3 * (c) 2003-2006 Advanced Micro Devices, Inc.
4 * Your use of this code is subject to the terms and conditions of the
5 * GNU general public license version 2. See "COPYING" or
6 * http://www.gnu.org/licenses/gpl.html
8 * Support : mark.langsdorf@amd.com
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@suse.cz>
14 * Licensed under the terms of the GNU GPL License version 2.
15 * Based upon datasheets & sample CPUs kindly provided by AMD.
17 * Valuable input gratefully received from Dave Jones, Pavel Machek,
18 * Dominik Brodowski, Jacob Shin, and others.
19 * Originally developed by Paul Devriendt.
20 * Processor information obtained from Chapter 9 (Power and Thermal Management)
21 * of the "BIOS and Kernel Developer's Guide for the AMD Athlon 64 and AMD
22 * Opteron Processors" available for download from www.amd.com
24 * Tables for specific CPUs can be inferred from
25 * http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/30430.pdf
28 #include <linux/kernel.h>
29 #include <linux/smp.h>
30 #include <linux/module.h>
31 #include <linux/init.h>
32 #include <linux/cpufreq.h>
33 #include <linux/slab.h>
34 #include <linux/string.h>
35 #include <linux/cpumask.h>
36 #include <linux/sched.h> /* for current / set_cpus_allowed() */
38 #include <linux/delay.h>
42 #include <linux/acpi.h>
43 #include <linux/mutex.h>
44 #include <acpi/processor.h>
46 #define PFX "powernow-k8: "
47 #define VERSION "version 2.20.00"
48 #include "powernow-k8.h"
50 /* serialize freq changes */
51 static DEFINE_MUTEX(fidvid_mutex);
53 static DEFINE_PER_CPU(struct powernow_k8_data *, powernow_data);
55 static int cpu_family = CPU_OPTERON;
58 static inline const struct cpumask *cpu_core_mask(int cpu)
64 /* Return a frequency in MHz, given an input fid */
65 static u32 find_freq_from_fid(u32 fid)
67 return 800 + (fid * 100);
70 /* Return a frequency in KHz, given an input fid */
71 static u32 find_khz_freq_from_fid(u32 fid)
73 return 1000 * find_freq_from_fid(fid);
76 static u32 find_khz_freq_from_pstate(struct cpufreq_frequency_table *data,
79 return data[pstate].frequency;
82 /* Return the vco fid for an input fid
84 * Each "low" fid has corresponding "high" fid, and you can get to "low" fids
85 * only from corresponding high fids. This returns "high" fid corresponding to
88 static u32 convert_fid_to_vco_fid(u32 fid)
90 if (fid < HI_FID_TABLE_BOTTOM)
97 * Return 1 if the pending bit is set. Unless we just instructed the processor
98 * to transition to a new state, seeing this bit set is really bad news.
100 static int pending_bit_stuck(void)
104 if (cpu_family == CPU_HW_PSTATE)
107 rdmsr(MSR_FIDVID_STATUS, lo, hi);
108 return lo & MSR_S_LO_CHANGE_PENDING ? 1 : 0;
112 * Update the global current fid / vid values from the status msr.
113 * Returns 1 on error.
115 static int query_current_values_with_pending_wait(struct powernow_k8_data *data)
120 if (cpu_family == CPU_HW_PSTATE) {
121 rdmsr(MSR_PSTATE_STATUS, lo, hi);
122 i = lo & HW_PSTATE_MASK;
123 data->currpstate = i;
126 * a workaround for family 11h erratum 311 might cause
127 * an "out-of-range Pstate if the core is in Pstate-0
129 if ((boot_cpu_data.x86 == 0x11) && (i >= data->numps))
130 data->currpstate = HW_PSTATE_0;
136 dprintk("detected change pending stuck\n");
139 rdmsr(MSR_FIDVID_STATUS, lo, hi);
140 } while (lo & MSR_S_LO_CHANGE_PENDING);
142 data->currvid = hi & MSR_S_HI_CURRENT_VID;
143 data->currfid = lo & MSR_S_LO_CURRENT_FID;
148 /* the isochronous relief time */
149 static void count_off_irt(struct powernow_k8_data *data)
151 udelay((1 << data->irt) * 10);
155 /* the voltage stabilization time */
156 static void count_off_vst(struct powernow_k8_data *data)
158 udelay(data->vstable * VST_UNITS_20US);
162 /* need to init the control msr to a safe value (for each cpu) */
163 static void fidvid_msr_init(void)
168 rdmsr(MSR_FIDVID_STATUS, lo, hi);
169 vid = hi & MSR_S_HI_CURRENT_VID;
170 fid = lo & MSR_S_LO_CURRENT_FID;
171 lo = fid | (vid << MSR_C_LO_VID_SHIFT);
172 hi = MSR_C_HI_STP_GNT_BENIGN;
