1 // SPDX-License-Identifier: MIT 2 /* 3 * Copyright © 2022 Intel Corporation 4 */ 5 6 #include "xe_rtp.h" 7 8 #include <kunit/visibility.h> 9 10 #include <uapi/drm/xe_drm.h> 11 12 #include "xe_configfs.h" 13 #include "xe_device.h" 14 #include "xe_gt.h" 15 #include "xe_gt_topology.h" 16 #include "xe_reg_sr.h" 17 #include "xe_sriov.h" 18 19 /** 20 * DOC: Register Table Processing 21 * 22 * Internal infrastructure to define how registers should be updated based on 23 * rules and actions. This can be used to define tables with multiple entries 24 * (one per register) that will be walked over at some point in time to apply 25 * the values to the registers that have matching rules. 26 */ 27 28 static bool has_samedia(const struct xe_device *xe) 29 { 30 return xe->info.media_verx100 >= 1300; 31 } 32 33 struct rule_match_ctx { 34 const struct xe_device *xe; 35 struct xe_gt *gt; 36 struct xe_hw_engine *hwe; 37 const struct xe_rtp_rule *rules; 38 const unsigned int n_rules; 39 unsigned int head; 40 int err; 41 }; 42 43 static bool rule_is_item(const struct xe_rtp_rule *r) 44 { 45 return r->match_type != XE_RTP_MATCH_OR; 46 } 47 48 static bool rule_match_item(struct rule_match_ctx *match_ctx) 49 { 50 const struct xe_device *xe = match_ctx->xe; 51 struct xe_gt *gt = match_ctx->gt; 52 struct xe_hw_engine *hwe = match_ctx->hwe; 53 const struct xe_rtp_rule *r = &match_ctx->rules[match_ctx->head]; 54 55 switch (r->match_type) { 56 case XE_RTP_MATCH_PLATFORM: 57 return xe->info.platform == r->platform; 58 case XE_RTP_MATCH_SUBPLATFORM: 59 return xe->info.platform == r->platform && 60 xe->info.subplatform == r->subplatform; 61 case XE_RTP_MATCH_PLATFORM_STEP: 62 if (drm_WARN_ON(&xe->drm, xe->info.step.platform == STEP_NONE)) 63 return false; 64 65 return xe->info.step.platform >= r->step_start && 66 xe->info.step.platform < r->step_end; 67 case XE_RTP_MATCH_GRAPHICS_VERSION: 68 if (drm_WARN_ON(&xe->drm, !gt)) 69 return false; 70 71 return xe->info.graphics_verx100 == r->ver_start && 72 (!has_samedia(xe) || !xe_gt_is_media_type(gt)); 73 case XE_RTP_MATCH_GRAPHICS_VERSION_RANGE: 74 if (drm_WARN_ON(&xe->drm, !gt)) 75 return false; 76 77 return xe->info.graphics_verx100 >= r->ver_start && 78 xe->info.graphics_verx100 <= r->ver_end && 79 (!has_samedia(xe) || !xe_gt_is_media_type(gt)); 80 case XE_RTP_MATCH_GRAPHICS_VERSION_ANY_GT: 81 if (drm_WARN_ON(&xe->drm, !gt)) 82 return false; 83 84 return xe->info.graphics_verx100 == r->ver_start; 85 case XE_RTP_MATCH_GRAPHICS_STEP: 86 if (drm_WARN_ON(&xe->drm, !gt)) 87 return false; 88 89 return xe->info.step.graphics >= r->step_start && 90 xe->info.step.graphics < r->step_end && 91 (!has_samedia(xe) || !xe_gt_is_media_type(gt)); 92 case XE_RTP_MATCH_MEDIA_VERSION: 93 if (drm_WARN_ON(&xe->drm, !gt)) 94 return false; 95 96 return xe->info.media_verx100 == r->ver_start && 97 (!has_samedia(xe) || xe_gt_is_media_type(gt)); 98 case XE_RTP_MATCH_MEDIA_VERSION_RANGE: 99 if (drm_WARN_ON(&xe->drm, !gt)) 100 return false; 101 102 return xe->info.media_verx100 >= r->ver_start && 103 xe->info.media_verx100 <= r->ver_end && 104 (!has_samedia(xe) || xe_gt_is_media_type(gt)); 105 case XE_RTP_MATCH_MEDIA_STEP: 106 if (drm_WARN_ON(&xe->drm, !gt)) 107 return false; 108 109 return xe->info.step.media >= r->step_start && 110 xe->info.step.media < r->step_end && 111 (!has_samedia(xe) || xe_gt_is_media_type(gt)); 112 case XE_RTP_MATCH_MEDIA_VERSION_ANY_GT: 113 if (drm_WARN_ON(&xe->drm, !gt)) 114 return false; 115 116 return xe->info.media_verx100 == r->ver_start; 117 case XE_RTP_MATCH_INTEGRATED: 118 return !xe->info.is_dgfx; 119 case XE_RTP_MATCH_DISCRETE: 120 return xe->info.is_dgfx; 121 case XE_RTP_MATCH_ENGINE_CLASS: 122 if (drm_WARN_ON(&xe->drm, !hwe)) 123 return false; 124 125 return hwe->class == r->engine_class; 126 case XE_RTP_MATCH_NOT_ENGINE_CLASS: 127 if (drm_WARN_ON(&xe->drm, !hwe)) 128 return false; 129 130 return hwe->class != r->engine_class; 131 case XE_RTP_MATCH_FUNC: 132 return r->match_func(xe, gt, hwe); 133 default: 134 drm_warn(&xe->drm, "Invalid RTP match %u\n", 135 r->match_type); 136 return false; 137 } 138 } 139 140 /* 141 * Match a conjunctive set of rules (rules joined by an implicit "AND"). 142 * 143 * Once one item evaluates to false, the remaining items are not evaluated 144 * anymore. Nevetheless, all rules are consumed to allow detecting syntax 145 * errors. 146 */ 147 static bool rule_match_and(struct rule_match_ctx *match_ctx, bool parse_only) 148 { 149 bool match = true; 150 unsigned int count = 0; 151 152 while (match_ctx->head < match_ctx->n_rules && 153 rule_is_item(&match_ctx->rules[match_ctx->head])) { 154 if (!parse_only) 155 match = rule_match_item(match_ctx); 156 157 if (!match) 158 parse_only = true; 159 160 match_ctx->head++; 161 count++; 162 } 163 164 if (drm_WARN_ON(&match_ctx->xe->drm, !count)) { 165 match_ctx->err = -EINVAL; 166 167 if (!parse_only) 168 match = false; 169 } 170 171 return match; 172 } 173 174 /* 175 * Match a disjunctive set of rules (subset of rules joined by 176 * "XE_RTP_MATCH_OR"). 177 * 178 * Once one subset evaluates to true, the remaining items are not evaluated 179 * anymore. Nevetheless, all rules are consumed to allow detecting syntax 180 * errors. 181 */ 182 static bool rule_match_or(struct rule_match_ctx *match_ctx) 183 { 184 bool match = rule_match_and(match_ctx, false); 185 186 while (match_ctx->head < match_ctx->n_rules && 187 match_ctx->rules[match_ctx->head].match_type == XE_RTP_MATCH_OR) { 188 /* Consume XE_RTP_MATCH_OR. */ 189 match_ctx->head++; 190 191 match = rule_match_and(match_ctx, match); 192 } 193 194 return match; 195 } 196 197 static bool rule_matches_with_err(const struct xe_device *xe, 198 struct xe_gt *gt, 199 struct xe_hw_engine *hwe, 200 const struct xe_rtp_rule *rules, 201 unsigned int n_rules, 202 int *err) 203 { 204 struct rule_match_ctx match_ctx = { 205 .xe = xe, 206 .gt = gt, 207 .hwe = hwe, 208 .rules = rules, 209 .n_rules = n_rules, 210 }; 211 bool match = rule_match_or(&match_ctx); 212 213 if (err) 214 *err = match_ctx.err; 215 216 return match; 217 } 218 219 static bool rule_matches(const struct xe_device *xe, 220 struct xe_gt *gt, 221 struct xe_hw_engine *hwe, 222 const struct xe_rtp_rule *rules, 223 unsigned int n_rules) 224 { 225 return rule_matches_with_err(xe, gt, hwe, rules, n_rules, NULL); 226 } 227 228 static void rtp_add_sr_entry(const struct xe_rtp_action *action, 229 struct xe_gt *gt, 230 u32 mmio_base, 231 struct xe_reg_sr *sr) 232 { 233 struct xe_reg_sr_entry sr_entry = { 234 .reg = action->reg, 235 .clr_bits = action->clr_bits, 236 .set_bits = action->set_bits, 237 .read_mask = action->read_mask, 238 }; 239 240 sr_entry.reg.addr += mmio_base; 241 xe_reg_sr_add(sr, &sr_entry, gt); 242 } 243 244 static bool rtp_process_one_sr(const struct xe_rtp_entry_sr *entry, 245 struct xe_device *xe, struct xe_gt *gt, 246 struct