Chameleon

Chameleon Svn Source Tree

Root/branches/Bungo/i386/libsaio/cpu.c

1/*␊
2 * Copyright 2008 Islam Ahmed Zaid. All rights reserved. <azismed@gmail.com>␊
3 * AsereBLN: 2009: cleanup and bugfix␊
4 */␊
5␊
6#include "libsaio.h"␊
7#include "platform.h"␊
8#include "cpu.h"␊
9#include "bootstruct.h"␊
10#include "boot.h"␊
11␊
12#ifndef DEBUG_CPU␊
13#define DEBUG_CPU 0␊
14#endif␊
15␊
16#if DEBUG_CPU␊
17#define DBG(x...)␉␉printf(x)␊
18#else␊
19#define DBG(x...)␉␉msglog(x)␊
20#endif␊
21␊
22/*␊
23 * timeRDTSC()␊
24 * This routine sets up PIT counter 2 to count down 1/20 of a second.␊
25 * It pauses until the value is latched in the counter␊
26 * and then reads the time stamp counter to return to the caller.␊
27 */␊
28uint64_t timeRDTSC(void)␊
29{␊
30␉int␉␉attempts = 0;␊
31␉uint64_t latchTime;␊
32␉uint64_t␉saveTime,intermediate;␊
33␉unsigned int timerValue, lastValue;␊
34␉//boolean_t␉int_enabled;␊
35␉/*␊
36␉ * Table of correction factors to account for␊
37␉ *␉ - timer counter quantization errors, and␊
38␉ *␉ - undercounts 0..5␊
39␉ */␊
40#define SAMPLE_CLKS_EXACT␉(((double) CLKNUM) / 20.0)␊
41#define SAMPLE_CLKS_INT␉␉((int) CLKNUM / 20)␊
42#define SAMPLE_NSECS␉␉(2000000000LL)␊
43#define SAMPLE_MULTIPLIER␉(((double)SAMPLE_NSECS)*SAMPLE_CLKS_EXACT)␊
44#define ROUND64(x)␉␉((uint64_t)((x) + 0.5))␊
45␉uint64_t␉scale[6] = {␊
46␉␉ROUND64(SAMPLE_MULTIPLIER/(double)(SAMPLE_CLKS_INT-0)), ␊
47␉␉ROUND64(SAMPLE_MULTIPLIER/(double)(SAMPLE_CLKS_INT-1)), ␊
48␉␉ROUND64(SAMPLE_MULTIPLIER/(double)(SAMPLE_CLKS_INT-2)), ␊
49␉␉ROUND64(SAMPLE_MULTIPLIER/(double)(SAMPLE_CLKS_INT-3)), ␊
50␉␉ROUND64(SAMPLE_MULTIPLIER/(double)(SAMPLE_CLKS_INT-4)), ␊
51␉␉ROUND64(SAMPLE_MULTIPLIER/(double)(SAMPLE_CLKS_INT-5))␊
52␉};␊
53␊
54␉//int_enabled = ml_set_interrupts_enabled(FALSE);␊
55␊
56restart:␊
57␉if (attempts >= 9) // increase to up to 9 attempts.␊
58␉{␊
59␉ // This will flash-reboot. TODO: Use tscPanic instead.␊
60␉␉printf("Timestamp counter calibation failed with %d attempts\n", attempts);␊
61␉}␊
62␉attempts++;␊
63␉enable_PIT2();␉␉// turn on PIT2␊
64␉set_PIT2(0);␉␉// reset timer 2 to be zero␊
65␉latchTime = rdtsc64();␉// get the time stamp to time ␊
66␉latchTime = get_PIT2(&timerValue) - latchTime; // time how long this takes␊
67␉set_PIT2(SAMPLE_CLKS_INT);␉// set up the timer for (almost) 1/20th a second␊
68␉saveTime = rdtsc64();␉// now time how long a 20th a second is...␊
69␉get_PIT2(&lastValue);␊
70␉get_PIT2(&lastValue);␉// read twice, first value may be unreliable␊
71␉do {␊
72␉␉intermediate = get_PIT2(&timerValue);␊
73␉␉if (timerValue > lastValue)␊
74␉␉{␊
75␉␉␉// Timer wrapped␊
76␉␉␉set_PIT2(0);␊
77␉␉␉disable_PIT2();␊
78␉␉␉goto restart;␊
79␉␉}␊
80␉␉lastValue = timerValue;␊
81␉} while (timerValue > 5);␊
82␉printf("timerValue␉ %d\n",timerValue);␊
83␉printf("intermediate 0x%016llx\n",intermediate);␊
84␉printf("saveTime␉ 0x%016llx\n",saveTime);␊
85 ␊
86␉intermediate -= saveTime;␉␉// raw count for about 1/20 second␊
87␉intermediate *= scale[timerValue];␉// rescale measured time spent␊
88␉intermediate /= SAMPLE_NSECS;␉// so its exactly 1/20 a second␊
89␉intermediate += latchTime;␉␉// add on our save fudge␊
90 ␊
91␉set_PIT2(0);␉␉␉// reset timer 2 to be zero␊
