, F, 1, S ), 279 CMD( MEDIA_VFE_STATE, S3D, !F, 0xFFFF, B, 280 .bits = {{ 281 .offset = 2, 282 .mask = MEDIA_VFE_STATE_MMIO_ACCESS_MASK, 283 .expected = 0, 284 }}, ), 285 CMD( GPGPU_OBJECT, S3D, !F, 0xFF, S ), 286 CMD( GPGPU_WALKER, S3D, !F, 0xFF, S ), 287 CMD( GFX_OP_3DSTATE_SO_DECL_LIST, S3D, !F, 0x1FF, S ), 288 CMD( GFX_OP_PIPE_CONTROL(5), S3D, !F, 0xFF, B, 289 .bits = {{ 290 .offset = 1, 291 .mask = (PIPE_CONTROL_MMIO_WRITE | PIPE_CONTROL_NOTIFY), 292 .expected = 0, 293 }, 294 { 295 .offset = 1, 296 .mask = (PIPE_CONTROL_GLOBAL_GTT_IVB | 297 PIPE_CONTROL_STORE_DATA_INDEX), 298 .expected = 0, 299 .condition_offset = 1, 300 .condition_mask = PIPE_CONTROL_POST_SYNC_OP_MASK, 301 }}, ), 302 }; 303 304 static const struct drm_i915_cmd_descriptor hsw_render_cmds[] = { 305 CMD( MI_SET_PREDICATE, SMI, F, 1, S ), 306 CMD( MI_RS_CONTROL, SMI, F, 1, S ), 307 CMD( MI_URB_ATOMIC_ALLOC, SMI, F, 1, S ), 308 CMD( MI_SET_APPID, SMI, F, 1, S ), 309 CMD( MI_RS_CONTEXT, SMI, F, 1, S ), 310 CMD( MI_LOAD_SCAN_LINES_INCL, SMI, !F, 0x3F, R ), 311 CMD( MI_LOAD_SCAN_LINES_EXCL, SMI, !F, 0x3F, R ), 312 CMD( MI_LOAD_REGISTER_REG, SMI, !F, 0xFF, W, 313 .reg = { .offset = 1, .mask = 0x007FFFFC, .step = 1 } ), 314 CMD( MI_RS_STORE_DATA_IMM, SMI, !F, 0xFF, S ), 315 CMD( MI_LOAD_URB_MEM, SMI, !F, 0xFF, S ), 316 CMD( MI_STORE_URB_MEM, SMI, !F, 0xFF, S ), 317 CMD( GFX_OP_3DSTATE_DX9_CONSTANTF_VS, S3D, !F, 0x7FF, S ), 318 CMD( GFX_OP_3DSTATE_DX9_CONSTANTF_PS, S3D, !F, 0x7FF, S ), 319 320 CMD( GFX_OP_3DSTATE_BINDING_TABLE_EDIT_VS, S3D, !F, 0x1FF, S ), 321 CMD( GFX_OP_3DSTATE_BINDING_TABLE_EDIT_GS, S3D, !F, 0x1FF, S ), 322 CMD( GFX_OP_3DSTATE_BINDING_TABLE_EDIT_HS, S3D, !F, 0x1FF, S ), 323 CMD( GFX_OP_3DSTATE_BINDING_TABLE_EDIT_DS, S3D, !F, 0x1FF, S ), 324 CMD( GFX_OP_3DSTATE_BINDING_TABLE_EDIT_PS, S3D, !F, 0x1FF, S ), 325 }; 326 327 static const struct drm_i915_cmd_descriptor gen7_video_cmds[] = { 328 CMD( MI_ARB_ON_OFF, SMI, F, 1, R ), 329 CMD( MI_SET_APPID, SMI, F, 1, S ), 330 CMD( MI_STORE_DWORD_IMM, SMI, !F, 0xFF, B, 331 .bits = {{ 332 .offset = 0, 333 .mask = MI_GLOBAL_GTT, 334 .expected = 0, 335 }}, ), 336 CMD( MI_UPDATE_GTT, SMI, !F, 0x3F, R ), 337 CMD( MI_FLUSH_DW, SMI, !F, 0x3F, B, 338 .bits = {{ 339 .offset = 0, 340 .mask = MI_FLUSH_DW_NOTIFY, 341 .expected = 0, 342 }, 343 { 344 .offset = 1, 345 .mask = MI_FLUSH_DW_USE_GTT, 346 .expected = 0, 347 .condition_offset = 0, 348 .condition_mask = MI_FLUSH_DW_OP_MASK, 349 }, 350 { 351 .offset = 0, 352 .mask = MI_FLUSH_DW_STORE_INDEX, 353 .expected = 0, 354 .condition_offset = 0, 355 .condition_mask = MI_FLUSH_DW_OP_MASK, 356 }}, ), 357 CMD( MI_CONDITIONAL_BATCH_BUFFER_END, SMI, !F, 0xFF, B, 358 .bits = {{ 359 .offset = 0, 360 .mask = MI_GLOBAL_GTT, 361 .expected = 0, 362 }}, ), 363 /* 364 * MFX_WAIT doesn't fit the way we handle length for most commands. 365 * It has a length field but it uses a non-standard length bias. 366 * It is always 1 dword though, so just treat it as fixed length. 367 */ 368 CMD( MFX_WAIT, SMFX, F, 1, S ), 369 }; 370 371 static const struct drm_i915_cmd_descriptor gen7_vecs_cmds[] = { 372 CMD( MI_ARB_ON_OFF, SMI, F, 1, R ), 373 CMD( MI_SET_APPID, SMI, F, 1, S ), 374 CMD( MI_STORE_DWORD_IMM, SMI, !F, 0xFF, B, 375 .bits = {{ 376 .offset = 0, 377 .mask = MI_GLOBAL_GTT, 378 .expected = 0, 379 }}, ), 380 CMD( MI_UPDATE_GTT, SMI, !F, 0x3F, R ), 381 CMD( MI_FLUSH_DW, SMI, !F, 0x3F, B, 382 .bits = {{ 383 .offset = 0, 384 .mask = MI_FLUSH_DW_NOTIFY, 385 .expected = 0, 386 }, 387 { 388 .offset = 1, 389 .mask = MI_FLUSH_DW_USE_GTT, 390 .expected = 0, 391 .condition_offset = 0, 392 .condition_mask = MI_FLUSH_DW_OP_MASK, 393 }, 394 { 395 .offset = 0, 396 .mask = MI_FLUSH_DW_STORE_INDEX, 397 .expected = 0, 398 .condition_offset = 0, 399 .condition_mask = MI_FLUSH_DW_OP_MASK, 400 }}, ), 401 CMD( MI_CONDITIONAL_BATCH_BUFFER_END, SMI, !F, 0xFF, B, 402 .bits = {{ 403 .offset = 0, 404 .mask = MI_GLOBAL_GTT, 405 .expected = 0, 406 }}, ), 407 }; 408 409 static const struct drm_i915_cmd_descriptor gen7_blt_cmds[] = { 410 CMD( MI_DISPLAY_FLIP, SMI, !F, 0xFF, R ), 411 CMD( MI_STORE_DWORD_IMM, SMI, !F, 0x3FF, B, 412 .bits = {{ 413 .offset = 0, 414 .mask = MI_GLOBAL_GTT, 415 .expected = 0, 416 }}, ), 417 CMD( MI_UPDATE_GTT, SMI, !F, 0x3F, R ), 418 CMD( MI_FLUSH_DW, SMI, !F, 0x3F, B, 419 .bits = {{ 420 .offset = 0, 421 .mask = MI_FLUSH_DW_NOTIFY, 422 .expected = 0, 423 }, 424 { 425 .offset = 1, 426 .mask = MI_FLUSH_DW_USE_GTT, 427 .expected = 0, 428 .condition_offset = 0, 429 .condition_mask = MI_FLUSH_DW_OP_MASK, 430 }, 431 { 432 .offset = 0, 433 .mask = MI_FLUSH_DW_STORE_INDEX, 434 .expected = 0, 435 .condition_offset = 0, 436 .condition_mask = MI_FLUSH_DW_OP_MASK, 437 }}, ), 438 CMD( COLOR_BLT, S2D, !F, 0x3F, S ), 439 CMD( SRC_COPY_BLT, S2D, !F, 0x3F, S ), 440 }; 441 442 static const struct drm_i915_cmd_descriptor hsw_blt_cmds[] = { 443 CMD( MI_LOAD_SCAN_LINES_INCL, SMI, !F, 0x3F, R ), 444 CMD( MI_LOAD_SCAN_LINES_EXCL, SMI, !F, 0x3F, R ), 445 }; 446 447 /* 448 * For Gen9 we can still rely on the h/w to enforce cmd security, and only 449 * need to re-enforce the register access checks. We therefore only need to 450 * teach the cmdparser how to find the end of each command, and identify 451 * register accesses. The table doesn't need to reject any commands, and so 452 * the only commands listed here are: 453 * 1) Those that touch registers 454 * 2) Those that do not have the default 8-bit length 455 * 456 * Note that the default MI length mask chosen for this table is 0xFF, not 457 * the 0x3F used on older devices. This is because the vast majority of MI 458 * cmds on Gen9 use a standard 8-bit Length field. 