Chameleon

Chameleon Svn Source Tree

Root/branches/azimutz/Cleancut/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 * DFE: Measures the TSC frequency in Hz (64-bit) using the ACPI PM timer
24 */
25static uint64_t measure_tsc_frequency(void)
26{
27 uint64_t tscStart;
28 uint64_t tscEnd;
29 uint64_t tscDelta = 0xffffffffffffffffULL;
30 unsigned long pollCount;
31 uint64_t retval = 0;
32 int i;
33
34 /* Time how many TSC ticks elapse in 30 msec using the 8254 PIT
35 * counter 2. We run this loop 3 times to make sure the cache
36 * is hot and we take the minimum delta from all of the runs.
37 * That is to say that we're biased towards measuring the minimum
38 * number of TSC ticks that occur while waiting for the timer to
39 * expire. That theoretically helps avoid inconsistencies when
40 * running under a VM if the TSC is not virtualized and the host
41 * steals time. The TSC is normally virtualized for VMware.
42 */
43 for(i = 0; i < 10; ++i)
44 {
45 enable_PIT2();
46 set_PIT2_mode0(CALIBRATE_LATCH);
47 tscStart = rdtsc64();
48 pollCount = poll_PIT2_gate();
49 tscEnd = rdtsc64();
50 /* The poll loop must have run at least a few times for accuracy */
51 if(pollCount <= 1)
52 continue;
53 /* The TSC must increment at LEAST once every millisecond. We
54 * should have waited exactly 30 msec so the TSC delta should
55 * be >= 30. Anything less and the processor is way too slow.
56 */
57 if((tscEnd - tscStart) <= CALIBRATE_TIME_MSEC)
58 continue;
59 // tscDelta = min(tscDelta, (tscEnd - tscStart))
60 if( (tscEnd - tscStart) < tscDelta )
61 tscDelta = tscEnd - tscStart;
62 }
63 /* tscDelta is now the least number of TSC ticks the processor made in
64 * a timespan of 0.03 s (e.g. 30 milliseconds)
65 * Linux thus divides by 30 which gives the answer in kiloHertz because
66 * 1 / ms = kHz. But we're xnu and most of the rest of the code uses
67 * Hz so we need to convert our milliseconds to seconds. Since we're
68 * dividing by the milliseconds, we simply multiply by 1000.
69 */
70
71 /* Unlike linux, we're not limited to 32-bit, but we do need to take care
72 * that we're going to multiply by 1000 first so we do need at least some
73 * arithmetic headroom. For now, 32-bit should be enough.
74 * Also unlike Linux, our compiler can do 64-bit integer arithmetic.
75 */
76 if(tscDelta > (1ULL<<32))
77 retval = 0;
78 else
79 {
80 retval = tscDelta * 1000 / 30;
81 }
82 disable_PIT2();
83 return retval;
84}
85
86/*
87 * Calculates the FSB and CPU frequencies using specific MSRs for each CPU
88 * - multi. is read from a specific MSR. In the case of Intel, there is:
89 * a max multi. (used to calculate the FSB freq.),
90 * and a current multi. (used to calculate the CPU freq.)
91 * - fsbFrequency = tscFrequency / multi
92 * - cpuFrequency = fsbFrequency * multi
93 */
94
95void scan_cpu(PlatformInfo_t *p)
96{
97uint64_ttscFrequency, fsbFrequency, cpuFrequency;
98uint64_tmsr, flex_ratio;
99uint8_tmaxcoef, maxdiv, currcoef, bus_ratio_max, currdiv;
100const char *newratio;
101int len, myfsb;
102uint8_t bus_ratio_min;
103uint32_t max_ratio, min_ratio;
104
105max_ratio = min_ratio = myfsb = bus_ratio_min = 0;
106maxcoef = maxdiv = bus_ratio_max = currcoef = currdiv = 0;
107
108/* get cpuid values */
109do_cpuid(0x00000000, p->CPU.CPUID[CPUID_0]);
