xref: /linux/drivers/cpufreq/powernow-k8.c (revision c4ee0af3fa0dc65f690fc908f02b8355f9576ea0)
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 #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/io.h>
39 #include <linux/delay.h>
40 
41 #include <asm/msr.h>
42 #include <asm/cpu_device_id.h>
43 
44 #include <linux/acpi.h>
45 #include <linux/mutex.h>
46 #include <acpi/processor.h>
47 
48 #define PFX "powernow-k8: "
49 #define VERSION "version 2.20.00"
50 #include "powernow-k8.h"
51 
52 /* serialize freq changes  */
53 static DEFINE_MUTEX(fidvid_mutex);
54 
55 static DEFINE_PER_CPU(struct powernow_k8_data *, powernow_data);
56 
57 static struct cpufreq_driver cpufreq_amd64_driver;
58 
59 #ifndef CONFIG_SMP
60 static inline const struct cpumask *cpu_core_mask(int cpu)
61 {
62 	return cpumask_of(0);
63 }
64 #endif
65 
66 /* Return a frequency in MHz, given an input fid */
67 static u32 find_freq_from_fid(u32 fid)
68 {
69 	return 800 + (fid * 100);
70 }
71 
72 /* Return a frequency in KHz, given an input fid */
73 static u32 find_khz_freq_from_fid(u32 fid)
74 {
75 	return 1000 * find_freq_from_fid(fid);
76 }
77 
78 /* Return the vco fid for an input fid
79  *
80  * Each "low" fid has corresponding "high" fid, and you can get to "low" fids
81  * only from corresponding high fids. This returns "high" fid corresponding to
82  * "low" one.
83  */
84 static u32 convert_fid_to_vco_fid(u32 fid)
85 {
86 	if (fid < HI_FID_TABLE_BOTTOM)
87 		return 8 + (2 * fid);
88 	else
89 		return fid;
90 }
91 
92 /*
93  * Return 1 if the pending bit is set. Unless we just instructed the processor
94  * to transition to a new state, seeing this bit set is really bad news.
95  */
96 static int pending_bit_stuck(void)
97 {
98 	u32 lo, hi;
99 
100 	rdmsr(MSR_FIDVID_STATUS, lo, hi);
101 	return lo & MSR_S_LO_CHANGE_PENDING ? 1 : 0;
102 }
103 
104 /*
105  * Update the global current fid / vid values from the status msr.
106  * Returns 1 on error.
107  */
108 static int query_current_values_with_pending_wait(struct powernow_k8_data *data)
109 {
110 	u32 lo, hi;
111 	u32 i = 0;
112 
113 	do {
114 		if (i++ > 10000) {
115 			pr_debug("detected change pending stuck\n");
116 			return 1;
117 		}
118 		rdmsr(MSR_FIDVID_STATUS, lo, hi);
119 	} while (lo & MSR_S_LO_CHANGE_PENDING);
120 
121 	data->currvid = hi & MSR_S_HI_CURRENT_VID;
122 	data->currfid = lo & MSR_S_LO_CURRENT_FID;
123 
124 	return 0;
125 }
126 
127 /* the isochronous relief time */
128 static void count_off_irt(struct powernow_k8_data *data)
129 {
130 	udelay((1 << data->irt) * 10);
131 	return;
132 }
133 
134 /* the voltage stabilization time */
135 static void count_off_vst(struct powernow_k8_data *data)
136 {
137 	udelay(data->vstable * VST_UNITS_20US);
138 	return;
139 }
140 
141 /* need to init the control msr to a safe value (for each cpu) */
142 static void fidvid_msr_init(void)
143 {
144 	u32 lo, hi;
145 	u8 fid, vid;
146 
147 	rdmsr(MSR_FIDVID_STATUS, lo, hi);
148 	vid = hi & MSR_S_HI_CURRENT_VID;
149 	fid = lo & MSR_S_LO_CURRENT_FID;
150 	lo = fid | (vid << MSR_C_LO_VID_SHIFT);
151 	hi = MSR_C_HI_STP_GNT_BENIGN;
152 	pr_debug("cpu%d, init lo 0x%x, hi 0x%x\n", smp_processor_id(), lo, hi);
153 	wrmsr(MSR_FIDVID_CTL, lo, hi);
154 }
155 
156 /* write the new fid value along with the other control fields to the msr */
157 static int write_new_fid(struct powernow_k8_data *data, u32 fid)
158 {
159 	u32 lo;
160 	u32 savevid = data->currvid;
161 	u32 i = 0;
162 
163 	if ((fid & INVALID_FID_MASK) || (data->currvid & INVALID_VID_MASK)) {
164 		printk(KERN_ERR PFX "internal error - overflow on fid write\n");
165 		return 1;
166 	}
167 
168 	lo = fid;
169 	lo |= (data->currvid << MSR_C_LO_VID_SHIFT);
170 	lo |= MSR_C_LO_INIT_FID_VID;
171 
172 	pr_debug("writing fid 0x%x, lo 0x%x, hi 0x%x\n",
173 		fid, lo, data->plllock * PLL_LOCK_CONVERSION);
174 
175 	do {
176 		wrmsr(MSR_FIDVID_CTL, lo, data->plllock * PLL_LOCK_CONVERSION);
177 		if (i++ > 100) {
178 			printk(KERN_ERR PFX
179 				"Hardware error - pending bit very stuck - "
180 				"no further pstate changes possible\n");
181 			return 1;
182 		}
183 	} while (query_current_values_with_pending_wait(data));
184 
185 	count_off_irt(data);
186 
187 	if (savevid != data->currvid) {
188 		printk(KERN_ERR PFX
189 			"vid change on fid trans, old 0x%x, new 0x%x\n",
190 			savevid, data->currvid);
191 		return 1;
192 	}
193 
194 	if (fid != data->currfid) {
195 		printk(KERN_ERR PFX
196 			"fid trans failed, fid 0x%x, curr 0x%x\n", fid,
197 			data->currfid);
198 		return 1;
199 	}
200 
201 	return 0;
202 }
203 
204 /* Write a new vid to the hardware */
205 static int write_new_vid(struct powernow_k8_data *data, u32 vid)
206 {
207 	u32 lo;
208 	u32 savefid = data->currfid;
209 	int i = 0;
210 
211 	if ((data->currfid & INVALID_FID_MASK) || (vid & INVALID_VID_MASK)) {
212 		printk(KERN_ERR PFX "internal error - overflow on vid write\n");
213 		return 1;
214 	}
215 
216 	lo = data->currfid;
217 	lo |= (vid << MSR_C_LO_VID_SHIFT);
218 	lo |= MSR_C_LO_INIT_FID_VID;
219 
220 	pr_debug("writing vid 0x%x, lo 0x%x, hi 0x%x\n",
221 		vid, lo, STOP_GRANT_5NS);
222 
223 	do {
224 		wrmsr(MSR_FIDVID_CTL, lo, STOP_GRANT_5NS);
225 		if (i++ > 100) {
226 			printk(KERN_ERR PFX "internal error - pending bit "
227 					"very stuck - no further pstate "
228 					"changes possible\n");
229 			return 1;
230 		}
231 	} while (query_current_values_with_pending_wait(data));
232 
233 	if (savefid != data->currfid) {
234 		printk(KERN_ERR PFX "fid changed on vid trans, old "
235 			"0x%x new 0x%x\n",
236 		       savefid, data->currfid);
237 		return 1;
238 	}
239 
240 	if (vid != data->currvid) {
241 		printk(KERN_ERR PFX "vid trans failed, vid 0x%x, "
242 				"curr 0x%x\n",
243 				vid, data->currvid);
244 		return 1;
245 	}
246 
247 	return 0;
248 }
249 
250 /*
251  * Reduce the vid by the max of step or reqvid.
