1 // SPDX-License-Identifier: MIT 2 /* 3 * Copyright © 2021 Intel Corporation 4 */ 5 6 #include "xe_hw_engine.h" 7 8 #include <linux/nospec.h> 9 10 #include <drm/drm_managed.h> 11 #include <drm/drm_print.h> 12 #include <uapi/drm/xe_drm.h> 13 #include <generated/xe_wa_oob.h> 14 15 #include "regs/xe_engine_regs.h" 16 #include "regs/xe_gt_regs.h" 17 #include "regs/xe_irq_regs.h" 18 #include "xe_assert.h" 19 #include "xe_bo.h" 20 #include "xe_configfs.h" 21 #include "xe_device.h" 22 #include "xe_execlist.h" 23 #include "xe_force_wake.h" 24 #include "xe_gsc.h" 25 #include "xe_gt.h" 26 #include "xe_gt_ccs_mode.h" 27 #include "xe_gt_clock.h" 28 #include "xe_gt_printk.h" 29 #include "xe_gt_mcr.h" 30 #include "xe_gt_topology.h" 31 #include "xe_guc_capture.h" 32 #include "xe_hw_engine_group.h" 33 #include "xe_hw_fence.h" 34 #include "xe_irq.h" 35 #include "xe_lrc.h" 36 #include "xe_mmio.h" 37 #include "xe_reg_sr.h" 38 #include "xe_reg_whitelist.h" 39 #include "xe_rtp.h" 40 #include "xe_sched_job.h" 41 #include "xe_sriov.h" 42 #include "xe_tuning.h" 43 #include "xe_uc_fw.h" 44 #include "xe_wa.h" 45 46 #define MAX_MMIO_BASES 3 47 struct engine_info { 48 const char *name; 49 unsigned int class : 8; 50 unsigned int instance : 8; 51 unsigned int irq_offset : 8; 52 enum xe_force_wake_domains domain; 53 u32 mmio_base; 54 }; 55 56 static const struct engine_info engine_infos[] = { 57 [XE_HW_ENGINE_RCS0] = { 58 .name = "rcs0", 59 .class = XE_ENGINE_CLASS_RENDER, 60 .instance = 0, 61 .irq_offset = ilog2(INTR_RCS0), 62 .domain = XE_FW_RENDER, 63 .mmio_base = RENDER_RING_BASE, 64 }, 65 [XE_HW_ENGINE_BCS0] = { 66 .name = "bcs0", 67 .class = XE_ENGINE_CLASS_COPY, 68 .instance = 0, 69 .irq_offset = ilog2(INTR_BCS(0)), 70 .domain = XE_FW_RENDER, 71 .mmio_base = BLT_RING_BASE, 72 }, 73 [XE_HW_ENGINE_BCS1] = { 74 .name = "bcs1", 75 .class = XE_ENGINE_CLASS_COPY, 76 .instance = 1, 77 .irq_offset = ilog2(INTR_BCS(1)), 78 .domain = XE_FW_RENDER, 79 .mmio_base = XEHPC_BCS1_RING_BASE, 80 }, 81 [XE_HW_ENGINE_BCS2] = { 82 .name = "bcs2", 83 .class = XE_ENGINE_CLASS_COPY, 84 .instance = 2, 85 .irq_offset = ilog2(INTR_BCS(2)), 86 .domain = XE_FW_RENDER, 87 .mmio_base = XEHPC_BCS2_RING_BASE, 88 }, 89 [XE_HW_ENGINE_BCS3] = { 90 .name = "bcs3", 91 .class = XE_ENGINE_CLASS_COPY, 92 .instance = 3, 93 .irq_offset = ilog2(INTR_BCS(3)), 94 .domain = XE_FW_RENDER, 95 .mmio_base = XEHPC_BCS3_RING_BASE, 96 }, 97 [XE_HW_ENGINE_BCS4] = { 98 .name = "bcs4", 99 .class = XE_ENGINE_CLASS_COPY, 100 .instance = 4, 101 .irq_offset = ilog2(INTR_BCS(4)), 102 .domain = XE_FW_RENDER, 103 .mmio_base = XEHPC_BCS4_RING_BASE, 104 }, 105 [XE_HW_ENGINE_BCS5] = { 106 .name = "bcs5", 107 .class = XE_ENGINE_CLASS_COPY, 108 .instance = 5, 109 .irq_offset = ilog2(INTR_BCS(5)), 110 .domain = XE_FW_RENDER, 111 .mmio_base = XEHPC_BCS5_RING_BASE, 112 }, 113 [XE_HW_ENGINE_BCS6] = { 114 .name = "bcs6", 115 .class = XE_ENGINE_CLASS_COPY, 116 .instance = 6, 117 .irq_offset = ilog2(INTR_BCS(6)), 118 .domain = XE_FW_RENDER, 119 .mmio_base = XEHPC_BCS6_RING_BASE, 120 }, 121 [XE_HW_ENGINE_BCS7] = { 122 .name = "bcs7", 123 .class = XE_ENGINE_CLASS_COPY, 124 .irq_offset = ilog2(INTR_BCS(7)), 125 .instance = 7, 126 .domain = XE_FW_RENDER, 127 .mmio_base = XEHPC_BCS7_RING_BASE, 128 }, 129 [XE_HW_ENGINE_BCS8] = { 130 .name = "bcs8", 131 .class = XE_ENGINE_CLASS_COPY, 132 .instance = 8, 133 .irq_offset = ilog2(INTR_BCS8), 134 .domain = XE_FW_RENDER, 135 .mmio_base = XEHPC_BCS8_RING_BASE, 136 }, 137 138 [XE_HW_ENGINE_VCS0] = { 139 .name = "vcs0", 140 .class = XE_ENGINE_CLASS_VIDEO_DECODE, 141 .instance = 0, 142 .irq_offset = 32 + ilog2(INTR_VCS(0)), 143 .domain = XE_FW_MEDIA_VDBOX0, 144 .mmio_base = BSD_RING_BASE, 145 }, 146 [XE_HW_ENGINE_VCS1] = { 147 .name = "vcs1", 148 .class = XE_ENGINE_CLASS_VIDEO_DECODE, 149 .instance = 1, 150 .irq_offset = 32 + ilog2(INTR_VCS(1)), 151 .domain = XE_FW_MEDIA_VDBOX1, 152 .mmio_base = BSD2_RING_BASE, 153 }, 154 [XE_HW_ENGINE_VCS2] = { 155 .name = "vcs2", 156 .class = XE_ENGINE_CLASS_VIDEO_DECODE, 157 .instance = 2, 158 .irq_offset = 32 + ilog2(INTR_VCS(2)), 159 .domain = XE_FW_MEDIA_VDBOX2, 160 .mmio_base = BSD3_RING_BASE, 161 }, 162 [XE_HW_ENGINE_VCS3] = { 163 .name = "vcs3", 164 .class = XE_ENGINE_CLASS_VIDEO_DECODE, 165 .instance = 3, 166 .irq_offset = 32 + ilog2(INTR_VCS(3)), 167 .domain = XE_FW_MEDIA_VDBOX3, 168 .mmio_base = BSD4_RING_BASE, 169 }, 170 [XE_HW_ENGINE_VCS4] = { 171 .name = "vcs4", 172 .class = XE_ENGINE_CLASS_VIDEO_DECODE, 173 .instance = 4, 174 .irq_offset = 32 + ilog2(INTR_VCS(4)), 175 .domain = XE_FW_MEDIA_VDBOX4, 176 .mmio_base = XEHP_BSD5_RING_BASE, 177 }, 178 [XE_HW_ENGINE_VCS5] = { 179 .name = "vcs5", 180 .class = XE_ENGINE_CLASS_VIDEO_DECODE, 181 .instance = 5, 182 .irq_offset = 32 + ilog2(INTR_VCS(5)), 183 .domain = XE_FW_MEDIA_VDBOX5, 184 .mmio_base = XEHP_BSD6_RING_BASE, 185 }, 186 [XE_HW_ENGINE_VCS6] = { 187 .name = "vcs6", 188 .class = XE_ENGINE_CLASS_VIDEO_DECODE, 189 .instance = 6, 190 .irq_offset = 32 + ilog2(INTR_VCS(6)), 191 .domain = XE_FW_MEDIA_VDBOX6, 192 .mmio_base = XEHP_BSD7_RING_BASE, 193 }, 194 [XE_HW_ENGINE_VCS7] = { 195 .name = "vcs7", 196 .class = XE_ENGINE_CLASS_VIDEO_DECODE, 197 .instance = 7, 198 .irq_offset = 32 + ilog2(INTR_VCS(7)), 199 .domain = XE_FW_MEDIA_VDBOX7, 200 .mmio_base = XEHP_BSD8_RING_BASE, 201 }, 202 [XE_HW_ENGINE_VECS0] = { 203 .name = "vecs0", 204 .class = XE_ENGINE_CLASS_VIDEO_ENHANCE, 205 .instance = 0, 206 .irq_offset = 32 + ilog2(INTR_VECS(0)), 207 .domain = XE_FW_MEDIA_VEBOX0, 208 .mmio_base = VEBOX_RING_BASE, 209 }, 210 [XE_HW_ENGINE_VECS1] = { 211 .name = "vecs1", 212 .class = XE_ENGINE_CLASS_VIDEO_ENHANCE, 213 .instance = 1, 214 .irq_offset = 32 + ilog2(INTR_VECS(1)), 215 .domain = XE_FW_MEDIA_VEBOX1, 216 .mmio_base = VEBOX2_RING_BASE, 217 }, 218 [XE_HW_ENGINE_VECS2] = { 219 .name = "vecs2", 220 .class = XE_ENGINE_CLASS_VIDEO_ENHANCE, 221 .instance = 2, 222 .irq_offset = 32 + ilog2(INTR_VECS(2)), 223 .domain = XE_FW_MEDIA_VEBOX2, 224 .mmio_base = XEHP_VEBOX3_RING_BASE, 225 }, 226 [XE_HW_ENGINE_VECS3] = { 227 .name = "vecs3", 228 .class = XE_ENGINE_CLASS_VIDEO_ENHANCE, 229 .instance = 3, 230 .irq_offset = 32 + ilog2(INTR_VECS(3)), 231 .domain = XE_FW_MEDIA_VEBOX3, 232 .mmio_base = XEHP_VEBOX4_RING_BASE, 233 }, 234 [XE_HW_ENGINE_CCS0] = { 235 .name = "ccs0", 236 .class = XE_ENGINE_CLASS_COMPUTE, 237 .instance = 0, 238 .irq_offset = ilog2(INTR_CCS(0)), 239 .domain = XE_FW_RENDER, 240 .mmio_base = COMPUTE0_RING_BASE, 241 }, 242 [XE_HW_ENGINE_CCS1] = { 243 .name = "ccs1", 244 .class = XE_ENGINE_CLASS_COMPUTE, 245 .instance = 1, 246 .irq_offset = ilog2(INTR_CCS(1)), 247 .domain = XE_FW_RENDER, 248 .mmio_base = COMPUTE1_RING_BASE, 249 }, 250 [XE_HW_ENGINE_CCS2] = { 251 .name = "ccs2", 252 .class = XE_ENGINE_CLASS_COMPUTE, 253 .instance = 2, 254 .irq_offset = ilog2(INTR_CCS(2)), 255 .domain = XE_FW_RENDER, 256 .mmio_base = COMPUTE2_RING_BASE, 257 }, 258 [XE_HW_ENGINE_CCS3] = { 259 .name = "ccs3", 260 .class = XE_ENGINE_CLASS_COMPUTE, 261 .instance = 3, 262 .irq_offset = ilog2(INTR_CCS(3)), 263 .domain = XE_FW_RENDER, 264 .mmio_base = COMPUTE3_RING_BASE, 265 }, 266 [XE_HW_ENGINE_GSCCS0] = { 267 .name = "gsccs0", 268 .class = XE_ENGINE_CLASS_OTHER, 269 .instance = OTHER_GSC_INSTANCE, 270 .domain = XE_FW_GSC, 271 .mmio_base = GSCCS_RING_BASE, 272 }, 273 }; 274 275 static void hw_engine_fini(void *arg) 276 { 277 struct xe_hw_engine *hwe = arg; 278 279 if (hwe->exl_port) 280 xe_execlist_port_destroy(hwe->exl_port); 281 282 hwe->gt = NULL; 283 } 284 285 /** 286 * xe_hw_engine_mmio_write32() - Write engine register 287 * @hwe: engine 288 * @reg: register to write into 289 * @val: desired 32-bit value to write 290 * 291 * This function will write val into an engine specific register. 292 * Forcewake must be held by the caller. 293 * 294 */ 295 void xe_hw_engine_mmio_write32(struct xe_hw_engine *hwe, 296 struct xe_reg reg, u32 val) 297 { 298 xe_gt_assert(hwe->gt, !(reg.addr & hwe->mmio_base)); 299 xe_force_wake_assert_held(gt_to_fw(hwe->gt), hwe->domain); 300 301 reg.addr += hwe->mmio_base; 302 303 xe_mmio_write32(&hwe->gt->mmio, reg, val); 304 } 305 306 /** 307 * xe_hw_engine_mmio_read32() - Read engine register 308 * @hwe: engine 309 * @reg: register to read from 310 * 311 * This function will read from an engine specific register. 312 * Forcewake must be held by the caller. 313 * 314 * Return: value of the 32-bit register. 