173 dprintk("cpu%d, init lo 0x%x, hi 0x%x\n", smp_processor_id(), lo, hi);
174 wrmsr(MSR_FIDVID_CTL, lo, hi);
177 /* write the new fid value along with the other control fields to the msr */
178 static int write_new_fid(struct powernow_k8_data *data, u32 fid)
181 u32 savevid = data->currvid;
184 if ((fid & INVALID_FID_MASK) || (data->currvid & INVALID_VID_MASK)) {
185 printk(KERN_ERR PFX "internal error - overflow on fid write\n");
190 lo |= (data->currvid << MSR_C_LO_VID_SHIFT);
191 lo |= MSR_C_LO_INIT_FID_VID;
193 dprintk("writing fid 0x%x, lo 0x%x, hi 0x%x\n",
194 fid, lo, data->plllock * PLL_LOCK_CONVERSION);
197 wrmsr(MSR_FIDVID_CTL, lo, data->plllock * PLL_LOCK_CONVERSION);
200 "Hardware error - pending bit very stuck - "
201 "no further pstate changes possible\n");
204 } while (query_current_values_with_pending_wait(data));
208 if (savevid != data->currvid) {
210 "vid change on fid trans, old 0x%x, new 0x%x\n",
211 savevid, data->currvid);
215 if (fid != data->currfid) {
217 "fid trans failed, fid 0x%x, curr 0x%x\n", fid,
225 /* Write a new vid to the hardware */
226 static int write_new_vid(struct powernow_k8_data *data, u32 vid)
229 u32 savefid = data->currfid;
232 if ((data->currfid & INVALID_FID_MASK) || (vid & INVALID_VID_MASK)) {
233 printk(KERN_ERR PFX "internal error - overflow on vid write\n");
238 lo |= (vid << MSR_C_LO_VID_SHIFT);
239 lo |= MSR_C_LO_INIT_FID_VID;
241 dprintk("writing vid 0x%x, lo 0x%x, hi 0x%x\n",
242 vid, lo, STOP_GRANT_5NS);
245 wrmsr(MSR_FIDVID_CTL, lo, STOP_GRANT_5NS);
247 printk(KERN_ERR PFX "internal error - pending bit "
248 "very stuck - no further pstate "
249 "changes possible\n");
252 } while (query_current_values_with_pending_wait(data));
254 if (savefid != data->currfid) {
255 printk(KERN_ERR PFX "fid changed on vid trans, old "
257 savefid, data->currfid);
261 if (vid != data->currvid) {
262 printk(KERN_ERR PFX "vid trans failed, vid 0x%x, "
272 * Reduce the vid by the max of step or reqvid.
273 * Decreasing vid codes represent increasing voltages:
274 * vid of 0 is 1.550V, vid of 0x1e is 0.800V, vid of VID_OFF is off.
276 static int decrease_vid_code_by_step(struct powernow_k8_data *data,
277 u32 reqvid, u32 step)
279 if ((data->currvid - reqvid) > step)
280 reqvid = data->currvid - step;
282 if (write_new_vid(data, reqvid))
290 /* Change hardware pstate by single MSR write */
291 static int transition_pstate(struct powernow_k8_data *data, u32 pstate)
293 wrmsr(MSR_PSTATE_CTRL, pstate, 0);
294 data->currpstate = pstate;
298 /* Change Opteron/Athlon64 fid and vid, by the 3 phases. */
299 static int transition_fid_vid(struct powernow_k8_data *data,
300 u32 reqfid, u32 reqvid)
302 if (core_voltage_pre_transition(data, reqvid, reqfid))
305 if (core_frequency_transition(data, reqfid))
308 if (core_voltage_post_transition(data, reqvid))
311 if (query_current_values_with_pending_wait(data))
314 if ((reqfid != data->currfid) || (reqvid != data->currvid)) {
315 printk(KERN_ERR PFX "failed (cpu%d): req 0x%x 0x%x, "
318 reqfid, reqvid, data->currfid, data->currvid);
322 dprintk("transitioned (cpu%d): new fid 0x%x, vid 0x%x\n",
323 smp_processor_id(), data->currfid, data->currvid);
328 /* Phase 1 - core voltage transition ... setup voltage */
329 static int core_voltage_pre_transition(struct powernow_k8_data *data,
330 u32 reqvid, u32 reqfid)
332 u32 rvosteps = data->rvo;
333 u32 savefid = data->currfid;
334 u32 maxvid, lo, rvomult = 1;
336 dprintk("ph1 (cpu%d): start, currfid 0x%x, currvid 0x%x, "
337 "reqvid 0x%x, rvo 0x%x\n",
339 data->currfid, data->currvid, reqvid, data->rvo);
341 if ((savefid < LO_FID_TABLE_TOP) && (reqfid < LO_FID_TABLE_TOP))
344 rdmsr(MSR_FIDVID_STATUS, lo, maxvid);
345 maxvid = 0x1f & (maxvid >> 16);
346 dprintk("ph1 maxvid=0x%x\n", maxvid);
347 if (reqvid < maxvid) /* lower numbers are higher voltages */
350 while (data->currvid > reqvid) {
351 dprintk("ph1: curr 0x%x, req vid 0x%x\n",
352 data->currvid, reqvid);
353 if (decrease_vid_code_by_step(data, reqvid, data->vidmvs))
357 while ((rvosteps > 0) &&
