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 | */␊ |
28 | uint64_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 | ␊ |
56 | restart:␊ |
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 | */␊ |
101 | static 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 | ␉␉{␊ |
129 | ␉␉␉continue;␊ |
130 | ␉␉}␊ |
131 | ␉␉/* The TSC must increment at LEAST once every millisecond.␊ |
132 | ␉␉ * We should have waited exactly 30 msec so the TSC delta should␊ |
133 | ␉␉ * be >= 30. Anything less and the processor is way too slow.␊ |
134 | ␉␉ */␊ |
135 | ␉␉if ((tscEnd - tscStart) <= CALIBRATE_TIME_MSEC)␊ |
136 | ␉␉{␊ |
137 | ␉␉␉continue;␊ |
138 | ␉␉}␊ |
139 | ␉␉// tscDelta = MIN(tscDelta, (tscEnd - tscStart))␊ |
140 | ␉␉if ( (tscEnd - tscStart) < tscDelta )␊ |
141 | ␉␉{␊ |
142 | ␉␉␉tscDelta = tscEnd - tscStart;␊ |
143 | ␉␉}␊ |
144 | ␉}␊ |
145 | ␉/* tscDelta is now the least number of TSC ticks the processor made in␊ |
146 | ␉ * a timespan of 0.03 s (e.g. 30 milliseconds)␊ |
147 | ␉ * Linux thus divides by 30 which gives the answer in kiloHertz because␊ |
148 | ␉ * 1 / ms = kHz. But we're xnu and most of the rest of the code uses␊ |
149 | ␉ * Hz so we need to convert our milliseconds to seconds. Since we're␊ |
150 | ␉ * dividing by the milliseconds, we simply multiply by 1000.␊ |
151 | ␉ */␊ |
152 | ␊ |
153 | ␉/* Unlike linux, we're not limited to 32-bit, but we do need to take care␊ |
154 | ␉ * that we're going to multiply by 1000 first so we do need at least some␊ |
155 | ␉ * arithmetic headroom. For now, 32-bit should be enough.␊ |
156 | ␉ * Also unlike Linux, our compiler can do 64-bit integer arithmetic.␊ |
157 | ␉ */␊ |
158 | ␉if (tscDelta > (1ULL<<32))␊ |
159 | ␉{␊ |
160 | ␉␉retval = 0;␊ |
161 | ␉}␊ |
162 | ␉else␊ |
163 | ␉{␊ |
164 | ␉␉retval = tscDelta * 1000 / 30;␊ |
165 | ␉}␊ |
166 | ␉disable_PIT2();␊ |
167 | ␉return retval;␊ |
168 | }␊ |
169 | ␊ |
170 | /*␊ |
171 | * Original comment/code:␊ |
172 | * "DFE: Measures the Max Performance Frequency in Hz (64-bit)"␊ |
173 | *␊ |
174 | * Measures the Actual Performance Frequency in Hz (64-bit)␊ |
175 | * (just a naming change, mperf --> aperf )␊ |
176 | */␊ |
177 | static uint64_t measure_aperf_frequency(void)␊ |
178 | {␊ |
179 | ␉uint64_t aperfStart;␊ |
180 | ␉uint64_t aperfEnd;␊ |
181 | ␉uint64_t aperfDelta = 0xffffffffffffffffULL;␊ |
182 | ␉unsigned long pollCount;␊ |
183 | ␉uint64_t retval = 0;␊ |
184 | ␉int i;␊ |
185 | ␉␊ |
186 | ␉/* Time how many APERF ticks elapse in 30 msec using the 8254 PIT␊ |
187 | ␉ * counter 2. We run this loop 3 times to make sure the cache␊ |
188 | ␉ * is hot and we take the minimum delta from all of the runs.␊ |
189 | ␉ * That is to say that we're biased towards measuring the minimum␊ |