252  * Decreasing vid codes represent increasing voltages:
253  * vid of 0 is 1.550V, vid of 0x1e is 0.800V, vid of VID_OFF is off.
254  */
255 static int decrease_vid_code_by_step(struct powernow_k8_data *data,
256 		u32 reqvid, u32 step)
257 {
258 	if ((data->currvid - reqvid) > step)
259 		reqvid = data->currvid - step;
260 
261 	if (write_new_vid(data, reqvid))
262 		return 1;
263 
264 	count_off_vst(data);
265 
266 	return 0;
267 }
268 
269 /* Change Opteron/Athlon64 fid and vid, by the 3 phases. */
270 static int transition_fid_vid(struct powernow_k8_data *data,
271 		u32 reqfid, u32 reqvid)
272 {
273 	if (core_voltage_pre_transition(data, reqvid, reqfid))
274 		return 1;
275 
276 	if (core_frequency_transition(data, reqfid))
277 		return 1;
278 
279 	if (core_voltage_post_transition(data, reqvid))
280 		return 1;
281 
282 	if (query_current_values_with_pending_wait(data))
283 		return 1;
284 
285 	if ((reqfid != data->currfid) || (reqvid != data->currvid)) {
286 		printk(KERN_ERR PFX "failed (cpu%d): req 0x%x 0x%x, "
287 				"curr 0x%x 0x%x\n",
288 				smp_processor_id(),
289 				reqfid, reqvid, data->currfid, data->currvid);
290 		return 1;
291 	}
292 
293 	pr_debug("transitioned (cpu%d): new fid 0x%x, vid 0x%x\n",
294 		smp_processor_id(), data->currfid, data->currvid);
295 
296 	return 0;
297 }
298 
299 /* Phase 1 - core voltage transition ... setup voltage */
300 static int core_voltage_pre_transition(struct powernow_k8_data *data,
301 		u32 reqvid, u32 reqfid)
302 {
303 	u32 rvosteps = data->rvo;
304 	u32 savefid = data->currfid;
305 	u32 maxvid, lo, rvomult = 1;
306 
307 	pr_debug("ph1 (cpu%d): start, currfid 0x%x, currvid 0x%x, "
308 		"reqvid 0x%x, rvo 0x%x\n",
309 		smp_processor_id(),
310 		data->currfid, data->currvid, reqvid, data->rvo);
311 
312 	if ((savefid < LO_FID_TABLE_TOP) && (reqfid < LO_FID_TABLE_TOP))
313 		rvomult = 2;
314 	rvosteps *= rvomult;
315 	rdmsr(MSR_FIDVID_STATUS, lo, maxvid);
316 	maxvid = 0x1f & (maxvid >> 16);
317 	pr_debug("ph1 maxvid=0x%x\n", maxvid);
318 	if (reqvid < maxvid) /* lower numbers are higher voltages */
319 		reqvid = maxvid;
320 
321 	while (data->currvid > reqvid) {
322 		pr_debug("ph1: curr 0x%x, req vid 0x%x\n",
323 			data->currvid, reqvid);
324 		if (decrease_vid_code_by_step(data, reqvid, data->vidmvs))
325 			return 1;
326 	}
327 
328 	while ((rvosteps > 0) &&
329 			((rvomult * data->rvo + data->currvid) > reqvid)) {
330 		if (data->currvid == maxvid) {
331 			rvosteps = 0;
332 		} else {
333 			pr_debug("ph1: changing vid for rvo, req 0x%x\n",
334 				data->currvid - 1);
335 			if (decrease_vid_code_by_step(data, data->currvid-1, 1))
336 				return 1;
337 			rvosteps--;
338 		}
339 	}
340 
341 	if (query_current_values_with_pending_wait(data))
342 		return 1;
343 
344 	if (savefid != data->currfid) {
345 		printk(KERN_ERR PFX "ph1 err, currfid changed 0x%x\n",
346 				data->currfid);
347 		return 1;
348 	}
349 
350 	pr_debug("ph1 complete, currfid 0x%x, currvid 0x%x\n",
351 		data->currfid, data->currvid);
352 
353 	return 0;
354 }
355 
356 /* Phase 2 - core frequency transition */
357 static int core_frequency_transition(struct powernow_k8_data *data, u32 reqfid)
358 {
359 	u32 vcoreqfid, vcocurrfid, vcofiddiff;
360 	u32 fid_interval, savevid = data->currvid;
361 
362 	if (data->currfid == reqfid) {
363 		printk(KERN_ERR PFX "ph2 null fid transition 0x%x\n",
364 				data->currfid);
365 		return 0;
366 	}
367 
368 	pr_debug("ph2 (cpu%d): starting, currfid 0x%x, currvid 0x%x, "
369 		"reqfid 0x%x\n",
370 		smp_processor_id(),
371 		data->currfid, data->currvid, reqfid);
372 
373 	vcoreqfid = convert_fid_to_vco_fid(reqfid);
374 	vcocurrfid = convert_fid_to_vco_fid(data->currfid);
375 	vcofiddiff = vcocurrfid > vcoreqfid ? vcocurrfid - vcoreqfid
376 	    : vcoreqfid - vcocurrfid;
377 
378 	if ((reqfid <= LO_FID_TABLE_TOP) && (data->currfid <= LO_FID_TABLE_TOP))
379 		vcofiddiff = 0;
380 
381 	while (vcofiddiff > 2) {
382 		(data->currfid & 1) ? (fid_interval = 1) : (fid_interval = 2);
383 
384 		if (reqfid > data->currfid) {
385 			if (data->currfid > LO_FID_TABLE_TOP) {
386 				if (write_new_fid(data,
387 						data->currfid + fid_interval))
388 					return 1;
389 			} else {
