xref: /openbmc/linux/drivers/cpufreq/powernow-k8.c (revision 9cdb81c7)
1 /*
2  *   (c) 2003-2012 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  *  Maintainer:
8  *  Andreas Herrmann <andreas.herrmann3@amd.com>
9  *
10  *  Based on the powernow-k7.c module written by Dave Jones.
11  *  (C) 2003 Dave Jones on behalf of SuSE Labs
12  *  (C) 2004 Dominik Brodowski <linux@brodo.de>
13  *  (C) 2004 Pavel Machek <pavel@ucw.cz>
14  *  Licensed under the terms of the GNU GPL License version 2.
15  *  Based upon datasheets & sample CPUs kindly provided by AMD.
16  *
17  *  Valuable input gratefully received from Dave Jones, Pavel Machek,
18  *  Dominik Brodowski, Jacob Shin, and others.
19  *  Originally developed by Paul Devriendt.
20  *
21  *  Processor information obtained from Chapter 9 (Power and Thermal
22  *  Management) of the "BIOS and Kernel Developer's Guide (BKDG) for
23  *  the AMD Athlon 64 and AMD Opteron Processors" and section "2.x
24  *  Power Management" in BKDGs for newer AMD CPU families.
25  *
26  *  Tables for specific CPUs can be inferred from AMD's processor
27  *  power and thermal data sheets, (e.g. 30417.pdf, 30430.pdf, 43375.pdf)
28  */
29 
30 #include <linux/kernel.h>
31 #include <linux/smp.h>
32 #include <linux/module.h>
33 #include <linux/init.h>
34 #include <linux/cpufreq.h>
35 #include <linux/slab.h>
36 #include <linux/string.h>
37 #include <linux/cpumask.h>
38 #include <linux/sched.h>	/* for current / set_cpus_allowed() */
39 #include <linux/io.h>
40 #include <linux/delay.h>
41 
42 #include <asm/msr.h>
43 #include <asm/cpu_device_id.h>
44 
45 #include <linux/acpi.h>
46 #include <linux/mutex.h>
47 #include <acpi/processor.h>
48 
49 #define PFX "powernow-k8: "
50 #define VERSION "version 2.20.00"
51 #include "powernow-k8.h"
52 #include "mperf.h"
53 
54 /* serialize freq changes  */
55 static DEFINE_MUTEX(fidvid_mutex);
56 
57 static DEFINE_PER_CPU(struct powernow_k8_data *, powernow_data);
58 
59 static int cpu_family = CPU_OPTERON;
60 
61 /* array to map SW pstate number to acpi state */
62 static u32 ps_to_as[8];
63 
64 /* core performance boost */
65 static bool cpb_capable, cpb_enabled;
66 static struct msr __percpu *msrs;
67 
68 static struct cpufreq_driver cpufreq_amd64_driver;
69 
70 #ifndef CONFIG_SMP
71 static inline const struct cpumask *cpu_core_mask(int cpu)
72 {
73 	return cpumask_of(0);
74 }
75 #endif
76 
77 /* Return a frequency in MHz, given an input fid */
78 static u32 find_freq_from_fid(u32 fid)
79 {
80 	return 800 + (fid * 100);
81 }
82 
83 /* Return a frequency in KHz, given an input fid */
84 static u32 find_khz_freq_from_fid(u32 fid)
85 {
86 	return 1000 * find_freq_from_fid(fid);
87 }
88 
89 static u32 find_khz_freq_from_pstate(struct cpufreq_frequency_table *data,
90 				     u32 pstate)
91 {
92 	return data[ps_to_as[pstate]].frequency;
93 }
94 
95 /* Return the vco fid for an input fid
96  *
97  * Each "low" fid has corresponding "high" fid, and you can get to "low" fids
98  * only from corresponding high fids. This returns "high" fid corresponding to
99  * "low" one.
100  */
101 static u32 convert_fid_to_vco_fid(u32 fid)
102 {
103 	if (fid < HI_FID_TABLE_BOTTOM)
104 		return 8 + (2 * fid);
105 	else
106 		return fid;
107 }
108 
109 /*
110  * Return 1 if the pending bit is set. Unless we just instructed the processor
111  * to transition to a new state, seeing this bit set is really bad news.
112  */
113 static int pending_bit_stuck(void)
114 {
115 	u32 lo, hi;
116 
117 	if (cpu_family == CPU_HW_PSTATE)
118 		return 0;
119 
120 	rdmsr(MSR_FIDVID_STATUS, lo, hi);
121 	return lo & MSR_S_LO_CHANGE_PENDING ? 1 : 0;
122 }
123 
124 /*
125  * Update the global current fid / vid values from the status msr.
126  * Returns 1 on error.
127  */
128 static int query_current_values_with_pending_wait(struct powernow_k8_data *data)
129 {
130 	u32 lo, hi;
131 	u32 i = 0;
132 
133 	if (cpu_family == CPU_HW_PSTATE) {
134 		rdmsr(MSR_PSTATE_STATUS, lo, hi);
135 		i = lo & HW_PSTATE_MASK;
136 		data->currpstate = i;
137 
138 		/*
139 		 * a workaround for family 11h erratum 311 might cause
140 		 * an "out-of-range Pstate if the core is in Pstate-0
141 		 */
142 		if ((boot_cpu_data.x86 == 0x11) && (i >= data->numps))
143 			data->currpstate = HW_PSTATE_0;
144 
145 		return 0;
146 	}
147 	do {
148 		if (i++ > 10000) {
149 			pr_debug("detected change pending stuck\n");
150 			return 1;
151 		}
152 		rdmsr(MSR_FIDVID_STATUS, lo, hi);
153 	} while (lo & MSR_S_LO_CHANGE_PENDING);
154 
155 	data->currvid = hi & MSR_S_HI_CURRENT_VID;
156 	data->currfid = lo & MSR_S_LO_CURRENT_FID;
157 
158 	return 0;
159 }
160 
161 /* the isochronous relief time */
162 static void count_off_irt(struct powernow_k8_data *data)
163 {
164 	udelay((1 << data->irt) * 10);
165 	return;
166 }
167 
168 /* the voltage stabilization time */
169 static void count_off_vst(struct powernow_k8_data *data)
170 {
171 	udelay(data->vstable * VST_UNITS_20US);
172 	return;
173 }
174 
175 /* need to init the control msr to a safe value (for each cpu) */
176 static void fidvid_msr_init(void)
177 {
178 	u32 lo, hi;
179 	u8 fid, vid;
180 
181 	rdmsr(MSR_FIDVID_STATUS, lo, hi);
182 	vid = hi & MSR_S_HI_CURRENT_VID;
183 	fid = lo & MSR_S_LO_CURRENT_FID;
184 	lo = fid | (vid << MSR_C_LO_VID_SHIFT);
185 	hi = MSR_C_HI_STP_GNT_BENIGN;
186 	pr_debug("cpu%d, init lo 0x%x, hi 0x%x\n", smp_processor_id(), lo, hi);
187 	wrmsr(MSR_FIDVID_CTL, lo, hi);
188 }
189 
190 /* write the new fid value along with the other control fields to the msr */
191 static int write_new_fid(struct powernow_k8_data *data, u32 fid)
192 {
193 	u32 lo;
194 	u32 savevid = data->currvid;
195 	u32 i = 0;
196 
197 	if ((fid & INVALID_FID_MASK) || (data->currvid & INVALID_VID_MASK)) {
198 		printk(KERN_ERR PFX "internal error - overflow on fid write\n");
199 		return 1;
200 	}
201 
202 	lo = fid;
203 	lo |= (data->currvid << MSR_C_LO_VID_SHIFT);
204 	lo |= MSR_C_LO_INIT_FID_VID;
205 