459 * All the Gen9 blitter instructions are standard 0xFF length mask, and 460 * none allow access to non-general registers, so in fact no BLT cmds are 461 * included in the table at all. 462 * 463 */ 464 static const struct drm_i915_cmd_descriptor gen9_blt_cmds[] = { 465 CMD( MI_NOOP, SMI, F, 1, S ), 466 CMD( MI_USER_INTERRUPT, SMI, F, 1, S ), 467 CMD( MI_WAIT_FOR_EVENT, SMI, F, 1, S ), 468 CMD( MI_FLUSH, SMI, F, 1, S ), 469 CMD( MI_ARB_CHECK, SMI, F, 1, S ), 470 CMD( MI_REPORT_HEAD, SMI, F, 1, S ), 471 CMD( MI_ARB_ON_OFF, SMI, F, 1, S ), 472 CMD( MI_SUSPEND_FLUSH, SMI, F, 1, S ), 473 CMD( MI_LOAD_SCAN_LINES_INCL, SMI, !F, 0x3F, S ), 474 CMD( MI_LOAD_SCAN_LINES_EXCL, SMI, !F, 0x3F, S ), 475 CMD( MI_STORE_DWORD_IMM, SMI, !F, 0x3FF, S ), 476 CMD( MI_LOAD_REGISTER_IMM(1), SMI, !F, 0xFF, W, 477 .reg = { .offset = 1, .mask = 0x007FFFFC, .step = 2 } ), 478 CMD( MI_UPDATE_GTT, SMI, !F, 0x3FF, S ), 479 CMD( MI_STORE_REGISTER_MEM_GEN8, SMI, F, 4, W, 480 .reg = { .offset = 1, .mask = 0x007FFFFC } ), 481 CMD( MI_FLUSH_DW, SMI, !F, 0x3F, S ), 482 CMD( MI_LOAD_REGISTER_MEM_GEN8, SMI, F, 4, W, 483 .reg = { .offset = 1, .mask = 0x007FFFFC } ), 484 CMD( MI_LOAD_REGISTER_REG, SMI, !F, 0xFF, W, 485 .reg = { .offset = 1, .mask = 0x007FFFFC, .step = 1 } ), 486 487 /* 488 * We allow BB_START but apply further checks. We just sanitize the 489 * basic fields here. 490 */ 491 #define MI_BB_START_OPERAND_MASK GENMASK(SMI-1, 0) 492 #define MI_BB_START_OPERAND_EXPECT (MI_BATCH_PPGTT_HSW | 1) 493 CMD( MI_BATCH_BUFFER_START_GEN8, SMI, !F, 0xFF, B, 494 .bits = {{ 495 .offset = 0, 496 .mask = MI_BB_START_OPERAND_MASK, 497 .expected = MI_BB_START_OPERAND_EXPECT, 498 }}, ), 499 }; 500 501 static const struct drm_i915_cmd_descriptor noop_desc = 502 CMD(MI_NOOP, SMI, F, 1, S); 503 504 #undef CMD 505 #undef SMI 506 #undef S3D 507 #undef S2D 508 #undef SMFX 509 #undef F 510 #undef S 511 #undef R 512 #undef W 513 #undef B 514 515 static const struct drm_i915_cmd_table gen7_render_cmd_table[] = { 516 { gen7_common_cmds, ARRAY_SIZE(gen7_common_cmds) }, 517 { gen7_render_cmds, ARRAY_SIZE(gen7_render_cmds) }, 518 }; 519 520 static const struct drm_i915_cmd_table hsw_render_ring_cmd_table[] = { 521 { gen7_common_cmds, ARRAY_SIZE(gen7_common_cmds) }, 522 { gen7_render_cmds, ARRAY_SIZE(gen7_render_cmds) }, 523 { hsw_render_cmds, ARRAY_SIZE(hsw_render_cmds) }, 524 }; 525 526 static const struct drm_i915_cmd_table gen7_video_cmd_table[] = { 527 { gen7_common_cmds, ARRAY_SIZE(gen7_common_cmds) }, 528 { gen7_video_cmds, ARRAY_SIZE(gen7_video_cmds) }, 529 }; 530 531 static const struct drm_i915_cmd_table hsw_vebox_cmd_table[] = { 532 { gen7_common_cmds, ARRAY_SIZE(gen7_common_cmds) }, 533 { gen7_vecs_cmds, ARRAY_SIZE(gen7_vecs_cmds) }, 534 }; 535 536 static const struct drm_i915_cmd_table gen7_blt_cmd_table[] = { 537 { gen7_common_cmds, ARRAY_SIZE(gen7_common_cmds) }, 538 { gen7_blt_cmds, ARRAY_SIZE(gen7_blt_cmds) }, 539 }; 540 541 static const struct drm_i915_cmd_table hsw_blt_ring_cmd_table[] = { 542 { gen7_common_cmds, ARRAY_SIZE(gen7_common_cmds) }, 543 { gen7_blt_cmds, ARRAY_SIZE(gen7_blt_cmds) }, 544 { hsw_blt_cmds, ARRAY_SIZE(hsw_blt_cmds) }, 545 }; 546 547 static const struct drm_i915_cmd_table gen9_blt_cmd_table[] = { 548 { gen9_blt_cmds, ARRAY_SIZE(gen9_blt_cmds) }, 549 }; 550 551 552 /* 553 * Register whitelists, sorted by increasing register offset. 554 */ 555 556 /* 557 * An individual whitelist entry granting access to register addr. If 558 * mask is non-zero the argument of immediate register writes will be 559 * AND-ed with mask, and the command will be rejected if the result 560 * doesn't match value. 561 * 562 * Registers with non-zero mask are only allowed to be written using 563 * LRI. 564 */ 565 struct drm_i915_reg_descriptor { 566 i915_reg_t addr; 567 u32 mask; 568 u32 value; 569 }; 570 571 /* Convenience macro for adding 32-bit registers. */ 572 #define REG32(_reg, ...) \ 573 { .addr = (_reg), __VA_ARGS__ } 574 575 /* 576 * Convenience macro for adding 64-bit registers. 577 * 578 * Some registers that userspace accesses are 64 bits. The register 579 * access commands only allow 32-bit accesses. Hence, we have to include 580 * entries for both halves of the 64-bit registers. 