Chameleon

Chameleon Svn Source Tree

Root/branches/ErmaC/Trunk/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{
27uint64_t tscStart;
28uint64_t tscEnd;
29uint64_t tscDelta = 0xffffffffffffffffULL;
30unsigned long pollCount;
31uint64_t retval = 0;
32int 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 */
43for(i = 0; i < 10; ++i)
44{
45enable_PIT2();
46set_PIT2_mode0(CALIBRATE_LATCH);
47tscStart = rdtsc64();
48pollCount = poll_PIT2_gate();
49tscEnd = rdtsc64();
50/* The poll loop must have run at least a few times for accuracy */
51if (pollCount <= 1)
52continue;
53/* The TSC must increment at LEAST once every millisecond.
54 * We should have waited exactly 30 msec so the TSC delta should
55 * be >= 30. Anything less and the processor is way too slow.
56 */
57if ((tscEnd - tscStart) <= CALIBRATE_TIME_MSEC)
58continue;
59// tscDelta = MIN(tscDelta, (tscEnd - tscStart))
60if ( (tscEnd - tscStart) < tscDelta )
61tscDelta = 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 */
76if (tscDelta > (1ULL<<32))
77retval = 0;
78else
79{
80retval = tscDelta * 1000 / 30;
81}
82disable_PIT2();
83return retval;
84}
85
86/*
87 * Original comment/code:
88 * "DFE: Measures the Max Performance Frequency in Hz (64-bit)"
89 *
90 * Measures the Actual Performance Frequency in Hz (64-bit)
91 * (just a naming change, mperf --> aperf )
92 */
93static uint64_t measure_aperf_frequency(void)
94{
95uint64_t aperfStart;
96uint64_t aperfEnd;
97uint64_t aperfDelta = 0xffffffffffffffffULL;
98unsigned long pollCount;
99uint64_t retval = 0;
100int i;
101
102/* Time how many APERF ticks elapse in 30 msec using the 8254 PIT
103 * counter 2. We run this loop 3 times to make sure the cache
104 * is hot and we take the minimum delta from all of the runs.
105 * That is to say that we're biased towards measuring the minimum
106 * number of APERF ticks that occur while waiting for the timer to
107 * expire.
108 */
109for(i = 0; i < 10; ++i)
110{
111enable_PIT2();
112set_PIT2_mode0(CALIBRATE_LATCH);
113aperfStart = rdmsr64(MSR_AMD_APERF);
114pollCount = poll_PIT2_gate();
115aperfEnd = rdmsr64(MSR_AMD_APERF);
116/* The poll loop must have run at least a few times for accuracy */
117if (pollCount <= 1)
118continue;
119/* The TSC must increment at LEAST once every millisecond.
120 * We should have waited exactly 30 msec so the APERF delta should
121 * be >= 30. Anything less and the processor is way too slow.
122 */
123if ((aperfEnd - aperfStart) <= CALIBRATE_TIME_MSEC)
124continue;
125// tscDelta = MIN(tscDelta, (tscEnd - tscStart))
126if ( (aperfEnd - aperfStart) < aperfDelta )
127aperfDelta = aperfEnd - aperfStart;
128}
129/* mperfDelta is now the least number of MPERF ticks the processor made in
130 * a timespan of 0.03 s (e.g. 30 milliseconds)
131 */
132
133if (aperfDelta > (1ULL<<32))
134retval = 0;
135else
136{
137retval = aperfDelta * 1000 / 30;
138}
139disable_PIT2();
140return retval;
141}
142
143/*
144 * Calculates the FSB and CPU frequencies using specific MSRs for each CPU
145 * - multi. is read from a specific MSR. In the case of Intel, there is:
146 * a max multi. (used to calculate the FSB freq.),
147 * and a current multi. (used to calculate the CPU freq.)