110do_cpuid(0x00000001, p->CPU.CPUID[CPUID_1]);
111do_cpuid(0x00000002, p->CPU.CPUID[CPUID_2]);
112do_cpuid(0x00000003, p->CPU.CPUID[CPUID_3]);
113do_cpuid2(0x00000004, 0, p->CPU.CPUID[CPUID_4]);
114do_cpuid(0x80000000, p->CPU.CPUID[CPUID_80]);
115if ((p->CPU.CPUID[CPUID_80][0] & 0x0000000f) >= 1) {
116do_cpuid(0x80000001, p->CPU.CPUID[CPUID_81]);
117}
118#if DEBUG_CPU
119{
120inti;
121printf("CPUID Raw Values:\n");
122for (i=0; i<CPUID_MAX; i++) {
123printf("%02d: %08x-%08x-%08x-%08x\n", i,
124p->CPU.CPUID[i][0], p->CPU.CPUID[i][1],
125p->CPU.CPUID[i][2], p->CPU.CPUID[i][3]);
126}
127}
128#endif
129p->CPU.Vendor= p->CPU.CPUID[CPUID_0][1];
130p->CPU.Signature= p->CPU.CPUID[CPUID_1][0];
131p->CPU.Stepping= bitfield(p->CPU.CPUID[CPUID_1][0], 3, 0);
132p->CPU.Model= bitfield(p->CPU.CPUID[CPUID_1][0], 7, 4);
133p->CPU.Family= bitfield(p->CPU.CPUID[CPUID_1][0], 11, 8);
134p->CPU.ExtModel= bitfield(p->CPU.CPUID[CPUID_1][0], 19, 16);
135p->CPU.ExtFamily= bitfield(p->CPU.CPUID[CPUID_1][0], 27, 20);
136
137 p->CPU.Model += (p->CPU.ExtModel << 4);
138
139 if (p->CPU.Vendor == 0x756E6547 /* Intel */ && p->CPU.Family == 0x06 && p->CPU.Model >= 0x1a){
140 msr = rdmsr64(MSR_CORE_THREAD_COUNT);// Undocumented MSR in Nehalem and newer CPUs
141 p->CPU.NoCores= bitfield((uint32_t)msr, 31, 16);// Using undocumented MSR to get actual values
142 p->CPU.NoThreads= bitfield((uint32_t)msr, 15, 0);// Using undocumented MSR to get actual values
143} else {
144 p->CPU.NoThreads= bitfield(p->CPU.CPUID[CPUID_1][1], 23, 16);// Use previous method for Cores and Threads
145 p->CPU.NoCores= bitfield(p->CPU.CPUID[CPUID_4][0], 31, 26) + 1;
146}
147
148/* get brand string (if supported) */
149/* Copyright: from Apple's XNU cpuid.c */
150if (p->CPU.CPUID[CPUID_80][0] > 0x80000004) {
151uint32_treg[4];
152 char str[128], *s;
153/*
154 * The brand string 48 bytes (max), guaranteed to
155 * be NUL terminated.
156 */
157do_cpuid(0x80000002, reg);
158bcopy((char *)reg, &str[0], 16);
159do_cpuid(0x80000003, reg);
160bcopy((char *)reg, &str[16], 16);
161do_cpuid(0x80000004, reg);
162bcopy((char *)reg, &str[32], 16);
163for (s = str; *s != '\0'; s++) {
164if (*s != ' ') break;
165}
166
167strlcpy(p->CPU.BrandString,s, sizeof(p->CPU.BrandString));
168
169if (!strncmp(p->CPU.BrandString, CPU_STRING_UNKNOWN, min(sizeof(p->CPU.BrandString), strlen(CPU_STRING_UNKNOWN) + 1))) {
170 /*
171 * This string means we have a firmware-programmable brand string,
172 * and the firmware couldn't figure out what sort of CPU we have.
173 */
174 p->CPU.BrandString[0] = '\0';
175 }
176}
177
178/* setup features */
179if ((bit(23) & p->CPU.CPUID[CPUID_1][3]) != 0) {
180p->CPU.Features |= CPU_FEATURE_MMX;
181}
182if ((bit(25) & p->CPU.CPUID[CPUID_1][3]) != 0) {
183p->CPU.Features |= CPU_FEATURE_SSE;
184}
185if ((bit(26) & p->CPU.CPUID[CPUID_1][3]) != 0) {
186p->CPU.Features |= CPU_FEATURE_SSE2;
187}
188if ((bit(0) & p->CPU.CPUID[CPUID_1][2]) != 0) {
189p->CPU.Features |= CPU_FEATURE_SSE3;
190}
191if ((bit(19) & p->CPU.CPUID[CPUID_1][2]) != 0) {
192p->CPU.Features |= CPU_FEATURE_SSE41;
193}
194if ((bit(20) & p->CPU.CPUID[CPUID_1][2]) != 0) {
195p->CPU.Features |= CPU_FEATURE_SSE42;
196}
197if ((bit(29) & p->CPU.CPUID[CPUID_81][3]) != 0) {
