| 1 | #include <acpi/lapic.h> |
| 2 | #include <boot/gdt.h> |
| 3 | #include <crypto/prng.h> |
| 4 | #include <irq/idt.h> |
| 5 | #include <limine.h> |
| 6 | #include <mem/alloc.h> |
| 7 | #include <mem/alloc_or_die.h> |
| 8 | #include <mem/domain.h> |
| 9 | #include <mem/tlb.h> |
| 10 | #include <sch/sched.h> |
| 11 | #include <smp/domain.h> |
| 12 | #include <smp/percpu.h> |
| 13 | #include <smp/smp.h> |
| 14 | #include <string.h> |
| 15 | #include <sync/spinlock.h> |
| 16 | #include <thread/dpc.h> |
| 17 | #include <thread/thread.h> |
| 18 | #include <time/time.h> |
| 19 | #include <time/tsc.h> |
| 20 | |
| 21 | /* TODO: This file is one that I had thought would stay small back when |
| 22 | * I first got SMP working, but day by day it grows and grows. Might |
| 23 | * want to refactor and clean this up into state machines */ |
| 24 | |
| 25 | static volatile uint64_t cr3 = 0; |
| 26 | static _Atomic uint32_t cores_awake = 0; |
| 27 | #define CPUID_LEAF_HYBRID 0x1A |
| 28 | |
| 29 | static void detect_cpu_features(struct cpu_capability *cap) { |
| 30 | uint32_t eax, ebx, ecx, edx; |
| 31 | |
| 32 | cap->feature_bits = 0; |
| 33 | |
| 34 | /* CPUID.1 */ |
| 35 | cpuid_count(leaf: 1, subleaf: 0, eax: &eax, ebx: &ebx, ecx: &ecx, edx: &edx); |
| 36 | |
| 37 | if (edx & (1 << 26)) |
| 38 | cap->feature_bits |= CPU_FEAT_SSE2; |
| 39 | if (ecx & (1 << 28)) |
| 40 | cap->feature_bits |= CPU_FEAT_AVX; |
| 41 | |
| 42 | /* CPUID.7.0 */ |
| 43 | cpuid_count(leaf: 7, subleaf: 0, eax: &eax, ebx: &ebx, ecx: &ecx, edx: &edx); |
| 44 | |
| 45 | if (ebx & (1 << 5)) |
| 46 | cap->feature_bits |= CPU_FEAT_AVX2; |
| 47 | if (ebx & (1 << 16)) |
| 48 | cap->feature_bits |= CPU_FEAT_AVX512F; |
| 49 | } |
| 50 | |
| 51 | static void detect_cpu_class(struct cpu_capability *cap) { |
| 52 | uint32_t eax, ebx, ecx, edx; |
| 53 | |
| 54 | cpuid_count(CPUID_LEAF_HYBRID, subleaf: 0, eax: &eax, ebx: &ebx, ecx: &ecx, edx: &edx); |
| 55 | |
| 56 | uint32_t core_type = (eax >> 24) & 0xFF; |
| 57 | |
| 58 | switch (core_type) { |
| 59 | case 0x40: cap->class = CPU_CLASS_PERFORMANCE; break; |
| 60 | case 0x20: cap->class = CPU_CLASS_EFFICIENCY; break; |
| 61 | default: cap->class = CPU_CLASS_UNKNOWN; break; |
| 62 | } |
| 63 | |
| 64 | cap->uarch_id = core_type; |
| 65 | } |
| 66 | |
| 67 | static uint32_t detect_uarch_id(void) { |
| 68 | uint32_t eax, ebx, ecx, edx; |
| 69 | |
| 70 | cpuid_count(leaf: 1, subleaf: 0, eax: &eax, ebx: &ebx, ecx: &ecx, edx: &edx); |
| 71 | |
| 72 | uint32_t family = (eax >> 8) & 0xF; |
| 73 | uint32_t model = (eax >> 4) & 0xF; |
| 74 | |
| 75 | if (family == 6) |
| 76 | model |= ((eax >> 16) & 0xF) << 4; |
| 77 | |
| 78 | switch (model) { |
| 79 | case 0x97: /* Alder Lake P */ |