315 */ 316 u32 xe_hw_engine_mmio_read32(struct xe_hw_engine *hwe, struct xe_reg reg) 317 { 318 xe_gt_assert(hwe->gt, !(reg.addr & hwe->mmio_base)); 319 xe_force_wake_assert_held(gt_to_fw(hwe->gt), hwe->domain); 320 321 reg.addr += hwe->mmio_base; 322 323 return xe_mmio_read32(&hwe->gt->mmio, reg); 324 } 325 326 void xe_hw_engine_enable_ring(struct xe_hw_engine *hwe) 327 { 328 u32 ccs_mask = 329 xe_hw_engine_mask_per_class(hwe->gt, XE_ENGINE_CLASS_COMPUTE); 330 u32 ring_mode = _MASKED_BIT_ENABLE(GFX_DISABLE_LEGACY_MODE); 331 332 if (hwe->class == XE_ENGINE_CLASS_COMPUTE && ccs_mask) 333 xe_mmio_write32(&hwe->gt->mmio, RCU_MODE, 334 _MASKED_BIT_ENABLE(RCU_MODE_CCS_ENABLE)); 335 336 xe_hw_engine_mmio_write32(hwe, RING_HWSTAM(0), ~0x0); 337 xe_hw_engine_mmio_write32(hwe, RING_HWS_PGA(0), 338 xe_bo_ggtt_addr(hwe->hwsp)); 339 340 if (xe_device_has_msix(gt_to_xe(hwe->gt))) 341 ring_mode |= _MASKED_BIT_ENABLE(GFX_MSIX_INTERRUPT_ENABLE); 342 xe_hw_engine_mmio_write32(hwe, RING_MODE(0), ring_mode); 343 xe_hw_engine_mmio_write32(hwe, RING_MI_MODE(0), 344 _MASKED_BIT_DISABLE(STOP_RING)); 345 xe_hw_engine_mmio_read32(hwe, RING_MI_MODE(0)); 346 } 347 348 static bool xe_hw_engine_match_fixed_cslice_mode(const struct xe_device *xe, 349 const struct xe_gt *gt, 350 const struct xe_hw_engine *hwe) 351 { 352 /* 353 * Xe3p no longer supports load balance mode, so "fixed cslice" mode 354 * is automatic and no RCU_MODE programming is required. 355 */ 356 if (GRAPHICS_VER(gt_to_xe(gt)) >= 35) 357 return false; 358 359 return xe_gt_ccs_mode_enabled(gt) && 360 xe_rtp_match_first_render_or_compute(xe, gt, hwe); 361 } 362 363 static bool xe_rtp_cfeg_wmtp_disabled(const struct xe_device *xe, 364 const struct xe_gt *gt, 365 const struct xe_hw_engine *hwe) 366 { 367 if (GRAPHICS_VER(xe) < 20) 368 return false; 369 370 if (hwe->class != XE_ENGINE_CLASS_COMPUTE && 371 hwe->class != XE_ENGINE_CLASS_RENDER) 372 return false; 373 374 return xe_mmio_read32(&hwe->gt->mmio, XEHP_FUSE4) & CFEG_WMTP_DISABLE; 375 } 376 377 void 378 xe_hw_engine_setup_default_lrc_state(struct xe_hw_engine *hwe) 379 { 380 struct xe_gt *gt = hwe->gt; 381 const u8 mocs_write_idx = gt->mocs.uc_index; 382 const u8 mocs_read_idx = gt->mocs.uc_index; 383 u32 blit_cctl_val = REG_FIELD_PREP(BLIT_CCTL_DST_MOCS_MASK, mocs_write_idx) | 384 REG_FIELD_PREP(BLIT_CCTL_SRC_MOCS_MASK, mocs_read_idx); 385 struct xe_rtp_process_ctx ctx = XE_RTP_PROCESS_CTX_INITIALIZER(hwe); 386 const struct xe_rtp_entry_sr lrc_setup[] = { 387 /* 388 * Some blitter commands do not have a field for MOCS, those 389 * commands will use MOCS index pointed by BLIT_CCTL. 390 * BLIT_CCTL registers are needed to be programmed to un-cached. 391 */ 392 { XE_RTP_NAME("BLIT_CCTL_default_MOCS"), 393 XE_RTP_RULES(GRAPHICS_VERSION_RANGE(1200, XE_RTP_END_VERSION_UNDEFINED), 394 ENGINE_CLASS(COPY)), 395 XE_RTP_ACTIONS(FIELD_SET(BLIT_CCTL(0), 396 BLIT_CCTL_DST_MOCS_MASK | 397 BLIT_CCTL_SRC_MOCS_MASK, 398 blit_cctl_val, 399 XE_RTP_ACTION_FLAG(ENGINE_BASE))) 400 }, 401 /* Disable WMTP if HW doesn't support it */ 402 { XE_RTP_NAME("DISABLE_WMTP_ON_UNSUPPORTED_HW"), 403 XE_RTP_RULES(FUNC(xe_rtp_cfeg_wmtp_disabled)), 404 XE_RTP_ACTIONS(FIELD_SET(CS_CHICKEN1(0), 405 PREEMPT_GPGPU_LEVEL_MASK, 406 PREEMPT_GPGPU_THREAD_GROUP_LEVEL)), 407 XE_RTP_ENTRY_FLAG(FOREACH_ENGINE) 408 }, 409 }; 410 411 xe_rtp_process_to_sr(&ctx, lrc_setup, ARRAY_SIZE(lrc_setup), &hwe->reg_lrc); 412 } 413 414 static void 415 hw_engine_setup_default_state(struct xe_hw_engine *hwe) 416 { 417 struct xe_gt *gt = hwe->gt; 418 struct xe_device *xe = gt_to_xe(gt); 419 /* 420 * RING_CMD_CCTL specifies the default MOCS entry that will be 421 * used by the command streamer when executing commands that 422 * don't have a way to explicitly specify a MOCS setting. 423 * The default should usually reference whichever MOCS entry 424 * corresponds to uncached behavior, although use of a WB cached 425 * entry is recommended by the spec in certain circumstances on 426 * specific platforms. 427 * Bspec: 72161 428 */ 429 const u8 mocs_write_idx = gt->mocs.uc_index; 430 const u8 mocs_read_idx = hwe->class == XE_ENGINE_CLASS_COMPUTE && IS_DGFX(xe) && 431 (GRAPHICS_VER(xe) >= 20 || xe->info.platform == XE_PVC) ? 