206 	pr_debug("writing fid 0x%x, lo 0x%x, hi 0x%x\n",
207 		fid, lo, data->plllock * PLL_LOCK_CONVERSION);
208 
209 	do {
210 		wrmsr(MSR_FIDVID_CTL, lo, data->plllock * PLL_LOCK_CONVERSION);
211 		if (i++ > 100) {
212 			printk(KERN_ERR PFX
213 				"Hardware error - pending bit very stuck - "
214 				"no further pstate changes possible\n");
215 			return 1;
216 		}
217 	} while (query_current_values_with_pending_wait(data));
218 
219 	count_off_irt(data);
220 
221 	if (savevid != data->currvid) {
222 		printk(KERN_ERR PFX
223 			"vid change on fid trans, old 0x%x, new 0x%x\n",
224 			savevid, data->currvid);
225 		return 1;
226 	}
227 
228 	if (fid != data->currfid) {
229 		printk(KERN_ERR PFX
230 			"fid trans failed, fid 0x%x, curr 0x%x\n", fid,
231 			data->currfid);
232 		return 1;
233 	}
234 
235 	return 0;
236 }
237 
238 /* Write a new vid to the hardware */
239 static int write_new_vid(struct powernow_k8_data *data, u32 vid)
240 {
241 	u32 lo;
242 	u32 savefid = data->currfid;
243 	int i = 0;
244 
245 	if ((data->currfid & INVALID_FID_MASK) || (vid & INVALID_VID_MASK)) {
246 		printk(KERN_ERR PFX "internal error - overflow on vid write\n");
247 		return 1;
248 	}
249 
250 	lo = data->currfid;
251 	lo |= (vid << MSR_C_LO_VID_SHIFT);
252 	lo |= MSR_C_LO_INIT_FID_VID;
253 
254 	pr_debug("writing vid 0x%x, lo 0x%x, hi 0x%x\n",
255 		vid, lo, STOP_GRANT_5NS);
256 
257 	do {
258 		wrmsr(MSR_FIDVID_CTL, lo, STOP_GRANT_5NS);
259 		if (i++ > 100) {
260 			printk(KERN_ERR PFX "internal error - pending bit "
261 					"very stuck - no further pstate "
262 					"changes possible\n");
263 			return 1;
264 		}
265 	} while (query_current_values_with_pending_wait(data));
266 
267 	if (savefid != data->currfid) {
268 		printk(KERN_ERR PFX "fid changed on vid trans, old "
269 			"0x%x new 0x%x\n",
270 		       savefid, data->currfid);
271 		return 1;
272 	}
273 
274 	if (vid != data->currvid) {
275 		printk(KERN_ERR PFX "vid trans failed, vid 0x%x, "
276 				"curr 0x%x\n",
277 				vid, data->currvid);
278 		return 1;
279 	}
280 
281 	return 0;
282 }
283 
284 /*
285  * Reduce the vid by the max of step or reqvid.
286  * Decreasing vid codes represent increasing voltages:
287  * vid of 0 is 1.550V, vid of 0x1e is 0.800V, vid of VID_OFF is off.
288  */
289 static int decrease_vid_code_by_step(struct powernow_k8_data *data,
290 		u32 reqvid, u32 step)
291 {
292 	if ((data->currvid - reqvid) > step)
293 		reqvid = data->currvid - step;
294 
295 	if (write_new_vid(data, reqvid))
296 		return 1;
297 
298 	count_off_vst(data);
299 
300 	return 0;
301 }
302 
303 /* Change hardware pstate by single MSR write */
304 static int transition_pstate(struct powernow_k8_data *data, u32 pstate)
305 {
306 	wrmsr(MSR_PSTATE_CTRL, pstate, 0);
307 	data->currpstate = pstate;
308 	return 0;
309 }
310 
311 /* Change Opteron/Athlon64 fid and vid, by the 3 phases. */
312 static int transition_fid_vid(struct powernow_k8_data *data,
313 		u32 reqfid, u32 reqvid)
314 {
315 	if (core_voltage_pre_transition(data, reqvid, reqfid))
316 		return 1;
317 
318 	if (core_frequency_transition(data, reqfid))
319 		return 1;
320 
321 	if (core_voltage_post_transition(data, reqvid))
322 		return 1;
323 
324 	if (query_current_values_with_pending_wait(data))
325 		return 1;
326 
327 	if ((reqfid != data->currfid) || (reqvid != data->currvid)) {
328 		printk(KERN_ERR PFX "failed (cpu%d): req 0x%x 0x%x, "
329 				"curr 0x%x 0x%x\n",
330 				smp_processor_id(),
331 				reqfid, reqvid, data->currfid, data->currvid);
332 		return 1;
333 	}
334 
335 	pr_debug("transitioned (cpu%d): new fid 0x%x, vid 0x%x\n",
336 		smp_processor_id(), data->currfid, data->currvid);
337 
338 	return 0;
339 }
340 
341 /* Phase 1 - core voltage transition ... setup voltage */
342 static int core_voltage_pre_transition(struct powernow_k8_data *data,
343 		u32 reqvid, u32 reqfid)
344 {
345 	u32 rvosteps = data->rvo;
346 	u32 savefid = data->currfid;
347 	u32 maxvid, lo, rvomult = 1;
348 
349 	pr_debug("ph1 (cpu%d): start, currfid 0x%x, currvid 0x%x, "
350 		"reqvid 0x%x, rvo 0x%x\n",
351 		smp_processor_id(),
352 		data->currfid, data->currvid, reqvid, data->rvo);
353 
354 	if ((savefid < LO_FID_TABLE_TOP) && (reqfid < LO_FID_TABLE_TOP))
355 		rvomult = 2;
356 	rvosteps *= rvomult;
357 	rdmsr(MSR_FIDVID_STATUS, lo, maxvid);
358 	maxvid = 0x1f & (maxvid >> 16);
359 	pr_debug("ph1 maxvid=0x%x\n", maxvid);
360 	if (reqvid < maxvid) /* lower numbers are higher voltages */
361 		reqvid = maxvid;
362 
363 	while (data->currvid > reqvid) {
364 		pr_debug("ph1: curr 0x%x, req vid 0x%x\n",
365 			data->currvid, reqvid);
366 		if (decrease_vid_code_by_step(data, reqvid, data->vidmvs))
367 			return 1;
368 	}
369 
370 	while ((rvosteps > 0) &&
371 			((rvomult * data->rvo + data->currvid) > reqvid)) {
372 		if (data->currvid == maxvid) {
373 			rvosteps = 0;
374 		} else {
375 			pr_debug("ph1: changing vid for rvo, req 0x%x\n",
376 				data->currvid - 1);
377 			if (decrease_vid_code_by_step(data, data->currvid-1, 1))
378 				return 1;
379 			rvosteps--;
380 		}
381 	}
382 
383 	if (query_current_values_with_pending_wait(data))
384 		return 1;
385 
386 	if (savefid != data->currfid) {
387 		printk(KERN_ERR PFX "ph1 err, currfid changed 0x%x\n",
388 				data->currfid);
389 		return 1;
390 	}
391 
392 	pr_debug("ph1 complete, currfid 0x%x, currvid 0x%x\n",
393 		data->currfid, data->currvid);
394 
395 	return 0;
396 }
397 
398 /* Phase 2 - core frequency transition */
399 static int core_frequency_transition(struct powernow_k8_data *data, u32 reqfid)
400 {
401 	u32 vcoreqfid, vcocurrfid, vcofiddiff;
402 	u32 fid_interval, savevid = data->currvid;
403 
404 	if (data->currfid == reqfid) {
405 		printk(KERN_ERR PFX "ph2 null fid transition 0x%x\n",
406 				data->currfid);
407 		return 0;
408 	}
409 
410 	pr_debug("ph2 (cpu%d): starting, currfid 0x%x, currvid 0x%x, "
411 		"reqfid 0x%x\n",
412 		smp_processor_id(),
413 		data->currfid, data->currvid, reqfid);
