2 * (c) 2003-2006 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
7 * Support : mark.langsdorf@amd.com
9 * Based on the powernow-k7.c module written by Dave Jones.
10 * (C) 2003 Dave Jones <davej@codemonkey.org.uk> on behalf of SuSE Labs
11 * (C) 2004 Dominik Brodowski <linux@brodo.de>
12 * (C) 2004 Pavel Machek <pavel@suse.cz>
13 * Licensed under the terms of the GNU GPL License version 2.
14 * Based upon datasheets & sample CPUs kindly provided by AMD.
16 * Valuable input gratefully received from Dave Jones, Pavel Machek,
17 * Dominik Brodowski, Jacob Shin, and others.
18 * Originally developed by Paul Devriendt.
19 * Processor information obtained from Chapter 9 (Power and Thermal Management)
20 * of the "BIOS and Kernel Developer's Guide for the AMD Athlon 64 and AMD
21 * Opteron Processors" available for download from www.amd.com
23 * Tables for specific CPUs can be inferred from
24 * http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/30430.pdf
27 #include <linux/kernel.h>
28 #include <linux/smp.h>
29 #include <linux/module.h>
30 #include <linux/init.h>
31 #include <linux/cpufreq.h>
32 #include <linux/slab.h>
33 #include <linux/string.h>
34 #include <linux/cpumask.h>
35 #include <linux/sched.h> /* for current / set_cpus_allowed() */
39 #include <asm/delay.h>
41 #ifdef CONFIG_X86_POWERNOW_K8_ACPI
42 #include <linux/acpi.h>
43 #include <linux/mutex.h>
44 #include <acpi/processor.h>
47 #define PFX "powernow-k8: "
48 #define BFX PFX "BIOS error: "
49 #define VERSION "version 2.20.00"
50 #include "powernow-k8.h"
52 /* serialize freq changes */
53 static DEFINE_MUTEX(fidvid_mutex);
55 static DEFINE_PER_CPU(struct powernow_k8_data *, powernow_data);
57 static int cpu_family = CPU_OPTERON;
60 DEFINE_PER_CPU(cpumask_t, cpu_core_map);
63 /* Return a frequency in MHz, given an input fid */
64 static u32 find_freq_from_fid(u32 fid)
66 return 800 + (fid * 100);
69 /* Return a frequency in KHz, given an input fid */
70 static u32 find_khz_freq_from_fid(u32 fid)
72 return 1000 * find_freq_from_fid(fid);
75 static u32 find_khz_freq_from_pstate(struct cpufreq_frequency_table *data, u32 pstate)
77 return data[pstate].frequency;
80 /* Return the vco fid for an input fid
82 * Each "low" fid has corresponding "high" fid, and you can get to "low" fids
83 * only from corresponding high fids. This returns "high" fid corresponding to
86 static u32 convert_fid_to_vco_fid(u32 fid)
88 if (fid < HI_FID_TABLE_BOTTOM)
95 * Return 1 if the pending bit is set. Unless we just instructed the processor
96 * to transition to a new state, seeing this bit set is really bad news.
98 static int pending_bit_stuck(void)
102 if (cpu_family == CPU_HW_PSTATE)
105 rdmsr(MSR_FIDVID_STATUS, lo, hi);
106 return lo & MSR_S_LO_CHANGE_PENDING ? 1 : 0;
110 * Update the global current fid / vid values from the status msr.
111 * Returns 1 on error.
113 static int query_current_values_with_pending_wait(struct powernow_k8_data *data)
118 if (cpu_family == CPU_HW_PSTATE) {
119 rdmsr(MSR_PSTATE_STATUS, lo, hi);
120 i = lo & HW_PSTATE_MASK;
121 data->currpstate = i;
126 dprintk("detected change pending stuck\n");
129 rdmsr(MSR_FIDVID_STATUS, lo, hi);
130 } while (lo & MSR_S_LO_CHANGE_PENDING);
132 data->currvid = hi & MSR_S_HI_CURRENT_VID;
133 data->currfid = lo & MSR_S_LO_CURRENT_FID;
138 /* the isochronous relief time */
139 static void count_off_irt(struct powernow_k8_data *data)
141 udelay((1 << data->irt) * 10);
145 /* the voltage stabilization time */
146 static void count_off_vst(struct powernow_k8_data *data)
148 udelay(data->vstable * VST_UNITS_20US);
152 /* need to init the control msr to a safe value (for each cpu) */
153 static void fidvid_msr_init(void)
158 rdmsr(MSR_FIDVID_STATUS, lo, hi);
159 vid = hi & MSR_S_HI_CURRENT_VID;
160 fid = lo & MSR_S_LO_CURRENT_FID;
161 lo = fid | (vid << MSR_C_LO_VID_SHIFT);
162 hi = MSR_C_HI_STP_GNT_BENIGN;
163 dprintk("cpu%d, init lo 0x%x, hi 0x%x\n", smp_processor_id(), lo, hi);
164 wrmsr(MSR_FIDVID_CTL, lo, hi);
167 /* write the new fid value along with the other control fields to the msr */
168 static int write_new_fid(struct powernow_k8_data *data, u32 fid)
171 u32 savevid = data->currvid;
174 if ((fid & INVALID_FID_MASK) || (data->currvid & INVALID_VID_MASK)) {
175 printk(KERN_ERR PFX "internal error - overflow on fid write\n");
