1 2 #include "x86/msr.h" 3 #include "x86/processor.h" 4 #include "x86/pmu.h" 5 #include "x86/apic-defs.h" 6 #include "x86/apic.h" 7 #include "x86/desc.h" 8 #include "x86/isr.h" 9 #include "vmalloc.h" 10 #include "alloc.h" 11 12 #include "libcflat.h" 13 #include <stdint.h> 14 15 #define N 1000000 16 17 // These values match the number of instructions and branches in the 18 // assembly block in check_emulated_instr(). 19 #define EXPECTED_INSTR 17 20 #define EXPECTED_BRNCH 5 21 22 typedef struct { 23 uint32_t ctr; 24 uint64_t config; 25 uint64_t count; 26 int idx; 27 } pmu_counter_t; 28 29 struct pmu_event { 30 const char *name; 31 uint32_t unit_sel; 32 int min; 33 int max; 34 } intel_gp_events[] = { 35 {"core cycles", 0x003c, 1*N, 50*N}, 36 {"instructions", 0x00c0, 10*N, 10.2*N}, 37 {"ref cycles", 0x013c, 1*N, 30*N}, 38 {"llc references", 0x4f2e, 1, 2*N}, 39 {"llc misses", 0x412e, 1, 1*N}, 40 {"branches", 0x00c4, 1*N, 1.1*N}, 41 {"branch misses", 0x00c5, 0, 0.1*N}, 42 }, amd_gp_events[] = { 43 {"core cycles", 0x0076, 1*N, 50*N}, 44 {"instructions", 0x00c0, 10*N, 10.2*N}, 45 {"branches", 0x00c2, 1*N, 1.1*N}, 46 {"branch misses", 0x00c3, 0, 0.1*N}, 47 }, fixed_events[] = { 48 {"fixed 1", MSR_CORE_PERF_FIXED_CTR0, 10*N, 10.2*N}, 49 {"fixed 2", MSR_CORE_PERF_FIXED_CTR0 + 1, 1*N, 30*N}, 50 {"fixed 3", MSR_CORE_PERF_FIXED_CTR0 + 2, 0.1*N, 30*N} 51 }; 52 53 char *buf; 54 55 static struct pmu_event *gp_events; 56 static unsigned int gp_events_size; 57 58 static inline void loop(void) 59 { 60 unsigned long tmp, tmp2, tmp3; 61 62 asm volatile("1: mov (%1), %2; add $64, %1; nop; nop; nop; nop; nop; nop; nop; loop 1b" 63 : "=c"(tmp), "=r"(tmp2), "=r"(tmp3): "0"(N), "1"(buf)); 64 65 } 66 67 volatile uint64_t irq_received; 68 69 static void cnt_overflow(isr_regs_t *regs) 70 { 71 irq_received++; 72 apic_write(APIC_LVTPC, apic_read(APIC_LVTPC) & ~APIC_LVT_MASKED); 73 apic_write(APIC_EOI, 0); 74 } 75 76 static bool check_irq(void) 77 { 78 int i; 79 irq_received = 0; 80 sti(); 81 for (i = 0; i < 100000 && !irq_received; i++) 82 asm volatile("pause"); 83 cli(); 84 return irq_received; 85 } 86 87 static bool is_gp(pmu_counter_t *evt) 88 { 89 if (!pmu.is_intel) 90 return true; 91 92 return evt->ctr < MSR_CORE_PERF_FIXED_CTR0 || 93 evt->ctr >= MSR_IA32_PMC0; 94 } 95 96 static int event_to_global_idx(pmu_counter_t *cnt) 97 { 98 if (pmu.is_intel) 99 return cnt->ctr - (is_gp(cnt) ? pmu.msr_gp_counter_base : 100 (MSR_CORE_PERF_FIXED_CTR0 - FIXED_CNT_INDEX)); 101 102 if (pmu.msr_gp_counter_base == MSR_F15H_PERF_CTR0) 103 return (cnt->ctr - pmu.msr_gp_counter_base) / 2; 104 else 105 return cnt->ctr - pmu.msr_gp_counter_base; 106 } 107 108 static struct pmu_event* get_counter_event(pmu_counter_t *cnt) 109 { 110 if (is_gp(cnt)) { 111 int i; 112 113 for (i = 0; i < gp_events_size; i++) 114 if (gp_events[i].unit_sel == (cnt->config & 0xffff)) 115 return &gp_events[i]; 116 } else 117 return &fixed_events[cnt->ctr - MSR_CORE_PERF_FIXED_CTR0]; 