581 */ 582 #define REG64(_reg) \ 583 { .addr = _reg }, \ 584 { .addr = _reg ## _UDW } 585 586 #define REG64_IDX(_reg, idx) \ 587 { .addr = _reg(idx) }, \ 588 { .addr = _reg ## _UDW(idx) } 589 590 static const struct drm_i915_reg_descriptor gen7_render_regs[] = { 591 REG64(GPGPU_THREADS_DISPATCHED), 592 REG64(HS_INVOCATION_COUNT), 593 REG64(DS_INVOCATION_COUNT), 594 REG64(IA_VERTICES_COUNT), 595 REG64(IA_PRIMITIVES_COUNT), 596 REG64(VS_INVOCATION_COUNT), 597 REG64(GS_INVOCATION_COUNT), 598 REG64(GS_PRIMITIVES_COUNT), 599 REG64(CL_INVOCATION_COUNT), 600 REG64(CL_PRIMITIVES_COUNT), 601 REG64(PS_INVOCATION_COUNT), 602 REG64(PS_DEPTH_COUNT), 603 REG64_IDX(RING_TIMESTAMP, RENDER_RING_BASE), 604 REG64(MI_PREDICATE_SRC0), 605 REG64(MI_PREDICATE_SRC1), 606 REG32(GEN7_3DPRIM_END_OFFSET), 607 REG32(GEN7_3DPRIM_START_VERTEX), 608 REG32(GEN7_3DPRIM_VERTEX_COUNT), 609 REG32(GEN7_3DPRIM_INSTANCE_COUNT), 610 REG32(GEN7_3DPRIM_START_INSTANCE), 611 REG32(GEN7_3DPRIM_BASE_VERTEX), 612 REG32(GEN7_GPGPU_DISPATCHDIMX), 613 REG32(GEN7_GPGPU_DISPATCHDIMY), 614 REG32(GEN7_GPGPU_DISPATCHDIMZ), 615 REG64_IDX(RING_TIMESTAMP, BSD_RING_BASE), 616 REG64_IDX(GEN7_SO_NUM_PRIMS_WRITTEN, 0), 617 REG64_IDX(GEN7_SO_NUM_PRIMS_WRITTEN, 1), 618 REG64_IDX(GEN7_SO_NUM_PRIMS_WRITTEN, 2), 619 REG64_IDX(GEN7_SO_NUM_PRIMS_WRITTEN, 3), 620 REG64_IDX(GEN7_SO_PRIM_STORAGE_NEEDED, 0), 621 REG64_IDX(GEN7_SO_PRIM_STORAGE_NEEDED, 1), 622 REG64_IDX(GEN7_SO_PRIM_STORAGE_NEEDED, 2), 623 REG64_IDX(GEN7_SO_PRIM_STORAGE_NEEDED, 3), 624 REG32(GEN7_SO_WRITE_OFFSET(0)), 625 REG32(GEN7_SO_WRITE_OFFSET(1)), 626 REG32(GEN7_SO_WRITE_OFFSET(2)), 627 REG32(GEN7_SO_WRITE_OFFSET(3)), 628 REG32(GEN7_L3SQCREG1), 629 REG32(GEN7_L3CNTLREG2), 630 REG32(GEN7_L3CNTLREG3), 631 REG64_IDX(RING_TIMESTAMP, BLT_RING_BASE), 632 }; 633 634 static const struct drm_i915_reg_descriptor hsw_render_regs[] = { 635 REG64_IDX(HSW_CS_GPR, 0), 636 REG64_IDX(HSW_CS_GPR, 1), 637 REG64_IDX(HSW_CS_GPR, 2), 638 REG64_IDX(HSW_CS_GPR, 3), 639 REG64_IDX(HSW_CS_GPR, 4), 640 REG64_IDX(HSW_CS_GPR, 5), 641 REG64_IDX(HSW_CS_GPR, 6), 642 REG64_IDX(HSW_CS_GPR, 7), 643 REG64_IDX(HSW_CS_GPR, 8), 644 REG64_IDX(HSW_CS_GPR, 9), 645 REG64_IDX(HSW_CS_GPR, 10), 646 REG64_IDX(HSW_CS_GPR, 11), 647 REG64_IDX(HSW_CS_GPR, 12), 648 REG64_IDX(HSW_CS_GPR, 13), 649 REG64_IDX(HSW_CS_GPR, 14), 650 REG64_IDX(HSW_CS_GPR, 15), 651 REG32(HSW_SCRATCH1, 652 .mask = ~HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE, 653 .value = 0), 654 REG32(HSW_ROW_CHICKEN3, 655 .mask = ~(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE << 16 | 656 HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE), 657 .value = 0), 658 }; 659 660 static const struct drm_i915_reg_descriptor gen7_blt_regs[] = { 661 REG64_IDX(RING_TIMESTAMP, RENDER_RING_BASE), 662 REG64_IDX(RING_TIMESTAMP, BSD_RING_BASE), 663 REG32(BCS_SWCTRL), 664 REG64_IDX(RING_TIMESTAMP, BLT_RING_BASE), 665 }; 666 667 static const struct drm_i915_reg_descriptor gen9_blt_regs[] = { 668 REG64_IDX(RING_TIMESTAMP, RENDER_RING_BASE), 669 REG64_IDX(RING_TIMESTAMP, BSD_RING_BASE), 670 REG32(BCS_SWCTRL), 671 REG64_IDX(RING_TIMESTAMP, BLT_RING_BASE), 672 REG64_IDX(BCS_GPR, 0), 673 REG64_IDX(BCS_GPR, 1), 674 REG64_IDX(BCS_GPR, 2), 675 REG64_IDX(BCS_GPR, 3), 676 REG64_IDX(BCS_GPR, 4), 677 REG64_IDX(BCS_GPR, 5), 678 REG64_IDX(BCS_GPR, 6), 679 REG64_IDX(BCS_GPR, 7), 680 REG64_IDX(BCS_GPR, 8), 681 REG64_IDX(BCS_GPR, 9), 682 REG64_IDX(BCS_GPR, 10), 683 REG64_IDX(BCS_GPR, 11), 684 REG64_IDX(BCS_GPR, 12), 685 REG64_IDX(BCS_GPR, 13), 686 REG64_IDX(BCS_GPR, 14), 687 REG64_IDX(BCS_GPR, 15), 688 }; 689 690 #undef REG64 691 #undef REG32 692 693 struct drm_i915_reg_table { 694 const struct drm_i915_reg_descriptor *regs; 695 int num_regs; 696 }; 697 698 static const struct drm_i915_reg_table ivb_render_reg_tables[] = { 699 { gen7_render_regs, ARRAY_SIZE(gen7_render_regs) }, 700 }; 701 702 static const struct drm_i915_reg_table ivb_blt_reg_tables[] = { 703 { gen7_blt_regs, ARRAY_SIZE(gen7_blt_regs) }, 704 }; 705 706 static const struct drm_i915_reg_table hsw_render_reg_tables[] = { 707 { gen7_render_regs, ARRAY_SIZE(gen7_render_regs) }, 708 { hsw_render_regs, ARRAY_SIZE(hsw_render_regs) }, 709 }; 710 711 static const struct drm_i915_reg_table hsw_blt_reg_tables[] = { 712 { gen7_blt_regs, ARRAY_SIZE(gen7_blt_regs) }, 713 }; 714 715 static const struct drm_i915_reg_table gen9_blt_reg_tables[] = { 716 { gen9_blt_regs, ARRAY_SIZE(gen9_blt_regs) }, 717 }; 718 719 static u32 gen7_render_get_cmd_length_mask(u32 cmd_header) 720 { 721 u32 client = cmd_header >> INSTR_CLIENT_SHIFT; 722 u32 subclient = 723 (cmd_header & INSTR_SUBCLIENT_MASK) >> INSTR_SUBCLIENT_SHIFT; 724 725 if (client == INSTR_MI_CLIENT) 726 return 0x3F; 727 else if (client == INSTR_RC_CLIENT) { 728 if (subclient == INSTR_MEDIA_SUBCLIENT) 729 return 0xFFFF; 730 else 731 return 0xFF; 732 } 733 734 DRM_DEBUG("CMD: Abnormal rcs cmd length! 0x%08X\n", cmd_header); 735 return 0; 736 } 737 738 static u32 gen7_bsd_get_cmd_length_mask(u32 cmd_header) 739 { 740 u32 client = cmd_header >> INSTR_CLIENT_SHIFT; 741 u32 subclient = 742 (cmd_header & INSTR_SUBCLIENT_MASK) >> INSTR_SUBCLIENT_SHIFT; 743 u32 op = (cmd_header & INSTR_26_TO_24_MASK) >> INSTR_26_TO_24_SHIFT; 744 745 if (client == INSTR_MI_CLIENT) 746 return 0x3F; 747 else if (client == INSTR_RC_CLIENT) { 748 if (subclient == INSTR_MEDIA_SUBCLIENT) { 749 if (op == 6) 750 return 0xFFFF; 751 else 752 return 0xFFF; 753 } else 754 return 0xFF; 755 } 756 757 DRM_DEBUG("CMD: Abnormal bsd cmd length! 