390 				if (write_new_fid
391 				    (data,
392 				     2 + convert_fid_to_vco_fid(data->currfid)))
393 					return 1;
394 			}
395 		} else {
396 			if (write_new_fid(data, data->currfid - fid_interval))
397 				return 1;
398 		}
399 
400 		vcocurrfid = convert_fid_to_vco_fid(data->currfid);
401 		vcofiddiff = vcocurrfid > vcoreqfid ? vcocurrfid - vcoreqfid
402 		    : vcoreqfid - vcocurrfid;
403 	}
404 
405 	if (write_new_fid(data, reqfid))
406 		return 1;
407 
408 	if (query_current_values_with_pending_wait(data))
409 		return 1;
410 
411 	if (data->currfid != reqfid) {
412 		printk(KERN_ERR PFX
413 			"ph2: mismatch, failed fid transition, "
414 			"curr 0x%x, req 0x%x\n",
415 			data->currfid, reqfid);
416 		return 1;
417 	}
418 
419 	if (savevid != data->currvid) {
420 		printk(KERN_ERR PFX "ph2: vid changed, save 0x%x, curr 0x%x\n",
421 			savevid, data->currvid);
422 		return 1;
423 	}
424 
425 	pr_debug("ph2 complete, currfid 0x%x, currvid 0x%x\n",
426 		data->currfid, data->currvid);
427 
428 	return 0;
429 }
430 
431 /* Phase 3 - core voltage transition flow ... jump to the final vid. */
432 static int core_voltage_post_transition(struct powernow_k8_data *data,
433 		u32 reqvid)
434 {
435 	u32 savefid = data->currfid;
436 	u32 savereqvid = reqvid;
437 
438 	pr_debug("ph3 (cpu%d): starting, currfid 0x%x, currvid 0x%x\n",
439 		smp_processor_id(),
440 		data->currfid, data->currvid);
441 
442 	if (reqvid != data->currvid) {
443 		if (write_new_vid(data, reqvid))
444 			return 1;
445 
446 		if (savefid != data->currfid) {
447 			printk(KERN_ERR PFX
448 			       "ph3: bad fid change, save 0x%x, curr 0x%x\n",
449 			       savefid, data->currfid);
450 			return 1;
451 		}
452 
453 		if (data->currvid != reqvid) {
454 			printk(KERN_ERR PFX
455 			       "ph3: failed vid transition\n, "
456 			       "req 0x%x, curr 0x%x",
457 			       reqvid, data->currvid);
458 			return 1;
459 		}
460 	}
461 
462 	if (query_current_values_with_pending_wait(data))
463 		return 1;
464 
465 	if (savereqvid != data->currvid) {
466 		pr_debug("ph3 failed, currvid 0x%x\n", data->currvid);
467 		return 1;
468 	}
469 
470 	if (savefid != data->currfid) {
471 		pr_debug("ph3 failed, currfid changed 0x%x\n",
472 			data->currfid);
473 		return 1;
474 	}
475 
476 	pr_debug("ph3 complete, currfid 0x%x, currvid 0x%x\n",
477 		data->currfid, data->currvid);
478 
479 	return 0;
480 }
481 
482 static const struct x86_cpu_id powernow_k8_ids[] = {
483 	/* IO based frequency switching */
484 	{ X86_VENDOR_AMD, 0xf },
485 	{}
486 };
487 MODULE_DEVICE_TABLE(x86cpu, powernow_k8_ids);
488 
489 static void check_supported_cpu(void *_rc)
490 {
491 	u32 eax, ebx, ecx, edx;
492 	int *rc = _rc;
493 
494 	*rc = -ENODEV;
495 
496 	eax = cpuid_eax(CPUID_PROCESSOR_SIGNATURE);
497 
498 	if ((eax & CPUID_XFAM) == CPUID_XFAM_K8) {
499 		if (((eax & CPUID_USE_XFAM_XMOD) != CPUID_USE_XFAM_XMOD) ||
500 		    ((eax & CPUID_XMOD) > CPUID_XMOD_REV_MASK)) {
501 			printk(KERN_INFO PFX
502 				"Processor cpuid %x not supported\n", eax);
503 			return;
504 		}
505 
506 		eax = cpuid_eax(CPUID_GET_MAX_CAPABILITIES);
507 		if (eax < CPUID_FREQ_VOLT_CAPABILITIES) {
508 			printk(KERN_INFO PFX
509 			       "No frequency change capabilities detected\n");
510 			return;
511 		}
512 
513 		cpuid(CPUID_FREQ_VOLT_CAPABILITIES, &eax, &ebx, &ecx, &edx);
514 		if ((edx & P_STATE_TRANSITION_CAPABLE)
515 			!= P_STATE_TRANSITION_CAPABLE) {
516 			printk(KERN_INFO PFX
517 				"Power state transitions not supported\n");
518 			return;
519 		}
520 		*rc = 0;
521 	}
522 }
523 
524 static int check_pst_table(struct powernow_k8_data *data, struct pst_s *pst,
525 		u8 maxvid)
526 {
527 	unsigned int j;
528 	u8 lastfid = 0xff;
529 
530 	for (j = 0; j < data->numps; j++) {
531 		if (pst[j].vid > LEAST_VID) {
532 			printk(KERN_ERR FW_BUG PFX "vid %d invalid : 0x%x\n",
533 			       j, pst[j].vid);
534 			return -EINVAL;
535 		}
536 		if (pst[j].vid < data->rvo) {
537 			/* vid + rvo >= 0 */
538 			printk(KERN_ERR FW_BUG PFX "0 vid exceeded with pstate"
539 			       " %d\n", j);
540 			return -ENODEV;
541 		}
542 		if (pst[j].vid < maxvid + data->rvo) {
543 			/* vid + rvo >= maxvid */