118 119 return (void*)0; 120 } 121 122 static void global_enable(pmu_counter_t *cnt) 123 { 124 if (!this_cpu_has_perf_global_ctrl()) 125 return; 126 127 cnt->idx = event_to_global_idx(cnt); 128 wrmsr(pmu.msr_global_ctl, rdmsr(pmu.msr_global_ctl) | BIT_ULL(cnt->idx)); 129 } 130 131 static void global_disable(pmu_counter_t *cnt) 132 { 133 if (!this_cpu_has_perf_global_ctrl()) 134 return; 135 136 wrmsr(pmu.msr_global_ctl, rdmsr(pmu.msr_global_ctl) & ~BIT_ULL(cnt->idx)); 137 } 138 139 static void __start_event(pmu_counter_t *evt, uint64_t count) 140 { 141 evt->count = count; 142 wrmsr(evt->ctr, evt->count); 143 if (is_gp(evt)) { 144 wrmsr(MSR_GP_EVENT_SELECTx(event_to_global_idx(evt)), 145 evt->config | EVNTSEL_EN); 146 } else { 147 uint32_t ctrl = rdmsr(MSR_CORE_PERF_FIXED_CTR_CTRL); 148 int shift = (evt->ctr - MSR_CORE_PERF_FIXED_CTR0) * 4; 149 uint32_t usrospmi = 0; 150 151 if (evt->config & EVNTSEL_OS) 152 usrospmi |= (1 << 0); 153 if (evt->config & EVNTSEL_USR) 154 usrospmi |= (1 << 1); 155 if (evt->config & EVNTSEL_INT) 156 usrospmi |= (1 << 3); // PMI on overflow 157 ctrl = (ctrl & ~(0xf << shift)) | (usrospmi << shift); 158 wrmsr(MSR_CORE_PERF_FIXED_CTR_CTRL, ctrl); 159 } 160 global_enable(evt); 161 apic_write(APIC_LVTPC, PMI_VECTOR); 162 } 163 164 static void start_event(pmu_counter_t *evt) 165 { 166 __start_event(evt, 0); 167 } 168 169 static void stop_event(pmu_counter_t *evt) 170 { 171 global_disable(evt); 172 if (is_gp(evt)) { 173 wrmsr(MSR_GP_EVENT_SELECTx(event_to_global_idx(evt)), 174 evt->config & ~EVNTSEL_EN); 175 } else { 176 uint32_t ctrl = rdmsr(MSR_CORE_PERF_FIXED_CTR_CTRL); 177 int shift = (evt->ctr - MSR_CORE_PERF_FIXED_CTR0) * 4; 178 wrmsr(MSR_CORE_PERF_FIXED_CTR_CTRL, ctrl & ~(0xf << shift)); 179 } 180 evt->count = rdmsr(evt->ctr); 181 } 182 183 static noinline void measure_many(pmu_counter_t *evt, int count) 184 { 185 int i; 186 for (i = 0; i < count; i++) 187 start_event(&evt[i]); 188 loop(); 189 for (i = 0; i < count; i++) 190 stop_event(&evt[i]); 191 } 192 193 static void measure_one(pmu_counter_t *evt) 194 { 195 measure_many(evt, 1); 196 } 197 198 static noinline void __measure(pmu_counter_t *evt, uint64_t count) 199 { 200 __start_event(evt, count); 201 loop(); 202 stop_event(evt); 203 } 204 205 static bool verify_event(uint64_t count, struct pmu_event *e) 206 { 207 // printf("%d <= %ld <= %d\n", e->min, count, e->max); 208 return count >= e->min && count <= e->max; 209 } 210 211 static bool verify_counter(pmu_counter_t *cnt) 212 { 213 return verify_event(cnt->count, get_counter_event(cnt)); 214 } 215 216 static void check_gp_counter(struct pmu_event *evt) 217 { 218 pmu_counter_t cnt = { 219 .config = EVNTSEL_OS | EVNTSEL_USR | evt->unit_sel, 220 }; 221 int i; 222 223 for (i = 0; i < pmu.nr_gp_counters; i++) { 224 cnt.ctr = MSR_GP_COUNTERx(i); 225 measure_one(&cnt); 226 report(verify_event(cnt.count, evt), "%s-%d", evt->name, i); 227 } 228 } 229 230 static