Chameleon

Chameleon Svn Source Tree

Root/branches/azimutz/Chazi/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 "boot.h"
10
11#ifndef DEBUG_CPU
12#define DEBUG_CPU 0
13#endif
14
15#if DEBUG_CPU
16#define DBG(x...)printf(x)
17#else
18#define DBG(x...)msglog(x)
19#endif
20
21/*
22 * DFE: Measures the TSC frequency in Hz (64-bit) using the ACPI PM timer
23 */
24static uint64_t measure_tsc_frequency(void)
25{
26 uint64_t tscStart;
27 uint64_t tscEnd;
28 uint64_t tscDelta = 0xffffffffffffffffULL;
29 unsigned long pollCount;
30 uint64_t retval = 0;
31 int i;
32
33 /* Time how many TSC ticks elapse in 30 msec using the 8254 PIT
34 * counter 2. We run this loop 3 times to make sure the cache
35 * is hot and we take the minimum delta from all of the runs.
36 * That is to say that we're biased towards measuring the minimum
37 * number of TSC ticks that occur while waiting for the timer to
38 * expire. That theoretically helps avoid inconsistencies when
39 * running under a VM if the TSC is not virtualized and the host
40 * steals time. The TSC is normally virtualized for VMware.
41 */
42 for(i = 0; i < 10; ++i)
43 {
44 enable_PIT2();
45 set_PIT2_mode0(CALIBRATE_LATCH);
46 tscStart = rdtsc64();
47 pollCount = poll_PIT2_gate();
48 tscEnd = rdtsc64();
49 /* The poll loop must have run at least a few times for accuracy */
50 if(pollCount <= 1)
51 continue;
52 /* The TSC must increment at LEAST once every millisecond. We
53 * should have waited exactly 30 msec so the TSC delta should
54 * be >= 30. Anything less and the processor is way too slow.
55 */
56 if((tscEnd - tscStart) <= CALIBRATE_TIME_MSEC)
57 continue;
58 // tscDelta = MIN(tscDelta, (tscEnd - tscStart))
59 if( (tscEnd - tscStart) < tscDelta )
60 tscDelta = tscEnd - tscStart;
61 }
62 /* tscDelta is now the least number of TSC ticks the processor made in
63 * a timespan of 0.03 s (e.g. 30 milliseconds)
64 * Linux thus divides by 30 which gives the answer in kiloHertz because
65 * 1 / ms = kHz. But we're xnu and most of the rest of the code uses
66 * Hz so we need to convert our milliseconds to seconds. Since we're
67 * dividing by the milliseconds, we simply multiply by 1000.
68 */
69
70 /* Unlike linux, we're not limited to 32-bit, but we do need to take care
71 * that we're going to multiply by 1000 first so we do need at least some
72 * arithmetic headroom. For now, 32-bit should be enough.
73 * Also unlike Linux, our compiler can do 64-bit integer arithmetic.
74 */
75 if(tscDelta > (1ULL<<32))
76 retval = 0;
77 else
78 {
79 retval = tscDelta * 1000 / 30;
80 }
81 disable_PIT2();
82 return retval;
83}
84
85/*
86 * Calculates the FSB and CPU frequencies using specific MSRs for each CPU
87 * - multi. is read from a specific MSR. In the case of Intel, there is:
88 * a max multi. (used to calculate the FSB freq.),
89 * and a current multi. (used to calculate the CPU freq.)