544 			printk(KERN_ERR FW_BUG PFX "maxvid exceeded with pstate"
545 			       " %d\n", j);
546 			return -ENODEV;
547 		}
548 		if (pst[j].fid > MAX_FID) {
549 			printk(KERN_ERR FW_BUG PFX "maxfid exceeded with pstate"
550 			       " %d\n", j);
551 			return -ENODEV;
552 		}
553 		if (j && (pst[j].fid < HI_FID_TABLE_BOTTOM)) {
554 			/* Only first fid is allowed to be in "low" range */
555 			printk(KERN_ERR FW_BUG PFX "two low fids - %d : "
556 			       "0x%x\n", j, pst[j].fid);
557 			return -EINVAL;
558 		}
559 		if (pst[j].fid < lastfid)
560 			lastfid = pst[j].fid;
561 	}
562 	if (lastfid & 1) {
563 		printk(KERN_ERR FW_BUG PFX "lastfid invalid\n");
564 		return -EINVAL;
565 	}
566 	if (lastfid > LO_FID_TABLE_TOP)
567 		printk(KERN_INFO FW_BUG PFX
568 			"first fid not from lo freq table\n");
569 
570 	return 0;
571 }
572 
573 static void invalidate_entry(struct cpufreq_frequency_table *powernow_table,
574 		unsigned int entry)
575 {
576 	powernow_table[entry].frequency = CPUFREQ_ENTRY_INVALID;
577 }
578 
579 static void print_basics(struct powernow_k8_data *data)
580 {
581 	int j;
582 	for (j = 0; j < data->numps; j++) {
583 		if (data->powernow_table[j].frequency !=
584 				CPUFREQ_ENTRY_INVALID) {
585 				printk(KERN_INFO PFX
586 					"fid 0x%x (%d MHz), vid 0x%x\n",
587 					data->powernow_table[j].driver_data & 0xff,
588 					data->powernow_table[j].frequency/1000,
589 					data->powernow_table[j].driver_data >> 8);
590 		}
591 	}
592 	if (data->batps)
593 		printk(KERN_INFO PFX "Only %d pstates on battery\n",
594 				data->batps);
595 }
596 
597 static int fill_powernow_table(struct powernow_k8_data *data,
598 		struct pst_s *pst, u8 maxvid)
599 {
600 	struct cpufreq_frequency_table *powernow_table;
601 	unsigned int j;
602 
603 	if (data->batps) {
604 		/* use ACPI support to get full speed on mains power */
605 		printk(KERN_WARNING PFX
606 			"Only %d pstates usable (use ACPI driver for full "
607 			"range\n", data->batps);
608 		data->numps = data->batps;
609 	}
610 
611 	for (j = 1; j < data->numps; j++) {
612 		if (pst[j-1].fid >= pst[j].fid) {
613 			printk(KERN_ERR PFX "PST out of sequence\n");
614 			return -EINVAL;
615 		}
616 	}
617 
618 	if (data->numps < 2) {
619 		printk(KERN_ERR PFX "no p states to transition\n");
620 		return -ENODEV;
621 	}
622 
623 	if (check_pst_table(data, pst, maxvid))
624 		return -EINVAL;
625 
626 	powernow_table = kmalloc((sizeof(*powernow_table)
627 		* (data->numps + 1)), GFP_KERNEL);
628 	if (!powernow_table) {
629 		printk(KERN_ERR PFX "powernow_table memory alloc failure\n");
630 		return -ENOMEM;
631 	}
632 
633 	for (j = 0; j < data->numps; j++) {
634 		int freq;
635 		powernow_table[j].driver_data = pst[j].fid; /* lower 8 bits */
636 		powernow_table[j].driver_data |= (pst[j].vid << 8); /* upper 8 bits */
637 		freq = find_khz_freq_from_fid(pst[j].fid);
638 		powernow_table[j].frequency = freq;
639 	}
640 	powernow_table[data->numps].frequency = CPUFREQ_TABLE_END;
641 	powernow_table[data->numps].driver_data = 0;
642 
643 	if (query_current_values_with_pending_wait(data)) {
644 		kfree(powernow_table);
645 		return -EIO;
646 	}
647 
648 	pr_debug("cfid 0x%x, cvid 0x%x\n", data->currfid, data->currvid);
649 	data->powernow_table = powernow_table;
650 	if (cpumask_first(cpu_core_mask(data->cpu)) == data->cpu)
651 		print_basics(data);
652 
653 	for (j = 0; j < data->numps; j++)
654 		if ((pst[j].fid == data->currfid) &&
655 		    (pst[j].vid == data->currvid))
656 			return 0;
657 
658 	pr_debug("currfid/vid do not match PST, ignoring\n");
659 	return 0;
660 }
661 
662 /* Find and validate the PSB/PST table in BIOS. */
663 static int find_psb_table(struct powernow_k8_data *data)
664 {
665 	struct psb_s *psb;
666 	unsigned int i;
667 	u32 mvs;
668 	u8 maxvid;
669 	u32 cpst = 0;
670 	u32 thiscpuid;
671 
672 	for (i = 0xc0000; i < 0xffff0; i += 0x10) {
673 		/* Scan BIOS looking for the signature. */
674 		/* It can not be at ffff0 - it is too big. */
675 
676 		psb = phys_to_virt(i);
677 		if (memcmp(psb, PSB_ID_STRING, PSB_ID_STRING_LEN) != 0)
678 			continue;
679 
680 		pr_debug("found PSB header at 0x%p\n", psb);
681 
682 		pr_debug("table vers: 0x%x\n", psb->tableversion);
