xref: /linux/drivers/cpufreq/powernow-k8.c (revision f9bff0e31881d03badf191d3b0005839391f5f2b)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  *   (c) 2003-2012 Advanced Micro Devices, Inc.
4  *
5  *  Maintainer:
6  *  Andreas Herrmann <herrmann.der.user@googlemail.com>
7  *
8  *  Based on the powernow-k7.c module written by Dave Jones.
9  *  (C) 2003 Dave Jones on behalf of SuSE Labs
10  *  (C) 2004 Dominik Brodowski <linux@brodo.de>
11  *  (C) 2004 Pavel Machek <pavel@ucw.cz>
12  *  Based upon datasheets & sample CPUs kindly provided by AMD.
13  *
14  *  Valuable input gratefully received from Dave Jones, Pavel Machek,
15  *  Dominik Brodowski, Jacob Shin, and others.
16  *  Originally developed by Paul Devriendt.
17  *
18  *  Processor information obtained from Chapter 9 (Power and Thermal
19  *  Management) of the "BIOS and Kernel Developer's Guide (BKDG) for
20  *  the AMD Athlon 64 and AMD Opteron Processors" and section "2.x
21  *  Power Management" in BKDGs for newer AMD CPU families.
22  *
23  *  Tables for specific CPUs can be inferred from AMD's processor
24  *  power and thermal data sheets, (e.g. 30417.pdf, 30430.pdf, 43375.pdf)
25  */
26 
27 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
28 
29 #include <linux/kernel.h>
30 #include <linux/smp.h>
31 #include <linux/module.h>
32 #include <linux/init.h>
33 #include <linux/cpufreq.h>
34 #include <linux/slab.h>
35 #include <linux/string.h>
36 #include <linux/cpumask.h>
37 #include <linux/io.h>
38 #include <linux/delay.h>
39 
40 #include <asm/msr.h>
41 #include <asm/cpu_device_id.h>
42 
43 #include <linux/acpi.h>
44 #include <linux/mutex.h>
45 #include <acpi/processor.h>
46 
47 #define VERSION "version 2.20.00"
48 #include "powernow-k8.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 struct cpufreq_driver cpufreq_amd64_driver;
56 
57 /* Return a frequency in MHz, given an input fid */
58 static u32 find_freq_from_fid(u32 fid)
59 {
60 	return 800 + (fid * 100);
61 }
62 
63 /* Return a frequency in KHz, given an input fid */
64 static u32 find_khz_freq_from_fid(u32 fid)
65 {
66 	return 1000 * find_freq_from_fid(fid);
67 }
68 
69 /* Return the vco fid for an input fid
70  *
71  * Each "low" fid has corresponding "high" fid, and you can get to "low" fids
72  * only from corresponding high fids. This returns "high" fid corresponding to
73  * "low" one.
74  */
75 static u32 convert_fid_to_vco_fid(u32 fid)
76 {
77 	if (fid < HI_FID_TABLE_BOTTOM)
78 		return 8 + (2 * fid);
79 	else
80 		return fid;
81 }
82 
83 /*
84  * Return 1 if the pending bit is set. Unless we just instructed the processor
85  * to transition to a new state, seeing this bit set is really bad news.
86  */
87 static int pending_bit_stuck(void)
88 {
89 	u32 lo, hi __always_unused;
90 
91 	rdmsr(MSR_FIDVID_STATUS, lo, hi);
92 	return lo & MSR_S_LO_CHANGE_PENDING ? 1 : 0;
93 }
94 
95 /*
96  * Update the global current fid / vid values from the status msr.
97  * Returns 1 on error.
98  */
99 static int query_current_values_with_pending_wait(struct powernow_k8_data *data)
100 {
101 	u32 lo, hi;
102 	u32 i = 0;
103 
104 	do {
105 		if (i++ > 10000) {
106 			pr_debug("detected change pending stuck\n");
107 			return 1;
108 		}
109 		rdmsr(MSR_FIDVID_STATUS, lo, hi);
110 	} while (lo & MSR_S_LO_CHANGE_PENDING);
111 
112 	data->currvid = hi & MSR_S_HI_CURRENT_VID;
113 	data->currfid = lo & MSR_S_LO_CURRENT_FID;
114 
115 	return 0;
116 }
117 
118 /* the isochronous relief time */
119 static void count_off_irt(struct powernow_k8_data *data)
120 {
121 	udelay((1 << data->irt) * 10);
122 }
123 
124 /* the voltage stabilization time */
125 static void count_off_vst(struct powernow_k8_data *data)
126 {
127 	udelay(data->vstable * VST_UNITS_20US);
128 }
129 
130 /* need to init the control msr to a safe value (for each cpu) */
131 static void fidvid_msr_init(void)
132 {
133 	u32 lo, hi;
134 	u8 fid, vid;
135 
136 	rdmsr(MSR_FIDVID_STATUS, lo, hi);
137 	vid = hi & MSR_S_HI_CURRENT_VID;
138 	fid = lo & MSR_S_LO_CURRENT_FID;
139 	lo = fid | (vid << MSR_C_LO_VID_SHIFT);
140 	hi = MSR_C_HI_STP_GNT_BENIGN;
141 	pr_debug("cpu%d, init lo 0x%x, hi 0x%x\n", smp_processor_id(), lo, hi);
142 	wrmsr(MSR_FIDVID_CTL, lo, hi);
143 }
144 
145 /* write the new fid value along with the other control fields to the msr */
146 static int write_new_fid(struct powernow_k8_data *data, u32 fid)
147 {
148 	u32 lo;
149 	u32 savevid = data->currvid;
150 	u32 i = 0;
151 