92␉disable_PIT2();␉␉␉// turn off PIT 2␊
93␉␊
94␉//ml_set_interrupts_enabled(int_enabled);␊
95␉return intermediate;␊
96}␊
97␊
98/*␊
99 * DFE: Measures the TSC frequency in Hz (64-bit) using the ACPI PM timer␊
100 */␊
101static uint64_t measure_tsc_frequency(void)␊
102{␊
103␉uint64_t tscStart;␊
104␉uint64_t tscEnd;␊
105␉uint64_t tscDelta = 0xffffffffffffffffULL;␊
106␉unsigned long pollCount;␊
107␉uint64_t retval = 0;␊
108␉int i;␊
109␊
110␉/* Time how many TSC ticks elapse in 30 msec using the 8254 PIT␊
111␉ * counter 2. We run this loop 3 times to make sure the cache␊
112␉ * is hot and we take the minimum delta from all of the runs.␊
113␉ * That is to say that we're biased towards measuring the minimum␊
114␉ * number of TSC ticks that occur while waiting for the timer to␊
115␉ * expire. That theoretically helps avoid inconsistencies when␊
116␉ * running under a VM if the TSC is not virtualized and the host␊
117␉ * steals time.␉ The TSC is normally virtualized for VMware.␊
118␉ */␊
119␉for(i = 0; i < 10; ++i)␊
120␉{␊
121␉␉enable_PIT2();␊
122␉␉set_PIT2_mode0(CALIBRATE_LATCH);␊
123␉␉tscStart = rdtsc64();␊
124␉␉pollCount = poll_PIT2_gate();␊
125␉␉tscEnd = rdtsc64();␊
126␉␉/* The poll loop must have run at least a few times for accuracy */␊
127␉␉if (pollCount <= 1) {␊
128␉␉␉continue;␊
129␉␉}␊
130␉␉/* The TSC must increment at LEAST once every millisecond.␊
131␉␉ * We should have waited exactly 30 msec so the TSC delta should␊
132␉␉ * be >= 30. Anything less and the processor is way too slow.␊
133␉␉ */␊
134␉␉if ((tscEnd - tscStart) <= CALIBRATE_TIME_MSEC) {␊
135␉␉␉continue;␊
136␉␉}␊
137␉␉// tscDelta = MIN(tscDelta, (tscEnd - tscStart))␊
138␉␉if ( (tscEnd - tscStart) < tscDelta ) {␊
139␉␉␉tscDelta = tscEnd - tscStart;␊
140␉␉}␊
141␉}␊
142␉/* tscDelta is now the least number of TSC ticks the processor made in␊
143␉ * a timespan of 0.03 s (e.g. 30 milliseconds)␊
144␉ * Linux thus divides by 30 which gives the answer in kiloHertz because␊
145␉ * 1 / ms = kHz. But we're xnu and most of the rest of the code uses␊
146␉ * Hz so we need to convert our milliseconds to seconds. Since we're␊
147␉ * dividing by the milliseconds, we simply multiply by 1000.␊
148␉ */␊
149␊
150␉/* Unlike linux, we're not limited to 32-bit, but we do need to take care␊
151␉ * that we're going to multiply by 1000 first so we do need at least some␊
152␉ * arithmetic headroom. For now, 32-bit should be enough.␊
153␉ * Also unlike Linux, our compiler can do 64-bit integer arithmetic.␊
154␉ */␊
155␉if (tscDelta > (1ULL<<32)) {␊
156␉␉retval = 0;␊
157␉} else {␊
158␉␉retval = tscDelta * 1000 / 30;␊
159␉}␊
160␉disable_PIT2();␊
161␉return retval;␊
162}␊
163␊
164/*␊
165 * Original comment/code:␊
166 * "DFE: Measures the Max Performance Frequency in Hz (64-bit)"␊
167 *␊
168 * Measures the Actual Performance Frequency in Hz (64-bit)␊
169 * (just a naming change, mperf --> aperf )␊
170 */␊
171static uint64_t measure_aperf_frequency(void)␊
172{␊
173␉uint64_t aperfStart;␊
174␉uint64_t aperfEnd;␊
175␉uint64_t aperfDelta = 0xffffffffffffffffULL;␊
176␉unsigned long pollCount;␊
177␉uint64_t retval = 0;␊
178␉int i;␊
179␊
180␉/* Time how many APERF ticks elapse in 30 msec using the 8254 PIT␊
