arch/i386/kernel/cpu/cpufreq/powernow-k8.c

   1 /*
   2  *   (c) 2003, 2004, 2005 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  *  Support : mark.langsdorf@amd.com
   8  *
   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.
  15  *
  16  *  Valuable input gratefully received from Dave Jones, Pavel Machek,
  17  *  Dominik Brodowski, 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
  22  *
  23  *  Tables for specific CPUs can be infrerred from
  24  *     http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/30430.pdf
  25  */
  26
  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
  36 #include <asm/msr.h>
  37 #include <asm/io.h>
  38 #include <asm/delay.h>
  39
  40 #ifdef CONFIG_X86_POWERNOW_K8_ACPI
  41 #include <linux/acpi.h>
  42 #include <acpi/processor.h>
  43 #endif
  44
  45 #define PFX "powernow-k8: "
  46 #define BFX PFX "BIOS error: "
  47 #define VERSION "version 1.50.3"
  48 #include "powernow-k8.h"
  49
  50 /* serialize freq changes  */
  51 static DECLARE_MUTEX(fidvid_sem);
  52
  53 static struct powernow_k8_data *powernow_data[NR_CPUS];
  54
  55 #ifndef CONFIG_SMP
  56 static cpumask_t cpu_core_map[1];
  57 #endif
  58
  59 /* Return a frequency in MHz, given an input fid */
  60 static u32 find_freq_from_fid(u32 fid)
  61 {
  62         return 800 + (fid * 100);
  63 }
  64
  65 /* Return a frequency in KHz, given an input fid */
  66 static u32 find_khz_freq_from_fid(u32 fid)
  67 {
  68         return 1000 * find_freq_from_fid(fid);
  69 }
  70
  71 /* Return a voltage in miliVolts, given an input vid */
  72 static u32 find_millivolts_from_vid(struct powernow_k8_data *data, u32 vid)
  73 {
  74         return 1550-vid*25;
  75 }
  76
  77 /* Return the vco fid for an input fid
  78  *
  79  * Each "low" fid has corresponding "high" fid, and you can get to "low" fids
  80  * only from corresponding high fids. This returns "high" fid corresponding to
  81  * "low" one.
  82  */
  83 static u32 convert_fid_to_vco_fid(u32 fid)
  84 {
  85         if (fid < HI_FID_TABLE_BOTTOM) {
  86                 return 8 + (2 * fid);
  87         } else {
  88                 return fid;
  89         }
  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++ > 0x1000000) {
 115                         printk(KERN_ERR PFX "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 stabalization 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         dprintk("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
 157 /* write the new fid value along with the other control fields to the msr */
 158 static int write_new_fid(struct powernow_k8_data *data, u32 fid)
 159 {
 160         u32 lo;
 161         u32 savevid = data->currvid;
 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 | (data->currvid << MSR_C_LO_VID_SHIFT) | MSR_C_LO_INIT_FID_VID;
 169
 170         dprintk("writing fid 0x%x, lo 0x%x, hi 0x%x\n",
 171                 fid, lo, data->plllock * PLL_LOCK_CONVERSION);
 172
 173         wrmsr(MSR_FIDVID_CTL, lo, data->plllock * PLL_LOCK_CONVERSION);
 174
 175         if (query_current_values_with_pending_wait(data))
 176                 return 1;
 177
 178         count_off_irt(data);
 179
 180         if (savevid != data->currvid) {
 181                 printk(KERN_ERR PFX "vid change on fid trans, old 0x%x, new 0x%x\n",
 182                        savevid, data->currvid);
 183                 return 1;
 184         }
 185
 186         if (fid != data->currfid) {
 187                 printk(KERN_ERR PFX "fid trans failed, fid 0x%x, curr 0x%x\n", fid,
 188                         data->currfid);
 189                 return 1;
 190         }
 191
 192         return 0;
 193 }
 194
 195 /* Write a new vid to the hardware */
 196 static int write_new_vid(struct powernow_k8_data *data, u32 vid)
 197 {
 198         u32 lo;
 199         u32 savefid = data->currfid;
 200
 201         if ((data->currfid & INVALID_FID_MASK) || (vid & INVALID_VID_MASK)) {
 202                 printk(KERN_ERR PFX "internal error - overflow on vid write\n");
 203                 return 1;
 204         }
 205
 206         lo = data->currfid | (vid << MSR_C_LO_VID_SHIFT) | MSR_C_LO_INIT_FID_VID;
 207
 208         dprintk("writing vid 0x%x, lo 0x%x, hi 0x%x\n",
 209                 vid, lo, STOP_GRANT_5NS);
 210
 211         wrmsr(MSR_FIDVID_CTL, lo, STOP_GRANT_5NS);
 212
 213         if (query_current_values_with_pending_wait(data))
 214                 return 1;
 215
 216         if (savefid != data->currfid) {
 217                 printk(KERN_ERR PFX "fid changed on vid trans, old 0x%x new 0x%x\n",
 218                        savefid, data->currfid);
 219                 return 1;
 220         }
 221
 222         if (vid != data->currvid) {
 223                 printk(KERN_ERR PFX "vid trans failed, vid 0x%x, curr 0x%x\n", vid,
 224                                 data->currvid);
 225                 return 1;
 226         }
 227
 228         return 0;
 229 }
 230
 231 /*
 232  * Reduce the vid by the max of step or reqvid.
 233  * Decreasing vid codes represent increasing voltages:
 234  * vid of 0 is 1.550V, vid of 0x1e is 0.800V, vid of VID_OFF is off.