152 	if ((fid & INVALID_FID_MASK) || (data->currvid & INVALID_VID_MASK)) {
153 		pr_err("internal error - overflow on fid write\n");
154 		return 1;
155 	}
156 
157 	lo = fid;
158 	lo |= (data->currvid << MSR_C_LO_VID_SHIFT);
159 	lo |= MSR_C_LO_INIT_FID_VID;
160 
161 	pr_debug("writing fid 0x%x, lo 0x%x, hi 0x%x\n",
162 		fid, lo, data->plllock * PLL_LOCK_CONVERSION);
163 
164 	do {
165 		wrmsr(MSR_FIDVID_CTL, lo, data->plllock * PLL_LOCK_CONVERSION);
166 		if (i++ > 100) {
167 			pr_err("Hardware error - pending bit very stuck - no further pstate changes possible\n");
168 			return 1;
169 		}
170 	} while (query_current_values_with_pending_wait(data));
171 
172 	count_off_irt(data);
173 
174 	if (savevid != data->currvid) {
175 		pr_err("vid change on fid trans, old 0x%x, new 0x%x\n",
176 		       savevid, data->currvid);
177 		return 1;
178 	}
179 
180 	if (fid != data->currfid) {
181 		pr_err("fid trans failed, fid 0x%x, curr 0x%x\n", fid,
182 			data->currfid);
183 		return 1;
184 	}
185 
186 	return 0;
187 }
188 
189 /* Write a new vid to the hardware */
190 static int write_new_vid(struct powernow_k8_data *data, u32 vid)
191 {
192 	u32 lo;
193 	u32 savefid = data->currfid;
194 	int i = 0;
195 
196 	if ((data->currfid & INVALID_FID_MASK) || (vid & INVALID_VID_MASK)) {
197 		pr_err("internal error - overflow on vid write\n");
198 		return 1;
199 	}
200 
201 	lo = data->currfid;
202 	lo |= (vid << MSR_C_LO_VID_SHIFT);
203 	lo |= MSR_C_LO_INIT_FID_VID;
204 
205 	pr_debug("writing vid 0x%x, lo 0x%x, hi 0x%x\n",
206 		vid, lo, STOP_GRANT_5NS);
207 
208 	do {
209 		wrmsr(MSR_FIDVID_CTL, lo, STOP_GRANT_5NS);
210 		if (i++ > 100) {
211 			pr_err("internal error - pending bit very stuck - no further pstate changes possible\n");
212 			return 1;
213 		}
214 	} while (query_current_values_with_pending_wait(data));
215 
216 	if (savefid != data->currfid) {
217 		pr_err("fid changed on vid trans, old 0x%x new 0x%x\n",
218 			savefid, data->currfid);
219 		return 1;
220 	}
221 
222 	if (vid != data->currvid) {
223 		pr_err("vid trans failed, vid 0x%x, curr 0x%x\n",
224 				vid, data->currvid);
225 		return 1;
226 	}
227 
228 	return 0;
229 }
230 
231 /*
232  * Reduce the vid by the max of step or reqvid.
233  * Decreasing vid codes represent increasing voltages:
234  * vid of 0 is 1.550V, vid of 0x1e is 0.800V, vid of VID_OFF is off.
235  */
236 static int decrease_vid_code_by_step(struct powernow_k8_data *data,
237 		u32 reqvid, u32 step)
238 {
239 	if ((data->currvid - reqvid) > step)
240 		reqvid = data->currvid - step;
241 
242 	if (write_new_vid(data, reqvid))
243 		return 1;
244 
245 	count_off_vst(data);
246 
247 	return 0;
248 }
249 
250 /* Change Opteron/Athlon64 fid and vid, by the 3 phases. */
251 static int transition_fid_vid(struct powernow_k8_data *data,
252 		u32 reqfid, u32 reqvid)
253 {
254 	if (core_voltage_pre_transition(data, reqvid, reqfid))
255 		return 1;
256 
257 	if (core_frequency_transition(data, reqfid))
258 		return 1;
259 
260 	if (core_voltage_post_transition(data, reqvid))
261 		return 1;
262 
263 	if (query_current_values_with_pending_wait(data))
264 		return 1;
265 
266 	if ((reqfid != data->currfid) || (reqvid != data->currvid)) {
267 		pr_err("failed (cpu%d): req 0x%x 0x%x, curr 0x%x 0x%x\n",
268 				smp_processor_id(),
269 				reqfid, reqvid, data->currfid, data->currvid);
270 		return 1;
271 	}
272 
273 	pr_debug("transitioned (cpu%d): new fid 0x%x, vid 0x%x\n",
274 		smp_processor_id(), data->currfid, data->currvid);
275 
276 	return 0;
277 }
278 
279 /* Phase 1 - core voltage transition ... setup voltage */
280 static int core_voltage_pre_transition(struct powernow_k8_data *data,
281 		u32 reqvid, u32 reqfid)
282 {
283 	u32 rvosteps = data->rvo;
284 	u32 savefid = data->currfid;
285 	u32 maxvid, lo __always_unused, rvomult = 1;
286 
287 	pr_debug("ph1 (cpu%d): start, currfid 0x%x, currvid 0x%x, reqvid 0x%x, rvo 0x%x\n",
288 		smp_processor_id(),
289 		data->currfid, data->currvid, reqvid, data->rvo);
290 
291 	if ((savefid < LO_FID_TABLE_TOP) && (reqfid < LO_FID_TABLE_TOP))
292 		rvomult = 2;
293 	rvosteps *= rvomult;
294 	rdmsr(MSR_FIDVID_STATUS, lo, maxvid);
295 	maxvid = 0x1f & (maxvid >> 16);
296 	pr_debug("ph1 maxvid=0x%x\n", maxvid);
297 	if (reqvid < maxvid) /* lower numbers are higher voltages */
298 		reqvid = maxvid;
299 
300 	while (data->currvid > reqvid) {
301 		pr_debug("ph1: curr 0x%x, req vid 0x%x\n",
302 			data->currvid, reqvid);
303 		if (decrease_vid_code_by_step(data, reqvid, data->vidmvs))