| 80 | case 0x9A: /* Raptor Lake P */ return UARCH_GOLDEN_COVE; |
| 81 | case 0x9C: /* Alder Lake E */ return UARCH_GRACEMONT; |
| 82 | default: return UARCH_UNKNOWN; |
| 83 | } |
| 84 | } |
| 85 | |
| 86 | static void detect_pipeline_width(struct cpu_capability *cap) { |
| 87 | switch (cap->uarch_id) { |
| 88 | case UARCH_GOLDEN_COVE: |
| 89 | cap->issue_width = 6; |
| 90 | cap->retire_width = 6; |
| 91 | break; |
| 92 | |
| 93 | case UARCH_GRACEMONT: |
| 94 | cap->issue_width = 3; |
| 95 | cap->retire_width = 3; |
| 96 | break; |
| 97 | |
| 98 | default: |
| 99 | cap->issue_width = 2; |
| 100 | cap->retire_width = 2; |
| 101 | break; |
| 102 | } |
| 103 | } |
| 104 | |
| 105 | static const char *cpu_class_str(enum cpu_class c) { |
| 106 | switch (c) { |
| 107 | case CPU_CLASS_PERFORMANCE: return "P-core" ; |
| 108 | case CPU_CLASS_EFFICIENCY: return "E-core" ; |
| 109 | default: return "unknown" ; |
| 110 | } |
| 111 | } |
| 112 | |
| 113 | static const char *uarch_str(uint32_t uarch) { |
| 114 | switch (uarch) { |
| 115 | case UARCH_GOLDEN_COVE: return "Golden Cove" ; |
| 116 | case UARCH_GRACEMONT: return "Gracemont" ; |
| 117 | default: return "unknown" ; |
| 118 | } |
| 119 | } |
| 120 | |
| 121 | static void dump_cpu_features(uint64_t f) { |
| 122 | char buf[128]; |
| 123 | buf[0] = '\0'; |
| 124 | |
| 125 | if (f & CPU_FEAT_SSE2) |
| 126 | strcat(dest: buf, src: " SSE2" ); |
| 127 | if (f & CPU_FEAT_AVX) |
| 128 | strcat(dest: buf, src: " AVX" ); |
| 129 | if (f & CPU_FEAT_AVX2) |
| 130 | strcat(dest: buf, src: " AVX2" ); |
| 131 | if (f & CPU_FEAT_AVX512F) |
| 132 | strcat(dest: buf, src: " AVX-512F" ); |
| 133 | |
| 134 | if (buf[0] == '\0') |
| 135 | strcpy(dest: buf, src: " (none)" ); |
| 136 | |
| 137 | log_msg(LOG_INFO, " Features:%s" , buf); |
| 138 | } |
| 139 | |
| 140 | void smp_dump_core(struct core *c) { |
| 141 | log_msg(LOG_INFO, "CPU%zu: pkg=%u core=%u smt=%u numa=%zu domain_cpu=%zu" , |
| 142 | c->id, c->package_id, c->core_id, c->smt_id, c->numa_node, |
| 143 | c->domain_cpu_id); |
| 144 | |
| 145 | /* SMT / topology */ |
| 146 | log_msg(LOG_INFO, " Topology: smt_mask=0x%x llc_shared=%u" , c->smt_mask, |
| 147 | c->llc.cores_sharing); |
| 148 | |
| 149 | /* Cache */ |
| 150 | log_msg(LOG_INFO, " LLC: L%u size=%uKB line=%uB type=%u" , c->llc.level, |
| 151 | c->llc.size_kb, c->llc.line_size, c->llc.type); |
| 152 | |
| 153 | /* CPU capability block */ |
| 154 | struct cpu_capability *cap = &c->cap; |
| 155 | |
| 156 | log_msg(LOG_INFO, " Class: %s (%s)" , cpu_class_str(cap->class), |
| 157 | uarch_str(cap->uarch_id)); |
| 158 | |
| 159 | log_msg(LOG_INFO, " Widths: issue=%u retire=%u" , cap->issue_width, |
| 160 | cap->retire_width); |
| 161 | |