414 
415 	vcoreqfid = convert_fid_to_vco_fid(reqfid);
416 	vcocurrfid = convert_fid_to_vco_fid(data->currfid);
417 	vcofiddiff = vcocurrfid > vcoreqfid ? vcocurrfid - vcoreqfid
418 	    : vcoreqfid - vcocurrfid;
419 
420 	if ((reqfid <= LO_FID_TABLE_TOP) && (data->currfid <= LO_FID_TABLE_TOP))
421 		vcofiddiff = 0;
422 
423 	while (vcofiddiff > 2) {
424 		(data->currfid & 1) ? (fid_interval = 1) : (fid_interval = 2);
425 
426 		if (reqfid > data->currfid) {
427 			if (data->currfid > LO_FID_TABLE_TOP) {
428 				if (write_new_fid(data,
429 						data->currfid + fid_interval))
430 					return 1;
431 			} else {
432 				if (write_new_fid
433 				    (data,
434 				     2 + convert_fid_to_vco_fid(data->currfid)))
435 					return 1;
436 			}
437 		} else {
438 			if (write_new_fid(data, data->currfid - fid_interval))
439 				return 1;
440 		}
441 
442 		vcocurrfid = convert_fid_to_vco_fid(data->currfid);
443 		vcofiddiff = vcocurrfid > vcoreqfid ? vcocurrfid - vcoreqfid
444 		    : vcoreqfid - vcocurrfid;
445 	}
446 
447 	if (write_new_fid(data, reqfid))
448 		return 1;
449 
450 	if (query_current_values_with_pending_wait(data))
451 		return 1;
452 
453 	if (data->currfid != reqfid) {
454 		printk(KERN_ERR PFX
455 			"ph2: mismatch, failed fid transition, "
456 			"curr 0x%x, req 0x%x\n",
457 			data->currfid, reqfid);
458 		return 1;
459 	}
460 
461 	if (savevid != data->currvid) {
462 		printk(KERN_ERR PFX "ph2: vid changed, save 0x%x, curr 0x%x\n",
463 			savevid, data->currvid);
464 		return 1;
465 	}
466 
467 	pr_debug("ph2 complete, currfid 0x%x, currvid 0x%x\n",
468 		data->currfid, data->currvid);
469 
470 	return 0;
471 }
472 
473 /* Phase 3 - core voltage transition flow ... jump to the final vid. */
474 static int core_voltage_post_transition(struct powernow_k8_data *data,
475 		u32 reqvid)
476 {
477 	u32 savefid = data->currfid;
478 	u32 savereqvid = reqvid;
479 
480 	pr_debug("ph3 (cpu%d): starting, currfid 0x%x, currvid 0x%x\n",
481 		smp_processor_id(),
482 		data->currfid, data->currvid);
483 
484 	if (reqvid != data->currvid) {
485 		if (write_new_vid(data, reqvid))
486 			return 1;
487 
488 		if (savefid != data->currfid) {
489 			printk(KERN_ERR PFX
490 			       "ph3: bad fid change, save 0x%x, curr 0x%x\n",
491 			       savefid, data->currfid);
492 			return 1;
493 		}
494 
495 		if (data->currvid != reqvid) {
496 			printk(KERN_ERR PFX
497 			       "ph3: failed vid transition\n, "
498 			       "req 0x%x, curr 0x%x",
499 			       reqvid, data->currvid);
500 			return 1;
501 		}
502 	}
503 
504 	if (query_current_values_with_pending_wait(data))
505 		return 1;
506 
507 	if (savereqvid != data->currvid) {
508 		pr_debug("ph3 failed, currvid 0x%x\n", data->currvid);
509 		return 1;
510 	}
511 
512 	if (savefid != data->currfid) {
513 		pr_debug("ph3 failed, currfid changed 0x%x\n",
514 			data->currfid);
515 		return 1;
516 	}
517 
518 	pr_debug("ph3 complete, currfid 0x%x, currvid 0x%x\n",
519 		data->currfid, data->currvid);
520 
521 	return 0;
522 }
523 
524 static const struct x86_cpu_id powernow_k8_ids[] = {
525 	/* IO based frequency switching */
526 	{ X86_VENDOR_AMD, 0xf },
527 	/* MSR based frequency switching supported */
528 	X86_FEATURE_MATCH(X86_FEATURE_HW_PSTATE),
529 	{}
530 };
531 MODULE_DEVICE_TABLE(x86cpu, powernow_k8_ids);
532 
533 static void check_supported_cpu(void *_rc)
534 {
535 	u32 eax, ebx, ecx, edx;
536 	int *rc = _rc;
537 
538 	*rc = -ENODEV;
539 
540 	eax = cpuid_eax(CPUID_PROCESSOR_SIGNATURE);
541 
542 	if ((eax & CPUID_XFAM) == CPUID_XFAM_K8) {
543 		if (((eax & CPUID_USE_XFAM_XMOD) != CPUID_USE_XFAM_XMOD) ||
544 		    ((eax & CPUID_XMOD) > CPUID_XMOD_REV_MASK)) {
545 			printk(KERN_INFO PFX
546 				"Processor cpuid %x not supported\n", eax);
547 			return;
548 		}
549 
550 		eax = cpuid_eax(CPUID_GET_MAX_CAPABILITIES);
551 		if (eax < CPUID_FREQ_VOLT_CAPABILITIES) {
552 			printk(KERN_INFO PFX
553 			       "No frequency change capabilities detected\n");
554 			return;
555 		}
556 
557 		cpuid(CPUID_FREQ_VOLT_CAPABILITIES, &eax, &ebx, &ecx, &edx);
558 		if ((edx & P_STATE_TRANSITION_CAPABLE)
559 			!= P_STATE_TRANSITION_CAPABLE) {
560 			printk(KERN_INFO PFX
561 				"Power state transitions not supported\n");
562 			return;
563 		}
564 	} else { /* must be a HW Pstate capable processor */
565 		cpuid(CPUID_FREQ_VOLT_CAPABILITIES, &eax, &ebx, &ecx, &edx);
566 		if ((edx & USE_HW_PSTATE) == USE_HW_PSTATE)
567 			cpu_family = CPU_HW_PSTATE;
568 		else
569 			return;
570 	}
571 
572 	*rc = 0;
573 }
574 
575 static int check_pst_table(struct powernow_k8_data *data, struct pst_s *pst,
576 		u8 maxvid)
577 {
578 	unsigned int j;
579 	u8 lastfid = 0xff;
580 
581 	for (j = 0; j < data->numps; j++) {
582 		if (pst[j].vid > LEAST_VID) {
583 			printk(KERN_ERR FW_BUG PFX "vid %d invalid : 0x%x\n",
584 			       j, pst[j].vid);
585 			return -EINVAL;
586 		}
587 		if (pst[j].vid < data->rvo) {
588 			/* vid + rvo >= 0 */
589 			printk(KERN_ERR FW_BUG PFX "0 vid exceeded with pstate"
590 			       " %d\n", j);
591 			return -ENODEV;
592 		}
593 		if (pst[j].vid < maxvid + data->rvo) {
594 			/* vid + rvo >= maxvid */
595 			printk(KERN_ERR FW_BUG PFX "maxvid exceeded with pstate"
596 			       " %d\n", j);
597 			return -ENODEV;
598 		}
599 		if (pst[j].fid > MAX_FID) {
600 			printk(KERN_ERR FW_BUG PFX "maxfid exceeded with pstate"
601 			       " %d\n", j);
602 			return -ENODEV;
603 		}
604 		if (j && (pst[j].fid < HI_FID_TABLE_BOTTOM)) {
605 			/* Only first fid is allowed to be in "low" range */
606 			printk(KERN_ERR FW_BUG PFX "two low fids - %d : "
607 			       "0x%x\n", j, pst[j].fid);
608 			return -EINVAL;
609 		}
610 		if (pst[j].fid < lastfid)
611 			lastfid = pst[j].fid;
612 	}
613 	if (lastfid & 1) {
614 		printk(KERN_ERR FW_BUG PFX "lastfid invalid\n");
615 		return -EINVAL;
616 	}
617 	if (lastfid > LO_FID_TABLE_TOP)
618 		printk(KERN_INFO FW_BUG PFX
619 			"first fid not from lo freq table\n");
