xref: /linux/drivers/cpufreq/powernow-k8.c (revision 905e46acd3272d04566fec49afbd7ad9e2ed9ae3)
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 <herrmann.der.user@googlemail.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 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
31 
32 #include <linux/kernel.h>
33 #include <linux/smp.h>
34 #include <linux/module.h>
35 #include <linux/init.h>
36 #include <linux/cpufreq.h>
37 #include <linux/slab.h>
38 #include <linux/string.h>
39 #include <linux/cpumask.h>
40 #include <linux/io.h>
41 #include <linux/delay.h>
42 
43 #include <asm/msr.h>
44 #include <asm/cpu_device_id.h>
45 
46 #include <linux/acpi.h>
47 #include <linux/mutex.h>
48 #include <acpi/processor.h>
49 
50 #define VERSION "version 2.20.00"
51 #include "powernow-k8.h"
52 
53 /* serialize freq changes  */
54 static DEFINE_MUTEX(fidvid_mutex);
55 
56 static DEFINE_PER_CPU(struct powernow_k8_data *, powernow_data);
57 
58 static struct cpufreq_driver cpufreq_amd64_driver;
59 
60 /* Return a frequency in MHz, given an input fid */
61 static u32 find_freq_from_fid(u32 fid)
62 {
63 	return 800 + (fid * 100);
64 }
65 
66 /* Return a frequency in KHz, given an input fid */
67 static u32 find_khz_freq_from_fid(u32 fid)
68 {
69 	return 1000 * find_freq_from_fid(fid);
70 }
71 
72 /* Return the vco fid for an input fid
73  *
74  * Each "low" fid has corresponding "high" fid, and you can get to "low" fids
75  * only from corresponding high fids. This returns "high" fid corresponding to
76  * "low" one.
77  */
78 static u32 convert_fid_to_vco_fid(u32 fid)
79 {
80 	if (fid < HI_FID_TABLE_BOTTOM)
81 		return 8 + (2 * fid);
82 	else
83 		return fid;
84 }
85 
86 /*
87  * Return 1 if the pending bit is set. Unless we just instructed the processor
88  * to transition to a new state, seeing this bit set is really bad news.
89  */
90 static int pending_bit_stuck(void)
91 {
92 	u32 lo, hi;
93 
94 	rdmsr(MSR_FIDVID_STATUS, lo, hi);
95 	return lo & MSR_S_LO_CHANGE_PENDING ? 1 : 0;
96 }
97 
98 /*
99  * Update the global current fid / vid values from the status msr.
100  * Returns 1 on error.
101  */
102 static int query_current_values_with_pending_wait(struct powernow_k8_data *data)
103 {
104 	u32 lo, hi;
105 	u32 i = 0;
106 
107 	do {
108 		if (i++ > 10000) {
109 			pr_debug("detected change pending stuck\n");
110 			return 1;
111 		}
112 		rdmsr(MSR_FIDVID_STATUS, lo, hi);
113 	} while (lo & MSR_S_LO_CHANGE_PENDING);
114 
115 	data->currvid = hi & MSR_S_HI_CURRENT_VID;
116 	data->currfid = lo & MSR_S_LO_CURRENT_FID;
117 
118 	return 0;
119 }
120 
121 /* the isochronous relief time */
122 static void count_off_irt(struct powernow_k8_data *data)
123 {
124 	udelay((1 << data->irt) * 10);
125 	return;
126 }
127 
128 /* the voltage stabilization time */
129 static void count_off_vst(struct powernow_k8_data *data)
130 {
131 	udelay(data->vstable * VST_UNITS_20US);
132 	return;
133 }
134 
135 /* need to init the control msr to a safe value (for each cpu) */
136 static void fidvid_msr_init(void)
137 {
138 	u32 lo, hi;
139 	u8 fid, vid;
140 
141 	rdmsr(MSR_FIDVID_STATUS, lo, hi);
142 	vid = hi & MSR_S_HI_CURRENT_VID;
143 	fid = lo & MSR_S_LO_CURRENT_FID;
144 	lo = fid | (vid << MSR_C_LO_VID_SHIFT);
145 	hi = MSR_C_HI_STP_GNT_BENIGN;
146 	pr_debug("cpu%d, init lo 0x%x, hi 0x%x\n", smp_processor_id(), lo, hi);
147 	wrmsr(MSR_FIDVID_CTL, lo, hi);
148 }
149 
150 /* write the new fid value along with the other control fields to the msr */
151 static int write_new_fid(struct powernow_k8_data *data, u32 fid)
152 {
153 	u32 lo;
154 	u32 savevid = data->currvid;
155 	u32 i = 0;
156 
157 	if ((fid & INVALID_FID_MASK) || (data->currvid & INVALID_VID_MASK)) {
158 		pr_err("internal error - overflow on fid write\n");
159 		return 1;
160 	}
161 
162 	lo = fid;
163 	lo |= (data->currvid << MSR_C_LO_VID_SHIFT);
164 	lo |= MSR_C_LO_INIT_FID_VID;
165 
166 	pr_debug("writing fid 0x%x, lo 0x%x, hi 0x%x\n",
167 		fid, lo, data->plllock * PLL_LOCK_CONVERSION);
168 
169 	do {
170 		wrmsr(MSR_FIDVID_CTL, lo, data->plllock * PLL_LOCK_CONVERSION);
171 		if (i++ > 100) {
172 			pr_err("Hardware error - pending bit very stuck - no further pstate changes possible\n");
173 			return 1;
174 		}
175 	} while (query_current_values_with_pending_wait(data));
176 
177 	count_off_irt(data);
178 
179 	if (savevid != data->currvid) {
180 		pr_err("vid change on fid trans, old 0x%x, new 0x%x\n",
181 		       savevid, data->currvid);
182 		return 1;
183 	}
184 
185 	if (fid != data->currfid) {
186 		pr_err("fid trans failed, fid 0x%x, curr 0x%x\n", fid,
187 			data->currfid);
188 		return 1;
189 	}
190 
191 	return 0;
192 }
193 
194 /* Write a new vid to the hardware */
195 static int write_new_vid(struct powernow_k8_data *data, u32 vid)
196 {
197 	u32 lo;
198 	u32 savefid = data->currfid;
199 	int i = 0;
200 
201 	if ((data->currfid & INVALID_FID_MASK) || (vid & INVALID_VID_MASK)) {
202 		pr_err("internal error - overflow on vid write\n");
203 		return 1;
204 	}
205 
206 	lo = data->currfid;
207 	lo |= (vid << MSR_C_LO_VID_SHIFT);
208 	lo |= MSR_C_LO_INIT_FID_VID;
209 
210 	pr_debug("writing vid 0x%x, lo 0x%x, hi 0x%x\n",
211 		vid, lo, STOP_GRANT_5NS);
212 
213 	do {
214 		wrmsr(MSR_FIDVID_CTL, lo, STOP_GRANT_5NS);
215 		if (i++ > 100) {
216 			pr_err("internal error - pending bit very stuck - no further pstate changes possible\n");
217 			return 1;
218 		}
219 	} while (query_current_values_with_pending_wait(data));
220 
221 	if (savefid != data->currfid) {
222 		pr_err("fid changed on vid trans, old 0x%x new 0x%x\n",
223 			savefid, data->currfid);
224 		return 1;
225 	}
226 
227 	if (vid != data->currvid) {
228 		pr_err("vid trans failed, vid 0x%x, curr 0x%x\n",
229 				vid, data->currvid);
230 		return 1;
231 	}
232 
233 	return 0;
234 }
235 
236 /*
237  * Reduce the vid by the max of step or reqvid.