void check_gp_counters(void) 231 { 232 int i; 233 234 for (i = 0; i < gp_events_size; i++) 235 if (pmu_gp_counter_is_available(i)) 236 check_gp_counter(&gp_events[i]); 237 else 238 printf("GP event '%s' is disabled\n", 239 gp_events[i].name); 240 } 241 242 static void check_fixed_counters(void) 243 { 244 pmu_counter_t cnt = { 245 .config = EVNTSEL_OS | EVNTSEL_USR, 246 }; 247 int i; 248 249 for (i = 0; i < pmu.nr_fixed_counters; i++) { 250 cnt.ctr = fixed_events[i].unit_sel; 251 measure_one(&cnt); 252 report(verify_event(cnt.count, &fixed_events[i]), "fixed-%d", i); 253 } 254 } 255 256 static void check_counters_many(void) 257 { 258 pmu_counter_t cnt[10]; 259 int i, n; 260 261 for (i = 0, n = 0; n < pmu.nr_gp_counters; i++) { 262 if (!pmu_gp_counter_is_available(i)) 263 continue; 264 265 cnt[n].ctr = MSR_GP_COUNTERx(n); 266 cnt[n].config = EVNTSEL_OS | EVNTSEL_USR | 267 gp_events[i % gp_events_size].unit_sel; 268 n++; 269 } 270 for (i = 0; i < pmu.nr_fixed_counters; i++) { 271 cnt[n].ctr = fixed_events[i].unit_sel; 272 cnt[n].config = EVNTSEL_OS | EVNTSEL_USR; 273 n++; 274 } 275 276 measure_many(cnt, n); 277 278 for (i = 0; i < n; i++) 279 if (!verify_counter(&cnt[i])) 280 break; 281 282 report(i == n, "all counters"); 283 } 284 285 static uint64_t measure_for_overflow(pmu_counter_t *cnt) 286 { 287 __measure(cnt, 0); 288 /* 289 * To generate overflow, i.e. roll over to '0', the initial count just 290 * needs to be preset to the negative expected count. However, as per 291 * Intel's SDM, the preset count needs to be incremented by 1 to ensure 292 * the overflow interrupt is generated immediately instead of possibly 293 * waiting for the overflow to propagate through the counter. 294 */ 295 assert(cnt->count > 1); 296 return 1 - cnt->count; 297 } 298 299 static void check_counter_overflow(void) 300 { 301 uint64_t overflow_preset; 302 int i; 303 pmu_counter_t cnt = { 304 .ctr = MSR_GP_COUNTERx(0), 305 .config = EVNTSEL_OS | EVNTSEL_USR | gp_events[1].unit_sel /* instructions */, 306 }; 307 overflow_preset = measure_for_overflow(&cnt); 308 309 /* clear status before test */ 310 if (this_cpu_has_perf_global_status()) 311 pmu_clear_global_status(); 312 313 report_prefix_push("overflow"); 314 315 for (i = 0; i < pmu.nr_gp_counters + 1; i++) { 316 uint64_t status; 317 int idx; 318 319 cnt.count = overflow_preset; 320 if (pmu_use_full_writes()) 321 cnt.count &= (1ull << pmu.gp_counter_width) - 1; 322 323 if (i == pmu.nr_gp_counters) { 324 if (!pmu.is_intel) 325 break; 326 327 cnt.ctr = fixed_events[0].unit_sel; 328 cnt.count = measure_for_overflow(&cnt); 329 cnt.count &= (1ull << pmu.gp_counter_width) - 1; 330 } else { 331 cnt.ctr = MSR_GP_COUNTERx(i); 332 } 333 334 if (i % 2) 335 cnt.config |= EVNTSEL_INT; 336 else 337 cnt.config &= ~EVNTSEL_INT; 338 idx = event_to_global_idx(&cnt); 339 __measure(&cnt, cnt.count); 340 if (pmu.is_intel) 341 report(cnt.count == 1, "cntr-%d", i); 342 else 343 report(cnt.count == 0xffffffffffff || cnt.count < 7, "cntr-%d", i); 344 345 if (!this_cpu_has_perf_global_status()) 346 continue; 347 348 status = rdmsr(pmu.msr_global_status); 349 report(status & (1ull << idx), "status-%d", i); 350 wrmsr(pmu.msr_global_status_clr, status); 351 status = rdmsr(pmu.msr_global_status); 352 report(!