179 lo = fid | (data->currvid << MSR_C_LO_VID_SHIFT) | MSR_C_LO_INIT_FID_VID;
181 dprintk("writing fid 0x%x, lo 0x%x, hi 0x%x\n",
182 fid, lo, data->plllock * PLL_LOCK_CONVERSION);
185 wrmsr(MSR_FIDVID_CTL, lo, data->plllock * PLL_LOCK_CONVERSION);
187 printk(KERN_ERR PFX "Hardware error - pending bit very stuck - no further pstate changes possible\n");
190 } while (query_current_values_with_pending_wait(data));
194 if (savevid != data->currvid) {
195 printk(KERN_ERR PFX "vid change on fid trans, old 0x%x, new 0x%x\n",
196 savevid, data->currvid);
200 if (fid != data->currfid) {
201 printk(KERN_ERR PFX "fid trans failed, fid 0x%x, curr 0x%x\n", fid,
209 /* Write a new vid to the hardware */
210 static int write_new_vid(struct powernow_k8_data *data, u32 vid)
213 u32 savefid = data->currfid;
216 if ((data->currfid & INVALID_FID_MASK) || (vid & INVALID_VID_MASK)) {
217 printk(KERN_ERR PFX "internal error - overflow on vid write\n");
221 lo = data->currfid | (vid << MSR_C_LO_VID_SHIFT) | MSR_C_LO_INIT_FID_VID;
223 dprintk("writing vid 0x%x, lo 0x%x, hi 0x%x\n",
224 vid, lo, STOP_GRANT_5NS);
227 wrmsr(MSR_FIDVID_CTL, lo, STOP_GRANT_5NS);
229 printk(KERN_ERR PFX "internal error - pending bit very stuck - no further pstate changes possible\n");
232 } while (query_current_values_with_pending_wait(data));
234 if (savefid != data->currfid) {
235 printk(KERN_ERR PFX "fid changed on vid trans, old 0x%x new 0x%x\n",
236 savefid, data->currfid);
240 if (vid != data->currvid) {
241 printk(KERN_ERR PFX "vid trans failed, vid 0x%x, curr 0x%x\n", vid,
250 * Reduce the vid by the max of step or reqvid.
251 * Decreasing vid codes represent increasing voltages:
252 * vid of 0 is 1.550V, vid of 0x1e is 0.800V, vid of VID_OFF is off.
254 static int decrease_vid_code_by_step(struct powernow_k8_data *data, u32 reqvid, u32 step)
256 if ((data->currvid - reqvid) > step)
257 reqvid = data->currvid - step;
259 if (write_new_vid(data, reqvid))
267 /* Change hardware pstate by single MSR write */
268 static int transition_pstate(struct powernow_k8_data *data, u32 pstate)
270 wrmsr(MSR_PSTATE_CTRL, pstate, 0);
271 data->currpstate = pstate;
275 /* Change Opteron/Athlon64 fid and vid, by the 3 phases. */
276 static int transition_fid_vid(struct powernow_k8_data *data, u32 reqfid, u32 reqvid)
278 if (core_voltage_pre_transition(data, reqvid))
281 if (core_frequency_transition(data, reqfid))
284 if (core_voltage_post_transition(data, reqvid))
287 if (query_current_values_with_pending_wait(data))
290 if ((reqfid != data->currfid) || (reqvid != data->currvid)) {
291 printk(KERN_ERR PFX "failed (cpu%d): req 0x%x 0x%x, curr 0x%x 0x%x\n",
293 reqfid, reqvid, data->currfid, data->currvid);
297 dprintk("transitioned (cpu%d): new fid 0x%x, vid 0x%x\n",
298 smp_processor_id(), data->currfid, data->currvid);
303 /* Phase 1 - core voltage transition ... setup voltage */
304 static int core_voltage_pre_transition(struct powernow_k8_data *data, u32 reqvid)
306 u32 rvosteps = data->rvo;
307 u32 savefid = data->currfid;
310 dprintk("ph1 (cpu%d): start, currfid 0x%x, currvid 0x%x, reqvid 0x%x, rvo 0x%x\n",
312 data->currfid, data->currvid, reqvid, data->rvo);
314 rdmsr(MSR_FIDVID_STATUS, lo, maxvid);
315 maxvid = 0x1f & (maxvid >> 16);
316 dprintk("ph1 maxvid=0x%x\n", maxvid);
317 if (reqvid < maxvid) /* lower numbers are higher voltages */
320 while (data->currvid > reqvid) {
321 dprintk("ph1: curr 0x%x, req vid 0x%x\n",
322 data->currvid, reqvid);
323 if (decrease_vid_code_by_step(data, reqvid, data->vidmvs))
327 while ((rvosteps > 0) && ((data->rvo + data->currvid) > reqvid)) {
328 if (data->currvid == maxvid) {
331 dprintk("ph1: changing vid for rvo, req 0x%x\n",
333 if (decrease_vid_code_by_step(data, data->currvid - 1, 1))
339 if (query_current_values_with_pending_wait(data))
342 if (savefid != data->currfid) {
343 printk(KERN_ERR PFX "ph1 err, currfid changed 0x%x\n", data->currfid);
347 dprintk("ph1 complete, currfid 0x%x, currvid 0x%x\n",
348 data->currfid, data->currvid);
353 /* Phase 2 - core frequency transition */
354 static int core_frequency_transition(struct powernow_k8_data *data, u32 reqfid)
356 u32 vcoreqfid, vcocurrfid, vcofiddiff, fid_interval, savevid = data->currvid;
358 if ((reqfid < HI_FID_TABLE_BOTTOM) && (data->currfid < HI_FID_TABLE_BOTTOM)) {
359 printk(KERN_ERR PFX "ph2: illegal lo-lo transition 0x%x 0x%x\n",
