[CPUFREQ] powernow-k8: Limit Pstate transition latency check
[linux-2.6.git] / arch / x86 / kernel / cpu / cpufreq / powernow-k8.c
1 /*
2  *   (c) 2003-2010 Advanced Micro Devices, Inc.
3  *  Your use of this code is subject to the terms and conditions of the
4  *  GNU general public license version 2. See "COPYING" or
5  *  http://www.gnu.org/licenses/gpl.html
6  *
7  *  Support : mark.langsdorf@amd.com
8  *
9  *  Based on the powernow-k7.c module written by Dave Jones.
10  *  (C) 2003 Dave Jones on behalf of SuSE Labs
11  *  (C) 2004 Dominik Brodowski <linux@brodo.de>
12  *  (C) 2004 Pavel Machek <pavel@suse.cz>
13  *  Licensed under the terms of the GNU GPL License version 2.
14  *  Based upon datasheets & sample CPUs kindly provided by AMD.
15  *
16  *  Valuable input gratefully received from Dave Jones, Pavel Machek,
17  *  Dominik Brodowski, Jacob Shin, and others.
18  *  Originally developed by Paul Devriendt.
19  *  Processor information obtained from Chapter 9 (Power and Thermal Management)
20  *  of the "BIOS and Kernel Developer's Guide for the AMD Athlon 64 and AMD
21  *  Opteron Processors" available for download from www.amd.com
22  *
23  *  Tables for specific CPUs can be inferred from
24  *     http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/30430.pdf
25  */
26
27 #include <linux/kernel.h>
28 #include <linux/smp.h>
29 #include <linux/module.h>
30 #include <linux/init.h>
31 #include <linux/cpufreq.h>
32 #include <linux/slab.h>
33 #include <linux/string.h>
34 #include <linux/cpumask.h>
35 #include <linux/sched.h>        /* for current / set_cpus_allowed() */
36 #include <linux/io.h>
37 #include <linux/delay.h>
38
39 #include <asm/msr.h>
40
41 #include <linux/acpi.h>
42 #include <linux/mutex.h>
43 #include <acpi/processor.h>
44
45 #define PFX "powernow-k8: "
46 #define VERSION "version 2.20.00"
47 #include "powernow-k8.h"
48 #include "mperf.h"
49
50 /* serialize freq changes  */
51 static DEFINE_MUTEX(fidvid_mutex);
52
53 static DEFINE_PER_CPU(struct powernow_k8_data *, powernow_data);
54
55 static int cpu_family = CPU_OPTERON;
56
57 /* core performance boost */
58 static bool cpb_capable, cpb_enabled;
59 static struct msr __percpu *msrs;
60
61 static struct cpufreq_driver cpufreq_amd64_driver;
62
63 #ifndef CONFIG_SMP
64 static inline const struct cpumask *cpu_core_mask(int cpu)
65 {
66         return cpumask_of(0);
67 }
68 #endif
69
70 /* Return a frequency in MHz, given an input fid */
71 static u32 find_freq_from_fid(u32 fid)
72 {
73         return 800 + (fid * 100);
74 }
75
76 /* Return a frequency in KHz, given an input fid */
77 static u32 find_khz_freq_from_fid(u32 fid)
78 {
79         return 1000 * find_freq_from_fid(fid);
80 }
81
82 static u32 find_khz_freq_from_pstate(struct cpufreq_frequency_table *data,
83                 u32 pstate)
84 {
85         return data[pstate].frequency;
86 }
87
88 /* Return the vco fid for an input fid
89  *
90  * Each "low" fid has corresponding "high" fid, and you can get to "low" fids
91  * only from corresponding high fids. This returns "high" fid corresponding to
92  * "low" one.
93  */
94 static u32 convert_fid_to_vco_fid(u32 fid)
95 {
96         if (fid < HI_FID_TABLE_BOTTOM)
97                 return 8 + (2 * fid);
98         else
99                 return fid;
100 }
101
102 /*
103  * Return 1 if the pending bit is set. Unless we just instructed the processor
104  * to transition to a new state, seeing this bit set is really bad news.
105  */
106 static int pending_bit_stuck(void)
107 {
108         u32 lo, hi;
109
110         if (cpu_family == CPU_HW_PSTATE)
111                 return 0;
112
113         rdmsr(MSR_FIDVID_STATUS, lo, hi);
114         return lo & MSR_S_LO_CHANGE_PENDING ? 1 : 0;
115 }
116
117 /*
118  * Update the global current fid / vid values from the status msr.
119  * Returns 1 on error.
120  */
121 static int query_current_values_with_pending_wait(struct powernow_k8_data *data)
122 {
123         u32 lo, hi;
124         u32 i = 0;
125
126         if (cpu_family == CPU_HW_PSTATE) {
127                 rdmsr(MSR_PSTATE_STATUS, lo, hi);
128                 i = lo & HW_PSTATE_MASK;
129                 data->currpstate = i;
130
131                 /*
132                  * a workaround for family 11h erratum 311 might cause
133                  * an "out-of-range Pstate if the core is in Pstate-0
134                  */
135                 if ((boot_cpu_data.x86 == 0x11) && (i >= data->numps))
136                         data->currpstate = HW_PSTATE_0;
137
138                 return 0;
139         }
140         do {
141                 if (i++ > 10000) {
142                         dprintk("detected change pending stuck\n");
143                         return 1;
144                 }
145                 rdmsr(MSR_FIDVID_STATUS, lo, hi);
146         } while (lo & MSR_S_LO_CHANGE_PENDING);
147
148         data->currvid = hi & MSR_S_HI_CURRENT_VID;
149         data->currfid = lo & MSR_S_LO_CURRENT_FID;
150
151         return 0;
152 }
153
154 /* the isochronous relief time */
155 static void count_off_irt(struct powernow_k8_data *data)
156 {
157         udelay((1 << data->irt) * 10);
158         return;
159 }
160
161 /* the voltage stabilization time */
162 static void count_off_vst(struct powernow_k8_data *data)
163 {
164         udelay(data->vstable * VST_UNITS_20US);
165         return;
166 }
167
168 /* need to init the control msr to a safe value (for each cpu) */
169 static void fidvid_msr_init(void)
170 {
171         u32 lo, hi;
172         u8 fid, vid;
173
174         rdmsr(MSR_FIDVID_STATUS, lo, hi);
175         vid = hi & MSR_S_HI_CURRENT_VID;
176         fid = lo & MSR_S_LO_CURRENT_FID;
177         lo = fid | (vid << MSR_C_LO_VID_SHIFT);
178         hi = MSR_C_HI_STP_GNT_BENIGN;
179         dprintk("cpu%d, init lo 0x%x, hi 0x%x\n", smp_processor_id(), lo, hi);
180         wrmsr(MSR_FIDVID_CTL, lo, hi);
181 }
182
183 /* write the new fid value along with the other control fields to the msr */
184 static int write_new_fid(struct powernow_k8_data *data, u32 fid)
185 {
186         u32 lo;
187         u32 savevid = data->currvid;
188         u32 i = 0;
189
190         if ((fid & INVALID_FID_MASK) || (data->currvid & INVALID_VID_MASK)) {
191                 printk(KERN_ERR PFX "internal error - overflow on fid write\n");
192                 return 1;
193         }
194
195         lo = fid;
196         lo |= (data->currvid << MSR_C_LO_VID_SHIFT);
197         lo |= MSR_C_LO_INIT_FID_VID;
198
199         dprintk("writing fid 0x%x, lo 0x%x, hi 0x%x\n",
200                 fid, lo, data->plllock * PLL_LOCK_CONVERSION);
201
202         do {
203                 wrmsr(MSR_FIDVID_CTL, lo, data->plllock * PLL_LOCK_CONVERSION);
204                 if (i++ > 100) {
205                         printk(KERN_ERR PFX
206                                 "Hardware error - pending bit very stuck - "
207                                 "no further pstate changes possible\n");
208                         return 1;
209                 }
210         } while (query_current_values_with_pending_wait(data));
211
212         count_off_irt(data);
213
214         if (savevid != data->currvid) {
215                 printk(KERN_ERR PFX
216                         "vid change on fid trans, old 0x%x, new 0x%x\n",
217                         savevid, data->currvid);
218                 return 1;
219         }
220
221         if (fid != data->currfid) {
222                 printk(KERN_ERR PFX
223                         "fid trans failed, fid 0x%x, curr 0x%x\n", fid,
224                         data->currfid);
225                 return 1;
226         }
227