148 * - fsbFrequency = tscFrequency / multi
149 * - cpuFrequency = fsbFrequency * multi
150 */
151void scan_cpu(PlatformInfo_t *p)
152{
153uint64_ttscFrequency, fsbFrequency, cpuFrequency;
154uint64_tmsr, flex_ratio;
155uint8_tmaxcoef, maxdiv, currcoef, bus_ratio_max, currdiv;
156const char*newratio;
157intlen, myfsb;
158uint8_tbus_ratio_min;
159uint32_tmax_ratio, min_ratio;
160
161max_ratio = min_ratio = myfsb = bus_ratio_min = 0;
162maxcoef = maxdiv = bus_ratio_max = currcoef = currdiv = 0;
163
164/* get cpuid values */
165do_cpuid(0x00000000, p->CPU.CPUID[CPUID_0]);
166do_cpuid(0x00000001, p->CPU.CPUID[CPUID_1]);
167do_cpuid(0x00000002, p->CPU.CPUID[CPUID_2]);
168do_cpuid(0x00000003, p->CPU.CPUID[CPUID_3]);
169do_cpuid2(0x00000004, 0, p->CPU.CPUID[CPUID_4]);
170do_cpuid(0x80000000, p->CPU.CPUID[CPUID_80]);
171if ((p->CPU.CPUID[CPUID_80][0] & 0x0000000f) >= 8) {
172do_cpuid(0x80000008, p->CPU.CPUID[CPUID_88]);
173do_cpuid(0x80000001, p->CPU.CPUID[CPUID_81]);
174}
175else if ((p->CPU.CPUID[CPUID_80][0] & 0x0000000f) >= 1) {
176do_cpuid(0x80000001, p->CPU.CPUID[CPUID_81]);
177}
178
179#if DEBUG_CPU
180{
181inti;
182printf("CPUID Raw Values:\n");
183for (i=0; i<CPUID_MAX; i++) {
184printf("%02d: %08x-%08x-%08x-%08x\n", i,
185 p->CPU.CPUID[i][0], p->CPU.CPUID[i][1],
186 p->CPU.CPUID[i][2], p->CPU.CPUID[i][3]);
187}
188}
189#endif
190
191p->CPU.Vendor= p->CPU.CPUID[CPUID_0][1];
192p->CPU.Signature= p->CPU.CPUID[CPUID_1][0];
193p->CPU.Stepping= bitfield(p->CPU.CPUID[CPUID_1][0], 3, 0);
194p->CPU.Model= bitfield(p->CPU.CPUID[CPUID_1][0], 7, 4);
195p->CPU.Family= bitfield(p->CPU.CPUID[CPUID_1][0], 11, 8);
196p->CPU.ExtModel= bitfield(p->CPU.CPUID[CPUID_1][0], 19, 16);
197p->CPU.ExtFamily= bitfield(p->CPU.CPUID[CPUID_1][0], 27, 20);
198
199p->CPU.Model += (p->CPU.ExtModel << 4);
200
201if (p->CPU.Vendor == CPUID_VENDOR_INTEL &&
202p->CPU.Family == 0x06 &&
203p->CPU.Model >= CPUID_MODEL_NEHALEM &&
204p->CPU.Model != CPUID_MODEL_ATOM// MSR is *NOT* available on the Intel Atom CPU
205)
206{
207msr = rdmsr64(MSR_CORE_THREAD_COUNT);// Undocumented MSR in Nehalem and newer CPUs
208p->CPU.NoCores= bitfield((uint32_t)msr, 31, 16);// Using undocumented MSR to get actual values
209p->CPU.NoThreads= bitfield((uint32_t)msr, 15, 0);// Using undocumented MSR to get actual values
210}
211else if (p->CPU.Vendor == CPUID_VENDOR_AMD)
212{
213p->CPU.NoThreads= bitfield(p->CPU.CPUID[CPUID_1][1], 23, 16);
214p->CPU.NoCores= bitfield(p->CPU.CPUID[CPUID_88][2], 7, 0) + 1;
215}
216else
217{
218// Use previous method for Cores and Threads
219p->CPU.NoThreads= bitfield(p->CPU.CPUID[CPUID_1][1], 23, 16);
220p->CPU.NoCores= bitfield(p->CPU.CPUID[CPUID_4][0], 31, 26) + 1;
221}
222
223/* get brand string (if supported) */
224/* Copyright: from Apple's XNU cpuid.c */
225if (p->CPU.CPUID[CPUID_80][0] > 0x80000004) {
226uint32_treg[4];
227charstr[128], *s;
228/*
229 * The brand string 48 bytes (max), guaranteed to
230 * be NULL terminated.
231 */
232do_cpuid(0x80000002, reg);
233bcopy((char *)reg, &str[0], 16);
234do_cpuid(0x80000003, reg);
235bcopy((char *)reg, &str[16], 16);
236do_cpuid(0x80000004, reg);
237bcopy((char *)reg, &str[32], 16);
238for (s = str; *s != '\0'; s++) {
239if (*s != ' ') break;
240}
241
242strlcpy(p->CPU.BrandString, s, sizeof(p->CPU.BrandString));
243
244if (!strncmp(p->CPU.BrandString, CPU_STRING_UNKNOWN, MIN(sizeof(p->CPU.BrandString), strlen(CPU_STRING_UNKNOWN) + 1))) {
245/*
246 * This string means we have a firmware-programmable brand string,
247 * and the firmware couldn't figure out what sort of CPU we have.