620 
621 	return 0;
622 }
623 
624 static void invalidate_entry(struct cpufreq_frequency_table *powernow_table,
625 		unsigned int entry)
626 {
627 	powernow_table[entry].frequency = CPUFREQ_ENTRY_INVALID;
628 }
629 
630 static void print_basics(struct powernow_k8_data *data)
631 {
632 	int j;
633 	for (j = 0; j < data->numps; j++) {
634 		if (data->powernow_table[j].frequency !=
635 				CPUFREQ_ENTRY_INVALID) {
636 			if (cpu_family == CPU_HW_PSTATE) {
637 				printk(KERN_INFO PFX
638 					"   %d : pstate %d (%d MHz)\n", j,
639 					data->powernow_table[j].index,
640 					data->powernow_table[j].frequency/1000);
641 			} else {
642 				printk(KERN_INFO PFX
643 					"fid 0x%x (%d MHz), vid 0x%x\n",
644 					data->powernow_table[j].index & 0xff,
645 					data->powernow_table[j].frequency/1000,
646 					data->powernow_table[j].index >> 8);
647 			}
648 		}
649 	}
650 	if (data->batps)
651 		printk(KERN_INFO PFX "Only %d pstates on battery\n",
652 				data->batps);
653 }
654 
655 static u32 freq_from_fid_did(u32 fid, u32 did)
656 {
657 	u32 mhz = 0;
658 
659 	if (boot_cpu_data.x86 == 0x10)
660 		mhz = (100 * (fid + 0x10)) >> did;
661 	else if (boot_cpu_data.x86 == 0x11)
662 		mhz = (100 * (fid + 8)) >> did;
663 	else
664 		BUG();
665 
666 	return mhz * 1000;
667 }
668 
669 static int fill_powernow_table(struct powernow_k8_data *data,
670 		struct pst_s *pst, u8 maxvid)
671 {
672 	struct cpufreq_frequency_table *powernow_table;
673 	unsigned int j;
674 
675 	if (data->batps) {
676 		/* use ACPI support to get full speed on mains power */
677 		printk(KERN_WARNING PFX
678 			"Only %d pstates usable (use ACPI driver for full "
679 			"range\n", data->batps);
680 		data->numps = data->batps;
681 	}
682 
683 	for (j = 1; j < data->numps; j++) {
684 		if (pst[j-1].fid >= pst[j].fid) {
685 			printk(KERN_ERR PFX "PST out of sequence\n");
686 			return -EINVAL;
687 		}
688 	}
689 
690 	if (data->numps < 2) {
691 		printk(KERN_ERR PFX "no p states to transition\n");
692 		return -ENODEV;
693 	}
694 
695 	if (check_pst_table(data, pst, maxvid))
696 		return -EINVAL;
697 
698 	powernow_table = kmalloc((sizeof(struct cpufreq_frequency_table)
699 		* (data->numps + 1)), GFP_KERNEL);
700 	if (!powernow_table) {
701 		printk(KERN_ERR PFX "powernow_table memory alloc failure\n");
702 		return -ENOMEM;
703 	}
704 
705 	for (j = 0; j < data->numps; j++) {
706 		int freq;
707 		powernow_table[j].index = pst[j].fid; /* lower 8 bits */
708 		powernow_table[j].index |= (pst[j].vid << 8); /* upper 8 bits */
709 		freq = find_khz_freq_from_fid(pst[j].fid);
710 		powernow_table[j].frequency = freq;
711 	}
712 	powernow_table[data->numps].frequency = CPUFREQ_TABLE_END;
713 	powernow_table[data->numps].index = 0;
714 
715 	if (query_current_values_with_pending_wait(data)) {
716 		kfree(powernow_table);
717 		return -EIO;
718 	}
719 
720 	pr_debug("cfid 0x%x, cvid 0x%x\n", data->currfid, data->currvid);
721 	data->powernow_table = powernow_table;
722 	if (cpumask_first(cpu_core_mask(data->cpu)) == data->cpu)
723 		print_basics(data);
724 
725 	for (j = 0; j < data->numps; j++)
726 		if ((pst[j].fid == data->currfid) &&
727 		    (pst[j].vid == data->currvid))
728 			return 0;
729 
730 	pr_debug("currfid/vid do not match PST, ignoring\n");
731 	return 0;
732 }
733 
734 /* Find and validate the PSB/PST table in BIOS. */
735 static int find_psb_table(struct powernow_k8_data *data)
736 {
737 	struct psb_s *psb;
738 	unsigned int i;
739 	u32 mvs;
740 	u8 maxvid;
741 	u32 cpst = 0;
742 	u32 thiscpuid;
743 
744 	for (i = 0xc0000; i < 0xffff0; i += 0x10) {
745 		/* Scan BIOS looking for the signature. */
746 		/* It can not be at ffff0 - it is too big. */
747 
748 		psb = phys_to_virt(i);
749 		if (memcmp(psb, PSB_ID_STRING, PSB_ID_STRING_LEN) != 0)
750 			continue;
751 
752 		pr_debug("found PSB header at 0x%p\n", psb);
753 
754 		pr_debug("table vers: 0x%x\n", psb->tableversion);
755 		if (psb->tableversion != PSB_VERSION_1_4) {
756 			printk(KERN_ERR FW_BUG PFX "PSB table is not v1.4\n");
757 			return -ENODEV;
758 		}
759 
760 		pr_debug("flags: 0x%x\n", psb->flags1);
761 		if (psb->flags1) {
762 			printk(KERN_ERR FW_BUG PFX "unknown flags\n");
763 			return -ENODEV;
764 		}
765 
766 		data->vstable = psb->vstable;
767 		pr_debug("voltage stabilization time: %d(*20us)\n",
768 				data->vstable);
769 
770 		pr_debug("flags2: 0x%x\n", psb->flags2);
771 		data->rvo = psb->flags2 & 3;
772 		data->irt = ((psb->flags2) >> 2) & 3;
773 		mvs = ((psb->flags2) >> 4) & 3;
774 		data->vidmvs = 1 << mvs;
775 		data->batps = ((psb->flags2) >> 6) & 3;
776 
777 		pr_debug("ramp voltage offset: %d\n", data->rvo);
778 		pr_debug("isochronous relief time: %d\n", data->irt);
779 		pr_debug("maximum voltage step: %d - 0x%x\n", mvs, data->vidmvs);
780 
781 		pr_debug("numpst: 0x%x\n", psb->num_tables);
782 		cpst = psb->num_tables;
783 		if ((psb->cpuid == 0x00000fc0) ||
784 		    (psb->cpuid == 0x00000fe0)) {
785 			thiscpuid = cpuid_eax(CPUID_PROCESSOR_SIGNATURE);
786 			if ((thiscpuid == 0x00000fc0) ||
787 			    (thiscpuid == 0x00000fe0))
788 				cpst = 1;
789 		}
790 		if (cpst != 1) {
791 			printk(KERN_ERR FW_BUG PFX "numpst must be 1\n");
792 			return -ENODEV;
793 		}
794 
795 		data->plllock = psb->plllocktime;
796 		pr_debug("plllocktime: 0x%x (units 1us)\n", psb->plllocktime);
797 		pr_debug("maxfid: 0x%x\n", psb->maxfid);
798 		pr_debug("maxvid: 0x%x\n", psb->maxvid);
799 		maxvid = psb->maxvid;
800 
801 		data->numps = psb->numps;
802 		pr_debug("numpstates: 0x%x\n", data->numps);
803 		return fill_powernow_table(data,
804 				(struct pst_s *)(psb+1), maxvid);
805 	}
806 	/*
807 	 * If you see this message, complain to BIOS manufacturer. If
808 	 * he tells you "we do not support Linux" or some similar
809 	 * nonsense, remember that Windows 2000 uses the same legacy
810 	 * mechanism that the old Linux PSB driver uses. Tell them it
811 	 * is broken with Windows 2000.