360 reqfid, data->currfid);
364 if (data->currfid == reqfid) {
365 printk(KERN_ERR PFX "ph2 null fid transition 0x%x\n", data->currfid);
369 dprintk("ph2 (cpu%d): starting, currfid 0x%x, currvid 0x%x, reqfid 0x%x\n",
371 data->currfid, data->currvid, reqfid);
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;
378 while (vcofiddiff > 2) {
379 (data->currfid & 1) ? (fid_interval = 1) : (fid_interval = 2);
381 if (reqfid > data->currfid) {
382 if (data->currfid > LO_FID_TABLE_TOP) {
383 if (write_new_fid(data, data->currfid + fid_interval)) {
388 (data, 2 + convert_fid_to_vco_fid(data->currfid))) {
393 if (write_new_fid(data, data->currfid - fid_interval))
397 vcocurrfid = convert_fid_to_vco_fid(data->currfid);
398 vcofiddiff = vcocurrfid > vcoreqfid ? vcocurrfid - vcoreqfid
399 : vcoreqfid - vcocurrfid;
402 if (write_new_fid(data, reqfid))
405 if (query_current_values_with_pending_wait(data))
408 if (data->currfid != reqfid) {
410 "ph2: mismatch, failed fid transition, curr 0x%x, req 0x%x\n",
411 data->currfid, reqfid);
415 if (savevid != data->currvid) {
416 printk(KERN_ERR PFX "ph2: vid changed, save 0x%x, curr 0x%x\n",
417 savevid, data->currvid);
421 dprintk("ph2 complete, currfid 0x%x, currvid 0x%x\n",
422 data->currfid, data->currvid);
427 /* Phase 3 - core voltage transition flow ... jump to the final vid. */
428 static int core_voltage_post_transition(struct powernow_k8_data *data, u32 reqvid)
430 u32 savefid = data->currfid;
431 u32 savereqvid = reqvid;
433 dprintk("ph3 (cpu%d): starting, currfid 0x%x, currvid 0x%x\n",
435 data->currfid, data->currvid);
437 if (reqvid != data->currvid) {
438 if (write_new_vid(data, reqvid))
441 if (savefid != data->currfid) {
443 "ph3: bad fid change, save 0x%x, curr 0x%x\n",
444 savefid, data->currfid);
448 if (data->currvid != reqvid) {
450 "ph3: failed vid transition\n, req 0x%x, curr 0x%x",
451 reqvid, data->currvid);
456 if (query_current_values_with_pending_wait(data))
459 if (savereqvid != data->currvid) {
460 dprintk("ph3 failed, currvid 0x%x\n", data->currvid);
464 if (savefid != data->currfid) {
465 dprintk("ph3 failed, currfid changed 0x%x\n",
470 dprintk("ph3 complete, currfid 0x%x, currvid 0x%x\n",
471 data->currfid, data->currvid);
476 static int check_supported_cpu(unsigned int cpu)
479 u32 eax, ebx, ecx, edx;
482 oldmask = current->cpus_allowed;
483 set_cpus_allowed_ptr(current, &cpumask_of_cpu(cpu));
485 if (smp_processor_id() != cpu) {
486 printk(KERN_ERR PFX "limiting to cpu %u failed\n", cpu);
490 if (current_cpu_data.x86_vendor != X86_VENDOR_AMD)
493 eax = cpuid_eax(CPUID_PROCESSOR_SIGNATURE);
494 if (((eax & CPUID_XFAM) != CPUID_XFAM_K8) &&
495 ((eax & CPUID_XFAM) < CPUID_XFAM_10H))
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 "Processor cpuid %x not supported\n", eax);
505 eax = cpuid_eax(CPUID_GET_MAX_CAPABILITIES);
506 if (eax < CPUID_FREQ_VOLT_CAPABILITIES) {
508 "No frequency change capabilities detected\n");
512 cpuid(CPUID_FREQ_VOLT_CAPABILITIES, &eax, &ebx, &ecx, &edx);
513 if ((edx & P_STATE_TRANSITION_CAPABLE) != P_STATE_TRANSITION_CAPABLE) {
514 printk(KERN_INFO PFX "Power state transitions not supported\n");
517 } else { /* must be a HW Pstate capable processor */
518 cpuid(CPUID_FREQ_VOLT_CAPABILITIES, &eax, &ebx, &ecx, &edx);
519 if ((edx & USE_HW_PSTATE) == USE_HW_PSTATE)
520 cpu_family = CPU_HW_PSTATE;
528 set_cpus_allowed_ptr(current, &oldmask);
532 static int check_pst_table(struct powernow_k8_data *data, struct pst_s *pst, u8 maxvid)
537 for (j = 0; j < data->numps; j++) {
538 if (pst[j].vid > LEAST_VID) {
539 printk(KERN_ERR PFX "vid %d invalid : 0x%x\n", j, pst[j].vid);
542 if (pst[j].vid < data->rvo) { /* vid + rvo >= 0 */
543 printk(KERN_ERR BFX "0 vid exceeded with pstate %d\n", j);
546 if (pst[j].vid < maxvid + data->rvo) { /* vid + rvo >= maxvid */
547 printk(KERN_ERR BFX "maxvid exceeded with pstate %d\n", j);
550 if (pst[j].fid > MAX_FID) {
551 printk(KERN_ERR BFX "maxfid exceeded with pstate %d\n", j);
554 if (j && (pst[j].fid < HI_FID_TABLE_BOTTOM)) {
555 /* Only first fid is allowed to be in "low" range */
556 printk(KERN_ERR BFX "two low fids - %d : 0x%x\n", j, pst[j].fid);
559 if (pst[j].fid < lastfid)
560 lastfid = pst[j].fid;
563 printk(KERN_ERR BFX "lastfid invalid\n");
566 if (lastfid > LO_FID_TABLE_TOP)
567 printk(KERN_INFO BFX "first fid not from lo freq table\n");