228         return 0;
229 }
230
231 /* Write a new vid to the hardware */
232 static int write_new_vid(struct powernow_k8_data *data, u32 vid)
233 {
234         u32 lo;
235         u32 savefid = data->currfid;
236         int i = 0;
237
238         if ((data->currfid & INVALID_FID_MASK) || (vid & INVALID_VID_MASK)) {
239                 printk(KERN_ERR PFX "internal error - overflow on vid write\n");
240                 return 1;
241         }
242
243         lo = data->currfid;
244         lo |= (vid << MSR_C_LO_VID_SHIFT);
245         lo |= MSR_C_LO_INIT_FID_VID;
246
247         dprintk("writing vid 0x%x, lo 0x%x, hi 0x%x\n",
248                 vid, lo, STOP_GRANT_5NS);
249
250         do {
251                 wrmsr(MSR_FIDVID_CTL, lo, STOP_GRANT_5NS);
252                 if (i++ > 100) {
253                         printk(KERN_ERR PFX "internal error - pending bit "
254                                         "very stuck - no further pstate "
255                                         "changes possible\n");
256                         return 1;
257                 }
258         } while (query_current_values_with_pending_wait(data));
259
260         if (savefid != data->currfid) {
261                 printk(KERN_ERR PFX "fid changed on vid trans, old "
262                         "0x%x new 0x%x\n",
263                        savefid, data->currfid);
264                 return 1;
265         }
266
267         if (vid != data->currvid) {
268                 printk(KERN_ERR PFX "vid trans failed, vid 0x%x, "
269                                 "curr 0x%x\n",
270                                 vid, data->currvid);
271                 return 1;
272         }
273
274         return 0;
275 }
276
277 /*
278  * Reduce the vid by the max of step or reqvid.
279  * Decreasing vid codes represent increasing voltages:
280  * vid of 0 is 1.550V, vid of 0x1e is 0.800V, vid of VID_OFF is off.
281  */
282 static int decrease_vid_code_by_step(struct powernow_k8_data *data,
283                 u32 reqvid, u32 step)
284 {
285         if ((data->currvid - reqvid) > step)
286                 reqvid = data->currvid - step;
287
288         if (write_new_vid(data, reqvid))
289                 return 1;
290
291         count_off_vst(data);
292
293         return 0;
294 }
295
296 /* Change hardware pstate by single MSR write */
297 static int transition_pstate(struct powernow_k8_data *data, u32 pstate)
298 {
299         wrmsr(MSR_PSTATE_CTRL, pstate, 0);
300         data->currpstate = pstate;
301         return 0;
302 }
303
304 /* Change Opteron/Athlon64 fid and vid, by the 3 phases. */
305 static int transition_fid_vid(struct powernow_k8_data *data,
306                 u32 reqfid, u32 reqvid)
307 {
308         if (core_voltage_pre_transition(data, reqvid, reqfid))
309                 return 1;
310
311         if (core_frequency_transition(data, reqfid))
312                 return 1;
313
314         if (core_voltage_post_transition(data, reqvid))
315                 return 1;
316
317         if (query_current_values_with_pending_wait(data))
318                 return 1;
319
320         if ((reqfid != data->currfid) || (reqvid != data->currvid)) {
321                 printk(KERN_ERR PFX "failed (cpu%d): req 0x%x 0x%x, "
322                                 "curr 0x%x 0x%x\n",
323                                 smp_processor_id(),
324                                 reqfid, reqvid, data->currfid, data->currvid);
325                 return 1;
326         }
327
328         dprintk("transitioned (cpu%d): new fid 0x%x, vid 0x%x\n",
329                 smp_processor_id(), data->currfid, data->currvid);
330
331         return 0;
332 }
333
334 /* Phase 1 - core voltage transition ... setup voltage */
335 static int core_voltage_pre_transition(struct powernow_k8_data *data,
336                 u32 reqvid, u32 reqfid)
337 {
338         u32 rvosteps = data->rvo;
339         u32 savefid = data->currfid;
340         u32 maxvid, lo, rvomult = 1;
341
342         dprintk("ph1 (cpu%d): start, currfid 0x%x, currvid 0x%x, "
343                 "reqvid 0x%x, rvo 0x%x\n",
344                 smp_processor_id(),
345                 data->currfid, data->currvid, reqvid, data->rvo);
346
347         if ((savefid < LO_FID_TABLE_TOP) && (reqfid < LO_FID_TABLE_TOP))
348                 rvomult = 2;
349         rvosteps *= rvomult;
350         rdmsr(MSR_FIDVID_STATUS, lo, maxvid);
351         maxvid = 0x1f & (maxvid >> 16);
352         dprintk("ph1 maxvid=0x%x\n", maxvid);
353         if (reqvid < maxvid) /* lower numbers are higher voltages */
354                 reqvid = maxvid;
355
356         while (data->currvid > reqvid) {
357                 dprintk("ph1: curr 0x%x, req vid 0x%x\n",
358                         data->currvid, reqvid);
359                 if (decrease_vid_code_by_step(data, reqvid, data->vidmvs))
360                         return 1;
361         }
362
363         while ((rvosteps > 0) &&
364                         ((rvomult * data->rvo + data->currvid) > reqvid)) {
365                 if (data->currvid == maxvid) {
366                         rvosteps = 0;
367                 } else {
368                         dprintk("ph1: changing vid for rvo, req 0x%x\n",
369                                 data->currvid - 1);
370                         if (decrease_vid_code_by_step(data, data->currvid-1, 1))
371                                 return 1;
372                         rvosteps--;
373                 }
374         }
375
376         if (query_current_values_with_pending_wait(data))
377                 return 1;
378
379         if (savefid != data->currfid) {
380                 printk(KERN_ERR PFX "ph1 err, currfid changed 0x%x\n",
381                                 data->currfid);
382                 return 1;
383         }
384
385         dprintk("ph1 complete, currfid 0x%x, currvid 0x%x\n",
386                 data->currfid, data->currvid);
387
388         return 0;
389 }
390
391 /* Phase 2 - core frequency transition */
392 static int core_frequency_transition(struct powernow_k8_data *data, u32 reqfid)
393 {
394         u32 vcoreqfid, vcocurrfid, vcofiddiff;
395         u32 fid_interval, savevid = data->currvid;
396
397         if (data->currfid == reqfid) {
398                 printk(KERN_ERR PFX "ph2 null fid transition 0x%x\n",
399                                 data->currfid);
400                 return 0;
401         }
402
403         dprintk("ph2 (cpu%d): starting, currfid 0x%x, currvid 0x%x, "
404                 "reqfid 0x%x\n",
405                 smp_processor_id(),
406                 data->currfid, data->currvid, reqfid);
407
408         vcoreqfid = convert_fid_to_vco_fid(reqfid);
409         vcocurrfid = convert_fid_to_vco_fid(data->currfid);
410         vcofiddiff = vcocurrfid > vcoreqfid ? vcocurrfid - vcoreqfid
411             : vcoreqfid - vcocurrfid;
412
413         if ((reqfid <= LO_FID_TABLE_TOP) && (data->currfid <= LO_FID_TABLE_TOP))
414                 vcofiddiff = 0;
415
416         while (vcofiddiff > 2) {
417                 (data->currfid & 1) ? (fid_interval = 1) : (fid_interval = 2);
418
419                 if (reqfid > data->currfid) {
420                         if (data->currfid > LO_FID_TABLE_TOP) {
421                                 if (write_new_fid(data,
422                                                 data->currfid + fid_interval))
423                                         return 1;
424                         } else {
425                                 if (write_new_fid
426                                     (data,
427                                      2 + convert_fid_to_vco_fid(data->currfid)))