812 	 *
813 	 * The reference to the AMD documentation is chapter 9 in the
814 	 * BIOS and Kernel Developer's Guide, which is available on
815 	 * www.amd.com
816 	 */
817 	printk(KERN_ERR FW_BUG PFX "No PSB or ACPI _PSS objects\n");
818 	printk(KERN_ERR PFX "Make sure that your BIOS is up to date"
819 		" and Cool'N'Quiet support is enabled in BIOS setup\n");
820 	return -ENODEV;
821 }
822 
823 static void powernow_k8_acpi_pst_values(struct powernow_k8_data *data,
824 		unsigned int index)
825 {
826 	u64 control;
827 
828 	if (!data->acpi_data.state_count || (cpu_family == CPU_HW_PSTATE))
829 		return;
830 
831 	control = data->acpi_data.states[index].control;
832 	data->irt = (control >> IRT_SHIFT) & IRT_MASK;
833 	data->rvo = (control >> RVO_SHIFT) & RVO_MASK;
834 	data->exttype = (control >> EXT_TYPE_SHIFT) & EXT_TYPE_MASK;
835 	data->plllock = (control >> PLL_L_SHIFT) & PLL_L_MASK;
836 	data->vidmvs = 1 << ((control >> MVS_SHIFT) & MVS_MASK);
837 	data->vstable = (control >> VST_SHIFT) & VST_MASK;
838 }
839 
840 static int powernow_k8_cpu_init_acpi(struct powernow_k8_data *data)
841 {
842 	struct cpufreq_frequency_table *powernow_table;
843 	int ret_val = -ENODEV;
844 	u64 control, status;
845 
846 	if (acpi_processor_register_performance(&data->acpi_data, data->cpu)) {
847 		pr_debug("register performance failed: bad ACPI data\n");
848 		return -EIO;
849 	}
850 
851 	/* verify the data contained in the ACPI structures */
852 	if (data->acpi_data.state_count <= 1) {
853 		pr_debug("No ACPI P-States\n");
854 		goto err_out;
855 	}
856 
857 	control = data->acpi_data.control_register.space_id;
858 	status = data->acpi_data.status_register.space_id;
859 
860 	if ((control != ACPI_ADR_SPACE_FIXED_HARDWARE) ||
861 	    (status != ACPI_ADR_SPACE_FIXED_HARDWARE)) {
862 		pr_debug("Invalid control/status registers (%llx - %llx)\n",
863 			control, status);
864 		goto err_out;
865 	}
866 
867 	/* fill in data->powernow_table */
868 	powernow_table = kmalloc((sizeof(struct cpufreq_frequency_table)
869 		* (data->acpi_data.state_count + 1)), GFP_KERNEL);
870 	if (!powernow_table) {
871 		pr_debug("powernow_table memory alloc failure\n");
872 		goto err_out;
873 	}
874 
875 	/* fill in data */
876 	data->numps = data->acpi_data.state_count;
877 	powernow_k8_acpi_pst_values(data, 0);
878 
879 	if (cpu_family == CPU_HW_PSTATE)
880 		ret_val = fill_powernow_table_pstate(data, powernow_table);
881 	else
882 		ret_val = fill_powernow_table_fidvid(data, powernow_table);
883 	if (ret_val)
884 		goto err_out_mem;
885 
886 	powernow_table[data->acpi_data.state_count].frequency =
887 		CPUFREQ_TABLE_END;
888 	powernow_table[data->acpi_data.state_count].index = 0;
889 	data->powernow_table = powernow_table;
890 
891 	if (cpumask_first(cpu_core_mask(data->cpu)) == data->cpu)
892 		print_basics(data);
893 
894 	/* notify BIOS that we exist */
895 	acpi_processor_notify_smm(THIS_MODULE);
896 
897 	if (!zalloc_cpumask_var(&data->acpi_data.shared_cpu_map, GFP_KERNEL)) {
898 		printk(KERN_ERR PFX
899 				"unable to alloc powernow_k8_data cpumask\n");
900 		ret_val = -ENOMEM;
901 		goto err_out_mem;
902 	}
903 
904 	return 0;
905 
906 err_out_mem:
907 	kfree(powernow_table);
908 
909 err_out:
910 	acpi_processor_unregister_performance(&data->acpi_data, data->cpu);
911 
912 	/* data->acpi_data.state_count informs us at ->exit()
913 	 * whether ACPI was used */
914 	data->acpi_data.state_count = 0;
915 
916 	return ret_val;
917 }
918 
919 static int fill_powernow_table_pstate(struct powernow_k8_data *data,
920 		struct cpufreq_frequency_table *powernow_table)
921 {
922 	int i;
923 	u32 hi = 0, lo = 0;
924 	rdmsr(MSR_PSTATE_CUR_LIMIT, lo, hi);
925 	data->max_hw_pstate = (lo & HW_PSTATE_MAX_MASK) >> HW_PSTATE_MAX_SHIFT;
926 
927 	for (i = 0; i < data->acpi_data.state_count; i++) {
928 		u32 index;
929 
930 		index = data->acpi_data.states[i].control & HW_PSTATE_MASK;
931 		if (index > data->max_hw_pstate) {
932 			printk(KERN_ERR PFX "invalid pstate %d - "
933 					"bad value %d.\n", i, index);
934 			printk(KERN_ERR PFX "Please report to BIOS "
935 					"manufacturer\n");
936 			invalidate_entry(powernow_table, i);
937 			continue;
938 		}
939 
940 		ps_to_as[index] = i;
941 
942 		/* Frequency may be rounded for these */
943 		if ((boot_cpu_data.x86 == 0x10 && boot_cpu_data.x86_model < 10)
944 				 || boot_cpu_data.x86 == 0x11) {
945 
946 			rdmsr(MSR_PSTATE_DEF_BASE + index, lo, hi);
947 			if (!(hi & HW_PSTATE_VALID_MASK)) {
948 				pr_debug("invalid pstate %d, ignoring\n", index);
949 				invalidate_entry(powernow_table, i);
950 				continue;
951 			}
952 
953 			powernow_table[i].frequency =
954 				freq_from_fid_did(lo & 0x3f, (lo >> 6) & 7);
955 		} else
956 			powernow_table[i].frequency =
957 				data->acpi_data.states[i].core_frequency * 1000;
958 
959 		powernow_table[i].index = index;
960 	}
961 	return 0;
962 }
963 
964 static int fill_powernow_table_fidvid(struct powernow_k8_data *data,
965 		struct cpufreq_frequency_table *powernow_table)
966 {
967 	int i;
968 
969 	for (i = 0; i < data->acpi_data.state_count; i++) {
970 		u32 fid;
971 		u32 vid;
972 		u32 freq, index;
973 		u64 status, control;
974 
975 		if (data->exttype) {
976 			status =  data->acpi_data.states[i].status;
977 			fid = status & EXT_FID_MASK;
978 			vid = (status >> VID_SHIFT) & EXT_VID_MASK;
979 		} else {
980 			control =  data->acpi_data.states[i].control;
981 			fid = control & FID_MASK;
982 			vid = (control >> VID_SHIFT) & VID_MASK;
983 		}
984 
985 		pr_debug("   %d : fid 0x%x, vid 0x%x\n", i, fid, vid);
986 
987 		index = fid | (vid<<8);
988 		powernow_table[i].index = index;
989 
990 		freq = find_khz_freq_from_fid(fid);
991 		powernow_table[i].frequency = freq;
992 
993 		/* verify frequency is OK */
994 		if ((freq > (MAX_FREQ * 1000)) || (freq < (MIN_FREQ * 1000))) {
995 			pr_debug("invalid freq %u kHz, ignoring\n", freq);
996 			invalidate_entry(powernow_table, i);
997 			continue;
998 		}
999 
1000 		/* verify voltage is OK -
1001 		 * BIOSs are using "off" to indicate invalid */
1002 		if (vid == VID_OFF) {
1003 			pr_debug("invalid vid %u, ignoring\n", vid);
1004 			invalidate_entry(powernow_table, i);
1005 			continue;
1006 		}
1007 
1008 		if (freq != (data->acpi_data.states[i].core_frequency * 1000)) {
1009 			printk(KERN_INFO PFX "invalid freq entries "
1010 				"%u kHz vs. %u kHz\n", freq,
1011 				(unsigned int)
1012 				(data->acpi_data.states[i].core_frequency
1013 				 * 1000));
1014 			invalidate_entry(powernow_table, i);
1015 			continue;
1016 		}
1017 	}
1018 	return 0;
1019 }
1020 
1021 static void powernow_k8_cpu_exit_acpi(struct powernow_k8_data *data)
1022 {
1023 	if (data->acpi_data.state_count)
1024 		acpi_processor_unregister_performance(&data->acpi_data,
1025 				data->cpu);
1026 	free_cpumask_var(data->acpi_data.shared_cpu_map);
1027 }
1028 
1029 static int get_transition_latency(struct powernow_k8_data *data)
1030 {
1031 	int max_latency = 0;
1032 	int i;
1033 	for (i = 0; i < data->acpi_data.state_count; i++) {
1034 		int cur_latency = data->acpi_data.states[i].transition_latency
1035 			+ data->acpi_data.states[i].bus_master_latency;
1036 		if (cur_latency > max_latency)
1037 			max_latency = cur_latency;
1038 	}
1039 	if (max_latency == 0) {
1040 		/*
1041 		 * Fam 11h and later may return 0 as transition latency. This
1042 		 * is intended and means "very fast". While cpufreq core and
1043 		 * governors currently can handle that gracefully, better set it
1044 		 * to 1 to avoid problems in the future.