198p->CPU.Features |= CPU_FEATURE_EM64T;
199}
200if ((bit(5) & p->CPU.CPUID[CPUID_1][3]) != 0) {
201p->CPU.Features |= CPU_FEATURE_MSR;
202}
203//if ((bit(28) & p->CPU.CPUID[CPUID_1][3]) != 0) {
204if (p->CPU.NoThreads > p->CPU.NoCores) {
205p->CPU.Features |= CPU_FEATURE_HTT;
206}
207
208tscFrequency = measure_tsc_frequency();
209fsbFrequency = 0;
210cpuFrequency = 0;
211
212if ((p->CPU.Vendor == 0x756E6547 /* Intel */) && ((p->CPU.Family == 0x06) || (p->CPU.Family == 0x0f))) {
213int intelCPU = p->CPU.Model;
214if ((p->CPU.Family == 0x06 && p->CPU.Model >= 0x0c) || (p->CPU.Family == 0x0f && p->CPU.Model >= 0x03)) {
215/* Nehalem CPU model */
216if (p->CPU.Family == 0x06 && (p->CPU.Model == CPU_MODEL_NEHALEM ||
217 p->CPU.Model == CPU_MODEL_FIELDS ||
218 p->CPU.Model == CPU_MODEL_DALES ||
219 p->CPU.Model == CPU_MODEL_DALES_32NM ||
220 p->CPU.Model == CPU_MODEL_WESTMERE ||
221 p->CPU.Model == CPU_MODEL_NEHALEM_EX ||
222 p->CPU.Model == CPU_MODEL_WESTMERE_EX ||
223 p->CPU.Model == CPU_MODEL_SANDY ||
224 p->CPU.Model == CPU_MODEL_SANDY_XEON)) {
225msr = rdmsr64(MSR_PLATFORM_INFO);
226DBG("msr(%d): platform_info %08x\n", __LINE__, msr & 0xffffffff);
227bus_ratio_max = (msr >> 8) & 0xff;
228bus_ratio_min = (msr >> 40) & 0xff; //valv: not sure about this one (Remarq.1)
229msr = rdmsr64(MSR_FLEX_RATIO);
230DBG("msr(%d): flex_ratio %08x\n", __LINE__, msr & 0xffffffff);
231if ((msr >> 16) & 0x01) {
232flex_ratio = (msr >> 8) & 0xff;
233/* bcc9: at least on the gigabyte h67ma-ud2h,
234 where the cpu multipler can't be changed to
235 allow overclocking, the flex_ratio msr has unexpected (to OSX)
236 contents. These contents cause mach_kernel to
237 fail to compute the bus ratio correctly, instead
238 causing the system to crash since tscGranularity
239 is inadvertently set to 0.
240*/
241if (flex_ratio == 0) {
242/* Clear bit 16 (evidently the
243 presence bit) */
244wrmsr64(MSR_FLEX_RATIO, (msr & 0xFFFFFFFFFFFEFFFFULL));
245msr = rdmsr64(MSR_FLEX_RATIO);
246verbose("Unusable flex ratio detected. Patched MSR now %08x\n", msr & 0xffffffff);
247} else {
248if (bus_ratio_max > flex_ratio) {
249bus_ratio_max = flex_ratio;
250}
251}
252}
253
254if (bus_ratio_max) {
255fsbFrequency = (tscFrequency / bus_ratio_max);
256}
257//valv: Turbo Ratio Limit
258if ((intelCPU != 0x2e) && (intelCPU != 0x2f)) {
259msr = rdmsr64(MSR_TURBO_RATIO_LIMIT);
260cpuFrequency = bus_ratio_max * fsbFrequency;
261max_ratio = bus_ratio_max * 10;
262} else {
263cpuFrequency = tscFrequency;
264}
265if ((getValueForKey(kbusratio, &newratio, &len, &bootInfo->bootConfig)) && (len <= 4)) {
266max_ratio = atoi(newratio);
267max_ratio = (max_ratio * 10);
268if (len >= 3) max_ratio = (max_ratio + 5);
269
270verbose("Bus-Ratio: min=%d, max=%s\n", bus_ratio_min, newratio);
271
272// extreme overclockers may love 320 ;)
273if ((max_ratio >= min_ratio) && (max_ratio <= 320)) {
274cpuFrequency = (fsbFrequency * max_ratio) / 10;
275if (len >= 3) maxdiv = 1;
276else maxdiv = 0;
277} else {
278max_ratio = (bus_ratio_max * 10);
279}
280}
281//valv: to be uncommented if Remarq.1 didn't stick
282/*if(bus_ratio_max > 0) bus_ratio = flex_ratio;*/
283p->CPU.MaxRatio = max_ratio;
284p->CPU.MinRatio = min_ratio;
285
286myfsb = fsbFrequency / 1000000;
