1 /*- 2 * Copyright (c) 1998-2003 Poul-Henning Kamp 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice, this list of conditions and the following disclaimer. 10 * 2. Redistributions in binary form must reproduce the above copyright 11 * notice, this list of conditions and the following disclaimer in the 12 * documentation and/or other materials provided with the distribution. 13 * 14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 24 * SUCH DAMAGE. 25 */ 26 27 #include <sys/cdefs.h> 28 __FBSDID("$FreeBSD$"); 29 30 #include "opt_compat.h" 31 #include "opt_clock.h" 32 33 #include <sys/param.h> 34 #include <sys/bus.h> 35 #include <sys/cpu.h> 36 #include <sys/limits.h> 37 #include <sys/malloc.h> 38 #include <sys/systm.h> 39 #include <sys/sysctl.h> 40 #include <sys/time.h> 41 #include <sys/timetc.h> 42 #include <sys/kernel.h> 43 #include <sys/power.h> 44 #include <sys/smp.h> 45 #include <sys/vdso.h> 46 #include <machine/clock.h> 47 #include <machine/cputypes.h> 48 #include <machine/md_var.h> 49 #include <machine/specialreg.h> 50 #include <x86/vmware.h> 51 #include <dev/acpica/acpi_hpet.h> 52 53 #include "cpufreq_if.h" 54 55 uint64_t tsc_freq; 56 int tsc_is_invariant; 57 int tsc_perf_stat; 58 59 static eventhandler_tag tsc_levels_tag, tsc_pre_tag, tsc_post_tag; 60 61 SYSCTL_INT(_kern_timecounter, OID_AUTO, invariant_tsc, CTLFLAG_RDTUN, 62 &tsc_is_invariant, 0, "Indicates whether the TSC is P-state invariant"); 63 64 #ifdef SMP 65 int smp_tsc; 66 SYSCTL_INT(_kern_timecounter, OID_AUTO, smp_tsc, CTLFLAG_RDTUN, &smp_tsc, 0, 67 "Indicates whether the TSC is safe to use in SMP mode"); 68 69 int smp_tsc_adjust = 0; 70 SYSCTL_INT(_kern_timecounter, OID_AUTO, smp_tsc_adjust, CTLFLAG_RDTUN, 71 &smp_tsc_adjust, 0, "Try to adjust TSC on APs to match BSP"); 72 #endif 73 74 static int tsc_shift = 1; 75 SYSCTL_INT(_kern_timecounter, OID_AUTO, tsc_shift, CTLFLAG_RDTUN, 76 &tsc_shift, 0, "Shift to pre-apply for the maximum TSC frequency"); 77 78 static int tsc_disabled; 79 SYSCTL_INT(_machdep, OID_AUTO, disable_tsc, CTLFLAG_RDTUN, &tsc_disabled, 0, 80 "Disable x86 Time Stamp Counter"); 81 82 static int tsc_skip_calibration; 83 SYSCTL_INT(_machdep, OID_AUTO, disable_tsc_calibration, CTLFLAG_RDTUN, 84 &tsc_skip_calibration, 0, "Disable TSC frequency calibration"); 85 86 static void tsc_freq_changed(void *arg, const struct cf_level *level, 87 int status); 88 static void tsc_freq_changing(void *arg, const struct cf_level *level, 89 int *status); 90 static unsigned tsc_get_timecount(struct timecounter *tc); 91 static inline unsigned tsc_get_timecount_low(struct timecounter *tc); 92 static unsigned tsc_get_timecount_lfence(struct timecounter *tc); 93 static unsigned tsc_get_timecount_low_lfence(struct timecounter *tc); 94 static unsigned tsc_get_timecount_mfence(struct timecounter *tc); 95 static unsigned tsc_get_timecount_low_mfence(struct timecounter *tc); 96 static void tsc_levels_changed(void *arg, int unit); 97 static