432 gt->mocs.wb_index : gt->mocs.uc_index; 433 u32 ring_cmd_cctl_val = REG_FIELD_PREP(CMD_CCTL_WRITE_OVERRIDE_MASK, mocs_write_idx) | 434 REG_FIELD_PREP(CMD_CCTL_READ_OVERRIDE_MASK, mocs_read_idx); 435 struct xe_rtp_process_ctx ctx = XE_RTP_PROCESS_CTX_INITIALIZER(hwe); 436 const struct xe_rtp_entry_sr engine_entries[] = { 437 { XE_RTP_NAME("RING_CMD_CCTL_default_MOCS"), 438 XE_RTP_RULES(GRAPHICS_VERSION_RANGE(1200, XE_RTP_END_VERSION_UNDEFINED)), 439 XE_RTP_ACTIONS(FIELD_SET(RING_CMD_CCTL(0), 440 CMD_CCTL_WRITE_OVERRIDE_MASK | 441 CMD_CCTL_READ_OVERRIDE_MASK, 442 ring_cmd_cctl_val, 443 XE_RTP_ACTION_FLAG(ENGINE_BASE))) 444 }, 445 /* 446 * To allow the GSC engine to go idle on MTL we need to enable 447 * idle messaging and set the hysteresis value (we use 0xA=5us 448 * as recommended in spec). On platforms after MTL this is 449 * enabled by default. 450 */ 451 { XE_RTP_NAME("MTL GSCCS IDLE MSG enable"), 452 XE_RTP_RULES(MEDIA_VERSION(1300), ENGINE_CLASS(OTHER)), 453 XE_RTP_ACTIONS(CLR(RING_PSMI_CTL(0), 454 IDLE_MSG_DISABLE, 455 XE_RTP_ACTION_FLAG(ENGINE_BASE)), 456 FIELD_SET(RING_PWRCTX_MAXCNT(0), 457 IDLE_WAIT_TIME, 458 0xA, 459 XE_RTP_ACTION_FLAG(ENGINE_BASE))) 460 }, 461 /* Enable Priority Mem Read */ 462 { XE_RTP_NAME("Priority_Mem_Read"), 463 XE_RTP_RULES(GRAPHICS_VERSION_RANGE(2001, XE_RTP_END_VERSION_UNDEFINED)), 464 XE_RTP_ACTIONS(SET(CSFE_CHICKEN1(0), CS_PRIORITY_MEM_READ, 465 XE_RTP_ACTION_FLAG(ENGINE_BASE))) 466 }, 467 /* Use Fixed slice CCS mode */ 468 { XE_RTP_NAME("RCU_MODE_FIXED_SLICE_CCS_MODE"), 469 XE_RTP_RULES(FUNC(xe_hw_engine_match_fixed_cslice_mode)), 470 XE_RTP_ACTIONS(FIELD_SET(RCU_MODE, RCU_MODE_FIXED_SLICE_CCS_MODE, 471 RCU_MODE_FIXED_SLICE_CCS_MODE)) 472 }, 473 }; 474 475 xe_rtp_process_to_sr(&ctx, engine_entries, ARRAY_SIZE(engine_entries), &hwe->reg_sr); 476 } 477 478 static const struct engine_info *find_engine_info(enum xe_engine_class class, int instance) 479 { 480 const struct engine_info *info; 481 enum xe_hw_engine_id id; 482 483 for (id = 0; id < XE_NUM_HW_ENGINES; ++id) { 484 info = &engine_infos[id]; 485 if (info->class == class && info->instance == instance) 486 return info; 487 } 488 489 return NULL; 490 } 491 492 static u16 get_msix_irq_offset(struct xe_gt *gt, enum xe_engine_class class) 493 { 494 /* For MSI-X, hw engines report to offset of engine instance zero */ 495 const struct engine_info *info = find_engine_info(class, 0); 496 497 xe_gt_assert(gt, info); 498 499 return info ? info->irq_offset : 0; 500 } 501 502 static void hw_engine_init_early(struct xe_gt *gt, struct xe_hw_engine *hwe, 503 enum xe_hw_engine_id id) 504 { 505 const struct engine_info *info; 506 507 if (WARN_ON(id >= ARRAY_SIZE(engine_infos) || !engine_infos[id].name)) 508 return; 509 510 if (!(gt->info.engine_mask & BIT(id))) 511 return; 512 513 info = &engine_infos[id]; 514 515 xe_gt_assert(gt, !hwe->gt); 516 517 hwe->gt = gt; 518 hwe->class = info->class; 519 hwe->instance = info->instance; 520 hwe->mmio_base = info->mmio_base; 521 hwe->irq_offset = xe_device_has_msix(gt_to_xe(gt)) ? 522 get_msix_irq_offset(gt, info->class) : 523 info->irq_offset; 524 hwe->domain = info->domain; 525 hwe->name = info->name; 526 hwe->fence_irq = >->fence_irq[info->class]; 527 hwe->engine_id = id; 528 529 hwe->eclass = >->eclass[hwe->class]; 530 if (!hwe->eclass->sched_props.job_timeout_ms) { 531 hwe->eclass->sched_props.job_timeout_ms = 5 * 1000; 532 hwe->eclass->sched_props.job_timeout_min = XE_HW_ENGINE_JOB_TIMEOUT_MIN; 533 hwe->eclass->sched_props.job_timeout_max = XE_HW_ENGINE_JOB_TIMEOUT_MAX; 534 hwe->eclass->sched_props.timeslice_us = 1 * 1000; 535 hwe->eclass->sched_props.timeslice_min = XE_HW_ENGINE_TIMESLICE_MIN; 536 hwe->eclass->sched_props.timeslice_max = XE_HW_ENGINE_TIMESLICE_MAX; 537 hwe->eclass->sched_props.preempt_timeout_us = XE_HW_ENGINE_PREEMPT_TIMEOUT; 538 hwe->eclass->sched_props.preempt_timeout_min = XE_HW_ENGINE_PREEMPT_TIMEOUT_MIN; 539 hwe->eclass->sched_props.preempt_timeout_max = XE_HW_ENGINE_PREEMPT_TIMEOUT_MAX; 540 541 /* 542 * The GSC engine can accept submissions while the GSC shim is 543 * being reset, during which time the submission is stalled. In 544 * the worst case, the shim reset can take up to the maximum GSC 545 * command execution time (250ms), so