683 		if (psb->tableversion != PSB_VERSION_1_4) {
684 			printk(KERN_ERR FW_BUG PFX "PSB table is not v1.4\n");
685 			return -ENODEV;
686 		}
687 
688 		pr_debug("flags: 0x%x\n", psb->flags1);
689 		if (psb->flags1) {
690 			printk(KERN_ERR FW_BUG PFX "unknown flags\n");
691 			return -ENODEV;
692 		}
693 
694 		data->vstable = psb->vstable;
695 		pr_debug("voltage stabilization time: %d(*20us)\n",
696 				data->vstable);
697 
698 		pr_debug("flags2: 0x%x\n", psb->flags2);
699 		data->rvo = psb->flags2 & 3;
700 		data->irt = ((psb->flags2) >> 2) & 3;
701 		mvs = ((psb->flags2) >> 4) & 3;
702 		data->vidmvs = 1 << mvs;
703 		data->batps = ((psb->flags2) >> 6) & 3;
704 
705 		pr_debug("ramp voltage offset: %d\n", data->rvo);
706 		pr_debug("isochronous relief time: %d\n", data->irt);
707 		pr_debug("maximum voltage step: %d - 0x%x\n", mvs, data->vidmvs);
708 
709 		pr_debug("numpst: 0x%x\n", psb->num_tables);
710 		cpst = psb->num_tables;
711 		if ((psb->cpuid == 0x00000fc0) ||
712 		    (psb->cpuid == 0x00000fe0)) {
713 			thiscpuid = cpuid_eax(CPUID_PROCESSOR_SIGNATURE);
714 			if ((thiscpuid == 0x00000fc0) ||
715 			    (thiscpuid == 0x00000fe0))
716 				cpst = 1;
717 		}
718 		if (cpst != 1) {
719 			printk(KERN_ERR FW_BUG PFX "numpst must be 1\n");
720 			return -ENODEV;
721 		}
722 
723 		data->plllock = psb->plllocktime;
724 		pr_debug("plllocktime: 0x%x (units 1us)\n", psb->plllocktime);
725 		pr_debug("maxfid: 0x%x\n", psb->maxfid);
726 		pr_debug("maxvid: 0x%x\n", psb->maxvid);
727 		maxvid = psb->maxvid;
728 
729 		data->numps = psb->numps;
730 		pr_debug("numpstates: 0x%x\n", data->numps);
731 		return fill_powernow_table(data,
732 				(struct pst_s *)(psb+1), maxvid);
733 	}
734 	/*
735 	 * If you see this message, complain to BIOS manufacturer. If
736 	 * he tells you "we do not support Linux" or some similar
737 	 * nonsense, remember that Windows 2000 uses the same legacy
738 	 * mechanism that the old Linux PSB driver uses. Tell them it
739 	 * is broken with Windows 2000.
740 	 *
741 	 * The reference to the AMD documentation is chapter 9 in the
742 	 * BIOS and Kernel Developer's Guide, which is available on
743 	 * www.amd.com
744 	 */
745 	printk(KERN_ERR FW_BUG PFX "No PSB or ACPI _PSS objects\n");
746 	printk(KERN_ERR PFX "Make sure that your BIOS is up to date"
747 		" and Cool'N'Quiet support is enabled in BIOS setup\n");
748 	return -ENODEV;
749 }
750 
751 static void powernow_k8_acpi_pst_values(struct powernow_k8_data *data,
752 		unsigned int index)
753 {
754 	u64 control;
755 
756 	if (!data->acpi_data.state_count)
757 		return;
758 
759 	control = data->acpi_data.states[index].control;
760 	data->irt = (control >> IRT_SHIFT) & IRT_MASK;
761 	data->rvo = (control >> RVO_SHIFT) & RVO_MASK;
762 	data->exttype = (control >> EXT_TYPE_SHIFT) & EXT_TYPE_MASK;
763 	data->plllock = (control >> PLL_L_SHIFT) & PLL_L_MASK;
764 	data->vidmvs = 1 << ((control >> MVS_SHIFT) & MVS_MASK);
765 	data->vstable = (control >> VST_SHIFT) & VST_MASK;
766 }
767 
768 static int powernow_k8_cpu_init_acpi(struct powernow_k8_data *data)
769 {
770 	struct cpufreq_frequency_table *powernow_table;
771 	int ret_val = -ENODEV;
772 	u64 control, status;
773 
774 	if (acpi_processor_register_performance(&data->acpi_data, data->cpu)) {
775 		pr_debug("register performance failed: bad ACPI data\n");
776 		return -EIO;
777 	}
778 
779 	/* verify the data contained in the ACPI structures */
780 	if (data->acpi_data.state_count <= 1) {
781 		pr_debug("No ACPI P-States\n");
782 		goto err_out;
783 	}
784 
785 	control = data->acpi_data.control_register.space_id;
786 	status = data->acpi_data.status_register.space_id;
787 
788 	if ((control != ACPI_ADR_SPACE_FIXED_HARDWARE) ||
789 	    (status != ACPI_ADR_SPACE_FIXED_HARDWARE)) {
790 		pr_debug("Invalid control/status registers (%llx - %llx)\n",
791 			control, status);
792 		goto err_out;
793 	}
794 
795 	/* fill in data->powernow_table */
796 	powernow_table = kmalloc((sizeof(*powernow_table)
797 		* (data->acpi_data.state_count + 1)), GFP_KERNEL);
798 	if (!powernow_table) {
799 		pr_debug("powernow_table memory alloc failure\n");
800 		goto err_out;
801 	}
802 
803 	/* fill in data */
804 	data->numps = data->acpi_data.state_count;