238  * Decreasing vid codes represent increasing voltages:
239  * vid of 0 is 1.550V, vid of 0x1e is 0.800V, vid of VID_OFF is off.
240  */
241 static int decrease_vid_code_by_step(struct powernow_k8_data *data,
242 		u32 reqvid, u32 step)
243 {
244 	if ((data->currvid - reqvid) > step)
245 		reqvid = data->currvid - step;
246 
247 	if (write_new_vid(data, reqvid))
248 		return 1;
249 
250 	count_off_vst(data);
251 
252 	return 0;
253 }
254 
255 /* Change Opteron/Athlon64 fid and vid, by the 3 phases. */
256 static int transition_fid_vid(struct powernow_k8_data *data,
257 		u32 reqfid, u32 reqvid)
258 {
259 	if (core_voltage_pre_transition(data, reqvid, reqfid))
260 		return 1;
261 
262 	if (core_frequency_transition(data, reqfid))
263 		return 1;
264 
265 	if (core_voltage_post_transition(data, reqvid))
266 		return 1;
267 
268 	if (query_current_values_with_pending_wait(data))
269 		return 1;
270 
271 	if ((reqfid != data->currfid) || (reqvid != data->currvid)) {
272 		pr_err("failed (cpu%d): req 0x%x 0x%x, curr 0x%x 0x%x\n",
273 				smp_processor_id(),
274 				reqfid, reqvid, data->currfid, data->currvid);
275 		return 1;
276 	}
277 
278 	pr_debug("transitioned (cpu%d): new fid 0x%x, vid 0x%x\n",
279 		smp_processor_id(), data->currfid, data->currvid);
280 
281 	return 0;
282 }
283 
284 /* Phase 1 - core voltage transition ... setup voltage */
285 static int core_voltage_pre_transition(struct powernow_k8_data *data,
286 		u32 reqvid, u32 reqfid)
287 {
288 	u32 rvosteps = data->rvo;
289 	u32 savefid = data->currfid;
290 	u32 maxvid, lo, rvomult = 1;
291 
292 	pr_debug("ph1 (cpu%d): start, currfid 0x%x, currvid 0x%x, reqvid 0x%x, rvo 0x%x\n",
293 		smp_processor_id(),
294 		data->currfid, data->currvid, reqvid, data->rvo);
295 
296 	if ((savefid < LO_FID_TABLE_TOP) && (reqfid < LO_FID_TABLE_TOP))
297 		rvomult = 2;
298 	rvosteps *= rvomult;
299 	rdmsr(MSR_FIDVID_STATUS, lo, maxvid);
300 	maxvid = 0x1f & (maxvid >> 16);
301 	pr_debug("ph1 maxvid=0x%x\n", maxvid);
302 	if (reqvid < maxvid) /* lower numbers are higher voltages */
303 		reqvid = maxvid;
304 
305 	while (data->currvid > reqvid) {
306 		pr_debug("ph1: curr 0x%x, req vid 0x%x\n",
307 			data->currvid, reqvid);
308 		if (decrease_vid_code_by_step(data, reqvid, data->vidmvs))
309 			return 1;
310 	}
311 
312 	while ((rvosteps > 0) &&
313 			((rvomult * data->rvo + data->currvid) > reqvid)) {
314 		if (data->currvid == maxvid) {
315 			rvosteps = 0;
316 		} else {
317 			pr_debug("ph1: changing vid for rvo, req 0x%x\n",
318 				data->currvid - 1);
319 			if (decrease_vid_code_by_step(data, data->currvid-1, 1))
320 				return 1;
321 			rvosteps--;
322 		}
323 	}
324 
325 	if (query_current_values_with_pending_wait(data))
326 		return 1;
327 
328 	if (savefid != data->currfid) {
329 		pr_err("ph1 err, currfid changed 0x%x\n", data->currfid);
330 		return 1;
331 	}
332 
333 	pr_debug("ph1 complete, currfid 0x%x, currvid 0x%x\n",
334 		data->currfid, data->currvid);
335 
336 	return 0;
337 }
338 
339 /* Phase 2 - core frequency transition */
340 static int core_frequency_transition(struct powernow_k8_data *data, u32 reqfid)
341 {
342 	u32 vcoreqfid, vcocurrfid, vcofiddiff;
343 	u32 fid_interval, savevid = data->currvid;
344 
345 	if (data->currfid == reqfid) {
346 		pr_err("ph2 null fid transition 0x%x\n", data->currfid);
347 		return 0;
348 	}
349 
350 	pr_debug("ph2 (cpu%d): starting, currfid 0x%x, currvid 0x%x, reqfid 0x%x\n",
351 		smp_processor_id(),
352 		data->currfid, data->currvid, reqfid);
353 
354 	vcoreqfid = convert_fid_to_vco_fid(reqfid);
355 	vcocurrfid = convert_fid_to_vco_fid(data->currfid);
356 	vcofiddiff = vcocurrfid > vcoreqfid ? vcocurrfid - vcoreqfid
357 	    : vcoreqfid - vcocurrfid;
358 
359 	if ((reqfid <= LO_FID_TABLE_TOP) && (data->currfid <= LO_FID_TABLE_TOP))
360 		vcofiddiff = 0;
361 
362 	while (vcofiddiff > 2) {
363 		(data->currfid & 1) ? (fid_interval = 1) : (fid_interval = 2);
364 
365 		if (reqfid > data->currfid) {
366 			if (data->currfid > LO_FID_TABLE_TOP) {
367 				if (write_new_fid(data,