428                                         return 1;
429                         }
430                 } else {
431                         if (write_new_fid(data, data->currfid - fid_interval))
432                                 return 1;
433                 }
434
435                 vcocurrfid = convert_fid_to_vco_fid(data->currfid);
436                 vcofiddiff = vcocurrfid > vcoreqfid ? vcocurrfid - vcoreqfid
437                     : vcoreqfid - vcocurrfid;
438         }
439
440         if (write_new_fid(data, reqfid))
441                 return 1;
442
443         if (query_current_values_with_pending_wait(data))
444                 return 1;
445
446         if (data->currfid != reqfid) {
447                 printk(KERN_ERR PFX
448                         "ph2: mismatch, failed fid transition, "
449                         "curr 0x%x, req 0x%x\n",
450                         data->currfid, reqfid);
451                 return 1;
452         }
453
454         if (savevid != data->currvid) {
455                 printk(KERN_ERR PFX "ph2: vid changed, save 0x%x, curr 0x%x\n",
456                         savevid, data->currvid);
457                 return 1;
458         }
459
460         dprintk("ph2 complete, currfid 0x%x, currvid 0x%x\n",
461                 data->currfid, data->currvid);
462
463         return 0;
464 }
465
466 /* Phase 3 - core voltage transition flow ... jump to the final vid. */
467 static int core_voltage_post_transition(struct powernow_k8_data *data,
468                 u32 reqvid)
469 {
470         u32 savefid = data->currfid;
471         u32 savereqvid = reqvid;
472
473         dprintk("ph3 (cpu%d): starting, currfid 0x%x, currvid 0x%x\n",
474                 smp_processor_id(),
475                 data->currfid, data->currvid);
476
477         if (reqvid != data->currvid) {
478                 if (write_new_vid(data, reqvid))
479                         return 1;
480
481                 if (savefid != data->currfid) {
482                         printk(KERN_ERR PFX
483                                "ph3: bad fid change, save 0x%x, curr 0x%x\n",
484                                savefid, data->currfid);
485                         return 1;
486                 }
487
488                 if (data->currvid != reqvid) {
489                         printk(KERN_ERR PFX
490                                "ph3: failed vid transition\n, "
491                                "req 0x%x, curr 0x%x",
492                                reqvid, data->currvid);
493                         return 1;
494                 }
495         }
496
497         if (query_current_values_with_pending_wait(data))
498                 return 1;
499
500         if (savereqvid != data->currvid) {
501                 dprintk("ph3 failed, currvid 0x%x\n", data->currvid);
502                 return 1;
503         }
504
505         if (savefid != data->currfid) {
506                 dprintk("ph3 failed, currfid changed 0x%x\n",
507                         data->currfid);
508                 return 1;
509         }
510
511         dprintk("ph3 complete, currfid 0x%x, currvid 0x%x\n",
512                 data->currfid, data->currvid);
513
514         return 0;
515 }
516
517 static void check_supported_cpu(void *_rc)
518 {
519         u32 eax, ebx, ecx, edx;
520         int *rc = _rc;
521
522         *rc = -ENODEV;
523
524         if (current_cpu_data.x86_vendor != X86_VENDOR_AMD)
525                 return;
526
527         eax = cpuid_eax(CPUID_PROCESSOR_SIGNATURE);
528         if (((eax & CPUID_XFAM) != CPUID_XFAM_K8) &&
529             ((eax & CPUID_XFAM) < CPUID_XFAM_10H))
530                 return;
531
532         if ((eax & CPUID_XFAM) == CPUID_XFAM_K8) {
533                 if (((eax & CPUID_USE_XFAM_XMOD) != CPUID_USE_XFAM_XMOD) ||
534                     ((eax & CPUID_XMOD) > CPUID_XMOD_REV_MASK)) {
535                         printk(KERN_INFO PFX
536                                 "Processor cpuid %x not supported\n", eax);
537                         return;
538                 }
539
540                 eax = cpuid_eax(CPUID_GET_MAX_CAPABILITIES);
541                 if (eax < CPUID_FREQ_VOLT_CAPABILITIES) {
542                         printk(KERN_INFO PFX
543                                "No frequency change capabilities detected\n");
544                         return;
545                 }
546
547                 cpuid(CPUID_FREQ_VOLT_CAPABILITIES, &eax, &ebx, &ecx, &edx);
548                 if ((edx & P_STATE_TRANSITION_CAPABLE)
549                         != P_STATE_TRANSITION_CAPABLE) {
550                         printk(KERN_INFO PFX
551                                 "Power state transitions not supported\n");
552                         return;
553                 }
554         } else { /* must be a HW Pstate capable processor */
555                 cpuid(CPUID_FREQ_VOLT_CAPABILITIES, &eax, &ebx, &ecx, &edx);
556                 if ((edx & USE_HW_PSTATE) == USE_HW_PSTATE)
557                         cpu_family = CPU_HW_PSTATE;
558                 else
559                         return;
560         }
561
562         *rc = 0;
563 }
564
565 static int check_pst_table(struct powernow_k8_data *data, struct pst_s *pst,
566                 u8 maxvid)
567 {
568         unsigned int j;
569         u8 lastfid = 0xff;
570
571         for (j = 0; j < data->numps; j++) {
572                 if (pst[j].vid > LEAST_VID) {
573                         printk(KERN_ERR FW_BUG PFX "vid %d invalid : 0x%x\n",
574                                j, pst[j].vid);
575                         return -EINVAL;
576                 }
577                 if (pst[j].vid < data->rvo) {
578                         /* vid + rvo >= 0 */
579                         printk(KERN_ERR FW_BUG PFX "0 vid exceeded with pstate"
580                                " %d\n", j);
581                         return -ENODEV;
582                 }
583                 if (pst[j].vid < maxvid + data->rvo) {
584                         /* vid + rvo >= maxvid */
585                         printk(KERN_ERR FW_BUG PFX "maxvid exceeded with pstate"
586                                " %d\n", j);
587                         return -ENODEV;
588                 }
589                 if (pst[j].fid > MAX_FID) {
590                         printk(KERN_ERR FW_BUG PFX "maxfid exceeded with pstate"
591                                " %d\n", j);
592                         return -ENODEV;
593                 }
594                 if (j && (pst[j].fid < HI_FID_TABLE_BOTTOM)) {
595                         /* Only first fid is allowed to be in "low" range */
596                         printk(KERN_ERR FW_BUG PFX "two low fids - %d : "
597                                "0x%x\n", j, pst[j].fid);
598                         return -EINVAL;
599                 }
600                 if (pst[j].fid < lastfid)
601                         lastfid = pst[j].fid;
602         }
603         if (lastfid & 1) {
604                 printk(KERN_ERR FW_BUG PFX "lastfid invalid\n");
605                 return -EINVAL;
606         }
607         if (lastfid > LO_FID_TABLE_TOP)
608                 printk(KERN_INFO FW_BUG PFX
609                         "first fid not from lo freq table\n");
610
611         return 0;
612 }
613
614 static void invalidate_entry(struct cpufreq_frequency_table *powernow_table,
615                 unsigned int entry)
616 {
617         powernow_table[entry].frequency = CPUFREQ_ENTRY_INVALID;
618 }
619
620 static void print_basics(struct powernow_k8_data *data)
621 {
622         int j;
623         for (j = 0; j < data->numps; j++) {