uint32_t x86_tsc_vdso_timehands(struct vdso_timehands *vdso_th, 98 struct timecounter *tc); 99 #ifdef COMPAT_FREEBSD32 100 static uint32_t x86_tsc_vdso_timehands32(struct vdso_timehands32 *vdso_th32, 101 struct timecounter *tc); 102 #endif 103 104 static struct timecounter tsc_timecounter = { 105 .tc_get_timecount = tsc_get_timecount, 106 .tc_counter_mask = ~0u, 107 .tc_name = "TSC", 108 .tc_quality = 800, /* adjusted in code */ 109 .tc_fill_vdso_timehands = x86_tsc_vdso_timehands, 110 #ifdef COMPAT_FREEBSD32 111 .tc_fill_vdso_timehands32 = x86_tsc_vdso_timehands32, 112 #endif 113 }; 114 115 static void 116 tsc_freq_vmware(void) 117 { 118 u_int regs[4]; 119 120 if (hv_high >= 0x40000010) { 121 do_cpuid(0x40000010, regs); 122 tsc_freq = regs[0] * 1000; 123 } else { 124 vmware_hvcall(VMW_HVCMD_GETHZ, regs); 125 if (regs[1] != UINT_MAX) 126 tsc_freq = regs[0] | ((uint64_t)regs[1] << 32); 127 } 128 tsc_is_invariant = 1; 129 } 130 131 static void 132 tsc_freq_intel(void) 133 { 134 char brand[48]; 135 u_int regs[4]; 136 uint64_t freq; 137 char *p; 138 u_int i; 139 140 /* 141 * Intel Processor Identification and the CPUID Instruction 142 * Application Note 485. 143 * http://www.intel.com/assets/pdf/appnote/241618.pdf 144 */ 145 if (cpu_exthigh >= 0x80000004) { 146 p = brand; 147 for (i = 0x80000002; i < 0x80000005; i++) { 148 do_cpuid(i, regs); 149 memcpy(p, regs, sizeof(regs)); 150 p += sizeof(regs); 151 } 152 p = NULL; 153 for (i = 0; i < sizeof(brand) - 1; i++) 154 if (brand[i] == 'H' && brand[i + 1] == 'z') 155 p = brand + i; 156 if (p != NULL) { 157 p -= 5; 158 switch (p[4]) { 159 case 'M': 160 i = 1; 161 break; 162 case 'G': 163 i = 1000; 164 break; 165 case 'T': 166 i = 1000000; 167 break; 168 default: 169 return; 170 } 171 #define C2D(c) ((c) - '0') 172 if (p[1] == '.') { 173 freq = C2D(p[0]) * 1000; 174 freq += C2D(p[2]) * 100; 175 freq += C2D(p[3]) * 10; 176 freq *= i * 1000; 177 } else { 178 freq = C2D(p[0]) * 1000; 179 freq += C2D(p[1]) * 100; 180 freq += C2D(p[2]) * 10; 181 freq += C2D(p[3]); 182 freq *= i * 1000000; 183 } 184 #undef C2D 185 tsc_freq = freq; 186 } 187 } 188 } 189 190 static void 191 probe_tsc_freq(void) 192 { 193 u_int regs[4]; 194 uint64_t tsc1, tsc2; 195 196 if (cpu_high >= 6) { 197 do_cpuid(6, regs); 198 if ((regs[2] & CPUID_PERF_STAT) != 0) { 199 /* 200 * XXX Some emulators expose host CPUID without actual 201 * support for these MSRs. We must test whether they 202 * really work. 203 */ 204 wrmsr(MSR_MPERF, 0); 205 wrmsr(MSR_APERF, 0); 206 DELAY(10); 207 if (rdmsr(MSR_MPERF) > 0 && rdmsr(MSR_APERF) > 0) 208 tsc_perf_stat = 1; 209 } 210 } 211 212 if (vm_guest == VM_GUEST_VMWARE) { 213 tsc_freq_vmware(); 214 return; 215 } 216 217 switch (cpu_vendor_id) { 218 case CPU_VENDOR_AMD: 219 if ((amd_pminfo & AMDPM_TSC_INVARIANT) != 0 || 220 (vm_guest == VM_GUEST_NO && 221 CPUID_TO_FAMILY(cpu_id) >= 0x10)) 222 tsc_is_invariant = 1; 223 if (cpu_feature & CPUID_SSE2) { 224 