572 static void print_basics(struct powernow_k8_data *data)
575 for (j = 0; j < data->numps; j++) {
576 if (data->powernow_table[j].frequency != CPUFREQ_ENTRY_INVALID) {
577 if (cpu_family == CPU_HW_PSTATE) {
578 printk(KERN_INFO PFX " %d : pstate %d (%d MHz)\n",
580 data->powernow_table[j].index,
581 data->powernow_table[j].frequency/1000);
583 printk(KERN_INFO PFX " %d : fid 0x%x (%d MHz), vid 0x%x\n",
585 data->powernow_table[j].index & 0xff,
586 data->powernow_table[j].frequency/1000,
587 data->powernow_table[j].index >> 8);
592 printk(KERN_INFO PFX "Only %d pstates on battery\n", data->batps);
595 static int fill_powernow_table(struct powernow_k8_data *data, struct pst_s *pst, u8 maxvid)
597 struct cpufreq_frequency_table *powernow_table;
600 if (data->batps) { /* use ACPI support to get full speed on mains power */
601 printk(KERN_WARNING PFX "Only %d pstates usable (use ACPI driver for full range\n", data->batps);
602 data->numps = data->batps;
605 for ( j=1; j<data->numps; j++ ) {
606 if (pst[j-1].fid >= pst[j].fid) {
607 printk(KERN_ERR PFX "PST out of sequence\n");
612 if (data->numps < 2) {
613 printk(KERN_ERR PFX "no p states to transition\n");
617 if (check_pst_table(data, pst, maxvid))
620 powernow_table = kmalloc((sizeof(struct cpufreq_frequency_table)
621 * (data->numps + 1)), GFP_KERNEL);
622 if (!powernow_table) {
623 printk(KERN_ERR PFX "powernow_table memory alloc failure\n");
627 for (j = 0; j < data->numps; j++) {
628 powernow_table[j].index = pst[j].fid; /* lower 8 bits */
629 powernow_table[j].index |= (pst[j].vid << 8); /* upper 8 bits */
630 powernow_table[j].frequency = find_khz_freq_from_fid(pst[j].fid);
632 powernow_table[data->numps].frequency = CPUFREQ_TABLE_END;
633 powernow_table[data->numps].index = 0;
635 if (query_current_values_with_pending_wait(data)) {
636 kfree(powernow_table);
640 dprintk("cfid 0x%x, cvid 0x%x\n", data->currfid, data->currvid);
641 data->powernow_table = powernow_table;
642 if (first_cpu(per_cpu(cpu_core_map, data->cpu)) == data->cpu)
645 for (j = 0; j < data->numps; j++)
646 if ((pst[j].fid==data->currfid) && (pst[j].vid==data->currvid))
649 dprintk("currfid/vid do not match PST, ignoring\n");
653 /* Find and validate the PSB/PST table in BIOS. */
654 static int find_psb_table(struct powernow_k8_data *data)
663 for (i = 0xc0000; i < 0xffff0; i += 0x10) {
664 /* Scan BIOS looking for the signature. */
665 /* It can not be at ffff0 - it is too big. */
667 psb = phys_to_virt(i);
668 if (memcmp(psb, PSB_ID_STRING, PSB_ID_STRING_LEN) != 0)
671 dprintk("found PSB header at 0x%p\n", psb);
673 dprintk("table vers: 0x%x\n", psb->tableversion);
674 if (psb->tableversion != PSB_VERSION_1_4) {
675 printk(KERN_ERR BFX "PSB table is not v1.4\n");
679 dprintk("flags: 0x%x\n", psb->flags1);
681 printk(KERN_ERR BFX "unknown flags\n");
685 data->vstable = psb->vstable;
686 dprintk("voltage stabilization time: %d(*20us)\n", data->vstable);
688 dprintk("flags2: 0x%x\n", psb->flags2);
689 data->rvo = psb->flags2 & 3;
690 data->irt = ((psb->flags2) >> 2) & 3;
691 mvs = ((psb->flags2) >> 4) & 3;
692 data->vidmvs = 1 << mvs;
693 data->batps = ((psb->flags2) >> 6) & 3;
695 dprintk("ramp voltage offset: %d\n", data->rvo);
696 dprintk("isochronous relief time: %d\n", data->irt);
697 dprintk("maximum voltage step: %d - 0x%x\n", mvs, data->vidmvs);
699 dprintk("numpst: 0x%x\n", psb->num_tables);
700 cpst = psb->num_tables;
701 if ((psb->cpuid == 0x00000fc0) || (psb->cpuid == 0x00000fe0) ){
702 thiscpuid = cpuid_eax(CPUID_PROCESSOR_SIGNATURE);
703 if ((thiscpuid == 0x00000fc0) || (thiscpuid == 0x00000fe0) ) {
708 printk(KERN_ERR BFX "numpst must be 1\n");
712 data->plllock = psb->plllocktime;
713 dprintk("plllocktime: 0x%x (units 1us)\n", psb->plllocktime);
714 dprintk("maxfid: 0x%x\n", psb->maxfid);
715 dprintk("maxvid: 0x%x\n", psb->maxvid);
716 maxvid = psb->maxvid;
718 data->numps = psb->numps;
719 dprintk("numpstates: 0x%x\n", data->numps);
720 return fill_powernow_table(data, (struct pst_s *)(psb+1), maxvid);
723 * If you see this message, complain to BIOS manufacturer. If
724 * he tells you "we do not support Linux" or some similar
725 * nonsense, remember that Windows 2000 uses the same legacy
726 * mechanism that the old Linux PSB driver uses. Tell them it
727 * is broken with Windows 2000.