 235  */
 236 static int decrease_vid_code_by_step(struct powernow_k8_data *data, u32 reqvid, u32 step)
 237 {
 238         if ((data->currvid - reqvid) > step)
 239                 reqvid = data->currvid - step;
 240
 241         if (write_new_vid(data, reqvid))
 242                 return 1;
 243
 244         count_off_vst(data);
 245
 246         return 0;
 247 }
 248
 249 /* Change the fid and vid, by the 3 phases. */
 250 static int transition_fid_vid(struct powernow_k8_data *data, u32 reqfid, u32 reqvid)
 251 {
 252         if (core_voltage_pre_transition(data, reqvid))
 253                 return 1;
 254
 255         if (core_frequency_transition(data, reqfid))
 256                 return 1;
 257
 258         if (core_voltage_post_transition(data, reqvid))
 259                 return 1;
 260
 261         if (query_current_values_with_pending_wait(data))
 262                 return 1;
 263
 264         if ((reqfid != data->currfid) || (reqvid != data->currvid)) {
 265                 printk(KERN_ERR PFX "failed (cpu%d): req 0x%x 0x%x, curr 0x%x 0x%x\n",
 266                                 smp_processor_id(),
 267                                 reqfid, reqvid, data->currfid, data->currvid);
 268                 return 1;
 269         }
 270
 271         dprintk("transitioned (cpu%d): new fid 0x%x, vid 0x%x\n",
 272                 smp_processor_id(), data->currfid, data->currvid);
 273
 274         return 0;
 275 }
 276
 277 /* Phase 1 - core voltage transition ... setup voltage */
 278 static int core_voltage_pre_transition(struct powernow_k8_data *data, u32 reqvid)
 279 {
 280         u32 rvosteps = data->rvo;
 281         u32 savefid = data->currfid;
 282         u32 maxvid, lo;
 283
 284         dprintk("ph1 (cpu%d): start, currfid 0x%x, currvid 0x%x, reqvid 0x%x, rvo 0x%x\n",
 285                 smp_processor_id(),
 286                 data->currfid, data->currvid, reqvid, data->rvo);
 287
 288         rdmsr(MSR_FIDVID_STATUS, lo, maxvid);
 289         maxvid = 0x1f & (maxvid >> 16);
 290         dprintk("ph1 maxvid=0x%x\n", maxvid);
 291         if (reqvid < maxvid) /* lower numbers are higher voltages */
 292                 reqvid = maxvid;
 293
 294         while (data->currvid > reqvid) {
 295                 dprintk("ph1: curr 0x%x, req vid 0x%x\n",
 296                         data->currvid, reqvid);
 297                 if (decrease_vid_code_by_step(data, reqvid, data->vidmvs))
 298                         return 1;
 299         }
 300
 301         while ((rvosteps > 0) && ((data->rvo + data->currvid) > reqvid)) {
 302                 if (data->currvid == maxvid) {
 303                         rvosteps = 0;
 304                 } else {
 305                         dprintk("ph1: changing vid for rvo, req 0x%x\n",
 306                                 data->currvid - 1);
 307                         if (decrease_vid_code_by_step(data, data->currvid - 1, 1))
 308                                 return 1;
 309                         rvosteps--;
 310                 }
 311         }
 312
 313         if (query_current_values_with_pending_wait(data))
 314                 return 1;
 315
 316         if (savefid != data->currfid) {
 317                 printk(KERN_ERR PFX "ph1 err, currfid changed 0x%x\n", data->currfid);
 318                 return 1;
 319         }
 320
 321         dprintk("ph1 complete, currfid 0x%x, currvid 0x%x\n",
 322                 data->currfid, data->currvid);
 323
 324         return 0;
 325 }
 326
 327 /* Phase 2 - core frequency transition */
 328 static int core_frequency_transition(struct powernow_k8_data *data, u32 reqfid)
 329 {
 330         u32 vcoreqfid, vcocurrfid, vcofiddiff, savevid = data->currvid;
 331
 332         if ((reqfid < HI_FID_TABLE_BOTTOM) && (data->currfid < HI_FID_TABLE_BOTTOM)) {
 333                 printk(KERN_ERR PFX "ph2: illegal lo-lo transition 0x%x 0x%x\n",
 334                         reqfid, data->currfid);
 335                 return 1;
 336         }
 337
 338         if (data->currfid == reqfid) {
 339                 printk(KERN_ERR PFX "ph2 null fid transition 0x%x\n", data->currfid);
 340                 return 0;
 341         }
 342
 343         dprintk("ph2 (cpu%d): starting, currfid 0x%x, currvid 0x%x, reqfid 0x%x\n",
 344                 smp_processor_id(),
 345                 data->currfid, data->currvid, reqfid);
 346
 347         vcoreqfid = convert_fid_to_vco_fid(reqfid);
 348         vcocurrfid = convert_fid_to_vco_fid(data->currfid);
 349         vcofiddiff = vcocurrfid > vcoreqfid ? vcocurrfid - vcoreqfid
 350             : vcoreqfid - vcocurrfid;
 351
 352         while (vcofiddiff > 2) {
 353                 if (reqfid > data->currfid) {
 354                         if (data->currfid > LO_FID_TABLE_TOP) {
 355                                 if (write_new_fid(data, data->currfid + 2)) {
 356                                         return 1;
 357                                 }
 358                         } else {
 359                                 if (write_new_fid
 360                                     (data, 2 + convert_fid_to_vco_fid(data->currfid))) {
 361                                         return 1;
 362                                 }
 363                         }
 364                 } else {
 365                         if (write_new_fid(data, data->currfid - 2))
 366                                 return 1;
 367                 }
 368
 369                 vcocurrfid = convert_fid_to_vco_fid(data->currfid);
 370                 vcofiddiff = vcocurrfid > vcoreqfid ? vcocurrfid - vcoreqfid
 371                     : vcoreqfid - vcocurrfid;
 372         }
 373
 374         if (write_new_fid(data, reqfid))
 375                 return 1;
 376
