Chameleon

Chameleon Svn Source Tree

Root/branches/Chimera/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 = 0;
154uint64_tfsbFrequency = 0;
155uint64_tcpuFrequency =0;
156uint64_tmsr = 0;
157uint64_tflex_ratio = 0;
158uint32_tmax_ratio = 0;
159uint32_tmin_ratio = 0;
160uint8_tbus_ratio_max = 0;
161uint8_tbus_ratio_min = 0;
162uint8_tcurrdiv = 0;
163uint8_tcurrcoef = 0;
164uint8_tmaxdiv = 0;
165uint8_tmaxcoef = 0;
166
167const char*newratio;
168intlen = 0;
169
170/* get cpuid values */
171do_cpuid(0x00000000, p->CPU.CPUID[CPUID_0]);
172do_cpuid(0x00000001, p->CPU.CPUID[CPUID_1]);
173do_cpuid(0x00000002, p->CPU.CPUID[CPUID_2]);
174do_cpuid(0x00000003, p->CPU.CPUID[CPUID_3]);
175do_cpuid2(0x00000004, 0, p->CPU.CPUID[CPUID_4]);
176do_cpuid(0x80000000, p->CPU.CPUID[CPUID_80]);
177if ((p->CPU.CPUID[CPUID_80][0] & 0x0000000f) >= 8) {
178do_cpuid(0x80000008, p->CPU.CPUID[CPUID_88]);
179do_cpuid(0x80000001, p->CPU.CPUID[CPUID_81]);
180}
181else if ((p->CPU.CPUID[CPUID_80][0] & 0x0000000f) >= 1) {
182do_cpuid(0x80000001, p->CPU.CPUID[CPUID_81]);
183}
184
185#if DEBUG_CPU
186{
187inti;
188printf("CPUID Raw Values:\n");
189for (i=0; i<CPUID_MAX; i++) {
190printf("%02d: %08x-%08x-%08x-%08x\n", i,
191 p->CPU.CPUID[i][0], p->CPU.CPUID[i][1],
192 p->CPU.CPUID[i][2], p->CPU.CPUID[i][3]);
193}
194}
195#endif
196
197p->CPU.Vendor= p->CPU.CPUID[CPUID_0][1];
198p->CPU.Signature= p->CPU.CPUID[CPUID_1][0];
199p->CPU.Stepping= bitfield(p->CPU.CPUID[CPUID_1][0], 3, 0);
200p->CPU.Model= bitfield(p->CPU.CPUID[CPUID_1][0], 7, 4);
201p->CPU.Family= bitfield(p->CPU.CPUID[CPUID_1][0], 11, 8);
202p->CPU.ExtModel= bitfield(p->CPU.CPUID[CPUID_1][0], 19, 16);
203p->CPU.ExtFamily= bitfield(p->CPU.CPUID[CPUID_1][0], 27, 20);
204
205p->CPU.Model += (p->CPU.ExtModel << 4);
206
207if (p->CPU.Vendor == CPUID_VENDOR_INTEL &&
208p->CPU.Family == 0x06 &&
209p->CPU.Model >= CPU_MODEL_NEHALEM &&
210p->CPU.Model != CPU_MODEL_ATOM// MSR is *NOT* available on the Intel Atom CPU
211)
212{
213msr = rdmsr64(MSR_CORE_THREAD_COUNT);// Undocumented MSR in Nehalem and newer CPUs
214p->CPU.NoCores= bitfield((uint32_t)msr, 31, 16);// Using undocumented MSR to get actual values
215p->CPU.NoThreads= bitfield((uint32_t)msr, 15, 0);// Using undocumented MSR to get actual values
216}
217else if (p->CPU.Vendor == CPUID_VENDOR_AMD)
218{
219p->CPU.NoThreads= bitfield(p->CPU.CPUID[CPUID_1][1], 23, 16);
220p->CPU.NoCores= bitfield(p->CPU.CPUID[CPUID_88][2], 7, 0) + 1;
221}
222else
223{
224// Use previous method for Cores and Threads
225p->CPU.NoThreads= bitfield(p->CPU.CPUID[CPUID_1][1], 23, 16);
226p->CPU.NoCores= bitfield(p->CPU.CPUID[CPUID_4][0], 31, 26) + 1;
227}
228
229/* get brand string (if supported) */
230/* Copyright: from Apple's XNU cpuid.c */
231if (p->CPU.CPUID[CPUID_80][0] > 0x80000004) {
232uint32_treg[4];
233charstr[128], *s;
234/*
235 * The brand string 48 bytes (max), guaranteed to
236 * be NULL terminated.