624                 if (data->powernow_table[j].frequency !=
625                                 CPUFREQ_ENTRY_INVALID) {
626                         if (cpu_family == CPU_HW_PSTATE) {
627                                 printk(KERN_INFO PFX
628                                         "   %d : pstate %d (%d MHz)\n", j,
629                                         data->powernow_table[j].index,
630                                         data->powernow_table[j].frequency/1000);
631                         } else {
632                                 printk(KERN_INFO PFX
633                                         "   %d : fid 0x%x (%d MHz), vid 0x%x\n",
634                                         j,
635                                         data->powernow_table[j].index & 0xff,
636                                         data->powernow_table[j].frequency/1000,
637                                         data->powernow_table[j].index >> 8);
638                         }
639                 }
640         }
641         if (data->batps)
642                 printk(KERN_INFO PFX "Only %d pstates on battery\n",
643                                 data->batps);
644 }
645
646 static u32 freq_from_fid_did(u32 fid, u32 did)
647 {
648         u32 mhz = 0;
649
650         if (boot_cpu_data.x86 == 0x10)
651                 mhz = (100 * (fid + 0x10)) >> did;
652         else if (boot_cpu_data.x86 == 0x11)
653                 mhz = (100 * (fid + 8)) >> did;
654         else
655                 BUG();
656
657         return mhz * 1000;
658 }
659
660 static int fill_powernow_table(struct powernow_k8_data *data,
661                 struct pst_s *pst, u8 maxvid)
662 {
663         struct cpufreq_frequency_table *powernow_table;
664         unsigned int j;
665
666         if (data->batps) {
667                 /* use ACPI support to get full speed on mains power */
668                 printk(KERN_WARNING PFX
669                         "Only %d pstates usable (use ACPI driver for full "
670                         "range\n", data->batps);
671                 data->numps = data->batps;
672         }
673
674         for (j = 1; j < data->numps; j++) {
675                 if (pst[j-1].fid >= pst[j].fid) {
676                         printk(KERN_ERR PFX "PST out of sequence\n");
677                         return -EINVAL;
678                 }
679         }
680
681         if (data->numps < 2) {
682                 printk(KERN_ERR PFX "no p states to transition\n");
683                 return -ENODEV;
684         }
685
686         if (check_pst_table(data, pst, maxvid))
687                 return -EINVAL;
688
689         powernow_table = kmalloc((sizeof(struct cpufreq_frequency_table)
690                 * (data->numps + 1)), GFP_KERNEL);
691         if (!powernow_table) {
692                 printk(KERN_ERR PFX "powernow_table memory alloc failure\n");
693                 return -ENOMEM;
694         }
695
696         for (j = 0; j < data->numps; j++) {
697                 int freq;
698                 powernow_table[j].index = pst[j].fid; /* lower 8 bits */
699                 powernow_table[j].index |= (pst[j].vid << 8); /* upper 8 bits */
700                 freq = find_khz_freq_from_fid(pst[j].fid);
701                 powernow_table[j].frequency = freq;
702         }
703         powernow_table[data->numps].frequency = CPUFREQ_TABLE_END;
704         powernow_table[data->numps].index = 0;
705
706         if (query_current_values_with_pending_wait(data)) {
707                 kfree(powernow_table);
708                 return -EIO;
709         }
710
711         dprintk("cfid 0x%x, cvid 0x%x\n", data->currfid, data->currvid);
712         data->powernow_table = powernow_table;
713         if (cpumask_first(cpu_core_mask(data->cpu)) == data->cpu)
714                 print_basics(data);
715
716         for (j = 0; j < data->numps; j++)
717                 if ((pst[j].fid == data->currfid) &&
718                     (pst[j].vid == data->currvid))
719                         return 0;
720
721         dprintk("currfid/vid do not match PST, ignoring\n");
722         return 0;
723 }
724
725 /* Find and validate the PSB/PST table in BIOS. */
726 static int find_psb_table(struct powernow_k8_data *data)
727 {
728         struct psb_s *psb;
729         unsigned int i;
730         u32 mvs;
731         u8 maxvid;
732         u32 cpst = 0;
733         u32 thiscpuid;
734
735         for (i = 0xc0000; i < 0xffff0; i += 0x10) {
736                 /* Scan BIOS looking for the signature. */
737                 /* It can not be at ffff0 - it is too big. */
738
739                 psb = phys_to_virt(i);
740                 if (memcmp(psb, PSB_ID_STRING, PSB_ID_STRING_LEN) != 0)
741                         continue;
742
743                 dprintk("found PSB header at 0x%p\n", psb);
744
745                 dprintk("table vers: 0x%x\n", psb->tableversion);
746                 if (psb->tableversion != PSB_VERSION_1_4) {
747                         printk(KERN_ERR FW_BUG PFX "PSB table is not v1.4\n");
748                         return -ENODEV;
749                 }
750
751                 dprintk("flags: 0x%x\n", psb->flags1);
752                 if (psb->flags1) {
753                         printk(KERN_ERR FW_BUG PFX "unknown flags\n");
754                         return -ENODEV;
755                 }
756
757                 data->vstable = psb->vstable;
758                 dprintk("voltage stabilization time: %d(*20us)\n",
759                                 data->vstable);
760
761                 dprintk("flags2: 0x%x\n", psb->flags2);
762                 data->rvo = psb->flags2 & 3;
763                 data->irt = ((psb->flags2) >> 2) & 3;
764                 mvs = ((psb->flags2) >> 4) & 3;
765                 data->vidmvs = 1 << mvs;
766                 data->batps = ((psb->flags2) >> 6) & 3;
767
768                 dprintk("ramp voltage offset: %d\n", data->rvo);
769                 dprintk("isochronous relief time: %d\n", data->irt);
770                 dprintk("maximum voltage step: %d - 0x%x\n", mvs, data->vidmvs);
771
772                 dprintk("numpst: 0x%x\n", psb->num_tables);
773                 cpst = psb->num_tables;
774                 if ((psb->cpuid == 0x00000fc0) ||
775                     (psb->cpuid == 0x00000fe0)) {
776                         thiscpuid = cpuid_eax(CPUID_PROCESSOR_SIGNATURE);
777                         if ((thiscpuid == 0x00000fc0) ||
778                             (thiscpuid == 0x00000fe0))
779                                 cpst = 1;
780                 }
781                 if (cpst != 1) {
782                         printk(KERN_ERR FW_BUG PFX "numpst must be 1\n");
783                         return -ENODEV;
784                 }
785
786                 data->plllock = psb->plllocktime;
787                 dprintk("plllocktime: 0x%x (units 1us)\n", psb->plllocktime);
788                 dprintk("maxfid: 0x%x\n", psb->maxfid);
789                 dprintk("maxvid: 0x%x\n", psb->maxvid);
790                 maxvid = psb->maxvid;
791
792                 data->numps = psb->numps;
793                 dprintk("numpstates: 0x%x\n", data->numps);
794                 return fill_powernow_table(data,
795                                 (struct pst_s *)(psb+1), maxvid);
796         }
797         /*
798          * If you see this message, complain to BIOS manufacturer. If
799          * he tells you "we do not support Linux" or some similar
800          * nonsense, remember that Windows 2000 uses the same legacy
801          * mechanism that the old Linux PSB driver uses. Tell them it
802          * is broken with Windows 2000.