 377         if (query_current_values_with_pending_wait(data))
 378                 return 1;
 379
 380         if (data->currfid != reqfid) {
 381                 printk(KERN_ERR PFX
 382                         "ph2: mismatch, failed fid transition, curr 0x%x, req 0x%x\n",
 383                         data->currfid, reqfid);
 384                 return 1;
 385         }
 386
 387         if (savevid != data->currvid) {
 388                 printk(KERN_ERR PFX "ph2: vid changed, save 0x%x, curr 0x%x\n",
 389                         savevid, data->currvid);
 390                 return 1;
 391         }
 392
 393         dprintk("ph2 complete, currfid 0x%x, currvid 0x%x\n",
 394                 data->currfid, data->currvid);
 395
 396         return 0;
 397 }
 398
 399 /* Phase 3 - core voltage transition flow ... jump to the final vid. */
 400 static int core_voltage_post_transition(struct powernow_k8_data *data, u32 reqvid)
 401 {
 402         u32 savefid = data->currfid;
 403         u32 savereqvid = reqvid;
 404
 405         dprintk("ph3 (cpu%d): starting, currfid 0x%x, currvid 0x%x\n",
 406                 smp_processor_id(),
 407                 data->currfid, data->currvid);
 408
 409         if (reqvid != data->currvid) {
 410                 if (write_new_vid(data, reqvid))
 411                         return 1;
 412
 413                 if (savefid != data->currfid) {
 414                         printk(KERN_ERR PFX
 415                                "ph3: bad fid change, save 0x%x, curr 0x%x\n",
 416                                savefid, data->currfid);
 417                         return 1;
 418                 }
 419
 420                 if (data->currvid != reqvid) {
 421                         printk(KERN_ERR PFX
 422                                "ph3: failed vid transition\n, req 0x%x, curr 0x%x",
 423                                reqvid, data->currvid);
 424                         return 1;
 425                 }
 426         }
 427
 428         if (query_current_values_with_pending_wait(data))
 429                 return 1;
 430
 431         if (savereqvid != data->currvid) {
 432                 dprintk("ph3 failed, currvid 0x%x\n", data->currvid);
 433                 return 1;
 434         }
 435
 436         if (savefid != data->currfid) {
 437                 dprintk("ph3 failed, currfid changed 0x%x\n",
 438                         data->currfid);
 439                 return 1;
 440         }
 441
 442         dprintk("ph3 complete, currfid 0x%x, currvid 0x%x\n",
 443                 data->currfid, data->currvid);
 444
 445         return 0;
 446 }
 447
 448 static int check_supported_cpu(unsigned int cpu)
 449 {
 450         cpumask_t oldmask = CPU_MASK_ALL;
 451         u32 eax, ebx, ecx, edx;
 452         unsigned int rc = 0;
 453
 454         oldmask = current->cpus_allowed;
 455         set_cpus_allowed(current, cpumask_of_cpu(cpu));
 456         schedule();
 457
 458         if (smp_processor_id() != cpu) {
 459                 printk(KERN_ERR "limiting to cpu %u failed\n", cpu);
 460                 goto out;
 461         }
 462
 463         if (current_cpu_data.x86_vendor != X86_VENDOR_AMD)
 464                 goto out;
 465
 466         eax = cpuid_eax(CPUID_PROCESSOR_SIGNATURE);
 467         if (((eax & CPUID_USE_XFAM_XMOD) != CPUID_USE_XFAM_XMOD) ||
 468             ((eax & CPUID_XFAM) != CPUID_XFAM_K8) ||
 469             ((eax & CPUID_XMOD) > CPUID_XMOD_REV_F)) {
 470                 printk(KERN_INFO PFX "Processor cpuid %x not supported\n", eax);
 471                 goto out;
 472         }
 473
 474         eax = cpuid_eax(CPUID_GET_MAX_CAPABILITIES);
 475         if (eax < CPUID_FREQ_VOLT_CAPABILITIES) {
 476                 printk(KERN_INFO PFX
 477                        "No frequency change capabilities detected\n");
 478                 goto out;
 479         }
 480
 481         cpuid(CPUID_FREQ_VOLT_CAPABILITIES, &eax, &ebx, &ecx, &edx);
 482         if ((edx & P_STATE_TRANSITION_CAPABLE) != P_STATE_TRANSITION_CAPABLE) {
 483                 printk(KERN_INFO PFX "Power state transitions not supported\n");
 484                 goto out;
 485         }
 486
 487         rc = 1;
 488
 489 out:
 490         set_cpus_allowed(current, oldmask);
 491         schedule();
 492         return rc;
 493
 494 }
 495
 496 static int check_pst_table(struct powernow_k8_data *data, struct pst_s *pst, u8 maxvid)
 497 {
 498         unsigned int j;
 499         u8 lastfid = 0xff;
 500
 501         for (j = 0; j < data->numps; j++) {
 502                 if (pst[j].vid > LEAST_VID) {
 503                         printk(KERN_ERR PFX "vid %d invalid : 0x%x\n", j, pst[j].vid);
 504                         return -EINVAL;
 505                 }
 506                 if (pst[j].vid < data->rvo) {   /* vid + rvo >= 0 */
 507                         printk(KERN_ERR BFX "0 vid exceeded with pstate %d\n", j);
 508                         return -ENODEV;
 509                 }
 510                 if (pst[j].vid < maxvid + data->rvo) {  /* vid + rvo >= maxvid */
 511                         printk(KERN_ERR BFX "maxvid exceeded with pstate %d\n", j);
 512                         return -ENODEV;
 513                 }
 514                 if ((pst[j].fid > MAX_FID)
 515                     || (pst[j].fid & 1)
 516                     || (j && (pst[j].fid < HI_FID_TABLE_BOTTOM))) {
 517                         /* Only first fid is allowed to be in "low" range */
 518                         printk(KERN_ERR PFX "two low fids - %d : 0x%x\n", j, pst[j].fid);
 519                         return -EINVAL;
 520                 }
 521                 if (pst[j].fid < lastfid)
 522                         lastfid = pst[j].fid;
 523         }
 524         if (lastfid & 1) {