237 */
238do_cpuid(0x80000002, reg);
239bcopy((char *)reg, &str[0], 16);
240do_cpuid(0x80000003, reg);
241bcopy((char *)reg, &str[16], 16);
242do_cpuid(0x80000004, reg);
243bcopy((char *)reg, &str[32], 16);
244for (s = str; *s != '\0'; s++) {
245if (*s != ' ') break;
246}
247
248strlcpy(p->CPU.BrandString, s, sizeof(p->CPU.BrandString));
249
250if (!strncmp(p->CPU.BrandString, CPU_STRING_UNKNOWN, MIN(sizeof(p->CPU.BrandString), strlen(CPU_STRING_UNKNOWN) + 1))) {
251/*
252 * This string means we have a firmware-programmable brand string,
253 * and the firmware couldn't figure out what sort of CPU we have.
254 */
255p->CPU.BrandString[0] = '\0';
256}
257}
258
259/* setup features */
260if ((bit(23) & p->CPU.CPUID[CPUID_1][3]) != 0) {
261p->CPU.Features |= CPU_FEATURE_MMX;
262}
263if ((bit(25) & p->CPU.CPUID[CPUID_1][3]) != 0) {
264p->CPU.Features |= CPU_FEATURE_SSE;
265}
266if ((bit(26) & p->CPU.CPUID[CPUID_1][3]) != 0) {
267p->CPU.Features |= CPU_FEATURE_SSE2;
268}
269if ((bit(0) & p->CPU.CPUID[CPUID_1][2]) != 0) {
270p->CPU.Features |= CPU_FEATURE_SSE3;
271}
272if ((bit(19) & p->CPU.CPUID[CPUID_1][2]) != 0) {
273p->CPU.Features |= CPU_FEATURE_SSE41;
274}
275if ((bit(20) & p->CPU.CPUID[CPUID_1][2]) != 0) {
276p->CPU.Features |= CPU_FEATURE_SSE42;
277}
278if ((bit(29) & p->CPU.CPUID[CPUID_81][3]) != 0) {
279p->CPU.Features |= CPU_FEATURE_EM64T;
280}
281if ((bit(5) & p->CPU.CPUID[CPUID_1][3]) != 0) {
282p->CPU.Features |= CPU_FEATURE_MSR;
283}
284//if ((bit(28) & p->CPU.CPUID[CPUID_1][3]) != 0) {
285if (p->CPU.NoThreads > p->CPU.NoCores) {
286p->CPU.Features |= CPU_FEATURE_HTT;
287}
288
289tscFrequency = measure_tsc_frequency();
290
291if ((p->CPU.Vendor == CPUID_VENDOR_INTEL) && ((p->CPU.Family == 0x06) || (p->CPU.Family == 0x0f))) {
292int intelCPU = p->CPU.Model;
293if ((p->CPU.Family == 0x06 && p->CPU.Model >= 0x0c) || (p->CPU.Family == 0x0f && p->CPU.Model >= 0x03)) {
294/* Nehalem CPU model */
295if (p->CPU.Family == 0x06 && (p->CPU.Model == CPU_MODEL_NEHALEM||
296 p->CPU.Model == CPU_MODEL_FIELDS||
297 p->CPU.Model == CPU_MODEL_DALES||
298 p->CPU.Model == CPU_MODEL_DALES_32NM||
299 p->CPU.Model == CPU_MODEL_WESTMERE||
300 p->CPU.Model == CPU_MODEL_NEHALEM_EX||
301 p->CPU.Model == CPU_MODEL_WESTMERE_EX ||
302 p->CPU.Model == CPU_MODEL_SANDYBRIDGE ||
303 p->CPU.Model == CPU_MODEL_JAKETOWN||
304 p->CPU.Model == CPU_MODEL_IVYBRIDGE)) {
305msr = rdmsr64(MSR_PLATFORM_INFO);
306//DBG("MSR_PLATFORM_INFO: 0x%x\n", bitfield(msr, 63, 0));
307bus_ratio_max = bitfield(msr, 15, 8);//MacMan: Changed bitfield to match Apple tsc.c
308bus_ratio_min = bitfield(msr, 47, 40);//MacMan: Changed bitfield to match Apple tsc.c
309msr = rdmsr64(MSR_FLEX_RATIO);
310//DBG("MSR_FLEX_RATIO: 0x%x\n", bitfield(msr, 31, 0));
311if (bitfield(msr, 16, 16)) {
312flex_ratio = bitfield(msr, 15, 8);//MacMan: Changed bitfield to match Apple tsc.c
313if (flex_ratio == 0) {
314/* Clear bit 16 (evidently the presence bit) */
315wrmsr64(MSR_FLEX_RATIO, (msr & 0xFFFFFFFFFFFEFFFFULL));
316msr = rdmsr64(MSR_FLEX_RATIO);
317//DBG("Null flex ratio detected. MSR_FLEX_RATIO now: %08x\n", bitfield(msr, 31, 0));
318} else {
319if (bus_ratio_max > flex_ratio) {
320bus_ratio_max = flex_ratio;
321}
322}
323}
324
325if (bus_ratio_max) {
326fsbFrequency = (tscFrequency / bus_ratio_max);
327}
328
329//MacMan: Turbo Ratio Limit
330switch (intelCPU)