181␉ * counter 2. We run this loop 3 times to make sure the cache␊
182␉ * is hot and we take the minimum delta from all of the runs.␊
183␉ * That is to say that we're biased towards measuring the minimum␊
184␉ * number of APERF ticks that occur while waiting for the timer to␊
185␉ * expire.␊
186␉ */␊
187␉for(i = 0; i < 10; ++i)␊
188␉{␊
189␉␉enable_PIT2();␊
190␉␉set_PIT2_mode0(CALIBRATE_LATCH);␊
191␉␉aperfStart = rdmsr64(MSR_AMD_APERF);␊
192␉␉pollCount = poll_PIT2_gate();␊
193␉␉aperfEnd = rdmsr64(MSR_AMD_APERF);␊
194␉␉/* The poll loop must have run at least a few times for accuracy */␊
195␉␉if (pollCount <= 1) {␊
196␉␉␉continue;␊
197␉␉}␊
198␉␉/* The TSC must increment at LEAST once every millisecond.␊
199␉␉ * We should have waited exactly 30 msec so the APERF delta should␊
200␉␉ * be >= 30. Anything less and the processor is way too slow.␊
201␉␉ */␊
202␉␉if ((aperfEnd - aperfStart) <= CALIBRATE_TIME_MSEC) {␊
203␉␉␉continue;␊
204␉␉}␊
205␉␉// tscDelta = MIN(tscDelta, (tscEnd - tscStart))␊
206␉␉if ( (aperfEnd - aperfStart) < aperfDelta ) {␊
207␉␉␉aperfDelta = aperfEnd - aperfStart;␊
208␉␉}␊
209␉}␊
210␉/* mperfDelta is now the least number of MPERF ticks the processor made in␊
211␉ * a timespan of 0.03 s (e.g. 30 milliseconds)␊
212␉ */␊
213␊
214␉if (aperfDelta > (1ULL<<32)) {␊
215␉␉retval = 0;␊
216␉} else {␊
217␉␉retval = aperfDelta * 1000 / 30;␊
218␉}␊
219␉disable_PIT2();␊
220␉return retval;␊
221}␊
222␊
223/*␊
224 * Calculates the FSB and CPU frequencies using specific MSRs for each CPU␊
225 * - multi. is read from a specific MSR. In the case of Intel, there is:␊
226 *␉ a max multi. (used to calculate the FSB freq.),␊
227 *␉ and a current multi. (used to calculate the CPU freq.)␊
228 * - fsbFrequency = tscFrequency / multi␊
229 * - cpuFrequency = fsbFrequency * multi␊
230 */␊
231void scan_cpu(PlatformInfo_t *p)␊
232{␊
233␉uint64_t␉tscFrequency = 0;␊
234␉uint64_t␉fsbFrequency = 0;␊
235␉uint64_t␉cpuFrequency = 0;␊
236␉uint64_t␉msr = 0;␊
237␉uint64_t␉flex_ratio = 0;␊
238␉uint32_t␉max_ratio = 0;␊
239␉uint32_t␉min_ratio = 0;␊
240␉uint8_t␉␉bus_ratio_max = 0;␊
241␉uint8_t␉␉currdiv = 0;␊
242␉uint8_t␉␉currcoef = 0;␊
243␉uint8_t␉␉maxdiv = 0;␊
244␉uint8_t␉␉maxcoef = 0;␊
245␉const char␉*newratio;␊
246␉int␉␉len = 0;␊
247␉int␉␉myfsb = 0;␊
248␉uint8_t␉␉bus_ratio_min = 0;␊
249␉uint32_t␉reg[4];␊
250␉char␉␉str[128];␊
251␊
252␉/* get cpuid values */␊
253␉do_cpuid(0x00000000, p->CPU.CPUID[CPUID_0]);␊
254␉do_cpuid(0x00000001, p->CPU.CPUID[CPUID_1]);␊
255␊
256␉do_cpuid(0x00000002, p->CPU.CPUID[CPUID_2]);␊
257␉do_cpuid(0x00000003, p->CPU.CPUID[CPUID_3]);␊
258␉do_cpuid2(0x00000004, 0, p->CPU.CPUID[CPUID_4]);␊
259␊
260␉do_cpuid(0x80000000, p->CPU.CPUID[CPUID_80]);␊
261␉if (p->CPU.CPUID[CPUID_0][0] >= 0x5) {␊
262␉␉do_cpuid(5, p->CPU.CPUID[CPUID_5]);␊
263␉}␊
264␉if (p->CPU.CPUID[CPUID_0][0] >= 6) {␊
265␉␉do_cpuid(6, p->CPU.CPUID[CPUID_6]);␊
266␉}␊
267␉if ((p->CPU.CPUID[CPUID_80][0] & 0x0000000f) >= 8) {␊
268␉␉do_cpuid(0x80000008, p->CPU.CPUID[CPUID_88]);␊
269␉␉do_cpuid(0x80000001, p->CPU.CPUID[CPUID_81]);␊
270␉} else if ((p->CPU.CPUID[CPUID_80][0] & 0x0000000f) >= 1) {␊
271␉␉do_cpuid(0x80000001, p->CPU.CPUID[CPUID_81]);␊
272␉}␊
273␊
