Chameleon

Chameleon Svn Source Tree

Root/branches/meklort/i386/libsaio/cpu.c

Source at commit 546 created 13 years 7 months ago.
By meklort, More code shrinkage
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␊
10#ifndef DEBUG_CPU␊
11#define DEBUG_CPU 0␊
12#endif␊
13␊
14#if DEBUG_CPU␊
15#define DBG(x...)␉␉printf(x)␊
16#else␊
17#define DBG(x...)␉␉msglog(x)␊
18#endif␊
19␊
20/*␊
21 * DFE: Measures the TSC frequency in Hz (64-bit) using the ACPI PM timer␊
22 */␊
23static uint64_t measure_tsc_frequency(void)␊
24{␊
25 uint64_t tscStart;␊
26 uint64_t tscEnd;␊
27 uint64_t tscDelta = 0xffffffffffffffffULL;␊
28 unsigned long pollCount;␊
29 uint64_t retval = 0;␊
30 int i;␊
31␊
32 /* Time how many TSC ticks elapse in 30 msec using the 8254 PIT␊
33 * counter 2. We run this loop 3 times to make sure the cache␊
34 * is hot and we take the minimum delta from all of the runs.␊
35 * That is to say that we're biased towards measuring the minimum␊
36 * number of TSC ticks that occur while waiting for the timer to␊
37 * expire. That theoretically helps avoid inconsistencies when␊
38 * running under a VM if the TSC is not virtualized and the host␊
39 * steals time. The TSC is normally virtualized for VMware.␊
40 */␊
41 for(i = 0; i < 10; ++i)␊
42 {␊
43 enable_PIT2();␊
44 set_PIT2_mode0(CALIBRATE_LATCH);␊
45 tscStart = rdtsc64();␊
46 pollCount = poll_PIT2_gate();␊
47 tscEnd = rdtsc64();␊
48 /* The poll loop must have run at least a few times for accuracy */␊
49 if(pollCount <= 1)␊
50 continue;␊
51 /* The TSC must increment at LEAST once every millisecond. We␊
52 * should have waited exactly 30 msec so the TSC delta should␊
53 * be >= 30. Anything less and the processor is way too slow.␊
54 */␊
55 if((tscEnd - tscStart) <= CALIBRATE_TIME_MSEC)␊
56 continue;␊
57 // tscDelta = min(tscDelta, (tscEnd - tscStart))␊
58 if( (tscEnd - tscStart) < tscDelta )␊
59 tscDelta = tscEnd - tscStart;␊
60 }␊
61 /* tscDelta is now the least number of TSC ticks the processor made in␊
62 * a timespan of 0.03 s (e.g. 30 milliseconds)␊
63 * Linux thus divides by 30 which gives the answer in kiloHertz because␊
64 * 1 / ms = kHz. But we're xnu and most of the rest of the code uses␊
65 * Hz so we need to convert our milliseconds to seconds. Since we're␊
66 * dividing by the milliseconds, we simply multiply by 1000.␊
67 */␊
68␊
69 /* Unlike linux, we're not limited to 32-bit, but we do need to take care␊
70 * that we're going to multiply by 1000 first so we do need at least some␊
71 * arithmetic headroom. For now, 32-bit should be enough.␊
72 * Also unlike Linux, our compiler can do 64-bit integer arithmetic.␊
73 */␊
74 if(tscDelta > (1ULL<<32))␊
75 retval = 0;␊
76 else␊
77 {␊
78 retval = tscDelta * 1000 / 30;␊
79 }␊
80 disable_PIT2();␊
81 return retval;␊
