drivers/cpufreq/powernow-k8.c

   1 /*
   2  *   (c) 2003-2012 Advanced Micro Devices, Inc.
   3  *  Your use of this code is subject to the terms and conditions of the
   4  *  GNU general public license version 2. See "COPYING" or
   5  *  http://www.gnu.org/licenses/gpl.html
   6  *
   7  *  Maintainer:
   8  *  Andreas Herrmann <andreas.herrmann3@amd.com>
   9  *
  10  *  Based on the powernow-k7.c module written by Dave Jones.
  11  *  (C) 2003 Dave Jones on behalf of SuSE Labs
  12  *  (C) 2004 Dominik Brodowski <linux@brodo.de>
  13  *  (C) 2004 Pavel Machek <pavel@ucw.cz>
  14  *  Licensed under the terms of the GNU GPL License version 2.
  15  *  Based upon datasheets & sample CPUs kindly provided by AMD.
  16  *
  17  *  Valuable input gratefully received from Dave Jones, Pavel Machek,
  18  *  Dominik Brodowski, Jacob Shin, and others.
  19  *  Originally developed by Paul Devriendt.
  20  *
  21  *  Processor information obtained from Chapter 9 (Power and Thermal
  22  *  Management) of the "BIOS and Kernel Developer's Guide (BKDG) for
  23  *  the AMD Athlon 64 and AMD Opteron Processors" and section "2.x
  24  *  Power Management" in BKDGs for newer AMD CPU families.
  25  *
  26  *  Tables for specific CPUs can be inferred from AMD's processor
  27  *  power and thermal data sheets, (e.g. 30417.pdf, 30430.pdf, 43375.pdf)
  28  */
  29
  30 #include <linux/kernel.h>
  31 #include <linux/smp.h>
  32 #include <linux/module.h>
  33 #include <linux/init.h>
  34 #include <linux/cpufreq.h>
  35 #include <linux/slab.h>
  36 #include <linux/string.h>
  37 #include <linux/cpumask.h>
  38 #include <linux/sched.h>        /* for current / set_cpus_allowed() */
  39 #include <linux/io.h>
  40 #include <linux/delay.h>
  41
  42 #include <asm/msr.h>
  43
  44 #include <linux/acpi.h>
  45 #include <linux/mutex.h>
  46 #include <acpi/processor.h>
  47
  48 #define PFX "powernow-k8: "
  49 #define VERSION "version 2.20.00"
  50 #include "powernow-k8.h"
  51 #include "mperf.h"
  52
  53 /* serialize freq changes  */
  54 static DEFINE_MUTEX(fidvid_mutex);
  55
  56 static DEFINE_PER_CPU(struct powernow_k8_data *, powernow_data);
  57
  58 static int cpu_family = CPU_OPTERON;
  59
  60 /* array to map SW pstate number to acpi state */
  61 static u32 ps_to_as[8];
  62
  63 /* core performance boost */
  64 static bool cpb_capable, cpb_enabled;
  65 static struct msr __percpu *msrs;
  66
  67 static struct cpufreq_driver cpufreq_amd64_driver;
  68
  69 #ifndef CONFIG_SMP
  70 static inline const struct cpumask *cpu_core_mask(int cpu)
  71 {
  72         return cpumask_of(0);
  73 }
  74 #endif
  75
  76 /* Return a frequency in MHz, given an input fid */
  77 static u32 find_freq_from_fid(u32 fid)
  78 {
  79         return 800 + (fid * 100);
  80 }
  81
  82 /* Return a frequency in KHz, given an input fid */
  83 static u32 find_khz_freq_from_fid(u32 fid)
  84 {
  85         return 1000 * find_freq_from_fid(fid);
  86 }
  87
  88 static u32 find_khz_freq_from_pstate(struct cpufreq_frequency_table *data,
  89                                      u32 pstate)
  90 {
  91         return data[ps_to_as[pstate]].frequency;
  92 }
  93
  94 /* Return the vco fid for an input fid
  95  *
  96  * Each "low" fid has corresponding "high" fid, and you can get to "low" fids
  97  * only from corresponding high fids. This returns "high" fid corresponding to
  98  * "low" one.
  99  */
 100 static u32 convert_fid_to_vco_fid(u32 fid)
 101 {
 102         if (fid < HI_FID_TABLE_BOTTOM)
 103                 return 8 + (2 * fid);
 104         else
 105                 return fid;
 106 }
 107
 108 /*
 109  * Return 1 if the pending bit is set. Unless we just instructed the processor
 110  * to transition to a new state, seeing this bit set is really bad news.
 111  */
 112 static int pending_bit_stuck(void)
 113 {
 114         u32 lo, hi;
 115
 116         if (cpu_family == CPU_HW_PSTATE)
 117                 return 0;
 118
 119         rdmsr(MSR_FIDVID_STATUS, lo, hi);
 120         return lo & MSR_S_LO_CHANGE_PENDING ? 1 : 0;
 121 }
 122
 123 /*
 124  * Update the global current fid / vid values from the status msr.
 125  * Returns 1 on error.
 126  */
 127 static int query_current_values_with_pending_wait(struct powernow_k8_data *data)
 128 {
 129         u32 lo, hi;
 130         u32 i = 0;
 131
 132         if (cpu_family == CPU_HW_PSTATE) {
 133                 rdmsr(MSR_PSTATE_STATUS, lo, hi);
 134                 i = lo & HW_PSTATE_MASK;
 135                 data->currpstate = i;
 136
 137                 /*
 138                  * a workaround for family 11h erratum 311 might cause
 139                  * an "out-of-range Pstate if the core is in Pstate-0
 140                  */
 141                 if ((boot_cpu_data.x86 == 0x11) && (i >= data->numps))
 142                         data->currpstate = HW_PSTATE_0;
 143
 144                 return 0;
 145         }
 146         do {
 147                 if (i++ > 10000) {
 148                         pr_debug("detected change pending stuck\n");
 149                         return 1;
 150                 }
 151                 rdmsr(MSR_FIDVID_STATUS, lo, hi);
 152         } while (lo & MSR_S_LO_CHANGE_PENDING);
 153
 154         data->currvid = hi & MSR_S_HI_CURRENT_VID;
 155         data->currfid = lo & MSR_S_LO_CURRENT_FID;
 156
 157         return 0;
 158 }
 159
 160 /* the isochronous relief time */
 161 static void count_off_irt(struct powernow_k8_data *data)
 162 {
 163         udelay((1 << data->irt) * 10);
 164         return;
 165 }
 166
 167 /* the voltage stabilization time */
 168 static void count_off_vst(struct powernow_k8_data *data)
 169 {
 170         udelay(data->vstable * VST_UNITS_20US);
 171         return;
 172 }
 173
 174 /* need to init the control msr to a safe value (for each cpu) */
 175 static void fidvid_msr_init(void)
 176 {
 177         u32 lo, hi;
 178         u8 fid, vid;
 179
 180         rdmsr(MSR_FIDVID_STATUS, lo, hi);
 181         vid = hi & MSR_S_HI_CURRENT_VID;
 182         fid = lo & MSR_S_LO_CURRENT_FID;
 183         lo = fid | (vid << MSR_C_LO_VID_SHIFT);
 184         hi = MSR_C_HI_STP_GNT_BENIGN;
 185         pr_debug("cpu%d, init lo 0x%x, hi 0x%x\n", smp_processor_id(), lo, hi);
 186         wrmsr(MSR_FIDVID_CTL, lo, hi);
 187 }
 188
 189 /* write the new fid value along with the other control fields to the msr */
 190 static int write_new_fid(struct powernow_k8_data *data, u32 fid)
 191 {
 192         u32 lo;
 193         u32 savevid = data->currvid;
 194         u32 i = 0;
 195
 196         if ((fid & INVALID_FID_MASK) || (data->currvid & INVALID_VID_MASK)) {
 197                 printk(KERN_ERR PFX "internal error - overflow on fid write\n");
 198                 return 1;
 199         }
 200
 201         lo = fid;
 202         lo |= (data->currvid << MSR_C_LO_VID_SHIFT);
 203         lo |= MSR_C_LO_INIT_FID_VID;
 204
 205         pr_debug("writing fid 0x%x, lo 0x%x, hi 0x%x\n",
 206                 fid, lo, data->plllock * PLL_LOCK_CONVERSION);
 207
 208         do {
 209                 wrmsr(MSR_FIDVID_CTL, lo, data->plllock * PLL_LOCK_CONVERSION);
 210                 if (i++ > 100) {
 211                         printk(KERN_ERR PFX
 212                                 "Hardware error - pending bit very stuck - "
 213                                 "no further pstate changes possible\n");
 214                         return 1;
 215                 }
 216         } while (query_current_values_with_pending_wait(data));
 217
 218         count_off_irt(data);
 219
 220         if (savevid != data->currvid) {
 221                 printk(KERN_ERR PFX
 222                         "vid change on fid trans, old 0x%x, new 0x%x\n",
 223                         savevid, data->currvid);
 224                 return 1;
 225         }
 226
 227         if (fid != data->currfid) {