(status & (1ull << idx)), "status clear-%d", i); 353 report(check_irq() == (i % 2), "irq-%d", i); 354 } 355 356 report_prefix_pop(); 357 } 358 359 static void check_gp_counter_cmask(void) 360 { 361 pmu_counter_t cnt = { 362 .ctr = MSR_GP_COUNTERx(0), 363 .config = EVNTSEL_OS | EVNTSEL_USR | gp_events[1].unit_sel /* instructions */, 364 }; 365 cnt.config |= (0x2 << EVNTSEL_CMASK_SHIFT); 366 measure_one(&cnt); 367 report(cnt.count < gp_events[1].min, "cmask"); 368 } 369 370 static void do_rdpmc_fast(void *ptr) 371 { 372 pmu_counter_t *cnt = ptr; 373 uint32_t idx = (uint32_t)cnt->idx | (1u << 31); 374 375 if (!is_gp(cnt)) 376 idx |= 1 << 30; 377 378 cnt->count = rdpmc(idx); 379 } 380 381 382 static void check_rdpmc(void) 383 { 384 uint64_t val = 0xff0123456789ull; 385 bool exc; 386 int i; 387 388 report_prefix_push("rdpmc"); 389 390 for (i = 0; i < pmu.nr_gp_counters; i++) { 391 uint64_t x; 392 pmu_counter_t cnt = { 393 .ctr = MSR_GP_COUNTERx(i), 394 .idx = i 395 }; 396 397 /* 398 * Without full-width writes, only the low 32 bits are writable, 399 * and the value is sign-extended. 400 */ 401 if (pmu.msr_gp_counter_base == MSR_IA32_PERFCTR0) 402 x = (uint64_t)(int64_t)(int32_t)val; 403 else 404 x = (uint64_t)(int64_t)val; 405 406 /* Mask according to the number of supported bits */ 407 x &= (1ull << pmu.gp_counter_width) - 1; 408 409 wrmsr(MSR_GP_COUNTERx(i), val); 410 report(rdpmc(i) == x, "cntr-%d", i); 411 412 exc = test_for_exception(GP_VECTOR, do_rdpmc_fast, &cnt); 413 if (exc) 414 report_skip("fast-%d", i); 415 else 416 report(cnt.count == (u32)val, "fast-%d", i); 417 } 418 for (i = 0; i < pmu.nr_fixed_counters; i++) { 419 uint64_t x = val & ((1ull << pmu.fixed_counter_width) - 1); 420 pmu_counter_t cnt = { 421 .ctr = MSR_CORE_PERF_FIXED_CTR0 + i, 422 .idx = i 423 }; 424 425 wrmsr(MSR_PERF_FIXED_CTRx(i), x); 426 report(rdpmc(i | (1 << 30)) == x, "fixed cntr-%d", i); 427 428 exc = test_for_exception(GP_VECTOR, do_rdpmc_fast, &cnt); 429 if (exc) 430 report_skip("fixed fast-%d", i); 431 else 432 report(cnt.count == (u32)x, "fixed fast-%d", i); 433 } 434 435 report_prefix_pop(); 436 } 437 438 static void check_running_counter_wrmsr(void) 439 { 440 uint64_t status; 441 uint64_t count; 442 pmu_counter_t evt = { 443 .ctr = MSR_GP_COUNTERx(0), 444 .config = EVNTSEL_OS | EVNTSEL_USR | gp_events[1].unit_sel, 445 }; 446 447 report_prefix_push("running counter wrmsr"); 448 449 start_event(&evt); 450 loop(); 451 wrmsr(MSR_GP_COUNTERx(0), 0); 452 stop_event(&evt); 453 report(evt.count < gp_events[1].min, "cntr"); 454 455 /* clear status before overflow test */ 456 if (this_cpu_has_perf_global_status()) 