304 			return 1;
305 	}
306 
307 	while ((rvosteps > 0) &&
308 			((rvomult * data->rvo + data->currvid) > reqvid)) {
309 		if (data->currvid == maxvid) {
310 			rvosteps = 0;
311 		} else {
312 			pr_debug("ph1: changing vid for rvo, req 0x%x\n",
313 				data->currvid - 1);
314 			if (decrease_vid_code_by_step(data, data->currvid-1, 1))
315 				return 1;
316 			rvosteps--;
317 		}
318 	}
319 
320 	if (query_current_values_with_pending_wait(data))
321 		return 1;
322 
323 	if (savefid != data->currfid) {
324 		pr_err("ph1 err, currfid changed 0x%x\n", data->currfid);
325 		return 1;
326 	}
327 
328 	pr_debug("ph1 complete, currfid 0x%x, currvid 0x%x\n",
329 		data->currfid, data->currvid);
330 
331 	return 0;
332 }
333 
334 /* Phase 2 - core frequency transition */
335 static int core_frequency_transition(struct powernow_k8_data *data, u32 reqfid)
336 {
337 	u32 vcoreqfid, vcocurrfid, vcofiddiff;
338 	u32 fid_interval, savevid = data->currvid;
339 
340 	if (data->currfid == reqfid) {
341 		pr_err("ph2 null fid transition 0x%x\n", data->currfid);
342 		return 0;
343 	}
344 
345 	pr_debug("ph2 (cpu%d): starting, currfid 0x%x, currvid 0x%x, reqfid 0x%x\n",
346 		smp_processor_id(),
347 		data->currfid, data->currvid, reqfid);
348 
349 	vcoreqfid = convert_fid_to_vco_fid(reqfid);
350 	vcocurrfid = convert_fid_to_vco_fid(data->currfid);
351 	vcofiddiff = vcocurrfid > vcoreqfid ? vcocurrfid - vcoreqfid
352 	    : vcoreqfid - vcocurrfid;
353 
354 	if ((reqfid <= LO_FID_TABLE_TOP) && (data->currfid <= LO_FID_TABLE_TOP))
355 		vcofiddiff = 0;
356 
357 	while (vcofiddiff > 2) {
358 		(data->currfid & 1) ? (fid_interval = 1) : (fid_interval = 2);
359 
360 		if (reqfid > data->currfid) {
361 			if (data->currfid > LO_FID_TABLE_TOP) {
362 				if (write_new_fid(data,
363 						data->currfid + fid_interval))
364 					return 1;
365 			} else {
366 				if (write_new_fid
367 				    (data,
368 				     2 + convert_fid_to_vco_fid(data->currfid)))
369 					return 1;
370 			}
371 		} else {
372 			if (write_new_fid(data, data->currfid - fid_interval))
373 				return 1;
374 		}
375 
376 		vcocurrfid = convert_fid_to_vco_fid(data->currfid);
377 		vcofiddiff = vcocurrfid > vcoreqfid ? vcocurrfid - vcoreqfid
378 		    : vcoreqfid - vcocurrfid;
379 	}
380 
381 	if (write_new_fid(data, reqfid))
382 		return 1;
383 
384 	if (query_current_values_with_pending_wait(data))
385 		return 1;
386 
387 	if (data->currfid != reqfid) {
388 		pr_err("ph2: mismatch, failed fid transition, curr 0x%x, req 0x%x\n",
389 			data->currfid, reqfid);
390 		return 1;
391 	}
392 
393 	if (savevid != data->currvid) {
394 		pr_err("ph2: vid changed, save 0x%x, curr 0x%x\n",
395 			savevid, data->currvid);
396 		return 1;
397 	}
398 
399 	pr_debug("ph2 complete, currfid 0x%x, currvid 0x%x\n",
400 		data->currfid, data->currvid);
401 
402 	return 0;
403 }
404 
405 /* Phase 3 - core voltage transition flow ... jump to the final vid. */
406 static int core_voltage_post_transition(struct powernow_k8_data *data,
407 		u32 reqvid)
408 {
409 	u32 savefid = data->currfid;
410 	u32 savereqvid = reqvid;
411 
412 	pr_debug("ph3 (cpu%d): starting, currfid 0x%x, currvid 0x%x\n",
413 		smp_processor_id(),
414 		data->currfid, data->currvid);
415 
416 	if (reqvid != data->currvid) {
417 		if (write_new_vid(data, reqvid))
418 			return 1;
419 
420 		if (savefid != data->currfid) {
421 			pr_err("ph3: bad fid change, save 0x%x, curr 0x%x\n",
422 				savefid, data->currfid);
423 			return 1;
424 		}
425 
426 		if (data->currvid != reqvid) {
427 			pr_err("ph3: failed vid transition\n, req 0x%x, curr 0x%x",
428 				reqvid, data->currvid);
429 			return 1;
430 		}
431 	}
432 
433 	if (query_current_values_with_pending_wait(data))
434 		return 1;
435 
436 	if (savereqvid != data->currvid) {
437 		pr_debug("ph3 failed, currvid 0x%x\n", data->currvid);
438 		return 1;
439 	}
440 
441 	if (savefid != data->currfid) {
442 		pr_debug("ph3 failed, currfid changed 0x%x\n",
443 			data->currfid);
444 		return 1;
445 	}
446 
447 	pr_debug("ph3 complete, currfid 0x%x, currvid 0x%x\n",
448 		data->currfid, data->currvid);
449 
450 	return 0;
451 }
452 
453 static const struct x86_cpu_id powernow_k8_ids[] = {
454 	/* IO based frequency switching */
455 	X86_MATCH_VENDOR_FAM(AMD, 0xf, NULL),
456 	{}
457 };
458 MODULE_DEVICE_TABLE(x86cpu, powernow_k8_ids);
459 
460 static void check_supported_cpu(void *_rc)
461 {
462 	u32 eax, ebx, ecx, edx;
463 	int *rc = _rc;
464 
465 	*rc = -ENODEV;