 228                 printk(KERN_ERR PFX
 229                         "fid trans failed, fid 0x%x, curr 0x%x\n", fid,
 230                         data->currfid);
 231                 return 1;
 232         }
 233
 234         return 0;
 235 }
 236
 237 /* Write a new vid to the hardware */
 238 static int write_new_vid(struct powernow_k8_data *data, u32 vid)
 239 {
 240         u32 lo;
 241         u32 savefid = data->currfid;
 242         int i = 0;
 243
 244         if ((data->currfid & INVALID_FID_MASK) || (vid & INVALID_VID_MASK)) {
 245                 printk(KERN_ERR PFX "internal error - overflow on vid write\n");
 246                 return 1;
 247         }
 248
 249         lo = data->currfid;
 250         lo |= (vid << MSR_C_LO_VID_SHIFT);
 251         lo |= MSR_C_LO_INIT_FID_VID;
 252
 253         pr_debug("writing vid 0x%x, lo 0x%x, hi 0x%x\n",
 254                 vid, lo, STOP_GRANT_5NS);
 255
 256         do {
 257                 wrmsr(MSR_FIDVID_CTL, lo, STOP_GRANT_5NS);
 258                 if (i++ > 100) {
 259                         printk(KERN_ERR PFX "internal error - pending bit "
 260                                         "very stuck - no further pstate "
 261                                         "changes possible\n");
 262                         return 1;
 263                 }
 264         } while (query_current_values_with_pending_wait(data));
 265
 266         if (savefid != data->currfid) {
 267                 printk(KERN_ERR PFX "fid changed on vid trans, old "
 268                         "0x%x new 0x%x\n",
 269                        savefid, data->currfid);
 270                 return 1;
 271         }
 272
 273         if (vid != data->currvid) {
 274                 printk(KERN_ERR PFX "vid trans failed, vid 0x%x, "
 275                                 "curr 0x%x\n",
 276                                 vid, data->currvid);
 277                 return 1;
 278         }
 279
 280         return 0;
 281 }
 282
 283 /*
 284  * Reduce the vid by the max of step or reqvid.
 285  * Decreasing vid codes represent increasing voltages:
 286  * vid of 0 is 1.550V, vid of 0x1e is 0.800V, vid of VID_OFF is off.
 287  */
 288 static int decrease_vid_code_by_step(struct powernow_k8_data *data,
 289                 u32 reqvid, u32 step)
 290 {
 291         if ((data->currvid - reqvid) > step)
 292                 reqvid = data->currvid - step;
 293
 294         if (write_new_vid(data, reqvid))
 295                 return 1;
 296
 297         count_off_vst(data);
 298
 299         return 0;
 300 }
 301
 302 /* Change hardware pstate by single MSR write */
 303 static int transition_pstate(struct powernow_k8_data *data, u32 pstate)
 304 {
 305         wrmsr(MSR_PSTATE_CTRL, pstate, 0);
 306         data->currpstate = pstate;
 307         return 0;
 308 }
 309
 310 /* Change Opteron/Athlon64 fid and vid, by the 3 phases. */
 311 static int transition_fid_vid(struct powernow_k8_data *data,
 312                 u32 reqfid, u32 reqvid)
 313 {
 314         if (core_voltage_pre_transition(data, reqvid, reqfid))
 315                 return 1;
 316
 317         if (core_frequency_transition(data, reqfid))
 318                 return 1;
 319
 320         if (core_voltage_post_transition(data, reqvid))
 321                 return 1;
 322
 323         if (query_current_values_with_pending_wait(data))
 324                 return 1;
 325
 326         if ((reqfid != data->currfid) || (reqvid != data->currvid)) {
 327                 printk(KERN_ERR PFX "failed (cpu%d): req 0x%x 0x%x, "
 328                                 "curr 0x%x 0x%x\n",
 329                                 smp_processor_id(),
 330                                 reqfid, reqvid, data->currfid, data->currvid);
 331                 return 1;
 332         }
 333
 334         pr_debug("transitioned (cpu%d): new fid 0x%x, vid 0x%x\n",
 335                 smp_processor_id(), data->currfid, data->currvid);
 336
 337         return 0;
 338 }
 339
 340 /* Phase 1 - core voltage transition ... setup voltage */
 341 static int core_voltage_pre_transition(struct powernow_k8_data *data,
 342                 u32 reqvid, u32 reqfid)
 343 {
 344         u32 rvosteps = data->rvo;
 345         u32 savefid = data->currfid;
 346         u32 maxvid, lo, rvomult = 1;
 347
 348         pr_debug("ph1 (cpu%d): start, currfid 0x%x, currvid 0x%x, "
 349                 "reqvid 0x%x, rvo 0x%x\n",
 350                 smp_processor_id(),
 351                 data->currfid, data->currvid, reqvid, data->rvo);
 352
 353         if ((savefid < LO_FID_TABLE_TOP) && (reqfid < LO_FID_TABLE_TOP))
 354                 rvomult = 2;
 355         rvosteps *= rvomult;
 356         rdmsr(MSR_FIDVID_STATUS, lo, maxvid);
 357         maxvid = 0x1f & (maxvid >> 16);
 358         pr_debug("ph1 maxvid=0x%x\n", maxvid);
 359         if (reqvid < maxvid) /* lower numbers are higher voltages */
 360                 reqvid = maxvid;
 361
 362         while (data->currvid > reqvid) {
 363                 pr_debug("ph1: curr 0x%x, req vid 0x%x\n",
 364                         data->currvid, reqvid);
 365                 if (decrease_vid_code_by_step(data, reqvid, data->vidmvs))
 366                         return 1;
 367         }
 368
 369         while ((rvosteps > 0) &&
 370                         ((rvomult * data->rvo + data->currvid) > reqvid)) {
 371                 if (data->currvid == maxvid) {
 372                         rvosteps = 0;
 373                 } else {
 374                         pr_debug("ph1: changing vid for rvo, req 0x%x\n",
 375                                 data->currvid - 1);
 376                         if (decrease_vid_code_by_step(data, data->currvid-1, 1))
 377                                 return 1;
 378                         rvosteps--;
 379                 }
 380         }
 381
 382         if (query_current_values_with_pending_wait(data))
 383                 return 1;
 384
 385         if (savefid != data->currfid) {
 386                 printk(KERN_ERR PFX "ph1 err, currfid changed 0x%x\n",
 387                                 data->currfid);
 388                 return 1;
 389         }
 390
 391         pr_debug("ph1 complete, currfid 0x%x, currvid 0x%x\n",
 392                 data->currfid, data->currvid);
 393
 394         return 0;
 395 }
 396
 397 /* Phase 2 - core frequency transition */
 398 static int core_frequency_transition(struct powernow_k8_data *data, u32 reqfid)
 399 {
 400         u32 vcoreqfid, vcocurrfid, vcofiddiff;
 401         u32 fid_interval, savevid = data->currvid;
 402
 403         if (data->currfid == reqfid) {
 404                 printk(KERN_ERR PFX "ph2 null fid transition 0x%x\n",
 405                                 data->currfid);
 406                 return 0;
 407         }
 408
 409         pr_debug("ph2 (cpu%d): starting, currfid 0x%x, currvid 0x%x, "
 410                 "reqfid 0x%x\n",
 411                 smp_processor_id(),
 412                 data->currfid, data->currvid, reqfid);
 413
 414         vcoreqfid = convert_fid_to_vco_fid(reqfid);
 415         vcocurrfid = convert_fid_to_vco_fid(data->currfid);
 416         vcofiddiff = vcocurrfid > vcoreqfid ? vcocurrfid - vcoreqfid
 417             : vcoreqfid - vcocurrfid;
 418
 419         if ((reqfid <= LO_FID_TABLE_TOP) && (data->currfid <= LO_FID_TABLE_TOP))
 420                 vcofiddiff = 0;
 421
 422         while (vcofiddiff > 2) {
 423                 (data->currfid & 1) ? (fid_interval = 1) : (fid_interval = 2);
 424
 425                 if (reqfid > data->currfid) {
 426                         if (data->currfid > LO_FID_TABLE_TOP) {
 427                                 if (write_new_fid(data,
 428                                                 data->currfid + fid_interval))
 429                                         return 1;
 430                         } else {