274// #if DEBUG_CPU␊
275␉{␊
276␉␉int␉␉i;␊
277␉␉DBG("CPUID Raw Values:\n");␊
278␉␉for (i = 0; i < CPUID_MAX; i++) {␊
279␉␉␉DBG("%02d: %08x-%08x-%08x-%08x\n", i,␊
280␉␉␉␉ p->CPU.CPUID[i][0], p->CPU.CPUID[i][1],␊
281␉␉␉␉ p->CPU.CPUID[i][2], p->CPU.CPUID[i][3]);␊
282␉␉}␊
283␉}␊
284// #endif␊
285␊
286/*␊
287 EAX (Intel):␊
288 31 28 27 20 19 16 1514 1312 11 8 7 4 3 0␊
289 +--------+----------------+--------+----+----+--------+--------+--------+␊
290 |########|Extended family |Extmodel|####|type|familyid| model |stepping|␊
291 +--------+----------------+--------+----+----+--------+--------+--------+␊
292␊
293 EAX (AMD):␊
294 31 28 27 20 19 16 1514 1312 11 8 7 4 3 0␊
295 +--------+----------------+--------+----+----+--------+--------+--------+␊
296 |########|Extended family |Extmodel|####|####|familyid| model |stepping|␊
297 +--------+----------------+--------+----+----+--------+--------+--------+␊
298*/␊
299␊
300␉p->CPU.Vendor␉␉= p->CPU.CPUID[CPUID_0][1];␊
301␉p->CPU.Signature␉= p->CPU.CPUID[CPUID_1][0];␊
302␉p->CPU.Stepping␉␉= (uint8_t)bitfield(p->CPU.CPUID[CPUID_1][0], 3, 0);␉// stepping = cpu_feat_eax & 0xF;␊
303␉p->CPU.Model␉␉= (uint8_t)bitfield(p->CPU.CPUID[CPUID_1][0], 7, 4);␉// model = (cpu_feat_eax >> 4) & 0xF;␊
304␉p->CPU.Family␉␉= (uint8_t)bitfield(p->CPU.CPUID[CPUID_1][0], 11, 8);␉// family = (cpu_feat_eax >> 8) & 0xF;␊
305␉//p->CPU.Type␉␉= (uint8_t)bitfield(p->CPU.CPUID[CPUID_1][0], 13, 12);␉// type = (cpu_feat_eax >> 12) & 0x3;␊
306␉p->CPU.ExtModel␉␉= (uint8_t)bitfield(p->CPU.CPUID[CPUID_1][0], 19, 16);␉// ext_model = (cpu_feat_eax >> 16) & 0xF;␊
307␉p->CPU.ExtFamily␉= (uint8_t)bitfield(p->CPU.CPUID[CPUID_1][0], 27, 20);␉// ext_family = (cpu_feat_eax >> 20) & 0xFF;␊
308␊
309␉p->CPU.Model += (p->CPU.ExtModel << 4);␊
310␊
311␉if (p->CPU.Vendor == CPUID_VENDOR_INTEL)␊
312␉{␊
313␉␉/*␊
314␉␉ * Find the number of enabled cores and threads␊
315␉␉ * (which determines whether SMT/Hyperthreading is active).␊
316␉␉ */␊
317␉␉switch (p->CPU.Model)␊
318␉␉{␊
319␉␉␉case CPU_MODEL_NEHALEM:␊
320␉␉␉case CPU_MODEL_FIELDS:␊
321␉␉␉case CPU_MODEL_DALES:␊
322␉␉␉case CPU_MODEL_NEHALEM_EX:␊
323␉␉␉case CPU_MODEL_JAKETOWN:␊
324␉␉␉case CPU_MODEL_SANDYBRIDGE:␊
325␉␉␉case CPU_MODEL_IVYBRIDGE:␊
326␉␉␉case CPU_MODEL_HASWELL:␊
327␉␉␉case CPU_MODEL_HASWELL_SVR:␊
328␉␉␉//case CPU_MODEL_HASWELL_H:␊
329␉␉␉case CPU_MODEL_HASWELL_ULT:␊
330␉␉␉case CPU_MODEL_CRYSTALWELL:␊
331␉␉␉␉msr = rdmsr64(MSR_CORE_THREAD_COUNT);␊
332␉␉␉␉p->CPU.NoCores␉␉= (uint8_t)bitfield((uint32_t)msr, 31, 16);␊
333␉␉␉␉p->CPU.NoThreads␉= (uint8_t)bitfield((uint32_t)msr, 15, 0);␊
334␉␉␉␉break;␊
335␊
336␉␉␉case CPU_MODEL_DALES_32NM:␊
337␉␉␉case CPU_MODEL_WESTMERE:␊
338␉␉␉case CPU_MODEL_WESTMERE_EX:␊
339␉␉␉␉msr = rdmsr64(MSR_CORE_THREAD_COUNT);␊
340␉␉␉␉p->CPU.NoCores␉␉= (uint8_t)bitfield((uint32_t)msr, 19, 16);␊
341␉␉␉␉p->CPU.NoThreads␉= (uint8_t)bitfield((uint32_t)msr, 15, 0);␊
342␉␉␉␉break;␊
343␊
344␉␉␉default:␊
345␉␉␉␉p->CPU.NoCores = 0;␊
346␉␉␉␉break;␊
347␉␉} // end switch␊
348␉}␊
349␊
350␉if (p->CPU.NoCores == 0)␊
351␉{␊
352␉␉p->CPU.NoCores␉␉= (uint8_t)(p->CPU.CoresPerPackage & 0xff);␊
353␉␉p->CPU.NoThreads␉= (uint8_t)(p->CPU.LogicalPerPackage & 0xff);␊
354␉}␊
355␊