248 */
249p->CPU.BrandString[0] = '\0';
250}
251}
252
253/* setup features */
254if ((bit(23) & p->CPU.CPUID[CPUID_1][3]) != 0) {
255p->CPU.Features |= CPU_FEATURE_MMX;
256}
257if ((bit(25) & p->CPU.CPUID[CPUID_1][3]) != 0) {
258p->CPU.Features |= CPU_FEATURE_SSE;
259}
260if ((bit(26) & p->CPU.CPUID[CPUID_1][3]) != 0) {
261p->CPU.Features |= CPU_FEATURE_SSE2;
262}
263if ((bit(0) & p->CPU.CPUID[CPUID_1][2]) != 0) {
264p->CPU.Features |= CPU_FEATURE_SSE3;
265}
266if ((bit(19) & p->CPU.CPUID[CPUID_1][2]) != 0) {
267p->CPU.Features |= CPU_FEATURE_SSE41;
268}
269if ((bit(20) & p->CPU.CPUID[CPUID_1][2]) != 0) {
270p->CPU.Features |= CPU_FEATURE_SSE42;
271}
272if ((bit(29) & p->CPU.CPUID[CPUID_81][3]) != 0) {
273p->CPU.Features |= CPU_FEATURE_EM64T;
274}
275if ((bit(5) & p->CPU.CPUID[CPUID_1][3]) != 0) {
276p->CPU.Features |= CPU_FEATURE_MSR;
277}
278//if ((bit(28) & p->CPU.CPUID[CPUID_1][3]) != 0) {
279if (p->CPU.NoThreads > p->CPU.NoCores) {
280p->CPU.Features |= CPU_FEATURE_HTT;
281}
282
283tscFrequency = measure_tsc_frequency();
284fsbFrequency = 0;
285cpuFrequency = 0;
286
287if ((p->CPU.Vendor == CPUID_VENDOR_INTEL) && ((p->CPU.Family == 0x06) || (p->CPU.Family == 0x0f))) {
288int intelCPU = p->CPU.Model;
289if ((p->CPU.Family == 0x06 && p->CPU.Model >= 0x0c) || (p->CPU.Family == 0x0f && p->CPU.Model >= 0x03)) {
290/* Nehalem CPU model */
291if (p->CPU.Family == 0x06 && (p->CPU.Model == CPU_MODEL_NEHALEM||
292 p->CPU.Model == CPU_MODEL_FIELDS||
293 p->CPU.Model == CPU_MODEL_DALES||
294 p->CPU.Model == CPU_MODEL_DALES_32NM||
295 p->CPU.Model == CPU_MODEL_WESTMERE||
296 p->CPU.Model == CPU_MODEL_NEHALEM_EX||
297 p->CPU.Model == CPU_MODEL_WESTMERE_EX ||
298 p->CPU.Model == CPU_MODEL_SANDYBRIDGE ||
299 p->CPU.Model == CPU_MODEL_JAKETOWN||
300 p->CPU.Model == CPU_MODEL_IVYBRIDGE)) {
301msr = rdmsr64(MSR_PLATFORM_INFO);
302DBG("msr(%d): platform_info %08x\n", __LINE__, bitfield(msr, 31, 0));
303bus_ratio_max = bitfield(msr, 14, 8);
304bus_ratio_min = bitfield(msr, 46, 40); //valv: not sure about this one (Remarq.1)
305msr = rdmsr64(MSR_FLEX_RATIO);
306DBG("msr(%d): flex_ratio %08x\n", __LINE__, bitfield(msr, 31, 0));
307if (bitfield(msr, 16, 16)) {
308flex_ratio = bitfield(msr, 14, 8);
309/* bcc9: at least on the gigabyte h67ma-ud2h,
310 where the cpu multipler can't be changed to
311 allow overclocking, the flex_ratio msr has unexpected (to OSX)
312 contents.These contents cause mach_kernel to
313 fail to compute the bus ratio correctly, instead
314 causing the system to crash since tscGranularity
315 is inadvertently set to 0.