0x%08X\n", cmd_header); 758 return 0; 759 } 760 761 static u32 gen7_blt_get_cmd_length_mask(u32 cmd_header) 762 { 763 u32 client = cmd_header >> INSTR_CLIENT_SHIFT; 764 765 if (client == INSTR_MI_CLIENT) 766 return 0x3F; 767 else if (client == INSTR_BC_CLIENT) 768 return 0xFF; 769 770 DRM_DEBUG("CMD: Abnormal blt cmd length! 0x%08X\n", cmd_header); 771 return 0; 772 } 773 774 static u32 gen9_blt_get_cmd_length_mask(u32 cmd_header) 775 { 776 u32 client = cmd_header >> INSTR_CLIENT_SHIFT; 777 778 if (client == INSTR_MI_CLIENT || client == INSTR_BC_CLIENT) 779 return 0xFF; 780 781 DRM_DEBUG("CMD: Abnormal blt cmd length! 0x%08X\n", cmd_header); 782 return 0; 783 } 784 785 static bool validate_cmds_sorted(const struct intel_engine_cs *engine, 786 const struct drm_i915_cmd_table *cmd_tables, 787 int cmd_table_count) 788 { 789 int i; 790 bool ret = true; 791 792 if (!cmd_tables || cmd_table_count == 0) 793 return true; 794 795 for (i = 0; i < cmd_table_count; i++) { 796 const struct drm_i915_cmd_table *table = &cmd_tables[i]; 797 u32 previous = 0; 798 int j; 799 800 for (j = 0; j < table->count; j++) { 801 const struct drm_i915_cmd_descriptor *desc = 802 &table->table[j]; 803 u32 curr = desc->cmd.value & desc->cmd.mask; 804 805 if (curr < previous) { 806 drm_err(&engine->i915->drm, 807 "CMD: %s [%d] command table not sorted: " 808 "table=%d entry=%d cmd=0x%08X prev=0x%08X\n", 809 engine->name, engine->id, 810 i, j, curr, previous); 811 ret = false; 812 } 813 814 previous = curr; 815 } 816 } 817 818 return ret; 819 } 820 821 static bool check_sorted(const struct intel_engine_cs *engine, 822 const struct drm_i915_reg_descriptor *reg_table, 823 int reg_count) 824 { 825 int i; 826 u32 previous = 0; 827 bool ret = true; 828 829 for (i = 0; i < reg_count; i++) { 830 u32 curr = i915_mmio_reg_offset(reg_table[i].addr); 831 832 if (curr < previous) { 833 drm_err(&engine->i915->drm, 834 "CMD: %s [%d] register table not sorted: " 835 "entry=%d reg=0x%08X prev=0x%08X\n", 836 engine->name, engine->id, 837 i, curr, previous); 838 ret = false; 839 } 840 841 previous = curr; 842 } 843 844 return ret; 845 } 846 847 static bool validate_regs_sorted(struct intel_engine_cs *engine) 848 { 849 int i; 850 const struct drm_i915_reg_table *table; 851 852 for (i = 0; i < engine->reg_table_count; i++) { 853 table = &engine->reg_tables[i]; 854 if (!check_sorted(engine, table->regs, table->num_regs)) 855 return false; 856 } 857 858 return true; 859 } 860 861 struct cmd_node { 862 const struct drm_i915_cmd_descriptor *desc; 863 struct hlist_node node; 864 }; 865 866 /* 867 * Different command ranges have different numbers of bits for the opcode. For 868 * example, MI commands use bits 31:23 while 3D commands use bits 31:16. The 869 * problem is that, for example, MI commands use bits 22:16 for other fields 870 * such as GGTT vs PPGTT bits. If we include those bits in the mask then when 871 * we mask a command from a batch it could hash to the wrong bucket due to 872 * non-opcode bits being set. But if we don't include those bits, some 3D 873 * commands may hash to the same bucket due to not including opcode bits that 874 * make the command unique. For now, we will risk hashing to the same bucket. 875 */ 876 static inline u32 cmd_header_key(u32 x) 877 { 878 switch (x >> INSTR_CLIENT_SHIFT) { 879 default: 880 case INSTR_MI_CLIENT: 881 return x >> STD_MI_OPCODE_SHIFT; 882 case INSTR_RC_CLIENT: 883 return x >> STD_3D_OPCODE_SHIFT; 884 case INSTR_BC_CLIENT: 885 return x >> STD_2D_OPCODE_SHIFT; 886 } 887 } 888 889 static int init_hash_table(struct intel_engine_cs *engine, 890 const struct drm_i915_cmd_table *cmd_tables, 891 int cmd_table_count) 892 { 893 int i, j; 894 895 hash_init(engine->cmd_hash); 896 897 for (i = 0; i < cmd_table_count; i++) { 898 const struct drm_i915_cmd_table *table = &cmd_tables[i]; 899 900 for (j = 0; j < table->count; j++) { 901 const struct drm_i915_cmd_descriptor *desc = 902 &table->table[j]; 903 struct cmd_node *desc_node = 904 kmalloc(sizeof(*desc_node), GFP_KERNEL); 905 906 if (!desc_node) 907 return -ENOMEM; 908 909 desc_node->desc = desc; 910 hash_add(engine->cmd_hash, &desc_node->node, 911 cmd_header_key(desc->cmd.value)); 912 } 913 } 914 915 return 0; 916 } 917 918 static void fini_hash_table(struct intel_engine_cs *engine) 919 { 920 struct hlist_node *tmp; 921 struct cmd_node *desc_node; 922 int i; 923 924 hash_for_each_safe(engine->cmd_hash, i, tmp, desc_node, node) { 925 hash_del(&desc_node->node); 926 kfree(desc_node); 927 } 928 } 929 930 /** 931 * intel_engine_init_cmd_parser() - set cmd parser related fields for an engine 932 * @engine: the engine to initialize 933 * 934 * Optionally initializes fields related to batch buffer command parsing in the 935 * struct intel_engine_cs based on whether the platform requires software 936 * command parsing. 