729 * The reference to the AMD documentation is chapter 9 in the
730 * BIOS and Kernel Developer's Guide, which is available on
733 printk(KERN_ERR PFX "BIOS error - no PSB or ACPI _PSS objects\n");
737 #ifdef CONFIG_X86_POWERNOW_K8_ACPI
738 static void powernow_k8_acpi_pst_values(struct powernow_k8_data *data, unsigned int index)
740 if (!data->acpi_data->state_count || (cpu_family == CPU_HW_PSTATE))
743 data->irt = (data->acpi_data->states[index].control >> IRT_SHIFT) & IRT_MASK;
744 data->rvo = (data->acpi_data->states[index].control >> RVO_SHIFT) & RVO_MASK;
745 data->exttype = (data->acpi_data->states[index].control >> EXT_TYPE_SHIFT) & EXT_TYPE_MASK;
746 data->plllock = (data->acpi_data->states[index].control >> PLL_L_SHIFT) & PLL_L_MASK;
747 data->vidmvs = 1 << ((data->acpi_data->states[index].control >> MVS_SHIFT) & MVS_MASK);
748 data->vstable = (data->acpi_data->states[index].control >> VST_SHIFT) & VST_MASK;
752 static struct acpi_processor_performance *acpi_perf_data;
753 static int preregister_valid;
755 static int powernow_k8_cpu_preinit_acpi(void)
757 acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
761 if (acpi_processor_preregister_performance(acpi_perf_data))
764 preregister_valid = 1;
768 static int powernow_k8_cpu_init_acpi(struct powernow_k8_data *data)
770 struct cpufreq_frequency_table *powernow_table;
774 data->acpi_data = percpu_ptr(acpi_perf_data, cpu);
775 if (acpi_processor_register_performance(data->acpi_data, data->cpu)) {
776 dprintk("register performance failed: bad ACPI data\n");
780 /* verify the data contained in the ACPI structures */
781 if (data->acpi_data->state_count <= 1) {
782 dprintk("No ACPI P-States\n");
786 if ((data->acpi_data->control_register.space_id != ACPI_ADR_SPACE_FIXED_HARDWARE) ||
787 (data->acpi_data->status_register.space_id != ACPI_ADR_SPACE_FIXED_HARDWARE)) {
788 dprintk("Invalid control/status registers (%x - %x)\n",
789 data->acpi_data->control_register.space_id,
790 data->acpi_data->status_register.space_id);
794 /* fill in data->powernow_table */
795 powernow_table = kmalloc((sizeof(struct cpufreq_frequency_table)
796 * (data->acpi_data->state_count + 1)), GFP_KERNEL);
797 if (!powernow_table) {
798 dprintk("powernow_table memory alloc failure\n");
802 if (cpu_family == CPU_HW_PSTATE)
803 ret_val = fill_powernow_table_pstate(data, powernow_table);
805 ret_val = fill_powernow_table_fidvid(data, powernow_table);
809 powernow_table[data->acpi_data->state_count].frequency = CPUFREQ_TABLE_END;
810 powernow_table[data->acpi_data->state_count].index = 0;
811 data->powernow_table = powernow_table;
814 data->numps = data->acpi_data->state_count;
815 if (first_cpu(per_cpu(cpu_core_map, data->cpu)) == data->cpu)
817 powernow_k8_acpi_pst_values(data, 0);
819 /* notify BIOS that we exist */
820 acpi_processor_notify_smm(THIS_MODULE);
822 /* determine affinity, from ACPI if available */
823 if (preregister_valid) {
824 if ((data->acpi_data->shared_type == CPUFREQ_SHARED_TYPE_ALL) ||
825 (data->acpi_data->shared_type == CPUFREQ_SHARED_TYPE_ANY))
826 data->starting_core_affinity = data->acpi_data->shared_cpu_map;
828 data->starting_core_affinity = cpumask_of_cpu(data->cpu);
830 /* best guess from family if not */
831 if (cpu_family == CPU_HW_PSTATE)
832 data->starting_core_affinity = cpumask_of_cpu(data->cpu);
834 data->starting_core_affinity = per_cpu(cpu_core_map, data->cpu);
840 kfree(powernow_table);
843 acpi_processor_unregister_performance(data->acpi_data, data->cpu);
845 /* data->acpi_data.state_count informs us at ->exit() whether ACPI was used */
846 data->acpi_data->state_count = 0;
851 static int fill_powernow_table_pstate(struct powernow_k8_data *data, struct cpufreq_frequency_table *powernow_table)
855 rdmsr(MSR_PSTATE_CUR_LIMIT, hi, lo);