356␉/* get BrandString (if supported) */␊
357␉/* Copyright: from Apple's XNU cpuid.c */␊
358␉if (p->CPU.CPUID[CPUID_80][0] > 0x80000004)␊
359␉{␊
360␉␉char␉ *s;␊
361␉␉bzero(str, 128);␊
362␉␉/*␊
363␉␉ * The BrandString 48 bytes (max), guaranteed to␊
364␉␉ * be NULL terminated.␊
365␉␉ */␊
366␉␉do_cpuid(0x80000002, reg);␊
367␉␉memcpy(&str[0], (char *)reg, 16);␊
368␉␉do_cpuid(0x80000003, reg);␊
369␉␉memcpy(&str[16], (char *)reg, 16);␊
370␉␉do_cpuid(0x80000004, reg);␊
371␉␉memcpy(&str[32], (char *)reg, 16);␊
372␉␉for (s = str; *s != '\0'; s++)␊
373␉␉{␊
374␉␉␉if (*s != ' ')␊
375␉␉␉{␊
376␉␉␉␉break;␊
377␉␉␉}␊
378␉␉}␊
379␉␉strlcpy(p->CPU.BrandString, s, 48);␊
380␊
381␉␉if (!strncmp(p->CPU.BrandString, CPU_STRING_UNKNOWN, MIN(sizeof(p->CPU.BrandString), strlen(CPU_STRING_UNKNOWN) + 1)))␊
382␉␉{␊
383␉␉␉/*␊
384␉␉␉ * This string means we have a firmware-programmable brand string,␊
385␉␉␉ * and the firmware couldn't figure out what sort of CPU we have.␊
386␉␉␉ */␊
387␉␉␉p->CPU.BrandString[0] = '\0';␊
388␉␉}␊
389␉␉p->CPU.BrandString[47] = '\0';␊
390//␉␉DBG("Brandstring = %s\n", p->CPU.BrandString);␊
391␉}␊
392␊
393␉//workaround for N270. I don't know why it detected wrong␊
394␉// MSR is *NOT* available on the Intel Atom CPU␊
395␉if ((p->CPU.Model == CPU_MODEL_ATOM) && (strstr(p->CPU.BrandString, "270")))␊
396␉{␊
397␉␉p->CPU.NoCores␉␉= 1;␊
398␉␉p->CPU.NoThreads␉= 2;␊
399␉}␊
400␊
401␉if (p->CPU.Vendor == CPUID_VENDOR_AMD)␊
402␉{␊
403␉␉p->CPU.NoThreads␉= (uint8_t)bitfield(p->CPU.CPUID[CPUID_1][1], 23, 16);␊
404␉␉p->CPU.NoCores␉␉= (uint8_t)bitfield(p->CPU.CPUID[CPUID_88][2], 7, 0) + 1;␊
405␉}␊
406␊
407␉/* setup features */␊
408␉if ((bit(23) & p->CPU.CPUID[CPUID_1][3]) != 0) {␊
409␉␉p->CPU.Features |= CPU_FEATURE_MMX;␊
410␉}␊
411␉if ((bit(25) & p->CPU.CPUID[CPUID_1][3]) != 0) {␊
412␉␉p->CPU.Features |= CPU_FEATURE_SSE;␊
413␉}␊
414␉if ((bit(26) & p->CPU.CPUID[CPUID_1][3]) != 0) {␊
415␉␉p->CPU.Features |= CPU_FEATURE_SSE2;␊
416␉}␊
417␉if ((bit(0) & p->CPU.CPUID[CPUID_1][2]) != 0) {␊
418␉␉p->CPU.Features |= CPU_FEATURE_SSE3;␊
419␉}␊
420␉if ((bit(19) & p->CPU.CPUID[CPUID_1][2]) != 0) {␊
421␉␉p->CPU.Features |= CPU_FEATURE_SSE41;␊
422␉}␊
423␉if ((bit(20) & p->CPU.CPUID[CPUID_1][2]) != 0) {␊
424␉␉p->CPU.Features |= CPU_FEATURE_SSE42;␊
425␉}␊
426␉if ((bit(29) & p->CPU.CPUID[CPUID_81][3]) != 0) {␊
427␉␉p->CPU.Features |= CPU_FEATURE_EM64T;␊
428␉}␊
429␉if ((bit(5) & p->CPU.CPUID[CPUID_1][3]) != 0) {␊
430␉␉p->CPU.Features |= CPU_FEATURE_MSR;␊
431␉}␊
432␉//if ((bit(28) & p->CPU.CPUID[CPUID_1][3]) != 0) {␊
433␉if (p->CPU.NoThreads > p->CPU.NoCores) {␊
434␉␉p->CPU.Features |= CPU_FEATURE_HTT;␊
435␉}␊
436␊
437␉tscFrequency = measure_tsc_frequency();␊
438␉DBG("cpu freq classic = 0x%016llx\n", tscFrequency);␊
439␉/* if usual method failed */␊
440␉if ( tscFrequency < 1000 )␉//TEST␊
441␉{␊
442␉␉tscFrequency = timeRDTSC() * 20;//measure_tsc_frequency();␊
443␉␉// DBG("cpu freq timeRDTSC = 0x%016llx\n", tscFrequency);␊
444␉} else {␊
445␉␉// DBG("cpu freq timeRDTSC = 0x%016llxn", timeRDTSC() * 20);␊
446␉}␊
447␊
448␉fsbFrequency = 0;␊
449␉cpuFrequency = 0;␊
450␊
451␉if (p->CPU.Vendor == CPUID_VENDOR_INTEL && ((p->CPU.Family == 0x06 && p->CPU.Model >= 0x0c) || (p->CPU.Family == 0x0f && p->CPU.Model >= 0x03))) {␊