82}␊
83␊
84/*␊
85 * Calculates the FSB and CPU frequencies using specific MSRs for each CPU␊
86 * - multi. is read from a specific MSR. In the case of Intel, there is:␊
87 * a max multi. (used to calculate the FSB freq.),␊
88 * and a current multi. (used to calculate the CPU freq.)␊
89 * - fsbFrequency = tscFrequency / multi␊
90 * - cpuFrequency = fsbFrequency * multi␊
91 */␊
92␊
93void scan_cpu(PlatformInfo_t *p)␊
94{␊
95␉int i = 0;␊
96␉uint64_t␉tscFrequency, fsbFrequency, cpuFrequency;␊
97␉uint64_t␉msr, flex_ratio;␊
98␉uint8_t␉␉maxcoef, maxdiv, currcoef, currdiv;␊
99␊
100␉maxcoef = maxdiv = currcoef = currdiv = 0;␊
101␊
102␉/* get cpuid values */␊
103␉for( ; i < 3; i++)␊
104␉{␊
105␉␉do_cpuid(i, p->CPU.CPUID[i]);␊
106␉}␊
107␉␊
108␉do_cpuid2(0x00000004, 0, p->CPU.CPUID[CPUID_4]);␊
109␉do_cpuid(0x80000000, p->CPU.CPUID[CPUID_80]);␊
110␉if ((p->CPU.CPUID[CPUID_80][0] & 0x0000000f) >= 1) {␊
111␉␉do_cpuid(0x80000001, p->CPU.CPUID[CPUID_81]);␊
112␉}␊
113#if DEBUG_CPU␊
114␉{␊
115␉␉int␉␉i;␊
116␉␉printf("CPUID Raw Values:\n");␊
117␉␉for (i=0; i<CPUID_MAX; i++) {␊
118␉␉␉printf("%02d: %08x-%08x-%08x-%08x\n", i,␊
119␉␉␉␉p->CPU.CPUID[i][0], p->CPU.CPUID[i][1],␊
120␉␉␉␉p->CPU.CPUID[i][2], p->CPU.CPUID[i][3]);␊
121␉␉}␊
122␉}␊
123#endif␊
124␉p->CPU.Vendor␉␉= p->CPU.CPUID[CPUID_0][1];␊
125␉p->CPU.Signature␉= p->CPU.CPUID[CPUID_1][0];␊
126␉p->CPU.Stepping␉␉= bitfield(p->CPU.CPUID[CPUID_1][0], 3, 0);␊
127␉p->CPU.Model␉␉= bitfield(p->CPU.CPUID[CPUID_1][0], 7, 4);␊
128␉p->CPU.Family␉␉= bitfield(p->CPU.CPUID[CPUID_1][0], 11, 8);␊
129␉p->CPU.ExtModel␉␉= bitfield(p->CPU.CPUID[CPUID_1][0], 19, 16);␊
130␉p->CPU.ExtFamily␉= bitfield(p->CPU.CPUID[CPUID_1][0], 27, 20);␊
131␉p->CPU.NoThreads␉= bitfield(p->CPU.CPUID[CPUID_1][1], 23, 16);␊
132␉p->CPU.NoCores␉␉= bitfield(p->CPU.CPUID[CPUID_4][0], 31, 26) + 1;␊
133␊
134␉p->CPU.Model += (p->CPU.ExtModel << 4);␊
135␉␊
136␉/* get brand string (if supported) */␊
137␉/* Copyright: from Apple's XNU cpuid.c */␊
138␉if (p->CPU.CPUID[CPUID_80][0] > 0x80000004) {␊
139␉␉uint32_t␉reg[4];␊
140 char str[128], *s;␊
141␉␉/*␊
142␉␉ * The brand string 48 bytes (max), guaranteed to␊
143␉␉ * be NUL terminated.␊
144␉␉ */␊
145␉␉do_cpuid(0x80000002, reg);␊
146␉␉bcopy((char *)reg, &str[0], 16);␊
147␉␉do_cpuid(0x80000003, reg);␊
148␉␉bcopy((char *)reg, &str[16], 16);␊
149␉␉do_cpuid(0x80000004, reg);␊
150␉␉bcopy((char *)reg, &str[32], 16);␊
151␉␉for (s = str; *s != '\0'; s++) {␊
152␉␉␉if (*s != ' ') break;␊
153␉␉}␊
154␉␉␊
155␉␉strlcpy(p->CPU.BrandString,␉s, sizeof(p->CPU.BrandString));␊
156␉␉␊
157␉␉if (!strncmp(p->CPU.BrandString, CPU_STRING_UNKNOWN, min(sizeof(p->CPU.BrandString), strlen(CPU_STRING_UNKNOWN) + 1))) {␊
158␉␉␉ /*␊
159␉␉␉ * This string means we have a firmware-programmable brand string,␊