1045 		 */
1046 		if (boot_cpu_data.x86 < 0x11)
1047 			printk(KERN_ERR FW_WARN PFX "Invalid zero transition "
1048 				"latency\n");
1049 		max_latency = 1;
1050 	}
1051 	/* value in usecs, needs to be in nanoseconds */
1052 	return 1000 * max_latency;
1053 }
1054 
1055 /* Take a frequency, and issue the fid/vid transition command */
1056 static int transition_frequency_fidvid(struct powernow_k8_data *data,
1057 		unsigned int index)
1058 {
1059 	u32 fid = 0;
1060 	u32 vid = 0;
1061 	int res, i;
1062 	struct cpufreq_freqs freqs;
1063 
1064 	pr_debug("cpu %d transition to index %u\n", smp_processor_id(), index);
1065 
1066 	/* fid/vid correctness check for k8 */
1067 	/* fid are the lower 8 bits of the index we stored into
1068 	 * the cpufreq frequency table in find_psb_table, vid
1069 	 * are the upper 8 bits.
1070 	 */
1071 	fid = data->powernow_table[index].index & 0xFF;
1072 	vid = (data->powernow_table[index].index & 0xFF00) >> 8;
1073 
1074 	pr_debug("table matched fid 0x%x, giving vid 0x%x\n", fid, vid);
1075 
1076 	if (query_current_values_with_pending_wait(data))
1077 		return 1;
1078 
1079 	if ((data->currvid == vid) && (data->currfid == fid)) {
1080 		pr_debug("target matches current values (fid 0x%x, vid 0x%x)\n",
1081 			fid, vid);
1082 		return 0;
1083 	}
1084 
1085 	pr_debug("cpu %d, changing to fid 0x%x, vid 0x%x\n",
1086 		smp_processor_id(), fid, vid);
1087 	freqs.old = find_khz_freq_from_fid(data->currfid);
1088 	freqs.new = find_khz_freq_from_fid(fid);
1089 
1090 	for_each_cpu(i, data->available_cores) {
1091 		freqs.cpu = i;
1092 		cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
1093 	}
1094 
1095 	res = transition_fid_vid(data, fid, vid);
1096 	if (res)
1097 		return res;
1098 
1099 	freqs.new = find_khz_freq_from_fid(data->currfid);
1100 
1101 	for_each_cpu(i, data->available_cores) {
1102 		freqs.cpu = i;
1103 		cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
1104 	}
1105 	return res;
1106 }
1107 
1108 /* Take a frequency, and issue the hardware pstate transition command */
1109 static int transition_frequency_pstate(struct powernow_k8_data *data,
1110 		unsigned int index)
1111 {
1112 	u32 pstate = 0;
1113 	int res, i;
1114 	struct cpufreq_freqs freqs;
1115 
1116 	pr_debug("cpu %d transition to index %u\n", smp_processor_id(), index);
1117 
1118 	/* get MSR index for hardware pstate transition */
1119 	pstate = index & HW_PSTATE_MASK;
1120 	if (pstate > data->max_hw_pstate)
1121 		return -EINVAL;
1122 
1123 	freqs.old = find_khz_freq_from_pstate(data->powernow_table,
1124 			data->currpstate);
1125 	freqs.new = find_khz_freq_from_pstate(data->powernow_table, pstate);
1126 
1127 	for_each_cpu(i, data->available_cores) {
1128 		freqs.cpu = i;
1129 		cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
1130 	}
1131 
1132 	res = transition_pstate(data, pstate);
1133 	freqs.new = find_khz_freq_from_pstate(data->powernow_table, pstate);
1134 
1135 	for_each_cpu(i, data->available_cores) {
1136 		freqs.cpu = i;
1137 		cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
1138 	}
1139 	return res;
1140 }
1141 
1142 /* Driver entry point to switch to the target frequency */
1143 static int powernowk8_target(struct cpufreq_policy *pol,
1144 		unsigned targfreq, unsigned relation)
1145 {
1146 	cpumask_var_t oldmask;
1147 	struct powernow_k8_data *data = per_cpu(powernow_data, pol->cpu);
1148 	u32 checkfid;
1149 	u32 checkvid;
1150 	unsigned int newstate;
1151 	int ret = -EIO;
1152 
1153 	if (!data)
1154 		return -EINVAL;
1155 
1156 	checkfid = data->currfid;
1157 	checkvid = data->currvid;
1158 
1159 	/* only run on specific CPU from here on. */
1160 	/* This is poor form: use a workqueue or smp_call_function_single */
1161 	if (!alloc_cpumask_var(&oldmask, GFP_KERNEL))
1162 		return -ENOMEM;
1163 
1164 	cpumask_copy(oldmask, tsk_cpus_allowed(current));
1165 	set_cpus_allowed_ptr(current, cpumask_of(pol->cpu));
1166 
1167 	if (smp_processor_id() != pol->cpu) {
1168 		printk(KERN_ERR PFX "limiting to cpu %u failed\n", pol->cpu);
1169 		goto err_out;
1170 	}
1171 
1172 	if (pending_bit_stuck()) {
1173 		printk(KERN_ERR PFX "failing targ, change pending bit set\n");
1174 		goto err_out;
1175 	}
1176 
1177 	pr_debug("targ: cpu %d, %d kHz, min %d, max %d, relation %d\n",
1178 		pol->cpu, targfreq, pol->min, pol->max, relation);
1179 
1180 	if (query_current_values_with_pending_wait(data))
1181 		goto err_out;
1182 
1183 	if (cpu_family != CPU_HW_PSTATE) {
1184 		pr_debug("targ: curr fid 0x%x, vid 0x%x\n",
1185 		data->currfid, data->currvid);
1186 
1187 		if ((checkvid != data->currvid) ||
1188 		    (checkfid != data->currfid)) {
1189 			printk(KERN_INFO PFX
1190 				"error - out of sync, fix 0x%x 0x%x, "
1191 				"vid 0x%x 0x%x\n",
1192 				checkfid, data->currfid,
1193 				checkvid, data->currvid);
1194 		}
1195 	}
1196 
1197 	if (cpufreq_frequency_table_target(pol, data->powernow_table,
1198 				targfreq, relation, &newstate))
1199 		goto err_out;
1200 
1201 	mutex_lock(&fidvid_mutex);
1202 
1203 	powernow_k8_acpi_pst_values(data, newstate);
1204 
1205 	if (cpu_family == CPU_HW_PSTATE)
1206 		ret = transition_frequency_pstate(data,
1207 			data->powernow_table[newstate].index);
1208 	else
1209 		ret = transition_frequency_fidvid(data, newstate);
1210 	if (ret) {
1211 		printk(KERN_ERR PFX "transition frequency failed\n");
1212 		ret = 1;
1213 		mutex_unlock(&fidvid_mutex);
1214 		goto err_out;
1215 	}
1216 	mutex_unlock(&fidvid_mutex);
1217 
1218 	if (cpu_family == CPU_HW_PSTATE)
1219 		pol->cur = find_khz_freq_from_pstate(data->powernow_table,
1220 				data->powernow_table[newstate].index);
1221 	else
1222 		pol->cur = find_khz_freq_from_fid(data->currfid);
1223 	ret = 0;
1224 
1225 err_out:
1226 	set_cpus_allowed_ptr(current, oldmask);
1227 	free_cpumask_var(oldmask);
1228 	return ret;
1229 }
1230 
1231 /* Driver entry point to verify the policy and range of frequencies */
1232 static int powernowk8_verify(struct cpufreq_policy *pol)
1233 {
1234 	struct powernow_k8_data *data = per_cpu(powernow_data, pol->cpu);
1235 
1236 	if (!data)