937 */ 938 void intel_engine_init_cmd_parser(struct intel_engine_cs *engine) 939 { 940 const struct drm_i915_cmd_table *cmd_tables; 941 int cmd_table_count; 942 int ret; 943 944 if (!IS_GEN(engine->i915, 7) && !(IS_GEN(engine->i915, 9) && 945 engine->class == COPY_ENGINE_CLASS)) 946 return; 947 948 switch (engine->class) { 949 case RENDER_CLASS: 950 if (IS_HASWELL(engine->i915)) { 951 cmd_tables = hsw_render_ring_cmd_table; 952 cmd_table_count = 953 ARRAY_SIZE(hsw_render_ring_cmd_table); 954 } else { 955 cmd_tables = gen7_render_cmd_table; 956 cmd_table_count = ARRAY_SIZE(gen7_render_cmd_table); 957 } 958 959 if (IS_HASWELL(engine->i915)) { 960 engine->reg_tables = hsw_render_reg_tables; 961 engine->reg_table_count = ARRAY_SIZE(hsw_render_reg_tables); 962 } else { 963 engine->reg_tables = ivb_render_reg_tables; 964 engine->reg_table_count = ARRAY_SIZE(ivb_render_reg_tables); 965 } 966 engine->get_cmd_length_mask = gen7_render_get_cmd_length_mask; 967 break; 968 case VIDEO_DECODE_CLASS: 969 cmd_tables = gen7_video_cmd_table; 970 cmd_table_count = ARRAY_SIZE(gen7_video_cmd_table); 971 engine->get_cmd_length_mask = gen7_bsd_get_cmd_length_mask; 972 break; 973 case COPY_ENGINE_CLASS: 974 engine->get_cmd_length_mask = gen7_blt_get_cmd_length_mask; 975 if (IS_GEN(engine->i915, 9)) { 976 cmd_tables = gen9_blt_cmd_table; 977 cmd_table_count = ARRAY_SIZE(gen9_blt_cmd_table); 978 engine->get_cmd_length_mask = 979 gen9_blt_get_cmd_length_mask; 980 981 /* BCS Engine unsafe without parser */ 982 engine->flags |= I915_ENGINE_REQUIRES_CMD_PARSER; 983 } else if (IS_HASWELL(engine->i915)) { 984 cmd_tables = hsw_blt_ring_cmd_table; 985 cmd_table_count = ARRAY_SIZE(hsw_blt_ring_cmd_table); 986 } else { 987 cmd_tables = gen7_blt_cmd_table; 988 cmd_table_count = ARRAY_SIZE(gen7_blt_cmd_table); 989 } 990 991 if (IS_GEN(engine->i915, 9)) { 992 engine->reg_tables = gen9_blt_reg_tables; 993 engine->reg_table_count = 994 ARRAY_SIZE(gen9_blt_reg_tables); 995 } else if (IS_HASWELL(engine->i915)) { 996 engine->reg_tables = hsw_blt_reg_tables; 997 engine->reg_table_count = ARRAY_SIZE(hsw_blt_reg_tables); 998 } else { 999 engine->reg_tables = ivb_blt_reg_tables; 1000 engine->reg_table_count = ARRAY_SIZE(ivb_blt_reg_tables); 1001 } 1002 break; 1003 case VIDEO_ENHANCEMENT_CLASS: 1004 cmd_tables = hsw_vebox_cmd_table; 1005 cmd_table_count = ARRAY_SIZE(hsw_vebox_cmd_table); 1006 /* VECS can use the same length_mask function as VCS */ 1007 engine->get_cmd_length_mask = gen7_bsd_get_cmd_length_mask; 1008 break; 1009 default: 1010 MISSING_CASE(engine->class); 1011 return; 1012 } 1013 1014 if (!validate_cmds_sorted(engine, cmd_tables, cmd_table_count)) { 1015 drm_err(&engine->i915->drm, 1016 "%s: command descriptions are not sorted\n", 1017 engine->name); 1018 return; 1019 } 1020 if (!validate_regs_sorted(engine)) { 1021 drm_err(&engine->i915->drm, 1022 "%s: registers are not sorted\n", engine->name); 1023 return; 1024 } 1025 1026 ret = init_hash_table(engine, cmd_tables, cmd_table_count); 1027 if (ret) { 1028 drm_err(&engine->i915->drm, 1029 "%s: initialised failed!\n", engine->name); 1030 fini_hash_table(engine); 1031 return; 1032 } 1033 1034 engine->flags |= I915_ENGINE_USING_CMD_PARSER; 1035 } 1036 1037 /** 1038 * intel_engine_cleanup_cmd_parser() - clean up cmd parser related fields 1039 * @engine: the engine to clean up 1040 * 1041 * Releases any resources related to command parsing that may have been 1042 * initialized for the specified engine. 1043 */ 1044 void intel_engine_cleanup_cmd_parser(struct intel_engine_cs *engine) 1045 { 1046 if (!intel_engine_using_cmd_parser(engine)) 1047 return; 1048 1049 fini_hash_table(engine); 1050 } 1051 1052 static const struct drm_i915_cmd_descriptor* 1053 find_cmd_in_table(struct intel_engine_cs *engine, 1054 u32 cmd_header) 1055 { 1056 struct cmd_node *desc_node; 1057 1058 hash_for_each_possible(engine->cmd_hash, desc_node, node, 1059 cmd_header_key(cmd_header)) { 1060 const struct drm_i915_cmd_descriptor *desc = desc_node->desc; 1061 if (((cmd_header ^ desc->cmd.value) & desc->cmd.mask) == 0) 1062 return desc; 1063 } 1064 1065 return NULL; 1066 } 1067 1068 /* 1069 * Returns a pointer to a descriptor for the command specified by cmd_header. 1070 * 1071 * The caller must supply space for a default descriptor via the default_desc 1072 * parameter. If no descriptor for the specified command exists in the engine's 1073 * command parser tables, this function fills in default_desc based on the 1074 * engine's default length encoding and returns default_desc. 1075 */ 1076 static const struct drm_i915_cmd_descriptor* 1077 find_cmd(struct intel_engine_cs *engine, 1078 u32 cmd_header, 1079 const struct drm_i915_cmd_descriptor *desc, 1080 struct drm_i915_cmd_descriptor *default_desc) 1081 { 1082 u32 mask; 1083 1084 if (((cmd_header ^ desc->cmd.value) & desc->cmd.mask) == 0) 1085 return desc; 1086 1087 desc = find_cmd_in_table(engine, cmd_header); 1088 if (desc) 1089 return desc; 1090 1091 mask = engine->get_cmd_length_mask(cmd_header); 1092 if (!mask) 1093 return NULL; 1094 1095 default_desc->cmd.value = cmd_header; 1096 default_desc->cmd.mask = ~0u << MIN_OPCODE_SHIFT; 1097 default_desc->length.mask = mask; 1098 default_desc->flags = CMD_DESC_SKIP; 1099 return default_desc; 1100 } 1101 1102 static const struct drm_i915_reg_descriptor * 1103 __find_reg(const struct drm_i915_reg_descriptor *table, int count, u32 addr) 1104 { 1105 int start = 0, end = count; 1106 while (start < end) { 1107 int mid = start + (end - start) / 2; 1108 int ret = addr - i915_mmio_reg_offset(table[mid].addr); 1109 if (ret < 0) 1110 end = mid; 1111 else if (ret > 0) 1112 start = mid + 1; 1113 else 1114 return &table[mid]; 1115 } 1116 return NULL; 1117 } 1118 1119 static const struct drm_i915_reg_descriptor * 1120 find_reg(const struct intel_engine_cs *engine, u32 addr) 1121 { 1122 const struct drm_i915_reg_table *table = engine->reg_tables; 1123 const struct drm_i915_reg_descriptor *reg = NULL; 1124 int count = engine->reg_table_count; 1125 1126 for (; !reg && (count > 0); ++table, --count) 1127 reg = __find_reg(table->regs, table->num_regs, addr); 1128 1129 return reg; 1130 } 1131 1132 /* Returns a vmap'd pointer to dst_obj, which the caller must unmap */ 1133 static u32 *copy_batch(struct drm_i915_gem_object *dst_obj, 1134 struct drm_i915_gem_object *src_obj, 1135 u32 offset, u32 length) 1136 { 1137 bool needs_clflush; 1138 void *dst, *src; 1139 int ret; 1140 1141 dst = i915_gem_object_pin_map(dst_obj, I915_MAP_FORCE_WB); 1142 if (IS_ERR(dst)) 1143 return dst; 1144 1145 ret = i915_gem_object_pin_pages(src_obj); 1146 if (ret) { 1147 i915_gem_object_unpin_map(dst_obj); 1148 return ERR_PTR(ret); 1149 } 1150 1151 needs_clflush = 1152 !(src_obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_READ); 1153 1154 src = ERR_PTR(-ENODEV); 1155 if (needs_clflush && i915_has_memcpy_from_wc()) { 1156 src = i915_gem_object_pin_map(src_obj, I915_MAP_WC); 1157 if (!IS_ERR(src)) { 1158 i915_unaligned_memcpy_from_wc(dst, 1159 src + offset, 1160 length); 1161 i915_gem_object_unpin_map(src_obj); 1162 } 1163 } 1164 if (IS_ERR(src)) { 1165 void *ptr; 1166 int x, n; 1167 1168 /* 1169 * We can avoid clflushing partial cachelines before the write 1170 * if we only every write full cache-lines. Since we know that 1171 * both the source and destination are in multiples of 1172 * PAGE_SIZE, we can simply round up to the next cacheline. 1173 * We don't care about copying too much here as we only 1174 * validate up to the end of the batch. 1175 */ 1176 if (!(dst_obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_READ)) 1177 length = round_up(length, 1178 boot_cpu_data.x86_clflush_size); 1179 1180 ptr = dst; 1181 x = offset_in_page(offset); 1182 for (n = offset >> PAGE_SHIFT; length; n++) { 1183 int len = min_t(int, length, PAGE_SIZE - x); 1184 1185 src = kmap_atomic(i915_gem_object_get_page(src_obj, n)); 1186 if (needs_clflush) 1187 drm_clflush_virt_range(src + x, len); 1188 memcpy(ptr, src + x, len); 1189 kunmap_atomic(src); 1190 1191 ptr += len; 1192 length -= len; 1193 x = 0; 1194 } 1195 } 1196 1197 i915_gem_object_unpin_pages(src_obj); 1198 1199 /* dst_obj is returned with vmap pinned */ 1200 return dst; 1201 } 1202 1203 static bool check_cmd(const struct intel_engine_cs *engine, 1204 const struct drm_i915_cmd_descriptor *desc, 1205 const u32 *cmd, u32 length) 1206 { 1207 if (desc->flags & CMD_DESC_SKIP) 1208 return true; 1209 1210 if (desc->flags & CMD_DESC_REJECT) { 1211 DRM_DEBUG("CMD: Rejected command: 0x%08X\n", *cmd); 1212 return false; 1213 } 1214 1215 if (desc->flags & CMD_DESC_REGISTER) { 1216 /* 1217 * Get the distance between individual register offset 1218 * fields if the command can perform more than one 1219 * access at a time. 1220 */ 1221 const u32 step = desc->reg.step ? desc->reg.step : length; 1222 u32 offset; 1223 1224 for (offset = desc->reg.offset; offset < length; 1225 offset += step) { 1226 const u32 reg_addr = cmd[offset] & desc->reg.mask; 1227 const struct drm_i915_reg_descriptor *reg = 1228 find_reg(engine, reg_addr); 1229 1230 if (!reg) { 1231 DRM_DEBUG("CMD: Rejected register 0x%08X in command: 0x%08X (%s)\n", 1232 reg_addr, *cmd, engine->name); 1233 return false; 1234 } 1235 1236 /* 1237 * Check the value written to the register against the 1238 * allowed mask/value pair given in the whitelist entry. 