the request start can be 546 * delayed by that much; the request itself can take that long 547 * without being preemptible, which means worst case it can 548 * theoretically take up to 500ms for a preemption to go through 549 * on the GSC engine. Adding to that an extra 100ms as a safety 550 * margin, we get a minimum recommended timeout of 600ms. 551 * The preempt_timeout value can't be tuned for OTHER_CLASS 552 * because the class is reserved for kernel usage, so we just 553 * need to make sure that the starting value is above that 554 * threshold; since our default value (640ms) is greater than 555 * 600ms, the only way we can go below is via a kconfig setting. 556 * If that happens, log it in dmesg and update the value. 557 */ 558 if (hwe->class == XE_ENGINE_CLASS_OTHER) { 559 const u32 min_preempt_timeout = 600 * 1000; 560 if (hwe->eclass->sched_props.preempt_timeout_us < min_preempt_timeout) { 561 hwe->eclass->sched_props.preempt_timeout_us = min_preempt_timeout; 562 xe_gt_notice(gt, "Increasing preempt_timeout for GSC to 600ms\n"); 563 } 564 } 565 566 /* Record default props */ 567 hwe->eclass->defaults = hwe->eclass->sched_props; 568 } 569 570 xe_reg_sr_init(&hwe->reg_sr, hwe->name, gt_to_xe(gt)); 571 xe_tuning_process_engine(hwe); 572 xe_wa_process_engine(hwe); 573 hw_engine_setup_default_state(hwe); 574 575 xe_reg_sr_init(&hwe->reg_whitelist, hwe->name, gt_to_xe(gt)); 576 xe_reg_whitelist_process_engine(hwe); 577 } 578 579 static void adjust_idledly(struct xe_hw_engine *hwe) 580 { 581 struct xe_gt *gt = hwe->gt; 582 u32 idledly, maxcnt; 583 u32 idledly_units_ps = 8 * gt->info.timestamp_base; 584 u32 maxcnt_units_ns = 640; 585 bool inhibit_switch = 0; 586 587 if (!IS_SRIOV_VF(gt_to_xe(hwe->gt)) && XE_GT_WA(gt, 16023105232)) { 588 idledly = xe_mmio_read32(>->mmio, RING_IDLEDLY(hwe->mmio_base)); 589 maxcnt = xe_mmio_read32(>->mmio, RING_PWRCTX_MAXCNT(hwe->mmio_base)); 590 591 inhibit_switch = idledly & INHIBIT_SWITCH_UNTIL_PREEMPTED; 592 idledly = REG_FIELD_GET(IDLE_DELAY, idledly); 593 idledly = DIV_ROUND_CLOSEST(idledly * idledly_units_ps, 1000); 594 maxcnt = REG_FIELD_GET(IDLE_WAIT_TIME, maxcnt); 595 maxcnt *= maxcnt_units_ns; 596 597 if (xe_gt_WARN_ON(gt, idledly >= maxcnt || inhibit_switch)) { 598 idledly = DIV_ROUND_CLOSEST(((maxcnt - 1) * maxcnt_units_ns), 599 idledly_units_ps); 600 idledly = DIV_ROUND_CLOSEST(idledly, 1000); 601 xe_mmio_write32(>->mmio, RING_IDLEDLY(hwe->mmio_base), idledly); 602 } 603 } 604 } 605 606 static int hw_engine_init(struct xe_gt *gt, struct xe_hw_engine *hwe, 607 enum xe_hw_engine_id id) 608 { 609 struct xe_device *xe = gt_to_xe(gt); 610 struct xe_tile *tile = gt_to_tile(gt); 611 int err; 612 613 xe_gt_assert(gt, id < ARRAY_SIZE(engine_infos) && engine_infos[id].name); 614 xe_gt_assert(gt, gt->info.engine_mask & BIT(id)); 615 616 xe_reg_sr_apply_mmio(&hwe->reg_sr, gt); 617 618 hwe->hwsp = xe_managed_bo_create_pin_map(xe, tile, SZ_4K, 619 XE_BO_FLAG_VRAM_IF_DGFX(tile) | 620 XE_BO_FLAG_GGTT | 621 XE_BO_FLAG_GGTT_INVALIDATE); 622 if (IS_ERR(hwe->hwsp)) { 623 err = PTR_ERR(hwe->hwsp); 624 goto err_name; 625 } 626 627 if (!xe_device_uc_enabled(xe)) { 628 hwe->exl_port = xe_execlist_port_create(xe, hwe); 629 if (IS_ERR(hwe->exl_port)) { 630 err = PTR_ERR(hwe->exl_port); 631 goto err_hwsp; 632 } 633 } else { 634 /* GSCCS has a special interrupt for reset */ 635 if (hwe->class == XE_ENGINE_CLASS_OTHER) 636 hwe->irq_handler = xe_gsc_hwe_irq_handler; 637 638 if (!IS_SRIOV_VF(xe)) 639 xe_hw_engine_enable_ring(hwe); 640 } 641 642 /* We reserve the highest BCS instance for USM */ 643 if (xe->info.has_usm && hwe->class == XE_ENGINE_CLASS_COPY) 644 gt->usm.reserved_bcs_instance = hwe->instance; 645 646 /* Ensure IDLEDLY is lower than MAXCNT */ 647 adjust_idledly(hwe); 648 649 return devm_add_action_or_reset(xe->drm.dev, hw_engine_fini, hwe); 650 651 err_hwsp: 652 xe_bo_unpin_map_no_vm(hwe->hwsp); 653 err_name: 654 hwe->name = NULL; 655 656 return err; 657 } 658 659 static void hw_engine_setup_logical_mapping(struct xe_gt *gt) 660 { 661 int class; 662 663 /* FIXME: Doing a simple logical mapping that works for most hardware */ 664 for (class = 0; class < XE_ENGINE_CLASS_MAX; ++class) { 665 struct xe_hw_engine *hwe; 666 enum xe_hw_engine_id id; 667 int logical_instance = 0; 668 669 for_each_hw_engine(hwe, gt, id) 670 if (hwe->class == class) 671 hwe->logical_instance = logical_instance++; 672 } 673 } 674 675 static void read_media_fuses(struct xe_gt *gt) 676 { 677 struct xe_device *xe = gt_to_xe(gt); 678 u32 media_fuse; 679 u16 vdbox_mask; 680 u16 vebox_mask; 681 int i, j; 682 683 xe_force_wake_assert_held(gt_to_fw(gt), XE_FW_GT); 684 685 media_fuse = xe_mmio_read32(>->mmio, GT_VEBOX_VDBOX_DISABLE); 686 687 /* 688 * Pre-Xe_HP platforms had register bits representing absent engines, 689 * whereas Xe_HP and beyond have bits representing present engines. 690 * Invert the polarity on old platforms so that we can use common 691 * handling below. 692 */ 693 if (GRAPHICS_VERx100(xe) < 1250) 694 media_fuse = ~media_fuse; 695 696 vdbox_mask = REG_FIELD_GET(GT_VDBOX_DISABLE_MASK, media_fuse); 697 vebox_mask = REG_FIELD_GET(GT_VEBOX_DISABLE_MASK, media_fuse); 698 699 for (i = XE_HW_ENGINE_VCS0, j = 0; i <= XE_HW_ENGINE_VCS7; ++i, ++j) { 700 if (!(gt->info.engine_mask & BIT(i))) 701 continue; 702 703 if (!(BIT(j) & vdbox_mask)) { 704 gt->info.engine_mask &= ~BIT(i); 705 xe_gt_info(gt, "vcs%u fused off\n", j); 706 } 707 } 708 709 for (i = XE_HW_ENGINE_VECS0, j = 0; i <= XE_HW_ENGINE_VECS3; ++i, ++j) { 710 if (!(gt->info.engine_mask & BIT(i))) 711 continue; 712 713 if (!(BIT(j) & vebox_mask)) { 714 gt->info.engine_mask &= ~BIT(i); 715 xe_gt_info(gt, "vecs%u fused off\n", j); 716 } 717 } 718 } 719 720 static u32 infer_svccopy_from_meml3(struct xe_gt *gt) 721 { 722 u32 meml3 = REG_FIELD_GET(MEML3_EN_MASK, 723 xe_mmio_read32(>->mmio, MIRROR_FUSE3)); 724 u32 svccopy_mask = 0; 725 726 /* 727 * Each of the four meml3 bits determines the fusing of two service 728 * copy engines. 729 */ 730 for (int i = 0; i < 4; i++) 731 svccopy_mask |= (meml3 & BIT(i)) ? 0b11 << 2 * i : 0; 732 733 return svccopy_mask; 734 } 735 736 static u32 read_svccopy_fuses(struct xe_gt *gt) 737 { 738 return REG_FIELD_GET(FUSE_SERVICE_COPY_ENABLE_MASK, 739 xe_mmio_read32(>->mmio, SERVICE_COPY_ENABLE)); 740 } 741 742 static void read_copy_fuses(struct xe_gt *gt) 743 { 744 struct xe_device *xe = gt_to_xe(gt); 745 u32 bcs_mask; 746 747 xe_force_wake_assert_held(gt_to_fw(gt), XE_FW_GT); 748 749 if (GRAPHICS_VER(xe) >= 35) 750 bcs_mask = read_svccopy_fuses(gt); 751 else if (GRAPHICS_VERx100(xe) == 1260) 752 bcs_mask = infer_svccopy_from_meml3(gt); 753 else 754 return; 755 756 /* Only BCS1-BCS8 may be fused off */ 757 bcs_mask <<= XE_HW_ENGINE_BCS1; 758 for (int i = XE_HW_ENGINE_BCS1; i <= XE_HW_ENGINE_BCS8; ++i) { 759 if (!(gt->info.engine_mask & BIT(i))) 760 continue; 761 762 if (!(bcs_mask & BIT(i))) { 763 gt->info.engine_mask &= ~BIT(i); 764 xe_gt_info(gt, "bcs%u fused off\n", 765 i - XE_HW_ENGINE_BCS0); 766 } 767 } 768 } 769 770 static void read_compute_fuses_from_dss(struct xe_gt *gt) 771 { 772 /* 773 * CCS fusing based on DSS masks only applies to platforms that can 774 * have more than one CCS. 775 */ 776 if (hweight64(gt->info.engine_mask & 777 GENMASK_ULL(XE_HW_ENGINE_CCS3, XE_HW_ENGINE_CCS0)) <= 1) 778 return; 779 780 /* 781 * CCS availability on Xe_HP is inferred from the presence of DSS in 782 * each quadrant. 783 */ 784 for (int i = XE_HW_ENGINE_CCS0, j = 0; i <= XE_HW_ENGINE_CCS3; ++i, ++j) { 785 if (!(gt->info.engine_mask & BIT(i))) 786 continue; 787 788 if (!xe_gt_topology_has_dss_in_quadrant(gt, j)) { 789 gt->info.engine_mask &= ~BIT(i); 790 xe_gt_info(gt, "ccs%u fused off\n", j); 791 } 792 } 793 } 794 795 static void read_compute_fuses_from_reg(struct xe_gt *gt) 796 { 797 u32 ccs_mask; 798 799 ccs_mask = xe_mmio_read32(>->mmio, XEHP_FUSE4); 800 ccs_mask = REG_FIELD_GET(CCS_EN_MASK, ccs_mask); 801 802 for (int i = XE_HW_ENGINE_CCS0, j = 0; i <= XE_HW_ENGINE_CCS3; ++i, ++j) { 803 if (!(gt->info.engine_mask & BIT(i))) 804 continue; 805 806 if ((ccs_mask & BIT(j)) == 0) { 807 gt->info.engine_mask &= ~BIT(i); 808 xe_gt_info(gt, "ccs%u fused off\n", j); 809 } 810 } 811 } 812 813 static void read_compute_fuses(struct xe_gt *gt) 814 { 815 if (GRAPHICS_VER(gt_to_xe(gt)) >= 20) 816 read_compute_fuses_from_reg(gt); 817 else 818 read_compute_fuses_from_dss(gt); 819 } 820 821 static void check_gsc_availability(struct xe_gt *gt) 822 { 823 if (!