803          *
804          * The reference to the AMD documentation is chapter 9 in the
805          * BIOS and Kernel Developer's Guide, which is available on
806          * www.amd.com
807          */
808         printk(KERN_ERR FW_BUG PFX "No PSB or ACPI _PSS objects\n");
809         return -ENODEV;
810 }
811
812 static void powernow_k8_acpi_pst_values(struct powernow_k8_data *data,
813                 unsigned int index)
814 {
815         u64 control;
816
817         if (!data->acpi_data.state_count || (cpu_family == CPU_HW_PSTATE))
818                 return;
819
820         control = data->acpi_data.states[index].control;
821         data->irt = (control >> IRT_SHIFT) & IRT_MASK;
822         data->rvo = (control >> RVO_SHIFT) & RVO_MASK;
823         data->exttype = (control >> EXT_TYPE_SHIFT) & EXT_TYPE_MASK;
824         data->plllock = (control >> PLL_L_SHIFT) & PLL_L_MASK;
825         data->vidmvs = 1 << ((control >> MVS_SHIFT) & MVS_MASK);
826         data->vstable = (control >> VST_SHIFT) & VST_MASK;
827 }
828
829 static int powernow_k8_cpu_init_acpi(struct powernow_k8_data *data)
830 {
831         struct cpufreq_frequency_table *powernow_table;
832         int ret_val = -ENODEV;
833         u64 control, status;
834
835         if (acpi_processor_register_performance(&data->acpi_data, data->cpu)) {
836                 dprintk("register performance failed: bad ACPI data\n");
837                 return -EIO;
838         }
839
840         /* verify the data contained in the ACPI structures */
841         if (data->acpi_data.state_count <= 1) {
842                 dprintk("No ACPI P-States\n");
843                 goto err_out;
844         }
845
846         control = data->acpi_data.control_register.space_id;
847         status = data->acpi_data.status_register.space_id;
848
849         if ((control != ACPI_ADR_SPACE_FIXED_HARDWARE) ||
850             (status != ACPI_ADR_SPACE_FIXED_HARDWARE)) {
851                 dprintk("Invalid control/status registers (%x - %x)\n",
852                         control, status);
853                 goto err_out;
854         }
855
856         /* fill in data->powernow_table */
857         powernow_table = kmalloc((sizeof(struct cpufreq_frequency_table)
858                 * (data->acpi_data.state_count + 1)), GFP_KERNEL);
859         if (!powernow_table) {
860                 dprintk("powernow_table memory alloc failure\n");
861                 goto err_out;
862         }
863
864         /* fill in data */
865         data->numps = data->acpi_data.state_count;
866         powernow_k8_acpi_pst_values(data, 0);
867
868         if (cpu_family == CPU_HW_PSTATE)
869                 ret_val = fill_powernow_table_pstate(data, powernow_table);
870         else
871                 ret_val = fill_powernow_table_fidvid(data, powernow_table);
872         if (ret_val)
873                 goto err_out_mem;
874
875         powernow_table[data->acpi_data.state_count].frequency =
876                 CPUFREQ_TABLE_END;
877         powernow_table[data->acpi_data.state_count].index = 0;
878         data->powernow_table = powernow_table;
879
880         if (cpumask_first(cpu_core_mask(data->cpu)) == data->cpu)
881                 print_basics(data);
882
883         /* notify BIOS that we exist */
884         acpi_processor_notify_smm(THIS_MODULE);
885
886         if (!zalloc_cpumask_var(&data->acpi_data.shared_cpu_map, GFP_KERNEL)) {
887                 printk(KERN_ERR PFX
888                                 "unable to alloc powernow_k8_data cpumask\n");
889                 ret_val = -ENOMEM;
890                 goto err_out_mem;
891         }
892
893         return 0;
894
895 err_out_mem:
896         kfree(powernow_table);
897
898 err_out:
899         acpi_processor_unregister_performance(&data->acpi_data, data->cpu);
900
901         /* data->acpi_data.state_count informs us at ->exit()
902          * whether ACPI was used */
903         data->acpi_data.state_count = 0;
904
905         return ret_val;
906 }
907
908 static int fill_powernow_table_pstate(struct powernow_k8_data *data,
909                 struct cpufreq_frequency_table *powernow_table)
910 {
911         int i;
912         u32 hi = 0, lo = 0;
913         rdmsr(MSR_PSTATE_CUR_LIMIT, hi, lo);
914         data->max_hw_pstate = (hi & HW_PSTATE_MAX_MASK) >> HW_PSTATE_MAX_SHIFT;
915
916         for (i = 0; i < data->acpi_data.state_count; i++) {
917                 u32 index;
918
919                 index = data->acpi_data.states[i].control & HW_PSTATE_MASK;
920                 if (index > data->max_hw_pstate) {
921                         printk(KERN_ERR PFX "invalid pstate %d - "
922                                         "bad value %d.\n", i, index);
923                         printk(KERN_ERR PFX "Please report to BIOS "
924                                         "manufacturer\n");
925                         invalidate_entry(powernow_table, i);
926                         continue;
927                 }
928                 rdmsr(MSR_PSTATE_DEF_BASE + index, lo, hi);
929                 if (!(hi & HW_PSTATE_VALID_MASK)) {
930                         dprintk("invalid pstate %d, ignoring\n", index);
931                         invalidate_entry(powernow_table, i);
932                         continue;
933                 }
934
935                 powernow_table[i].index = index;
936
937                 /* Frequency may be rounded for these */
938                 if ((boot_cpu_data.x86 == 0x10 && boot_cpu_data.x86_model < 10)
939                                  || boot_cpu_data.x86 == 0x11) {
940                         powernow_table[i].frequency =
941                                 freq_from_fid_did(lo & 0x3f, (lo >> 6) & 7);
942                 } else
943                         powernow_table[i].frequency =
944                                 data->acpi_data.states[i].core_frequency * 1000;
945         }
946         return 0;
947 }
948
949 static int fill_powernow_table_fidvid(struct powernow_k8_data *data,
950                 struct cpufreq_frequency_table *powernow_table)
951 {
952         int i;
953
954         for (i = 0; i < data->acpi_data.state_count; i++) {
955                 u32 fid;
956                 u32 vid;
957                 u32 freq, index;
958                 u64 status, control;
959
960                 if (data->exttype) {
961                         status =  data->acpi_data.states[i].status;
962                         fid = status & EXT_FID_MASK;
963                         vid = (status >> VID_SHIFT) & EXT_VID_MASK;
964                 } else {
965                         control =  data->acpi_data.states[i].control;
966                         fid = control & FID_MASK;
967                         vid = (control >> VID_SHIFT) & VID_MASK;
968                 }
969
970                 dprintk("   %d : fid 0x%x, vid 0x%x\n", i, fid, vid);
971
972                 index = fid | (vid<<8);
973                 powernow_table[i].index = index;
974
975                 freq = find_khz_freq_from_fid(fid);
976                 powernow_table[i].frequency = freq;
977
978                 /* verify frequency is OK */
979                 if ((freq > (MAX_FREQ * 1000)) || (freq < (MIN_FREQ * 1000))) {
980                         dprintk("invalid freq %u kHz, ignoring\n", freq);
981                         invalidate_entry(powernow_table, i);
982                         continue;
983                 }
984
985                 /* verify voltage is OK -
986                  * BIOSs are using "off" to indicate invalid */
987                 if (vid == VID_OFF) {
988                         dprintk("invalid vid %u, ignoring\n", vid);
989                         invalidate_entry(powernow_table, i);
990                         continue;
991                 }
992
993                 if (freq != (data->acpi_data.states[i].core_frequency * 1000)) {
994                         printk(KERN_INFO PFX "invalid freq entries "
995                                 "%u kHz vs. %u kHz\n", freq,
996                                 (unsigned int)
997                                 (data->acpi_data.states[i].core_frequency
998                                  * 1000));
999                         invalidate_entry(powernow_table, i);
1000                         continue;
1001                 }
1002         }
1003         return 0;
1004 }
1005
1006 static void powernow_k8_cpu_exit_acpi(struct powernow_k8_data *data)
1007 {
1008         if (data->acpi_data.state_count)
1009                 acpi_processor_unregister_performance(&data->acpi_data,
1010                                 data->cpu);
1011         free_cpumask_var(data->acpi_data.shared_cpu_map);
1012 }
1013
1014 static int get_transition_latency(struct powernow_k8_data *data)
1015 {
1016         int max_latency = 0;
1017         int i;
1018         for (i = 0; i < data->acpi_data.state_count; i++) {
1019                 int cur_latency = data->acpi_data.states[i].transition_latency
1020                         + data->acpi_data.states[i].bus_master_latency;
1021                 if (cur_latency > max_latency)
1022                         max_latency = cur_latency;
1023         }
1024         if (max_latency == 0) {
1025                 /*
1026                  * Fam 11h and later may return 0 as transition latency. This
1027                  * is intended and means "very fast". While cpufreq core and
1028                  * governors currently can handle that gracefully, better set it
1029                  * to 1 to avoid problems in the future.