 431                                 if (write_new_fid
 432                                     (data,
 433                                      2 + convert_fid_to_vco_fid(data->currfid)))
 434                                         return 1;
 435                         }
 436                 } else {
 437                         if (write_new_fid(data, data->currfid - fid_interval))
 438                                 return 1;
 439                 }
 440
 441                 vcocurrfid = convert_fid_to_vco_fid(data->currfid);
 442                 vcofiddiff = vcocurrfid > vcoreqfid ? vcocurrfid - vcoreqfid
 443                     : vcoreqfid - vcocurrfid;
 444         }
 445
 446         if (write_new_fid(data, reqfid))
 447                 return 1;
 448
 449         if (query_current_values_with_pending_wait(data))
 450                 return 1;
 451
 452         if (data->currfid != reqfid) {
 453                 printk(KERN_ERR PFX
 454                         "ph2: mismatch, failed fid transition, "
 455                         "curr 0x%x, req 0x%x\n",
 456                         data->currfid, reqfid);
 457                 return 1;
 458         }
 459
 460         if (savevid != data->currvid) {
 461                 printk(KERN_ERR PFX "ph2: vid changed, save 0x%x, curr 0x%x\n",
 462                         savevid, data->currvid);
 463                 return 1;
 464         }
 465
 466         pr_debug("ph2 complete, currfid 0x%x, currvid 0x%x\n",
 467                 data->currfid, data->currvid);
 468
 469         return 0;
 470 }
 471
 472 /* Phase 3 - core voltage transition flow ... jump to the final vid. */
 473 static int core_voltage_post_transition(struct powernow_k8_data *data,
 474                 u32 reqvid)
 475 {
 476         u32 savefid = data->currfid;
 477         u32 savereqvid = reqvid;
 478
 479         pr_debug("ph3 (cpu%d): starting, currfid 0x%x, currvid 0x%x\n",
 480                 smp_processor_id(),
 481                 data->currfid, data->currvid);
 482
 483         if (reqvid != data->currvid) {
 484                 if (write_new_vid(data, reqvid))
 485                         return 1;
 486
 487                 if (savefid != data->currfid) {
 488                         printk(KERN_ERR PFX
 489                                "ph3: bad fid change, save 0x%x, curr 0x%x\n",
 490                                savefid, data->currfid);
 491                         return 1;
 492                 }
 493
 494                 if (data->currvid != reqvid) {
 495                         printk(KERN_ERR PFX
 496                                "ph3: failed vid transition\n, "
 497                                "req 0x%x, curr 0x%x",
 498                                reqvid, data->currvid);
 499                         return 1;
 500                 }
 501         }
 502
 503         if (query_current_values_with_pending_wait(data))
 504                 return 1;
 505
 506         if (savereqvid != data->currvid) {
 507                 pr_debug("ph3 failed, currvid 0x%x\n", data->currvid);
 508                 return 1;
 509         }
 510
 511         if (savefid != data->currfid) {
 512                 pr_debug("ph3 failed, currfid changed 0x%x\n",
 513                         data->currfid);
 514                 return 1;
 515         }
 516
 517         pr_debug("ph3 complete, currfid 0x%x, currvid 0x%x\n",
 518                 data->currfid, data->currvid);
 519
 520         return 0;
 521 }
 522
 523 static void check_supported_cpu(void *_rc)
 524 {
 525         u32 eax, ebx, ecx, edx;
 526         int *rc = _rc;
 527
 528         *rc = -ENODEV;
 529
 530         if (__this_cpu_read(cpu_info.x86_vendor) != X86_VENDOR_AMD)
 531                 return;
 532
 533         eax = cpuid_eax(CPUID_PROCESSOR_SIGNATURE);
 534         if (((eax & CPUID_XFAM) != CPUID_XFAM_K8) &&
 535             ((eax & CPUID_XFAM) < CPUID_XFAM_10H))
 536                 return;
 537
 538         if ((eax & CPUID_XFAM) == CPUID_XFAM_K8) {
 539                 if (((eax & CPUID_USE_XFAM_XMOD) != CPUID_USE_XFAM_XMOD) ||
 540                     ((eax & CPUID_XMOD) > CPUID_XMOD_REV_MASK)) {
 541                         printk(KERN_INFO PFX
 542                                 "Processor cpuid %x not supported\n", eax);
 543                         return;
 544                 }
 545
 546                 eax = cpuid_eax(CPUID_GET_MAX_CAPABILITIES);
 547                 if (eax < CPUID_FREQ_VOLT_CAPABILITIES) {
 548                         printk(KERN_INFO PFX
 549                                "No frequency change capabilities detected\n");
 550                         return;
 551                 }
 552
 553                 cpuid(CPUID_FREQ_VOLT_CAPABILITIES, &eax, &ebx, &ecx, &edx);
 554                 if ((edx & P_STATE_TRANSITION_CAPABLE)
 555                         != P_STATE_TRANSITION_CAPABLE) {
 556                         printk(KERN_INFO PFX
 557                                 "Power state transitions not supported\n");
 558                         return;
 559                 }
 560         } else { /* must be a HW Pstate capable processor */
 561                 cpuid(CPUID_FREQ_VOLT_CAPABILITIES, &eax, &ebx, &ecx, &edx);
 562                 if ((edx & USE_HW_PSTATE) == USE_HW_PSTATE)
 563                         cpu_family = CPU_HW_PSTATE;
 564                 else
 565                         return;
 566         }
 567
 568         *rc = 0;
 569 }
 570
 571 static int check_pst_table(struct powernow_k8_data *data, struct pst_s *pst,
 572                 u8 maxvid)
 573 {
 574         unsigned int j;
 575         u8 lastfid = 0xff;
 576
 577         for (j = 0; j < data->numps; j++) {
 578                 if (pst[j].vid > LEAST_VID) {
 579                         printk(KERN_ERR FW_BUG PFX "vid %d invalid : 0x%x\n",
 580                                j, pst[j].vid);
 581                         return -EINVAL;
 582                 }
 583                 if (pst[j].vid < data->rvo) {
 584                         /* vid + rvo >= 0 */
 585                         printk(KERN_ERR FW_BUG PFX "0 vid exceeded with pstate"
 586                                " %d\n", j);
 587                         return -ENODEV;
 588                 }
 589                 if (pst[j].vid < maxvid + data->rvo) {
 590                         /* vid + rvo >= maxvid */
 591                         printk(KERN_ERR FW_BUG PFX "maxvid exceeded with pstate"
 592                                " %d\n", j);
 593                         return -ENODEV;
 594                 }
 595                 if (pst[j].fid > MAX_FID) {
 596                         printk(KERN_ERR FW_BUG PFX "maxfid exceeded with pstate"
 597                                " %d\n", j);
 598                         return -ENODEV;
 599                 }
 600                 if (j && (pst[j].fid < HI_FID_TABLE_BOTTOM)) {
 601                         /* Only first fid is allowed to be in "low" range */
 602                         printk(KERN_ERR FW_BUG PFX "two low fids - %d : "
 603                                "0x%x\n", j, pst[j].fid);
 604                         return -EINVAL;
 605                 }
 606                 if (pst[j].fid < lastfid)
 607                         lastfid = pst[j].fid;
 608         }
 609         if (lastfid & 1) {
 610                 printk(KERN_ERR FW_BUG PFX "lastfid invalid\n");
 611                 return -EINVAL;
 612         }
 613         if (lastfid > LO_FID_TABLE_TOP)
 614                 printk(KERN_INFO FW_BUG PFX
 615                         "first fid not from lo freq table\n");
 616
 617         return 0;
 618 }
 619
 620 static void invalidate_entry(struct cpufreq_frequency_table *powernow_table,
 621                 unsigned int entry)
 622 {
 623         powernow_table[entry].frequency = CPUFREQ_ENTRY_INVALID;
 624 }
 625
 626 static void print_basics(struct powernow_k8_data *data)