190 | ␉ * number of APERF ticks that occur while waiting for the timer to␊ |
191 | ␉ * expire.␊ |
192 | ␉ */␊ |
193 | ␉for(i = 0; i < 10; ++i)␊ |
194 | ␉{␊ |
195 | ␉␉enable_PIT2();␊ |
196 | ␉␉set_PIT2_mode0(CALIBRATE_LATCH);␊ |
197 | ␉␉aperfStart = rdmsr64(MSR_AMD_APERF);␊ |
198 | ␉␉pollCount = poll_PIT2_gate();␊ |
199 | ␉␉aperfEnd = rdmsr64(MSR_AMD_APERF);␊ |
200 | ␉␉/* The poll loop must have run at least a few times for accuracy */␊ |
201 | ␉␉if (pollCount <= 1)␊ |
202 | ␉␉{␊ |
203 | ␉␉␉continue;␊ |
204 | ␉␉}␊ |
205 | ␉␉/* The TSC must increment at LEAST once every millisecond.␊ |
206 | ␉␉ * We should have waited exactly 30 msec so the APERF delta should␊ |
207 | ␉␉ * be >= 30. Anything less and the processor is way too slow.␊ |
208 | ␉␉ */␊ |
209 | ␉␉if ((aperfEnd - aperfStart) <= CALIBRATE_TIME_MSEC)␊ |
210 | ␉␉{␊ |
211 | ␉␉␉continue;␊ |
212 | ␉␉}␊ |
213 | ␉␉// tscDelta = MIN(tscDelta, (tscEnd - tscStart))␊ |
214 | ␉␉if ( (aperfEnd - aperfStart) < aperfDelta )␊ |
215 | ␉␉{␊ |
216 | ␉␉␉aperfDelta = aperfEnd - aperfStart;␊ |
217 | ␉␉}␊ |
218 | ␉}␊ |
219 | ␉/* mperfDelta is now the least number of MPERF ticks the processor made in␊ |
220 | ␉ * a timespan of 0.03 s (e.g. 30 milliseconds)␊ |
221 | ␉ */␊ |
222 | ␉␊ |
223 | ␉if (aperfDelta > (1ULL<<32))␊ |
224 | ␉{␊ |
225 | ␉␉retval = 0;␊ |
226 | ␉}␊ |
227 | ␉else␊ |
228 | ␉{␊ |
229 | ␉␉retval = aperfDelta * 1000 / 30;␊ |
230 | ␉}␊ |
231 | ␉disable_PIT2();␊ |
232 | ␉return retval;␊ |
233 | }␊ |
234 | ␊ |
235 | /*␊ |
236 | * Calculates the FSB and CPU frequencies using specific MSRs for each CPU␊ |
237 | * - multi. is read from a specific MSR. In the case of Intel, there is:␊ |
238 | *␉ a max multi. (used to calculate the FSB freq.),␊ |
239 | *␉ and a current multi. (used to calculate the CPU freq.)␊ |
240 | * - fsbFrequency = tscFrequency / multi␊ |
241 | * - cpuFrequency = fsbFrequency * multi␊ |
242 | */␊ |
243 | void scan_cpu(PlatformInfo_t *p)␊ |
244 | {␊ |
245 | ␉uint64_t␉tscFrequency, fsbFrequency, cpuFrequency;␊ |
246 | ␉uint64_t␉msr, flex_ratio;␊ |
247 | ␉uint8_t␉␉maxcoef, maxdiv, currcoef, bus_ratio_max, currdiv;␊ |
248 | ␉const char␉*newratio;␊ |
249 | ␉int␉␉␉len, myfsb;␊ |
250 | ␉uint8_t␉␉bus_ratio_min;␊ |
251 | ␉uint32_t␉max_ratio, min_ratio;␊ |
252 | ␊ |
253 | ␉max_ratio = min_ratio = myfsb = bus_ratio_min = 0;␊ |
254 | ␉maxcoef = maxdiv = bus_ratio_max = currcoef = currdiv = 0;␊ |
255 | ␊ |
256 | ␉/* get cpuid values */␊ |
257 | ␉do_cpuid(0x00000000, p->CPU.CPUID[CPUID_0]);␊ |
258 | ␉do_cpuid(0x00000001, p->CPU.CPUID[CPUID_1]);␊ |
259 | ␉do_cpuid(0x00000002, p->CPU.CPUID[CPUID_2]);␊ |
260 | ␉do_cpuid(0x00000003, p->CPU.CPUID[CPUID_3]);␊ |
261 | ␉do_cpuid2(0x00000004, 0, p->CPU.CPUID[CPUID_4]);␊ |
262 | ␉do_cpuid(0x80000000, p->CPU.CPUID[CPUID_80]);␊ |