 525                 printk(KERN_ERR PFX "lastfid invalid\n");
 526                 return -EINVAL;
 527         }
 528         if (lastfid > LO_FID_TABLE_TOP)
 529                 printk(KERN_INFO PFX  "first fid not from lo freq table\n");
 530
 531         return 0;
 532 }
 533
 534 static void print_basics(struct powernow_k8_data *data)
 535 {
 536         int j;
 537         for (j = 0; j < data->numps; j++) {
 538                 if (data->powernow_table[j].frequency != CPUFREQ_ENTRY_INVALID)
 539                         printk(KERN_INFO PFX "   %d : fid 0x%x (%d MHz), vid 0x%x (%d mV)\n", j,
 540                                 data->powernow_table[j].index & 0xff,
 541                                 data->powernow_table[j].frequency/1000,
 542                                 data->powernow_table[j].index >> 8,
 543                                 find_millivolts_from_vid(data, data->powernow_table[j].index >> 8));
 544         }
 545         if (data->batps)
 546                 printk(KERN_INFO PFX "Only %d pstates on battery\n", data->batps);
 547 }
 548
 549 static int fill_powernow_table(struct powernow_k8_data *data, struct pst_s *pst, u8 maxvid)
 550 {
 551         struct cpufreq_frequency_table *powernow_table;
 552         unsigned int j;
 553
 554         if (data->batps) {    /* use ACPI support to get full speed on mains power */
 555                 printk(KERN_WARNING PFX "Only %d pstates usable (use ACPI driver for full range\n", data->batps);
 556                 data->numps = data->batps;
 557         }
 558
 559         for ( j=1; j<data->numps; j++ ) {
 560                 if (pst[j-1].fid >= pst[j].fid) {
 561                         printk(KERN_ERR PFX "PST out of sequence\n");
 562                         return -EINVAL;
 563                 }
 564         }
 565
 566         if (data->numps < 2) {
 567                 printk(KERN_ERR PFX "no p states to transition\n");
 568                 return -ENODEV;
 569         }
 570
 571         if (check_pst_table(data, pst, maxvid))
 572                 return -EINVAL;
 573
 574         powernow_table = kmalloc((sizeof(struct cpufreq_frequency_table)
 575                 * (data->numps + 1)), GFP_KERNEL);
 576         if (!powernow_table) {
 577                 printk(KERN_ERR PFX "powernow_table memory alloc failure\n");
 578                 return -ENOMEM;
 579         }
 580
 581         for (j = 0; j < data->numps; j++) {
 582                 powernow_table[j].index = pst[j].fid; /* lower 8 bits */
 583                 powernow_table[j].index |= (pst[j].vid << 8); /* upper 8 bits */
 584                 powernow_table[j].frequency = find_khz_freq_from_fid(pst[j].fid);
 585         }
 586         powernow_table[data->numps].frequency = CPUFREQ_TABLE_END;
 587         powernow_table[data->numps].index = 0;
 588
 589         if (query_current_values_with_pending_wait(data)) {
 590                 kfree(powernow_table);
 591                 return -EIO;
 592         }
 593
 594         dprintk("cfid 0x%x, cvid 0x%x\n", data->currfid, data->currvid);
 595         data->powernow_table = powernow_table;
 596         print_basics(data);
 597
 598         for (j = 0; j < data->numps; j++)
 599                 if ((pst[j].fid==data->currfid) && (pst[j].vid==data->currvid))
 600                         return 0;
 601
 602         dprintk("currfid/vid do not match PST, ignoring\n");
 603         return 0;
 604 }
 605
 606 /* Find and validate the PSB/PST table in BIOS. */
 607 static int find_psb_table(struct powernow_k8_data *data)
 608 {
 609         struct psb_s *psb;
 610         unsigned int i;
 611         u32 mvs;
 612         u8 maxvid;
 613         u32 cpst = 0;
 614         u32 thiscpuid;
 615
 616         for (i = 0xc0000; i < 0xffff0; i += 0x10) {
 617                 /* Scan BIOS looking for the signature. */
 618                 /* It can not be at ffff0 - it is too big. */
 619
 620                 psb = phys_to_virt(i);
 621                 if (memcmp(psb, PSB_ID_STRING, PSB_ID_STRING_LEN) != 0)
 622                         continue;
 623
 624                 dprintk("found PSB header at 0x%p\n", psb);
 625
 626                 dprintk("table vers: 0x%x\n", psb->tableversion);
 627                 if (psb->tableversion != PSB_VERSION_1_4) {
 628                         printk(KERN_INFO BFX "PSB table is not v1.4\n");
 629                         return -ENODEV;
 630                 }
 631
 632                 dprintk("flags: 0x%x\n", psb->flags1);
 633                 if (psb->flags1) {
 634                         printk(KERN_ERR BFX "unknown flags\n");
 635                         return -ENODEV;
 636                 }
 637
 638                 data->vstable = psb->vstable;
 639                 dprintk("voltage stabilization time: %d(*20us)\n", data->vstable);
 640
 641                 dprintk("flags2: 0x%x\n", psb->flags2);
 642                 data->rvo = psb->flags2 & 3;
 643                 data->irt = ((psb->flags2) >> 2) & 3;
 644                 mvs = ((psb->flags2) >> 4) & 3;
 645                 data->vidmvs = 1 << mvs;
 646                 data->batps = ((psb->flags2) >> 6) & 3;
 647
 648                 dprintk("ramp voltage offset: %d\n", data->rvo);
 649                 dprintk("isochronous relief time: %d\n", data->irt);
 650                 dprintk("maximum voltage step: %d - 0x%x\n", mvs, data->vidmvs);
 651
 652                 dprintk("numpst: 0x%x\n", psb->num_tables);
 653                 cpst = psb->num_tables;
 654                 if ((psb->cpuid == 0x00000fc0) || (psb->cpuid == 0x00000fe0) ){
 655                         thiscpuid = cpuid_eax(CPUID_PROCESSOR_SIGNATURE);
 656                         if ((thiscpuid == 0x00000fc0) || (thiscpuid == 0x00000fe0) ) {