316 */
317if (flex_ratio == 0) {
318/* Clear bit 16 (evidently the presence bit) */
319wrmsr64(MSR_FLEX_RATIO, (msr & 0xFFFFFFFFFFFEFFFFULL));
320msr = rdmsr64(MSR_FLEX_RATIO);
321verbose("Unusable flex ratio detected. Patched MSR now %08x\n", bitfield(msr, 31, 0));
322} else {
323if (bus_ratio_max > flex_ratio) {
324bus_ratio_max = flex_ratio;
325}
326}
327}
328
329if (bus_ratio_max) {
330fsbFrequency = (tscFrequency / bus_ratio_max);
331}
332//valv: Turbo Ratio Limit
333if ((intelCPU != 0x2e) && (intelCPU != 0x2f)) {
334msr = rdmsr64(MSR_TURBO_RATIO_LIMIT);
335cpuFrequency = bus_ratio_max * fsbFrequency;
336max_ratio = bus_ratio_max * 10;
337} else {
338cpuFrequency = tscFrequency;
339}
340if ((getValueForKey(kbusratio, &newratio, &len, &bootInfo->chameleonConfig)) && (len <= 4)) {
341max_ratio = atoi(newratio);
342max_ratio = (max_ratio * 10);
343if (len >= 3) max_ratio = (max_ratio + 5);
344
345verbose("Bus-Ratio: min=%d, max=%s\n", bus_ratio_min, newratio);
346
347// extreme overclockers may love 320 ;)
348if ((max_ratio >= min_ratio) && (max_ratio <= 320)) {
349cpuFrequency = (fsbFrequency * max_ratio) / 10;
350if (len >= 3) maxdiv = 1;
351else maxdiv = 0;
352} else {
353max_ratio = (bus_ratio_max * 10);
354}
355}
356//valv: to be uncommented if Remarq.1 didn't stick
357/*if (bus_ratio_max > 0) bus_ratio = flex_ratio;*/
358p->CPU.MaxRatio = max_ratio;
359p->CPU.MinRatio = min_ratio;
360
361myfsb = fsbFrequency / 1000000;
362verbose("Sticking with [BCLK: %dMhz, Bus-Ratio: %d]\n", myfsb, max_ratio);
363currcoef = bus_ratio_max;
364} else {
365msr = rdmsr64(MSR_IA32_PERF_STATUS);
366DBG("msr(%d): ia32_perf_stat 0x%08x\n", __LINE__, bitfield(msr, 31, 0));
367currcoef = bitfield(msr, 12, 8);
368/* Non-integer bus ratio for the max-multi*/
369maxdiv = bitfield(msr, 46, 46);
370/* Non-integer bus ratio for the current-multi (undocumented)*/
371currdiv = bitfield(msr, 14, 14);
372
373// This will always be model >= 3
374if ((p->CPU.Family == 0x06 && p->CPU.Model >= 0x0e) || (p->CPU.Family == 0x0f))
375{
376/* On these models, maxcoef defines TSC freq */
377maxcoef = bitfield(msr, 44, 40);
378} else {
379/* On lower models, currcoef defines TSC freq */
380/* XXX */
381maxcoef = currcoef;
382}
383
384if (maxcoef) {
385if (maxdiv) {
386fsbFrequency = ((tscFrequency * 2) / ((maxcoef * 2) + 1));
387} else {
388fsbFrequency = (tscFrequency / maxcoef);
389}
390if (currdiv) {
391cpuFrequency = (fsbFrequency * ((currcoef * 2) + 1) / 2);
392} else {
393cpuFrequency = (fsbFrequency * currcoef);
394}
395DBG("max: %d%s current: %d%s\n", maxcoef, maxdiv ? ".5" : "",currcoef, currdiv ? ".5" : "");
396}
397}
398}
399/* Mobile CPU */
400if (rdmsr64(MSR_IA32_PLATFORM_ID) & (1<<28)) {
401p->CPU.Features |= CPU_FEATURE_MOBILE;
402}
403}
404else if ((p->CPU.Vendor == CPUID_VENDOR_AMD) && (p->CPU.Family == 0x0f))
405{
406switch(p->CPU.ExtFamily)
407{
408case 0x00: /* K8 */
409msr = rdmsr64(K8_FIDVID_STATUS);
410maxcoef = bitfield(msr, 21, 16) / 2 + 4;
411currcoef = bitfield(msr, 5, 0) / 2 + 4;
412break;
413
414case 0x01: /* K10 */
415msr = rdmsr64(K10_COFVID_STATUS);
416do_cpuid2(0x00000006, 0, p->CPU.CPUID[CPUID_6]);
417// EffFreq: effective frequency interface
418if (bitfield(p->CPU.CPUID[CPUID_6][2], 0, 0) == 1)
419{
420//uint64_t mperf = measure_mperf_frequency();
421uint64_t aperf = measure_aperf_frequency();
422cpuFrequency = aperf;
423}
424// NOTE: tsc runs at the maccoeff (non turbo)
425//*not* at the turbo frequency.
