Merge branch 'linus' into release
[linux-2.6] / arch / x86 / kernel / cpu / cpufreq / p4-clockmod.c
1 /*
2  *      Pentium 4/Xeon CPU on demand clock modulation/speed scaling
3  *      (C) 2002 - 2003 Dominik Brodowski <linux@brodo.de>
4  *      (C) 2002 Zwane Mwaikambo <zwane@commfireservices.com>
5  *      (C) 2002 Arjan van de Ven <arjanv@redhat.com>
6  *      (C) 2002 Tora T. Engstad
7  *      All Rights Reserved
8  *
9  *      This program is free software; you can redistribute it and/or
10  *      modify it under the terms of the GNU General Public License
11  *      as published by the Free Software Foundation; either version
12  *      2 of the License, or (at your option) any later version.
13  *
14  *      The author(s) of this software shall not be held liable for damages
15  *      of any nature resulting due to the use of this software. This
16  *      software is provided AS-IS with no warranties.
17  *
18  *      Date            Errata                  Description
19  *      20020525        N44, O17        12.5% or 25% DC causes lockup
20  *
21  */
22
23 #include <linux/kernel.h>
24 #include <linux/module.h>
25 #include <linux/init.h>
26 #include <linux/smp.h>
27 #include <linux/cpufreq.h>
28 #include <linux/slab.h>
29 #include <linux/cpumask.h>
30 #include <linux/timex.h>
31
32 #include <asm/processor.h>
33 #include <asm/msr.h>
34 #include <asm/timer.h>
35
36 #include "speedstep-lib.h"
37
38 #define PFX     "p4-clockmod: "
39 #define dprintk(msg...) cpufreq_debug_printk(CPUFREQ_DEBUG_DRIVER, \
40                 "p4-clockmod", msg)
41
42 /*
43  * Duty Cycle (3bits), note DC_DISABLE is not specified in
44  * intel docs i just use it to mean disable
45  */
46 enum {
47         DC_RESV, DC_DFLT, DC_25PT, DC_38PT, DC_50PT,
48         DC_64PT, DC_75PT, DC_88PT, DC_DISABLE
49 };
50
51 #define DC_ENTRIES      8
52
53
54 static int has_N44_O17_errata[NR_CPUS];
55 static unsigned int stock_freq;
56 static struct cpufreq_driver p4clockmod_driver;
57 static unsigned int cpufreq_p4_get(unsigned int cpu);
58
59 static int cpufreq_p4_setdc(unsigned int cpu, unsigned int newstate)
60 {
61         u32 l, h;
62
63         if (!cpu_online(cpu) ||
64             (newstate > DC_DISABLE) || (newstate == DC_RESV))
65                 return -EINVAL;
66
67         rdmsr_on_cpu(cpu, MSR_IA32_THERM_STATUS, &l, &h);
68
69         if (l & 0x01)
70                 dprintk("CPU#%d currently thermal throttled\n", cpu);
71
72         if (has_N44_O17_errata[cpu] &&
73             (newstate == DC_25PT || newstate == DC_DFLT))
74                 newstate = DC_38PT;
75
76         rdmsr_on_cpu(cpu, MSR_IA32_THERM_CONTROL, &l, &h);
77         if (newstate == DC_DISABLE) {
78                 dprintk("CPU#%d disabling modulation\n", cpu);
79                 wrmsr_on_cpu(cpu, MSR_IA32_THERM_CONTROL, l & ~(1<<4), h);
80         } else {
81                 dprintk("CPU#%d setting duty cycle to %d%%\n",
82                         cpu, ((125 * newstate) / 10));
83                 /* bits 63 - 5  : reserved
84                  * bit  4       : enable/disable
85                  * bits 3-1     : duty cycle
86                  * bit  0       : reserved
87                  */
88                 l = (l & ~14);
89                 l = l | (1<<4) | ((newstate & 0x7)<<1);
90                 wrmsr_on_cpu(cpu, MSR_IA32_THERM_CONTROL, l, h);
91         }
92
93         return 0;
94 }
95
96
97 static struct cpufreq_frequency_table p4clockmod_table[] = {
98         {DC_RESV, CPUFREQ_ENTRY_INVALID},
99         {DC_DFLT, 0},
100         {DC_25PT, 0},
101         {DC_38PT, 0},
102         {DC_50PT, 0},
103         {DC_64PT, 0},
104         {DC_75PT, 0},
105         {DC_88PT, 0},
106         {DC_DISABLE, 0},
107         {DC_RESV, CPUFREQ_TABLE_END},
108 };
109
110
111 static int cpufreq_p4_target(struct cpufreq_policy *policy,
112                              unsigned int target_freq,
113                              unsigned int relation)
114 {
115         unsigned int    newstate = DC_RESV;