160␉␉␉ * and the firmware couldn't figure out what sort of CPU we have.␊
161␉␉␉ */␊
162␉␉␉ p->CPU.BrandString[0] = '\0';␊
163␉␉ }␊
164␉}␊
165␉␊
166␉/* setup features */␊
167␉p->CPU.Features |= (CPU_FEATURE_MMX | CPU_FEATURE_SSE | CPU_FEATURE_SSE2 | CPU_FEATURE_MSR) & p->CPU.CPUID[CPUID_1][3];␊
168␉p->CPU.Features |= (CPU_FEATURE_SSE3 | CPU_FEATURE_SSE41 | CPU_FEATURE_SSE42) & p->CPU.CPUID[CPUID_1][2];␉␊
169␉p->CPU.Features |= (CPU_FEATURE_EM64T) & p->CPU.CPUID[CPUID_81][3];␊
170␊
171␊
172␉//if ((CPU_FEATURE_HTT & p->CPU.CPUID[CPUID_1][3]) != 0) {␊
173␉if (p->CPU.NoThreads > p->CPU.NoCores) {␊
174␉␉p->CPU.Features |= CPU_FEATURE_HTT;␊
175␉}␊
176␉ ␊
177␊
178␉tscFrequency = measure_tsc_frequency();␊
179␉fsbFrequency = 0;␊
180␉cpuFrequency = 0;␊
181␊
182␉if ((p->CPU.Vendor == 0x756E6547 /* Intel */) && ␊
183␉␉((p->CPU.Family == 0x06) || ␊
184␉␉ (p->CPU.Family == 0x0f)))␊
185␉{␊
186␉␉if ((p->CPU.Family == 0x06 && p->CPU.Model >= 0x0c) || ␊
187␉␉␉(p->CPU.Family == 0x0f && p->CPU.Model >= 0x03))␊
188␉␉{␊
189␉␉␉/* Nehalem CPU model */␊
190␉␉␉if (p->CPU.Family == 0x06 && (p->CPU.Model == 0x1a || p->CPU.Model == 0x1e ||␊
191␉␉␉␉␉␉␉␉␉␉ p->CPU.Model == 0x1f || p->CPU.Model == 0x25 ||␊
192␉␉␉␉␉␉␉␉␉␉ p->CPU.Model == 0x2c)) ␊
193␉␉␉{␊
194␉␉␉␉msr = rdmsr64(MSR_PLATFORM_INFO);␊
195␉␉␉␉DBG("msr(%d): platform_info %08x\n", __LINE__, msr & 0xffffffff);␊
196␉␉␉␉currcoef = (msr >> 8) & 0xff;␊
197␉␉␉␉msr = rdmsr64(MSR_FLEX_RATIO);␊
198␉␉␉␉DBG("msr(%d): flex_ratio %08x\n", __LINE__, msr & 0xffffffff);␊
199␉␉␉␉if ((msr >> 16) & 0x01)␊
200␉␉␉␉{␊
201␉␉␉␉␉flex_ratio = (msr >> 8) & 0xff;␊
202␉␉␉␉␉if (currcoef > flex_ratio) ␊
203␉␉␉␉␉{␊
204␉␉␉␉␉␉currcoef = flex_ratio;␊
205␉␉␉␉␉}␊
206␉␉␉␉}␊
207␊
208␉␉␉␉if (currcoef) {␊
209␉␉␉␉␉fsbFrequency = (tscFrequency / currcoef);␊
210␉␉␉␉}␊
211␉␉␉␉cpuFrequency = tscFrequency;␊
212␉␉␉} ␊
213␉␉␉else␊
214␉␉␉{␊
215␉␉␉␉msr = rdmsr64(MSR_IA32_PERF_STATUS);␊
216␉␉␉␉DBG("msr(%d): ia32_perf_stat 0x%08x\n", __LINE__, msr & 0xffffffff);␊
217␉␉␉␉currcoef = (msr >> 8) & 0x1f;␊
218␉␉␉␉/* Non-integer bus ratio for the max-multi*/␊
219␉␉␉␉maxdiv = (msr >> 46) & 0x01;␊
220␉␉␉␉/* Non-integer bus ratio for the current-multi (undocumented)*/␊
221␉␉␉␉currdiv = (msr >> 14) & 0x01;␊
222␊
223␉␉␉␉if ((p->CPU.Family == 0x06 && p->CPU.Model >= 0x0e) || ␊
224␉␉␉␉␉(p->CPU.Family == 0x0f)) // This will always be model >= 3␊
225␉␉␉␉{␊
226␉␉␉␉␉/* On these models, maxcoef defines TSC freq */␊
227␉␉␉␉␉maxcoef = (msr >> 40) & 0x1f;␊
228␉␉␉␉} ␊
229␉␉␉␉else ␊
230␉␉␉␉{␊
231␉␉␉␉␉/* On lower models, currcoef defines TSC freq */␊
232␉␉␉␉␉/* XXX */␊
233␉␉␉␉␉maxcoef = currcoef;␊
234␉␉␉␉}␊
235␊
236␉␉␉␉if (maxcoef) ␊
237␉␉␉␉{␊
238␉␉␉␉␉if (maxdiv)␊
239␉␉␉␉␉{␊
240␉␉␉␉␉␉fsbFrequency = ((tscFrequency * 2) / ((maxcoef * 2) + 1));␊
241␉␉␉␉␉}␊
242␉␉␉␉␉else ␊