 657                                 cpst = 1;
 658                         }
 659                 }
 660                 if (cpst != 1) {
 661                         printk(KERN_ERR BFX "numpst must be 1\n");
 662                         return -ENODEV;
 663                 }
 664
 665                 data->plllock = psb->plllocktime;
 666                 dprintk("plllocktime: 0x%x (units 1us)\n", psb->plllocktime);
 667                 dprintk("maxfid: 0x%x\n", psb->maxfid);
 668                 dprintk("maxvid: 0x%x\n", psb->maxvid);
 669                 maxvid = psb->maxvid;
 670
 671                 data->numps = psb->numps;
 672                 dprintk("numpstates: 0x%x\n", data->numps);
 673                 return fill_powernow_table(data, (struct pst_s *)(psb+1), maxvid);
 674         }
 675         /*
 676          * If you see this message, complain to BIOS manufacturer. If
 677          * he tells you "we do not support Linux" or some similar
 678          * nonsense, remember that Windows 2000 uses the same legacy
 679          * mechanism that the old Linux PSB driver uses. Tell them it
 680          * is broken with Windows 2000.
 681          *
 682          * The reference to the AMD documentation is chapter 9 in the
 683          * BIOS and Kernel Developer's Guide, which is available on
 684          * www.amd.com
 685          */
 686         printk(KERN_INFO PFX "BIOS error - no PSB or ACPI _PSS objects\n");
 687         return -ENODEV;
 688 }
 689
 690 #ifdef CONFIG_X86_POWERNOW_K8_ACPI
 691 static void powernow_k8_acpi_pst_values(struct powernow_k8_data *data, unsigned int index)
 692 {
 693         if (!data->acpi_data.state_count)
 694                 return;
 695
 696         data->irt = (data->acpi_data.states[index].control >> IRT_SHIFT) & IRT_MASK;
 697         data->rvo = (data->acpi_data.states[index].control >> RVO_SHIFT) & RVO_MASK;
 698         data->exttype = (data->acpi_data.states[index].control >> EXT_TYPE_SHIFT) & EXT_TYPE_MASK;
 699         data->plllock = (data->acpi_data.states[index].control >> PLL_L_SHIFT) & PLL_L_MASK;
 700         data->vidmvs = 1 << ((data->acpi_data.states[index].control >> MVS_SHIFT) & MVS_MASK);
 701         data->vstable = (data->acpi_data.states[index].control >> VST_SHIFT) & VST_MASK;
 702 }
 703
 704 static int powernow_k8_cpu_init_acpi(struct powernow_k8_data *data)
 705 {
 706         int i;
 707         int cntlofreq = 0;
 708         struct cpufreq_frequency_table *powernow_table;
 709
 710         if (acpi_processor_register_performance(&data->acpi_data, data->cpu)) {
 711                 dprintk("register performance failed: bad ACPI data\n");
 712                 return -EIO;
 713         }
 714
 715         /* verify the data contained in the ACPI structures */
 716         if (data->acpi_data.state_count <= 1) {
 717                 dprintk("No ACPI P-States\n");
 718                 goto err_out;
 719         }
 720
 721         if ((data->acpi_data.control_register.space_id != ACPI_ADR_SPACE_FIXED_HARDWARE) ||
 722                 (data->acpi_data.status_register.space_id != ACPI_ADR_SPACE_FIXED_HARDWARE)) {
 723                 dprintk("Invalid control/status registers (%x - %x)\n",
 724                         data->acpi_data.control_register.space_id,
 725                         data->acpi_data.status_register.space_id);
 726                 goto err_out;
 727         }
 728
 729         /* fill in data->powernow_table */
 730         powernow_table = kmalloc((sizeof(struct cpufreq_frequency_table)
 731                 * (data->acpi_data.state_count + 1)), GFP_KERNEL);
 732         if (!powernow_table) {
 733                 dprintk("powernow_table memory alloc failure\n");
 734                 goto err_out;
 735         }
 736
 737         for (i = 0; i < data->acpi_data.state_count; i++) {
 738                 u32 fid;
 739                 u32 vid;
 740
 741                 if (data->exttype) {
 742                         fid = data->acpi_data.states[i].status & FID_MASK;
 743                         vid = (data->acpi_data.states[i].status >> VID_SHIFT) & VID_MASK;
 744                 } else {
 745                         fid = data->acpi_data.states[i].control & FID_MASK;
 746                         vid = (data->acpi_data.states[i].control >> VID_SHIFT) & VID_MASK;
 747                 }
 748
 749                 dprintk("   %d : fid 0x%x, vid 0x%x\n", i, fid, vid);
 750
 751                 powernow_table[i].index = fid; /* lower 8 bits */
 752                 powernow_table[i].index |= (vid << 8); /* upper 8 bits */
 753                 powernow_table[i].frequency = find_khz_freq_from_fid(fid);
 754
 755                 /* verify frequency is OK */
 756                 if ((powernow_table[i].frequency > (MAX_FREQ * 1000)) ||
 757                         (powernow_table[i].frequency < (MIN_FREQ * 1000))) {
 758                         dprintk("invalid freq %u kHz, ignoring\n", powernow_table[i].frequency);
 759                         powernow_table[i].frequency = CPUFREQ_ENTRY_INVALID;
 760                         continue;
 761                 }
 762
 763                 /* verify voltage is OK - BIOSs are using "off" to indicate invalid */
 764                 if (vid == VID_OFF) {
 765                         dprintk("invalid vid %u, ignoring\n", vid);
 766                         powernow_table[i].frequency = CPUFREQ_ENTRY_INVALID;
 767                         continue;
 768                 }
 769
 770                 /* verify only 1 entry from the lo frequency table */
 771                 if (fid < HI_FID_TABLE_BOTTOM) {
 772                         if (cntlofreq) {
 773                                 /* if both entries are the same, ignore this
 774                                  * one...