116         struct cpufreq_freqs freqs;
117         int i;
118
119         if (cpufreq_frequency_table_target(policy, &p4clockmod_table[0],
120                                 target_freq, relation, &newstate))
121                 return -EINVAL;
122
123         freqs.old = cpufreq_p4_get(policy->cpu);
124         freqs.new = stock_freq * p4clockmod_table[newstate].index / 8;
125
126         if (freqs.new == freqs.old)
127                 return 0;
128
129         /* notifiers */
130         for_each_cpu(i, policy->cpus) {
131                 freqs.cpu = i;
132                 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
133         }
134
135         /* run on each logical CPU,
136          * see section 13.15.3 of IA32 Intel Architecture Software
137          * Developer's Manual, Volume 3
138          */
139         for_each_cpu(i, policy->cpus)
140                 cpufreq_p4_setdc(i, p4clockmod_table[newstate].index);
141
142         /* notifiers */
143         for_each_cpu(i, policy->cpus) {
144                 freqs.cpu = i;
145                 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
146         }
147
148         return 0;
149 }
150
151
152 static int cpufreq_p4_verify(struct cpufreq_policy *policy)
153 {
154         return cpufreq_frequency_table_verify(policy, &p4clockmod_table[0]);
155 }
156
157
158 static unsigned int cpufreq_p4_get_frequency(struct cpuinfo_x86 *c)
159 {
160         if (c->x86 == 0x06) {
161                 if (cpu_has(c, X86_FEATURE_EST))
162                         printk(KERN_WARNING PFX "Warning: EST-capable CPU "
163                                "detected. The acpi-cpufreq module offers "
164                                "voltage scaling in addition of frequency "
165                                "scaling. You should use that instead of "
166                                "p4-clockmod, if possible.\n");
167                 switch (c->x86_model) {
168                 case 0x0E: /* Core */
169                 case 0x0F: /* Core Duo */
170                 case 0x16: /* Celeron Core */
171                         p4clockmod_driver.flags |= CPUFREQ_CONST_LOOPS;
172                         return speedstep_get_frequency(SPEEDSTEP_CPU_PCORE);
173                 case 0x0D: /* Pentium M (Dothan) */
174                         p4clockmod_driver.flags |= CPUFREQ_CONST_LOOPS;
175                         /* fall through */
176                 case 0x09: /* Pentium M (Banias) */
177                         return speedstep_get_frequency(SPEEDSTEP_CPU_PM);
178                 }
179         }
180
181         if (c->x86 != 0xF) {
182                 if (!cpu_has(c, X86_FEATURE_EST))
183                         printk(KERN_WARNING PFX "Unknown CPU. "
184                                 "Please send an e-mail to "
185                                 "<cpufreq@vger.kernel.org>\n");
186                 return 0;
187         }
188
189         /* on P-4s, the TSC runs with constant frequency independent whether
190          * throttling is active or not. */
191         p4clockmod_driver.flags |= CPUFREQ_CONST_LOOPS;
192
193         if (speedstep_detect_processor() == SPEEDSTEP_CPU_P4M) {
194                 printk(KERN_WARNING PFX "Warning: Pentium 4-M detected. "
195                        "The speedstep-ich or acpi cpufreq modules offer "
196                        "voltage scaling in addition of frequency scaling. "
197                        "You should use either one instead of p4-clockmod, "
198                        "if possible.\n");
199                 return speedstep_get_frequency(SPEEDSTEP_CPU_P4M);
200         }
201
202         return speedstep_get_frequency(SPEEDSTEP_CPU_P4D);
203 }
204
205
206
207 static int cpufreq_p4_cpu_init(struct cpufreq_policy *policy)
208 {
209         struct cpuinfo_x86 *c = &cpu_data(policy->cpu);
210         int cpuid = 0;
211         unsigned int i;
212
213 #ifdef CONFIG_SMP
214         cpumask_copy(policy->cpus, &per_cpu(cpu_sibling_map, policy->cpu));
215 #endif
216