358 ((rvomult * data->rvo + data->currvid) > reqvid)) {
359 if (data->currvid == maxvid) {
362 dprintk("ph1: changing vid for rvo, req 0x%x\n",
364 if (decrease_vid_code_by_step(data, data->currvid-1, 1))
370 if (query_current_values_with_pending_wait(data))
373 if (savefid != data->currfid) {
374 printk(KERN_ERR PFX "ph1 err, currfid changed 0x%x\n",
379 dprintk("ph1 complete, currfid 0x%x, currvid 0x%x\n",
380 data->currfid, data->currvid);
385 /* Phase 2 - core frequency transition */
386 static int core_frequency_transition(struct powernow_k8_data *data, u32 reqfid)
388 u32 vcoreqfid, vcocurrfid, vcofiddiff;
389 u32 fid_interval, savevid = data->currvid;
391 if (data->currfid == reqfid) {
392 printk(KERN_ERR PFX "ph2 null fid transition 0x%x\n",
397 dprintk("ph2 (cpu%d): starting, currfid 0x%x, currvid 0x%x, "
400 data->currfid, data->currvid, reqfid);
402 vcoreqfid = convert_fid_to_vco_fid(reqfid);
403 vcocurrfid = convert_fid_to_vco_fid(data->currfid);
404 vcofiddiff = vcocurrfid > vcoreqfid ? vcocurrfid - vcoreqfid
405 : vcoreqfid - vcocurrfid;
407 if ((reqfid <= LO_FID_TABLE_TOP) && (data->currfid <= LO_FID_TABLE_TOP))
410 while (vcofiddiff > 2) {
411 (data->currfid & 1) ? (fid_interval = 1) : (fid_interval = 2);
413 if (reqfid > data->currfid) {
414 if (data->currfid > LO_FID_TABLE_TOP) {
415 if (write_new_fid(data,
416 data->currfid + fid_interval))
421 2 + convert_fid_to_vco_fid(data->currfid)))
425 if (write_new_fid(data, data->currfid - fid_interval))
429 vcocurrfid = convert_fid_to_vco_fid(data->currfid);
430 vcofiddiff = vcocurrfid > vcoreqfid ? vcocurrfid - vcoreqfid
431 : vcoreqfid - vcocurrfid;
434 if (write_new_fid(data, reqfid))
437 if (query_current_values_with_pending_wait(data))
440 if (data->currfid != reqfid) {
442 "ph2: mismatch, failed fid transition, "
443 "curr 0x%x, req 0x%x\n",
444 data->currfid, reqfid);
448 if (savevid != data->currvid) {
449 printk(KERN_ERR PFX "ph2: vid changed, save 0x%x, curr 0x%x\n",
450 savevid, data->currvid);
454 dprintk("ph2 complete, currfid 0x%x, currvid 0x%x\n",
455 data->currfid, data->currvid);
460 /* Phase 3 - core voltage transition flow ... jump to the final vid. */
461 static int core_voltage_post_transition(struct powernow_k8_data *data,
464 u32 savefid = data->currfid;
465 u32 savereqvid = reqvid;
467 dprintk("ph3 (cpu%d): starting, currfid 0x%x, currvid 0x%x\n",
469 data->currfid, data->currvid);
471 if (reqvid != data->currvid) {
472 if (write_new_vid(data, reqvid))
475 if (savefid != data->currfid) {
477 "ph3: bad fid change, save 0x%x, curr 0x%x\n",
478 savefid, data->currfid);
482 if (data->currvid != reqvid) {
484 "ph3: failed vid transition\n, "
485 "req 0x%x, curr 0x%x",
486 reqvid, data->currvid);
491 if (query_current_values_with_pending_wait(data))
494 if (savereqvid != data->currvid) {
495 dprintk("ph3 failed, currvid 0x%x\n", data->currvid);
499 if (savefid != data->currfid) {
500 dprintk("ph3 failed, currfid changed 0x%x\n",
505 dprintk("ph3 complete, currfid 0x%x, currvid 0x%x\n",
506 data->currfid, data->currvid);
511 static void check_supported_cpu(void *_rc)
513 u32 eax, ebx, ecx, edx;
518 if (current_cpu_data.x86_vendor != X86_VENDOR_AMD)
521 eax = cpuid_eax(CPUID_PROCESSOR_SIGNATURE);
522 if (((eax & CPUID_XFAM) != CPUID_XFAM_K8) &&
523 ((eax & CPUID_XFAM) < CPUID_XFAM_10H))
526 if ((eax & CPUID_XFAM) == CPUID_XFAM_K8) {
527 if (((eax & CPUID_USE_XFAM_XMOD) != CPUID_USE_XFAM_XMOD) ||
528 ((eax & CPUID_XMOD) > CPUID_XMOD_REV_MASK)) {
530 "Processor cpuid %x not supported\n", eax);
534 eax = cpuid_eax(CPUID_GET_MAX_CAPABILITIES);
535 if (eax < CPUID_FREQ_VOLT_CAPABILITIES) {
537 "No frequency change capabilities detected\n");
541 cpuid(CPUID_FREQ_VOLT_CAPABILITIES, &eax, &ebx, &ecx, &edx);
542 if ((edx & P_STATE_TRANSITION_CAPABLE)
543 != P_STATE_TRANSITION_CAPABLE) {
545 "Power state transitions not supported\n");
548 } else { /* must be a HW Pstate capable processor */
549 cpuid(CPUID_FREQ_VOLT_CAPABILITIES, &eax, &ebx, &ecx, &edx);
550 if ((edx & USE_HW_PSTATE) == USE_HW_PSTATE)
551 cpu_family = CPU_HW_PSTATE;