 775                                  */
 776                                 if ((powernow_table[i].frequency != powernow_table[cntlofreq].frequency) ||
 777                                     (powernow_table[i].index != powernow_table[cntlofreq].index)) {
 778                                         printk(KERN_ERR PFX "Too many lo freq table entries\n");
 779                                         goto err_out_mem;
 780                                 }
 781
 782                                 dprintk("double low frequency table entry, ignoring it.\n");
 783                                 powernow_table[i].frequency = CPUFREQ_ENTRY_INVALID;
 784                                 continue;
 785                         } else
 786                                 cntlofreq = i;
 787                 }
 788
 789                 if (powernow_table[i].frequency != (data->acpi_data.states[i].core_frequency * 1000)) {
 790                         printk(KERN_INFO PFX "invalid freq entries %u kHz vs. %u kHz\n",
 791                                 powernow_table[i].frequency,
 792                                 (unsigned int) (data->acpi_data.states[i].core_frequency * 1000));
 793                         powernow_table[i].frequency = CPUFREQ_ENTRY_INVALID;
 794                         continue;
 795                 }
 796         }
 797
 798         powernow_table[data->acpi_data.state_count].frequency = CPUFREQ_TABLE_END;
 799         powernow_table[data->acpi_data.state_count].index = 0;
 800         data->powernow_table = powernow_table;
 801
 802         /* fill in data */
 803         data->numps = data->acpi_data.state_count;
 804         print_basics(data);
 805         powernow_k8_acpi_pst_values(data, 0);
 806
 807         /* notify BIOS that we exist */
 808         acpi_processor_notify_smm(THIS_MODULE);
 809
 810         return 0;
 811
 812 err_out_mem:
 813         kfree(powernow_table);
 814
 815 err_out:
 816         acpi_processor_unregister_performance(&data->acpi_data, data->cpu);
 817
 818         /* data->acpi_data.state_count informs us at ->exit() whether ACPI was used */
 819         data->acpi_data.state_count = 0;
 820
 821         return -ENODEV;
 822 }
 823
 824 static void powernow_k8_cpu_exit_acpi(struct powernow_k8_data *data)
 825 {
 826         if (data->acpi_data.state_count)
 827                 acpi_processor_unregister_performance(&data->acpi_data, data->cpu);
 828 }
 829
 830 #else
 831 static int powernow_k8_cpu_init_acpi(struct powernow_k8_data *data) { return -ENODEV; }
 832 static void powernow_k8_cpu_exit_acpi(struct powernow_k8_data *data) { return; }
 833 static void powernow_k8_acpi_pst_values(struct powernow_k8_data *data, unsigned int index) { return; }
 834 #endif /* CONFIG_X86_POWERNOW_K8_ACPI */
 835
 836 /* Take a frequency, and issue the fid/vid transition command */
 837 static int transition_frequency(struct powernow_k8_data *data, unsigned int index)
 838 {
 839         u32 fid;
 840         u32 vid;
 841         int res, i;
 842         struct cpufreq_freqs freqs;
 843
 844         dprintk("cpu %d transition to index %u\n", smp_processor_id(), index);
 845
 846         /* fid are the lower 8 bits of the index we stored into
 847          * the cpufreq frequency table in find_psb_table, vid are
 848          * the upper 8 bits.