263 | ␉if (p->CPU.CPUID[CPUID_0][0] >= 0x5)␊ |
264 | ␉{␊ |
265 | ␉␉do_cpuid(5, p->CPU.CPUID[CPUID_5]);␊ |
266 | ␉}␊ |
267 | ␉if (p->CPU.CPUID[CPUID_0][0] >= 6)␊ |
268 | ␉{␊ |
269 | ␉␉do_cpuid(6, p->CPU.CPUID[CPUID_6]);␊ |
270 | ␉}␊ |
271 | ␉if ((p->CPU.CPUID[CPUID_80][0] & 0x0000000f) >= 8)␊ |
272 | ␉{␊ |
273 | ␉␉do_cpuid(0x80000008, p->CPU.CPUID[CPUID_88]);␊ |
274 | ␉␉do_cpuid(0x80000001, p->CPU.CPUID[CPUID_81]);␊ |
275 | ␉}␊ |
276 | ␉else if ((p->CPU.CPUID[CPUID_80][0] & 0x0000000f) >= 1)␊ |
277 | ␉{␊ |
278 | ␉␉do_cpuid(0x80000001, p->CPU.CPUID[CPUID_81]);␊ |
279 | ␉}␊ |
280 | ␊ |
281 | #if DEBUG_CPU␊ |
282 | ␉{␊ |
283 | ␉␉int␉␉i;␊ |
284 | ␉␉printf("CPUID Raw Values:\n");␊ |
285 | ␉␉for (i=0; i<CPUID_MAX; i++)␊ |
286 | ␉␉{␊ |
287 | ␉␉␉printf("%02d: %08x-%08x-%08x-%08x\n", i,␊ |
288 | ␉␉␉␉ p->CPU.CPUID[i][0], p->CPU.CPUID[i][1],␊ |
289 | ␉␉␉␉ p->CPU.CPUID[i][2], p->CPU.CPUID[i][3]);␊ |
290 | ␉␉}␊ |
291 | ␉}␊ |
292 | #endif␊ |
293 | ␊ |
294 | /*␊ |
295 | EAX (Intel):␊ |
296 | 31 28 27 20 19 16 1514 1312 11 8 7 4 3 0␊ |
297 | +--------+----------------+--------+----+----+--------+--------+--------+␊ |
298 | |########|Extended family |Extmodel|####|type|familyid| model |stepping|␊ |
299 | +--------+----------------+--------+----+----+--------+--------+--------+␊ |
300 | ␊ |
301 | EAX (AMD):␊ |
302 | 31 28 27 20 19 16 1514 1312 11 8 7 4 3 0␊ |
303 | +--------+----------------+--------+----+----+--------+--------+--------+␊ |
304 | |########|Extended family |Extmodel|####|####|familyid| model |stepping|␊ |
305 | +--------+----------------+--------+----+----+--------+--------+--------+␊ |
306 | */␊ |
307 | ␊ |
308 | ␉p->CPU.Vendor␉␉= p->CPU.CPUID[CPUID_0][1];␊ |
309 | ␉p->CPU.Signature␉= p->CPU.CPUID[CPUID_1][0];␊ |
310 | ␉p->CPU.Stepping␉␉= bitfield(p->CPU.CPUID[CPUID_1][0], 3, 0);␊ |
311 | ␉p->CPU.Model␉␉= bitfield(p->CPU.CPUID[CPUID_1][0], 7, 4);␊ |
312 | ␉p->CPU.Family␉␉= bitfield(p->CPU.CPUID[CPUID_1][0], 11, 8);␊ |
313 | ␉p->CPU.Type␉ = bitfield(p->CPU.CPUID[CPUID_1][0], 13, 12);␊ |
314 | ␉p->CPU.ExtModel␉␉= bitfield(p->CPU.CPUID[CPUID_1][0], 19, 16);␊ |
315 | ␉p->CPU.ExtFamily␉= bitfield(p->CPU.CPUID[CPUID_1][0], 27, 20);␊ |
316 | ␊ |
317 | ␉p->CPU.Model += (p->CPU.ExtModel << 4);␊ |
318 | ␊ |
319 | ␉if (p->CPU.Vendor == CPUID_VENDOR_INTEL &&␊ |
320 | ␉␉p->CPU.Family == 0x06 &&␊ |
321 | ␉␉p->CPU.Model >= CPU_MODEL_NEHALEM &&␊ |
322 | ␉␉p->CPU.Model != CPU_MODEL_ATOM␉␉// MSR is *NOT* available on the Intel Atom CPU␊ |
323 | ␉␉)␊ |
324 | ␉{␊ |
325 | ␉␉msr = rdmsr64(MSR_CORE_THREAD_COUNT);␉␉␉␉␉// Undocumented MSR in Nehalem and newer CPUs␊ |
326 | ␉␉p->CPU.NoCores␉␉= bitfield((uint32_t)msr, 31, 16);␉// Using undocumented MSR to get actual values␊ |