457 pmu_clear_global_status(); 458 459 start_event(&evt); 460 461 count = -1; 462 if (pmu_use_full_writes()) 463 count &= (1ull << pmu.gp_counter_width) - 1; 464 465 wrmsr(MSR_GP_COUNTERx(0), count); 466 467 loop(); 468 stop_event(&evt); 469 470 if (this_cpu_has_perf_global_status()) { 471 status = rdmsr(pmu.msr_global_status); 472 report(status & 1, "status msr bit"); 473 } 474 475 report_prefix_pop(); 476 } 477 478 static void check_emulated_instr(void) 479 { 480 uint64_t status, instr_start, brnch_start; 481 uint64_t gp_counter_width = (1ull << pmu.gp_counter_width) - 1; 482 unsigned int branch_idx = pmu.is_intel ? 5 : 2; 483 pmu_counter_t brnch_cnt = { 484 .ctr = MSR_GP_COUNTERx(0), 485 /* branch instructions */ 486 .config = EVNTSEL_OS | EVNTSEL_USR | gp_events[branch_idx].unit_sel, 487 }; 488 pmu_counter_t instr_cnt = { 489 .ctr = MSR_GP_COUNTERx(1), 490 /* instructions */ 491 .config = EVNTSEL_OS | EVNTSEL_USR | gp_events[1].unit_sel, 492 }; 493 report_prefix_push("emulated instruction"); 494 495 if (this_cpu_has_perf_global_status()) 496 pmu_clear_global_status(); 497 498 start_event(&brnch_cnt); 499 start_event(&instr_cnt); 500 501 brnch_start = -EXPECTED_BRNCH; 502 instr_start = -EXPECTED_INSTR; 503 wrmsr(MSR_GP_COUNTERx(0), brnch_start & gp_counter_width); 504 wrmsr(MSR_GP_COUNTERx(1), instr_start & gp_counter_width); 505 // KVM_FEP is a magic prefix that forces emulation so 506 // 'KVM_FEP "jne label\n"' just counts as a single instruction. 507 asm volatile( 508 "mov $0x0, %%eax\n" 509 "cmp $0x0, %%eax\n" 510 KVM_FEP "jne label\n" 511 KVM_FEP "jne label\n" 512 KVM_FEP "jne label\n" 513 KVM_FEP "jne label\n" 514 KVM_FEP "jne label\n" 515 "mov $0xa, %%eax\n" 516 "cpuid\n" 517 "mov $0xa, %%eax\n" 518 "cpuid\n" 519 "mov $0xa, %%eax\n" 520 "cpuid\n" 521 "mov $0xa, %%eax\n" 522 "cpuid\n" 523 "mov $0xa, %%eax\n" 524 "cpuid\n" 525 "label:\n" 526 : 527 : 528 : "eax", "ebx", "ecx", "edx"); 529 530 if (this_cpu_has_perf_global_ctrl()) 531 wrmsr(pmu.msr_global_ctl, 0); 532 533 stop_event(&brnch_cnt); 534 stop_event(&instr_cnt); 535 536 // Check that the end count - start count is at least the expected 537 // number of instructions and branches. 538 report(instr_cnt.count - instr_start >= EXPECTED_INSTR, 539 "instruction count"); 540 report(brnch_cnt.count - brnch_start >= EXPECTED_BRNCH, 541 "branch count"); 542 if (this_cpu_has_perf_global_status()) { 543 // Additionally check that those counters overflowed properly. 544 status = rdmsr(pmu.msr_global_status); 545 report(status & 1, "branch counter overflow"); 546 report(status & 2, "instruction counter overflow"); 547 } 548 549 report_prefix_pop(); 550 } 551 552 #define XBEGIN_STARTED (~0u) 553 static void check_tsx_cycles(void) 554 { 555 pmu_counter_t cnt; 556 unsigned int i, ret = 0; 557 558 if (!this_cpu_has(X86_FEATURE_RTM)) 559 return; 560 561 report_prefix_push("TSX cycles"); 562 563 for (i = 0; i < pmu.nr_gp_counters; i++) { 564 