 627 {
 628         int j;
 629         for (j = 0; j < data->numps; j++) {
 630                 if (data->powernow_table[j].frequency !=
 631                                 CPUFREQ_ENTRY_INVALID) {
 632                         if (cpu_family == CPU_HW_PSTATE) {
 633                                 printk(KERN_INFO PFX
 634                                         "   %d : pstate %d (%d MHz)\n", j,
 635                                         data->powernow_table[j].index,
 636                                         data->powernow_table[j].frequency/1000);
 637                         } else {
 638                                 printk(KERN_INFO PFX
 639                                         "fid 0x%x (%d MHz), vid 0x%x\n",
 640                                         data->powernow_table[j].index & 0xff,
 641                                         data->powernow_table[j].frequency/1000,
 642                                         data->powernow_table[j].index >> 8);
 643                         }
 644                 }
 645         }
 646         if (data->batps)
 647                 printk(KERN_INFO PFX "Only %d pstates on battery\n",
 648                                 data->batps);
 649 }
 650
 651 static u32 freq_from_fid_did(u32 fid, u32 did)
 652 {
 653         u32 mhz = 0;
 654
 655         if (boot_cpu_data.x86 == 0x10)
 656                 mhz = (100 * (fid + 0x10)) >> did;
 657         else if (boot_cpu_data.x86 == 0x11)
 658                 mhz = (100 * (fid + 8)) >> did;
 659         else
 660                 BUG();
 661
 662         return mhz * 1000;
 663 }
 664
 665 static int fill_powernow_table(struct powernow_k8_data *data,
 666                 struct pst_s *pst, u8 maxvid)
 667 {
 668         struct cpufreq_frequency_table *powernow_table;
 669         unsigned int j;
 670
 671         if (data->batps) {
 672                 /* use ACPI support to get full speed on mains power */
 673                 printk(KERN_WARNING PFX
 674                         "Only %d pstates usable (use ACPI driver for full "
 675                         "range\n", data->batps);
 676                 data->numps = data->batps;
 677         }
 678
 679         for (j = 1; j < data->numps; j++) {
 680                 if (pst[j-1].fid >= pst[j].fid) {
 681                         printk(KERN_ERR PFX "PST out of sequence\n");
 682                         return -EINVAL;
 683                 }
 684         }
 685
 686         if (data->numps < 2) {
 687                 printk(KERN_ERR PFX "no p states to transition\n");
 688                 return -ENODEV;
 689         }
 690
 691         if (check_pst_table(data, pst, maxvid))
 692                 return -EINVAL;
 693
 694         powernow_table = kmalloc((sizeof(struct cpufreq_frequency_table)
 695                 * (data->numps + 1)), GFP_KERNEL);
 696         if (!powernow_table) {
 697                 printk(KERN_ERR PFX "powernow_table memory alloc failure\n");
 698                 return -ENOMEM;
 699         }
 700
 701         for (j = 0; j < data->numps; j++) {
 702                 int freq;
 703                 powernow_table[j].index = pst[j].fid; /* lower 8 bits */
 704                 powernow_table[j].index |= (pst[j].vid << 8); /* upper 8 bits */
 705                 freq = find_khz_freq_from_fid(pst[j].fid);
 706                 powernow_table[j].frequency = freq;
 707         }
 708         powernow_table[data->numps].frequency = CPUFREQ_TABLE_END;
 709         powernow_table[data->numps].index = 0;
 710
 711         if (query_current_values_with_pending_wait(data)) {
 712                 kfree(powernow_table);
 713                 return -EIO;
 714         }
 715
 716         pr_debug("cfid 0x%x, cvid 0x%x\n", data->currfid, data->currvid);
 717         data->powernow_table = powernow_table;
 718         if (cpumask_first(cpu_core_mask(data->cpu)) == data->cpu)
 719                 print_basics(data);
 720
 721         for (j = 0; j < data->numps; j++)
 722                 if ((pst[j].fid == data->currfid) &&
 723                     (pst[j].vid == data->currvid))
 724                         return 0;
 725
 726         pr_debug("currfid/vid do not match PST, ignoring\n");
 727         return 0;
 728 }
 729
 730 /* Find and validate the PSB/PST table in BIOS. */
 731 static int find_psb_table(struct powernow_k8_data *data)
 732 {
 733         struct psb_s *psb;
 734         unsigned int i;
 735         u32 mvs;
 736         u8 maxvid;
 737         u32 cpst = 0;
 738         u32 thiscpuid;
 739
 740         for (i = 0xc0000; i < 0xffff0; i += 0x10) {
 741                 /* Scan BIOS looking for the signature. */
 742                 /* It can not be at ffff0 - it is too big. */
 743
 744                 psb = phys_to_virt(i);
 745                 if (memcmp(psb, PSB_ID_STRING, PSB_ID_STRING_LEN) != 0)
 746                         continue;
 747
 748                 pr_debug("found PSB header at 0x%p\n", psb);
 749
 750                 pr_debug("table vers: 0x%x\n", psb->tableversion);
 751                 if (psb->tableversion != PSB_VERSION_1_4) {
 752                         printk(KERN_ERR FW_BUG PFX "PSB table is not v1.4\n");
 753                         return -ENODEV;
 754                 }
 755
 756                 pr_debug("flags: 0x%x\n", psb->flags1);
 757                 if (psb->flags1) {
 758                         printk(KERN_ERR FW_BUG PFX "unknown flags\n");
 759                         return -ENODEV;
 760                 }
 761
 762                 data->vstable = psb->vstable;
 763                 pr_debug("voltage stabilization time: %d(*20us)\n",
 764                                 data->vstable);
 765
 766                 pr_debug("flags2: 0x%x\n", psb->flags2);
 767                 data->rvo = psb->flags2 & 3;
 768                 data->irt = ((psb->flags2) >> 2) & 3;
 769                 mvs = ((psb->flags2) >> 4) & 3;
 770                 data->vidmvs = 1 << mvs;
 771                 data->batps = ((psb->flags2) >> 6) & 3;
 772
 773                 pr_debug("ramp voltage offset: %d\n", data->rvo);
 774                 pr_debug("isochronous relief time: %d\n", data->irt);
 775                 pr_debug("maximum voltage step: %d - 0x%x\n", mvs, data->vidmvs);
 776
 777                 pr_debug("numpst: 0x%x\n", psb->num_tables);
 778                 cpst = psb->num_tables;
 779                 if ((psb->cpuid == 0x00000fc0) ||
 780                     (psb->cpuid == 0x00000fe0)) {
 781                         thiscpuid = cpuid_eax(CPUID_PROCESSOR_SIGNATURE);
 782                         if ((thiscpuid == 0x00000fc0) ||
 783                             (thiscpuid == 0x00000fe0))
 784                                 cpst = 1;
 785                 }
 786                 if (cpst != 1) {
 787                         printk(KERN_ERR FW_BUG PFX "numpst must be 1\n");
 788                         return -ENODEV;
 789                 }
 790
 791                 data->plllock = psb->plllocktime;
 792                 pr_debug("plllocktime: 0x%x (units 1us)\n", psb->plllocktime);
 793                 pr_debug("maxfid: 0x%x\n", psb->maxfid);
 794                 pr_debug("maxvid: 0x%x\n", psb->maxvid);
 795                 maxvid = psb->maxvid;
 796
 797                 data->numps = psb->numps;
 798                 pr_debug("numpstates: 0x%x\n", data->numps);
 799                 return fill_powernow_table(data,
 800                                 (struct pst_s *)(psb+1), maxvid);
 801         }
 802         /*
 803          * If you see this message, complain to BIOS manufacturer. If
 804          * he tells you "we do not support Linux" or some similar
 805          * nonsense, remember that Windows 2000 uses the same legacy
 806          * mechanism that the old Linux PSB driver uses. Tell them it
 807          * is broken with Windows 2000.