466 
467 	eax = cpuid_eax(CPUID_PROCESSOR_SIGNATURE);
468 
469 	if ((eax & CPUID_XFAM) == CPUID_XFAM_K8) {
470 		if (((eax & CPUID_USE_XFAM_XMOD) != CPUID_USE_XFAM_XMOD) ||
471 		    ((eax & CPUID_XMOD) > CPUID_XMOD_REV_MASK)) {
472 			pr_info("Processor cpuid %x not supported\n", eax);
473 			return;
474 		}
475 
476 		eax = cpuid_eax(CPUID_GET_MAX_CAPABILITIES);
477 		if (eax < CPUID_FREQ_VOLT_CAPABILITIES) {
478 			pr_info("No frequency change capabilities detected\n");
479 			return;
480 		}
481 
482 		cpuid(CPUID_FREQ_VOLT_CAPABILITIES, &eax, &ebx, &ecx, &edx);
483 		if ((edx & P_STATE_TRANSITION_CAPABLE)
484 			!= P_STATE_TRANSITION_CAPABLE) {
485 			pr_info("Power state transitions not supported\n");
486 			return;
487 		}
488 		*rc = 0;
489 	}
490 }
491 
492 static int check_pst_table(struct powernow_k8_data *data, struct pst_s *pst,
493 		u8 maxvid)
494 {
495 	unsigned int j;
496 	u8 lastfid = 0xff;
497 
498 	for (j = 0; j < data->numps; j++) {
499 		if (pst[j].vid > LEAST_VID) {
500 			pr_err(FW_BUG "vid %d invalid : 0x%x\n", j,
501 				pst[j].vid);
502 			return -EINVAL;
503 		}
504 		if (pst[j].vid < data->rvo) {
505 			/* vid + rvo >= 0 */
506 			pr_err(FW_BUG "0 vid exceeded with pstate %d\n", j);
507 			return -ENODEV;
508 		}
509 		if (pst[j].vid < maxvid + data->rvo) {
510 			/* vid + rvo >= maxvid */
511 			pr_err(FW_BUG "maxvid exceeded with pstate %d\n", j);
512 			return -ENODEV;
513 		}
514 		if (pst[j].fid > MAX_FID) {
515 			pr_err(FW_BUG "maxfid exceeded with pstate %d\n", j);
516 			return -ENODEV;
517 		}
518 		if (j && (pst[j].fid < HI_FID_TABLE_BOTTOM)) {
519 			/* Only first fid is allowed to be in "low" range */
520 			pr_err(FW_BUG "two low fids - %d : 0x%x\n", j,
521 				pst[j].fid);
522 			return -EINVAL;
523 		}
524 		if (pst[j].fid < lastfid)
525 			lastfid = pst[j].fid;
526 	}
527 	if (lastfid & 1) {
528 		pr_err(FW_BUG "lastfid invalid\n");
529 		return -EINVAL;
530 	}
531 	if (lastfid > LO_FID_TABLE_TOP)
532 		pr_info(FW_BUG "first fid not from lo freq table\n");
533 
534 	return 0;
535 }
536 
537 static void invalidate_entry(struct cpufreq_frequency_table *powernow_table,
538 		unsigned int entry)
539 {
540 	powernow_table[entry].frequency = CPUFREQ_ENTRY_INVALID;
541 }
542 
543 static void print_basics(struct powernow_k8_data *data)
544 {
545 	int j;
546 	for (j = 0; j < data->numps; j++) {
547 		if (data->powernow_table[j].frequency !=
548 				CPUFREQ_ENTRY_INVALID) {
549 			pr_info("fid 0x%x (%d MHz), vid 0x%x\n",
550 				data->powernow_table[j].driver_data & 0xff,
551 				data->powernow_table[j].frequency/1000,
552 				data->powernow_table[j].driver_data >> 8);
553 		}
554 	}
555 	if (data->batps)
556 		pr_info("Only %d pstates on battery\n", data->batps);
557 }
558 
559 static int fill_powernow_table(struct powernow_k8_data *data,
560 		struct pst_s *pst, u8 maxvid)
561 {
562 	struct cpufreq_frequency_table *powernow_table;
563 	unsigned int j;
564 
565 	if (data->batps) {
566 		/* use ACPI support to get full speed on mains power */
567 		pr_warn("Only %d pstates usable (use ACPI driver for full range\n",
568 			data->batps);
569 		data->numps = data->batps;
570 	}
571 
572 	for (j = 1; j < data->numps; j++) {
573 		if (pst[j-1].fid >= pst[j].fid) {
574 			pr_err("PST out of sequence\n");
575 			return -EINVAL;
576 		}
577 	}
578 
579 	if (data->numps < 2) {
580 		pr_err("no p states to transition\n");
581 		return -ENODEV;
582 	}
583 
584 	if (check_pst_table(data, pst, maxvid))
585 		return -EINVAL;
586 
587 	powernow_table = kzalloc((sizeof(*powernow_table)
588 		* (data->numps + 1)), GFP_KERNEL);
589 	if (!powernow_table)
590 		return -ENOMEM;
591 
592 	for (j = 0; j < data->numps; j++) {
593 		int freq;
594 		powernow_table[j].driver_data = pst[j].fid; /* lower 8 bits */
595 		powernow_table[j].driver_data |= (pst[j].vid << 8); /* upper 8 bits */
596 		freq = find_khz_freq_from_fid(pst[j].fid);
597 		powernow_table[j].frequency = freq;
598 	}
599 	powernow_table[data->numps].frequency = CPUFREQ_TABLE_END;
600 	powernow_table[data->numps].driver_data = 0;
601 
602 	if (query_current_values_with_pending_wait(data)) {
603 		kfree(powernow_table);
604 		return -EIO;
605 	}
606 
607 	pr_debug("cfid 0x%x, cvid 0x%x\n", data->currfid, data->currvid);
608 	data->powernow_table = powernow_table;
609 	if (cpumask_first(topology_core_cpumask(data->cpu)) == data->cpu)
610 		print_basics(data);
611 
612 	for (j = 0; j < data->numps; j++)