217         /* Errata workaround */
218         cpuid = (c->x86 << 8) | (c->x86_model << 4) | c->x86_mask;
219         switch (cpuid) {
220         case 0x0f07:
221         case 0x0f0a:
222         case 0x0f11:
223         case 0x0f12:
224                 has_N44_O17_errata[policy->cpu] = 1;
225                 dprintk("has errata -- disabling low frequencies\n");
226         }
227
228         if (speedstep_detect_processor() == SPEEDSTEP_CPU_P4D &&
229             c->x86_model < 2) {
230                 /* switch to maximum frequency and measure result */
231                 cpufreq_p4_setdc(policy->cpu, DC_DISABLE);
232                 recalibrate_cpu_khz();
233         }
234         /* get max frequency */
235         stock_freq = cpufreq_p4_get_frequency(c);
236         if (!stock_freq)
237                 return -EINVAL;
238
239         /* table init */
240         for (i = 1; (p4clockmod_table[i].frequency != CPUFREQ_TABLE_END); i++) {
241                 if ((i < 2) && (has_N44_O17_errata[policy->cpu]))
242                         p4clockmod_table[i].frequency = CPUFREQ_ENTRY_INVALID;
243                 else
244                         p4clockmod_table[i].frequency = (stock_freq * i)/8;
245         }
246         cpufreq_frequency_table_get_attr(p4clockmod_table, policy->cpu);
247
248         /* cpuinfo and default policy values */
249
250         /* the transition latency is set to be 1 higher than the maximum
251          * transition latency of the ondemand governor */
252         policy->cpuinfo.transition_latency = 10000001;
253         policy->cur = stock_freq;
254
255         return cpufreq_frequency_table_cpuinfo(policy, &p4clockmod_table[0]);
256 }
257
258
259 static int cpufreq_p4_cpu_exit(struct cpufreq_policy *policy)
260 {
261         cpufreq_frequency_table_put_attr(policy->cpu);
262         return 0;
263 }
264
265 static unsigned int cpufreq_p4_get(unsigned int cpu)
266 {
267         u32 l, h;
268
269         rdmsr_on_cpu(cpu, MSR_IA32_THERM_CONTROL, &l, &h);
270
271         if (l & 0x10) {
272                 l = l >> 1;
273                 l &= 0x7;
274         } else
275                 l = DC_DISABLE;
276
277         if (l != DC_DISABLE)
278                 return stock_freq * l / 8;
279
280         return stock_freq;
281 }
282
283 static struct freq_attr *p4clockmod_attr[] = {
284         &cpufreq_freq_attr_scaling_available_freqs,
285         NULL,
286 };
287
288 static struct cpufreq_driver p4clockmod_driver = {
289         .verify         = cpufreq_p4_verify,
290         .target         = cpufreq_p4_target,
291         .init           = cpufreq_p4_cpu_init,
292         .exit           = cpufreq_p4_cpu_exit,
293         .get            = cpufreq_p4_get,
294         .name           = "p4-clockmod",
295         .owner          = THIS_MODULE,
296         .attr           = p4clockmod_attr,
297 };
298
299
300 static int __init cpufreq_p4_init(void)
301 {
302         struct cpuinfo_x86 *c = &cpu_data(0);
303         int ret;
304
305         /*
306          * THERM_CONTROL is architectural for IA32 now, so
307          * we can rely on the capability checks
308          */
309         if (c->x86_vendor != X86_VENDOR_INTEL)
310                 return -ENODEV;
311
312         if (!test_cpu_cap(c, X86_FEATURE_ACPI) ||
313                                 !test_cpu_cap(c, X86_FEATURE_ACC))
314                 return -ENODEV;
315
316         ret = cpufreq_register_driver(&p4clockmod_driver);
317         if (!ret)
318                 printk(KERN_INFO PFX "P4/Xeon(TM) CPU On-Demand Clock "
319                                 "Modulation available\n");
320
321         return ret;
322 }
323
324
325 static void __exit cpufreq_p4_exit(void)
326 {
327         cpufreq_unregister_driver(&p4clockmod_driver);
328 }
329
330
331 MODULE_AUTHOR("Zwane Mwaikambo <zwane@commfireservices.com>");
332 MODULE_DESCRIPTION("cpufreq driver for Pentium(TM) 4/Xeon(TM)");
333 MODULE_LICENSE("GPL");
334
335 late_initcall(cpufreq_p4_init);
336 module_exit(cpufreq_p4_exit);