 808          *
 809          * The reference to the AMD documentation is chapter 9 in the
 810          * BIOS and Kernel Developer's Guide, which is available on
 811          * www.amd.com
 812          */
 813         printk(KERN_ERR FW_BUG PFX "No PSB or ACPI _PSS objects\n");
 814         printk(KERN_ERR PFX "Make sure that your BIOS is up to date"
 815                 " and Cool'N'Quiet support is enabled in BIOS setup\n");
 816         return -ENODEV;
 817 }
 818
 819 static void powernow_k8_acpi_pst_values(struct powernow_k8_data *data,
 820                 unsigned int index)
 821 {
 822         u64 control;
 823
 824         if (!data->acpi_data.state_count || (cpu_family == CPU_HW_PSTATE))
 825                 return;
 826
 827         control = data->acpi_data.states[index].control;
 828         data->irt = (control >> IRT_SHIFT) & IRT_MASK;
 829         data->rvo = (control >> RVO_SHIFT) & RVO_MASK;
 830         data->exttype = (control >> EXT_TYPE_SHIFT) & EXT_TYPE_MASK;
 831         data->plllock = (control >> PLL_L_SHIFT) & PLL_L_MASK;
 832         data->vidmvs = 1 << ((control >> MVS_SHIFT) & MVS_MASK);
 833         data->vstable = (control >> VST_SHIFT) & VST_MASK;
 834 }
 835
 836 static int powernow_k8_cpu_init_acpi(struct powernow_k8_data *data)
 837 {
 838         struct cpufreq_frequency_table *powernow_table;
 839         int ret_val = -ENODEV;
 840         u64 control, status;
 841
 842         if (acpi_processor_register_performance(&data->acpi_data, data->cpu)) {
 843                 pr_debug("register performance failed: bad ACPI data\n");
 844                 return -EIO;
 845         }
 846
 847         /* verify the data contained in the ACPI structures */
 848         if (data->acpi_data.state_count <= 1) {
 849                 pr_debug("No ACPI P-States\n");
 850                 goto err_out;
 851         }
 852
 853         control = data->acpi_data.control_register.space_id;
 854         status = data->acpi_data.status_register.space_id;
 855
 856         if ((control != ACPI_ADR_SPACE_FIXED_HARDWARE) ||
 857             (status != ACPI_ADR_SPACE_FIXED_HARDWARE)) {
 858                 pr_debug("Invalid control/status registers (%llx - %llx)\n",
 859                         control, status);
 860                 goto err_out;
 861         }
 862
 863         /* fill in data->powernow_table */
 864         powernow_table = kmalloc((sizeof(struct cpufreq_frequency_table)
 865                 * (data->acpi_data.state_count + 1)), GFP_KERNEL);
 866         if (!powernow_table) {
 867                 pr_debug("powernow_table memory alloc failure\n");
 868                 goto err_out;
 869         }
 870
 871         /* fill in data */
 872         data->numps = data->acpi_data.state_count;
 873         powernow_k8_acpi_pst_values(data, 0);
 874
 875         if (cpu_family == CPU_HW_PSTATE)
 876                 ret_val = fill_powernow_table_pstate(data, powernow_table);
 877         else
 878                 ret_val = fill_powernow_table_fidvid(data, powernow_table);
 879         if (ret_val)
 880                 goto err_out_mem;
 881
 882         powernow_table[data->acpi_data.state_count].frequency =
 883                 CPUFREQ_TABLE_END;
 884         powernow_table[data->acpi_data.state_count].index = 0;
 885         data->powernow_table = powernow_table;
 886
 887         if (cpumask_first(cpu_core_mask(data->cpu)) == data->cpu)
 888                 print_basics(data);
 889
 890         /* notify BIOS that we exist */
 891         acpi_processor_notify_smm(THIS_MODULE);
 892
 893         if (!zalloc_cpumask_var(&data->acpi_data.shared_cpu_map, GFP_KERNEL)) {
 894                 printk(KERN_ERR PFX
 895                                 "unable to alloc powernow_k8_data cpumask\n");
 896                 ret_val = -ENOMEM;
 897                 goto err_out_mem;
 898         }
 899
 900         return 0;
 901
 902 err_out_mem:
 903         kfree(powernow_table);
 904
 905 err_out:
 906         acpi_processor_unregister_performance(&data->acpi_data, data->cpu);
 907
 908         /* data->acpi_data.state_count informs us at ->exit()
 909          * whether ACPI was used */
 910         data->acpi_data.state_count = 0;
 911
 912         return ret_val;
 913 }
 914
 915 static int fill_powernow_table_pstate(struct powernow_k8_data *data,
 916                 struct cpufreq_frequency_table *powernow_table)
 917 {
 918         int i;
 919         u32 hi = 0, lo = 0;
 920         rdmsr(MSR_PSTATE_CUR_LIMIT, lo, hi);
 921         data->max_hw_pstate = (lo & HW_PSTATE_MAX_MASK) >> HW_PSTATE_MAX_SHIFT;
 922
 923         for (i = 0; i < data->acpi_data.state_count; i++) {
 924                 u32 index;
 925
 926                 index = data->acpi_data.states[i].control & HW_PSTATE_MASK;
 927                 if (index > data->max_hw_pstate) {
 928                         printk(KERN_ERR PFX "invalid pstate %d - "
 929                                         "bad value %d.\n", i, index);
 930                         printk(KERN_ERR PFX "Please report to BIOS "
 931                                         "manufacturer\n");
 932                         invalidate_entry(powernow_table, i);
 933                         continue;
 934                 }
 935
 936                 ps_to_as[index] = i;
 937
 938                 /* Frequency may be rounded for these */
 939                 if ((boot_cpu_data.x86 == 0x10 && boot_cpu_data.x86_model < 10)
 940                                  || boot_cpu_data.x86 == 0x11) {
 941
 942                         rdmsr(MSR_PSTATE_DEF_BASE + index, lo, hi);
 943                         if (!(hi & HW_PSTATE_VALID_MASK)) {
 944                                 pr_debug("invalid pstate %d, ignoring\n", index);
 945                                 invalidate_entry(powernow_table, i);
 946                                 continue;
 947                         }
 948
 949                         powernow_table[i].frequency =
 950                                 freq_from_fid_did(lo & 0x3f, (lo >> 6) & 7);
 951                 } else
 952                         powernow_table[i].frequency =
 953                                 data->acpi_data.states[i].core_frequency * 1000;
 954
 955                 powernow_table[i].index = index;
 956         }
 957         return 0;
 958 }
 959
 960 static int fill_powernow_table_fidvid(struct powernow_k8_data *data,
 961                 struct cpufreq_frequency_table *powernow_table)
 962 {
 963         int i;
 964
 965         for (i = 0; i < data->acpi_data.state_count; i++) {
 966                 u32 fid;
 967                 u32 vid;
 968                 u32 freq, index;
 969                 u64 status, control;
 970
 971                 if (data->exttype) {
 972                         status =  data->acpi_data.states[i].status;