856 data->max_hw_pstate = (hi & HW_PSTATE_MAX_MASK) >> HW_PSTATE_MAX_SHIFT;
858 for (i = 0; i < data->acpi_data->state_count; i++) {
861 index = data->acpi_data->states[i].control & HW_PSTATE_MASK;
862 if (index > data->max_hw_pstate) {
863 printk(KERN_ERR PFX "invalid pstate %d - bad value %d.\n", i, index);
864 printk(KERN_ERR PFX "Please report to BIOS manufacturer\n");
865 powernow_table[i].frequency = CPUFREQ_ENTRY_INVALID;
868 rdmsr(MSR_PSTATE_DEF_BASE + index, lo, hi);
869 if (!(hi & HW_PSTATE_VALID_MASK)) {
870 dprintk("invalid pstate %d, ignoring\n", index);
871 powernow_table[i].frequency = CPUFREQ_ENTRY_INVALID;
875 powernow_table[i].index = index;
877 powernow_table[i].frequency = data->acpi_data->states[i].core_frequency * 1000;
882 static int fill_powernow_table_fidvid(struct powernow_k8_data *data, struct cpufreq_frequency_table *powernow_table)
886 for (i = 0; i < data->acpi_data->state_count; i++) {
891 fid = data->acpi_data->states[i].status & EXT_FID_MASK;
892 vid = (data->acpi_data->states[i].status >> VID_SHIFT) & EXT_VID_MASK;
894 fid = data->acpi_data->states[i].control & FID_MASK;
895 vid = (data->acpi_data->states[i].control >> VID_SHIFT) & VID_MASK;
898 dprintk(" %d : fid 0x%x, vid 0x%x\n", i, fid, vid);
900 powernow_table[i].index = fid; /* lower 8 bits */
901 powernow_table[i].index |= (vid << 8); /* upper 8 bits */
902 powernow_table[i].frequency = find_khz_freq_from_fid(fid);
904 /* verify frequency is OK */
905 if ((powernow_table[i].frequency > (MAX_FREQ * 1000)) ||
906 (powernow_table[i].frequency < (MIN_FREQ * 1000))) {
907 dprintk("invalid freq %u kHz, ignoring\n", powernow_table[i].frequency);
908 powernow_table[i].frequency = CPUFREQ_ENTRY_INVALID;
912 /* verify voltage is OK - BIOSs are using "off" to indicate invalid */
913 if (vid == VID_OFF) {
914 dprintk("invalid vid %u, ignoring\n", vid);
915 powernow_table[i].frequency = CPUFREQ_ENTRY_INVALID;
919 /* verify only 1 entry from the lo frequency table */
920 if (fid < HI_FID_TABLE_BOTTOM) {
922 /* if both entries are the same, ignore this one ... */
923 if ((powernow_table[i].frequency != powernow_table[cntlofreq].frequency) ||
924 (powernow_table[i].index != powernow_table[cntlofreq].index)) {
925 printk(KERN_ERR PFX "Too many lo freq table entries\n");
929 dprintk("double low frequency table entry, ignoring it.\n");
930 powernow_table[i].frequency = CPUFREQ_ENTRY_INVALID;
936 if (powernow_table[i].frequency != (data->acpi_data->states[i].core_frequency * 1000)) {
937 printk(KERN_INFO PFX "invalid freq entries %u kHz vs. %u kHz\n",
938 powernow_table[i].frequency,
939 (unsigned int) (data->acpi_data->states[i].core_frequency * 1000));
940 powernow_table[i].frequency = CPUFREQ_ENTRY_INVALID;
947 static void powernow_k8_cpu_exit_acpi(struct powernow_k8_data *data)
949 if (data->acpi_data->state_count)
950 acpi_processor_unregister_performance(data->acpi_data, data->cpu);
954 static int powernow_k8_cpu_preinit_acpi(void) { return -ENODEV; }
955 static int powernow_k8_cpu_init_acpi(struct powernow_k8_data *data) { return -ENODEV; }
956 static void powernow_k8_cpu_exit_acpi(struct powernow_k8_data *data) { return; }
957 static void powernow_k8_acpi_pst_values(struct powernow_k8_data *data, unsigned int index) { return; }
958 #endif /* CONFIG_X86_POWERNOW_K8_ACPI */
960 /* Take a frequency, and issue the fid/vid transition command */
961 static int transition_frequency_fidvid(struct powernow_k8_data *data, unsigned int index)
966 struct cpufreq_freqs freqs;
968 dprintk("cpu %d transition to index %u\n", smp_processor_id(), index);
970 /* fid/vid correctness check for k8 */
971 /* fid are the lower 8 bits of the index we stored into
972 * the cpufreq frequency table in find_psb_table, vid
973 * are the upper 8 bits.