tsc_timecounter.tc_get_timecount = 225 tsc_get_timecount_mfence; 226 } 227 break; 228 case CPU_VENDOR_INTEL: 229 if ((amd_pminfo & AMDPM_TSC_INVARIANT) != 0 || 230 (vm_guest == VM_GUEST_NO && 231 ((CPUID_TO_FAMILY(cpu_id) == 0x6 && 232 CPUID_TO_MODEL(cpu_id) >= 0xe) || 233 (CPUID_TO_FAMILY(cpu_id) == 0xf && 234 CPUID_TO_MODEL(cpu_id) >= 0x3)))) 235 tsc_is_invariant = 1; 236 if (cpu_feature & CPUID_SSE2) { 237 tsc_timecounter.tc_get_timecount = 238 tsc_get_timecount_lfence; 239 } 240 break; 241 case CPU_VENDOR_CENTAUR: 242 if (vm_guest == VM_GUEST_NO && 243 CPUID_TO_FAMILY(cpu_id) == 0x6 && 244 CPUID_TO_MODEL(cpu_id) >= 0xf && 245 (rdmsr(0x1203) & 0x100000000ULL) == 0) 246 tsc_is_invariant = 1; 247 if (cpu_feature & CPUID_SSE2) { 248 tsc_timecounter.tc_get_timecount = 249 tsc_get_timecount_lfence; 250 } 251 break; 252 } 253 254 if (tsc_skip_calibration) { 255 if (cpu_vendor_id == CPU_VENDOR_INTEL) 256 tsc_freq_intel(); 257 return; 258 } 259 260 if (bootverbose) 261 printf("Calibrating TSC clock ... "); 262 tsc1 = rdtsc(); 263 DELAY(1000000); 264 tsc2 = rdtsc(); 265 tsc_freq = tsc2 - tsc1; 266 if (bootverbose) 267 printf("TSC clock: %ju Hz\n", (intmax_t)tsc_freq); 268 } 269 270 void 271 init_TSC(void) 272 { 273 274 if ((cpu_feature & CPUID_TSC) == 0 || tsc_disabled) 275 return; 276 277 #ifdef __i386__ 278 /* The TSC is known to be broken on certain CPUs. */ 279 switch (cpu_vendor_id) { 280 case CPU_VENDOR_AMD: 281 switch (cpu_id & 0xFF0) { 282 case 0x500: 283 /* K5 Model 0 */ 284 return; 285 } 286 break; 287 case CPU_VENDOR_CENTAUR: 288 switch (cpu_id & 0xff0) { 289 case 0x540: 290 /* 291 * http://www.centtech.com/c6_data_sheet.pdf 292 * 293 * I-12 RDTSC may return incoherent values in EDX:EAX 294 * I-13 RDTSC hangs when certain event counters are used 295 */ 296 return; 297 } 298 break; 299 case CPU_VENDOR_NSC: 300 switch (cpu_id & 0xff0) { 301 case 0x540: 302 if ((cpu_id & CPUID_STEPPING) == 0) 303 return; 304 break; 305 } 306 break; 307 } 308 #endif 309 310 probe_tsc_freq(); 311 312 /* 313 * Inform CPU accounting about our boot-time clock rate. This will 314 * be updated if someone loads a cpufreq driver after boot that 315 * discovers a new max frequency. 316 */ 317 if (tsc_freq != 0) 318 set_cputicker(rdtsc, tsc_freq, !tsc_is_invariant); 319 320 if (tsc_is_invariant) 321 return; 322 323 /* Register to find out about changes in CPU frequency. */ 324 tsc_pre_tag = EVENTHANDLER_REGISTER(cpufreq_pre_change, 325 tsc_freq_changing, NULL, EVENTHANDLER_PRI_FIRST); 326 tsc_post_tag = EVENTHANDLER_REGISTER(cpufreq_post_change, 327 tsc_freq_changed, NULL, EVENTHANDLER_PRI_FIRST); 328 tsc_levels_tag = EVENTHANDLER_REGISTER(cpufreq_levels_changed, 329 tsc_levels_changed, NULL, EVENTHANDLER_PRI_ANY); 330 } 331 332 #ifdef SMP 333 334 /* 335 * RDTSC is not a serializing instruction, and does not drain 336 * instruction stream, so we need to drain the stream before executing 337 * it. It could be fixed by use of RDTSCP, except the instruction is 338 * not available everywhere. 