 973                         fid = status & EXT_FID_MASK;
 974                         vid = (status >> VID_SHIFT) & EXT_VID_MASK;
 975                 } else {
 976                         control =  data->acpi_data.states[i].control;
 977                         fid = control & FID_MASK;
 978                         vid = (control >> VID_SHIFT) & VID_MASK;
 979                 }
 980
 981                 pr_debug("   %d : fid 0x%x, vid 0x%x\n", i, fid, vid);
 982
 983                 index = fid | (vid<<8);
 984                 powernow_table[i].index = index;
 985
 986                 freq = find_khz_freq_from_fid(fid);
 987                 powernow_table[i].frequency = freq;
 988
 989                 /* verify frequency is OK */
 990                 if ((freq > (MAX_FREQ * 1000)) || (freq < (MIN_FREQ * 1000))) {
 991                         pr_debug("invalid freq %u kHz, ignoring\n", freq);
 992                         invalidate_entry(powernow_table, i);
 993                         continue;
 994                 }
 995
 996                 /* verify voltage is OK -
 997                  * BIOSs are using "off" to indicate invalid */
 998                 if (vid == VID_OFF) {
 999                         pr_debug("invalid vid %u, ignoring\n", vid);
1000                         invalidate_entry(powernow_table, i);
1001                         continue;
1002                 }
1003
1004                 if (freq != (data->acpi_data.states[i].core_frequency * 1000)) {
1005                         printk(KERN_INFO PFX "invalid freq entries "
1006                                 "%u kHz vs. %u kHz\n", freq,
1007                                 (unsigned int)
1008                                 (data->acpi_data.states[i].core_frequency
1009                                  * 1000));
1010                         invalidate_entry(powernow_table, i);
1011                         continue;
1012                 }
1013         }
1014         return 0;
1015 }
1016
1017 static void powernow_k8_cpu_exit_acpi(struct powernow_k8_data *data)
1018 {
1019         if (data->acpi_data.state_count)
1020                 acpi_processor_unregister_performance(&data->acpi_data,
1021                                 data->cpu);
1022         free_cpumask_var(data->acpi_data.shared_cpu_map);
1023 }
1024
1025 static int get_transition_latency(struct powernow_k8_data *data)
1026 {
1027         int max_latency = 0;
1028         int i;
1029         for (i = 0; i < data->acpi_data.state_count; i++) {
1030                 int cur_latency = data->acpi_data.states[i].transition_latency
1031                         + data->acpi_data.states[i].bus_master_latency;
1032                 if (cur_latency > max_latency)
1033                         max_latency = cur_latency;
1034         }
1035         if (max_latency == 0) {
1036                 /*
1037                  * Fam 11h and later may return 0 as transition latency. This
1038                  * is intended and means "very fast". While cpufreq core and
1039                  * governors currently can handle that gracefully, better set it
1040                  * to 1 to avoid problems in the future.
1041                  */
1042                 if (boot_cpu_data.x86 < 0x11)
1043                         printk(KERN_ERR FW_WARN PFX "Invalid zero transition "
1044                                 "latency\n");
1045                 max_latency = 1;
1046         }
1047         /* value in usecs, needs to be in nanoseconds */
1048         return 1000 * max_latency;
1049 }
1050
1051 /* Take a frequency, and issue the fid/vid transition command */
1052 static int transition_frequency_fidvid(struct powernow_k8_data *data,
1053                 unsigned int index)
1054 {
1055         u32 fid = 0;
1056         u32 vid = 0;
1057         int res, i;
1058         struct cpufreq_freqs freqs;
1059
1060         pr_debug("cpu %d transition to index %u\n", smp_processor_id(), index);
1061
1062         /* fid/vid correctness check for k8 */
1063         /* fid are the lower 8 bits of the index we stored into
1064          * the cpufreq frequency table in find_psb_table, vid
1065          * are the upper 8 bits.
1066          */
1067         fid = data->powernow_table[index].index & 0xFF;
1068         vid = (data->powernow_table[index].index & 0xFF00) >> 8;
1069
1070         pr_debug("table matched fid 0x%x, giving vid 0x%x\n", fid, vid);
1071
1072         if (query_current_values_with_pending_wait(data))
1073                 return 1;
1074
1075         if ((data->currvid == vid) && (data->currfid == fid)) {
1076                 pr_debug("target matches current values (fid 0x%x, vid 0x%x)\n",
1077                         fid, vid);
1078                 return 0;
1079         }
1080
1081         pr_debug("cpu %d, changing to fid 0x%x, vid 0x%x\n",
1082                 smp_processor_id(), fid, vid);
1083         freqs.old = find_khz_freq_from_fid(data->currfid);
1084         freqs.new = find_khz_freq_from_fid(fid);
1085
1086         for_each_cpu(i, data->available_cores) {
1087                 freqs.cpu = i;
1088                 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
1089         }
1090
1091         res = transition_fid_vid(data, fid, vid);
1092         if (res)
1093                 return res;
1094
1095         freqs.new = find_khz_freq_from_fid(data->currfid);
1096
1097         for_each_cpu(i, data->available_cores) {
1098                 freqs.cpu = i;
1099                 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
1100         }
1101         return res;
1102 }
1103
1104 /* Take a frequency, and issue the hardware pstate transition command */
1105 static int transition_frequency_pstate(struct powernow_k8_data *data,
1106                 unsigned int index)
1107 {
1108         u32 pstate = 0;
1109         int res, i;
1110         struct cpufreq_freqs freqs;
1111
1112         pr_debug("cpu %d transition to index %u\n", smp_processor_id(), index);
1113
1114         /* get MSR index for hardware pstate transition */
1115         pstate = index & HW_PSTATE_MASK;
1116         if (pstate > data->max_hw_pstate)
1117                 return -EINVAL;
1118
1119         freqs.old = find_khz_freq_from_pstate(data->powernow_table,
1120                         data->currpstate);
1121         freqs.new = find_khz_freq_from_pstate(data->powernow_table, pstate);
1122
1123         for_each_cpu(i, data->available_cores) {
1124                 freqs.cpu = i;
1125                 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
1126         }
1127
1128         res = transition_pstate(data, pstate);
1129         freqs.new = find_khz_freq_from_pstate(data->powernow_table, pstate);
1130
1131         for_each_cpu(i, data->available_cores) {