452␉␉int intelCPU = p->CPU.Model;␊
453␉␉if (p->CPU.Family == 0x06) {␊
454␉␉␉/* Nehalem CPU model */␊
455␉␉␉switch (p->CPU.Model) {␊
456␉␉␉␉case CPU_MODEL_NEHALEM:␊
457␉␉␉␉case CPU_MODEL_FIELDS:␊
458␉␉␉␉case CPU_MODEL_DALES:␊
459␉␉␉␉case CPU_MODEL_DALES_32NM:␊
460␉␉␉␉case CPU_MODEL_WESTMERE:␊
461␉␉␉␉case CPU_MODEL_NEHALEM_EX:␊
462␉␉␉␉case CPU_MODEL_WESTMERE_EX:␊
463/* --------------------------------------------------------- */␊
464␉␉␉␉case CPU_MODEL_SANDYBRIDGE:␊
465␉␉␉␉case CPU_MODEL_JAKETOWN:␊
466␉␉␉␉case CPU_MODEL_IVYBRIDGE_XEON:␊
467␉␉␉␉case CPU_MODEL_IVYBRIDGE:␊
468␉␉␉␉case CPU_MODEL_HASWELL:␊
469␉␉␉␉case CPU_MODEL_HASWELL_SVR:␊
470␊
471␉␉␉␉case CPU_MODEL_HASWELL_ULT:␊
472␉␉␉␉case CPU_MODEL_CRYSTALWELL:␊
473/* --------------------------------------------------------- */␊
474␉␉␉␉␉msr = rdmsr64(MSR_PLATFORM_INFO);␊
475␉␉␉␉␉DBG("msr(%d): platform_info %08x\n", __LINE__, bitfield(msr, 31, 0));␊
476␉␉␉␉␉bus_ratio_max = bitfield(msr, 15, 8);␊
477␉␉␉␉␉bus_ratio_min = bitfield(msr, 47, 40); //valv: not sure about this one (Remarq.1)␊
478␉␉␉␉␉msr = rdmsr64(MSR_FLEX_RATIO);␊
479␉␉␉␉␉DBG("msr(%d): flex_ratio %08x\n", __LINE__, bitfield(msr, 31, 0));␊
480␉␉␉␉␉if (bitfield(msr, 16, 16))␊
481␉␉␉␉␉{␊
482␉␉␉␉␉␉flex_ratio = bitfield(msr, 15, 8);␊
483␉␉␉␉␉␉/* bcc9: at least on the gigabyte h67ma-ud2h,␊
484␉␉␉␉␉␉ where the cpu multipler can't be changed to␊
485␉␉␉␉␉␉ allow overclocking, the flex_ratio msr has unexpected (to OSX)␊
486␉␉␉␉␉␉ contents.␉These contents cause mach_kernel to␊
487␉␉␉␉␉␉ fail to compute the bus ratio correctly, instead␊
488␉␉␉␉␉␉ causing the system to crash since tscGranularity␊
489␉␉␉␉␉␉ is inadvertently set to 0.␊
490␉␉␉␉␉␉ */␊
491␉␉␉␉␉␉if (flex_ratio == 0)␊
492␉␉␉␉␉␉{␊
493␉␉␉␉␉␉␉/* Clear bit 16 (evidently the presence bit) */␊
494␉␉␉␉␉␉␉wrmsr64(MSR_FLEX_RATIO, (msr & 0xFFFFFFFFFFFEFFFFULL));␊
495␉␉␉␉␉␉␉msr = rdmsr64(MSR_FLEX_RATIO);␊
496␉␉␉␉␉␉␉DBG("Unusable flex ratio detected. Patched MSR now %08x\n", bitfield(msr, 31, 0));␊
497␉␉␉␉␉␉}␊
498␉␉␉␉␉␉else␊
499␉␉␉␉␉␉{␊
500␉␉␉␉␉␉␉if (bus_ratio_max > flex_ratio)␊
501␉␉␉␉␉␉␉{␊
502␉␉␉␉␉␉␉␉bus_ratio_max = flex_ratio;␊
503␉␉␉␉␉␉␉}␊
504␉␉␉␉␉␉}␊
505␉␉␉␉␉}␊
506␊
507␉␉␉␉␉if (bus_ratio_max)␊
508␉␉␉␉␉{␊
509␉␉␉␉␉␉fsbFrequency = (tscFrequency / bus_ratio_max);␊
510␉␉␉␉␉}␊
511␊
512␉␉␉␉␉//valv: Turbo Ratio Limit␊
513␉␉␉␉␉if ((intelCPU != 0x2e) && (intelCPU != 0x2f))␊
514␉␉␉␉␉{␊
515␉␉␉␉␉␉msr = rdmsr64(MSR_TURBO_RATIO_LIMIT);␊
516␊
517␉␉␉␉␉␉cpuFrequency = bus_ratio_max * fsbFrequency;␊
518␉␉␉␉␉␉max_ratio = bus_ratio_max * 10;␊
519␉␉␉␉␉}␊
520␉␉␉␉␉else␊
521␉␉␉␉␉{␊
522␉␉␉␉␉␉cpuFrequency = tscFrequency;␊
523␉␉␉␉␉}␊
524␉␉␉␉␉if ((getValueForKey(kbusratio, &newratio, &len, &bootInfo->chameleonConfig)) && (len <= 4))␊
525␉␉␉␉␉{␊
526␉␉␉␉␉␉max_ratio = atoi(newratio);␊
527␉␉␉␉␉␉max_ratio = (max_ratio * 10);␊
528␉␉␉␉␉␉if (len >= 3)␊
529␉␉␉␉␉␉{␊
530␉␉␉␉␉␉␉max_ratio = (max_ratio + 5);␊
531␉␉␉␉␉␉}␊
532␊
533␉␉␉␉␉␉verbose("Bus-Ratio: min=%d, max=%s\n", bus_ratio_min, newratio);␊
534␊
535␉␉␉␉␉␉// extreme overclockers may love 320 ;)␊
536␉␉␉␉␉␉if ((max_ratio >= min_ratio) && (max_ratio <= 320))␊
537␉␉␉␉␉␉{␊
538␉␉␉␉␉␉␉cpuFrequency = (fsbFrequency * max_ratio) / 10;␊