327 | ␉␉p->CPU.NoThreads␉= bitfield((uint32_t)msr, 15, 0);␉// Using undocumented MSR to get actual values␊ |
328 | ␉}␊ |
329 | ␉else if (p->CPU.Vendor == CPUID_VENDOR_AMD)␊ |
330 | ␉{␊ |
331 | ␉␉p->CPU.NoThreads␉= bitfield(p->CPU.CPUID[CPUID_1][1], 23, 16);␊ |
332 | ␉␉p->CPU.NoCores␉␉= bitfield(p->CPU.CPUID[CPUID_88][2], 7, 0) + 1;␊ |
333 | ␉}␊ |
334 | ␉else␊ |
335 | ␉{␊ |
336 | ␉␉// Use previous method for Cores and Threads␊ |
337 | ␉␉p->CPU.NoThreads␉= bitfield(p->CPU.CPUID[CPUID_1][1], 23, 16);␊ |
338 | ␉␉p->CPU.NoCores␉␉= bitfield(p->CPU.CPUID[CPUID_4][0], 31, 26) + 1;␊ |
339 | ␉}␊ |
340 | ␊ |
341 | ␉/* get brand string (if supported) */␊ |
342 | ␉/* Copyright: from Apple's XNU cpuid.c */␊ |
343 | ␉if (p->CPU.CPUID[CPUID_80][0] > 0x80000004)␊ |
344 | ␉{␊ |
345 | ␉␉uint32_t␉reg[4];␊ |
346 | ␉␉char␉␉str[128], *s;␊ |
347 | ␉␉/*␊ |
348 | ␉␉ * The brand string 48 bytes (max), guaranteed to␊ |
349 | ␉␉ * be NULL terminated.␊ |
350 | ␉␉ */␊ |
351 | ␉␉do_cpuid(0x80000002, reg);␊ |
352 | ␉␉bcopy((char *)reg, &str[0], 16);␊ |
353 | ␉␉do_cpuid(0x80000003, reg);␊ |
354 | ␉␉bcopy((char *)reg, &str[16], 16);␊ |
355 | ␉␉do_cpuid(0x80000004, reg);␊ |
356 | ␉␉bcopy((char *)reg, &str[32], 16);␊ |
357 | ␉␉for (s = str; *s != '\0'; s++)␊ |
358 | ␉␉{␊ |
359 | ␉␉␉if (*s != ' ')␊ |
360 | ␉␉␉{␊ |
361 | ␉␉␉␉break;␊ |
362 | ␉␉␉}␊ |
363 | ␉␉}␊ |
364 | ␉␉␊ |
365 | ␉␉strlcpy(p->CPU.BrandString, s, sizeof(p->CPU.BrandString));␊ |
366 | ␉␉␊ |
367 | ␉␉if (!strncmp(p->CPU.BrandString, CPU_STRING_UNKNOWN, MIN(sizeof(p->CPU.BrandString), strlen(CPU_STRING_UNKNOWN) + 1)))␊ |
368 | ␉␉{␊ |
369 | ␉␉␉/*␊ |
370 | ␉␉␉ * This string means we have a firmware-programmable brand string,␊ |
371 | ␉␉␉ * and the firmware couldn't figure out what sort of CPU we have.␊ |
372 | ␉␉␉ */␊ |
373 | ␉␉␉p->CPU.BrandString[0] = '\0';␊ |
374 | ␉␉}␊ |
375 | ␉}␊ |
376 | ␉␊ |
377 | ␉/* setup features */␊ |
378 | ␉if ((bit(23) & p->CPU.CPUID[CPUID_1][3]) != 0)␊ |
379 | ␉{␊ |
380 | ␉␉p->CPU.Features |= CPU_FEATURE_MMX;␊ |
381 | ␉}␊ |
382 | ␉if ((bit(25) & p->CPU.CPUID[CPUID_1][3]) != 0)␊ |
383 | ␉{␊ |
384 | ␉␉p->CPU.Features |= CPU_FEATURE_SSE;␊ |
385 | ␉}␊ |
386 | ␉if ((bit(26) & p->CPU.CPUID[CPUID_1][3]) != 0)␊ |
387 | ␉{␊ |
388 | ␉␉p->CPU.Features |= CPU_FEATURE_SSE2;␊ |
389 | ␉}␊ |
390 | ␉if ((bit(0) & p->CPU.CPUID[CPUID_1][2]) != 0)␊ |
391 | ␉{␊ |
392 | ␉␉p->CPU.Features |= CPU_FEATURE_SSE3;␊ |
393 | ␉}␊ |
394 | ␉if ((bit(19) & p->CPU.CPUID[CPUID_1][2]) != 0)␊ |
395 | ␉{␊ |
396 | ␉␉p->CPU.Features |= CPU_FEATURE_SSE41;␊ |
397 | ␉}␊ |
398 | ␉if ((bit(20) & p->CPU.CPUID[CPUID_1][2]) != 0)␊ |
399 | ␉{␊ |
400 | ␉␉p->CPU.Features |= CPU_FEATURE_SSE42;␊ |
401 | ␉}␊ |
402 | ␉if ((bit(29) & p->CPU.CPUID[CPUID_81][3]) != 0)␊ |
403 | ␉{␊ |
404 | ␉␉p->CPU.Features |= CPU_FEATURE_EM64T;␊ |
405 | ␉}␊ |