805 	powernow_k8_acpi_pst_values(data, 0);
806 
807 	ret_val = fill_powernow_table_fidvid(data, powernow_table);
808 	if (ret_val)
809 		goto err_out_mem;
810 
811 	powernow_table[data->acpi_data.state_count].frequency =
812 		CPUFREQ_TABLE_END;
813 	powernow_table[data->acpi_data.state_count].driver_data = 0;
814 	data->powernow_table = powernow_table;
815 
816 	if (cpumask_first(cpu_core_mask(data->cpu)) == data->cpu)
817 		print_basics(data);
818 
819 	/* notify BIOS that we exist */
820 	acpi_processor_notify_smm(THIS_MODULE);
821 
822 	if (!zalloc_cpumask_var(&data->acpi_data.shared_cpu_map, GFP_KERNEL)) {
823 		printk(KERN_ERR PFX
824 				"unable to alloc powernow_k8_data cpumask\n");
825 		ret_val = -ENOMEM;
826 		goto err_out_mem;
827 	}
828 
829 	return 0;
830 
831 err_out_mem:
832 	kfree(powernow_table);
833 
834 err_out:
835 	acpi_processor_unregister_performance(&data->acpi_data, data->cpu);
836 
837 	/* data->acpi_data.state_count informs us at ->exit()
838 	 * whether ACPI was used */
839 	data->acpi_data.state_count = 0;
840 
841 	return ret_val;
842 }
843 
844 static int fill_powernow_table_fidvid(struct powernow_k8_data *data,
845 		struct cpufreq_frequency_table *powernow_table)
846 {
847 	int i;
848 
849 	for (i = 0; i < data->acpi_data.state_count; i++) {
850 		u32 fid;
851 		u32 vid;
852 		u32 freq, index;
853 		u64 status, control;
854 
855 		if (data->exttype) {
856 			status =  data->acpi_data.states[i].status;
857 			fid = status & EXT_FID_MASK;
858 			vid = (status >> VID_SHIFT) & EXT_VID_MASK;
859 		} else {
860 			control =  data->acpi_data.states[i].control;
861 			fid = control & FID_MASK;
862 			vid = (control >> VID_SHIFT) & VID_MASK;
863 		}
864 
865 		pr_debug("   %d : fid 0x%x, vid 0x%x\n", i, fid, vid);
866 
867 		index = fid | (vid<<8);
868 		powernow_table[i].driver_data = index;
869 
870 		freq = find_khz_freq_from_fid(fid);
871 		powernow_table[i].frequency = freq;
872 
873 		/* verify frequency is OK */
874 		if ((freq > (MAX_FREQ * 1000)) || (freq < (MIN_FREQ * 1000))) {
875 			pr_debug("invalid freq %u kHz, ignoring\n", freq);
876 			invalidate_entry(powernow_table, i);
877 			continue;
878 		}
879 
880 		/* verify voltage is OK -
881 		 * BIOSs are using "off" to indicate invalid */
882 		if (vid == VID_OFF) {
883 			pr_debug("invalid vid %u, ignoring\n", vid);
884 			invalidate_entry(powernow_table, i);
885 			continue;
886 		}
887 
888 		if (freq != (data->acpi_data.states[i].core_frequency * 1000)) {
889 			printk(KERN_INFO PFX "invalid freq entries "
890 				"%u kHz vs. %u kHz\n", freq,
891 				(unsigned int)
892 				(data->acpi_data.states[i].core_frequency
893 				 * 1000));
894 			invalidate_entry(powernow_table, i);
895 			continue;
896 		}
897 	}
898 	return 0;
899 }
900 
901 static void powernow_k8_cpu_exit_acpi(struct powernow_k8_data *data)
902 {
903 	if (data->acpi_data.state_count)
904 		acpi_processor_unregister_performance(&data->acpi_data,
905 				data->cpu);
906 	free_cpumask_var(data->acpi_data.shared_cpu_map);
907 }
908 
909 static int get_transition_latency(struct powernow_k8_data *data)
910 {
911 	int max_latency = 0;
912 	int i;
913 	for (i = 0; i < data->acpi_data.state_count; i++) {
914 		int cur_latency = data->acpi_data.states[i].transition_latency
915 			+ data->acpi_data.states[i].bus_master_latency;
916 		if (cur_latency > max_latency)
917 			max_latency = cur_latency;
918 	}
919 	if (max_latency == 0) {
920 		pr_err(FW_WARN PFX "Invalid zero transition latency\n");
921 		max_latency = 1;
922 	}
923 	/* value in usecs, needs to be in nanoseconds */
924 	return 1000 * max_latency;
925 }
926 
927 /* Take a frequency, and issue the fid/vid transition command */
928 static int transition_frequency_fidvid(struct powernow_k8_data *data,
929 		unsigned int index)
930 {
931 	struct cpufreq_policy *policy;
932 	u32 fid = 0;
933 	u32 vid = 0;
934 	int res;
935 	struct cpufreq_freqs freqs;
936 
937 	pr_debug("cpu %d transition to index %u\n", smp_processor_id(), index);
938 
939 	/* fid/vid correctness check for k8 */
940 	/* fid are the lower 8 bits of the index we stored into
941 	 * the cpufreq frequency table in find_psb_table, vid
942 	 * are the upper 8 bits.