539␉␉␉␉␉␉␉if (len >= 3)␊
540␉␉␉␉␉␉␉{␊
541␉␉␉␉␉␉␉␉maxdiv = 1;␊
542␉␉␉␉␉␉␉}␊
543␉␉␉␉␉␉␉else␊
544␉␉␉␉␉␉␉{␊
545␉␉␉␉␉␉␉␉maxdiv = 0;␊
546␉␉␉␉␉␉␉}␊
547␉␉␉␉␉␉}␊
548␉␉␉␉␉␉else␊
549␉␉␉␉␉␉{␊
550␉␉␉␉␉␉␉max_ratio = (bus_ratio_max * 10);␊
551␉␉␉␉␉␉}␊
552␉␉␉␉␉}␊
553␉␉␉␉␉//valv: to be uncommented if Remarq.1 didn't stick␊
554␉␉␉␉␉/*if (bus_ratio_max > 0) bus_ratio = flex_ratio;*/␊
555␉␉␉␉␉p->CPU.MaxRatio = max_ratio;␊
556␉␉␉␉␉p->CPU.MinRatio = min_ratio;␊
557␊
558␉␉␉␉myfsb = fsbFrequency / 1000000;␊
559␉␉␉␉verbose("Sticking with [BCLK: %dMhz, Bus-Ratio: %d]\n", myfsb, max_ratio/10); // Bungo: fixed wrong Bus-Ratio readout␊
560␉␉␉␉currcoef = bus_ratio_max;␊
561␊
562␉␉␉␉break;␊
563␊
564␉␉␉default:␊
565␉␉␉␉msr = rdmsr64(MSR_IA32_PERF_STATUS);␊
566␉␉␉␉DBG("msr(%d): ia32_perf_stat 0x%08x\n", __LINE__, bitfield(msr, 31, 0));␊
567␉␉␉␉currcoef = bitfield(msr, 12, 8); // Bungo: reverted to 2263 state because of wrong old CPUs freq. calculating␊
568␉␉␉␉/* Non-integer bus ratio for the max-multi*/␊
569␉␉␉␉maxdiv = bitfield(msr, 46, 46);␊
570␉␉␉␉/* Non-integer bus ratio for the current-multi (undocumented)*/␊
571␉␉␉␉currdiv = bitfield(msr, 14, 14);␊
572␊
573␉␉␉␉// This will always be model >= 3␊
574␉␉␉␉if ((p->CPU.Family == 0x06 && p->CPU.Model >= 0x0e) || (p->CPU.Family == 0x0f)) {␊
575␉␉␉␉␉/* On these models, maxcoef defines TSC freq */␊
576␉␉␉␉␉maxcoef = bitfield(msr, 44, 40);␊
577␉␉␉␉} else {␊
578␉␉␉␉␉/* On lower models, currcoef defines TSC freq */␊
579␉␉␉␉␉/* XXX */␊
580␉␉␉␉␉maxcoef = currcoef;␊
581␉␉␉␉}␊
582␊
583␉␉␉␉if (maxcoef) {␊
584␉␉␉␉␉if (maxdiv) {␊
585␉␉␉␉␉␉fsbFrequency = ((tscFrequency * 2) / ((maxcoef * 2) + 1));␊
586␉␉␉␉␉} else {␊
587␉␉␉␉␉␉fsbFrequency = (tscFrequency / maxcoef);␊
588␉␉␉␉␉}␊
589␉␉␉␉␉if (currdiv) {␊
590␉␉␉␉␉␉cpuFrequency = (fsbFrequency * ((currcoef * 2) + 1) / 2);␊
591␉␉␉␉␉} else {␊
592␉␉␉␉␉␉cpuFrequency = (fsbFrequency * currcoef);␊
593␉␉␉␉␉}␊
594␉␉␉␉␉DBG("max: %d%s current: %d%s\n", maxcoef, maxdiv ? ".5" : "",currcoef, currdiv ? ".5" : "");␊
595␉␉␉␉}␊
596␉␉␉␉break;␊
597␉␉␉}␊
598␉␉}␊
599␉␉/* Mobile CPU */␊
600␉␉if (rdmsr64(MSR_IA32_PLATFORM_ID) & (1<<28)) {␊
601␉␉␉p->CPU.Features |= CPU_FEATURE_MOBILE;␊
602␉␉}␊
603␉} else if ((p->CPU.Vendor == CPUID_VENDOR_AMD) && (p->CPU.Family == 0x0f)) {␊
604␉␉switch(p->CPU.ExtFamily) {␊
605␉␉␉case 0x00: /* K8 */␊
606␉␉␉␉msr = rdmsr64(K8_FIDVID_STATUS);␊
607␉␉␉␉maxcoef = bitfield(msr, 21, 16) / 2 + 4;␊
608␉␉␉␉currcoef = bitfield(msr, 5, 0) / 2 + 4;␊
609␉␉␉␉break;␊
610␊
611␉␉␉case 0x01: /* K10 */␊
612␉␉␉␉msr = rdmsr64(K10_COFVID_STATUS);␊
613␉␉␉␉do_cpuid2(0x00000006, 0, p->CPU.CPUID[CPUID_6]);␊
614␉␉␉␉// EffFreq: effective frequency interface␊
615␉␉␉␉if (bitfield(p->CPU.CPUID[CPUID_6][2], 0, 0) == 1) {␊
616␉␉␉␉␉//uint64_t mperf = measure_mperf_frequency();␊
617␉␉␉␉␉uint64_t aperf = measure_aperf_frequency();␊
618␉␉␉␉␉cpuFrequency = aperf;␊
619␉␉␉␉}␊
620␉␉␉␉// NOTE: tsc runs at the maccoeff (non turbo)␊
621␉␉␉␉//␉␉␉*not* at the turbo frequency.␊
622␉␉␉␉maxcoef␉ = bitfield(msr, 54, 49) / 2 + 4;␊
623␉␉␉␉currcoef = bitfield(msr, 5, 0) + 0x10;␊
624␉␉␉␉currdiv = 2 << bitfield(msr, 8, 6);␊
625␊
626␉␉␉␉break;␊
627␊
628␉␉␉case 0x05: /* K14 */␊
629␉␉␉␉msr = rdmsr64(K10_COFVID_STATUS);␊