975 fid = data->powernow_table[index].index & 0xFF;
976 vid = (data->powernow_table[index].index & 0xFF00) >> 8;
978 dprintk("table matched fid 0x%x, giving vid 0x%x\n", fid, vid);
980 if (query_current_values_with_pending_wait(data))
983 if ((data->currvid == vid) && (data->currfid == fid)) {
984 dprintk("target matches current values (fid 0x%x, vid 0x%x)\n",
989 if ((fid < HI_FID_TABLE_BOTTOM) && (data->currfid < HI_FID_TABLE_BOTTOM)) {
991 "ignoring illegal change in lo freq table-%x to 0x%x\n",
996 dprintk("cpu %d, changing to fid 0x%x, vid 0x%x\n",
997 smp_processor_id(), fid, vid);
998 freqs.old = find_khz_freq_from_fid(data->currfid);
999 freqs.new = find_khz_freq_from_fid(fid);
1001 for_each_cpu_mask_nr(i, *(data->available_cores)) {
1003 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
1006 res = transition_fid_vid(data, fid, vid);
1007 freqs.new = find_khz_freq_from_fid(data->currfid);
1009 for_each_cpu_mask_nr(i, *(data->available_cores)) {
1011 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
1016 /* Take a frequency, and issue the hardware pstate transition command */
1017 static int transition_frequency_pstate(struct powernow_k8_data *data, unsigned int index)
1021 struct cpufreq_freqs freqs;
1023 dprintk("cpu %d transition to index %u\n", smp_processor_id(), index);
1025 /* get MSR index for hardware pstate transition */
1026 pstate = index & HW_PSTATE_MASK;
1027 if (pstate > data->max_hw_pstate)
1029 freqs.old = find_khz_freq_from_pstate(data->powernow_table, data->currpstate);
1030 freqs.new = find_khz_freq_from_pstate(data->powernow_table, pstate);
1032 for_each_cpu_mask_nr(i, *(data->available_cores)) {
1034 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
1037 res = transition_pstate(data, pstate);
1038 freqs.new = find_khz_freq_from_pstate(data->powernow_table, pstate);
1040 for_each_cpu_mask_nr(i, *(data->available_cores)) {
1042 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
1047 /* Driver entry point to switch to the target frequency */
1048 static int powernowk8_target(struct cpufreq_policy *pol, unsigned targfreq, unsigned relation)
1051 struct powernow_k8_data *data = per_cpu(powernow_data, pol->cpu);
1054 unsigned int newstate;
1060 checkfid = data->currfid;
1061 checkvid = data->currvid;
1063 /* only run on specific CPU from here on */
1064 oldmask = current->cpus_allowed;
1065 set_cpus_allowed_ptr(current, &cpumask_of_cpu(pol->cpu));
1067 if (smp_processor_id() != pol->cpu) {
1068 printk(KERN_ERR PFX "limiting to cpu %u failed\n", pol->cpu);
1072 if (pending_bit_stuck()) {
1073 printk(KERN_ERR PFX "failing targ, change pending bit set\n");
1077 dprintk("targ: cpu %d, %d kHz, min %d, max %d, relation %d\n",
1078 pol->cpu, targfreq, pol->min, pol->max, relation);
1080 if (query_current_values_with_pending_wait(data))
1083 if (cpu_family != CPU_HW_PSTATE) {
1084 dprintk("targ: curr fid 0x%x, vid 0x%x\n",
1085 data->currfid, data->currvid);
1087 if ((checkvid != data->currvid) || (checkfid != data->currfid)) {
1088 printk(KERN_INFO PFX
1089 "error - out of sync, fix 0x%x 0x%x, vid 0x%x 0x%x\n",
1090 checkfid, data->currfid, checkvid, data->currvid);
1094 if (cpufreq_frequency_table_target(pol, data->powernow_table, targfreq, relation, &newstate))
1097 mutex_lock(&fidvid_mutex);
1099 powernow_k8_acpi_pst_values(data, newstate);
1101 if (cpu_family == CPU_HW_PSTATE)
1102 ret = transition_frequency_pstate(data, newstate);
1104 ret = transition_frequency_fidvid(data, newstate);
1106 printk(KERN_ERR PFX "transition frequency failed\n");
1108 mutex_unlock(&fidvid_mutex);
1111 mutex_unlock(&fidvid_mutex);
1113 if (cpu_family == CPU_HW_PSTATE)
1114 pol->cur = find_khz_freq_from_pstate(data->powernow_table, newstate);
1116 pol->cur = find_khz_freq_from_fid(data->currfid);
1120 set_cpus_allowed_ptr(current, &oldmask);
1124 /* Driver entry point to verify the policy and range of frequencies */
1125 static int powernowk8_verify(struct cpufreq_policy *pol)
1127 struct powernow_k8_data *data = per_cpu(powernow_data, pol->cpu);
1132 return cpufreq_frequency_table_verify(pol, data->powernow_table);
1135 /* per CPU init entry point to the driver */
1136 static int __cpuinit powernowk8_cpu_init(struct cpufreq_policy *pol)
1138 struct powernow_k8_data *data;
1139 cpumask_t oldmask = CPU_MASK_ALL;
1142 if (!cpu_online(pol->cpu))
1145 if (!check_supported_cpu(pol->cpu))
1148 data = kzalloc(sizeof(struct powernow_k8_data), GFP_KERNEL);
1150 printk(KERN_ERR PFX "unable to alloc powernow_k8_data");
1154 data->cpu = pol->cpu;
1156 if (powernow_k8_cpu_init_acpi(data)) {
1158 * Use the PSB BIOS structure. This is only availabe on
1159 * an UP version, and is deprecated by AMD.