943 	 */
944 	fid = data->powernow_table[index].driver_data & 0xFF;
945 	vid = (data->powernow_table[index].driver_data & 0xFF00) >> 8;
946 
947 	pr_debug("table matched fid 0x%x, giving vid 0x%x\n", fid, vid);
948 
949 	if (query_current_values_with_pending_wait(data))
950 		return 1;
951 
952 	if ((data->currvid == vid) && (data->currfid == fid)) {
953 		pr_debug("target matches current values (fid 0x%x, vid 0x%x)\n",
954 			fid, vid);
955 		return 0;
956 	}
957 
958 	pr_debug("cpu %d, changing to fid 0x%x, vid 0x%x\n",
959 		smp_processor_id(), fid, vid);
960 	freqs.old = find_khz_freq_from_fid(data->currfid);
961 	freqs.new = find_khz_freq_from_fid(fid);
962 
963 	policy = cpufreq_cpu_get(smp_processor_id());
964 	cpufreq_cpu_put(policy);
965 
966 	cpufreq_notify_transition(policy, &freqs, CPUFREQ_PRECHANGE);
967 
968 	res = transition_fid_vid(data, fid, vid);
969 	if (res)
970 		freqs.new = freqs.old;
971 	else
972 		freqs.new = find_khz_freq_from_fid(data->currfid);
973 
974 	cpufreq_notify_transition(policy, &freqs, CPUFREQ_POSTCHANGE);
975 	return res;
976 }
977 
978 struct powernowk8_target_arg {
979 	struct cpufreq_policy		*pol;
980 	unsigned			newstate;
981 };
982 
983 static long powernowk8_target_fn(void *arg)
984 {
985 	struct powernowk8_target_arg *pta = arg;
986 	struct cpufreq_policy *pol = pta->pol;
987 	unsigned newstate = pta->newstate;
988 	struct powernow_k8_data *data = per_cpu(powernow_data, pol->cpu);
989 	u32 checkfid;
990 	u32 checkvid;
991 	int ret;
992 
993 	if (!data)
994 		return -EINVAL;
995 
996 	checkfid = data->currfid;
997 	checkvid = data->currvid;
998 
999 	if (pending_bit_stuck()) {
1000 		printk(KERN_ERR PFX "failing targ, change pending bit set\n");
1001 		return -EIO;
1002 	}
1003 
1004 	pr_debug("targ: cpu %d, %d kHz, min %d, max %d\n",
1005 		pol->cpu, data->powernow_table[newstate].frequency, pol->min,
1006 		pol->max);
1007 
1008 	if (query_current_values_with_pending_wait(data))
1009 		return -EIO;
1010 
1011 	pr_debug("targ: curr fid 0x%x, vid 0x%x\n",
1012 		 data->currfid, data->currvid);
1013 
1014 	if ((checkvid != data->currvid) ||
1015 	    (checkfid != data->currfid)) {
1016 		pr_info(PFX
1017 		       "error - out of sync, fix 0x%x 0x%x, vid 0x%x 0x%x\n",
1018 		       checkfid, data->currfid,
1019 		       checkvid, data->currvid);
1020 	}
1021 
1022 	mutex_lock(&fidvid_mutex);
1023 
1024 	powernow_k8_acpi_pst_values(data, newstate);
1025 
1026 	ret = transition_frequency_fidvid(data, newstate);
1027 
1028 	if (ret) {
1029 		printk(KERN_ERR PFX "transition frequency failed\n");
1030 		mutex_unlock(&fidvid_mutex);
1031 		return 1;
1032 	}
1033 	mutex_unlock(&fidvid_mutex);
1034 
1035 	pol->cur = find_khz_freq_from_fid(data->currfid);
1036 
1037 	return 0;
1038 }
1039 
1040 /* Driver entry point to switch to the target frequency */
1041 static int powernowk8_target(struct cpufreq_policy *pol, unsigned index)
1042 {
1043 	struct powernowk8_target_arg pta = { .pol = pol, .newstate = index };
1044 
1045 	return work_on_cpu(pol->cpu, powernowk8_target_fn, &pta);
1046 }
1047 
1048 struct init_on_cpu {
1049 	struct powernow_k8_data *data;
1050 	int rc;
1051 };
1052 
1053 static void powernowk8_cpu_init_on_cpu(void *_init_on_cpu)
1054 {
1055 	struct init_on_cpu *init_on_cpu = _init_on_cpu;
1056 
1057 	if (pending_bit_stuck()) {
1058 		printk(KERN_ERR PFX "failing init, change pending bit set\n");
1059 		init_on_cpu->rc = -ENODEV;
1060 		return;
1061 	}
1062 
1063 	if (query_current_values_with_pending_wait(init_on_cpu->data)) {
1064 		init_on_cpu->rc = -ENODEV;
1065 		return;
1066 	}
1067 
1068 	fidvid_msr_init();
1069 
1070 	init_on_cpu->rc = 0;
1071 }
1072 
1073 static const char missing_pss_msg[] =
1074 	KERN_ERR
1075 	FW_BUG PFX "No compatible ACPI _PSS objects found.\n"
1076 	FW_BUG PFX "First, make sure Cool'N'Quiet is enabled in the BIOS.\n"
1077 	FW_BUG PFX "If that doesn't help, try upgrading your BIOS.\n";
1078 
1079 /* per CPU init entry point to the driver */
1080 static int powernowk8_cpu_init(struct cpufreq_policy *pol)
1081 {
1082 	struct powernow_k8_data *data;
1083 	struct init_on_cpu init_on_cpu;
1084 	int rc;
1085 
1086 	smp_call_function_single(pol->cpu, check_supported_cpu, &rc, 1);
1087 	if (rc)
1088 		return -ENODEV;
1089 
1090 	data = kzalloc(sizeof(*data), GFP_KERNEL);
1091 	if (!data) {
1092 		printk(KERN_ERR PFX "unable to alloc powernow_k8_data");
1093 		return -ENOMEM;
1094 	}
1095 
1096 	data->cpu = pol->cpu;
1097 
1098 	if (powernow_k8_cpu_init_acpi(data)) {
1099 		/*
1100 		 * Use the PSB BIOS structure. This is only available on
1101 		 * an UP version, and is deprecated by AMD.
1102 		 */
1103 		if (num_online_cpus() != 1) {
1104 			printk_once(missing_pss_msg);
1105 			goto err_out;
1106 		}
1107 		if (pol->cpu != 0) {
1108 			printk(KERN_ERR FW_BUG PFX "No ACPI _PSS objects for "
1109 			       "CPU other than CPU0. Complain to your BIOS "
1110 			       "vendor.\n");
1111 			goto err_out;
1112 		}
1113 		rc = find_psb_table(data);
1114 		if (rc)
1115 			goto err_out;
1116 
1117 		/* Take a crude guess here.