339 * 340 * Use CPUID for draining in the boot-time SMP constistency test. The 341 * timecounters use MFENCE for AMD CPUs, and LFENCE for others (Intel 342 * and VIA) when SSE2 is present, and nothing on older machines which 343 * also do not issue RDTSC prematurely. There, testing for SSE2 and 344 * vendor is too cumbersome, and we learn about TSC presence from CPUID. 345 * 346 * Do not use do_cpuid(), since we do not need CPUID results, which 347 * have to be written into memory with do_cpuid(). 348 */ 349 #define TSC_READ(x) \ 350 static void \ 351 tsc_read_##x(void *arg) \ 352 { \ 353 uint64_t *tsc = arg; \ 354 u_int cpu = PCPU_GET(cpuid); \ 355 \ 356 __asm __volatile("cpuid" : : : "eax", "ebx", "ecx", "edx"); \ 357 tsc[cpu * 3 + x] = rdtsc(); \ 358 } 359 TSC_READ(0) 360 TSC_READ(1) 361 TSC_READ(2) 362 #undef TSC_READ 363 364 #define N 1000 365 366 static void 367 comp_smp_tsc(void *arg) 368 { 369 uint64_t *tsc; 370 int64_t d1, d2; 371 u_int cpu = PCPU_GET(cpuid); 372 u_int i, j, size; 373 374 size = (mp_maxid + 1) * 3; 375 for (i = 0, tsc = arg; i < N; i++, tsc += size) 376 CPU_FOREACH(j) { 377 if (j == cpu) 378 continue; 379 d1 = tsc[cpu * 3 + 1] - tsc[j * 3]; 380 d2 = tsc[cpu * 3 + 2] - tsc[j * 3 + 1]; 381 if (d1 <= 0 || d2 <= 0) { 382 smp_tsc = 0; 383 return; 384 } 385 } 386 } 387 388 static void 389 adj_smp_tsc(void *arg) 390 { 391 uint64_t *tsc; 392 int64_t d, min, max; 393 u_int cpu = PCPU_GET(cpuid); 394 u_int first, i, size; 395 396 first = CPU_FIRST(); 397 if (cpu == first) 398 return; 399 min = INT64_MIN; 400 max = INT64_MAX; 401 size = (mp_maxid + 1) * 3; 402 for (i = 0, tsc = arg; i < N; i++, tsc += size) { 403 d = tsc[first * 3] - tsc[cpu * 3 + 1]; 404 if (d > min) 405 min = d; 406 d = tsc[first * 3 + 1] - tsc[cpu * 3 + 2]; 407 if (d > min) 408 min = d; 409 d = tsc[first * 3 + 1] - tsc[cpu * 3]; 410 if (d < max) 411 max = d; 412 d = tsc[first * 3 + 2] - tsc[cpu * 3 + 1]; 413 if (d < max) 414 max = d; 415 } 416 if (min > max) 417 return; 418 d = min / 2 + max / 2; 419 __asm __volatile ( 420 "movl $0x10, %%ecx\n\t" 421 "rdmsr\n\t" 422 "addl %%edi, %%eax\n\t" 423 "adcl %%esi, %%edx\n\t" 424 "wrmsr\n" 425 : /* No output */ 426 : "D" ((uint32_t)d), "S" ((uint32_t)(d >> 32)) 427 : "ax", "cx", "dx", "cc" 428 ); 429 } 430 431 static int 432 test_tsc(void) 433 { 434 uint64_t *data, *tsc; 435 u_int i, size, adj; 436 437 if ((!smp_tsc && !tsc_is_invariant) || vm_guest) 438 return (-100); 439 size = (mp_maxid + 1) * 3; 440 data = malloc(sizeof(*data) * size * N, M_TEMP, M_WAITOK); 441 adj = 0; 442 retry: 443 for (i = 0, tsc = data; i < N; i++, tsc += size) 444 smp_rendezvous(tsc_read_0, tsc_read_1, tsc_read_2, tsc); 445 smp_tsc = 1; /* XXX */ 446 smp_rendezvous(smp_no_rendevous_barrier, comp_smp_tsc, 447 smp_no_rendevous_barrier, data); 448 if (!smp_tsc && adj < smp_tsc_adjust) { 449 adj++; 450 smp_rendezvous(smp_no_rendevous_barrier, adj_smp_tsc, 451 