613 		if ((pst[j].fid == data->currfid) &&
614 		    (pst[j].vid == data->currvid))
615 			return 0;
616 
617 	pr_debug("currfid/vid do not match PST, ignoring\n");
618 	return 0;
619 }
620 
621 /* Find and validate the PSB/PST table in BIOS. */
622 static int find_psb_table(struct powernow_k8_data *data)
623 {
624 	struct psb_s *psb;
625 	unsigned int i;
626 	u32 mvs;
627 	u8 maxvid;
628 	u32 cpst = 0;
629 	u32 thiscpuid;
630 
631 	for (i = 0xc0000; i < 0xffff0; i += 0x10) {
632 		/* Scan BIOS looking for the signature. */
633 		/* It can not be at ffff0 - it is too big. */
634 
635 		psb = phys_to_virt(i);
636 		if (memcmp(psb, PSB_ID_STRING, PSB_ID_STRING_LEN) != 0)
637 			continue;
638 
639 		pr_debug("found PSB header at 0x%p\n", psb);
640 
641 		pr_debug("table vers: 0x%x\n", psb->tableversion);
642 		if (psb->tableversion != PSB_VERSION_1_4) {
643 			pr_err(FW_BUG "PSB table is not v1.4\n");
644 			return -ENODEV;
645 		}
646 
647 		pr_debug("flags: 0x%x\n", psb->flags1);
648 		if (psb->flags1) {
649 			pr_err(FW_BUG "unknown flags\n");
650 			return -ENODEV;
651 		}
652 
653 		data->vstable = psb->vstable;
654 		pr_debug("voltage stabilization time: %d(*20us)\n",
655 				data->vstable);
656 
657 		pr_debug("flags2: 0x%x\n", psb->flags2);
658 		data->rvo = psb->flags2 & 3;
659 		data->irt = ((psb->flags2) >> 2) & 3;
660 		mvs = ((psb->flags2) >> 4) & 3;
661 		data->vidmvs = 1 << mvs;
662 		data->batps = ((psb->flags2) >> 6) & 3;
663 
664 		pr_debug("ramp voltage offset: %d\n", data->rvo);
665 		pr_debug("isochronous relief time: %d\n", data->irt);
666 		pr_debug("maximum voltage step: %d - 0x%x\n", mvs, data->vidmvs);
667 
668 		pr_debug("numpst: 0x%x\n", psb->num_tables);
669 		cpst = psb->num_tables;
670 		if ((psb->cpuid == 0x00000fc0) ||
671 		    (psb->cpuid == 0x00000fe0)) {
672 			thiscpuid = cpuid_eax(CPUID_PROCESSOR_SIGNATURE);
673 			if ((thiscpuid == 0x00000fc0) ||
674 			    (thiscpuid == 0x00000fe0))
675 				cpst = 1;
676 		}
677 		if (cpst != 1) {
678 			pr_err(FW_BUG "numpst must be 1\n");
679 			return -ENODEV;
680 		}
681 
682 		data->plllock = psb->plllocktime;
683 		pr_debug("plllocktime: 0x%x (units 1us)\n", psb->plllocktime);
684 		pr_debug("maxfid: 0x%x\n", psb->maxfid);
685 		pr_debug("maxvid: 0x%x\n", psb->maxvid);
686 		maxvid = psb->maxvid;
687 
688 		data->numps = psb->numps;
689 		pr_debug("numpstates: 0x%x\n", data->numps);
690 		return fill_powernow_table(data,
691 				(struct pst_s *)(psb+1), maxvid);
692 	}
693 	/*
694 	 * If you see this message, complain to BIOS manufacturer. If
695 	 * he tells you "we do not support Linux" or some similar
696 	 * nonsense, remember that Windows 2000 uses the same legacy
697 	 * mechanism that the old Linux PSB driver uses. Tell them it
698 	 * is broken with Windows 2000.
699 	 *
700 	 * The reference to the AMD documentation is chapter 9 in the
701 	 * BIOS and Kernel Developer's Guide, which is available on
702 	 * www.amd.com
703 	 */
704 	pr_err(FW_BUG "No PSB or ACPI _PSS objects\n");
705 	pr_err("Make sure that your BIOS is up to date and Cool'N'Quiet support is enabled in BIOS setup\n");
706 	return -ENODEV;
707 }
708 
709 static void powernow_k8_acpi_pst_values(struct powernow_k8_data *data,
710 		unsigned int index)
711 {
712 	u64 control;
713 
714 	if (!data->acpi_data.state_count)
715 		return;
716 
717 	control = data->acpi_data.states[index].control;
718 	data->irt = (control >> IRT_SHIFT) & IRT_MASK;
719 	data->rvo = (control >> RVO_SHIFT) & RVO_MASK;
720 	data->exttype = (control >> EXT_TYPE_SHIFT) & EXT_TYPE_MASK;
721 	data->plllock = (control >> PLL_L_SHIFT) & PLL_L_MASK;
722 	data->vidmvs = 1 << ((control >> MVS_SHIFT) & MVS_MASK);
723 	data->vstable = (control >> VST_SHIFT) & VST_MASK;
724 }
725 
726 static int powernow_k8_cpu_init_acpi(struct powernow_k8_data *data)
727 {
728 	struct cpufreq_frequency_table *powernow_table;
729 	int ret_val = -ENODEV;
730 	u64 control, status;
731 
732 	if (acpi_processor_register_performance(&data->acpi_data, data->cpu)) {
733 		pr_debug("register performance failed: bad ACPI data\n");
734 		return -EIO;
735 	}
736 
737 	/* verify the data contained in the ACPI structures */
738 	if (data->acpi_data.state_count <= 1) {
739 		pr_debug("No ACPI P-States\n");
740 		goto err_out;
741 	}
742 
743 	control = data->acpi_data.control_register.space_id;
744 	status = data->acpi_data.status_register.space_id;
745 
746 	if ((control != ACPI_ADR_SPACE_FIXED_HARDWARE) ||