xe_hw_engine *hwe, struct xe_reg_sr *sr) 247 { 248 const struct xe_rtp_action *action; 249 u32 mmio_base; 250 unsigned int i; 251 252 if (!rule_matches(xe, gt, hwe, entry->rules, entry->n_rules)) 253 return false; 254 255 for (i = 0, action = &entry->actions[0]; i < entry->n_actions; action++, i++) { 256 if ((entry->flags & XE_RTP_ENTRY_FLAG_FOREACH_ENGINE) || 257 (action->flags & XE_RTP_ACTION_FLAG_ENGINE_BASE)) 258 mmio_base = hwe->mmio_base; 259 else 260 mmio_base = 0; 261 262 rtp_add_sr_entry(action, gt, mmio_base, sr); 263 } 264 265 return true; 266 } 267 268 static void rtp_get_context(struct xe_rtp_process_ctx *ctx, 269 struct xe_hw_engine **hwe, 270 struct xe_gt **gt, 271 struct xe_device **xe) 272 { 273 switch (ctx->type) { 274 case XE_RTP_PROCESS_TYPE_DEVICE: 275 *hwe = NULL; 276 *gt = NULL; 277 *xe = ctx->xe; 278 break; 279 case XE_RTP_PROCESS_TYPE_GT: 280 *hwe = NULL; 281 *gt = ctx->gt; 282 *xe = gt_to_xe(*gt); 283 break; 284 case XE_RTP_PROCESS_TYPE_ENGINE: 285 *hwe = ctx->hwe; 286 *gt = (*hwe)->gt; 287 *xe = gt_to_xe(*gt); 288 break; 289 } 290 } 291 292 /** 293 * xe_rtp_process_ctx_enable_active_tracking - Enable tracking of active entries 294 * 295 * Set additional metadata to track what entries are considered "active", i.e. 296 * their rules match the condition. Bits are never cleared: entries with 297 * matching rules set the corresponding bit in the bitmap. 298 * 299 * @ctx: The context for processing the table 300 * @active_entries: bitmap to store the active entries 301 * @n_entries: number of entries to be processed 302 */ 303 void xe_rtp_process_ctx_enable_active_tracking(struct xe_rtp_process_ctx *ctx, 304 unsigned long *active_entries, 305 size_t n_entries) 306 { 307 ctx->active_entries = active_entries; 308 ctx->n_entries = n_entries; 309 } 310 EXPORT_SYMBOL_IF_KUNIT(xe_rtp_process_ctx_enable_active_tracking); 311 312 static void rtp_mark_active(struct xe_device *xe, 313 struct xe_rtp_process_ctx *ctx, 314 unsigned int idx) 315 { 316 if (!ctx->active_entries) 317 return; 318 319 if (drm_WARN_ON(&xe->drm, idx >= ctx->n_entries)) 320 return; 321 322 bitmap_set(ctx->active_entries, idx, 1); 323 } 324 325 /** 326 * xe_rtp_process_to_sr - Process all rtp @entries, adding the matching ones to 327 * the save-restore argument. 328 * @ctx: The context for processing the table, with one of device, gt or hwe 329 * @entries: Table with RTP definitions 330 * @n_entries: Number of entries to process, usually ARRAY_SIZE(entries) 331 * @sr: Save-restore struct where matching rules execute the action. This can be 332 * viewed as the "coalesced view" of multiple the tables. The bits for each 333 * register set are expected not to collide with previously added entries 334 * @process_in_vf: Whether this RTP table should get processed for SR-IOV VF 335 * devices. Should generally only be 'true' for LRC tables. 336 * 337 * Walk the table pointed by @entries (with an empty sentinel) and add all 338 * entries with matching rules to @sr. If @hwe is not NULL, its mmio_base is 339 * used to calculate the right register offset 340 */ 341 void xe_rtp_process_to_sr(struct xe_rtp_process_ctx *ctx, 342 const struct xe_rtp_entry_sr *entries, 343 size_t n_entries, 344 struct xe_reg_sr *sr, 345 bool process_in_vf) 346 { 347 const struct xe_rtp_entry_sr *entry; 348 struct xe_hw_engine *hwe = NULL; 349 struct xe_gt *gt = NULL; 350 struct xe_device *xe = NULL; 351 352 rtp_get_context(ctx, &hwe, >, &xe); 