1239 */ 1240 if (reg->mask) { 1241 if (desc->cmd.value == MI_LOAD_REGISTER_MEM) { 1242 DRM_DEBUG("CMD: Rejected LRM to masked register 0x%08X\n", 1243 reg_addr); 1244 return false; 1245 } 1246 1247 if (desc->cmd.value == MI_LOAD_REGISTER_REG) { 1248 DRM_DEBUG("CMD: Rejected LRR to masked register 0x%08X\n", 1249 reg_addr); 1250 return false; 1251 } 1252 1253 if (desc->cmd.value == MI_LOAD_REGISTER_IMM(1) && 1254 (offset + 2 > length || 1255 (cmd[offset + 1] & reg->mask) != reg->value)) { 1256 DRM_DEBUG("CMD: Rejected LRI to masked register 0x%08X\n", 1257 reg_addr); 1258 return false; 1259 } 1260 } 1261 } 1262 } 1263 1264 if (desc->flags & CMD_DESC_BITMASK) { 1265 int i; 1266 1267 for (i = 0; i < MAX_CMD_DESC_BITMASKS; i++) { 1268 u32 dword; 1269 1270 if (desc->bits[i].mask == 0) 1271 break; 1272 1273 if (desc->bits[i].condition_mask != 0) { 1274 u32 offset = 1275 desc->bits[i].condition_offset; 1276 u32 condition = cmd[offset] & 1277 desc->bits[i].condition_mask; 1278 1279 if (condition == 0) 1280 continue; 1281 } 1282 1283 if (desc->bits[i].offset >= length) { 1284 DRM_DEBUG("CMD: Rejected command 0x%08X, too short to check bitmask (%s)\n", 1285 *cmd, engine->name); 1286 return false; 1287 } 1288 1289 dword = cmd[desc->bits[i].offset] & 1290 desc->bits[i].mask; 1291 1292 if (dword != desc->bits[i].expected) { 1293 DRM_DEBUG("CMD: Rejected command 0x%08X for bitmask 0x%08X (exp=0x%08X act=0x%08X) (%s)\n", 1294 *cmd, 1295 desc->bits[i].mask, 1296 desc->bits[i].expected, 1297 dword, engine->name); 1298 return false; 1299 } 1300 } 1301 } 1302 1303 return true; 1304 } 1305 1306 static int check_bbstart(u32 *cmd, u32 offset, u32 length, 1307 u32 batch_length, 1308 u64 batch_addr, 1309 u64 shadow_addr, 1310 const unsigned long *jump_whitelist) 1311 { 1312 u64 jump_offset, jump_target; 1313 u32 target_cmd_offset, target_cmd_index; 1314 1315 /* For igt compatibility on older platforms */ 1316 if (!jump_whitelist) { 1317 DRM_DEBUG("CMD: Rejecting BB_START for ggtt based submission\n"); 1318 return -EACCES; 1319 } 1320 1321 if (length != 3) { 1322 DRM_DEBUG("CMD: Recursive BB_START with bad length(%u)\n", 1323 length); 1324 return -EINVAL; 1325 } 1326 1327 jump_target = *(u64 *)(cmd + 1); 1328 jump_offset = jump_target - batch_addr; 1329 1330 /* 1331 * Any underflow of jump_target is guaranteed to be outside the range 1332 * of a u32, so >= test catches both too large and too small 1333 */ 1334 if (jump_offset >= batch_length) { 1335 DRM_DEBUG("CMD: BB_START to 0x%llx jumps out of BB\n", 1336 jump_target); 1337 return -EINVAL; 1338 } 1339 1340 /* 1341 * This cannot overflow a u32 because we already checked jump_offset 1342 * is within the BB, and the batch_length is a u32 1343 */ 1344 target_cmd_offset = lower_32_bits(jump_offset); 1345 target_cmd_index = target_cmd_offset / sizeof(u32); 1346 1347 *(u64 *)(cmd + 1) = shadow_addr + target_cmd_offset; 1348 1349 if (target_cmd_index == offset) 1350 return 0; 1351 1352 if (IS_ERR(jump_whitelist)) 1353 return PTR_ERR(jump_whitelist); 1354 1355 if (!test_bit(target_cmd_index, jump_whitelist)) { 1356 DRM_DEBUG("CMD: BB_START to 0x%llx not a previously executed cmd\n", 1357 jump_target); 1358 return -EINVAL; 1359 } 1360 1361 return 0; 1362 } 1363 1364 static unsigned long *alloc_whitelist(u32 batch_length) 1365 { 1366 unsigned long *jmp; 1367 1368 /* 1369 * We expect batch_length to be less than 256KiB for known users, 1370 * i.e. we need at most an 8KiB bitmap allocation which should be 1371 * reasonably cheap due to kmalloc caches. 1372 */ 1373 1374 /* Prefer to report transient allocation failure rather than hit oom */ 1375 jmp = bitmap_zalloc(DIV_ROUND_UP(batch_length, sizeof(u32)), 1376 GFP_KERNEL | __GFP_RETRY_MAYFAIL | __GFP_NOWARN); 1377 if (!jmp) 1378 return ERR_PTR(-ENOMEM); 1379 1380 return jmp; 1381 } 1382 1383 #define LENGTH_BIAS 2 1384 1385 static bool shadow_needs_clflush(struct drm_i915_gem_object *obj) 1386 { 1387 return !(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_WRITE); 1388 } 1389 1390 /** 1391 * intel_engine_cmd_parser() - parse a batch buffer for privilege violations 1392 * @engine: the engine on which the batch is to execute 1393 * @batch: the batch buffer in question 1394 * @batch_offset: byte offset in the batch at which execution starts 1395 * @batch_length: length of the commands in batch_obj 1396 * @shadow: validated copy of the batch buffer in question 1397 * @trampoline: whether to emit a conditional trampoline at the end of the batch 1398 * 1399 * Parses the specified batch buffer looking for privilege violations as 1400 * described in the overview. 1401 * 1402 * Return: non-zero if the parser finds violations or otherwise fails; -EACCES 1403 * if the batch appears legal but should use hardware parsing 1404 */ 1405 int intel_engine_cmd_parser(struct intel_engine_cs *engine, 1406 struct i915_vma *batch, 1407 u32 batch_offset, 1408 u32 batch_length, 1409 struct i915_vma *shadow, 1410 bool trampoline) 1411 { 1412 u32 *cmd, *batch_end, offset = 0; 1413 struct drm_i915_cmd_descriptor default_desc = noop_desc; 1414 const struct drm_i915_cmd_descriptor *desc = &default_desc; 1415 unsigned long *jump_whitelist; 1416 u64 batch_addr, shadow_addr; 1417 int ret = 0; 1418 1419 GEM_BUG_ON(!IS_ALIGNED(batch_offset, sizeof(*cmd))); 1420 GEM_BUG_ON(!IS_ALIGNED(batch_length, sizeof(*cmd))); 1421 GEM_BUG_ON(range_overflows_t(u64, batch_offset, batch_length, 1422 batch->size)); 1423 GEM_BUG_ON(!batch_length); 1424 1425 cmd = copy_batch(shadow->obj, batch->obj, batch_offset, batch_length); 1426 if (IS_ERR(cmd)) { 1427 DRM_DEBUG("CMD: Failed to copy batch\n"); 1428 return PTR_ERR(cmd); 1429 } 1430 1431 jump_whitelist = NULL; 1432 if (!trampoline) 