426maxcoef = bitfield(msr, 54, 49) / 2 + 4;
427currcoef = bitfield(msr, 5, 0) + 0x10;
428currdiv = 2 << bitfield(msr, 8, 6);
429
430break;
431
432case 0x05: /* K14 */
433msr = rdmsr64(K10_COFVID_STATUS);
434currcoef = (bitfield(msr, 54, 49) + 0x10) << 2;
435currdiv = (bitfield(msr, 8, 4) + 1) << 2;
436currdiv += bitfield(msr, 3, 0);
437
438break;
439
440case 0x02: /* K11 */
441// not implimented
442break;
443}
444
445if (maxcoef)
446{
447if (currdiv)
448{
449if (!currcoef) currcoef = maxcoef;
450if (!cpuFrequency)
451fsbFrequency = ((tscFrequency * currdiv) / currcoef);
452else
453fsbFrequency = ((cpuFrequency * currdiv) / currcoef);
454
455DBG("%d.%d\n", currcoef / currdiv, ((currcoef % currdiv) * 100) / currdiv);
456} else {
457if (!cpuFrequency)
458fsbFrequency = (tscFrequency / maxcoef);
459else
460fsbFrequency = (cpuFrequency / maxcoef);
461DBG("%d\n", currcoef);
462}
463}
464else if (currcoef)
465{
466if (currdiv)
467{
468fsbFrequency = ((tscFrequency * currdiv) / currcoef);
469DBG("%d.%d\n", currcoef / currdiv, ((currcoef % currdiv) * 100) / currdiv);
470} else {
471fsbFrequency = (tscFrequency / currcoef);
472DBG("%d\n", currcoef);
473}
474}
475if (!cpuFrequency) cpuFrequency = tscFrequency;
476}
477
478#if 0
479if (!fsbFrequency) {
480fsbFrequency = (DEFAULT_FSB * 1000);
481cpuFrequency = tscFrequency;
482DBG("0 ! using the default value for FSB !\n");
483}
484#endif
485
486p->CPU.MaxCoef = maxcoef;
487p->CPU.MaxDiv = maxdiv;
488p->CPU.CurrCoef = currcoef;
489p->CPU.CurrDiv = currdiv;
490p->CPU.TSCFrequency = tscFrequency;
491p->CPU.FSBFrequency = fsbFrequency;
492p->CPU.CPUFrequency = cpuFrequency;
493
494// keep formatted with spaces instead of tabs
495DBG("CPU: Brand String: %s\n", p->CPU.BrandString);
496 DBG("CPU: Vendor/Family/ExtFamily: 0x%x/0x%x/0x%x\n", p->CPU.Vendor, p->CPU.Family, p->CPU.ExtFamily);
497 DBG("CPU: Model/ExtModel/Stepping: 0x%x/0x%x/0x%x\n", p->CPU.Model, p->CPU.ExtModel, p->CPU.Stepping);
498 DBG("CPU: MaxCoef/CurrCoef: 0x%x/0x%x\n", p->CPU.MaxCoef, p->CPU.CurrCoef);
499 DBG("CPU: MaxDiv/CurrDiv: 0x%x/0x%x\n", p->CPU.MaxDiv, p->CPU.CurrDiv);
500 DBG("CPU: TSCFreq: %dMHz\n", p->CPU.TSCFrequency / 1000000);
501 DBG("CPU: FSBFreq: %dMHz\n", p->CPU.FSBFrequency / 1000000);
502 DBG("CPU: CPUFreq: %dMHz\n", p->CPU.CPUFrequency / 1000000);
503 DBG("CPU: NoCores/NoThreads: %d/%d\n", p->CPU.NoCores, p->CPU.NoThreads);
504 DBG("CPU: Features: 0x%08x\n", p->CPU.Features);
505#if DEBUG_CPU
506pause();
507#endif
508}
509

Archive Download this file

Revision: 1918