(gt->info.engine_mask & BIT(XE_HW_ENGINE_GSCCS0))) 824 return; 825 826 /* 827 * The GSCCS is only used to communicate with the GSC FW, so if we don't 828 * have the FW there is nothing we need the engine for and can therefore 829 * skip its initialization. 830 */ 831 if (!xe_uc_fw_is_available(>->uc.gsc.fw)) { 832 gt->info.engine_mask &= ~BIT(XE_HW_ENGINE_GSCCS0); 833 834 /* interrupts where previously enabled, so turn them off */ 835 xe_mmio_write32(>->mmio, GUNIT_GSC_INTR_ENABLE, 0); 836 xe_mmio_write32(>->mmio, GUNIT_GSC_INTR_MASK, ~0); 837 838 xe_gt_dbg(gt, "GSC FW not used, disabling gsccs\n"); 839 } 840 } 841 842 static void check_sw_disable(struct xe_gt *gt) 843 { 844 struct xe_device *xe = gt_to_xe(gt); 845 u64 sw_allowed = xe_configfs_get_engines_allowed(to_pci_dev(xe->drm.dev)); 846 enum xe_hw_engine_id id; 847 848 for (id = 0; id < XE_NUM_HW_ENGINES; ++id) { 849 if (!(gt->info.engine_mask & BIT(id))) 850 continue; 851 852 if (!(sw_allowed & BIT(id))) { 853 gt->info.engine_mask &= ~BIT(id); 854 xe_gt_info(gt, "%s disabled via configfs\n", 855 engine_infos[id].name); 856 } 857 } 858 } 859 860 int xe_hw_engines_init_early(struct xe_gt *gt) 861 { 862 int i; 863 864 read_media_fuses(gt); 865 read_copy_fuses(gt); 866 read_compute_fuses(gt); 867 check_gsc_availability(gt); 868 check_sw_disable(gt); 869 870 BUILD_BUG_ON(XE_HW_ENGINE_PREEMPT_TIMEOUT < XE_HW_ENGINE_PREEMPT_TIMEOUT_MIN); 871 BUILD_BUG_ON(XE_HW_ENGINE_PREEMPT_TIMEOUT > XE_HW_ENGINE_PREEMPT_TIMEOUT_MAX); 872 873 for (i = 0; i < ARRAY_SIZE(gt->hw_engines); i++) 874 hw_engine_init_early(gt, >->hw_engines[i], i); 875 876 return 0; 877 } 878 879 int xe_hw_engines_init(struct xe_gt *gt) 880 { 881 int err; 882 struct xe_hw_engine *hwe; 883 enum xe_hw_engine_id id; 884 885 for_each_hw_engine(hwe, gt, id) { 886 err = hw_engine_init(gt, hwe, id); 887 if (err) 888 return err; 889 } 890 891 hw_engine_setup_logical_mapping(gt); 892 err = xe_hw_engine_setup_groups(gt); 893 if (err) 894 return err; 895 896 return 0; 897 } 898 899 void xe_hw_engine_handle_irq(struct xe_hw_engine *hwe, u16 intr_vec) 900 { 901 wake_up_all(>_to_xe(hwe->gt)->ufence_wq); 902 903 if (hwe->irq_handler) 904 hwe->irq_handler(hwe, intr_vec); 905 906 if (intr_vec & GT_MI_USER_INTERRUPT) 907 xe_hw_fence_irq_run(hwe->fence_irq); 908 } 909 910 /** 911 * xe_hw_engine_snapshot_capture - Take a quick snapshot of the HW Engine. 912 * @hwe: Xe HW Engine. 913 * @q: The exec queue object. 914 * 915 * This can be printed out in a later stage like during dev_coredump 916 * analysis. 917 * 918 * Returns: a Xe HW Engine snapshot object that must be freed by the 919 * caller, using `xe_hw_engine_snapshot_free`. 920 */ 921 struct xe_hw_engine_snapshot * 922 xe_hw_engine_snapshot_capture(struct xe_hw_engine *hwe, struct xe_exec_queue *q) 923 { 924 struct xe_hw_engine_snapshot *snapshot; 925 struct __guc_capture_parsed_output *node; 926 927 if (!xe_hw_engine_is_valid(hwe)) 928 return NULL; 929 930 snapshot = kzalloc(sizeof(*snapshot), GFP_ATOMIC); 931 932 if (!snapshot) 933 return NULL; 934 935 snapshot->name = kstrdup(hwe->name, GFP_ATOMIC); 936 snapshot->hwe = hwe; 937 snapshot->logical_instance = hwe->logical_instance; 938 snapshot->forcewake.domain = hwe->domain; 939 snapshot->forcewake.ref = xe_force_wake_ref(gt_to_fw(hwe->gt), 940 hwe->domain); 941 snapshot->mmio_base = hwe->mmio_base; 942 snapshot->kernel_reserved = xe_hw_engine_is_reserved(hwe); 943 944 /* no more VF accessible data below this point */ 945 if (IS_SRIOV_VF(gt_to_xe(hwe->gt))) 946 return snapshot; 947 948 if (q) { 949 /* If got guc capture, set source to GuC */ 950 node = xe_guc_capture_get_matching_and_lock(q); 951 if (node) { 952 struct xe_device *xe = gt_to_xe(hwe->gt); 953 struct xe_devcoredump *coredump = &xe->devcoredump; 954 955 coredump->snapshot.matched_node = node; 956 xe_gt_dbg(hwe->gt, "Found and locked GuC-err-capture node"); 957 return snapshot; 958 } 959 } 960 961 /* otherwise, do manual capture */ 962 xe_engine_manual_capture(hwe, snapshot); 963 xe_gt_dbg(hwe->gt, "Proceeding with manual engine snapshot"); 964 965 return snapshot; 966 } 967 968 /** 969 * xe_hw_engine_snapshot_free - Free all allocated objects for a given snapshot. 