331{
332case CPU_MODEL_WESTMERE_EX:// Intel Xeon E7
333case CPU_MODEL_NEHALEM_EX:// Intel Xeon X75xx, Xeon X65xx, Xeon E75xx, Xeon E65xx
334{
335cpuFrequency = tscFrequency;
336DBG("cpu.c (%d)CPU_MODEL_NEHALEM_EX or CPU_MODEL_WESTMERE_EX Found\n", __LINE__);
337break;
338}
339case CPU_MODEL_SANDYBRIDGE:// Intel Core i3, i5, i7 LGA1155 (32nm)
340case CPU_MODEL_IVYBRIDGE:// Intel Core i3, i5, i7 LGA1155 (22nm)
341case CPU_MODEL_JAKETOWN:// Intel Core i7, Xeon E5 LGA2011 (32nm)
342{
343msr = rdmsr64(MSR_IA32_PERF_STATUS);
344currcoef = bitfield(msr, 15, 8);
345cpuFrequency = currcoef * fsbFrequency;
346maxcoef = bus_ratio_max;
347break;
348}
349default:
350{
351msr = rdmsr64(MSR_IA32_PERF_STATUS);
352currcoef = bitfield(msr, 7, 0);
353cpuFrequency = currcoef * fsbFrequency;
354maxcoef = bus_ratio_max;
355break;
356}
357}
358
359if ((getValueForKey(kbusratio, &newratio, &len, &bootInfo->chameleonConfig)) && (len <= 4)) {
360max_ratio = atoi(newratio);
361max_ratio = (max_ratio * 10);
362if (len >= 3) max_ratio = (max_ratio + 5);
363
364DBG("Bus-Ratio: min=%d, max=%s\n", bus_ratio_min, newratio);
365
366// extreme overclockers may love 320 ;)
367if ((max_ratio >= min_ratio) && (max_ratio <= 320)) {
368cpuFrequency = (fsbFrequency * max_ratio) / 10;
369if (len >= 3) maxdiv = 1;
370else maxdiv = 0;
371} else {
372max_ratio = (bus_ratio_max * 10);
373}
374}
375
376p->CPU.MaxRatio = bus_ratio_max;
377p->CPU.MinRatio = bus_ratio_min;
378
379} else {
380msr = rdmsr64(MSR_IA32_PERF_STATUS);
381DBG("MSR_IA32_PERF_STATUS: 0x%08x\n", bitfield(msr, 31, 0));
382currcoef = bitfield(msr, 12, 8);
383/* Non-integer bus ratio for the max-multi*/
384maxdiv = bitfield(msr, 46, 46);
385/* Non-integer bus ratio for the current-multi (undocumented)*/
386currdiv = bitfield(msr, 14, 14);
387
388// This will always be model >= 3
389if ((p->CPU.Family == 0x06 && p->CPU.Model >= 0x0e) || (p->CPU.Family == 0x0f))
390{
391/* On these models, maxcoef defines TSC freq */
392maxcoef = bitfield(msr, 44, 40);
393} else {
394/* On lower models, currcoef defines TSC freq */
395/* XXX */
396maxcoef = currcoef;
397}
398
399if (maxcoef) {
400if (maxdiv) {
401fsbFrequency = ((tscFrequency * 2) / ((maxcoef * 2) + 1));
402} else {
403fsbFrequency = (tscFrequency / maxcoef);
404}
405if (currdiv) {
406cpuFrequency = (fsbFrequency * ((currcoef * 2) + 1) / 2);
407} else {
408cpuFrequency = (fsbFrequency * currcoef);
409}
410DBG("max: %d%s current: %d%s\n", maxcoef, maxdiv ? ".5" : "",currcoef, currdiv ? ".5" : "");
411}
412}
413}
414/* Mobile CPU */
415if (rdmsr64(MSR_IA32_PLATFORM_ID) & (1<<28)) {
416p->CPU.Features |= CPU_FEATURE_MOBILE;
417}
418}
419else if ((p->CPU.Vendor == CPUID_VENDOR_AMD) && (p->CPU.Family == 0x0f))
420{
421switch(p->CPU.ExtFamily)
422{
423case 0x00: /* K8 */
424msr = rdmsr64(K8_FIDVID_STATUS);
425maxcoef = bitfield(msr, 21, 16) / 2 + 4;
426currcoef = bitfield(msr, 5, 0) / 2 + 4;
427break;
428
429case 0x01: /* K10 */
430msr = rdmsr64(K10_COFVID_STATUS);
431do_cpuid2(0x00000006, 0, p->CPU.CPUID[CPUID_6]);
432// EffFreq: effective frequency interface
433if (bitfield(p->CPU.CPUID[CPUID_6][2], 0, 0) == 1)
434{
435//uint64_t mperf = measure_mperf_frequency();
436uint64_t aperf = measure_aperf_frequency();
437cpuFrequency = aperf;
438}
439// NOTE: tsc runs at the maccoeff (non turbo)
440//*not* at the turbo frequency.