1118 		 * That guess was in microseconds, so multiply with 1000 */
1119 		pol->cpuinfo.transition_latency = (
1120 			 ((data->rvo + 8) * data->vstable * VST_UNITS_20US) +
1121 			 ((1 << data->irt) * 30)) * 1000;
1122 	} else /* ACPI _PSS objects available */
1123 		pol->cpuinfo.transition_latency = get_transition_latency(data);
1124 
1125 	/* only run on specific CPU from here on */
1126 	init_on_cpu.data = data;
1127 	smp_call_function_single(data->cpu, powernowk8_cpu_init_on_cpu,
1128 				 &init_on_cpu, 1);
1129 	rc = init_on_cpu.rc;
1130 	if (rc != 0)
1131 		goto err_out_exit_acpi;
1132 
1133 	cpumask_copy(pol->cpus, cpu_core_mask(pol->cpu));
1134 	data->available_cores = pol->cpus;
1135 
1136 	/* min/max the cpu is capable of */
1137 	if (cpufreq_table_validate_and_show(pol, data->powernow_table)) {
1138 		printk(KERN_ERR FW_BUG PFX "invalid powernow_table\n");
1139 		powernow_k8_cpu_exit_acpi(data);
1140 		kfree(data->powernow_table);
1141 		kfree(data);
1142 		return -EINVAL;
1143 	}
1144 
1145 	pr_debug("cpu_init done, current fid 0x%x, vid 0x%x\n",
1146 		 data->currfid, data->currvid);
1147 
1148 	per_cpu(powernow_data, pol->cpu) = data;
1149 
1150 	return 0;
1151 
1152 err_out_exit_acpi:
1153 	powernow_k8_cpu_exit_acpi(data);
1154 
1155 err_out:
1156 	kfree(data);
1157 	return -ENODEV;
1158 }
1159 
1160 static int powernowk8_cpu_exit(struct cpufreq_policy *pol)
1161 {
1162 	struct powernow_k8_data *data = per_cpu(powernow_data, pol->cpu);
1163 
1164 	if (!data)
1165 		return -EINVAL;
1166 
1167 	powernow_k8_cpu_exit_acpi(data);
1168 
1169 	cpufreq_frequency_table_put_attr(pol->cpu);
1170 
1171 	kfree(data->powernow_table);
1172 	kfree(data);
1173 	per_cpu(powernow_data, pol->cpu) = NULL;
1174 
1175 	return 0;
1176 }
1177 
1178 static void query_values_on_cpu(void *_err)
1179 {
1180 	int *err = _err;
1181 	struct powernow_k8_data *data = __this_cpu_read(powernow_data);
1182 
1183 	*err = query_current_values_with_pending_wait(data);
1184 }
1185 
1186 static unsigned int powernowk8_get(unsigned int cpu)
1187 {
1188 	struct powernow_k8_data *data = per_cpu(powernow_data, cpu);
1189 	unsigned int khz = 0;
1190 	int err;
1191 
1192 	if (!data)
1193 		return 0;
1194 
1195 	smp_call_function_single(cpu, query_values_on_cpu, &err, true);
1196 	if (err)
1197 		goto out;
1198 
1199 	khz = find_khz_freq_from_fid(data->currfid);
1200 
1201 
1202 out:
1203 	return khz;
1204 }
1205 
1206 static struct cpufreq_driver cpufreq_amd64_driver = {
1207 	.flags		= CPUFREQ_ASYNC_NOTIFICATION,
1208 	.verify		= cpufreq_generic_frequency_table_verify,
1209 	.target_index	= powernowk8_target,
1210 	.bios_limit	= acpi_processor_get_bios_limit,
1211 	.init		= powernowk8_cpu_init,
1212 	.exit		= powernowk8_cpu_exit,
1213 	.get		= powernowk8_get,
1214 	.name		= "powernow-k8",
1215 	.attr		= cpufreq_generic_attr,
1216 };
1217 
1218 static void __request_acpi_cpufreq(void)
1219 {
1220 	const char *cur_drv, *drv = "acpi-cpufreq";
1221 
1222 	cur_drv = cpufreq_get_current_driver();
1223 	if (!cur_drv)
1224 		goto request;
1225 
1226 	if (strncmp(cur_drv, drv, min_t(size_t, strlen(cur_drv), strlen(drv))))
1227 		pr_warn(PFX "WTF driver: %s\n", cur_drv);
1228 
1229 	return;
1230 
1231  request:
1232 	pr_warn(PFX "This CPU is not supported anymore, using acpi-cpufreq instead.\n");
1233 	request_module(drv);
1234 }
1235 
1236 /* driver entry point for init */
1237 static int powernowk8_init(void)
1238 {
1239 	unsigned int i, supported_cpus = 0;
1240 	int ret;
1241 
1242 	if (static_cpu_has(X86_FEATURE_HW_PSTATE)) {
1243 		__request_acpi_cpufreq();
1244 		return -ENODEV;
1245 	}
1246 
1247 	if (!x86_match_cpu(powernow_k8_ids))
1248 		return -ENODEV;
1249 
1250 	get_online_cpus();
1251 	for_each_online_cpu(i) {
1252 		smp_call_function_single(i, check_supported_cpu, &ret, 1);
1253 		if (!ret)
1254 			supported_cpus++;
1255 	}
1256 
1257 	if (supported_cpus != num_online_cpus()) {
1258 		put_online_cpus();
1259 		return -ENODEV;
1260 	}
1261 	put_online_cpus();
1262 
1263 	ret = cpufreq_register_driver(&cpufreq_amd64_driver);
1264 	if (ret)
1265 		return ret;
1266 
1267 	pr_info(PFX "Found %d %s (%d cpu cores) (" VERSION ")\n",
1268 		num_online_nodes(), boot_cpu_data.x86_model_id, supported_cpus);
1269 
1270 	return ret;
1271 }
1272 
1273 /* driver entry point for term */
1274 static void __exit powernowk8_exit(void)
1275 {
1276 	pr_debug("exit\n");
1277 
1278 	cpufreq_unregister_driver(&cpufreq_amd64_driver);
1279 }
1280 
1281 MODULE_AUTHOR("Paul Devriendt <paul.devriendt@amd.com> and "
1282 		"Mark Langsdorf <mark.langsdorf@amd.com>");
1283 MODULE_DESCRIPTION("AMD Athlon 64 and Opteron processor frequency driver.");
1284 MODULE_LICENSE("GPL");
1285 
1286 late_initcall(powernowk8_init);
1287 module_exit(powernowk8_exit);
1288