970 * @snapshot: Xe HW Engine snapshot object. 971 * 972 * This function free all the memory that needed to be allocated at capture 973 * time. 974 */ 975 void xe_hw_engine_snapshot_free(struct xe_hw_engine_snapshot *snapshot) 976 { 977 struct xe_gt *gt; 978 if (!snapshot) 979 return; 980 981 gt = snapshot->hwe->gt; 982 /* 983 * xe_guc_capture_put_matched_nodes is called here and from 984 * xe_devcoredump_snapshot_free, to cover the 2 calling paths 985 * of hw_engines - debugfs and devcoredump free. 986 */ 987 xe_guc_capture_put_matched_nodes(>->uc.guc); 988 989 kfree(snapshot->name); 990 kfree(snapshot); 991 } 992 993 /** 994 * xe_hw_engine_print - Xe HW Engine Print. 995 * @hwe: Hardware Engine. 996 * @p: drm_printer. 997 * 998 * This function quickly capture a snapshot and immediately print it out. 999 */ 1000 void xe_hw_engine_print(struct xe_hw_engine *hwe, struct drm_printer *p) 1001 { 1002 struct xe_hw_engine_snapshot *snapshot; 1003 1004 snapshot = xe_hw_engine_snapshot_capture(hwe, NULL); 1005 xe_engine_snapshot_print(snapshot, p); 1006 xe_hw_engine_snapshot_free(snapshot); 1007 } 1008 1009 u32 xe_hw_engine_mask_per_class(struct xe_gt *gt, 1010 enum xe_engine_class engine_class) 1011 { 1012 u32 mask = 0; 1013 enum xe_hw_engine_id id; 1014 1015 for (id = 0; id < XE_NUM_HW_ENGINES; ++id) { 1016 if (engine_infos[id].class == engine_class && 1017 gt->info.engine_mask & BIT(id)) 1018 mask |= BIT(engine_infos[id].instance); 1019 } 1020 return mask; 1021 } 1022 1023 bool xe_hw_engine_is_reserved(struct xe_hw_engine *hwe) 1024 { 1025 struct xe_gt *gt = hwe->gt; 1026 struct xe_device *xe = gt_to_xe(gt); 1027 1028 if (hwe->class == XE_ENGINE_CLASS_OTHER) 1029 return true; 1030 1031 /* Check for engines disabled by ccs_mode setting */ 1032 if (xe_gt_ccs_mode_enabled(gt) && 1033 hwe->class == XE_ENGINE_CLASS_COMPUTE && 1034 hwe->logical_instance >= gt->ccs_mode) 1035 return true; 1036 1037 return xe->info.has_usm && hwe->class == XE_ENGINE_CLASS_COPY && 1038 hwe->instance == gt->usm.reserved_bcs_instance; 1039 } 1040 1041 const char *xe_hw_engine_class_to_str(enum xe_engine_class class) 1042 { 1043 switch (class) { 1044 case XE_ENGINE_CLASS_RENDER: 1045 return "rcs"; 1046 case XE_ENGINE_CLASS_VIDEO_DECODE: 1047 return "vcs"; 1048 case XE_ENGINE_CLASS_VIDEO_ENHANCE: 1049 return "vecs"; 1050 case XE_ENGINE_CLASS_COPY: 1051 return "bcs"; 1052 case XE_ENGINE_CLASS_OTHER: 1053 return "other"; 1054 case XE_ENGINE_CLASS_COMPUTE: 1055 return "ccs"; 1056 case XE_ENGINE_CLASS_MAX: 1057 break; 1058 } 1059 1060 return NULL; 1061 } 1062 1063 u64 xe_hw_engine_read_timestamp(struct xe_hw_engine *hwe) 1064 { 1065 return xe_mmio_read64_2x32(&hwe->gt->mmio, RING_TIMESTAMP(hwe->mmio_base)); 1066 } 1067 1068 enum xe_force_wake_domains xe_hw_engine_to_fw_domain(struct xe_hw_engine *hwe) 1069 { 1070 return engine_infos[hwe->engine_id].domain; 1071 } 1072 1073 static const enum xe_engine_class user_to_xe_engine_class[] = { 1074 [DRM_XE_ENGINE_CLASS_RENDER] = XE_ENGINE_CLASS_RENDER, 1075 [DRM_XE_ENGINE_CLASS_COPY] = XE_ENGINE_CLASS_COPY, 1076 [DRM_XE_ENGINE_CLASS_VIDEO_DECODE] = XE_ENGINE_CLASS_VIDEO_DECODE, 1077 [DRM_XE_ENGINE_CLASS_VIDEO_ENHANCE] = XE_ENGINE_CLASS_VIDEO_ENHANCE, 1078 [DRM_XE_ENGINE_CLASS_COMPUTE] = XE_ENGINE_CLASS_COMPUTE, 1079 }; 1080 1081 /** 1082 * xe_hw_engine_lookup() - Lookup hardware engine for class:instance 1083 * @xe: xe device 1084 * @eci: engine class and instance 1085 * 1086 * This function will find a hardware engine for given engine 1087 * class and instance. 1088 * 1089 * Return: If found xe_hw_engine pointer, NULL otherwise. 1090 */ 1091 struct xe_hw_engine * 1092 xe_hw_engine_lookup(struct xe_device *xe, 1093 struct drm_xe_engine_class_instance eci) 1094 { 1095 struct xe_gt *gt = xe_device_get_gt(xe, eci.gt_id); 1096 unsigned int idx; 1097 1098 if (eci.engine_class >= ARRAY_SIZE(user_to_xe_engine_class)) 1099 return NULL; 1100 1101 if (!gt) 1102 return NULL; 1103 1104 idx = array_index_nospec(eci.engine_class, 1105 ARRAY_SIZE(user_to_xe_engine_class)); 1106 1107 return xe_gt_hw_engine(xe_device_get_gt(xe, eci.gt_id), 1108 user_to_xe_engine_class[idx], 1109 eci.engine_instance, true); 1110 } 1111