| 162 | log_msg(LOG_INFO, " Scores: perf=%u energy=%u" , cap->perf_score, |
| 163 | cap->energy_score); |
| 164 | |
| 165 | dump_cpu_features(f: cap->feature_bits); |
| 166 | } |
| 167 | |
| 168 | static void detect_llc(struct topology_cache_info *llc) { |
| 169 | uint32_t eax, ebx, ecx, edx; |
| 170 | for (uint32_t idx = 0;; idx++) { |
| 171 | |
| 172 | cpuid_count(leaf: 4, subleaf: idx, eax: &eax, ebx: &ebx, ecx: &ecx, edx: &edx); |
| 173 | |
| 174 | uint32_t cache_type = eax & 0x1F; |
| 175 | if (cache_type == 0) |
| 176 | break; |
| 177 | |
| 178 | uint32_t cache_level = (eax >> 5) & 0x7; |
| 179 | if (cache_level != 3) |
| 180 | continue; |
| 181 | |
| 182 | llc->level = cache_level; |
| 183 | llc->type = cache_type; |
| 184 | llc->line_size = (ebx & 0xFFF) + 1; |
| 185 | llc->cores_sharing = ((eax >> 14) & 0xFFF) + 1; |
| 186 | |
| 187 | uint32_t sets = ecx + 1; |
| 188 | uint32_t ways = ((ebx >> 22) & 0x3FF) + 1; |
| 189 | llc->size_kb = (ways * sets * llc->line_size) / 1024; |
| 190 | |
| 191 | break; |
| 192 | } |
| 193 | } |
| 194 | |
| 195 | static void init_smt_info(struct core *c) { |
| 196 | uint32_t eax, ebx, ecx, edx; |
| 197 | |
| 198 | uint32_t smt_width = 0; |
| 199 | uint32_t core_width = 0; |
| 200 | uint32_t apic_id; |
| 201 | |
| 202 | for (uint32_t level = 0;; level++) { |
| 203 | cpuid_count(leaf: 0xB, subleaf: level, eax: &eax, ebx: &ebx, ecx: &ecx, edx: &edx); |
| 204 | uint32_t level_type = (ecx >> 8) & 0xFF; |
| 205 | if (level_type == 0) |
| 206 | break; |
| 207 | |
| 208 | if (level_type == 1) { |
| 209 | smt_width = eax & 0x1F; |
| 210 | } else if (level_type == 2) { |
| 211 | core_width = eax & 0x1F; |
| 212 | } |
| 213 | } |
| 214 | |
| 215 | apic_id = c->id; |
| 216 | c->package_id = c->id >> core_width; |
| 217 | c->smt_mask = (1 << smt_width) - 1; |
| 218 | c->smt_id = apic_id & c->smt_mask; |
| 219 | c->core_id = (apic_id >> smt_width) & ((1 << (core_width - smt_width)) - 1); |
| 220 | } |
| 221 | |
| 222 | static void detect_cpu_capability(struct core *c) { |
| 223 | struct cpu_capability *cap = &c->cap; |
| 224 | |
| 225 | detect_cpu_features(cap); |
| 226 | |
| 227 | cap->uarch_id = detect_uarch_id(); |
| 228 | |
| 229 | detect_cpu_class(cap); |
| 230 | detect_pipeline_width(cap); |
| 231 | } |
| 232 | |
| 233 | static struct core *setup_cpu(uint64_t cpu) { |
| 234 | struct core *c = global.cores[cpu]; |
| 235 | kassert(c->id == cpu); |
| 236 | c->self = c; |
| 237 | c->current_irql = IRQL_PASSIVE_LEVEL; |
| 238 | c->tsc_hz = tsc_calibrate(); |
| 239 | init_smt_info(c); |
| 240 | detect_llc(llc: &c->llc); |
| 241 | detect_cpu_capability(c); |
| 242 | |
| 243 | wrmsr(MSR_GS_BASE, value: (uint64_t) c); |
| 244 | return c; |
| 245 | } |
| 246 | |