559 static int check_pst_table(struct powernow_k8_data *data, struct pst_s *pst,
565 for (j = 0; j < data->numps; j++) {
566 if (pst[j].vid > LEAST_VID) {
567 printk(KERN_ERR FW_BUG PFX "vid %d invalid : 0x%x\n",
571 if (pst[j].vid < data->rvo) {
573 printk(KERN_ERR FW_BUG PFX "0 vid exceeded with pstate"
577 if (pst[j].vid < maxvid + data->rvo) {
578 /* vid + rvo >= maxvid */
579 printk(KERN_ERR FW_BUG PFX "maxvid exceeded with pstate"
583 if (pst[j].fid > MAX_FID) {
584 printk(KERN_ERR FW_BUG PFX "maxfid exceeded with pstate"
588 if (j && (pst[j].fid < HI_FID_TABLE_BOTTOM)) {
589 /* Only first fid is allowed to be in "low" range */
590 printk(KERN_ERR FW_BUG PFX "two low fids - %d : "
591 "0x%x\n", j, pst[j].fid);
594 if (pst[j].fid < lastfid)
595 lastfid = pst[j].fid;
598 printk(KERN_ERR FW_BUG PFX "lastfid invalid\n");
601 if (lastfid > LO_FID_TABLE_TOP)
602 printk(KERN_INFO FW_BUG PFX
603 "first fid not from lo freq table\n");
608 static void invalidate_entry(struct cpufreq_frequency_table *powernow_table,
611 powernow_table[entry].frequency = CPUFREQ_ENTRY_INVALID;
614 static void print_basics(struct powernow_k8_data *data)
617 for (j = 0; j < data->numps; j++) {
618 if (data->powernow_table[j].frequency !=
619 CPUFREQ_ENTRY_INVALID) {
620 if (cpu_family == CPU_HW_PSTATE) {
622 " %d : pstate %d (%d MHz)\n", j,
623 data->powernow_table[j].index,
624 data->powernow_table[j].frequency/1000);
627 " %d : fid 0x%x (%d MHz), vid 0x%x\n",
629 data->powernow_table[j].index & 0xff,
630 data->powernow_table[j].frequency/1000,
631 data->powernow_table[j].index >> 8);
636 printk(KERN_INFO PFX "Only %d pstates on battery\n",
640 static u32 freq_from_fid_did(u32 fid, u32 did)
644 if (boot_cpu_data.x86 == 0x10)
645 mhz = (100 * (fid + 0x10)) >> did;
646 else if (boot_cpu_data.x86 == 0x11)
647 mhz = (100 * (fid + 8)) >> did;
654 static int fill_powernow_table(struct powernow_k8_data *data,
655 struct pst_s *pst, u8 maxvid)
657 struct cpufreq_frequency_table *powernow_table;
661 /* use ACPI support to get full speed on mains power */
662 printk(KERN_WARNING PFX
663 "Only %d pstates usable (use ACPI driver for full "
664 "range\n", data->batps);
665 data->numps = data->batps;
668 for (j = 1; j < data->numps; j++) {
669 if (pst[j-1].fid >= pst[j].fid) {
670 printk(KERN_ERR PFX "PST out of sequence\n");
675 if (data->numps < 2) {
676 printk(KERN_ERR PFX "no p states to transition\n");
680 if (check_pst_table(data, pst, maxvid))
683 powernow_table = kmalloc((sizeof(struct cpufreq_frequency_table)
684 * (data->numps + 1)), GFP_KERNEL);
685 if (!powernow_table) {
686 printk(KERN_ERR PFX "powernow_table memory alloc failure\n");
690 for (j = 0; j < data->numps; j++) {
692 powernow_table[j].index = pst[j].fid; /* lower 8 bits */
693 powernow_table[j].index |= (pst[j].vid << 8); /* upper 8 bits */
694 freq = find_khz_freq_from_fid(pst[j].fid);
695 powernow_table[j].frequency = freq;
697 powernow_table[data->numps].frequency = CPUFREQ_TABLE_END;
698 powernow_table[data->numps].index = 0;
700 if (query_current_values_with_pending_wait(data)) {
701 kfree(powernow_table);
705 dprintk("cfid 0x%x, cvid 0x%x\n", data->currfid, data->currvid);
706 data->powernow_table = powernow_table;
707 if (cpumask_first(cpu_core_mask(data->cpu)) == data->cpu)
710 for (j = 0; j < data->numps; j++)
711 if ((pst[j].fid == data->currfid) &&
712 (pst[j].vid == data->currvid))
715 dprintk("currfid/vid do not match PST, ignoring\n");
719 /* Find and validate the PSB/PST table in BIOS. */
720 static int find_psb_table(struct powernow_k8_data *data)
729 for (i = 0xc0000; i < 0xffff0; i += 0x10) {
730 /* Scan BIOS looking for the signature. */
731 /* It can not be at ffff0 - it is too big. */
733 psb = phys_to_virt(i);
734 if (memcmp(psb, PSB_ID_STRING, PSB_ID_STRING_LEN) != 0)
737 dprintk("found PSB header at 0x%p\n", psb);
739 dprintk("table vers: 0x%x\n", psb->tableversion);
740 if (psb->tableversion != PSB_VERSION_1_4) {
741 printk(KERN_ERR FW_BUG PFX "PSB table is not v1.4\n");