cnt.ctr = MSR_GP_COUNTERx(i); 565 566 if (i == 2) { 567 /* Transactional cycles committed only on gp counter 2 */ 568 cnt.config = EVNTSEL_OS | EVNTSEL_USR | 0x30000003c; 569 } else { 570 /* Transactional cycles */ 571 cnt.config = EVNTSEL_OS | EVNTSEL_USR | 0x10000003c; 572 } 573 574 start_event(&cnt); 575 576 asm volatile("xbegin 1f\n\t" 577 "1:\n\t" 578 : "+a" (ret) :: "memory"); 579 580 /* Generate a non-canonical #GP to trigger ABORT. */ 581 if (ret == XBEGIN_STARTED) 582 *(int *)NONCANONICAL = 0; 583 584 stop_event(&cnt); 585 586 report(cnt.count > 0, "gp cntr-%d with a value of %" PRId64 "", i, cnt.count); 587 } 588 589 report_prefix_pop(); 590 } 591 592 static void check_counters(void) 593 { 594 if (is_fep_available()) 595 check_emulated_instr(); 596 597 check_gp_counters(); 598 check_fixed_counters(); 599 check_rdpmc(); 600 check_counters_many(); 601 check_counter_overflow(); 602 check_gp_counter_cmask(); 603 check_running_counter_wrmsr(); 604 check_tsx_cycles(); 605 } 606 607 static void do_unsupported_width_counter_write(void *index) 608 { 609 wrmsr(MSR_IA32_PMC0 + *((int *) index), 0xffffff0123456789ull); 610 } 611 612 static void check_gp_counters_write_width(void) 613 { 614 u64 val_64 = 0xffffff0123456789ull; 615 u64 val_32 = val_64 & ((1ull << 32) - 1); 616 u64 val_max_width = val_64 & ((1ull << pmu.gp_counter_width) - 1); 617 int i; 618 619 /* 620 * MSR_IA32_PERFCTRn supports 64-bit writes, 621 * but only the lowest 32 bits are valid. 622 */ 623 for (i = 0; i < pmu.nr_gp_counters; i++) { 624 wrmsr(MSR_IA32_PERFCTR0 + i, val_32); 625 assert(rdmsr(MSR_IA32_PERFCTR0 + i) == val_32); 626 assert(rdmsr(MSR_IA32_PMC0 + i) == val_32); 627 628 wrmsr(MSR_IA32_PERFCTR0 + i, val_max_width); 629 assert(rdmsr(MSR_IA32_PERFCTR0 + i) == val_32); 630 assert(rdmsr(MSR_IA32_PMC0 + i) == val_32); 631 632 wrmsr(MSR_IA32_PERFCTR0 + i, val_64); 633 assert(rdmsr(MSR_IA32_PERFCTR0 + i) == val_32); 634 assert(rdmsr(MSR_IA32_PMC0 + i) == val_32); 635 } 636 637 /* 638 * MSR_IA32_PMCn supports writing values up to GP counter width, 639 * and only the lowest bits of GP counter width are valid. 640 */ 641 for (i = 0; i < pmu.nr_gp_counters; i++) { 642 wrmsr(MSR_IA32_PMC0 + i, val_32); 643 assert(rdmsr(MSR_IA32_PMC0 + i) == val_32); 644 assert(rdmsr(MSR_IA32_PERFCTR0 + i) == val_32); 645 646 wrmsr(MSR_IA32_PMC0 + i, val_max_width); 647 assert(rdmsr(MSR_IA32_PMC0 + i) == val_max_width); 648 assert(rdmsr(MSR_IA32_PERFCTR0 + i) == val_max_width); 649 650 report(test_for_exception(GP_VECTOR, 651 do_unsupported_width_counter_write, &i), 652 "writing unsupported width to MSR_IA32_PMC%d raises #GP", i); 653 } 654 } 655 656 /* 657 * Per the SDM, reference cycles are currently implemented using the 658 * core crystal clock, TSC, or bus clock. Calibrate to the TSC 659 * frequency to set reasonable expectations. 