406 | ␉if ((bit(5) & p->CPU.CPUID[CPUID_1][3]) != 0)␊ |
407 | ␉{␊ |
408 | ␉␉p->CPU.Features |= CPU_FEATURE_MSR;␊ |
409 | ␉}␊ |
410 | ␉//if ((bit(28) & p->CPU.CPUID[CPUID_1][3]) != 0) {␊ |
411 | ␉if (p->CPU.NoThreads > p->CPU.NoCores)␊ |
412 | ␉{␊ |
413 | ␉␉p->CPU.Features |= CPU_FEATURE_HTT;␊ |
414 | ␉}␊ |
415 | ␊ |
416 | ␉tscFrequency = measure_tsc_frequency();␊ |
417 | ␉/* if usual method failed */␊ |
418 | ␉if ( tscFrequency < 1000 ) { //TEST␊ |
419 | ␉␉tscFrequency = timeRDTSC() * 20;␊ |
420 | ␉}␊ |
421 | ␉fsbFrequency = 0;␊ |
422 | ␉cpuFrequency = 0;␊ |
423 | ␊ |
424 | ␉if ((p->CPU.Vendor == CPUID_VENDOR_INTEL) && ((p->CPU.Family == 0x06) || (p->CPU.Family == 0x0f))) {␊ |
425 | ␉␉int intelCPU = p->CPU.Model;␊ |
426 | ␉␉if ((p->CPU.Family == 0x06 && p->CPU.Model >= 0x0c) || (p->CPU.Family == 0x0f && p->CPU.Model >= 0x03))␉{␊ |
427 | ␉␉␉/* Nehalem CPU model */␊ |
428 | ␉␉␉if (p->CPU.Family == 0x06 && (p->CPU.Model == CPU_MODEL_NEHALEM␉␉||␊ |
429 | ␉␉␉␉␉␉␉␉␉␉ p->CPU.Model == CPU_MODEL_FIELDS␉||␊ |
430 | ␉␉␉␉␉␉␉␉␉␉ p->CPU.Model == CPU_MODEL_DALES␉||␊ |
431 | ␉␉␉␉␉␉␉␉␉␉ p->CPU.Model == CPU_MODEL_DALES_32NM␉||␊ |
432 | ␉␉␉␉␉␉␉␉␉␉ p->CPU.Model == CPU_MODEL_WESTMERE␉||␊ |
433 | ␉␉␉␉␉␉␉␉␉␉ p->CPU.Model == CPU_MODEL_NEHALEM_EX␉||␊ |
434 | ␉␉␉␉␉␉␉␉␉␉ p->CPU.Model == CPU_MODEL_WESTMERE_EX ||␊ |
435 | ␉␉␉␉␉␉␉␉␉␉ p->CPU.Model == CPU_MODEL_SANDYBRIDGE ||␊ |
436 | ␉␉␉␉␉␉␉␉␉␉ p->CPU.Model == CPU_MODEL_JAKETOWN ||␊ |
437 | ␉␉␉␉␉␉␉␉␉␉ p->CPU.Model == CPU_MODEL_IVYBRIDGE_XEON␉||␊ |
438 | ␉␉␉␉␉␉␉␉␉␉ p->CPU.Model == CPU_MODEL_IVYBRIDGE ||␊ |
439 | ␉␉␉␉␉␉␉␉␉␉ p->CPU.Model == CPU_MODEL_HASWELL ||␊ |
440 | ␉␉␉␉␉␉␉␉␉␉ p->CPU.Model == CPU_MODEL_HASWELL_MB ||␊ |
441 | ␉␉␉␉␉␉␉␉␉␉ //p->CPU.Model == CPU_MODEL_HASWELL_H ||␊ |
442 | ␉␉␉␉␉␉␉␉␉␉ p->CPU.Model == CPU_MODEL_HASWELL_ULT ||␊ |
443 | ␉␉␉␉␉␉␉␉␉␉ p->CPU.Model == CPU_MODEL_CRYSTALWELL ))␊ |
444 | ␉␉␉{␊ |
445 | ␉␉␉␉msr = rdmsr64(MSR_PLATFORM_INFO);␊ |
446 | ␉␉␉␉DBG("msr(%d): platform_info %08x\n", __LINE__, bitfield(msr, 31, 0));␊ |
447 | ␉␉␉␉bus_ratio_max = bitfield(msr, 15, 8);␊ |
448 | ␉␉␉␉bus_ratio_min = bitfield(msr, 47, 40); //valv: not sure about this one (Remarq.1)␊ |
449 | ␉␉␉␉msr = rdmsr64(MSR_FLEX_RATIO);␊ |
450 | ␉␉␉␉DBG("msr(%d): flex_ratio %08x\n", __LINE__, bitfield(msr, 31, 0));␊ |
451 | ␉␉␉␉if (bitfield(msr, 16, 16))␊ |
452 | ␉␉␉␉{␊ |
453 | ␉␉␉␉␉flex_ratio = bitfield(msr, 15, 8);␊ |
454 | ␉␉␉␉␉/* bcc9: at least on the gigabyte h67ma-ud2h,␊ |
455 | ␉␉␉␉␉ where the cpu multipler can't be changed to␊ |
456 | ␉␉␉␉␉ allow overclocking, the flex_ratio msr has unexpected (to OSX)␊ |
457 | ␉␉␉␉␉ contents.␉These contents cause mach_kernel to␊ |
458 | ␉␉␉␉␉ fail to compute the bus ratio correctly, instead␊ |
459 | ␉␉␉␉␉ causing the system to crash since tscGranularity␊ |