368 						data->currfid + fid_interval))
369 					return 1;
370 			} else {
371 				if (write_new_fid
372 				    (data,
373 				     2 + convert_fid_to_vco_fid(data->currfid)))
374 					return 1;
375 			}
376 		} else {
377 			if (write_new_fid(data, data->currfid - fid_interval))
378 				return 1;
379 		}
380 
381 		vcocurrfid = convert_fid_to_vco_fid(data->currfid);
382 		vcofiddiff = vcocurrfid > vcoreqfid ? vcocurrfid - vcoreqfid
383 		    : vcoreqfid - vcocurrfid;
384 	}
385 
386 	if (write_new_fid(data, reqfid))
387 		return 1;
388 
389 	if (query_current_values_with_pending_wait(data))
390 		return 1;
391 
392 	if (data->currfid != reqfid) {
393 		pr_err("ph2: mismatch, failed fid transition, curr 0x%x, req 0x%x\n",
394 			data->currfid, reqfid);
395 		return 1;
396 	}
397 
398 	if (savevid != data->currvid) {
399 		pr_err("ph2: vid changed, save 0x%x, curr 0x%x\n",
400 			savevid, data->currvid);
401 		return 1;
402 	}
403 
404 	pr_debug("ph2 complete, currfid 0x%x, currvid 0x%x\n",
405 		data->currfid, data->currvid);
406 
407 	return 0;
408 }
409 
410 /* Phase 3 - core voltage transition flow ... jump to the final vid. */
411 static int core_voltage_post_transition(struct powernow_k8_data *data,
412 		u32 reqvid)
413 {
414 	u32 savefid = data->currfid;
415 	u32 savereqvid = reqvid;
416 
417 	pr_debug("ph3 (cpu%d): starting, currfid 0x%x, currvid 0x%x\n",
418 		smp_processor_id(),
419 		data->currfid, data->currvid);
420 
421 	if (reqvid != data->currvid) {
422 		if (write_new_vid(data, reqvid))
423 			return 1;
424 
425 		if (savefid != data->currfid) {
426 			pr_err("ph3: bad fid change, save 0x%x, curr 0x%x\n",
427 				savefid, data->currfid);
428 			return 1;
429 		}
430 
431 		if (data->currvid != reqvid) {
432 			pr_err("ph3: failed vid transition\n, req 0x%x, curr 0x%x",
433 				reqvid, data->currvid);
434 			return 1;
435 		}
436 	}
437 
438 	if (query_current_values_with_pending_wait(data))
439 		return 1;
440 
441 	if (savereqvid != data->currvid) {
442 		pr_debug("ph3 failed, currvid 0x%x\n", data->currvid);
443 		return 1;
444 	}
445 
446 	if (savefid != data->currfid) {
447 		pr_debug("ph3 failed, currfid changed 0x%x\n",
448 			data->currfid);
449 		return 1;
450 	}
451 
452 	pr_debug("ph3 complete, currfid 0x%x, currvid 0x%x\n",
453 		data->currfid, data->currvid);
454 
455 	return 0;
456 }
457 
458 static const struct x86_cpu_id powernow_k8_ids[] = {
459 	/* IO based frequency switching */
460 	{ X86_VENDOR_AMD, 0xf },
461 	{}
462 };
463 MODULE_DEVICE_TABLE(x86cpu, powernow_k8_ids);
464 
465 static void check_supported_cpu(void *_rc)
466 {
467 	u32 eax, ebx, ecx, edx;
468 	int *rc = _rc;
469 
470 	*rc = -ENODEV;
471 
472 	eax = cpuid_eax(CPUID_PROCESSOR_SIGNATURE);
473 
474 	if ((eax & CPUID_XFAM) == CPUID_XFAM_K8) {
475 		if (((eax & CPUID_USE_XFAM_XMOD) != CPUID_USE_XFAM_XMOD) ||
476 		    ((eax & CPUID_XMOD) > CPUID_XMOD_REV_MASK)) {
477 			pr_info("Processor cpuid %x not supported\n", eax);
478 			return;
479 		}
480 
481 		eax = cpuid_eax(CPUID_GET_MAX_CAPABILITIES);
482 		if (eax < CPUID_FREQ_VOLT_CAPABILITIES) {
483 			pr_info("No frequency change capabilities detected\n");
484 			return;
485 		}
486 
487 		cpuid(CPUID_FREQ_VOLT_CAPABILITIES, &eax, &ebx, &ecx, &edx);
488 		if ((edx & P_STATE_TRANSITION_CAPABLE)
489 			!= P_STATE_TRANSITION_CAPABLE) {
490 			pr_info("Power state transitions not supported\n");
491 			return;
492 		}
493 		*rc = 0;
494 	}
495 }
496 
497 static int check_pst_table(struct powernow_k8_data *data, struct pst_s *pst,
498 		u8 maxvid)
499 {
500 	unsigned int j;
501 	u8 lastfid = 0xff;
502 
503 	for (j = 0; j < data->numps; j++) {
504 		if (pst[j].vid > LEAST_VID) {
505 			pr_err(FW_BUG "vid %d invalid : 0x%x\n", j,
506 				pst[j].vid);
507 			return -EINVAL;
508 		}
509 		if (pst[j].vid < data->rvo) {
510 			/* vid + rvo >= 0 */
511 			pr_err(FW_BUG "0 vid exceeded with pstate %d\n", j);
512 			return -ENODEV;
513 		}
514 		if (pst[j].vid < maxvid + data->rvo) {
515 			/* vid + rvo >= maxvid */
516 			pr_err(FW_BUG "maxvid exceeded with pstate %d\n", j);