660 */ 661 static void set_ref_cycle_expectations(void) 662 { 663 pmu_counter_t cnt = { 664 .ctr = MSR_IA32_PERFCTR0, 665 .config = EVNTSEL_OS | EVNTSEL_USR | intel_gp_events[2].unit_sel, 666 }; 667 uint64_t tsc_delta; 668 uint64_t t0, t1, t2, t3; 669 670 /* Bit 2 enumerates the availability of reference cycles events. */ 671 if (!pmu.nr_gp_counters || !pmu_gp_counter_is_available(2)) 672 return; 673 674 if (this_cpu_has_perf_global_ctrl()) 675 wrmsr(pmu.msr_global_ctl, 0); 676 677 t0 = fenced_rdtsc(); 678 start_event(&cnt); 679 t1 = fenced_rdtsc(); 680 681 /* 682 * This loop has to run long enough to dominate the VM-exit 683 * costs for playing with the PMU MSRs on start and stop. 684 * 685 * On a 2.6GHz Ice Lake, with the TSC frequency at 104 times 686 * the core crystal clock, this function calculated a guest 687 * TSC : ref cycles ratio of around 105 with ECX initialized 688 * to one billion. 689 */ 690 asm volatile("loop ." : "+c"((int){1000000000ull})); 691 692 t2 = fenced_rdtsc(); 693 stop_event(&cnt); 694 t3 = fenced_rdtsc(); 695 696 tsc_delta = ((t2 - t1) + (t3 - t0)) / 2; 697 698 if (!tsc_delta) 699 return; 700 701 intel_gp_events[2].min = (intel_gp_events[2].min * cnt.count) / tsc_delta; 702 intel_gp_events[2].max = (intel_gp_events[2].max * cnt.count) / tsc_delta; 703 } 704 705 static void check_invalid_rdpmc_gp(void) 706 { 707 uint64_t val; 708 709 report(rdpmc_safe(64, &val) == GP_VECTOR, 710 "Expected #GP on RDPMC(64)"); 711 } 712 713 int main(int ac, char **av) 714 { 715 setup_vm(); 716 handle_irq(PMI_VECTOR, cnt_overflow); 717 buf = malloc(N*64); 718 719 check_invalid_rdpmc_gp(); 720 721 if (pmu.is_intel) { 722 if (!pmu.version) { 723 report_skip("No Intel Arch PMU is detected!"); 724 return report_summary(); 725 } 726 gp_events = (struct pmu_event *)intel_gp_events; 727 gp_events_size = sizeof(intel_gp_events)/sizeof(intel_gp_events[0]); 728 report_prefix_push("Intel"); 729 set_ref_cycle_expectations(); 730 } else { 731 gp_events_size = sizeof(amd_gp_events)/sizeof(amd_gp_events[0]); 732 gp_events = (struct pmu_event *)amd_gp_events; 733 report_prefix_push("AMD"); 734 } 735 736 printf("PMU version: %d\n", pmu.version); 737 printf("GP counters: %d\n", pmu.nr_gp_counters); 738 printf("GP counter width: %d\n", pmu.gp_counter_width); 739 printf("Mask length: %d\n", pmu.gp_counter_mask_length); 740 printf("Fixed counters: %d\n", pmu.nr_fixed_counters); 741 printf("Fixed counter width: %d\n", pmu.fixed_counter_width); 742 743 apic_write(APIC_LVTPC, PMI_VECTOR); 744 745 check_counters(); 746 747 if (pmu_has_full_writes()) { 748 pmu.msr_gp_counter_base = MSR_IA32_PMC0; 749 750 report_prefix_push("full-width writes"); 751 check_counters(); 752 check_gp_counters_write_width(); 753 report_prefix_pop(); 754 } 755 756 if (!pmu.is_intel) { 757 report_prefix_push("K7"); 758 pmu.nr_gp_counters = AMD64_NUM_COUNTERS; 759 pmu.msr_gp_counter_base = MSR_K7_PERFCTR0; 760 pmu.msr_gp_event_select_base = MSR_K7_EVNTSEL0; 761 check_counters(); 762 report_prefix_pop(); 763 } 764 765 return report_summary(); 766 } 767