243␉␉␉␉␉{␊
244␉␉␉␉␉␉fsbFrequency = (tscFrequency / maxcoef);␊
245␉␉␉␉␉}␊
246␉␉␉␉␉␊
247␉␉␉␉␉if (currdiv) ␊
248␉␉␉␉␉{␊
249␉␉␉␉␉␉cpuFrequency = (fsbFrequency * ((currcoef * 2) + 1) / 2);␊
250␉␉␉␉␉}␊
251␉␉␉␉␉else ␊
252␉␉␉␉␉{␊
253␉␉␉␉␉␉cpuFrequency = (fsbFrequency * currcoef);␊
254␉␉␉␉␉}␊
255␉␉␉␉␉DBG("max: %d%s current: %d%s\n", maxcoef, maxdiv ? ".5" : "",currcoef, currdiv ? ".5" : "");␊
256␉␉␉␉}␊
257␉␉␉}␊
258␉␉}␊
259␉␉/* Mobile CPU ? */␊
260␉␉if (rdmsr64(0x17) & (1<<28))␊
261␉␉{␊
262␉␉␉p->CPU.Features |= CPU_FEATURE_MOBILE;␊
263␉␉}␊
264␉}␊
265#if 0␊
266␉else if((p->CPU.Vendor == 0x68747541 /* AMD */) && (p->CPU.Family == 0x0f))␊
267␉{␊
268␉␉if(p->CPU.ExtFamily == 0x00 /* K8 */)␊
269␉␉{␊
270␉␉␉msr = rdmsr64(K8_FIDVID_STATUS);␊
271␉␉␉currcoef = (msr & 0x3f) / 2 + 4;␊
272␉␉␉currdiv = (msr & 0x01) * 2;␊
273␉␉} ␊
274␉␉else if(p->CPU.ExtFamily >= 0x01 /* K10+ */)␊
275␉␉{␊
276␉␉␉msr = rdmsr64(K10_COFVID_STATUS);␊
277␉␉␉if(p->CPU.ExtFamily == 0x01 /* K10 */)␊
278␉␉␉␉currcoef = (msr & 0x3f) + 0x10;␊
279␉␉␉else /* K11+ */␊
280␉␉␉␉currcoef = (msr & 0x3f) + 0x08;␊
281␉␉␉currdiv = (2 << ((msr >> 6) & 0x07));␊
282␉␉}␊
283␊
284␉␉if (currcoef) ␊
285␉␉{␊
286␉␉␉if (currdiv) ␊
287␉␉␉{␊
288␉␉␉␉fsbFrequency = ((tscFrequency * currdiv) / currcoef);␊
289␉␉␉␉DBG("%d.%d\n", currcoef / currdiv, ((currcoef % currdiv) * 100) / currdiv);␊
290␉␉␉}␊
291␉␉␉else␊
292␉␉␉{␊
293␉␉␉␉fsbFrequency = (tscFrequency / currcoef);␊
294␉␉␉␉DBG("%d\n", currcoef);␊
295␉␉␉}␊
296␉␉␉fsbFrequency = (tscFrequency / currcoef);␊
297␉␉␉cpuFrequency = tscFrequency;␊
298␉␉}␊
299␉}␊
300␊
301␉if (!fsbFrequency) ␊
302␉{␊
303␉␉fsbFrequency = (DEFAULT_FSB * 1000);␊
304␉␉cpuFrequency = tscFrequency;␊
305␉␉DBG("0 ! using the default value for FSB !\n");␊
306␉}␊
307#endif␊
308␊
309␉p->CPU.MaxCoef = maxcoef;␊
310␉p->CPU.MaxDiv = maxdiv;␊
311␉p->CPU.CurrCoef = currcoef;␊
312␉p->CPU.CurrDiv = currdiv;␊
313␉p->CPU.TSCFrequency = tscFrequency;␊
314␉p->CPU.FSBFrequency = fsbFrequency;␊
315␉p->CPU.CPUFrequency = cpuFrequency;␊
316␊
317␉DBG("CPU: Vendor/Model/ExtModel: 0x%x/0x%x/0x%x\n", p->CPU.Vendor, p->CPU.Model, p->CPU.ExtModel);␊
318␉DBG("CPU: Family/ExtFamily: 0x%x/0x%x\n", p->CPU.Family, p->CPU.ExtFamily);␊
319␉DBG("CPU: MaxCoef/CurrCoef: 0x%x/0x%x\n", p->CPU.MaxCoef, p->CPU.CurrCoef);␊
320␉DBG("CPU: MaxDiv/CurrDiv: 0x%x/0x%x\n", p->CPU.MaxDiv, p->CPU.CurrDiv);␊
321␉DBG("CPU: TSCFreq: %dMHz\n", p->CPU.TSCFrequency / 1000000);␊
322␉DBG("CPU: FSBFreq: %dMHz\n", p->CPU.FSBFrequency / 1000000);␊
323␉DBG("CPU: CPUFreq: %dMHz\n", p->CPU.CPUFrequency / 1000000);␊
324␉DBG("CPU: NoCores/NoThreads: %d/%d\n", p->CPU.NoCores, p->CPU.NoThreads);␊
325␉DBG("CPU: Features: 0x%08x\n", p->CPU.Features);␊
326#if DEBUG_CPU␊
327␉pause();␊
328#endif␊
329}␊
330

Archive Download this file

Revision: 546