353 354 if (!process_in_vf && IS_SRIOV_VF(xe)) 355 return; 356 357 xe_assert(xe, entries); 358 359 for (entry = entries; entry - entries < n_entries; entry++) { 360 bool match = false; 361 362 if (entry->flags & XE_RTP_ENTRY_FLAG_FOREACH_ENGINE) { 363 struct xe_hw_engine *each_hwe; 364 enum xe_hw_engine_id id; 365 366 for_each_hw_engine(each_hwe, gt, id) 367 match |= rtp_process_one_sr(entry, xe, gt, 368 each_hwe, sr); 369 } else { 370 match = rtp_process_one_sr(entry, xe, gt, hwe, sr); 371 } 372 373 if (match) 374 rtp_mark_active(xe, ctx, entry - entries); 375 } 376 } 377 EXPORT_SYMBOL_IF_KUNIT(xe_rtp_process_to_sr); 378 379 /** 380 * xe_rtp_process - Process all rtp @entries, without running any action 381 * @ctx: The context for processing the table, with one of device, gt or hwe 382 * @entries: Table with RTP definitions 383 * 384 * Walk the table pointed by @entries (with an empty sentinel), executing the 385 * rules. One difference from xe_rtp_process_to_sr(): there is no action 386 * associated with each entry since this uses struct xe_rtp_entry. Its main use 387 * is for marking active workarounds via 388 * xe_rtp_process_ctx_enable_active_tracking(). 389 */ 390 void xe_rtp_process(struct xe_rtp_process_ctx *ctx, 391 const struct xe_rtp_entry *entries) 392 { 393 const struct xe_rtp_entry *entry; 394 struct xe_hw_engine *hwe; 395 struct xe_gt *gt; 396 struct xe_device *xe; 397 398 rtp_get_context(ctx, &hwe, >, &xe); 399 400 for (entry = entries; entry && entry->rules; entry++) { 401 if (!rule_matches(xe, gt, hwe, entry->rules, entry->n_rules)) 402 continue; 403 404 rtp_mark_active(xe, ctx, entry - entries); 405 } 406 } 407 EXPORT_SYMBOL_IF_KUNIT(xe_rtp_process); 408 409 bool xe_rtp_match_always(const struct xe_device *xe, 410 const struct xe_gt *gt, 411 const struct xe_hw_engine *hwe) 412 { 413 return true; 414 } 415 416 bool xe_rtp_match_even_instance(const struct xe_device *xe, 417 const struct xe_gt *gt, 418 const struct xe_hw_engine *hwe) 419 { 420 return hwe->instance % 2 == 0; 421 } 422 423 bool xe_rtp_match_first_render_or_compute(const struct xe_device *xe, 424 const struct xe_gt *gt, 425 const struct xe_hw_engine *hwe) 426 { 427 u64 render_compute_mask = gt->info.engine_mask & 428 (XE_HW_ENGINE_CCS_MASK | XE_HW_ENGINE_RCS_MASK); 429 430 return render_compute_mask && 431 hwe->engine_id == __ffs(render_compute_mask); 432 } 433 434 bool xe_rtp_match_not_sriov_vf(const struct xe_device *xe, 435 const struct xe_gt *gt, 436 const struct xe_hw_engine *hwe) 437 { 438 return !IS_SRIOV_VF(xe); 439 } 440 441 bool xe_rtp_match_psmi_enabled(const struct xe_device *xe, 442 const struct xe_gt *gt, 443 const struct xe_hw_engine *hwe) 444 { 445 return xe_configfs_get_psmi_enabled(to_pci_dev(xe->drm.dev)); 446 } 447 448 bool xe_rtp_match_gt_has_discontiguous_dss_groups(const struct xe_device *xe, 449 const struct xe_gt *gt, 450 const struct xe_hw_engine *hwe) 451 { 452 return xe_gt_has_discontiguous_dss_groups(gt); 453 } 454 455 bool xe_rtp_match_has_flat_ccs(const struct xe_device *xe, 456 const struct xe_gt *gt, 457 const struct xe_hw_engine *hwe) 458 { 459 return xe->info.has_flat_ccs; 460 } 461 462 bool xe_rtp_match_has_msix(const struct xe_device *xe, 463 const struct xe_gt *gt, 464 const struct xe_hw_engine *hwe) 465 { 466 return xe_device_has_msix(xe); 467 } 468 469 #if IS_ENABLED(CONFIG_DRM_XE_KUNIT_TEST) 470 #include "tests/xe_rtp.c" 471 #endif 472