517 			return -ENODEV;
518 		}
519 		if (pst[j].fid > MAX_FID) {
520 			pr_err(FW_BUG "maxfid exceeded with pstate %d\n", j);
521 			return -ENODEV;
522 		}
523 		if (j && (pst[j].fid < HI_FID_TABLE_BOTTOM)) {
524 			/* Only first fid is allowed to be in "low" range */
525 			pr_err(FW_BUG "two low fids - %d : 0x%x\n", j,
526 				pst[j].fid);
527 			return -EINVAL;
528 		}
529 		if (pst[j].fid < lastfid)
530 			lastfid = pst[j].fid;
531 	}
532 	if (lastfid & 1) {
533 		pr_err(FW_BUG "lastfid invalid\n");
534 		return -EINVAL;
535 	}
536 	if (lastfid > LO_FID_TABLE_TOP)
537 		pr_info(FW_BUG "first fid not from lo freq table\n");
538 
539 	return 0;
540 }
541 
542 static void invalidate_entry(struct cpufreq_frequency_table *powernow_table,
543 		unsigned int entry)
544 {
545 	powernow_table[entry].frequency = CPUFREQ_ENTRY_INVALID;
546 }
547 
548 static void print_basics(struct powernow_k8_data *data)
549 {
550 	int j;
551 	for (j = 0; j < data->numps; j++) {
552 		if (data->powernow_table[j].frequency !=
553 				CPUFREQ_ENTRY_INVALID) {
554 			pr_info("fid 0x%x (%d MHz), vid 0x%x\n",
555 				data->powernow_table[j].driver_data & 0xff,
556 				data->powernow_table[j].frequency/1000,
557 				data->powernow_table[j].driver_data >> 8);
558 		}
559 	}
560 	if (data->batps)
561 		pr_info("Only %d pstates on battery\n", data->batps);
562 }
563 
564 static int fill_powernow_table(struct powernow_k8_data *data,
565 		struct pst_s *pst, u8 maxvid)
566 {
567 	struct cpufreq_frequency_table *powernow_table;
568 	unsigned int j;
569 
570 	if (data->batps) {
571 		/* use ACPI support to get full speed on mains power */
572 		pr_warn("Only %d pstates usable (use ACPI driver for full range\n",
573 			data->batps);
574 		data->numps = data->batps;
575 	}
576 
577 	for (j = 1; j < data->numps; j++) {
578 		if (pst[j-1].fid >= pst[j].fid) {
579 			pr_err("PST out of sequence\n");
580 			return -EINVAL;
581 		}
582 	}
583 
584 	if (data->numps < 2) {
585 		pr_err("no p states to transition\n");
586 		return -ENODEV;
587 	}
588 
589 	if (check_pst_table(data, pst, maxvid))
590 		return -EINVAL;
591 
592 	powernow_table = kzalloc((sizeof(*powernow_table)
593 		* (data->numps + 1)), GFP_KERNEL);
594 	if (!powernow_table) {
595 		pr_err("powernow_table memory alloc failure\n");
596 		return -ENOMEM;
597 	}
598 
599 	for (j = 0; j < data->numps; j++) {
600 		int freq;
601 		powernow_table[j].driver_data = pst[j].fid; /* lower 8 bits */
602 		powernow_table[j].driver_data |= (pst[j].vid << 8); /* upper 8 bits */
603 		freq = find_khz_freq_from_fid(pst[j].fid);
604 		powernow_table[j].frequency = freq;
605 	}
606 	powernow_table[data->numps].frequency = CPUFREQ_TABLE_END;
607 	powernow_table[data->numps].driver_data = 0;
608 
609 	if (query_current_values_with_pending_wait(data)) {
610 		kfree(powernow_table);
611 		return -EIO;
612 	}
613 
614 	pr_debug("cfid 0x%x, cvid 0x%x\n", data->currfid, data->currvid);
615 	data->powernow_table = powernow_table;
616 	if (cpumask_first(topology_core_cpumask(data->cpu)) == data->cpu)
617 		print_basics(data);
618 
619 	for (j = 0; j < data->numps; j++)
620 		if ((pst[j].fid == data->currfid) &&
621 		    (pst[j].vid == data->currvid))
622 			return 0;
623 
624 	pr_debug("currfid/vid do not match PST, ignoring\n");
625 	return 0;
626 }
627 
628 /* Find and validate the PSB/PST table in BIOS. */
629 static int find_psb_table(struct powernow_k8_data *data)
630 {
631 	struct psb_s *psb;
632 	unsigned int i;
633 	u32 mvs;
634 	u8 maxvid;
635 	u32 cpst = 0;
636 	u32 thiscpuid;
637 
638 	for (i = 0xc0000; i < 0xffff0; i += 0x10) {
639 		/* Scan BIOS looking for the signature. */
640 		/* It can not be at ffff0 - it is too big. */
641 
642 		psb = phys_to_virt(i);
643 		if (memcmp(psb, PSB_ID_STRING, PSB_ID_STRING_LEN) != 0)
644 			continue;
645 
646 		pr_debug("found PSB header at 0x%p\n", psb);
647 
648 		pr_debug("table vers: 0x%x\n", psb->tableversion);
649 		if (psb->tableversion != PSB_VERSION_1_4) {
650 			pr_err(FW_BUG "PSB table is not v1.4\n");