| 247 | static inline void set_core_awake(void) { |
| 248 | atomic_fetch_add_explicit(&cores_awake, 1, memory_order_release); |
| 249 | if (atomic_load_explicit(&cores_awake, memory_order_acquire) == |
| 250 | (global.core_count - 1)) { |
| 251 | bootstage_advance(new: BOOTSTAGE_MID_MP); |
| 252 | } |
| 253 | } |
| 254 | |
| 255 | void smp_wakeup() { |
| 256 | disable_interrupts(); |
| 257 | |
| 258 | asm volatile("mov %0, %%cr3" ::"r" (cr3)); |
| 259 | |
| 260 | x2apic_init(); |
| 261 | uint64_t cpu = cpu_get_this_id(); |
| 262 | setup_cpu(cpu); |
| 263 | |
| 264 | gdt_install(); |
| 265 | wrmsr(MSR_GS_BASE, value: (uint64_t) global.cores[cpu]); |
| 266 | irq_load(); |
| 267 | |
| 268 | lapic_timer_init(core_id: cpu); |
| 269 | tsc_sync_check_ap(self: cpu); |
| 270 | set_core_awake(); |
| 271 | |
| 272 | scheduler_yield(); |
| 273 | } |
| 274 | |
| 275 | void smp_init() { |
| 276 | for (size_t i = 0; i < global.core_count; i++) { |
| 277 | size_t d = domain_for_cpu(cpu: i); |
| 278 | |
| 279 | if (i != 0) { |
| 280 | global.cores[i] = |
| 281 | alloc_or_die(kmalloc_from_domain(d, sizeof(struct core))); |
| 282 | |
| 283 | memset(global.cores[i], 0, sizeof(struct core)); |
| 284 | } |
| 285 | |
| 286 | global.cores[i]->irq_entered_irql = IRQL_NONE; |
| 287 | global.cores[i]->id = i; |
| 288 | global.cores[i]->numa_node = numa_node_for_cpu(cpu: i); |
| 289 | global.cores[i]->domain = global.domains[d]; |
| 290 | } |
| 291 | } |
| 292 | |
| 293 | void smp_wait_for_others_to_idle() { |
| 294 | /* wait for them to enter idle threads */ |
| 295 | size_t expected_idle = global.core_count - 1; |
| 296 | while (atomic_load(&global.idle_core_count) < expected_idle) { |
| 297 | cpu_relax(); |
| 298 | } |
| 299 | } |
| 300 | |
| 301 | void smp_wake(struct limine_mp_response *mpr) { |
| 302 | asm volatile("mov %%cr3, %0" : "=r" (cr3)); |
| 303 | for (uint64_t i = 1; i < mpr->cpu_count; i++) |
| 304 | mpr->cpus[i]->goto_address = smp_wakeup; |
| 305 | |
| 306 | smp_core(cond: TOPC_NONE)->tsc_hz = tsc_calibrate(); |
| 307 | if (global.core_count == 1) |
| 308 | return; |
| 309 | |
| 310 | tsc_sync_check_all_aps(); |
| 311 | /* wait for bootstage to progress, TODO: Find a better |
| 312 | * way to do this, reading a volatile and NOT using |
| 313 | * atomics here is problematic */ |
| 314 | while (global.current_bootstage != BOOTSTAGE_MID_MP) |
| 315 | cpu_relax(); |
| 316 | |
| 317 | smp_wait_for_others_to_idle(); |
| 318 | } |
| 319 | |
| 320 | void smp_setup_bsp(void) { |
| 321 | struct core *c = kmalloc(sizeof(struct core), ALLOC_FLAGS_ZERO); |
| 322 | if (!c) |
| 323 | panic("Could not allocate space for core structure on BSP" ); |
| 324 | |
| 325 | c->irq_entered_irql = IRQL_NONE; |
| 326 | c->id = 0; |