90 * - fsbFrequency = tscFrequency / multi
91 * - cpuFrequency = fsbFrequency * multi
92 */
93
94void scan_cpu(PlatformInfo_t *p)
95{
96const char *newratio;
97intlen, myfsb;
98uint8_tmaxcoef, maxdiv, currcoef, bus_ratio_max, bus_ratio_min, currdiv;
99uint32_tmax_ratio, min_ratio;
100uint64_ttscFrequency, fsbFrequency, cpuFrequency, msr, flex_ratio;
101
102max_ratio = min_ratio = myfsb = bus_ratio_min = 0;
103maxcoef = maxdiv = bus_ratio_max = currcoef = currdiv = 0;
104
105/* get cpuid values */
106do_cpuid(0x00000000, p->CPU.CPUID[CPUID_0]);
107do_cpuid(0x00000001, p->CPU.CPUID[CPUID_1]);
108do_cpuid(0x00000002, p->CPU.CPUID[CPUID_2]);
109do_cpuid(0x00000003, p->CPU.CPUID[CPUID_3]);
110do_cpuid2(0x00000004, 0, p->CPU.CPUID[CPUID_4]);
111do_cpuid(0x80000000, p->CPU.CPUID[CPUID_80]);
112if ((p->CPU.CPUID[CPUID_80][0] & 0x0000000f) >= 1) {
113do_cpuid(0x80000001, p->CPU.CPUID[CPUID_81]);
114}
115#if DEBUG_CPU
116{
117inti;
118printf("CPUID Raw Values:\n");
119for (i=0; i<CPUID_MAX; i++) {
120printf("%02d: %08x-%08x-%08x-%08x\n", i,
121p->CPU.CPUID[i][0], p->CPU.CPUID[i][1],
122p->CPU.CPUID[i][2], p->CPU.CPUID[i][3]);
123}
124}
125#endif
126p->CPU.Vendor= p->CPU.CPUID[CPUID_0][1];
127p->CPU.Signature= p->CPU.CPUID[CPUID_1][0];
128p->CPU.Stepping= bitfield(p->CPU.CPUID[CPUID_1][0], 3, 0);
129p->CPU.Model= bitfield(p->CPU.CPUID[CPUID_1][0], 7, 4);
130p->CPU.Family= bitfield(p->CPU.CPUID[CPUID_1][0], 11, 8);
131p->CPU.ExtModel= bitfield(p->CPU.CPUID[CPUID_1][0], 19, 16);
132p->CPU.ExtFamily= bitfield(p->CPU.CPUID[CPUID_1][0], 27, 20);
133
134p->CPU.Model += (p->CPU.ExtModel << 4);
135
136if (p->CPU.Vendor == 0x756E6547 /* Intel */ && p->CPU.Family == 0x06 && p->CPU.Model >= 0x1a){
137msr = rdmsr64(MSR_CORE_THREAD_COUNT);// Undocumented MSR in Nehalem and newer CPUs
138p->CPU.NoCores= bitfield((uint32_t)msr, 31, 16);// Using undocumented MSR to get actual values
139p->CPU.NoThreads= bitfield((uint32_t)msr, 15, 0);// Using undocumented MSR to get actual values
140} else {
141p->CPU.NoThreads= bitfield(p->CPU.CPUID[CPUID_1][1], 23, 16);// Use previous method for Cores and Threads
142p->CPU.NoCores= bitfield(p->CPU.CPUID[CPUID_4][0], 31, 26) + 1;
143}
144
145/* get brand string (if supported) */
146/* Copyright: from Apple's XNU cpuid.c */
147if (p->CPU.CPUID[CPUID_80][0] > 0x80000004) {
148uint32_treg[4];
149 char str[128], *s;
150/*
151 * The brand string 48 bytes (max), guaranteed to
152 * be NUL terminated.
153 */
154do_cpuid(0x80000002, reg);
155bcopy((char *)reg, &str[0], 16);
156do_cpuid(0x80000003, reg);
157bcopy((char *)reg, &str[16], 16);
158do_cpuid(0x80000004, reg);
159bcopy((char *)reg, &str[32], 16);
160for (s = str; *s != '\0'; s++) {
161if (*s != ' ') break;
162}
163
164strlcpy(p->CPU.BrandString,s, sizeof(p->CPU.BrandString));
165
166if (!strncmp(p->CPU.BrandString, CPU_STRING_UNKNOWN, MIN(sizeof(p->CPU.BrandString), strlen(CPU_STRING_UNKNOWN) + 1))) {
167 /*
168 * This string means we have a firmware-programmable brand string,
169 * and the firmware couldn't figure out what sort of CPU we have.