651 			return -ENODEV;
652 		}
653 
654 		pr_debug("flags: 0x%x\n", psb->flags1);
655 		if (psb->flags1) {
656 			pr_err(FW_BUG "unknown flags\n");
657 			return -ENODEV;
658 		}
659 
660 		data->vstable = psb->vstable;
661 		pr_debug("voltage stabilization time: %d(*20us)\n",
662 				data->vstable);
663 
664 		pr_debug("flags2: 0x%x\n", psb->flags2);
665 		data->rvo = psb->flags2 & 3;
666 		data->irt = ((psb->flags2) >> 2) & 3;
667 		mvs = ((psb->flags2) >> 4) & 3;
668 		data->vidmvs = 1 << mvs;
669 		data->batps = ((psb->flags2) >> 6) & 3;
670 
671 		pr_debug("ramp voltage offset: %d\n", data->rvo);
672 		pr_debug("isochronous relief time: %d\n", data->irt);
673 		pr_debug("maximum voltage step: %d - 0x%x\n", mvs, data->vidmvs);
674 
675 		pr_debug("numpst: 0x%x\n", psb->num_tables);
676 		cpst = psb->num_tables;
677 		if ((psb->cpuid == 0x00000fc0) ||
678 		    (psb->cpuid == 0x00000fe0)) {
679 			thiscpuid = cpuid_eax(CPUID_PROCESSOR_SIGNATURE);
680 			if ((thiscpuid == 0x00000fc0) ||
681 			    (thiscpuid == 0x00000fe0))
682 				cpst = 1;
683 		}
684 		if (cpst != 1) {
685 			pr_err(FW_BUG "numpst must be 1\n");
686 			return -ENODEV;
687 		}
688 
689 		data->plllock = psb->plllocktime;
690 		pr_debug("plllocktime: 0x%x (units 1us)\n", psb->plllocktime);
691 		pr_debug("maxfid: 0x%x\n", psb->maxfid);
692 		pr_debug("maxvid: 0x%x\n", psb->maxvid);
693 		maxvid = psb->maxvid;
694 
695 		data->numps = psb->numps;
696 		pr_debug("numpstates: 0x%x\n", data->numps);
697 		return fill_powernow_table(data,
698 				(struct pst_s *)(psb+1), maxvid);
699 	}
700 	/*
701 	 * If you see this message, complain to BIOS manufacturer. If
702 	 * he tells you "we do not support Linux" or some similar
703 	 * nonsense, remember that Windows 2000 uses the same legacy
704 	 * mechanism that the old Linux PSB driver uses. Tell them it
705 	 * is broken with Windows 2000.
706 	 *
707 	 * The reference to the AMD documentation is chapter 9 in the
708 	 * BIOS and Kernel Developer's Guide, which is available on
709 	 * www.amd.com
710 	 */
711 	pr_err(FW_BUG "No PSB or ACPI _PSS objects\n");
712 	pr_err("Make sure that your BIOS is up to date and Cool'N'Quiet support is enabled in BIOS setup\n");
713 	return -ENODEV;
714 }
715 
716 static void powernow_k8_acpi_pst_values(struct powernow_k8_data *data,
717 		unsigned int index)
718 {
719 	u64 control;
720 
721 	if (!data->acpi_data.state_count)
722 		return;
723 
724 	control = data->acpi_data.states[index].control;
725 	data->irt = (control >> IRT_SHIFT) & IRT_MASK;
726 	data->rvo = (control >> RVO_SHIFT) & RVO_MASK;
727 	data->exttype = (control >> EXT_TYPE_SHIFT) & EXT_TYPE_MASK;
728 	data->plllock = (control >> PLL_L_SHIFT) & PLL_L_MASK;
729 	data->vidmvs = 1 << ((control >> MVS_SHIFT) & MVS_MASK);
730 	data->vstable = (control >> VST_SHIFT) & VST_MASK;
731 }
732 
733 static int powernow_k8_cpu_init_acpi(struct powernow_k8_data *data)
734 {
735 	struct cpufreq_frequency_table *powernow_table;
736 	int ret_val = -ENODEV;
737 	u64 control, status;
738 
739 	if (acpi_processor_register_performance(&data->acpi_data, data->cpu)) {
740 		pr_debug("register performance failed: bad ACPI data\n");
741 		return -EIO;
742 	}
743 
744 	/* verify the data contained in the ACPI structures */
745 	if (data->acpi_data.state_count <= 1) {
746 		pr_debug("No ACPI P-States\n");
747 		goto err_out;
748 	}
749 
750 	control = data->acpi_data.control_register.space_id;
751 	status = data->acpi_data.status_register.space_id;
752 
753 	if ((control != ACPI_ADR_SPACE_FIXED_HARDWARE) ||
754 	    (status != ACPI_ADR_SPACE_FIXED_HARDWARE)) {
755 		pr_debug("Invalid control/status registers (%llx - %llx)\n",
756 			control, status);
757 		goto err_out;
758 	}
759 
760 	/* fill in data->powernow_table */
761 	powernow_table = kzalloc((sizeof(*powernow_table)
762 		* (data->acpi_data.state_count + 1)), GFP_KERNEL);
763 	if (!powernow_table) {
764 		pr_debug("powernow_table memory alloc failure\n");
765 		goto err_out;
766 	}
767 
768 	/* fill in data */
769 	data->numps = data->acpi_data.state_count;