1132                 freqs.cpu = i;
1133                 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
1134         }
1135         return res;
1136 }
1137
1138 /* Driver entry point to switch to the target frequency */
1139 static int powernowk8_target(struct cpufreq_policy *pol,
1140                 unsigned targfreq, unsigned relation)
1141 {
1142         cpumask_var_t oldmask;
1143         struct powernow_k8_data *data = per_cpu(powernow_data, pol->cpu);
1144         u32 checkfid;
1145         u32 checkvid;
1146         unsigned int newstate;
1147         int ret = -EIO;
1148
1149         if (!data)
1150                 return -EINVAL;
1151
1152         checkfid = data->currfid;
1153         checkvid = data->currvid;
1154
1155         /* only run on specific CPU from here on. */
1156         /* This is poor form: use a workqueue or smp_call_function_single */
1157         if (!alloc_cpumask_var(&oldmask, GFP_KERNEL))
1158                 return -ENOMEM;
1159
1160         cpumask_copy(oldmask, tsk_cpus_allowed(current));
1161         set_cpus_allowed_ptr(current, cpumask_of(pol->cpu));
1162
1163         if (smp_processor_id() != pol->cpu) {
1164                 printk(KERN_ERR PFX "limiting to cpu %u failed\n", pol->cpu);
1165                 goto err_out;
1166         }
1167
1168         if (pending_bit_stuck()) {
1169                 printk(KERN_ERR PFX "failing targ, change pending bit set\n");
1170                 goto err_out;
1171         }
1172
1173         pr_debug("targ: cpu %d, %d kHz, min %d, max %d, relation %d\n",
1174                 pol->cpu, targfreq, pol->min, pol->max, relation);
1175
1176         if (query_current_values_with_pending_wait(data))
1177                 goto err_out;
1178
1179         if (cpu_family != CPU_HW_PSTATE) {
1180                 pr_debug("targ: curr fid 0x%x, vid 0x%x\n",
1181                 data->currfid, data->currvid);
1182
1183                 if ((checkvid != data->currvid) ||
1184                     (checkfid != data->currfid)) {
1185                         printk(KERN_INFO PFX
1186                                 "error - out of sync, fix 0x%x 0x%x, "
1187                                 "vid 0x%x 0x%x\n",
1188                                 checkfid, data->currfid,
1189                                 checkvid, data->currvid);
1190                 }
1191         }
1192
1193         if (cpufreq_frequency_table_target(pol, data->powernow_table,
1194                                 targfreq, relation, &newstate))
1195                 goto err_out;
1196
1197         mutex_lock(&fidvid_mutex);
1198
1199         powernow_k8_acpi_pst_values(data, newstate);
1200
1201         if (cpu_family == CPU_HW_PSTATE)
1202                 ret = transition_frequency_pstate(data,
1203                         data->powernow_table[newstate].index);
1204         else
1205                 ret = transition_frequency_fidvid(data, newstate);
1206         if (ret) {
1207                 printk(KERN_ERR PFX "transition frequency failed\n");
1208                 ret = 1;
1209                 mutex_unlock(&fidvid_mutex);
1210                 goto err_out;
1211         }
1212         mutex_unlock(&fidvid_mutex);
1213
1214         if (cpu_family == CPU_HW_PSTATE)
1215                 pol->cur = find_khz_freq_from_pstate(data->powernow_table,
1216                                 data->powernow_table[newstate].index);
1217         else
1218                 pol->cur = find_khz_freq_from_fid(data->currfid);
1219         ret = 0;
1220
1221 err_out:
1222         set_cpus_allowed_ptr(current, oldmask);
1223         free_cpumask_var(oldmask);
1224         return ret;
1225 }
1226
1227 /* Driver entry point to verify the policy and range of frequencies */
1228 static int powernowk8_verify(struct cpufreq_policy *pol)
1229 {
1230         struct powernow_k8_data *data = per_cpu(powernow_data, pol->cpu);
1231
1232         if (!data)
1233                 return -EINVAL;
1234
1235         return cpufreq_frequency_table_verify(pol, data->powernow_table);
1236 }
1237
1238 struct init_on_cpu {
1239         struct powernow_k8_data *data;
1240         int rc;
1241 };
1242
1243 static void __cpuinit powernowk8_cpu_init_on_cpu(void *_init_on_cpu)
1244 {
1245         struct init_on_cpu *init_on_cpu = _init_on_cpu;
1246
1247         if (pending_bit_stuck()) {
1248                 printk(KERN_ERR PFX "failing init, change pending bit set\n");
1249                 init_on_cpu->rc = -ENODEV;
1250                 return;
1251         }
1252
1253         if (query_current_values_with_pending_wait(init_on_cpu->data)) {
1254                 init_on_cpu->rc = -ENODEV;
1255                 return;
1256         }
1257
1258         if (cpu_family == CPU_OPTERON)
1259                 fidvid_msr_init();
1260
1261         init_on_cpu->rc = 0;
1262 }
1263
1264 /* per CPU init entry point to the driver */
1265 static int __cpuinit powernowk8_cpu_init(struct cpufreq_policy *pol)
1266 {
1267         static const char ACPI_PSS_BIOS_BUG_MSG[] =
1268                 KERN_ERR FW_BUG PFX "No compatible ACPI _PSS objects found.\n"
1269                 FW_BUG PFX "Try again with latest BIOS.\n";
1270         struct powernow_k8_data *data;
1271         struct init_on_cpu init_on_cpu;
1272         int rc;
1273         struct cpuinfo_x86 *c = &cpu_data(pol->cpu);
1274
1275         if (!cpu_online(pol->cpu))
1276                 return -ENODEV;
1277
1278         smp_call_function_single(pol->cpu, check_supported_cpu, &rc, 1);
1279         if (rc)
1280                 return -ENODEV;
1281
1282         data = kzalloc(sizeof(struct powernow_k8_data), GFP_KERNEL);
1283         if (!data) {
1284                 printk(KERN_ERR PFX "unable to alloc powernow_k8_data");
1285                 return -ENOMEM;
1286         }
1287
1288         data->cpu = pol->cpu;
1289         data->currpstate = HW_PSTATE_INVALID;
1290
1291         if (powernow_k8_cpu_init_acpi(data)) {
1292                 /*
1293                  * Use the PSB BIOS structure. This is only available on
1294                  * an UP version, and is deprecated by AMD.
1295                  */
1296                 if (num_online_cpus() != 1) {
1297                         printk_once(ACPI_PSS_BIOS_BUG_MSG);
1298                         goto err_out;
1299                 }
1300                 if (pol->cpu != 0) {
1301                         printk(KERN_ERR FW_BUG PFX "No ACPI _PSS objects for "
1302                                "CPU other than CPU0. Complain to your BIOS "
1303                                "vendor.\n");
1304                         goto err_out;
1305                 }
1306                 rc = find_psb_table(data);
1307                 if (rc)
1308                         goto err_out;
1309
1310                 /* Take a crude guess here.