170 */
171 p->CPU.BrandString[0] = '\0';
172 }
173}
174
175/* setup features */
176if ((bit(23) & p->CPU.CPUID[CPUID_1][3]) != 0) {
177p->CPU.Features |= CPU_FEATURE_MMX;
178}
179if ((bit(25) & p->CPU.CPUID[CPUID_1][3]) != 0) {
180p->CPU.Features |= CPU_FEATURE_SSE;
181}
182if ((bit(26) & p->CPU.CPUID[CPUID_1][3]) != 0) {
183p->CPU.Features |= CPU_FEATURE_SSE2;
184}
185if ((bit(0) & p->CPU.CPUID[CPUID_1][2]) != 0) {
186p->CPU.Features |= CPU_FEATURE_SSE3;
187}
188if ((bit(19) & p->CPU.CPUID[CPUID_1][2]) != 0) {
189p->CPU.Features |= CPU_FEATURE_SSE41;
190}
191if ((bit(20) & p->CPU.CPUID[CPUID_1][2]) != 0) {
192p->CPU.Features |= CPU_FEATURE_SSE42;
193}
194if ((bit(29) & p->CPU.CPUID[CPUID_81][3]) != 0) {
195p->CPU.Features |= CPU_FEATURE_EM64T;
196}
197if ((bit(5) & p->CPU.CPUID[CPUID_1][3]) != 0) {
198p->CPU.Features |= CPU_FEATURE_MSR;
199}
200//if ((bit(28) & p->CPU.CPUID[CPUID_1][3]) != 0) {
201if (p->CPU.NoThreads > p->CPU.NoCores) {
202p->CPU.Features |= CPU_FEATURE_HTT;
203}
204
205tscFrequency = measure_tsc_frequency();
206fsbFrequency = 0;
207cpuFrequency = 0;
208
209if ((p->CPU.Vendor == 0x756E6547 /* Intel */) && ((p->CPU.Family == 0x06) || (p->CPU.Family == 0x0f))) {
210int intelCPU = p->CPU.Model;
211if ((p->CPU.Family == 0x06 && p->CPU.Model >= 0x0c) || (p->CPU.Family == 0x0f && p->CPU.Model >= 0x03)) {
212/* Nehalem CPU model */
213if (p->CPU.Family == 0x06 && (p->CPU.Model == CPU_MODEL_NEHALEM ||
214 p->CPU.Model == CPU_MODEL_FIELDS ||
215 p->CPU.Model == CPU_MODEL_DALES ||
216 p->CPU.Model == CPU_MODEL_DALES_32NM ||
217 p->CPU.Model == CPU_MODEL_WESTMERE ||
218 p->CPU.Model == CPU_MODEL_NEHALEM_EX ||
219 p->CPU.Model == CPU_MODEL_WESTMERE_EX ||
220 p->CPU.Model == CPU_MODEL_SANDY ||
221 p->CPU.Model == CPU_MODEL_SANDY_XEON)) {
222msr = rdmsr64(MSR_PLATFORM_INFO);
223DBG("msr(%d): platform_info %08x\n", __LINE__, msr & 0xffffffff);
224//kabylDBG("MSR_PLATFORM_INFO (0xCE): 0x%08X%08X\n", (uint32_t)(msr >> 32), (uint32_t)msr);
225bus_ratio_max = (msr >> 8) & 0xff;
226bus_ratio_min = (msr >> 40) & 0xff; //valv: not sure about this one (Remarq.1)
227msr = rdmsr64(MSR_FLEX_RATIO);
228DBG("msr(%d): flex_ratio %08x\n", __LINE__, msr & 0xffffffff);
229//kabylDBG("MSR_FLEX_RATIO (0x194): 0x%08X%08X\n", (uint32_t)(msr >> 32), (uint32_t)msr);
230if ((msr >> 16) & 0x01) {
231flex_ratio = (msr >> 8) & 0xff;
232/* bcc9: at least on the gigabyte h67ma-ud2h,
233 where the cpu multipler can't be changed to
234 allow overclocking, the flex_ratio msr has unexpected (to OSX)
235 contents. These contents cause mach_kernel to
236 fail to compute the bus ratio correctly, instead
237 causing the system to crash since tscGranularity
238 is inadvertently set to 0.