770 	powernow_k8_acpi_pst_values(data, 0);
771 
772 	ret_val = fill_powernow_table_fidvid(data, powernow_table);
773 	if (ret_val)
774 		goto err_out_mem;
775 
776 	powernow_table[data->acpi_data.state_count].frequency =
777 		CPUFREQ_TABLE_END;
778 	data->powernow_table = powernow_table;
779 
780 	if (cpumask_first(topology_core_cpumask(data->cpu)) == data->cpu)
781 		print_basics(data);
782 
783 	/* notify BIOS that we exist */
784 	acpi_processor_notify_smm(THIS_MODULE);
785 
786 	if (!zalloc_cpumask_var(&data->acpi_data.shared_cpu_map, GFP_KERNEL)) {
787 		pr_err("unable to alloc powernow_k8_data cpumask\n");
788 		ret_val = -ENOMEM;
789 		goto err_out_mem;
790 	}
791 
792 	return 0;
793 
794 err_out_mem:
795 	kfree(powernow_table);
796 
797 err_out:
798 	acpi_processor_unregister_performance(data->cpu);
799 
800 	/* data->acpi_data.state_count informs us at ->exit()
801 	 * whether ACPI was used */
802 	data->acpi_data.state_count = 0;
803 
804 	return ret_val;
805 }
806 
807 static int fill_powernow_table_fidvid(struct powernow_k8_data *data,
808 		struct cpufreq_frequency_table *powernow_table)
809 {
810 	int i;
811 
812 	for (i = 0; i < data->acpi_data.state_count; i++) {
813 		u32 fid;
814 		u32 vid;
815 		u32 freq, index;
816 		u64 status, control;
817 
818 		if (data->exttype) {
819 			status =  data->acpi_data.states[i].status;
820 			fid = status & EXT_FID_MASK;
821 			vid = (status >> VID_SHIFT) & EXT_VID_MASK;
822 		} else {
823 			control =  data->acpi_data.states[i].control;
824 			fid = control & FID_MASK;
825 			vid = (control >> VID_SHIFT) & VID_MASK;
826 		}
827 
828 		pr_debug("   %d : fid 0x%x, vid 0x%x\n", i, fid, vid);
829 
830 		index = fid | (vid<<8);
831 		powernow_table[i].driver_data = index;
832 
833 		freq = find_khz_freq_from_fid(fid);
834 		powernow_table[i].frequency = freq;
835 
836 		/* verify frequency is OK */
837 		if ((freq > (MAX_FREQ * 1000)) || (freq < (MIN_FREQ * 1000))) {
838 			pr_debug("invalid freq %u kHz, ignoring\n", freq);
839 			invalidate_entry(powernow_table, i);
840 			continue;
841 		}
842 
843 		/* verify voltage is OK -
844 		 * BIOSs are using "off" to indicate invalid */
845 		if (vid == VID_OFF) {
846 			pr_debug("invalid vid %u, ignoring\n", vid);
847 			invalidate_entry(powernow_table, i);
848 			continue;
849 		}
850 
851 		if (freq != (data->acpi_data.states[i].core_frequency * 1000)) {
852 			pr_info("invalid freq entries %u kHz vs. %u kHz\n",
853 				freq, (unsigned int)
854 				(data->acpi_data.states[i].core_frequency
855 				 * 1000));
856 			invalidate_entry(powernow_table, i);
857 			continue;
858 		}
859 	}
860 	return 0;
861 }
862 
863 static void powernow_k8_cpu_exit_acpi(struct powernow_k8_data *data)
864 {
865 	if (data->acpi_data.state_count)
866 		acpi_processor_unregister_performance(data->cpu);
867 	free_cpumask_var(data->acpi_data.shared_cpu_map);
868 }
869 
870 static int get_transition_latency(struct powernow_k8_data *data)
871 {
872 	int max_latency = 0;
873 	int i;
874 	for (i = 0; i < data->acpi_data.state_count; i++) {
875 		int cur_latency = data->acpi_data.states[i].transition_latency
876 			+ data->acpi_data.states[i].bus_master_latency;
877 		if (cur_latency > max_latency)
878 			max_latency = cur_latency;
879 	}
880 	if (max_latency == 0) {
881 		pr_err(FW_WARN "Invalid zero transition latency\n");
882 		max_latency = 1;
883 	}
884 	/* value in usecs, needs to be in nanoseconds */
885 	return 1000 * max_latency;
886 }
887 
888 /* Take a frequency, and issue the fid/vid transition command */
889 static int transition_frequency_fidvid(struct powernow_k8_data *data,
890 		unsigned int index)
891 {
892 	struct cpufreq_policy *policy;
893 	u32 fid = 0;
894 	u32 vid = 0;
895 	int res;
896 	struct cpufreq_freqs freqs;
897 
898 	pr_debug("cpu %d transition to index %u\n", smp_processor_id(), index);
899 
900 	/* fid/vid correctness check for k8 */
901 	/* fid are the lower 8 bits of the index we stored into
902 	 * the cpufreq frequency table in find_psb_table, vid
903 	 * are the upper 8 bits.