747 	    (status != ACPI_ADR_SPACE_FIXED_HARDWARE)) {
748 		pr_debug("Invalid control/status registers (%llx - %llx)\n",
749 			control, status);
750 		goto err_out;
751 	}
752 
753 	/* fill in data->powernow_table */
754 	powernow_table = kzalloc((sizeof(*powernow_table)
755 		* (data->acpi_data.state_count + 1)), GFP_KERNEL);
756 	if (!powernow_table)
757 		goto err_out;
758 
759 	/* fill in data */
760 	data->numps = data->acpi_data.state_count;
761 	powernow_k8_acpi_pst_values(data, 0);
762 
763 	ret_val = fill_powernow_table_fidvid(data, powernow_table);
764 	if (ret_val)
765 		goto err_out_mem;
766 
767 	powernow_table[data->acpi_data.state_count].frequency =
768 		CPUFREQ_TABLE_END;
769 	data->powernow_table = powernow_table;
770 
771 	if (cpumask_first(topology_core_cpumask(data->cpu)) == data->cpu)
772 		print_basics(data);
773 
774 	/* notify BIOS that we exist */
775 	acpi_processor_notify_smm(THIS_MODULE);
776 
777 	if (!zalloc_cpumask_var(&data->acpi_data.shared_cpu_map, GFP_KERNEL)) {
778 		pr_err("unable to alloc powernow_k8_data cpumask\n");
779 		ret_val = -ENOMEM;
780 		goto err_out_mem;
781 	}
782 
783 	return 0;
784 
785 err_out_mem:
786 	kfree(powernow_table);
787 
788 err_out:
789 	acpi_processor_unregister_performance(data->cpu);
790 
791 	/* data->acpi_data.state_count informs us at ->exit()
792 	 * whether ACPI was used */
793 	data->acpi_data.state_count = 0;
794 
795 	return ret_val;
796 }
797 
798 static int fill_powernow_table_fidvid(struct powernow_k8_data *data,
799 		struct cpufreq_frequency_table *powernow_table)
800 {
801 	int i;
802 
803 	for (i = 0; i < data->acpi_data.state_count; i++) {
804 		u32 fid;
805 		u32 vid;
806 		u32 freq, index;
807 		u64 status, control;
808 
809 		if (data->exttype) {
810 			status =  data->acpi_data.states[i].status;
811 			fid = status & EXT_FID_MASK;
812 			vid = (status >> VID_SHIFT) & EXT_VID_MASK;
813 		} else {
814 			control =  data->acpi_data.states[i].control;
815 			fid = control & FID_MASK;
816 			vid = (control >> VID_SHIFT) & VID_MASK;
817 		}
818 
819 		pr_debug("   %d : fid 0x%x, vid 0x%x\n", i, fid, vid);
820 
821 		index = fid | (vid<<8);
822 		powernow_table[i].driver_data = index;
823 
824 		freq = find_khz_freq_from_fid(fid);
825 		powernow_table[i].frequency = freq;
826 
827 		/* verify frequency is OK */
828 		if ((freq > (MAX_FREQ * 1000)) || (freq < (MIN_FREQ * 1000))) {
829 			pr_debug("invalid freq %u kHz, ignoring\n", freq);
830 			invalidate_entry(powernow_table, i);
831 			continue;
832 		}
833 
834 		/* verify voltage is OK -
835 		 * BIOSs are using "off" to indicate invalid */
836 		if (vid == VID_OFF) {
837 			pr_debug("invalid vid %u, ignoring\n", vid);
838 			invalidate_entry(powernow_table, i);
839 			continue;
840 		}
841 
842 		if (freq != (data->acpi_data.states[i].core_frequency * 1000)) {
843 			pr_info("invalid freq entries %u kHz vs. %u kHz\n",
844 				freq, (unsigned int)
845 				(data->acpi_data.states[i].core_frequency
846 				 * 1000));
847 			invalidate_entry(powernow_table, i);
848 			continue;
849 		}
850 	}
851 	return 0;
852 }
853 
854 static void powernow_k8_cpu_exit_acpi(struct powernow_k8_data *data)
855 {
856 	if (data->acpi_data.state_count)
857 		acpi_processor_unregister_performance(data->cpu);
858 	free_cpumask_var(data->acpi_data.shared_cpu_map);
859 }
860 
861 static int get_transition_latency(struct powernow_k8_data *data)
862 {
863 	int max_latency = 0;
864 	int i;
865 	for (i = 0; i < data->acpi_data.state_count; i++) {
866 		int cur_latency = data->acpi_data.states[i].transition_latency
867 			+ data->acpi_data.states[i].bus_master_latency;
868 		if (cur_latency > max_latency)
869 			max_latency = cur_latency;
870 	}
871 	if (max_latency == 0) {
872 		pr_err(FW_WARN "Invalid zero transition latency\n");
873 		max_latency = 1;
874 	}
875 	/* value in usecs, needs to be in nanoseconds */
876 	return 1000 * max_latency;
877 }
878 
879 /* Take a frequency, and issue the fid/vid transition command */
880 static int transition_frequency_fidvid(struct powernow_k8_data *data,
881 		unsigned int index,
882 		struct cpufreq_policy *policy)
883 {
884 	u32 fid = 0;
885 	u32 vid = 0;
886 	int res;
887 	struct cpufreq_freqs freqs;
888 
889 	pr_debug("cpu %d transition to index %u\n", smp_processor_id(), index);
890 
891 	/* fid/vid correctness check for k8 */
892 	/* fid are the lower 8 bits of the index we stored into
893 	 * the cpufreq frequency table in find_psb_table, vid
894 	 * are the upper 8 bits.