460 | ␉␉␉␉␉ is inadvertently set to 0.␊ |
461 | ␉␉␉␉␉ */␊ |
462 | ␉␉␉␉␉if (flex_ratio == 0)␊ |
463 | ␉␉␉␉␉{␊ |
464 | ␉␉␉␉␉␉/* Clear bit 16 (evidently the presence bit) */␊ |
465 | ␉␉␉␉␉␉wrmsr64(MSR_FLEX_RATIO, (msr & 0xFFFFFFFFFFFEFFFFULL));␊ |
466 | ␉␉␉␉␉␉msr = rdmsr64(MSR_FLEX_RATIO);␊ |
467 | ␉␉␉␉␉␉verbose("Unusable flex ratio detected. Patched MSR now %08x\n", bitfield(msr, 31, 0));␊ |
468 | ␉␉␉␉␉}␊ |
469 | ␉␉␉␉␉else␊ |
470 | ␉␉␉␉␉{␊ |
471 | ␉␉␉␉␉␉if (bus_ratio_max > flex_ratio)␊ |
472 | ␉␉␉␉␉␉{␊ |
473 | ␉␉␉␉␉␉␉bus_ratio_max = flex_ratio;␊ |
474 | ␉␉␉␉␉␉}␊ |
475 | ␉␉␉␉␉}␊ |
476 | ␉␉␉␉}␊ |
477 | ␊ |
478 | ␉␉␉␉if (bus_ratio_max)␊ |
479 | ␉␉␉␉{␊ |
480 | ␉␉␉␉␉fsbFrequency = (tscFrequency / bus_ratio_max);␊ |
481 | ␉␉␉␉}␊ |
482 | ␉␉␉␉//valv: Turbo Ratio Limit␊ |
483 | ␉␉␉␉if ((intelCPU != 0x2e) && (intelCPU != 0x2f))␊ |
484 | ␉␉␉␉{␊ |
485 | ␉␉␉␉␉msr = rdmsr64(MSR_TURBO_RATIO_LIMIT);␊ |
486 | ␉␉␉␉␉cpuFrequency = bus_ratio_max * fsbFrequency;␊ |
487 | ␉␉␉␉␉max_ratio = bus_ratio_max * 10;␊ |
488 | ␉␉␉␉}␊ |
489 | ␉␉␉␉else␊ |
490 | ␉␉␉␉{␊ |
491 | ␉␉␉␉␉cpuFrequency = tscFrequency;␊ |
492 | ␉␉␉␉}␊ |
493 | ␉␉␉␉if ((getValueForKey(kbusratio, &newratio, &len, &bootInfo->chameleonConfig)) && (len <= 4))␊ |
494 | ␉␉␉␉{␊ |
495 | ␉␉␉␉␉max_ratio = atoi(newratio);␊ |
496 | ␉␉␉␉␉max_ratio = (max_ratio * 10);␊ |
497 | ␉␉␉␉␉if (len >= 3)␊ |
498 | ␉␉␉␉␉{␊ |
499 | ␉␉␉␉␉␉max_ratio = (max_ratio + 5);␊ |
500 | ␉␉␉␉␉}␊ |
501 | ␊ |
502 | ␉␉␉␉␉verbose("Bus-Ratio: min=%d, max=%s\n", bus_ratio_min, newratio);␊ |
503 | ␊ |
504 | ␉␉␉␉␉// extreme overclockers may love 320 ;)␊ |
505 | ␉␉␉␉␉if ((max_ratio >= min_ratio) && (max_ratio <= 320))␊ |
506 | ␉␉␉␉␉{␊ |
507 | ␉␉␉␉␉␉cpuFrequency = (fsbFrequency * max_ratio) / 10;␊ |
508 | ␉␉␉␉␉␉if (len >= 3)␊ |
509 | ␉␉␉␉␉␉{␊ |
510 | ␉␉␉␉␉␉␉maxdiv = 1;␊ |
511 | ␉␉␉␉␉␉}␊ |
512 | ␉␉␉␉␉␉else␊ |
513 | ␉␉␉␉␉␉{␊ |
514 | ␉␉␉␉␉␉␉maxdiv = 0;␊ |
515 | ␉␉␉␉␉␉}␊ |
516 | ␉␉␉␉␉}␊ |
517 | ␉␉␉␉␉else␊ |
518 | ␉␉␉␉␉{␊ |
519 | ␉␉␉␉␉␉max_ratio = (bus_ratio_max * 10);␊ |
520 | ␉␉␉␉␉}␊ |
521 | ␉␉␉␉}␊ |
522 | ␉␉␉␉//valv: to be uncommented if Remarq.1 didn't stick␊ |
523 | ␉␉␉␉/*if (bus_ratio_max > 0) bus_ratio = flex_ratio;*/␊ |
524 | ␉␉␉␉p->CPU.MaxRatio = max_ratio;␊ |
525 | ␉␉␉␉p->CPU.MinRatio = min_ratio;␊ |
526 | ␊ |
527 | ␉␉␉␉myfsb = fsbFrequency / 1000000;␊ |
528 | ␉␉␉␉verbose("Sticking with [BCLK: %dMhz, Bus-Ratio: %d]\n", myfsb, max_ratio/10);␊ |
529 | ␉␉␉␉currcoef = bus_ratio_max;␊ |
530 | ␉␉␉} else {␊ |
531 | ␉␉␉␉msr = rdmsr64(MSR_IA32_PERF_STATUS);␊ |
532 | ␉␉␉␉DBG("msr(%d): ia32_perf_stat 0x%08x\n", __LINE__, bitfield(msr, 31, 0));␊ |
533 | ␉␉␉␉currcoef = bitfield(msr, 12, 8);␊ |
534 | ␉␉␉␉/* Non-integer bus ratio for the max-multi*/␊ |
535 | ␉␉␉␉maxdiv = bitfield(msr, 46, 46);␊ |