239*/
240if (flex_ratio == 0) {
241/* Clear bit 16 (evidently the
242 presence bit) */
243wrmsr64(MSR_FLEX_RATIO, (msr & 0xFFFFFFFFFFFEFFFFULL));
244msr = rdmsr64(MSR_FLEX_RATIO);
245verbose("Unusable flex ratio detected. Patched MSR now %08x\n", msr & 0xffffffff);
246} else {
247if (bus_ratio_max > flex_ratio) {
248bus_ratio_max = flex_ratio;
249}
250}
251}
252
253if (bus_ratio_max) {
254fsbFrequency = (tscFrequency / bus_ratio_max);
255}
256//valv: Turbo Ratio Limit
257if ((intelCPU != 0x2e) && (intelCPU != 0x2f)) {
258msr = rdmsr64(MSR_TURBO_RATIO_LIMIT);
259cpuFrequency = bus_ratio_max * fsbFrequency;
260max_ratio = bus_ratio_max * 10;
261} else {
262cpuFrequency = tscFrequency;
263}
264if ((getValueForKey(kBusRatioKey, &newratio, &len, &bootInfo->bootConfig)) && (len <= 4)) {
265max_ratio = atoi(newratio);
266max_ratio = (max_ratio * 10);
267if (len >= 3) max_ratio = (max_ratio + 5);
268
269verbose("Bus-Ratio: min=%d, max=%s\n", bus_ratio_min, newratio);
270
271// extreme overclockers may love 320 ;)
272if ((max_ratio >= min_ratio) && (max_ratio <= 320)) {
273cpuFrequency = (fsbFrequency * max_ratio) / 10;
274if (len >= 3) maxdiv = 1;
275else maxdiv = 0;
276} else {
277max_ratio = (bus_ratio_max * 10);
278}
279}
280//valv: to be uncommented if Remarq.1 didn't stick
281/*if(bus_ratio_max > 0) bus_ratio = flex_ratio;*/
282p->CPU.MaxRatio = max_ratio;
283p->CPU.MinRatio = min_ratio;
284
285myfsb = fsbFrequency / 1000000;
286verbose("Sticking with [BCLK: %dMhz, Bus-Ratio: %d]\n", myfsb, max_ratio);
287currcoef = bus_ratio_max;
288} else {
289msr = rdmsr64(MSR_IA32_PERF_STATUS);
290//Azi: __FILE__ %d, __LINE__ %d, __FUNCTION__ %s
291//msglog("%s - %d\n", __FUNCTION__, __LINE__);
292DBG("msr(%d): ia32_perf_stat 0x%08x\n", __LINE__, msr & 0xffffffff);
293//Azi: 0x00000F2700000F27
294//kabylDBG("MSR_IA32_PERF_STATUS (0x198): 0x%08X%08X\n", (uint32_t)(msr >> 32), (uint32_t)msr);
295currcoef = (msr >> 8) & 0x1f;
296/* Non-integer bus ratio for the max-multi*/
297maxdiv = (msr >> 46) & 0x01;
298/* Non-integer bus ratio for the current-multi (undocumented)*/
299currdiv = (msr >> 14) & 0x01;
300
301if ((p->CPU.Family == 0x06 && p->CPU.Model >= 0x0e) || (p->CPU.Family == 0x0f)) // This will always be model >= 3
302{
303/* On these models, maxcoef defines TSC freq */
304maxcoef = (msr >> 40) & 0x1f;
305} else {
306/* On lower models, currcoef defines TSC freq */
307/* XXX */
308maxcoef = currcoef;
309}
310
311if (maxcoef) {
312if (maxdiv) {