441maxcoef = bitfield(msr, 54, 49) / 2 + 4;
442currcoef = bitfield(msr, 5, 0) + 0x10;
443currdiv = 2 << bitfield(msr, 8, 6);
444
445break;
446
447case 0x05: /* K14 */
448msr = rdmsr64(K10_COFVID_STATUS);
449currcoef = (bitfield(msr, 54, 49) + 0x10) << 2;
450currdiv = (bitfield(msr, 8, 4) + 1) << 2;
451currdiv += bitfield(msr, 3, 0);
452
453break;
454
455case 0x02: /* K11 */
456// not implimented
457break;
458}
459
460if (maxcoef)
461{
462if (currdiv)
463{
464if (!currcoef) currcoef = maxcoef;
465if (!cpuFrequency)
466fsbFrequency = ((tscFrequency * currdiv) / currcoef);
467else
468fsbFrequency = ((cpuFrequency * currdiv) / currcoef);
469
470DBG("%d.%d\n", currcoef / currdiv, ((currcoef % currdiv) * 100) / currdiv);
471} else {
472if (!cpuFrequency)
473fsbFrequency = (tscFrequency / maxcoef);
474else
475fsbFrequency = (cpuFrequency / maxcoef);
476DBG("%d\n", currcoef);
477}
478}
479else if (currcoef)
480{
481if (currdiv)
482{
483fsbFrequency = ((tscFrequency * currdiv) / currcoef);
484DBG("%d.%d\n", currcoef / currdiv, ((currcoef % currdiv) * 100) / currdiv);
485} else {
486fsbFrequency = (tscFrequency / currcoef);
487DBG("%d\n", currcoef);
488}
489}
490if (!cpuFrequency) cpuFrequency = tscFrequency;
491}
492
493#if 0
494if (!fsbFrequency) {
495fsbFrequency = (DEFAULT_FSB * 1000);
496cpuFrequency = tscFrequency;
497DBG("0 ! using the default value for FSB !\n");
498}
499#endif
500
501p->CPU.MaxCoef = maxcoef;
502if (maxdiv == 0){
503p->CPU.MaxDiv = bus_ratio_max;
504}
505else {
506p->CPU.MaxDiv = maxdiv;
507}
508p->CPU.CurrCoef = currcoef;
509if (currdiv == 0){
510p->CPU.CurrDiv = currcoef;
511}
512else {
513p->CPU.CurrDiv = currdiv;
514}
515p->CPU.TSCFrequency = tscFrequency;
516p->CPU.FSBFrequency = fsbFrequency;
517p->CPU.CPUFrequency = cpuFrequency;
518
519// keep formatted with spaces instead of tabs
520DBG("CPU: Brand String: %s\n", p->CPU.BrandString);
521 DBG("CPU: Vendor: 0x%x\n", p->CPU.Vendor);
522DBG("CPU: Family / ExtFamily: 0x%x / 0x%x\n", p->CPU.Family, p->CPU.ExtFamily);
523 DBG("CPU: Model / ExtModel / Stepping: 0x%x / 0x%x / 0x%x\n", p->CPU.Model, p->CPU.ExtModel, p->CPU.Stepping);
524 DBG("CPU: Number of Cores / Threads: %d / %d\n", p->CPU.NoCores, p->CPU.NoThreads);
525 DBG("CPU: Features: 0x%08x\n", p->CPU.Features);
526 DBG("CPU: TSC Frequency: %d MHz\n", p->CPU.TSCFrequency / 1000000);
527 DBG("CPU: FSB Frequency: %d MHz\n", p->CPU.FSBFrequency / 1000000);
528 DBG("CPU: CPU Frequency: %d MHz\n", p->CPU.CPUFrequency / 1000000);
529DBG("CPU: Minimum Bus Ratio: %d\n", p->CPU.MinRatio);
530 DBG("CPU: Maximum Bus Ratio: %d\n", p->CPU.MaxRatio);
531DBG("CPU: Current Bus Ratio: %d\n", p->CPU.CurrCoef);
532//DBG("CPU: Maximum Multiplier: %d\n", p->CPU.MaxCoef);
533// DBG("CPU: Maximum Divider: %d\n", p->CPU.MaxDiv);
534//DBG("CPU: Current Divider: %d\n", p->CPU.CurrDiv);
535
536#if DEBUG_CPU
537pause();
538#endif
539}
540

Archive Download this file

Revision: 1930