745 dprintk("flags: 0x%x\n", psb->flags1);
747 printk(KERN_ERR FW_BUG PFX "unknown flags\n");
751 data->vstable = psb->vstable;
752 dprintk("voltage stabilization time: %d(*20us)\n",
755 dprintk("flags2: 0x%x\n", psb->flags2);
756 data->rvo = psb->flags2 & 3;
757 data->irt = ((psb->flags2) >> 2) & 3;
758 mvs = ((psb->flags2) >> 4) & 3;
759 data->vidmvs = 1 << mvs;
760 data->batps = ((psb->flags2) >> 6) & 3;
762 dprintk("ramp voltage offset: %d\n", data->rvo);
763 dprintk("isochronous relief time: %d\n", data->irt);
764 dprintk("maximum voltage step: %d - 0x%x\n", mvs, data->vidmvs);
766 dprintk("numpst: 0x%x\n", psb->num_tables);
767 cpst = psb->num_tables;
768 if ((psb->cpuid == 0x00000fc0) ||
769 (psb->cpuid == 0x00000fe0)) {
770 thiscpuid = cpuid_eax(CPUID_PROCESSOR_SIGNATURE);
771 if ((thiscpuid == 0x00000fc0) ||
772 (thiscpuid == 0x00000fe0))
776 printk(KERN_ERR FW_BUG PFX "numpst must be 1\n");
780 data->plllock = psb->plllocktime;
781 dprintk("plllocktime: 0x%x (units 1us)\n", psb->plllocktime);
782 dprintk("maxfid: 0x%x\n", psb->maxfid);
783 dprintk("maxvid: 0x%x\n", psb->maxvid);
784 maxvid = psb->maxvid;
786 data->numps = psb->numps;
787 dprintk("numpstates: 0x%x\n", data->numps);
788 return fill_powernow_table(data,
789 (struct pst_s *)(psb+1), maxvid);
792 * If you see this message, complain to BIOS manufacturer. If
793 * he tells you "we do not support Linux" or some similar
794 * nonsense, remember that Windows 2000 uses the same legacy
795 * mechanism that the old Linux PSB driver uses. Tell them it
796 * is broken with Windows 2000.
798 * The reference to the AMD documentation is chapter 9 in the
799 * BIOS and Kernel Developer's Guide, which is available on
802 printk(KERN_ERR FW_BUG PFX "No PSB or ACPI _PSS objects\n");
806 static void powernow_k8_acpi_pst_values(struct powernow_k8_data *data,
809 acpi_integer control;
811 if (!data->acpi_data.state_count || (cpu_family == CPU_HW_PSTATE))
814 control = data->acpi_data.states[index].control;
815 data->irt = (control >> IRT_SHIFT) & IRT_MASK;
816 data->rvo = (control >> RVO_SHIFT) & RVO_MASK;
817 data->exttype = (control >> EXT_TYPE_SHIFT) & EXT_TYPE_MASK;
818 data->plllock = (control >> PLL_L_SHIFT) & PLL_L_MASK;
819 data->vidmvs = 1 << ((control >> MVS_SHIFT) & MVS_MASK);
820 data->vstable = (control >> VST_SHIFT) & VST_MASK;
823 static int powernow_k8_cpu_init_acpi(struct powernow_k8_data *data)
825 struct cpufreq_frequency_table *powernow_table;
826 int ret_val = -ENODEV;
827 acpi_integer control, status;
829 if (acpi_processor_register_performance(&data->acpi_data, data->cpu)) {
830 dprintk("register performance failed: bad ACPI data\n");
834 /* verify the data contained in the ACPI structures */
835 if (data->acpi_data.state_count <= 1) {
836 dprintk("No ACPI P-States\n");
840 control = data->acpi_data.control_register.space_id;
841 status = data->acpi_data.status_register.space_id;
843 if ((control != ACPI_ADR_SPACE_FIXED_HARDWARE) ||
844 (status != ACPI_ADR_SPACE_FIXED_HARDWARE)) {
845 dprintk("Invalid control/status registers (%x - %x)\n",
850 /* fill in data->powernow_table */
851 powernow_table = kmalloc((sizeof(struct cpufreq_frequency_table)
852 * (data->acpi_data.state_count + 1)), GFP_KERNEL);
853 if (!powernow_table) {
854 dprintk("powernow_table memory alloc failure\n");
859 data->numps = data->acpi_data.state_count;
860 powernow_k8_acpi_pst_values(data, 0);
862 if (cpu_family == CPU_HW_PSTATE)
863 ret_val = fill_powernow_table_pstate(data, powernow_table);
865 ret_val = fill_powernow_table_fidvid(data, powernow_table);
869 powernow_table[data->acpi_data.state_count].frequency =
871 powernow_table[data->acpi_data.state_count].index = 0;
872 data->powernow_table = powernow_table;
874 if (cpumask_first(cpu_core_mask(data->cpu)) == data->cpu)
877 /* notify BIOS that we exist */
878 acpi_processor_notify_smm(THIS_MODULE);
880 if (!zalloc_cpumask_var(&data->acpi_data.shared_cpu_map, GFP_KERNEL)) {
882 "unable to alloc powernow_k8_data cpumask\n");
890 kfree(powernow_table);