1237 		return -EINVAL;
1238 
1239 	return cpufreq_frequency_table_verify(pol, data->powernow_table);
1240 }
1241 
1242 struct init_on_cpu {
1243 	struct powernow_k8_data *data;
1244 	int rc;
1245 };
1246 
1247 static void __cpuinit powernowk8_cpu_init_on_cpu(void *_init_on_cpu)
1248 {
1249 	struct init_on_cpu *init_on_cpu = _init_on_cpu;
1250 
1251 	if (pending_bit_stuck()) {
1252 		printk(KERN_ERR PFX "failing init, change pending bit set\n");
1253 		init_on_cpu->rc = -ENODEV;
1254 		return;
1255 	}
1256 
1257 	if (query_current_values_with_pending_wait(init_on_cpu->data)) {
1258 		init_on_cpu->rc = -ENODEV;
1259 		return;
1260 	}
1261 
1262 	if (cpu_family == CPU_OPTERON)
1263 		fidvid_msr_init();
1264 
1265 	init_on_cpu->rc = 0;
1266 }
1267 
1268 /* per CPU init entry point to the driver */
1269 static int __cpuinit powernowk8_cpu_init(struct cpufreq_policy *pol)
1270 {
1271 	static const char ACPI_PSS_BIOS_BUG_MSG[] =
1272 		KERN_ERR FW_BUG PFX "No compatible ACPI _PSS objects found.\n"
1273 		FW_BUG PFX "Try again with latest BIOS.\n";
1274 	struct powernow_k8_data *data;
1275 	struct init_on_cpu init_on_cpu;
1276 	int rc;
1277 	struct cpuinfo_x86 *c = &cpu_data(pol->cpu);
1278 
1279 	if (!cpu_online(pol->cpu))
1280 		return -ENODEV;
1281 
1282 	smp_call_function_single(pol->cpu, check_supported_cpu, &rc, 1);
1283 	if (rc)
1284 		return -ENODEV;
1285 
1286 	data = kzalloc(sizeof(struct powernow_k8_data), GFP_KERNEL);
1287 	if (!data) {
1288 		printk(KERN_ERR PFX "unable to alloc powernow_k8_data");
1289 		return -ENOMEM;
1290 	}
1291 
1292 	data->cpu = pol->cpu;
1293 	data->currpstate = HW_PSTATE_INVALID;
1294 
1295 	if (powernow_k8_cpu_init_acpi(data)) {
1296 		/*
1297 		 * Use the PSB BIOS structure. This is only available on
1298 		 * an UP version, and is deprecated by AMD.
1299 		 */
1300 		if (num_online_cpus() != 1) {
1301 			printk_once(ACPI_PSS_BIOS_BUG_MSG);
1302 			goto err_out;
1303 		}
1304 		if (pol->cpu != 0) {
1305 			printk(KERN_ERR FW_BUG PFX "No ACPI _PSS objects for "
1306 			       "CPU other than CPU0. Complain to your BIOS "
1307 			       "vendor.\n");
1308 			goto err_out;
1309 		}
1310 		rc = find_psb_table(data);
1311 		if (rc)
1312 			goto err_out;
1313 
1314 		/* Take a crude guess here.
1315 		 * That guess was in microseconds, so multiply with 1000 */
1316 		pol->cpuinfo.transition_latency = (
1317 			 ((data->rvo + 8) * data->vstable * VST_UNITS_20US) +
1318 			 ((1 << data->irt) * 30)) * 1000;
1319 	} else /* ACPI _PSS objects available */
1320 		pol->cpuinfo.transition_latency = get_transition_latency(data);
1321 
1322 	/* only run on specific CPU from here on */
1323 	init_on_cpu.data = data;
1324 	smp_call_function_single(data->cpu, powernowk8_cpu_init_on_cpu,
1325 				 &init_on_cpu, 1);
1326 	rc = init_on_cpu.rc;
1327 	if (rc != 0)
1328 		goto err_out_exit_acpi;
1329 
1330 	if (cpu_family == CPU_HW_PSTATE)
1331 		cpumask_copy(pol->cpus, cpumask_of(pol->cpu));
1332 	else
1333 		cpumask_copy(pol->cpus, cpu_core_mask(pol->cpu));
1334 	data->available_cores = pol->cpus;
1335 
1336 	if (cpu_family == CPU_HW_PSTATE)
1337 		pol->cur = find_khz_freq_from_pstate(data->powernow_table,
1338 				data->currpstate);
1339 	else
1340 		pol->cur = find_khz_freq_from_fid(data->currfid);
1341 	pr_debug("policy current frequency %d kHz\n", pol->cur);
1342 
1343 	/* min/max the cpu is capable of */
1344 	if (cpufreq_frequency_table_cpuinfo(pol, data->powernow_table)) {
1345 		printk(KERN_ERR FW_BUG PFX "invalid powernow_table\n");
1346 		powernow_k8_cpu_exit_acpi(data);
1347 		kfree(data->powernow_table);
1348 		kfree(data);
1349 		return -EINVAL;
1350 	}
1351 
1352 	/* Check for APERF/MPERF support in hardware */
1353 	if (cpu_has(c, X86_FEATURE_APERFMPERF))
1354 		cpufreq_amd64_driver.getavg = cpufreq_get_measured_perf;
1355 
1356 	cpufreq_frequency_table_get_attr(data->powernow_table, pol->cpu);
1357 
1358 	if (cpu_family == CPU_HW_PSTATE)
1359 		pr_debug("cpu_init done, current pstate 0x%x\n",
1360 				data->currpstate);
1361 	else
1362 		pr_debug("cpu_init done, current fid 0x%x, vid 0x%x\n",
1363 			data->currfid, data->currvid);
1364 
1365 	per_cpu(powernow_data, pol->cpu) = data;
1366 
1367 	return 0;
1368 
1369 err_out_exit_acpi:
1370 	powernow_k8_cpu_exit_acpi(data);
1371 
1372 err_out:
1373 	kfree(data);
1374 	return -ENODEV;
1375 }
1376 
1377 static int __devexit powernowk8_cpu_exit(struct cpufreq_policy *pol)
1378 {
1379 	struct powernow_k8_data *data = per_cpu(powernow_data, pol->cpu);
1380 
1381 	if (!data)
1382 		return -EINVAL;
1383 
1384 	powernow_k8_cpu_exit_acpi(data);
1385 
1386 	cpufreq_frequency_table_put_attr(pol->cpu);
1387 
1388 	kfree(data->powernow_table);
1389 	kfree(data);
1390 	per_cpu(powernow_data, pol->cpu) = NULL;
1391 
1392 	return 0;
1393 }
1394 
1395 static void query_values_on_cpu(void *_err)
1396 {
1397 	int *err = _err;
1398 	struct powernow_k8_data *data = __this_cpu_read(powernow_data);
1399 
1400 	*err = query_current_values_with_pending_wait(data);
1401 }
1402 
1403 static unsigned int powernowk8_get(unsigned int cpu)
1404 {
1405 	struct powernow_k8_data *data = per_cpu(powernow_data, cpu);
1406 	unsigned int khz = 0;
1407 	int err;
1408 
1409 	if (!data)
1410 		return 0;
1411 
1412 	smp_call_function_single(cpu, query_values_on_cpu, &err, true);
1413 	if (err)
1414 		goto out;
1415 
1416 	if (cpu_family == CPU_HW_PSTATE)
1417 		khz = find_khz_freq_from_pstate(data->powernow_table,
1418 						data->currpstate);
1419 	else
1420 		khz = find_khz_freq_from_fid(data->currfid);
1421 
1422 
1423 out:
1424 	return khz;
1425 }
1426 
1427 static void _cpb_toggle_msrs(bool t)
1428 {
1429 	int cpu;
1430 
1431 	get_online_cpus();
1432 
1433 	rdmsr_on_cpus(cpu_online_mask, MSR_K7_HWCR, msrs);
1434 
1435 	for_each_cpu(cpu, cpu_online_mask) {
1436 		struct msr *reg = per_cpu_ptr(msrs, cpu);
1437 		if (t)
1438 			reg->l &= ~BIT(25);
1439 		else
1440 			reg->l |= BIT(25);
1441 	}
1442 	wrmsr_on_cpus(cpu_online_mask, MSR_K7_HWCR, msrs);
1443 
1444 	put_online_cpus();
1445 }
1446 
1447 /*
1448  * Switch on/off core performance boosting.