| 327 | c->self = c; |
| 328 | c->current_irql = IRQL_PASSIVE_LEVEL; |
| 329 | wrmsr(MSR_GS_BASE, value: (uint64_t) c); |
| 330 | global.cores = |
| 331 | kmalloc(sizeof(struct core *) * global.core_count, ALLOC_FLAGS_ZERO); |
| 332 | |
| 333 | if (unlikely(!global.cores)) |
| 334 | panic("Could not allocate space for global core structures" ); |
| 335 | |
| 336 | global.shootdown_data = kmalloc( |
| 337 | sizeof(struct tlb_shootdown_cpu) * global.core_count, ALLOC_FLAGS_ZERO); |
| 338 | if (!global.shootdown_data) |
| 339 | panic("Could not allocate global shootdown data" ); |
| 340 | |
| 341 | global.cores[0] = c; |
| 342 | init_smt_info(c); |
| 343 | detect_llc(llc: &c->llc); |
| 344 | detect_cpu_capability(c); |
| 345 | tsc_mailboxes_init(); |
| 346 | } |
| 347 | |
| 348 | static atomic_uint tick_change_state = 0; |
| 349 | static bool enable = false; |
| 350 | static uint8_t entry = 0; |
| 351 | |
| 352 | static enum irq_result tick_op_isr(void *ctx, uint8_t vector, |
| 353 | struct irq_context *rsp) { |
| 354 | if (enable) { |
| 355 | scheduler_tick_enable(); |
| 356 | } else { |
| 357 | scheduler_tick_disable(); |
| 358 | } |
| 359 | |
| 360 | atomic_fetch_add(&tick_change_state, 1); |
| 361 | return IRQ_HANDLED; |
| 362 | } |
| 363 | |
| 364 | static void send_em_all_out(bool e) { |
| 365 | enable = e; |
| 366 | |
| 367 | atomic_store(&tick_change_state, 0); |
| 368 | size_t i; |
| 369 | for_each_cpu_id(i) { |
| 370 | if (i == 0) |
| 371 | continue; |
| 372 | |
| 373 | ipi_send(apic_id: i, vector: entry); |
| 374 | } |
| 375 | |
| 376 | /* wait for everyone to change their tick state */ |
| 377 | while (atomic_load(&tick_change_state) < (global.core_count - 1)) |
| 378 | cpu_relax(); |
| 379 | } |
| 380 | |
| 381 | void smp_disable_all_ticks() { |
| 382 | entry = irq_alloc_entry(); |
| 383 | irq_register(name: "tick_op" , vector: entry, handler: tick_op_isr, NULL, flags: IRQ_FLAG_NONE); |
| 384 | irq_set_chip(vector: entry, chip: lapic_get_chip(), NULL); |
| 385 | send_em_all_out(false); |
| 386 | } |
| 387 | |
| 388 | extern void nop_handler(void *, uint8_t, void *); |
| 389 | void smp_enable_all_ticks() { |
| 390 | send_em_all_out(true); |
| 391 | irq_free_entry(entry); |
| 392 | irq_set_chip(vector: entry, NULL, NULL); |
| 393 | } |
| 394 | |
| 395 | struct core *smp_bsp(void) { |
| 396 | return global.cores[0]; |
| 397 | } |
| 398 | |
| 399 | void smp_caller_verify(enum topology_caller caller) { |
| 400 | struct topology_contract tct = { |
| 401 | .caller = caller, |
| 402 | .scope = TOPOLOGY_LEVEL_SMT, |
| 403 | }; |
| 404 | |
| 405 | bool valid = topology_contract_verify(c: tct); |
| 406 | kassert(valid, "Caller 0x%lx was not satisfied" , caller); |
| 407 | } |
| 408 | |