1030                  */
1031                 if (boot_cpu_data.x86 < 0x11)
1032                         printk(KERN_ERR FW_WARN PFX "Invalid zero transition "
1033                                 "latency\n");
1034                 max_latency = 1;
1035         }
1036         /* value in usecs, needs to be in nanoseconds */
1037         return 1000 * max_latency;
1038 }
1039
1040 /* Take a frequency, and issue the fid/vid transition command */
1041 static int transition_frequency_fidvid(struct powernow_k8_data *data,
1042                 unsigned int index)
1043 {
1044         u32 fid = 0;
1045         u32 vid = 0;
1046         int res, i;
1047         struct cpufreq_freqs freqs;
1048
1049         dprintk("cpu %d transition to index %u\n", smp_processor_id(), index);
1050
1051         /* fid/vid correctness check for k8 */
1052         /* fid are the lower 8 bits of the index we stored into
1053          * the cpufreq frequency table in find_psb_table, vid
1054          * are the upper 8 bits.
1055          */
1056         fid = data->powernow_table[index].index & 0xFF;
1057         vid = (data->powernow_table[index].index & 0xFF00) >> 8;
1058
1059         dprintk("table matched fid 0x%x, giving vid 0x%x\n", fid, vid);
1060
1061         if (query_current_values_with_pending_wait(data))
1062                 return 1;
1063
1064         if ((data->currvid == vid) && (data->currfid == fid)) {
1065                 dprintk("target matches current values (fid 0x%x, vid 0x%x)\n",
1066                         fid, vid);
1067                 return 0;
1068         }
1069
1070         dprintk("cpu %d, changing to fid 0x%x, vid 0x%x\n",
1071                 smp_processor_id(), fid, vid);
1072         freqs.old = find_khz_freq_from_fid(data->currfid);
1073         freqs.new = find_khz_freq_from_fid(fid);
1074
1075         for_each_cpu(i, data->available_cores) {
1076                 freqs.cpu = i;
1077                 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
1078         }
1079
1080         res = transition_fid_vid(data, fid, vid);
1081         freqs.new = find_khz_freq_from_fid(data->currfid);
1082
1083         for_each_cpu(i, data->available_cores) {
1084                 freqs.cpu = i;
1085                 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
1086         }
1087         return res;
1088 }
1089
1090 /* Take a frequency, and issue the hardware pstate transition command */
1091 static int transition_frequency_pstate(struct powernow_k8_data *data,
1092                 unsigned int index)
1093 {
1094         u32 pstate = 0;
1095         int res, i;
1096         struct cpufreq_freqs freqs;
1097
1098         dprintk("cpu %d transition to index %u\n", smp_processor_id(), index);
1099
1100         /* get MSR index for hardware pstate transition */
1101         pstate = index & HW_PSTATE_MASK;
1102         if (pstate > data->max_hw_pstate)
1103                 return 0;
1104         freqs.old = find_khz_freq_from_pstate(data->powernow_table,
1105                         data->currpstate);
1106         freqs.new = find_khz_freq_from_pstate(data->powernow_table, pstate);
1107
1108         for_each_cpu(i, data->available_cores) {
1109                 freqs.cpu = i;
1110                 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
1111         }
1112
1113         res = transition_pstate(data, pstate);
1114         freqs.new = find_khz_freq_from_pstate(data->powernow_table, pstate);
1115
1116         for_each_cpu(i, data->available_cores) {
1117                 freqs.cpu = i;
1118                 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
1119         }
1120         return res;
1121 }
1122
1123 /* Driver entry point to switch to the target frequency */
1124 static int powernowk8_target(struct cpufreq_policy *pol,
1125                 unsigned targfreq, unsigned relation)
1126 {
1127         cpumask_var_t oldmask;
1128         struct powernow_k8_data *data = per_cpu(powernow_data, pol->cpu);
1129         u32 checkfid;
1130         u32 checkvid;
1131         unsigned int newstate;
1132         int ret = -EIO;
1133
1134         if (!data)
1135                 return -EINVAL;
1136
1137         checkfid = data->currfid;
1138         checkvid = data->currvid;
1139
1140         /* only run on specific CPU from here on. */
1141         /* This is poor form: use a workqueue or smp_call_function_single */
1142         if (!alloc_cpumask_var(&oldmask, GFP_KERNEL))
1143                 return -ENOMEM;
1144
1145         cpumask_copy(oldmask, tsk_cpus_allowed(current));
1146         set_cpus_allowed_ptr(current, cpumask_of(pol->cpu));
1147
1148         if (smp_processor_id() != pol->cpu) {
1149                 printk(KERN_ERR PFX "limiting to cpu %u failed\n", pol->cpu);
1150                 goto err_out;
1151         }
1152
1153         if (pending_bit_stuck()) {
1154                 printk(KERN_ERR PFX "failing targ, change pending bit set\n");
1155                 goto err_out;
1156         }
1157
1158         dprintk("targ: cpu %d, %d kHz, min %d, max %d, relation %d\n",
1159                 pol->cpu, targfreq, pol->min, pol->max, relation);
1160
1161         if (query_current_values_with_pending_wait(data))
1162                 goto err_out;
1163
1164         if (cpu_family != CPU_HW_PSTATE) {
1165                 dprintk("targ: curr fid 0x%x, vid 0x%x\n",
1166                 data->currfid, data->currvid);
1167
1168                 if ((checkvid != data->currvid) ||
1169                     (checkfid != data->currfid)) {
1170                         printk(KERN_INFO PFX
1171                                 "error - out of sync, fix 0x%x 0x%x, "
1172                                 "vid 0x%x 0x%x\n",
1173                                 checkfid, data->currfid,
1174                                 checkvid, data->currvid);
1175                 }
1176         }
1177
1178         if (cpufreq_frequency_table_target(pol, data->powernow_table,
1179                                 targfreq, relation, &newstate))
1180                 goto err_out;
1181
1182         mutex_lock(&fidvid_mutex);
1183
1184         powernow_k8_acpi_pst_values(data, newstate);
1185
1186         if (cpu_family == CPU_HW_PSTATE)
1187                 ret = transition_frequency_pstate(data, newstate);
1188         else
1189                 ret = transition_frequency_fidvid(data, newstate);
1190         if (ret) {
1191                 printk(KERN_ERR PFX "transition frequency failed\n");
1192                 ret = 1;
1193                 mutex_unlock(&fidvid_mutex);
1194                 goto err_out;
1195         }
1196         mutex_unlock(&fidvid_mutex);
1197
1198         if (cpu_family == CPU_HW_PSTATE)
1199                 pol->cur = find_khz_freq_from_pstate(data->powernow_table,
1200                                 newstate);
1201         else
1202                 pol->cur = find_khz_freq_from_fid(data->currfid);
1203         ret = 0;
1204
1205 err_out:
1206         set_cpus_allowed_ptr(current, oldmask);
1207         free_cpumask_var(oldmask);
1208         return ret;
1209 }
1210
1211 /* Driver entry point to verify the policy and range of frequencies */
1212 static int powernowk8_verify(struct cpufreq_policy *pol)
1213 {
1214         struct powernow_k8_data *data = per_cpu(powernow_data, pol->cpu);
1215
1216         if (!data)
1217                 return -EINVAL;
1218
1219         return cpufreq_frequency_table_verify(pol, data->powernow_table);
1220 }
1221