536 | ␉␉␉␉/* Non-integer bus ratio for the current-multi (undocumented)*/␊ |
537 | ␉␉␉␉currdiv = bitfield(msr, 14, 14);␊ |
538 | ␊ |
539 | ␉␉␉␉// This will always be model >= 3␊ |
540 | ␉␉␉␉if ((p->CPU.Family == 0x06 && p->CPU.Model >= 0x0e) || (p->CPU.Family == 0x0f))␊ |
541 | ␉␉␉␉{␊ |
542 | ␉␉␉␉␉/* On these models, maxcoef defines TSC freq */␊ |
543 | ␉␉␉␉␉maxcoef = bitfield(msr, 44, 40);␊ |
544 | ␉␉␉␉}␊ |
545 | ␉␉␉␉else␊ |
546 | ␉␉␉␉{␊ |
547 | ␉␉␉␉␉/* On lower models, currcoef defines TSC freq */␊ |
548 | ␉␉␉␉␉/* XXX */␊ |
549 | ␉␉␉␉␉maxcoef = currcoef;␊ |
550 | ␉␉␉␉}␊ |
551 | ␊ |
552 | ␉␉␉␉if (maxcoef)␊ |
553 | ␉␉␉␉{␊ |
554 | ␉␉␉␉␉if (maxdiv)␊ |
555 | ␉␉␉␉␉{␊ |
556 | ␉␉␉␉␉␉fsbFrequency = ((tscFrequency * 2) / ((maxcoef * 2) + 1));␊ |
557 | ␉␉␉␉␉}␊ |
558 | ␉␉␉␉␉else␊ |
559 | ␉␉␉␉␉{␊ |
560 | ␉␉␉␉␉␉fsbFrequency = (tscFrequency / maxcoef);␊ |
561 | ␉␉␉␉␉}␊ |
562 | ␉␉␉␉␉if (currdiv)␊ |
563 | ␉␉␉␉␉{␊ |
564 | ␉␉␉␉␉␉cpuFrequency = (fsbFrequency * ((currcoef * 2) + 1) / 2);␊ |
565 | ␉␉␉␉␉}␊ |
566 | ␉␉␉␉␉else␊ |
567 | ␉␉␉␉␉{␊ |
568 | ␉␉␉␉␉␉cpuFrequency = (fsbFrequency * currcoef);␊ |
569 | ␉␉␉␉␉}␊ |
570 | ␉␉␉␉␉DBG("max: %d%s current: %d%s\n", maxcoef, maxdiv ? ".5" : "",currcoef, currdiv ? ".5" : "");␊ |
571 | ␉␉␉␉}␊ |
572 | ␉␉␉}␊ |
573 | ␉␉}␊ |
574 | ␉␉/* Mobile CPU */␊ |
575 | ␉␉if (rdmsr64(MSR_IA32_PLATFORM_ID) & (1<<28)) {␊ |
576 | ␉␉␉p->CPU.Features |= CPU_FEATURE_MOBILE;␊ |
577 | ␉␉}␊ |
578 | ␉} else if ((p->CPU.Vendor == CPUID_VENDOR_AMD) && (p->CPU.Family == 0x0f)) {␊ |
579 | ␉␉switch(p->CPU.ExtFamily) {␊ |
580 | ␉␉␉case 0x00: /* K8 */␊ |
581 | ␉␉␉␉msr = rdmsr64(K8_FIDVID_STATUS);␊ |
582 | ␉␉␉␉maxcoef = bitfield(msr, 21, 16) / 2 + 4;␊ |
583 | ␉␉␉␉currcoef = bitfield(msr, 5, 0) / 2 + 4;␊ |
584 | ␉␉␉␉break;␊ |
585 | ␊ |
586 | ␉␉␉case 0x01: /* K10 */␊ |
587 | ␉␉␉␉msr = rdmsr64(K10_COFVID_STATUS);␊ |
588 | ␉␉␉␉do_cpuid2(0x00000006, 0, p->CPU.CPUID[CPUID_6]);␊ |
589 | ␉␉␉␉// EffFreq: effective frequency interface␊ |
590 | ␉␉␉␉if (bitfield(p->CPU.CPUID[CPUID_6][2], 0, 0) == 1)␊ |
591 | ␉␉␉␉{␊ |
592 | ␉␉␉␉␉//uint64_t mperf = measure_mperf_frequency();␊ |
593 | ␉␉␉␉␉uint64_t aperf = measure_aperf_frequency();␊ |
594 | ␉␉␉␉␉cpuFrequency = aperf;␊ |
595 | ␉␉␉␉}␊ |
596 | ␉␉␉␉// NOTE: tsc runs at the maccoeff (non turbo)␊ |
597 | ␉␉␉␉//␉␉␉*not* at the turbo frequency.␊ |
598 | ␉␉␉␉maxcoef␉ = bitfield(msr, 54, 49) / 2 + 4;␊ |
599 | ␉␉␉␉currcoef = bitfield(msr, 5, 0) + 0x10;␊ |
600 | ␉␉␉␉currdiv = 2 << bitfield(msr, 8, 6);␊ |
601 | ␊ |
602 | ␉␉␉␉break;␊ |
603 | ␊ |
604 | ␉␉␉case 0x05: /* K14 */␊ |
605 | ␉␉␉␉msr = rdmsr64(K10_COFVID_STATUS);␊ |
606 | ␉␉␉␉currcoef = (bitfield(msr, 54, 49) + 0x10) << 2;␊ |
607 | ␉␉␉␉currdiv = (bitfield(msr, 8, 4) + 1) << 2;␊ |
608 | ␉␉␉␉currdiv += bitfield(msr, 3, 0);␊ |