287verbose("Sticking with [BCLK: %dMhz, Bus-Ratio: %d]\n", myfsb, max_ratio);
288currcoef = bus_ratio_max;
289} else {
290msr = rdmsr64(MSR_IA32_PERF_STATUS);
291DBG("msr(%d): ia32_perf_stat 0x%08x\n", __LINE__, msr & 0xffffffff);
292currcoef = (msr >> 8) & 0x1f;
293/* Non-integer bus ratio for the max-multi*/
294maxdiv = (msr >> 46) & 0x01;
295/* Non-integer bus ratio for the current-multi (undocumented)*/
296currdiv = (msr >> 14) & 0x01;
297
298if ((p->CPU.Family == 0x06 && p->CPU.Model >= 0x0e) || (p->CPU.Family == 0x0f)) // This will always be model >= 3
299{
300/* On these models, maxcoef defines TSC freq */
301maxcoef = (msr >> 40) & 0x1f;
302} else {
303/* On lower models, currcoef defines TSC freq */
304/* XXX */
305maxcoef = currcoef;
306}
307
308if (maxcoef) {
309if (maxdiv) {
310fsbFrequency = ((tscFrequency * 2) / ((maxcoef * 2) + 1));
311} else {
312fsbFrequency = (tscFrequency / maxcoef);
313}
314if (currdiv) {
315cpuFrequency = (fsbFrequency * ((currcoef * 2) + 1) / 2);
316} else {
317cpuFrequency = (fsbFrequency * currcoef);
318}
319DBG("max: %d%s current: %d%s\n", maxcoef, maxdiv ? ".5" : "",currcoef, currdiv ? ".5" : "");
320}
321}
322}
323/* Mobile CPU ? */
324if (rdmsr64(0x17) & (1<<28)) {
325p->CPU.Features |= CPU_FEATURE_MOBILE;
326}
327}
328#if 0
329else if((p->CPU.Vendor == 0x68747541 /* AMD */) && (p->CPU.Family == 0x0f)) {
330if(p->CPU.ExtFamily == 0x00 /* K8 */) {
331msr = rdmsr64(K8_FIDVID_STATUS);
332currcoef = (msr & 0x3f) / 2 + 4;
333currdiv = (msr & 0x01) * 2;
334} else if(p->CPU.ExtFamily >= 0x01 /* K10+ */) {
335msr = rdmsr64(K10_COFVID_STATUS);
336if(p->CPU.ExtFamily == 0x01 /* K10 */)
337currcoef = (msr & 0x3f) + 0x10;
338else /* K11+ */
339currcoef = (msr & 0x3f) + 0x08;
340currdiv = (2 << ((msr >> 6) & 0x07));
341}
342
343if (currcoef) {
344if (currdiv) {
345fsbFrequency = ((tscFrequency * currdiv) / currcoef);
346DBG("%d.%d\n", currcoef / currdiv, ((currcoef % currdiv) * 100) / currdiv);
347} else {
348fsbFrequency = (tscFrequency / currcoef);
349DBG("%d\n", currcoef);
350}
351fsbFrequency = (tscFrequency / currcoef);
352cpuFrequency = tscFrequency;
353}
354}
355
356if (!fsbFrequency) {
357fsbFrequency = (DEFAULT_FSB * 1000);
358cpuFrequency = tscFrequency;
359DBG("0 ! using the default value for FSB !\n");
360}
361#endif
362
363p->CPU.MaxCoef = maxcoef;
364p->CPU.MaxDiv = maxdiv;
365p->CPU.CurrCoef = currcoef;
366p->CPU.CurrDiv = currdiv;
367p->CPU.TSCFrequency = tscFrequency;
368p->CPU.FSBFrequency = fsbFrequency;
369p->CPU.CPUFrequency = cpuFrequency;
370
371DBG("CPU: Vendor/Model/ExtModel: 0x%x/0x%x/0x%x\n", p->CPU.Vendor, p->CPU.Model, p->CPU.ExtModel);
372DBG("CPU: Family/ExtFamily: 0x%x/0x%x\n", p->CPU.Family, p->CPU.ExtFamily);
373DBG("CPU: MaxCoef/CurrCoef: 0x%x/0x%x\n", p->CPU.MaxCoef, p->CPU.CurrCoef);
374DBG("CPU: MaxDiv/CurrDiv: 0x%x/0x%x\n", p->CPU.MaxDiv, p->CPU.CurrDiv);
375DBG("CPU: TSCFreq: %dMHz\n", p->CPU.TSCFrequency / 1000000);
376DBG("CPU: FSBFreq: %dMHz\n", p->CPU.FSBFrequency / 1000000);
377DBG("CPU: CPUFreq: %dMHz\n", p->CPU.CPUFrequency / 1000000);
378DBG("CPU: NoCores/NoThreads: %d/%d\n", p->CPU.NoCores, p->CPU.NoThreads);
379DBG("CPU: Features: 0x%08x\n", p->CPU.Features);
380#if DEBUG_CPU
381pause();
382#endif
383}
384

Archive Download this file

Revision: 839