1161 if (num_online_cpus() != 1) {
1162 #ifndef CONFIG_ACPI_PROCESSOR
1163 printk(KERN_ERR PFX "ACPI Processor support is required "
1164 "for SMP systems but is absent. Please load the "
1165 "ACPI Processor module before starting this "
1168 printk(KERN_ERR PFX "Your BIOS does not provide ACPI "
1169 "_PSS objects in a way that Linux understands. "
1170 "Please report this to the Linux ACPI maintainers"
1171 " and complain to your BIOS vendor.\n");
1176 if (pol->cpu != 0) {
1177 printk(KERN_ERR PFX "No ACPI _PSS objects for CPU other than "
1178 "CPU0. Complain to your BIOS vendor.\n");
1182 rc = find_psb_table(data);
1189 /* only run on specific CPU from here on */
1190 oldmask = current->cpus_allowed;
1191 set_cpus_allowed_ptr(current, &cpumask_of_cpu(pol->cpu));
1193 if (smp_processor_id() != pol->cpu) {
1194 printk(KERN_ERR PFX "limiting to cpu %u failed\n", pol->cpu);
1198 if (pending_bit_stuck()) {
1199 printk(KERN_ERR PFX "failing init, change pending bit set\n");
1203 if (query_current_values_with_pending_wait(data))
1206 if (cpu_family == CPU_OPTERON)
1209 /* run on any CPU again */
1210 set_cpus_allowed_ptr(current, &oldmask);
1212 pol->cpus = data->starting_core_affinity;
1213 data->available_cores = &(pol->cpus);
1215 /* Take a crude guess here.
1216 * That guess was in microseconds, so multiply with 1000 */
1217 pol->cpuinfo.transition_latency = (((data->rvo + 8) * data->vstable * VST_UNITS_20US)
1218 + (3 * (1 << data->irt) * 10)) * 1000;
1220 if (cpu_family == CPU_HW_PSTATE)
1221 pol->cur = find_khz_freq_from_pstate(data->powernow_table, data->currpstate);
1223 pol->cur = find_khz_freq_from_fid(data->currfid);
1224 dprintk("policy current frequency %d kHz\n", pol->cur);
1226 /* min/max the cpu is capable of */
1227 if (cpufreq_frequency_table_cpuinfo(pol, data->powernow_table)) {
1228 printk(KERN_ERR PFX "invalid powernow_table\n");
1229 powernow_k8_cpu_exit_acpi(data);
1230 kfree(data->powernow_table);
1235 cpufreq_frequency_table_get_attr(data->powernow_table, pol->cpu);
1237 if (cpu_family == CPU_HW_PSTATE)
1238 dprintk("cpu_init done, current pstate 0x%x\n", data->currpstate);
1240 dprintk("cpu_init done, current fid 0x%x, vid 0x%x\n",
1241 data->currfid, data->currvid);
1243 per_cpu(powernow_data, pol->cpu) = data;
1248 set_cpus_allowed_ptr(current, &oldmask);
1249 powernow_k8_cpu_exit_acpi(data);
1255 static int __devexit powernowk8_cpu_exit (struct cpufreq_policy *pol)
1257 struct powernow_k8_data *data = per_cpu(powernow_data, pol->cpu);
1262 powernow_k8_cpu_exit_acpi(data);
1264 cpufreq_frequency_table_put_attr(pol->cpu);
1266 kfree(data->powernow_table);
1272 static unsigned int powernowk8_get (unsigned int cpu)
1274 struct powernow_k8_data *data;
1275 cpumask_t oldmask = current->cpus_allowed;
1276 unsigned int khz = 0;
1279 first = first_cpu(per_cpu(cpu_core_map, cpu));
1280 data = per_cpu(powernow_data, first);
1285 set_cpus_allowed_ptr(current, &cpumask_of_cpu(cpu));
1286 if (smp_processor_id() != cpu) {
1288 "limiting to CPU %d failed in powernowk8_get\n", cpu);
1289 set_cpus_allowed_ptr(current, &oldmask);
1293 if (query_current_values_with_pending_wait(data))
1296 if (cpu_family == CPU_HW_PSTATE)
1297 khz = find_khz_freq_from_pstate(data->powernow_table,
1300 khz = find_khz_freq_from_fid(data->currfid);
1304 set_cpus_allowed_ptr(current, &oldmask);
1308 static struct freq_attr* powernow_k8_attr[] = {
1309 &cpufreq_freq_attr_scaling_available_freqs,
1313 static struct cpufreq_driver cpufreq_amd64_driver = {
1314 .verify = powernowk8_verify,
1315 .target = powernowk8_target,
1316 .init = powernowk8_cpu_init,
1317 .exit = __devexit_p(powernowk8_cpu_exit),
1318 .get = powernowk8_get,
1319 .name = "powernow-k8",
1320 .owner = THIS_MODULE,
1321 .attr = powernow_k8_attr,
1324 /* driver entry point for init */
1325 static int __cpuinit powernowk8_init(void)
1327 unsigned int i, supported_cpus = 0;
1329 for_each_online_cpu(i) {
1330 if (check_supported_cpu(i))
1334 if (supported_cpus == num_online_cpus()) {
1335 powernow_k8_cpu_preinit_acpi();
1336 printk(KERN_INFO PFX "Found %d %s "
1337 "processors (%d cpu cores) (" VERSION ")\n",
1339 boot_cpu_data.x86_model_id, supported_cpus);
1340 return cpufreq_register_driver(&cpufreq_amd64_driver);
1346 /* driver entry point for term */
1347 static void __exit powernowk8_exit(void)
1351 cpufreq_unregister_driver(&cpufreq_amd64_driver);
1353 #ifdef CONFIG_X86_POWERNOW_K8_ACPI
1354 free_percpu(acpi_perf_data);
1358 MODULE_AUTHOR("Paul Devriendt <paul.devriendt@amd.com> and Mark Langsdorf <mark.langsdorf@amd.com>");
1359 MODULE_DESCRIPTION("AMD Athlon 64 and Opteron processor frequency driver.");
1360 MODULE_LICENSE("GPL");
1362 late_initcall(powernowk8_init);
1363 module_exit(powernowk8_exit);