609 | ␊ |
610 | ␉␉␉␉break;␊ |
611 | ␊ |
612 | ␉␉␉case 0x02: /* K11 */␊ |
613 | ␉␉␉␉// not implimented␊ |
614 | ␉␉␉␉break;␊ |
615 | ␉␉}␊ |
616 | ␊ |
617 | ␉␉if (maxcoef) {␊ |
618 | ␉␉␉if (currdiv) {␊ |
619 | ␉␉␉␉if (!currcoef) {␊ |
620 | ␉␉␉␉␉currcoef = maxcoef;␊ |
621 | ␉␉␉␉}␊ |
622 | ␊ |
623 | ␉␉␉␉if (!cpuFrequency) {␊ |
624 | ␉␉␉␉␉fsbFrequency = ((tscFrequency * currdiv) / currcoef);␊ |
625 | ␉␉␉␉} else {␊ |
626 | ␉␉␉␉␉fsbFrequency = ((cpuFrequency * currdiv) / currcoef);␊ |
627 | ␉␉␉␉}␊ |
628 | ␉␉␉␉DBG("%d.%d\n", currcoef / currdiv, ((currcoef % currdiv) * 100) / currdiv);␊ |
629 | ␉␉␉} else {␊ |
630 | ␉␉␉␉if (!cpuFrequency) {␊ |
631 | ␉␉␉␉␉fsbFrequency = (tscFrequency / maxcoef);␊ |
632 | ␉␉␉␉} else {␊ |
633 | ␉␉␉␉␉fsbFrequency = (cpuFrequency / maxcoef);␊ |
634 | ␉␉␉␉}␊ |
635 | ␉␉␉␉DBG("%d\n", currcoef);␊ |
636 | ␉␉␉}␊ |
637 | ␉␉} else if (currcoef) {␊ |
638 | ␉␉␉if (currdiv) {␊ |
639 | ␉␉␉␉fsbFrequency = ((tscFrequency * currdiv) / currcoef);␊ |
640 | ␉␉␉␉DBG("%d.%d\n", currcoef / currdiv, ((currcoef % currdiv) * 100) / currdiv);␊ |
641 | ␉␉␉} else {␊ |
642 | ␉␉␉␉fsbFrequency = (tscFrequency / currcoef);␊ |
643 | ␉␉␉␉DBG("%d\n", currcoef);␊ |
644 | ␉␉␉}␊ |
645 | ␉␉}␊ |
646 | ␉␉if (!cpuFrequency) cpuFrequency = tscFrequency;␊ |
647 | ␉}␊ |
648 | ␉␊ |
649 | #if 0␊ |
650 | ␉if (!fsbFrequency) {␊ |
651 | ␉␉fsbFrequency = (DEFAULT_FSB * 1000);␊ |
652 | ␉␉cpuFrequency = tscFrequency;␊ |
653 | ␉␉DBG("0 ! using the default value for FSB !\n");␊ |
654 | ␉}␊ |
655 | #endif␊ |
656 | ␊ |
657 | ␉p->CPU.MaxCoef = maxcoef;␊ |
658 | ␉p->CPU.MaxDiv = maxdiv;␊ |
659 | ␉p->CPU.CurrCoef = currcoef;␊ |
660 | ␉p->CPU.CurrDiv = currdiv;␊ |
661 | ␉p->CPU.TSCFrequency = tscFrequency;␊ |
662 | ␉p->CPU.FSBFrequency = fsbFrequency;␊ |
663 | ␉p->CPU.CPUFrequency = cpuFrequency;␊ |
664 | ␊ |
665 | ␉// keep formatted with spaces instead of tabs␊ |
666 | ␉DBG("CPU: Brand String: %s\n", p->CPU.BrandString);␊ |
667 | ␉DBG("CPU: Vendor/Family/ExtFamily: 0x%x/0x%x/0x%x\n", p->CPU.Vendor, p->CPU.Family, p->CPU.ExtFamily);␊ |
668 | ␉DBG("CPU: Model/ExtModel/Stepping: 0x%x/0x%x/0x%x\n", p->CPU.Model, p->CPU.ExtModel, p->CPU.Stepping);␊ |
669 | ␉DBG("CPU: MaxCoef/CurrCoef: 0x%x/0x%x\n", p->CPU.MaxCoef, p->CPU.CurrCoef);␊ |
670 | ␉DBG("CPU: MaxDiv/CurrDiv: 0x%x/0x%x\n", p->CPU.MaxDiv, p->CPU.CurrDiv);␊ |
671 | ␉DBG("CPU: TSCFreq: %dMHz\n", p->CPU.TSCFrequency / 1000000);␊ |
672 | ␉DBG("CPU: FSBFreq: %dMHz\n", p->CPU.FSBFrequency / 1000000);␊ |
673 | ␉DBG("CPU: CPUFreq: %dMHz\n", p->CPU.CPUFrequency / 1000000);␊ |
674 | ␉DBG("CPU: NoCores/NoThreads: %d/%d\n", p->CPU.NoCores, p->CPU.NoThreads);␊ |
675 | ␉DBG("CPU: Features: 0x%08x\n", p->CPU.Features);␊ |
676 | #if DEBUG_CPU␊ |
677 | ␉pause();␊ |
678 | #endif␊ |
679 | }␊ |
680 | |