630␉␉␉␉currcoef = (bitfield(msr, 54, 49) + 0x10) << 2;␊
631␉␉␉␉currdiv = (bitfield(msr, 8, 4) + 1) << 2;␊
632␉␉␉␉currdiv += bitfield(msr, 3, 0);␊
633␊
634␉␉␉␉break;␊
635␊
636␉␉␉case 0x02: /* K11 */␊
637␉␉␉␉// not implimented␊
638␉␉␉␉break;␊
639␉␉}␊
640␊
641␉␉if (maxcoef) {␊
642␉␉␉if (currdiv) {␊
643␉␉␉␉if (!currcoef) {␊
644␉␉␉␉␉currcoef = maxcoef;␊
645␉␉␉␉}␊
646␊
647␉␉␉␉if (!cpuFrequency) {␊
648␉␉␉␉␉fsbFrequency = ((tscFrequency * currdiv) / currcoef);␊
649␉␉␉␉} else {␊
650␉␉␉␉␉fsbFrequency = ((cpuFrequency * currdiv) / currcoef);␊
651␉␉␉␉}␊
652␉␉␉␉DBG("%d.%d\n", currcoef / currdiv, ((currcoef % currdiv) * 100) / currdiv);␊
653␉␉␉} else {␊
654␉␉␉␉if (!cpuFrequency) {␊
655␉␉␉␉␉fsbFrequency = (tscFrequency / maxcoef);␊
656␉␉␉␉} else {␊
657␉␉␉␉␉fsbFrequency = (cpuFrequency / maxcoef);␊
658␉␉␉␉}␊
659␉␉␉␉DBG("%d\n", currcoef);␊
660␉␉␉}␊
661␉␉} else if (currcoef) {␊
662␉␉␉if (currdiv) {␊
663␉␉␉␉fsbFrequency = ((tscFrequency * currdiv) / currcoef);␊
664␉␉␉␉DBG("%d.%d\n", currcoef / currdiv, ((currcoef % currdiv) * 100) / currdiv);␊
665␉␉␉} else {␊
666␉␉␉␉fsbFrequency = (tscFrequency / currcoef);␊
667␉␉␉␉DBG("%d\n", currcoef);␊
668␉␉␉}␊
669␉␉}␊
670␉␉if (!cpuFrequency) cpuFrequency = tscFrequency;␊
671␉}␊
672␉␊
673#if 0␊
674␉if (!fsbFrequency) {␊
675␉␉fsbFrequency = (DEFAULT_FSB * 1000);␊
676␉␉cpuFrequency = tscFrequency;␊
677␉␉DBG("0 ! using the default value for FSB !\n");␊
678␉}␊
679␊
680␉DBG("cpu freq = 0x%016llxn", timeRDTSC() * 20);␊
681␊
682#endif␊
683␊
684␉p->CPU.MaxCoef = maxcoef;␊
685␉p->CPU.MaxDiv = maxdiv;␊
686␉p->CPU.CurrCoef = currcoef;␊
687␉p->CPU.CurrDiv = currdiv;␊
688␉p->CPU.TSCFrequency = tscFrequency;␊
689␉p->CPU.FSBFrequency = fsbFrequency;␊
690␉p->CPU.CPUFrequency = cpuFrequency;␊
691␊
692␉// keep formatted with spaces instead of tabs␊
693␉DBG("\n---------------------------------------------\n");␊
694 ␉DBG("------------------ CPU INFO -----------------\n");␊
695␉DBG("---------------------------------------------\n");␊
696␉DBG("Brand String: %s\n", p->CPU.BrandString); // Processor name (BIOS)␊
697␉DBG("Vendor: 0x%x\n", p->CPU.Vendor); // Vendor ex: GenuineIntel␊
698␉DBG("Family: 0x%x\n", p->CPU.Family); // Family ex: 6 (06h)␊
699␉DBG("ExtFamily: 0x%x\n", p->CPU.ExtFamily);␊
700␉DBG("Signature: %x\n", p->CPU.Signature); // CPUID signature␊
701␉DBG("Model: 0x%x\n", p->CPU.Model); // Model ex: 37 (025h)␊
702␉DBG("ExtModel: 0x%x\n", p->CPU.ExtModel);␊
703␉DBG("Stepping: 0x%x\n", p->CPU.Stepping); // Stepping ex: 5 (05h)␊
704␉DBG("MaxCoef: 0x%x\n", p->CPU.MaxCoef);␊
705␉DBG("CurrCoef: 0x%x\n", p->CPU.CurrCoef);␊
706␉DBG("MaxDiv: 0x%x\n", p->CPU.MaxDiv);␊
707␉DBG("CurrDiv: 0x%x\n", p->CPU.CurrDiv);␊
708␉DBG("TSCFreq: %dMHz\n", p->CPU.TSCFrequency / 1000000);␊
709␉DBG("FSBFreq: %dMHz\n", p->CPU.FSBFrequency / 1000000);␊
710␉DBG("CPUFreq: %dMHz\n", p->CPU.CPUFrequency / 1000000);␊
711␉DBG("Cores: %d\n", p->CPU.NoCores); // Cores␊
712␉DBG("Logical processor: %d\n", p->CPU.NoThreads); // Logical procesor␊
713␉DBG("Features: 0x%08x\n", p->CPU.Features);␊
714␊
715␉DBG("\n---------------------------------------------\n");␊
716#if DEBUG_CPU␊
717␉pause();␊
718#endif␊
719}␊
720

Archive Download this file

Revision: 2469