 849          */
 850
 851         fid = data->powernow_table[index].index & 0xFF;
 852         vid = (data->powernow_table[index].index & 0xFF00) >> 8;
 853
 854         dprintk("table matched fid 0x%x, giving vid 0x%x\n", fid, vid);
 855
 856         if (query_current_values_with_pending_wait(data))
 857                 return 1;
 858
 859         if ((data->currvid == vid) && (data->currfid == fid)) {
 860                 dprintk("target matches current values (fid 0x%x, vid 0x%x)\n",
 861                         fid, vid);
 862                 return 0;
 863         }
 864
 865         if ((fid < HI_FID_TABLE_BOTTOM) && (data->currfid < HI_FID_TABLE_BOTTOM)) {
 866                 printk(KERN_ERR PFX
 867                        "ignoring illegal change in lo freq table-%x to 0x%x\n",
 868                        data->currfid, fid);
 869                 return 1;
 870         }
 871
 872         dprintk("cpu %d, changing to fid 0x%x, vid 0x%x\n",
 873                 smp_processor_id(), fid, vid);
 874
 875         freqs.cpu = data->cpu;
 876         freqs.old = find_khz_freq_from_fid(data->currfid);
 877         freqs.new = find_khz_freq_from_fid(fid);
 878         for_each_cpu_mask(i, cpu_core_map[data->cpu]) {
 879                 freqs.cpu = i;
 880                 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
 881         }
 882
 883         res = transition_fid_vid(data, fid, vid);
 884
 885         freqs.new = find_khz_freq_from_fid(data->currfid);
 886         for_each_cpu_mask(i, cpu_core_map[data->cpu]) {
 887                 freqs.cpu = i;
 888                 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
 889         }
 890         return res;
 891 }
 892
 893 /* Driver entry point to switch to the target frequency */
 894 static int powernowk8_target(struct cpufreq_policy *pol, unsigned targfreq, unsigned relation)
 895 {
 896         cpumask_t oldmask = CPU_MASK_ALL;
 897         struct powernow_k8_data *data = powernow_data[pol->cpu];
 898         u32 checkfid = data->currfid;
 899         u32 checkvid = data->currvid;
 900         unsigned int newstate;
 901         int ret = -EIO;
 902         int i;
 903
 904         /* only run on specific CPU from here on */
 905         oldmask = current->cpus_allowed;
 906         set_cpus_allowed(current, cpumask_of_cpu(pol->cpu));
 907         schedule();
 908
 909         if (smp_processor_id() != pol->cpu) {
 910                 printk(KERN_ERR "limiting to cpu %u failed\n", pol->cpu);
 911                 goto err_out;
 912         }
 913
 914         if (pending_bit_stuck()) {
 915                 printk(KERN_ERR PFX "failing targ, change pending bit set\n");
 916                 goto err_out;
 917         }
 918
 919         dprintk("targ: cpu %d, %d kHz, min %d, max %d, relation %d\n",
 920                 pol->cpu, targfreq, pol->min, pol->max, relation);
 921
 922         if (query_current_values_with_pending_wait(data)) {
 923                 ret = -EIO;
 924                 goto err_out;
 925         }
 926
 927         dprintk("targ: curr fid 0x%x, vid 0x%x\n",
 928                 data->currfid, data->currvid);
 929
 930         if ((checkvid != data->currvid) || (checkfid != data->currfid)) {
 931                 printk(KERN_INFO PFX
 932                         "error - out of sync, fix 0x%x 0x%x, vid 0x%x 0x%x\n",
 933                         checkfid, data->currfid, checkvid, data->currvid);
 934         }
 935
 936         if (cpufreq_frequency_table_target(pol, data->powernow_table, targfreq, relation, &newstate))
 937                 goto err_out;
 938
 939         down(&fidvid_sem);
 940
 941         powernow_k8_acpi_pst_values(data, newstate);
 942
 943         if (transition_frequency(data, newstate)) {
 944                 printk(KERN_ERR PFX "transition frequency failed\n");
 945                 ret = 1;
 946                 up(&fidvid_sem);
 947                 goto err_out;
 948         }
 949
 950         /* Update all the fid/vids of our siblings */
 951         for_each_cpu_mask(i, cpu_core_map[pol->cpu]) {
 952                 powernow_data[i]->currvid = data->currvid;
 953                 powernow_data[i]->currfid = data->currfid;
 954         }
 955         up(&fidvid_sem);
 956
 957         pol->cur = find_khz_freq_from_fid(data->currfid);
 958         ret = 0;
 959
 960 err_out:
 961         set_cpus_allowed(current, oldmask);
 962         schedule();
 963
 964         return ret;
 965 }
 966
 967 /* Driver entry point to verify the policy and range of frequencies */
 968 static int powernowk8_verify(struct cpufreq_policy *pol)
 969 {
 970         struct powernow_k8_data *data = powernow_data[pol->cpu];
 971
 972         return cpufreq_frequency_table_verify(pol, data->powernow_table);
 973 }
 974
 975 /* per CPU init entry point to the driver */
 976 static int __init powernowk8_cpu_init(struct cpufreq_policy *pol)
 977 {
 978         struct powernow_k8_data *data;
 979         cpumask_t oldmask = CPU_MASK_ALL;
 980         int rc, i;
 981
 982         if (!check_supported_cpu(pol->cpu))
 983                 return -ENODEV;
 984
 985         data = kmalloc(sizeof(struct powernow_k8_data), GFP_KERNEL);
 986         if (!data) {
 987                 printk(KERN_ERR PFX "unable to alloc powernow_k8_data");
 988                 return -ENOMEM;
 989         }
 990         memset(data,0,sizeof(struct powernow_k8_data));
 991
 992         data->cpu = pol->cpu;
 993
 994         if (powernow_k8_cpu_init_acpi(data)) {
 995                 /*
 996                  * Use the PSB BIOS structure. This is only availabe on
 997                  * an UP version, and is deprecated by AMD.
 998                  */
 999
1000                 if ((num_online_cpus() != 1) || (num_possible_cpus() != 1)) {
1001                         printk(KERN_ERR PFX "MP systems not supported by PSB BIOS structure\n");
1002                         kfree(data);
1003                         return -ENODEV;
1004                 }
1005                 if (pol->cpu != 0) {
1006                         printk(KERN_ERR PFX "init not cpu 0\n");
1007                         kfree(data);
1008                         return -ENODEV;
1009                 }
1010                 rc = find_psb_table(data);
1011                 if (rc) {
1012                         kfree(data);
1013                         return -ENODEV;
1014                 }
1015         }
1016
1017         /* only run on specific CPU from here on */
1018         oldmask = current->cpus_allowed;
1019         set_cpus_allowed(current, cpumask_of_cpu(pol->cpu));
1020         schedule();
1021
1022         if (smp_processor_id() != pol->cpu) {
1023                 printk(KERN_ERR "limiting to cpu %u failed\n", pol->cpu);
1024                 goto err_out;
1025         }
1026
1027         if (pending_bit_stuck()) {
1028                 printk(KERN_ERR PFX "failing init, change pending bit set\n");
1029                 goto err_out;
1030         }
1031
1032         if (query_current_values_with_pending_wait(data))
1033                 goto err_out;
1034
1035         fidvid_msr_init();
1036
1037         /* run on any CPU again */
1038         set_cpus_allowed(current, oldmask);
1039         schedule();
1040
1041         pol->governor = CPUFREQ_DEFAULT_GOVERNOR;
1042         pol->cpus = cpu_core_map[pol->cpu];
1043
1044         /* Take a crude guess here.