1222 struct init_on_cpu {
1223         struct powernow_k8_data *data;
1224         int rc;
1225 };
1226
1227 static void __cpuinit powernowk8_cpu_init_on_cpu(void *_init_on_cpu)
1228 {
1229         struct init_on_cpu *init_on_cpu = _init_on_cpu;
1230
1231         if (pending_bit_stuck()) {
1232                 printk(KERN_ERR PFX "failing init, change pending bit set\n");
1233                 init_on_cpu->rc = -ENODEV;
1234                 return;
1235         }
1236
1237         if (query_current_values_with_pending_wait(init_on_cpu->data)) {
1238                 init_on_cpu->rc = -ENODEV;
1239                 return;
1240         }
1241
1242         if (cpu_family == CPU_OPTERON)
1243                 fidvid_msr_init();
1244
1245         init_on_cpu->rc = 0;
1246 }
1247
1248 /* per CPU init entry point to the driver */
1249 static int __cpuinit powernowk8_cpu_init(struct cpufreq_policy *pol)
1250 {
1251         static const char ACPI_PSS_BIOS_BUG_MSG[] =
1252                 KERN_ERR FW_BUG PFX "No compatible ACPI _PSS objects found.\n"
1253                 FW_BUG PFX "Try again with latest BIOS.\n";
1254         struct powernow_k8_data *data;
1255         struct init_on_cpu init_on_cpu;
1256         int rc;
1257         struct cpuinfo_x86 *c = &cpu_data(pol->cpu);
1258
1259         if (!cpu_online(pol->cpu))
1260                 return -ENODEV;
1261
1262         smp_call_function_single(pol->cpu, check_supported_cpu, &rc, 1);
1263         if (rc)
1264                 return -ENODEV;
1265
1266         data = kzalloc(sizeof(struct powernow_k8_data), GFP_KERNEL);
1267         if (!data) {
1268                 printk(KERN_ERR PFX "unable to alloc powernow_k8_data");
1269                 return -ENOMEM;
1270         }
1271
1272         data->cpu = pol->cpu;
1273         data->currpstate = HW_PSTATE_INVALID;
1274
1275         if (powernow_k8_cpu_init_acpi(data)) {
1276                 /*
1277                  * Use the PSB BIOS structure. This is only availabe on
1278                  * an UP version, and is deprecated by AMD.
1279                  */
1280                 if (num_online_cpus() != 1) {
1281                         printk_once(ACPI_PSS_BIOS_BUG_MSG);
1282                         goto err_out;
1283                 }
1284                 if (pol->cpu != 0) {
1285                         printk(KERN_ERR FW_BUG PFX "No ACPI _PSS objects for "
1286                                "CPU other than CPU0. Complain to your BIOS "
1287                                "vendor.\n");
1288                         goto err_out;
1289                 }
1290                 rc = find_psb_table(data);
1291                 if (rc)
1292                         goto err_out;
1293
1294                 /* Take a crude guess here.
1295                  * That guess was in microseconds, so multiply with 1000 */
1296                 pol->cpuinfo.transition_latency = (
1297                          ((data->rvo + 8) * data->vstable * VST_UNITS_20US) +
1298                          ((1 << data->irt) * 30)) * 1000;
1299         } else /* ACPI _PSS objects available */
1300                 pol->cpuinfo.transition_latency = get_transition_latency(data);
1301
1302         /* only run on specific CPU from here on */
1303         init_on_cpu.data = data;
1304         smp_call_function_single(data->cpu, powernowk8_cpu_init_on_cpu,
1305                                  &init_on_cpu, 1);
1306         rc = init_on_cpu.rc;
1307         if (rc != 0)
1308                 goto err_out_exit_acpi;
1309
1310         if (cpu_family == CPU_HW_PSTATE)
1311                 cpumask_copy(pol->cpus, cpumask_of(pol->cpu));
1312         else
1313                 cpumask_copy(pol->cpus, cpu_core_mask(pol->cpu));
1314         data->available_cores = pol->cpus;
1315
1316         if (cpu_family == CPU_HW_PSTATE)
1317                 pol->cur = find_khz_freq_from_pstate(data->powernow_table,
1318                                 data->currpstate);
1319         else
1320                 pol->cur = find_khz_freq_from_fid(data->currfid);
1321         dprintk("policy current frequency %d kHz\n", pol->cur);
1322
1323         /* min/max the cpu is capable of */
1324         if (cpufreq_frequency_table_cpuinfo(pol, data->powernow_table)) {
1325                 printk(KERN_ERR FW_BUG PFX "invalid powernow_table\n");
1326                 powernow_k8_cpu_exit_acpi(data);
1327                 kfree(data->powernow_table);
1328                 kfree(data);
1329                 return -EINVAL;
1330         }
1331
1332         /* Check for APERF/MPERF support in hardware */
1333         if (cpu_has(c, X86_FEATURE_APERFMPERF))
1334                 cpufreq_amd64_driver.getavg = cpufreq_get_measured_perf;
1335
1336         cpufreq_frequency_table_get_attr(data->powernow_table, pol->cpu);
1337
1338         if (cpu_family == CPU_HW_PSTATE)
1339                 dprintk("cpu_init done, current pstate 0x%x\n",
1340                                 data->currpstate);
1341         else
1342                 dprintk("cpu_init done, current fid 0x%x, vid 0x%x\n",
1343                         data->currfid, data->currvid);
1344
1345         per_cpu(powernow_data, pol->cpu) = data;
1346
1347         return 0;
1348
1349 err_out_exit_acpi:
1350         powernow_k8_cpu_exit_acpi(data);
1351
1352 err_out:
1353         kfree(data);
1354         return -ENODEV;
1355 }
1356
1357 static int __devexit powernowk8_cpu_exit(struct cpufreq_policy *pol)
1358 {
1359         struct powernow_k8_data *data = per_cpu(powernow_data, pol->cpu);
1360
1361         if (!data)
1362                 return -EINVAL;
1363
1364         powernow_k8_cpu_exit_acpi(data);
1365
1366         cpufreq_frequency_table_put_attr(pol->cpu);
1367
1368         kfree(data->powernow_table);
1369         kfree(data);
1370         per_cpu(powernow_data, pol->cpu) = NULL;
1371
1372         return 0;
1373 }
1374
1375 static void query_values_on_cpu(void *_err)
1376 {
1377         int *err = _err;
1378         struct powernow_k8_data *data = __get_cpu_var(powernow_data);
1379
1380         *err = query_current_values_with_pending_wait(data);
1381 }
1382
1383 static unsigned int powernowk8_get(unsigned int cpu)
1384 {
1385         struct powernow_k8_data *data = per_cpu(powernow_data, cpu);
1386         unsigned int khz = 0;
1387         int err;
1388
1389         if (!data)
1390                 return 0;
1391
1392         smp_call_function_single(cpu, query_values_on_cpu, &err, true);
1393         if (err)
1394                 goto out;
1395
1396         if (cpu_family == CPU_HW_PSTATE)
1397                 khz = find_khz_freq_from_pstate(data->powernow_table,
1398                                                 data->currpstate);
1399         else
1400                 khz = find_khz_freq_from_fid(data->currfid);
1401
1402
1403 out:
1404         return khz;
1405 }
1406
1407 static void _cpb_toggle_msrs(bool t)
1408 {
1409         int cpu;
1410
1411         get_online_cpus();
1412
1413         rdmsr_on_cpus(cpu_online_mask, MSR_K7_HWCR, msrs);
1414
1415         for_each_cpu(cpu, cpu_online_mask) {
1416                 struct msr *reg = per_cpu_ptr(msrs, cpu);
1417                 if (t)
1418                         reg->l &= ~BIT(25);
1419                 else
1420                         reg->l |= BIT(25);
1421         }
1422         wrmsr_on_cpus(cpu_online_mask, MSR_K7_HWCR, msrs);
1423
1424         put_online_cpus();
1425 }
1426
1427 /*
1428  * Switch on/off core performance boosting.