893 acpi_processor_unregister_performance(&data->acpi_data, data->cpu);
895 /* data->acpi_data.state_count informs us at ->exit()
896 * whether ACPI was used */
897 data->acpi_data.state_count = 0;
902 static int fill_powernow_table_pstate(struct powernow_k8_data *data,
903 struct cpufreq_frequency_table *powernow_table)
907 rdmsr(MSR_PSTATE_CUR_LIMIT, hi, lo);
908 data->max_hw_pstate = (hi & HW_PSTATE_MAX_MASK) >> HW_PSTATE_MAX_SHIFT;
910 for (i = 0; i < data->acpi_data.state_count; i++) {
913 index = data->acpi_data.states[i].control & HW_PSTATE_MASK;
914 if (index > data->max_hw_pstate) {
915 printk(KERN_ERR PFX "invalid pstate %d - "
916 "bad value %d.\n", i, index);
917 printk(KERN_ERR PFX "Please report to BIOS "
919 invalidate_entry(powernow_table, i);
922 rdmsr(MSR_PSTATE_DEF_BASE + index, lo, hi);
923 if (!(hi & HW_PSTATE_VALID_MASK)) {
924 dprintk("invalid pstate %d, ignoring\n", index);
925 invalidate_entry(powernow_table, i);
929 powernow_table[i].index = index;
931 /* Frequency may be rounded for these */
932 if (boot_cpu_data.x86 == 0x10 || boot_cpu_data.x86 == 0x11) {
933 powernow_table[i].frequency =
934 freq_from_fid_did(lo & 0x3f, (lo >> 6) & 7);
936 powernow_table[i].frequency =
937 data->acpi_data.states[i].core_frequency * 1000;
942 static int fill_powernow_table_fidvid(struct powernow_k8_data *data,
943 struct cpufreq_frequency_table *powernow_table)
947 for (i = 0; i < data->acpi_data.state_count; i++) {
951 acpi_integer status, control;
954 status = data->acpi_data.states[i].status;
955 fid = status & EXT_FID_MASK;
956 vid = (status >> VID_SHIFT) & EXT_VID_MASK;
958 control = data->acpi_data.states[i].control;
959 fid = control & FID_MASK;
960 vid = (control >> VID_SHIFT) & VID_MASK;
963 dprintk(" %d : fid 0x%x, vid 0x%x\n", i, fid, vid);
965 index = fid | (vid<<8);
966 powernow_table[i].index = index;
968 freq = find_khz_freq_from_fid(fid);
969 powernow_table[i].frequency = freq;
971 /* verify frequency is OK */
972 if ((freq > (MAX_FREQ * 1000)) || (freq < (MIN_FREQ * 1000))) {
973 dprintk("invalid freq %u kHz, ignoring\n", freq);
974 invalidate_entry(powernow_table, i);
978 /* verify voltage is OK -
979 * BIOSs are using "off" to indicate invalid */
980 if (vid == VID_OFF) {
981 dprintk("invalid vid %u, ignoring\n", vid);
982 invalidate_entry(powernow_table, i);
986 if (freq != (data->acpi_data.states[i].core_frequency * 1000)) {
987 printk(KERN_INFO PFX "invalid freq entries "
988 "%u kHz vs. %u kHz\n", freq,
990 (data->acpi_data.states[i].core_frequency
992 invalidate_entry(powernow_table, i);
999 static void powernow_k8_cpu_exit_acpi(struct powernow_k8_data *data)
1001 if (data->acpi_data.state_count)
1002 acpi_processor_unregister_performance(&data->acpi_data,
1004 free_cpumask_var(data->acpi_data.shared_cpu_map);
1007 static int get_transition_latency(struct powernow_k8_data *data)
1009 int max_latency = 0;
1011 for (i = 0; i < data->acpi_data.state_count; i++) {
1012 int cur_latency = data->acpi_data.states[i].transition_latency
1013 + data->acpi_data.states[i].bus_master_latency;
1014 if (cur_latency > max_latency)
1015 max_latency = cur_latency;
1017 if (max_latency == 0) {
1019 * Fam 11h always returns 0 as transition latency.
1020 * This is intended and means "very fast". While cpufreq core
1021 * and governors currently can handle that gracefully, better
1022 * set it to 1 to avoid problems in the future.
1023 * For all others it's a BIOS bug.
1025 if (boot_cpu_data.x86 != 0x11)
1026 printk(KERN_ERR FW_WARN PFX "Invalid zero transition "
1030 /* value in usecs, needs to be in nanoseconds */
1031 return 1000 * max_latency;
1034 /* Take a frequency, and issue the fid/vid transition command */
1035 static int transition_frequency_fidvid(struct powernow_k8_data *data,
1041 struct cpufreq_freqs freqs;
1043 dprintk("cpu %d transition to index %u\n", smp_processor_id(), index);
1045 /* fid/vid correctness check for k8 */
1046 /* fid are the lower 8 bits of the index we stored into
1047 * the cpufreq frequency table in find_psb_table, vid
1048 * are the upper 8 bits.