1433 /* Defer failure until attempted use */ 1434 jump_whitelist = alloc_whitelist(batch_length); 1435 1436 shadow_addr = gen8_canonical_addr(shadow->node.start); 1437 batch_addr = gen8_canonical_addr(batch->node.start + batch_offset); 1438 1439 /* 1440 * We use the batch length as size because the shadow object is as 1441 * large or larger and copy_batch() will write MI_NOPs to the extra 1442 * space. Parsing should be faster in some cases this way. 1443 */ 1444 batch_end = cmd + batch_length / sizeof(*batch_end); 1445 do { 1446 u32 length; 1447 1448 if (*cmd == MI_BATCH_BUFFER_END) 1449 break; 1450 1451 desc = find_cmd(engine, *cmd, desc, &default_desc); 1452 if (!desc) { 1453 DRM_DEBUG("CMD: Unrecognized command: 0x%08X\n", *cmd); 1454 ret = -EINVAL; 1455 break; 1456 } 1457 1458 if (desc->flags & CMD_DESC_FIXED) 1459 length = desc->length.fixed; 1460 else 1461 length = (*cmd & desc->length.mask) + LENGTH_BIAS; 1462 1463 if ((batch_end - cmd) < length) { 1464 DRM_DEBUG("CMD: Command length exceeds batch length: 0x%08X length=%u batchlen=%td\n", 1465 *cmd, 1466 length, 1467 batch_end - cmd); 1468 ret = -EINVAL; 1469 break; 1470 } 1471 1472 if (!check_cmd(engine, desc, cmd, length)) { 1473 ret = -EACCES; 1474 break; 1475 } 1476 1477 if (desc->cmd.value == MI_BATCH_BUFFER_START) { 1478 ret = check_bbstart(cmd, offset, length, batch_length, 1479 batch_addr, shadow_addr, 1480 jump_whitelist); 1481 break; 1482 } 1483 1484 if (!IS_ERR_OR_NULL(jump_whitelist)) 1485 __set_bit(offset, jump_whitelist); 1486 1487 cmd += length; 1488 offset += length; 1489 if (cmd >= batch_end) { 1490 DRM_DEBUG("CMD: Got to the end of the buffer w/o a BBE cmd!\n"); 1491 ret = -EINVAL; 1492 break; 1493 } 1494 } while (1); 1495 1496 if (trampoline) { 1497 /* 1498 * With the trampoline, the shadow is executed twice. 1499 * 1500 * 1 - starting at offset 0, in privileged mode 1501 * 2 - starting at offset batch_len, as non-privileged 1502 * 1503 * Only if the batch is valid and safe to execute, do we 1504 * allow the first privileged execution to proceed. If not, 1505 * we terminate the first batch and use the second batchbuffer 1506 * entry to chain to the original unsafe non-privileged batch, 1507 * leaving it to the HW to validate. 1508 */ 1509 *batch_end = MI_BATCH_BUFFER_END; 1510 1511 if (ret) { 1512 /* Batch unsafe to execute with privileges, cancel! */ 1513 cmd = page_mask_bits(shadow->obj->mm.mapping); 1514 *cmd = MI_BATCH_BUFFER_END; 1515 1516 /* If batch is unsafe but valid, jump to the original */ 1517 if (ret == -EACCES) { 1518 unsigned int flags; 1519 1520 flags = MI_BATCH_NON_SECURE_I965; 1521 if (IS_HASWELL(engine->i915)) 1522 flags = MI_BATCH_NON_SECURE_HSW; 1523 1524 GEM_BUG_ON(!IS_GEN_RANGE(engine->i915, 6, 7)); 1525 __gen6_emit_bb_start(batch_end, 1526 batch_addr, 1527 flags); 1528 1529 ret = 0; /* allow execution */ 1530 } 1531 } 1532 1533 if (shadow_needs_clflush(shadow->obj)) 1534 drm_clflush_virt_range(batch_end, 8); 1535 } 1536 1537 if (shadow_needs_clflush(shadow->obj)) { 1538 void *ptr = page_mask_bits(shadow->obj->mm.mapping); 1539 1540 drm_clflush_virt_range(ptr, (void *)(cmd + 1) - ptr); 1541 } 1542 1543 if (!IS_ERR_OR_NULL(jump_whitelist)) 1544 kfree(jump_whitelist); 1545 i915_gem_object_unpin_map(shadow->obj); 1546 return ret; 1547 } 1548 1549 /** 1550 * i915_cmd_parser_get_version() - get the cmd parser version number 1551 * @dev_priv: i915 device private 1552 * 1553 * The cmd parser maintains a simple increasing integer version number suitable 1554 * for passing to userspace clients to determine what operations are permitted. 1555 * 1556 * Return: the current version number of the cmd parser 1557 */ 1558 int i915_cmd_parser_get_version(struct drm_i915_private *dev_priv) 1559 { 1560 struct intel_engine_cs *engine; 1561 bool active = false; 1562 1563 /* If the command parser is not enabled, report 0 - unsupported */ 1564 for_each_uabi_engine(engine, dev_priv) { 1565 if (intel_engine_using_cmd_parser(engine)) { 1566 active = true; 1567 break; 1568 } 1569 } 1570 if (!active) 1571 return 0; 1572 1573 /* 1574 * Command parser version history 1575 * 1576 * 1. Initial version. Checks batches and reports violations, but leaves 1577 * hardware parsing enabled (so does not allow new use cases). 1578 * 2. Allow access to the MI_PREDICATE_SRC0 and 1579 * MI_PREDICATE_SRC1 registers. 1580 * 3. Allow access to the GPGPU_THREADS_DISPATCHED register. 1581 * 4. L3 atomic chicken bits of HSW_SCRATCH1 and HSW_ROW_CHICKEN3. 1582 * 5. GPGPU dispatch compute indirect registers. 1583 * 6. TIMESTAMP register and Haswell CS GPR registers 1584 * 7. Allow MI_LOAD_REGISTER_REG between whitelisted registers. 1585 * 8. Don't report cmd_check() failures as EINVAL errors to userspace; 1586 * rely on the HW to NOOP disallowed commands as it would without 1587 * the parser enabled. 1588 * 9. Don't whitelist or handle oacontrol specially, as ownership 1589 * for oacontrol state is moving to i915-perf. 1590 * 10. Support for Gen9 BCS Parsing 1591 */ 1592 return 10; 1593 } 1594