895 	 */
896 	fid = data->powernow_table[index].driver_data & 0xFF;
897 	vid = (data->powernow_table[index].driver_data & 0xFF00) >> 8;
898 
899 	pr_debug("table matched fid 0x%x, giving vid 0x%x\n", fid, vid);
900 
901 	if (query_current_values_with_pending_wait(data))
902 		return 1;
903 
904 	if ((data->currvid == vid) && (data->currfid == fid)) {
905 		pr_debug("target matches current values (fid 0x%x, vid 0x%x)\n",
906 			fid, vid);
907 		return 0;
908 	}
909 
910 	pr_debug("cpu %d, changing to fid 0x%x, vid 0x%x\n",
911 		smp_processor_id(), fid, vid);
912 	freqs.old = find_khz_freq_from_fid(data->currfid);
913 	freqs.new = find_khz_freq_from_fid(fid);
914 
915 	cpufreq_freq_transition_begin(policy, &freqs);
916 	res = transition_fid_vid(data, fid, vid);
917 	cpufreq_freq_transition_end(policy, &freqs, res);
918 
919 	return res;
920 }
921 
922 struct powernowk8_target_arg {
923 	struct cpufreq_policy		*pol;
924 	unsigned			newstate;
925 };
926 
927 static long powernowk8_target_fn(void *arg)
928 {
929 	struct powernowk8_target_arg *pta = arg;
930 	struct cpufreq_policy *pol = pta->pol;
931 	unsigned newstate = pta->newstate;
932 	struct powernow_k8_data *data = per_cpu(powernow_data, pol->cpu);
933 	u32 checkfid;
934 	u32 checkvid;
935 	int ret;
936 
937 	if (!data)
938 		return -EINVAL;
939 
940 	checkfid = data->currfid;
941 	checkvid = data->currvid;
942 
943 	if (pending_bit_stuck()) {
944 		pr_err("failing targ, change pending bit set\n");
945 		return -EIO;
946 	}
947 
948 	pr_debug("targ: cpu %d, %d kHz, min %d, max %d\n",
949 		pol->cpu, data->powernow_table[newstate].frequency, pol->min,
950 		pol->max);
951 
952 	if (query_current_values_with_pending_wait(data))
953 		return -EIO;
954 
955 	pr_debug("targ: curr fid 0x%x, vid 0x%x\n",
956 		data->currfid, data->currvid);
957 
958 	if ((checkvid != data->currvid) ||
959 	    (checkfid != data->currfid)) {
960 		pr_info("error - out of sync, fix 0x%x 0x%x, vid 0x%x 0x%x\n",
961 		       checkfid, data->currfid,
962 		       checkvid, data->currvid);
963 	}
964 
965 	mutex_lock(&fidvid_mutex);
966 
967 	powernow_k8_acpi_pst_values(data, newstate);
968 
969 	ret = transition_frequency_fidvid(data, newstate, pol);
970 
971 	if (ret) {
972 		pr_err("transition frequency failed\n");
973 		mutex_unlock(&fidvid_mutex);
974 		return 1;
975 	}
976 	mutex_unlock(&fidvid_mutex);
977 
978 	pol->cur = find_khz_freq_from_fid(data->currfid);
979 
980 	return 0;
981 }
982 
983 /* Driver entry point to switch to the target frequency */
984 static int powernowk8_target(struct cpufreq_policy *pol, unsigned index)
985 {
986 	struct powernowk8_target_arg pta = { .pol = pol, .newstate = index };
987 
988 	return work_on_cpu(pol->cpu, powernowk8_target_fn, &pta);
989 }
990 
991 struct init_on_cpu {
992 	struct powernow_k8_data *data;
993 	int rc;
994 };
995 
996 static void powernowk8_cpu_init_on_cpu(void *_init_on_cpu)
997 {
998 	struct init_on_cpu *init_on_cpu = _init_on_cpu;
999 
1000 	if (pending_bit_stuck()) {
1001 		pr_err("failing init, change pending bit set\n");
1002 		init_on_cpu->rc = -ENODEV;
1003 		return;
1004 	}
1005 
1006 	if (query_current_values_with_pending_wait(init_on_cpu->data)) {
1007 		init_on_cpu->rc = -ENODEV;
1008 		return;
1009 	}
1010 
1011 	fidvid_msr_init();
1012 
1013 	init_on_cpu->rc = 0;
1014 }
1015 
1016 #define MISSING_PSS_MSG \
1017 	FW_BUG "No compatible ACPI _PSS objects found.\n" \
1018 	FW_BUG "First, make sure Cool'N'Quiet is enabled in the BIOS.\n" \
1019 	FW_BUG "If that doesn't help, try upgrading your BIOS.\n"
1020 
1021 /* per CPU init entry point to the driver */
1022 static int powernowk8_cpu_init(struct cpufreq_policy *pol)
1023 {
1024 	struct powernow_k8_data *data;
1025 	struct init_on_cpu init_on_cpu;
1026 	int rc, cpu;
1027 
1028 	smp_call_function_single(pol->cpu, check_supported_cpu, &rc, 1);
1029 	if (rc)
1030 		return -ENODEV;
1031 
1032 	data = kzalloc(sizeof(*data), GFP_KERNEL);
1033 	if (!data)
1034 		return -ENOMEM;
1035 
1036 	data->cpu = pol->cpu;
1037 
1038 	if (powernow_k8_cpu_init_acpi(data)) {
1039 		/*
1040 		 * Use the PSB BIOS structure. This is only available on
1041 		 * an UP version, and is deprecated by AMD.
1042 		 */
1043 		if (num_online_cpus() != 1) {
1044 			pr_err_once(MISSING_PSS_MSG);
1045 			goto err_out;
1046 		}
1047 		if (pol->cpu != 0) {
1048 			pr_err(FW_BUG "No ACPI _PSS objects for CPU other than CPU0. Complain to your BIOS vendor.\n");
1049 			goto err_out;
1050 		}
1051 		rc = find_psb_table(data);
1052 		if (rc)
1053 			goto err_out;
1054 
1055 		/* Take a crude guess here.