1050 fid = data->powernow_table[index].index & 0xFF;
1051 vid = (data->powernow_table[index].index & 0xFF00) >> 8;
1053 dprintk("table matched fid 0x%x, giving vid 0x%x\n", fid, vid);
1055 if (query_current_values_with_pending_wait(data))
1058 if ((data->currvid == vid) && (data->currfid == fid)) {
1059 dprintk("target matches current values (fid 0x%x, vid 0x%x)\n",
1064 dprintk("cpu %d, changing to fid 0x%x, vid 0x%x\n",
1065 smp_processor_id(), fid, vid);
1066 freqs.old = find_khz_freq_from_fid(data->currfid);
1067 freqs.new = find_khz_freq_from_fid(fid);
1069 for_each_cpu(i, data->available_cores) {
1071 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
1074 res = transition_fid_vid(data, fid, vid);
1075 freqs.new = find_khz_freq_from_fid(data->currfid);
1077 for_each_cpu(i, data->available_cores) {
1079 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
1084 /* Take a frequency, and issue the hardware pstate transition command */
1085 static int transition_frequency_pstate(struct powernow_k8_data *data,
1090 struct cpufreq_freqs freqs;
1092 dprintk("cpu %d transition to index %u\n", smp_processor_id(), index);
1094 /* get MSR index for hardware pstate transition */
1095 pstate = index & HW_PSTATE_MASK;
1096 if (pstate > data->max_hw_pstate)
1098 freqs.old = find_khz_freq_from_pstate(data->powernow_table,
1100 freqs.new = find_khz_freq_from_pstate(data->powernow_table, pstate);
1102 for_each_cpu(i, data->available_cores) {
1104 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
1107 res = transition_pstate(data, pstate);
1108 freqs.new = find_khz_freq_from_pstate(data->powernow_table, pstate);
1110 for_each_cpu(i, data->available_cores) {
1112 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
1117 /* Driver entry point to switch to the target frequency */
1118 static int powernowk8_target(struct cpufreq_policy *pol,
1119 unsigned targfreq, unsigned relation)
1121 cpumask_var_t oldmask;
1122 struct powernow_k8_data *data = per_cpu(powernow_data, pol->cpu);
1125 unsigned int newstate;
1131 checkfid = data->currfid;
1132 checkvid = data->currvid;
1134 /* only run on specific CPU from here on. */
1135 /* This is poor form: use a workqueue or smp_call_function_single */
1136 if (!alloc_cpumask_var(&oldmask, GFP_KERNEL))
1139 cpumask_copy(oldmask, tsk_cpus_allowed(current));
1140 set_cpus_allowed_ptr(current, cpumask_of(pol->cpu));
1142 if (smp_processor_id() != pol->cpu) {
1143 printk(KERN_ERR PFX "limiting to cpu %u failed\n", pol->cpu);
1147 if (pending_bit_stuck()) {
1148 printk(KERN_ERR PFX "failing targ, change pending bit set\n");
1152 dprintk("targ: cpu %d, %d kHz, min %d, max %d, relation %d\n",
1153 pol->cpu, targfreq, pol->min, pol->max, relation);
1155 if (query_current_values_with_pending_wait(data))
1158 if (cpu_family != CPU_HW_PSTATE) {
1159 dprintk("targ: curr fid 0x%x, vid 0x%x\n",
1160 data->currfid, data->currvid);
1162 if ((checkvid != data->currvid) ||
1163 (checkfid != data->currfid)) {
1164 printk(KERN_INFO PFX
1165 "error - out of sync, fix 0x%x 0x%x, "
1167 checkfid, data->currfid,
1168 checkvid, data->currvid);
1172 if (cpufreq_frequency_table_target(pol, data->powernow_table,
1173 targfreq, relation, &newstate))
1176 mutex_lock(&fidvid_mutex);
1178 powernow_k8_acpi_pst_values(data, newstate);
1180 if (cpu_family == CPU_HW_PSTATE)
1181 ret = transition_frequency_pstate(data, newstate);
1183 ret = transition_frequency_fidvid(data, newstate);
1185 printk(KERN_ERR PFX "transition frequency failed\n");
1187 mutex_unlock(&fidvid_mutex);
1190 mutex_unlock(&fidvid_mutex);
1192 if (cpu_family == CPU_HW_PSTATE)
1193 pol->cur = find_khz_freq_from_pstate(data->powernow_table,
1196 pol->cur = find_khz_freq_from_fid(data->currfid);
1200 set_cpus_allowed_ptr(current, oldmask);
1201 free_cpumask_var(oldmask);
1205 /* Driver entry point to verify the policy and range of frequencies */
1206 static int powernowk8_verify(struct cpufreq_policy *pol)
1208 struct powernow_k8_data *data = per_cpu(powernow_data, pol->cpu);
1213 return cpufreq_frequency_table_verify(pol, data->powernow_table);
1216 struct init_on_cpu {
1217 struct powernow_k8_data *data;
1221 static void __cpuinit powernowk8_cpu_init_on_cpu(void *_init_on_cpu)
1223 struct init_on_cpu *init_on_cpu = _init_on_cpu;
1225 if (pending_bit_stuck()) {
1226 printk(KERN_ERR PFX "failing init, change pending bit set\n");
1227 init_on_cpu->rc = -ENODEV;
1231 if (query_current_values_with_pending_wait(init_on_cpu->data)) {
1232 init_on_cpu->rc = -ENODEV;
1236 if (cpu_family == CPU_OPTERON)
1239 init_on_cpu->rc = 0;
1242 /* per CPU init entry point to the driver */
1243 static int __cpuinit powernowk8_cpu_init(struct cpufreq_policy *pol)
1245 static const char ACPI_PSS_BIOS_BUG_MSG[] =
1246 KERN_ERR FW_BUG PFX "No compatible ACPI _PSS objects found.\n"
1247 FW_BUG PFX "Try again with latest BIOS.\n";
1248 struct powernow_k8_data *data;
1249 struct init_on_cpu init_on_cpu;
1252 if (!cpu_online(pol->cpu))
1255 smp_call_function_single(pol->cpu, check_supported_cpu, &rc, 1);
1259 data = kzalloc(sizeof(struct powernow_k8_data), GFP_KERNEL);
1261 printk(KERN_ERR PFX "unable to alloc powernow_k8_data");
1265 data->cpu = pol->cpu;
1266 data->currpstate = HW_PSTATE_INVALID;
1268 if (powernow_k8_cpu_init_acpi(data)) {
1270 * Use the PSB BIOS structure. This is only availabe on
1271 * an UP version, and is deprecated by AMD.