smp_no_rendevous_barrier, data); 452 goto retry; 453 } 454 free(data, M_TEMP); 455 if (bootverbose) 456 printf("SMP: %sed TSC synchronization test%s\n", 457 smp_tsc ? "pass" : "fail", 458 adj > 0 ? " after adjustment" : ""); 459 if (smp_tsc && tsc_is_invariant) { 460 switch (cpu_vendor_id) { 461 case CPU_VENDOR_AMD: 462 /* 463 * Starting with Family 15h processors, TSC clock 464 * source is in the north bridge. Check whether 465 * we have a single-socket/multi-core platform. 466 * XXX Need more work for complex cases. 467 */ 468 if (CPUID_TO_FAMILY(cpu_id) < 0x15 || 469 (amd_feature2 & AMDID2_CMP) == 0 || 470 smp_cpus > (cpu_procinfo2 & AMDID_CMP_CORES) + 1) 471 break; 472 return (1000); 473 case CPU_VENDOR_INTEL: 474 /* 475 * XXX Assume Intel platforms have synchronized TSCs. 476 */ 477 return (1000); 478 } 479 return (800); 480 } 481 return (-100); 482 } 483 484 #undef N 485 486 #else 487 488 /* 489 * The function is not called, it is provided to avoid linking failure 490 * on uniprocessor kernel. 491 */ 492 static int 493 test_tsc(void) 494 { 495 496 return (0); 497 } 498 499 #endif /* SMP */ 500 501 static void 502 init_TSC_tc(void) 503 { 504 uint64_t max_freq; 505 int shift; 506 507 if ((cpu_feature & CPUID_TSC) == 0 || tsc_disabled) 508 return; 509 510 /* 511 * Limit timecounter frequency to fit in an int and prevent it from 512 * overflowing too fast. 513 */ 514 max_freq = UINT_MAX; 515 516 /* 517 * We can not use the TSC if we support APM. Precise timekeeping 518 * on an APM'ed machine is at best a fools pursuit, since 519 * any and all of the time spent in various SMM code can't 520 * be reliably accounted for. Reading the RTC is your only 521 * source of reliable time info. The i8254 loses too, of course, 522 * but we need to have some kind of time... 523 * We don't know at this point whether APM is going to be used 524 * or not, nor when it might be activated. Play it safe. 525 */ 526 if (power_pm_get_type() == POWER_PM_TYPE_APM) { 527 tsc_timecounter.tc_quality = -1000; 528 if (bootverbose) 529 printf("TSC timecounter disabled: APM enabled.\n"); 530 goto init; 531 } 532 533 /* 534 * Intel CPUs without a C-state invariant TSC can stop the TSC 535 * in either C2 or C3. Disable use of C2 and C3 while using 536 * the TSC as the timecounter. The timecounter can be changed 537 * to enable C2 and C3. 538 * 539 * Note that the TSC is used as the cputicker for computing 540 * thread runtime regardless of the timecounter setting, so 541 * using an alternate timecounter and enabling C2 or C3 can 542 * result incorrect runtimes for kernel idle threads (but not 543 * for any non-idle threads). 544 */ 545 if (cpu_deepest_sleep >= 2 && cpu_vendor_id == CPU_VENDOR_INTEL && 546 (amd_pminfo & AMDPM_TSC_INVARIANT) == 0) { 547 tsc_timecounter.tc_flags |= TC_FLAGS_C2STOP; 548 if (bootverbose) 549 printf("TSC timecounter disables C2 and C3.