1449  *
1450  * 0=disable
1451  * 1=enable.
1452  */
1453 static void cpb_toggle(bool t)
1454 {
1455 	if (!cpb_capable)
1456 		return;
1457 
1458 	if (t && !cpb_enabled) {
1459 		cpb_enabled = true;
1460 		_cpb_toggle_msrs(t);
1461 		printk(KERN_INFO PFX "Core Boosting enabled.\n");
1462 	} else if (!t && cpb_enabled) {
1463 		cpb_enabled = false;
1464 		_cpb_toggle_msrs(t);
1465 		printk(KERN_INFO PFX "Core Boosting disabled.\n");
1466 	}
1467 }
1468 
1469 static ssize_t store_cpb(struct cpufreq_policy *policy, const char *buf,
1470 				 size_t count)
1471 {
1472 	int ret = -EINVAL;
1473 	unsigned long val = 0;
1474 
1475 	ret = strict_strtoul(buf, 10, &val);
1476 	if (!ret && (val == 0 || val == 1) && cpb_capable)
1477 		cpb_toggle(val);
1478 	else
1479 		return -EINVAL;
1480 
1481 	return count;
1482 }
1483 
1484 static ssize_t show_cpb(struct cpufreq_policy *policy, char *buf)
1485 {
1486 	return sprintf(buf, "%u\n", cpb_enabled);
1487 }
1488 
1489 #define define_one_rw(_name) \
1490 static struct freq_attr _name = \
1491 __ATTR(_name, 0644, show_##_name, store_##_name)
1492 
1493 define_one_rw(cpb);
1494 
1495 static struct freq_attr *powernow_k8_attr[] = {
1496 	&cpufreq_freq_attr_scaling_available_freqs,
1497 	&cpb,
1498 	NULL,
1499 };
1500 
1501 static struct cpufreq_driver cpufreq_amd64_driver = {
1502 	.verify		= powernowk8_verify,
1503 	.target		= powernowk8_target,
1504 	.bios_limit	= acpi_processor_get_bios_limit,
1505 	.init		= powernowk8_cpu_init,
1506 	.exit		= __devexit_p(powernowk8_cpu_exit),
1507 	.get		= powernowk8_get,
1508 	.name		= "powernow-k8",
1509 	.owner		= THIS_MODULE,
1510 	.attr		= powernow_k8_attr,
1511 };
1512 
1513 /*
1514  * Clear the boost-disable flag on the CPU_DOWN path so that this cpu
1515  * cannot block the remaining ones from boosting. On the CPU_UP path we
1516  * simply keep the boost-disable flag in sync with the current global
1517  * state.
1518  */
1519 static int cpb_notify(struct notifier_block *nb, unsigned long action,
1520 		      void *hcpu)
1521 {
1522 	unsigned cpu = (long)hcpu;
1523 	u32 lo, hi;
1524 
1525 	switch (action) {
1526 	case CPU_UP_PREPARE:
1527 	case CPU_UP_PREPARE_FROZEN:
1528 
1529 		if (!cpb_enabled) {
1530 			rdmsr_on_cpu(cpu, MSR_K7_HWCR, &lo, &hi);
1531 			lo |= BIT(25);
1532 			wrmsr_on_cpu(cpu, MSR_K7_HWCR, lo, hi);
1533 		}
1534 		break;
1535 
1536 	case CPU_DOWN_PREPARE:
1537 	case CPU_DOWN_PREPARE_FROZEN:
1538 		rdmsr_on_cpu(cpu, MSR_K7_HWCR, &lo, &hi);
1539 		lo &= ~BIT(25);
1540 		wrmsr_on_cpu(cpu, MSR_K7_HWCR, lo, hi);
1541 		break;
1542 
1543 	default:
1544 		break;
1545 	}
1546 
1547 	return NOTIFY_OK;
1548 }
1549 
1550 static struct notifier_block cpb_nb = {
1551 	.notifier_call		= cpb_notify,
1552 };
1553 
1554 /* driver entry point for init */
1555 static int __cpuinit powernowk8_init(void)
1556 {
1557 	unsigned int i, supported_cpus = 0, cpu;
1558 	int rv;
1559 
1560 	if (!x86_match_cpu(powernow_k8_ids))
1561 		return -ENODEV;
1562 
1563 	for_each_online_cpu(i) {
1564 		int rc;
1565 		smp_call_function_single(i, check_supported_cpu, &rc, 1);
1566 		if (rc == 0)
1567 			supported_cpus++;
1568 	}
1569 
1570 	if (supported_cpus != num_online_cpus())
1571 		return -ENODEV;
1572 
1573 	printk(KERN_INFO PFX "Found %d %s (%d cpu cores) (" VERSION ")\n",
1574 		num_online_nodes(), boot_cpu_data.x86_model_id, supported_cpus);
1575 
1576 	if (boot_cpu_has(X86_FEATURE_CPB)) {
1577 
1578 		cpb_capable = true;
1579 
1580 		msrs = msrs_alloc();
1581 		if (!msrs) {
1582 			printk(KERN_ERR "%s: Error allocating msrs!\n", __func__);
1583 			return -ENOMEM;
1584 		}
1585 
1586 		register_cpu_notifier(&cpb_nb);
1587 
1588 		rdmsr_on_cpus(cpu_online_mask, MSR_K7_HWCR, msrs);
1589 
1590 		for_each_cpu(cpu, cpu_online_mask) {
1591 			struct msr *reg = per_cpu_ptr(msrs, cpu);
1592 			cpb_enabled |= !(!!(reg->l & BIT(25)));
1593 		}
1594 
1595 		printk(KERN_INFO PFX "Core Performance Boosting: %s.\n",
1596 			(cpb_enabled ? "on" : "off"));
1597 	}
1598 
1599 	rv = cpufreq_register_driver(&cpufreq_amd64_driver);
1600 	if (rv < 0 && boot_cpu_has(X86_FEATURE_CPB)) {
1601 		unregister_cpu_notifier(&cpb_nb);
1602 		msrs_free(msrs);
1603 		msrs = NULL;
1604 	}
1605 	return rv;
1606 }
1607 
1608 /* driver entry point for term */
1609 static void __exit powernowk8_exit(void)
1610 {
1611 	pr_debug("exit\n");
1612 
1613 	if (boot_cpu_has(X86_FEATURE_CPB)) {
1614 		msrs_free(msrs);
1615 		msrs = NULL;
1616 
1617 		unregister_cpu_notifier(&cpb_nb);
1618 	}
1619 
1620 	cpufreq_unregister_driver(&cpufreq_amd64_driver);
1621 }
1622 
1623 MODULE_AUTHOR("Paul Devriendt <paul.devriendt@amd.com> and "
1624 		"Mark Langsdorf <mark.langsdorf@amd.com>");
1625 MODULE_DESCRIPTION("AMD Athlon 64 and Opteron processor frequency driver.");
1626 MODULE_LICENSE("GPL");
1627 
1628 late_initcall(powernowk8_init);
1629 module_exit(powernowk8_exit);
1630