904 	 */
905 	fid = data->powernow_table[index].driver_data & 0xFF;
906 	vid = (data->powernow_table[index].driver_data & 0xFF00) >> 8;
907 
908 	pr_debug("table matched fid 0x%x, giving vid 0x%x\n", fid, vid);
909 
910 	if (query_current_values_with_pending_wait(data))
911 		return 1;
912 
913 	if ((data->currvid == vid) && (data->currfid == fid)) {
914 		pr_debug("target matches current values (fid 0x%x, vid 0x%x)\n",
915 			fid, vid);
916 		return 0;
917 	}
918 
919 	pr_debug("cpu %d, changing to fid 0x%x, vid 0x%x\n",
920 		smp_processor_id(), fid, vid);
921 	freqs.old = find_khz_freq_from_fid(data->currfid);
922 	freqs.new = find_khz_freq_from_fid(fid);
923 
924 	policy = cpufreq_cpu_get(smp_processor_id());
925 	cpufreq_cpu_put(policy);
926 
927 	cpufreq_freq_transition_begin(policy, &freqs);
928 	res = transition_fid_vid(data, fid, vid);
929 	cpufreq_freq_transition_end(policy, &freqs, res);
930 
931 	return res;
932 }
933 
934 struct powernowk8_target_arg {
935 	struct cpufreq_policy		*pol;
936 	unsigned			newstate;
937 };
938 
939 static long powernowk8_target_fn(void *arg)
940 {
941 	struct powernowk8_target_arg *pta = arg;
942 	struct cpufreq_policy *pol = pta->pol;
943 	unsigned newstate = pta->newstate;
944 	struct powernow_k8_data *data = per_cpu(powernow_data, pol->cpu);
945 	u32 checkfid;
946 	u32 checkvid;
947 	int ret;
948 
949 	if (!data)
950 		return -EINVAL;
951 
952 	checkfid = data->currfid;
953 	checkvid = data->currvid;
954 
955 	if (pending_bit_stuck()) {
956 		pr_err("failing targ, change pending bit set\n");
957 		return -EIO;
958 	}
959 
960 	pr_debug("targ: cpu %d, %d kHz, min %d, max %d\n",
961 		pol->cpu, data->powernow_table[newstate].frequency, pol->min,
962 		pol->max);
963 
964 	if (query_current_values_with_pending_wait(data))
965 		return -EIO;
966 
967 	pr_debug("targ: curr fid 0x%x, vid 0x%x\n",
968 		data->currfid, data->currvid);
969 
970 	if ((checkvid != data->currvid) ||
971 	    (checkfid != data->currfid)) {
972 		pr_info("error - out of sync, fix 0x%x 0x%x, vid 0x%x 0x%x\n",
973 		       checkfid, data->currfid,
974 		       checkvid, data->currvid);
975 	}
976 
977 	mutex_lock(&fidvid_mutex);
978 
979 	powernow_k8_acpi_pst_values(data, newstate);
980 
981 	ret = transition_frequency_fidvid(data, newstate);
982 
983 	if (ret) {
984 		pr_err("transition frequency failed\n");
985 		mutex_unlock(&fidvid_mutex);
986 		return 1;
987 	}
988 	mutex_unlock(&fidvid_mutex);
989 
990 	pol->cur = find_khz_freq_from_fid(data->currfid);
991 
992 	return 0;
993 }
994 
995 /* Driver entry point to switch to the target frequency */
996 static int powernowk8_target(struct cpufreq_policy *pol, unsigned index)
997 {
998 	struct powernowk8_target_arg pta = { .pol = pol, .newstate = index };
999 
1000 	return work_on_cpu(pol->cpu, powernowk8_target_fn, &pta);
1001 }
1002 
1003 struct init_on_cpu {
1004 	struct powernow_k8_data *data;
1005 	int rc;
1006 };
1007 
1008 static void powernowk8_cpu_init_on_cpu(void *_init_on_cpu)
1009 {
1010 	struct init_on_cpu *init_on_cpu = _init_on_cpu;
1011 
1012 	if (pending_bit_stuck()) {
1013 		pr_err("failing init, change pending bit set\n");
1014 		init_on_cpu->rc = -ENODEV;
1015 		return;
1016 	}
1017 
1018 	if (query_current_values_with_pending_wait(init_on_cpu->data)) {
1019 		init_on_cpu->rc = -ENODEV;
1020 		return;
1021 	}
1022 
1023 	fidvid_msr_init();
1024 
1025 	init_on_cpu->rc = 0;
1026 }
1027 
1028 #define MISSING_PSS_MSG \
1029 	FW_BUG "No compatible ACPI _PSS objects found.\n" \
1030 	FW_BUG "First, make sure Cool'N'Quiet is enabled in the BIOS.\n" \
1031 	FW_BUG "If that doesn't help, try upgrading your BIOS.\n"
1032 
1033 /* per CPU init entry point to the driver */
1034 static int powernowk8_cpu_init(struct cpufreq_policy *pol)
1035 {
1036 	struct powernow_k8_data *data;
1037 	struct init_on_cpu init_on_cpu;
1038 	int rc, cpu;
1039 
1040 	smp_call_function_single(pol->cpu, check_supported_cpu, &rc, 1);
1041 	if (rc)
1042 		return -ENODEV;
1043 
1044 	data = kzalloc(sizeof(*data), GFP_KERNEL);
1045 	if (!data) {
1046 		pr_err("unable to alloc powernow_k8_data");
1047 		return -ENOMEM;
1048 	}
1049 
1050 	data->cpu = pol->cpu;
1051 
1052 	if (powernow_k8_cpu_init_acpi(data)) {
1053 		/*
1054 		 * Use the PSB BIOS structure. This is only available on
1055 		 * an UP version, and is deprecated by AMD.
1056 		 */
1057 		if (num_online_cpus() != 1) {
1058 			pr_err_once(MISSING_PSS_MSG);
1059 			goto err_out;
1060 		}
1061 		if (pol->cpu != 0) {
1062 			pr_err(FW_BUG "No ACPI _PSS objects for CPU other than CPU0. Complain to your BIOS vendor.\n");
1063 			goto err_out;
1064 		}
1065 		rc = find_psb_table(data);
1066 		if (rc)
1067 			goto err_out;
1068 
1069 		/* Take a crude guess here.