\n"); 550 } 551 552 /* 553 * We can not use the TSC in SMP mode unless the TSCs on all CPUs 554 * are synchronized. If the user is sure that the system has 555 * synchronized TSCs, set kern.timecounter.smp_tsc tunable to a 556 * non-zero value. The TSC seems unreliable in virtualized SMP 557 * environments, so it is set to a negative quality in those cases. 558 */ 559 if (mp_ncpus > 1) 560 tsc_timecounter.tc_quality = test_tsc(); 561 else if (tsc_is_invariant) 562 tsc_timecounter.tc_quality = 1000; 563 max_freq >>= tsc_shift; 564 565 init: 566 for (shift = 0; shift <= 31 && (tsc_freq >> shift) > max_freq; shift++) 567 ; 568 if ((cpu_feature & CPUID_SSE2) != 0 && mp_ncpus > 1) { 569 if (cpu_vendor_id == CPU_VENDOR_AMD) { 570 tsc_timecounter.tc_get_timecount = shift > 0 ? 571 tsc_get_timecount_low_mfence : 572 tsc_get_timecount_mfence; 573 } else { 574 tsc_timecounter.tc_get_timecount = shift > 0 ? 575 tsc_get_timecount_low_lfence : 576 tsc_get_timecount_lfence; 577 } 578 } else { 579 tsc_timecounter.tc_get_timecount = shift > 0 ? 580 tsc_get_timecount_low : tsc_get_timecount; 581 } 582 if (shift > 0) { 583 tsc_timecounter.tc_name = "TSC-low"; 584 if (bootverbose) 585 printf("TSC timecounter discards lower %d bit(s)\n", 586 shift); 587 } 588 if (tsc_freq != 0) { 589 tsc_timecounter.tc_frequency = tsc_freq >> shift; 590 tsc_timecounter.tc_priv = (void *)(intptr_t)shift; 591 tc_init(&tsc_timecounter); 592 } 593 } 594 SYSINIT(tsc_tc, SI_SUB_SMP, SI_ORDER_ANY, init_TSC_tc, NULL); 595 596 /* 597 * When cpufreq levels change, find out about the (new) max frequency. We 598 * use this to update CPU accounting in case it got a lower estimate at boot. 599 */ 600 static void 601 tsc_levels_changed(void *arg, int unit) 602 { 603 device_t cf_dev; 604 struct cf_level *levels; 605 int count, error; 606 uint64_t max_freq; 607 608 /* Only use values from the first CPU, assuming all are equal. */ 609 if (unit != 0) 610 return; 611 612 /* Find the appropriate cpufreq device instance. */ 613 cf_dev = devclass_get_device(devclass_find("cpufreq"), unit); 614 if (cf_dev == NULL) { 615 printf("tsc_levels_changed() called but no cpufreq device?\n"); 616 return; 617 } 618 619 /* Get settings from the device and find the max frequency. */ 620 count = 64; 621 levels = malloc(count * sizeof(*levels), M_TEMP, M_NOWAIT); 622 if (levels == NULL) 623 return; 624 error = CPUFREQ_LEVELS(cf_dev, levels, &count); 625 if (error == 0 && count != 0) { 626 max_freq = (uint64_t)levels[0].total_set.freq * 1000000; 627 set_cputicker(rdtsc, max_freq, 1); 628 } else 629 printf("tsc_levels_changed: no max freq found\n"); 630 free(levels, M_TEMP); 631 } 632 633 /* 634 * If the TSC timecounter is in use, veto the pending change. It may be 635 * possible in the future to handle a dynamically-changing timecounter rate. 636 */ 637 static void 638 tsc_freq_changing(void *arg, const struct cf_level *level, int *status) 639 { 640 641 if (*status != 0 || timecounter != &tsc_timecounter) 642 return; 643 644 printf("timecounter TSC must