313fsbFrequency = ((tscFrequency * 2) / ((maxcoef * 2) + 1));
314} else {
315fsbFrequency = (tscFrequency / maxcoef);
316}
317if (currdiv) {
318cpuFrequency = (fsbFrequency * ((currcoef * 2) + 1) / 2);
319} else {
320cpuFrequency = (fsbFrequency * currcoef);
321}
322DBG("max: %d%s current: %d%s\n", maxcoef, maxdiv ? ".5" : "",currcoef, currdiv ? ".5" : "");
323}
324}
325}
326/* Mobile CPU ? */
327if (rdmsr64(0x17) & (1<<28)) {
328p->CPU.Features |= CPU_FEATURE_MOBILE;
329}
330}
331#if 0
332else if((p->CPU.Vendor == 0x68747541 /* AMD */) && (p->CPU.Family == 0x0f)) {
333if(p->CPU.ExtFamily == 0x00 /* K8 */) {
334msr = rdmsr64(K8_FIDVID_STATUS);
335currcoef = (msr & 0x3f) / 2 + 4;
336currdiv = (msr & 0x01) * 2;
337} else if(p->CPU.ExtFamily >= 0x01 /* K10+ */) {
338msr = rdmsr64(K10_COFVID_STATUS);
339if(p->CPU.ExtFamily == 0x01 /* K10 */)
340currcoef = (msr & 0x3f) + 0x10;
341else /* K11+ */
342currcoef = (msr & 0x3f) + 0x08;
343currdiv = (2 << ((msr >> 6) & 0x07));
344}
345
346if (currcoef) {
347if (currdiv) {
348fsbFrequency = ((tscFrequency * currdiv) / currcoef);
349DBG("%d.%d\n", currcoef / currdiv, ((currcoef % currdiv) * 100) / currdiv);
350} else {
351fsbFrequency = (tscFrequency / currcoef);
352DBG("%d\n", currcoef);
353}
354fsbFrequency = (tscFrequency / currcoef);
355cpuFrequency = tscFrequency;
356}
357}
358
359if (!fsbFrequency) {
360fsbFrequency = (DEFAULT_FSB * 1000);
361cpuFrequency = tscFrequency;
362DBG("0 ! using the default value for FSB !\n");
363}
364#endif
365
366p->CPU.MaxCoef = maxcoef;
367p->CPU.MaxDiv = maxdiv;
368p->CPU.CurrCoef = currcoef;
369p->CPU.CurrDiv = currdiv;
370p->CPU.TSCFrequency = tscFrequency;
371p->CPU.FSBFrequency = fsbFrequency;
372p->CPU.CPUFrequency = cpuFrequency;
373
374DBG("CPU: Vendor/Model/ExtModel: 0x%x/0x%x/0x%x\n", p->CPU.Vendor, p->CPU.Model, p->CPU.ExtModel);
375DBG("CPU: Family/ExtFamily: 0x%x/0x%x\n", p->CPU.Family, p->CPU.ExtFamily);
376DBG("CPU: MaxCoef/CurrCoef: 0x%x/0x%x\n", p->CPU.MaxCoef, p->CPU.CurrCoef);
377DBG("CPU: MaxDiv/CurrDiv: 0x%x/0x%x\n", p->CPU.MaxDiv, p->CPU.CurrDiv);
378DBG("CPU: TSCFreq: %dMHz\n", p->CPU.TSCFrequency / 1000000);
379DBG("CPU: FSBFreq: %dMHz\n", p->CPU.FSBFrequency / 1000000);
380DBG("CPU: CPUFreq: %dMHz\n", p->CPU.CPUFrequency / 1000000);
381DBG("CPU: NoCores/NoThreads: %d/%d\n", p->CPU.NoCores, p->CPU.NoThreads);
382DBG("CPU: Features: 0x%08x\n", p->CPU.Features);
383#if DEBUG_CPU
384pause();
385#endif
386}
387

Archive Download this file

Revision: 842