1070 		 * That guess was in microseconds, so multiply with 1000 */
1071 		pol->cpuinfo.transition_latency = (
1072 			 ((data->rvo + 8) * data->vstable * VST_UNITS_20US) +
1073 			 ((1 << data->irt) * 30)) * 1000;
1074 	} else /* ACPI _PSS objects available */
1075 		pol->cpuinfo.transition_latency = get_transition_latency(data);
1076 
1077 	/* only run on specific CPU from here on */
1078 	init_on_cpu.data = data;
1079 	smp_call_function_single(data->cpu, powernowk8_cpu_init_on_cpu,
1080 				 &init_on_cpu, 1);
1081 	rc = init_on_cpu.rc;
1082 	if (rc != 0)
1083 		goto err_out_exit_acpi;
1084 
1085 	cpumask_copy(pol->cpus, topology_core_cpumask(pol->cpu));
1086 	data->available_cores = pol->cpus;
1087 
1088 	/* min/max the cpu is capable of */
1089 	if (cpufreq_table_validate_and_show(pol, data->powernow_table)) {
1090 		pr_err(FW_BUG "invalid powernow_table\n");
1091 		powernow_k8_cpu_exit_acpi(data);
1092 		kfree(data->powernow_table);
1093 		kfree(data);
1094 		return -EINVAL;
1095 	}
1096 
1097 	pr_debug("cpu_init done, current fid 0x%x, vid 0x%x\n",
1098 		data->currfid, data->currvid);
1099 
1100 	/* Point all the CPUs in this policy to the same data */
1101 	for_each_cpu(cpu, pol->cpus)
1102 		per_cpu(powernow_data, cpu) = data;
1103 
1104 	return 0;
1105 
1106 err_out_exit_acpi:
1107 	powernow_k8_cpu_exit_acpi(data);
1108 
1109 err_out:
1110 	kfree(data);
1111 	return -ENODEV;
1112 }
1113 
1114 static int powernowk8_cpu_exit(struct cpufreq_policy *pol)
1115 {
1116 	struct powernow_k8_data *data = per_cpu(powernow_data, pol->cpu);
1117 	int cpu;
1118 
1119 	if (!data)
1120 		return -EINVAL;
1121 
1122 	powernow_k8_cpu_exit_acpi(data);
1123 
1124 	kfree(data->powernow_table);
1125 	kfree(data);
1126 	for_each_cpu(cpu, pol->cpus)
1127 		per_cpu(powernow_data, cpu) = NULL;
1128 
1129 	return 0;
1130 }
1131 
1132 static void query_values_on_cpu(void *_err)
1133 {
1134 	int *err = _err;
1135 	struct powernow_k8_data *data = __this_cpu_read(powernow_data);
1136 
1137 	*err = query_current_values_with_pending_wait(data);
1138 }
1139 
1140 static unsigned int powernowk8_get(unsigned int cpu)
1141 {
1142 	struct powernow_k8_data *data = per_cpu(powernow_data, cpu);
1143 	unsigned int khz = 0;
1144 	int err;
1145 
1146 	if (!data)
1147 		return 0;
1148 
1149 	smp_call_function_single(cpu, query_values_on_cpu, &err, true);
1150 	if (err)
1151 		goto out;
1152 
1153 	khz = find_khz_freq_from_fid(data->currfid);
1154 
1155 
1156 out:
1157 	return khz;
1158 }
1159 
1160 static struct cpufreq_driver cpufreq_amd64_driver = {
1161 	.flags		= CPUFREQ_ASYNC_NOTIFICATION,
1162 	.verify		= cpufreq_generic_frequency_table_verify,
1163 	.target_index	= powernowk8_target,
1164 	.bios_limit	= acpi_processor_get_bios_limit,
1165 	.init		= powernowk8_cpu_init,
1166 	.exit		= powernowk8_cpu_exit,
1167 	.get		= powernowk8_get,
1168 	.name		= "powernow-k8",
1169 	.attr		= cpufreq_generic_attr,
1170 };
1171 
1172 static void __request_acpi_cpufreq(void)
1173 {
1174 	const char drv[] = "acpi-cpufreq";
1175 	const char *cur_drv;
1176 
1177 	cur_drv = cpufreq_get_current_driver();
1178 	if (!cur_drv)
1179 		goto request;
1180 
1181 	if (strncmp(cur_drv, drv, min_t(size_t, strlen(cur_drv), strlen(drv))))
1182 		pr_warn("WTF driver: %s\n", cur_drv);
1183 
1184 	return;
1185 
1186  request:
1187 	pr_warn("This CPU is not supported anymore, using acpi-cpufreq instead.\n");
1188 	request_module(drv);
1189 }
1190 
1191 /* driver entry point for init */
1192 static int powernowk8_init(void)
1193 {
1194 	unsigned int i, supported_cpus = 0;
1195 	int ret;
1196 
1197 	if (static_cpu_has(X86_FEATURE_HW_PSTATE)) {
1198 		__request_acpi_cpufreq();
1199 		return -ENODEV;
1200 	}
1201 
1202 	if (!x86_match_cpu(powernow_k8_ids))
1203 		return -ENODEV;
1204 
1205 	get_online_cpus();
1206 	for_each_online_cpu(i) {
1207 		smp_call_function_single(i, check_supported_cpu, &ret, 1);
1208 		if (!ret)
1209 			supported_cpus++;
1210 	}
1211 
1212 	if (supported_cpus != num_online_cpus()) {
1213 		put_online_cpus();
1214 		return -ENODEV;
1215 	}
1216 	put_online_cpus();
1217 
1218 	ret = cpufreq_register_driver(&cpufreq_amd64_driver);
1219 	if (ret)
1220 		return ret;
1221 
1222 	pr_info("Found %d %s (%d cpu cores) (" VERSION ")\n",
1223 		num_online_nodes(), boot_cpu_data.x86_model_id, supported_cpus);
1224 
1225 	return ret;
1226 }
1227 
1228 /* driver entry point for term */
1229 static void __exit powernowk8_exit(void)
1230 {
1231 	pr_debug("exit\n");
1232 
1233 	cpufreq_unregister_driver(&cpufreq_amd64_driver);
1234 }
1235 
1236 MODULE_AUTHOR("Paul Devriendt <paul.devriendt@amd.com>");
1237 MODULE_AUTHOR("Mark Langsdorf <mark.langsdorf@amd.com>");
1238 MODULE_DESCRIPTION("AMD Athlon 64 and Opteron processor frequency driver.");
1239 MODULE_LICENSE("GPL");
1240 
1241 late_initcall(powernowk8_init);
1242 module_exit(powernowk8_exit);
1243