not be in use when " 645 "changing frequencies; change denied\n"); 646 *status = EBUSY; 647 } 648 649 /* Update TSC freq with the value indicated by the caller. */ 650 static void 651 tsc_freq_changed(void *arg, const struct cf_level *level, int status) 652 { 653 uint64_t freq; 654 655 /* If there was an error during the transition, don't do anything. */ 656 if (tsc_disabled || status != 0) 657 return; 658 659 /* Total setting for this level gives the new frequency in MHz. */ 660 freq = (uint64_t)level->total_set.freq * 1000000; 661 atomic_store_rel_64(&tsc_freq, freq); 662 tsc_timecounter.tc_frequency = 663 freq >> (int)(intptr_t)tsc_timecounter.tc_priv; 664 } 665 666 static int 667 sysctl_machdep_tsc_freq(SYSCTL_HANDLER_ARGS) 668 { 669 int error; 670 uint64_t freq; 671 672 freq = atomic_load_acq_64(&tsc_freq); 673 if (freq == 0) 674 return (EOPNOTSUPP); 675 error = sysctl_handle_64(oidp, &freq, 0, req); 676 if (error == 0 && req->newptr != NULL) { 677 atomic_store_rel_64(&tsc_freq, freq); 678 atomic_store_rel_64(&tsc_timecounter.tc_frequency, 679 freq >> (int)(intptr_t)tsc_timecounter.tc_priv); 680 } 681 return (error); 682 } 683 684 SYSCTL_PROC(_machdep, OID_AUTO, tsc_freq, CTLTYPE_U64 | CTLFLAG_RW, 685 0, 0, sysctl_machdep_tsc_freq, "QU", "Time Stamp Counter frequency"); 686 687 static u_int 688 tsc_get_timecount(struct timecounter *tc __unused) 689 { 690 691 return (rdtsc32()); 692 } 693 694 static inline u_int 695 tsc_get_timecount_low(struct timecounter *tc) 696 { 697 uint32_t rv; 698 699 __asm __volatile("rdtsc; shrd %%cl, %%edx, %0" 700 : "=a" (rv) : "c" ((int)(intptr_t)tc->tc_priv) : "edx"); 701 return (rv); 702 } 703 704 static u_int 705 tsc_get_timecount_lfence(struct timecounter *tc __unused) 706 { 707 708 lfence(); 709 return (rdtsc32()); 710 } 711 712 static u_int 713 tsc_get_timecount_low_lfence(struct timecounter *tc) 714 { 715 716 lfence(); 717 return (tsc_get_timecount_low(tc)); 718 } 719 720 static u_int 721 tsc_get_timecount_mfence(struct timecounter *tc __unused) 722 { 723 724 mfence(); 725 return (rdtsc32()); 726 } 727 728 static u_int 729 tsc_get_timecount_low_mfence(struct timecounter *tc) 730 { 731 732 mfence(); 733 return (tsc_get_timecount_low(tc)); 734 } 735 736 static uint32_t 737 x86_tsc_vdso_timehands(struct vdso_timehands *vdso_th, struct timecounter *tc) 738 { 739 740 vdso_th->th_algo = VDSO_TH_ALGO_X86_TSC; 741 vdso_th->th_x86_shift = (int)(intptr_t)tc->tc_priv; 742 vdso_th->th_x86_hpet_idx = 0xffffffff; 743 bzero(vdso_th->th_res, sizeof(vdso_th->th_res)); 744 return (1); 745 } 746 747 #ifdef COMPAT_FREEBSD32 748 static uint32_t 749 x86_tsc_vdso_timehands32(struct vdso_timehands32 *vdso_th32, 750 struct timecounter *tc) 751 { 752 753 vdso_th32->th_algo = VDSO_TH_ALGO_X86_TSC; 754 vdso_th32->th_x86_shift = (int)(intptr_t)tc->tc_priv; 755 vdso_th32->th_x86_hpet_idx = 0xffffffff; 756 bzero(vdso_th32->th_res, sizeof(vdso_th32->th_res)); 757 return (1); 758 } 759 #endif 760