1273 if (num_online_cpus() != 1) {
1274 printk_once(ACPI_PSS_BIOS_BUG_MSG);
1277 if (pol->cpu != 0) {
1278 printk(KERN_ERR FW_BUG PFX "No ACPI _PSS objects for "
1279 "CPU other than CPU0. Complain to your BIOS "
1283 rc = find_psb_table(data);
1287 /* Take a crude guess here.
1288 * That guess was in microseconds, so multiply with 1000 */
1289 pol->cpuinfo.transition_latency = (
1290 ((data->rvo + 8) * data->vstable * VST_UNITS_20US) +
1291 ((1 << data->irt) * 30)) * 1000;
1292 } else /* ACPI _PSS objects available */
1293 pol->cpuinfo.transition_latency = get_transition_latency(data);
1295 /* only run on specific CPU from here on */
1296 init_on_cpu.data = data;
1297 smp_call_function_single(data->cpu, powernowk8_cpu_init_on_cpu,
1299 rc = init_on_cpu.rc;
1301 goto err_out_exit_acpi;
1303 if (cpu_family == CPU_HW_PSTATE)
1304 cpumask_copy(pol->cpus, cpumask_of(pol->cpu));
1306 cpumask_copy(pol->cpus, cpu_core_mask(pol->cpu));
1307 data->available_cores = pol->cpus;
1309 if (cpu_family == CPU_HW_PSTATE)
1310 pol->cur = find_khz_freq_from_pstate(data->powernow_table,
1313 pol->cur = find_khz_freq_from_fid(data->currfid);
1314 dprintk("policy current frequency %d kHz\n", pol->cur);
1316 /* min/max the cpu is capable of */
1317 if (cpufreq_frequency_table_cpuinfo(pol, data->powernow_table)) {
1318 printk(KERN_ERR FW_BUG PFX "invalid powernow_table\n");
1319 powernow_k8_cpu_exit_acpi(data);
1320 kfree(data->powernow_table);
1325 cpufreq_frequency_table_get_attr(data->powernow_table, pol->cpu);
1327 if (cpu_family == CPU_HW_PSTATE)
1328 dprintk("cpu_init done, current pstate 0x%x\n",
1331 dprintk("cpu_init done, current fid 0x%x, vid 0x%x\n",
1332 data->currfid, data->currvid);
1334 per_cpu(powernow_data, pol->cpu) = data;
1339 powernow_k8_cpu_exit_acpi(data);
1346 static int __devexit powernowk8_cpu_exit(struct cpufreq_policy *pol)
1348 struct powernow_k8_data *data = per_cpu(powernow_data, pol->cpu);
1353 powernow_k8_cpu_exit_acpi(data);
1355 cpufreq_frequency_table_put_attr(pol->cpu);
1357 kfree(data->powernow_table);
1359 per_cpu(powernow_data, pol->cpu) = NULL;
1364 static void query_values_on_cpu(void *_err)
1367 struct powernow_k8_data *data = __get_cpu_var(powernow_data);
1369 *err = query_current_values_with_pending_wait(data);
1372 static unsigned int powernowk8_get(unsigned int cpu)
1374 struct powernow_k8_data *data = per_cpu(powernow_data, cpu);
1375 unsigned int khz = 0;
1381 smp_call_function_single(cpu, query_values_on_cpu, &err, true);
1385 if (cpu_family == CPU_HW_PSTATE)
1386 khz = find_khz_freq_from_pstate(data->powernow_table,
1389 khz = find_khz_freq_from_fid(data->currfid);
1396 static struct freq_attr *powernow_k8_attr[] = {
1397 &cpufreq_freq_attr_scaling_available_freqs,
1401 static struct cpufreq_driver cpufreq_amd64_driver = {
1402 .verify = powernowk8_verify,
1403 .target = powernowk8_target,
1404 .bios_limit = acpi_processor_get_bios_limit,
1405 .init = powernowk8_cpu_init,
1406 .exit = __devexit_p(powernowk8_cpu_exit),
1407 .get = powernowk8_get,
1408 .name = "powernow-k8",
1409 .owner = THIS_MODULE,
1410 .attr = powernow_k8_attr,
1413 /* driver entry point for init */
1414 static int __cpuinit powernowk8_init(void)
1416 unsigned int i, supported_cpus = 0;
1418 for_each_online_cpu(i) {
1420 smp_call_function_single(i, check_supported_cpu, &rc, 1);
1425 if (supported_cpus == num_online_cpus()) {
1426 printk(KERN_INFO PFX "Found %d %s "
1427 "processors (%d cpu cores) (" VERSION ")\n",
1429 boot_cpu_data.x86_model_id, supported_cpus);
1430 return cpufreq_register_driver(&cpufreq_amd64_driver);
1436 /* driver entry point for term */
1437 static void __exit powernowk8_exit(void)
1441 cpufreq_unregister_driver(&cpufreq_amd64_driver);
1444 MODULE_AUTHOR("Paul Devriendt <paul.devriendt@amd.com> and "
1445 "Mark Langsdorf <mark.langsdorf@amd.com>");
1446 MODULE_DESCRIPTION("AMD Athlon 64 and Opteron processor frequency driver.");
1447 MODULE_LICENSE("GPL");
1449 late_initcall(powernowk8_init);
1450 module_exit(powernowk8_exit);