1429  *
1430  * 0=disable
1431  * 1=enable.
1432  */
1433 static void cpb_toggle(bool t)
1434 {
1435         if (!cpb_capable)
1436                 return;
1437
1438         if (t && !cpb_enabled) {
1439                 cpb_enabled = true;
1440                 _cpb_toggle_msrs(t);
1441                 printk(KERN_INFO PFX "Core Boosting enabled.\n");
1442         } else if (!t && cpb_enabled) {
1443                 cpb_enabled = false;
1444                 _cpb_toggle_msrs(t);
1445                 printk(KERN_INFO PFX "Core Boosting disabled.\n");
1446         }
1447 }
1448
1449 static ssize_t store_cpb(struct cpufreq_policy *policy, const char *buf,
1450                                  size_t count)
1451 {
1452         int ret = -EINVAL;
1453         unsigned long val = 0;
1454
1455         ret = strict_strtoul(buf, 10, &val);
1456         if (!ret && (val == 0 || val == 1) && cpb_capable)
1457                 cpb_toggle(val);
1458         else
1459                 return -EINVAL;
1460
1461         return count;
1462 }
1463
1464 static ssize_t show_cpb(struct cpufreq_policy *policy, char *buf)
1465 {
1466         return sprintf(buf, "%u\n", cpb_enabled);
1467 }
1468
1469 #define define_one_rw(_name) \
1470 static struct freq_attr _name = \
1471 __ATTR(_name, 0644, show_##_name, store_##_name)
1472
1473 define_one_rw(cpb);
1474
1475 static struct freq_attr *powernow_k8_attr[] = {
1476         &cpufreq_freq_attr_scaling_available_freqs,
1477         &cpb,
1478         NULL,
1479 };
1480
1481 static struct cpufreq_driver cpufreq_amd64_driver = {
1482         .verify         = powernowk8_verify,
1483         .target         = powernowk8_target,
1484         .bios_limit     = acpi_processor_get_bios_limit,
1485         .init           = powernowk8_cpu_init,
1486         .exit           = __devexit_p(powernowk8_cpu_exit),
1487         .get            = powernowk8_get,
1488         .name           = "powernow-k8",
1489         .owner          = THIS_MODULE,
1490         .attr           = powernow_k8_attr,
1491 };
1492
1493 /*
1494  * Clear the boost-disable flag on the CPU_DOWN path so that this cpu
1495  * cannot block the remaining ones from boosting. On the CPU_UP path we
1496  * simply keep the boost-disable flag in sync with the current global
1497  * state.
1498  */
1499 static int cpb_notify(struct notifier_block *nb, unsigned long action,
1500                       void *hcpu)
1501 {
1502         unsigned cpu = (long)hcpu;
1503         u32 lo, hi;
1504
1505         switch (action) {
1506         case CPU_UP_PREPARE:
1507         case CPU_UP_PREPARE_FROZEN:
1508
1509                 if (!cpb_enabled) {
1510                         rdmsr_on_cpu(cpu, MSR_K7_HWCR, &lo, &hi);
1511                         lo |= BIT(25);
1512                         wrmsr_on_cpu(cpu, MSR_K7_HWCR, lo, hi);
1513                 }
1514                 break;
1515
1516         case CPU_DOWN_PREPARE:
1517         case CPU_DOWN_PREPARE_FROZEN:
1518                 rdmsr_on_cpu(cpu, MSR_K7_HWCR, &lo, &hi);
1519                 lo &= ~BIT(25);
1520                 wrmsr_on_cpu(cpu, MSR_K7_HWCR, lo, hi);
1521                 break;
1522
1523         default:
1524                 break;
1525         }
1526
1527         return NOTIFY_OK;
1528 }
1529
1530 static struct notifier_block cpb_nb = {
1531         .notifier_call          = cpb_notify,
1532 };
1533
1534 /* driver entry point for init */
1535 static int __cpuinit powernowk8_init(void)
1536 {
1537         unsigned int i, supported_cpus = 0, cpu;
1538
1539         for_each_online_cpu(i) {
1540                 int rc;
1541                 smp_call_function_single(i, check_supported_cpu, &rc, 1);
1542                 if (rc == 0)
1543                         supported_cpus++;
1544         }
1545
1546         if (supported_cpus != num_online_cpus())
1547                 return -ENODEV;
1548
1549         printk(KERN_INFO PFX "Found %d %s (%d cpu cores) (" VERSION ")\n",
1550                 num_online_nodes(), boot_cpu_data.x86_model_id, supported_cpus);
1551
1552         if (boot_cpu_has(X86_FEATURE_CPB)) {
1553
1554                 cpb_capable = true;
1555
1556                 register_cpu_notifier(&cpb_nb);
1557
1558                 msrs = msrs_alloc();
1559                 if (!msrs) {
1560                         printk(KERN_ERR "%s: Error allocating msrs!\n", __func__);
1561                         return -ENOMEM;
1562                 }
1563
1564                 rdmsr_on_cpus(cpu_online_mask, MSR_K7_HWCR, msrs);
1565
1566                 for_each_cpu(cpu, cpu_online_mask) {
1567                         struct msr *reg = per_cpu_ptr(msrs, cpu);
1568                         cpb_enabled |= !(!!(reg->l & BIT(25)));
1569                 }
1570
1571                 printk(KERN_INFO PFX "Core Performance Boosting: %s.\n",
1572                         (cpb_enabled ? "on" : "off"));
1573         }
1574
1575         return cpufreq_register_driver(&cpufreq_amd64_driver);
1576 }
1577
1578 /* driver entry point for term */
1579 static void __exit powernowk8_exit(void)
1580 {
1581         dprintk("exit\n");
1582
1583         if (boot_cpu_has(X86_FEATURE_CPB)) {
1584                 msrs_free(msrs);
1585                 msrs = NULL;
1586
1587                 unregister_cpu_notifier(&cpb_nb);
1588         }
1589
1590         cpufreq_unregister_driver(&cpufreq_amd64_driver);
1591 }
1592
1593 MODULE_AUTHOR("Paul Devriendt <paul.devriendt@amd.com> and "
1594                 "Mark Langsdorf <mark.langsdorf@amd.com>");
1595 MODULE_DESCRIPTION("AMD Athlon 64 and Opteron processor frequency driver.");
1596 MODULE_LICENSE("GPL");
1597
1598 late_initcall(powernowk8_init);
1599 module_exit(powernowk8_exit);