1056 		 * That guess was in microseconds, so multiply with 1000 */
1057 		pol->cpuinfo.transition_latency = (
1058 			 ((data->rvo + 8) * data->vstable * VST_UNITS_20US) +
1059 			 ((1 << data->irt) * 30)) * 1000;
1060 	} else /* ACPI _PSS objects available */
1061 		pol->cpuinfo.transition_latency = get_transition_latency(data);
1062 
1063 	/* only run on specific CPU from here on */
1064 	init_on_cpu.data = data;
1065 	smp_call_function_single(data->cpu, powernowk8_cpu_init_on_cpu,
1066 				 &init_on_cpu, 1);
1067 	rc = init_on_cpu.rc;
1068 	if (rc != 0)
1069 		goto err_out_exit_acpi;
1070 
1071 	cpumask_copy(pol->cpus, topology_core_cpumask(pol->cpu));
1072 	data->available_cores = pol->cpus;
1073 	pol->freq_table = data->powernow_table;
1074 
1075 	pr_debug("cpu_init done, current fid 0x%x, vid 0x%x\n",
1076 		data->currfid, data->currvid);
1077 
1078 	/* Point all the CPUs in this policy to the same data */
1079 	for_each_cpu(cpu, pol->cpus)
1080 		per_cpu(powernow_data, cpu) = data;
1081 
1082 	return 0;
1083 
1084 err_out_exit_acpi:
1085 	powernow_k8_cpu_exit_acpi(data);
1086 
1087 err_out:
1088 	kfree(data);
1089 	return -ENODEV;
1090 }
1091 
1092 static int powernowk8_cpu_exit(struct cpufreq_policy *pol)
1093 {
1094 	struct powernow_k8_data *data = per_cpu(powernow_data, pol->cpu);
1095 	int cpu;
1096 
1097 	if (!data)
1098 		return -EINVAL;
1099 
1100 	powernow_k8_cpu_exit_acpi(data);
1101 
1102 	kfree(data->powernow_table);
1103 	kfree(data);
1104 	for_each_cpu(cpu, pol->cpus)
1105 		per_cpu(powernow_data, cpu) = NULL;
1106 
1107 	return 0;
1108 }
1109 
1110 static void query_values_on_cpu(void *_err)
1111 {
1112 	int *err = _err;
1113 	struct powernow_k8_data *data = __this_cpu_read(powernow_data);
1114 
1115 	*err = query_current_values_with_pending_wait(data);
1116 }
1117 
1118 static unsigned int powernowk8_get(unsigned int cpu)
1119 {
1120 	struct powernow_k8_data *data = per_cpu(powernow_data, cpu);
1121 	unsigned int khz = 0;
1122 	int err;
1123 
1124 	if (!data)
1125 		return 0;
1126 
1127 	smp_call_function_single(cpu, query_values_on_cpu, &err, true);
1128 	if (err)
1129 		goto out;
1130 
1131 	khz = find_khz_freq_from_fid(data->currfid);
1132 
1133 
1134 out:
1135 	return khz;
1136 }
1137 
1138 static struct cpufreq_driver cpufreq_amd64_driver = {
1139 	.flags		= CPUFREQ_ASYNC_NOTIFICATION,
1140 	.verify		= cpufreq_generic_frequency_table_verify,
1141 	.target_index	= powernowk8_target,
1142 	.bios_limit	= acpi_processor_get_bios_limit,
1143 	.init		= powernowk8_cpu_init,
1144 	.exit		= powernowk8_cpu_exit,
1145 	.get		= powernowk8_get,
1146 	.name		= "powernow-k8",
1147 	.attr		= cpufreq_generic_attr,
1148 };
1149 
1150 static void __request_acpi_cpufreq(void)
1151 {
1152 	const char drv[] = "acpi-cpufreq";
1153 	const char *cur_drv;
1154 
1155 	cur_drv = cpufreq_get_current_driver();
1156 	if (!cur_drv)
1157 		goto request;
1158 
1159 	if (strncmp(cur_drv, drv, min_t(size_t, strlen(cur_drv), strlen(drv))))
1160 		pr_warn("WTF driver: %s\n", cur_drv);
1161 
1162 	return;
1163 
1164  request:
1165 	pr_warn("This CPU is not supported anymore, using acpi-cpufreq instead.\n");
1166 	request_module(drv);
1167 }
1168 
1169 /* driver entry point for init */
1170 static int powernowk8_init(void)
1171 {
1172 	unsigned int i, supported_cpus = 0;
1173 	int ret;
1174 
1175 	if (!x86_match_cpu(powernow_k8_ids))
1176 		return -ENODEV;
1177 
1178 	if (boot_cpu_has(X86_FEATURE_HW_PSTATE)) {
1179 		__request_acpi_cpufreq();
1180 		return -ENODEV;
1181 	}
1182 
1183 	cpus_read_lock();
1184 	for_each_online_cpu(i) {
1185 		smp_call_function_single(i, check_supported_cpu, &ret, 1);
1186 		if (!ret)
1187 			supported_cpus++;
1188 	}
1189 
1190 	if (supported_cpus != num_online_cpus()) {
1191 		cpus_read_unlock();
1192 		return -ENODEV;
1193 	}
1194 	cpus_read_unlock();
1195 
1196 	ret = cpufreq_register_driver(&cpufreq_amd64_driver);
1197 	if (ret)
1198 		return ret;
1199 
1200 	pr_info("Found %d %s (%d cpu cores) (" VERSION ")\n",
1201 		num_online_nodes(), boot_cpu_data.x86_model_id, supported_cpus);
1202 
1203 	return ret;
1204 }
1205 
1206 /* driver entry point for term */
1207 static void __exit powernowk8_exit(void)
1208 {
1209 	pr_debug("exit\n");
1210 
1211 	cpufreq_unregister_driver(&cpufreq_amd64_driver);
1212 }
1213 
1214 MODULE_AUTHOR("Paul Devriendt <paul.devriendt@amd.com>");
1215 MODULE_AUTHOR("Mark Langsdorf <mark.langsdorf@amd.com>");
1216 MODULE_DESCRIPTION("AMD Athlon 64 and Opteron processor frequency driver.");
1217 MODULE_LICENSE("GPL");
1218 
1219 late_initcall(powernowk8_init);
1220 module_exit(powernowk8_exit);
1221