1045          * That guess was in microseconds, so multiply with 1000 */
1046         pol->cpuinfo.transition_latency = (((data->rvo + 8) * data->vstable * VST_UNITS_20US)
1047             + (3 * (1 << data->irt) * 10)) * 1000;
1048
1049         pol->cur = find_khz_freq_from_fid(data->currfid);
1050         dprintk("policy current frequency %d kHz\n", pol->cur);
1051
1052         /* min/max the cpu is capable of */
1053         if (cpufreq_frequency_table_cpuinfo(pol, data->powernow_table)) {
1054                 printk(KERN_ERR PFX "invalid powernow_table\n");
1055                 powernow_k8_cpu_exit_acpi(data);
1056                 kfree(data->powernow_table);
1057                 kfree(data);
1058                 return -EINVAL;
1059         }
1060
1061         cpufreq_frequency_table_get_attr(data->powernow_table, pol->cpu);
1062
1063         printk("cpu_init done, current fid 0x%x, vid 0x%x\n",
1064                data->currfid, data->currvid);
1065
1066         for_each_cpu_mask(i, cpu_core_map[pol->cpu]) {
1067                 powernow_data[i] = data;
1068         }
1069
1070         return 0;
1071
1072 err_out:
1073         set_cpus_allowed(current, oldmask);
1074         schedule();
1075         powernow_k8_cpu_exit_acpi(data);
1076
1077         kfree(data);
1078         return -ENODEV;
1079 }
1080
1081 static int __devexit powernowk8_cpu_exit (struct cpufreq_policy *pol)
1082 {
1083         struct powernow_k8_data *data = powernow_data[pol->cpu];
1084
1085         if (!data)
1086                 return -EINVAL;
1087
1088         powernow_k8_cpu_exit_acpi(data);
1089
1090         cpufreq_frequency_table_put_attr(pol->cpu);
1091
1092         kfree(data->powernow_table);
1093         kfree(data);
1094
1095         return 0;
1096 }
1097
1098 static unsigned int powernowk8_get (unsigned int cpu)
1099 {
1100         struct powernow_k8_data *data = powernow_data[cpu];
1101         cpumask_t oldmask = current->cpus_allowed;
1102         unsigned int khz = 0;
1103
1104         set_cpus_allowed(current, cpumask_of_cpu(cpu));
1105         if (smp_processor_id() != cpu) {
1106                 printk(KERN_ERR PFX "limiting to CPU %d failed in powernowk8_get\n", cpu);
1107                 set_cpus_allowed(current, oldmask);
1108                 return 0;
1109         }
1110         preempt_disable();
1111
1112         if (query_current_values_with_pending_wait(data))
1113                 goto out;
1114
1115         khz = find_khz_freq_from_fid(data->currfid);
1116
1117  out:
1118         preempt_enable_no_resched();
1119         set_cpus_allowed(current, oldmask);
1120
1121         return khz;
1122 }
1123
1124 static struct freq_attr* powernow_k8_attr[] = {
1125         &cpufreq_freq_attr_scaling_available_freqs,
1126         NULL,
1127 };
1128
1129 static struct cpufreq_driver cpufreq_amd64_driver = {
1130         .verify = powernowk8_verify,
1131         .target = powernowk8_target,
1132         .init = powernowk8_cpu_init,
1133         .exit = __devexit_p(powernowk8_cpu_exit),
1134         .get = powernowk8_get,
1135         .name = "powernow-k8",
1136         .owner = THIS_MODULE,
1137         .attr = powernow_k8_attr,
1138 };
1139
1140 /* driver entry point for init */
1141 static int __init powernowk8_init(void)
1142 {
1143         unsigned int i, supported_cpus = 0;
1144
1145         for (i=0; i<NR_CPUS; i++) {
1146                 if (!cpu_online(i))
1147                         continue;
1148                 if (check_supported_cpu(i))
1149                         supported_cpus++;
1150         }
1151
1152         if (supported_cpus == num_online_cpus()) {
1153                 printk(KERN_INFO PFX "Found %d AMD Athlon 64 / Opteron processors (" VERSION ")\n",
1154                         supported_cpus);
1155                 return cpufreq_register_driver(&cpufreq_amd64_driver);
1156         }
1157
1158         return -ENODEV;
1159 }
1160
1161 /* driver entry point for term */
1162 static void __exit powernowk8_exit(void)
1163 {
1164         dprintk("exit\n");
1165
1166         cpufreq_unregister_driver(&cpufreq_amd64_driver);
1167 }
1168
1169 MODULE_AUTHOR("Paul Devriendt <paul.devriendt@amd.com> and Mark Langsdorf <mark.langsdorf@amd.com.");
1170 MODULE_DESCRIPTION("AMD Athlon 64 and Opteron processor frequency driver.");
1171 MODULE_LICENSE("GPL");
1172
1173 late_initcall(powernowk8_init);
1174 module_exit(powernowk8_exit);
1175