1311                  * That guess was in microseconds, so multiply with 1000 */
1312                 pol->cpuinfo.transition_latency = (
1313                          ((data->rvo + 8) * data->vstable * VST_UNITS_20US) +
1314                          ((1 << data->irt) * 30)) * 1000;
1315         } else /* ACPI _PSS objects available */
1316                 pol->cpuinfo.transition_latency = get_transition_latency(data);
1317
1318         /* only run on specific CPU from here on */
1319         init_on_cpu.data = data;
1320         smp_call_function_single(data->cpu, powernowk8_cpu_init_on_cpu,
1321                                  &init_on_cpu, 1);
1322         rc = init_on_cpu.rc;
1323         if (rc != 0)
1324                 goto err_out_exit_acpi;
1325
1326         if (cpu_family == CPU_HW_PSTATE)
1327                 cpumask_copy(pol->cpus, cpumask_of(pol->cpu));
1328         else
1329                 cpumask_copy(pol->cpus, cpu_core_mask(pol->cpu));
1330         data->available_cores = pol->cpus;
1331
1332         if (cpu_family == CPU_HW_PSTATE)
1333                 pol->cur = find_khz_freq_from_pstate(data->powernow_table,
1334                                 data->currpstate);
1335         else
1336                 pol->cur = find_khz_freq_from_fid(data->currfid);
1337         pr_debug("policy current frequency %d kHz\n", pol->cur);
1338
1339         /* min/max the cpu is capable of */
1340         if (cpufreq_frequency_table_cpuinfo(pol, data->powernow_table)) {
1341                 printk(KERN_ERR FW_BUG PFX "invalid powernow_table\n");
1342                 powernow_k8_cpu_exit_acpi(data);
1343                 kfree(data->powernow_table);
1344                 kfree(data);
1345                 return -EINVAL;
1346         }
1347
1348         /* Check for APERF/MPERF support in hardware */
1349         if (cpu_has(c, X86_FEATURE_APERFMPERF))
1350                 cpufreq_amd64_driver.getavg = cpufreq_get_measured_perf;
1351
1352         cpufreq_frequency_table_get_attr(data->powernow_table, pol->cpu);
1353
1354         if (cpu_family == CPU_HW_PSTATE)
1355                 pr_debug("cpu_init done, current pstate 0x%x\n",
1356                                 data->currpstate);
1357         else
1358                 pr_debug("cpu_init done, current fid 0x%x, vid 0x%x\n",
1359                         data->currfid, data->currvid);
1360
1361         per_cpu(powernow_data, pol->cpu) = data;
1362
1363         return 0;
1364
1365 err_out_exit_acpi:
1366         powernow_k8_cpu_exit_acpi(data);
1367
1368 err_out:
1369         kfree(data);
1370         return -ENODEV;
1371 }
1372
1373 static int __devexit powernowk8_cpu_exit(struct cpufreq_policy *pol)
1374 {
1375         struct powernow_k8_data *data = per_cpu(powernow_data, pol->cpu);
1376
1377         if (!data)
1378                 return -EINVAL;
1379
1380         powernow_k8_cpu_exit_acpi(data);
1381
1382         cpufreq_frequency_table_put_attr(pol->cpu);
1383
1384         kfree(data->powernow_table);
1385         kfree(data);
1386         per_cpu(powernow_data, pol->cpu) = NULL;
1387
1388         return 0;
1389 }
1390
1391 static void query_values_on_cpu(void *_err)
1392 {
1393         int *err = _err;
1394         struct powernow_k8_data *data = __this_cpu_read(powernow_data);
1395
1396         *err = query_current_values_with_pending_wait(data);
1397 }
1398
1399 static unsigned int powernowk8_get(unsigned int cpu)
1400 {
1401         struct powernow_k8_data *data = per_cpu(powernow_data, cpu);
1402         unsigned int khz = 0;
1403         int err;
1404
1405         if (!data)
1406                 return 0;
1407
1408         smp_call_function_single(cpu, query_values_on_cpu, &err, true);
1409         if (err)
1410                 goto out;
1411
1412         if (cpu_family == CPU_HW_PSTATE)
1413                 khz = find_khz_freq_from_pstate(data->powernow_table,
1414                                                 data->currpstate);
1415         else
1416                 khz = find_khz_freq_from_fid(data->currfid);
1417
1418
1419 out:
1420         return khz;
1421 }
1422
1423 static void _cpb_toggle_msrs(bool t)
1424 {
1425         int cpu;
1426
1427         get_online_cpus();
1428
1429         rdmsr_on_cpus(cpu_online_mask, MSR_K7_HWCR, msrs);
1430
1431         for_each_cpu(cpu, cpu_online_mask) {
1432                 struct msr *reg = per_cpu_ptr(msrs, cpu);
1433                 if (t)
1434                         reg->l &= ~BIT(25);
1435                 else
1436                         reg->l |= BIT(25);
1437         }
1438         wrmsr_on_cpus(cpu_online_mask, MSR_K7_HWCR, msrs);
1439
1440         put_online_cpus();
1441 }
1442
1443 /*
1444  * Switch on/off core performance boosting.
1445  *
1446  * 0=disable
1447  * 1=enable.
1448  */
1449 static void cpb_toggle(bool t)
1450 {
1451         if (!cpb_capable)
1452                 return;
1453
1454         if (t && !cpb_enabled) {
1455                 cpb_enabled = true;
1456                 _cpb_toggle_msrs(t);
1457                 printk(KERN_INFO PFX "Core Boosting enabled.\n");
1458         } else if (!t && cpb_enabled) {
1459                 cpb_enabled = false;
1460                 _cpb_toggle_msrs(t);
1461                 printk(KERN_INFO PFX "Core Boosting disabled.\n");
1462         }
1463 }
1464
1465 static ssize_t store_cpb(struct cpufreq_policy *policy, const char *buf,
1466                                  size_t count)
1467 {
1468         int ret = -EINVAL;
1469         unsigned long val = 0;
1470
1471         ret = strict_strtoul(buf, 10, &val);
1472         if (!ret && (val == 0 || val == 1) && cpb_capable)
1473                 cpb_toggle(val);
1474         else
1475                 return -EINVAL;
1476
1477         return count;
1478 }
1479
1480 static ssize_t show_cpb(struct cpufreq_policy *policy, char *buf)
1481 {
1482         return sprintf(buf, "%u\n", cpb_enabled);
1483 }
1484
1485 #define define_one_rw(_name) \
1486 static struct freq_attr _name = \
1487 __ATTR(_name, 0644, show_##_name, store_##_name)
1488
1489 define_one_rw(cpb);
1490
1491 static struct freq_attr *powernow_k8_attr[] = {
1492         &cpufreq_freq_attr_scaling_available_freqs,
1493         &cpb,
1494         NULL,
1495 };
1496
1497 static struct cpufreq_driver cpufreq_amd64_driver = {
1498         .verify         = powernowk8_verify,
1499         .target         = powernowk8_target,
1500         .bios_limit     = acpi_processor_get_bios_limit,
1501         .init           = powernowk8_cpu_init,
1502         .exit           = __devexit_p(powernowk8_cpu_exit),
1503         .get            = powernowk8_get,
1504         .name           = "powernow-k8",
1505         .owner          = THIS_MODULE,
1506         .attr           = powernow_k8_attr,
1507 };
1508
1509 /*
1510  * Clear the boost-disable flag on the CPU_DOWN path so that this cpu
1511  * cannot block the remaining ones from boosting. On the CPU_UP path we
1512  * simply keep the boost-disable flag in sync with the current global
1513  * state.
1514  */
1515 static int cpb_notify(struct notifier_block *nb, unsigned long action,
1516                       void *hcpu)
1517 {
1518         unsigned cpu = (long)hcpu;
1519         u32 lo, hi;
1520
1521         switch (action) {
1522         case CPU_UP_PREPARE:
1523         case CPU_UP_PREPARE_FROZEN:
1524
1525                 if (!cpb_enabled) {
1526                         rdmsr_on_cpu(cpu, MSR_K7_HWCR, &lo, &hi);
1527                         lo |= BIT(25);
1528                         wrmsr_on_cpu(cpu, MSR_K7_HWCR, lo, hi);
1529                 }
1530                 break;
1531
1532         case CPU_DOWN_PREPARE:
1533         case CPU_DOWN_PREPARE_FROZEN:
1534                 rdmsr_on_cpu(cpu, MSR_K7_HWCR, &lo, &hi);
1535                 lo &= ~BIT(25);
1536                 wrmsr_on_cpu(cpu, MSR_K7_HWCR, lo, hi);
1537                 break;
1538
1539         default:
1540                 break;
1541         }
1542
1543         return NOTIFY_OK;
1544 }
1545
1546 static struct notifier_block cpb_nb = {
1547         .notifier_call          = cpb_notify,
1548 };
1549
1550 /* driver entry point for init */
1551 static int __cpuinit powernowk8_init(void)
1552 {
1553         unsigned int i, supported_cpus = 0, cpu;
1554         int rv;
1555
1556         for_each_online_cpu(i) {
1557                 int rc;
1558                 smp_call_function_single(i, check_supported_cpu, &rc, 1);
1559                 if (rc == 0)
1560                         supported_cpus++;
1561         }
1562
1563         if (supported_cpus != num_online_cpus())
1564                 return -ENODEV;
1565
1566         printk(KERN_INFO PFX "Found %d %s (%d cpu cores) (" VERSION ")\n",
1567                 num_online_nodes(), boot_cpu_data.x86_model_id, supported_cpus);
1568
1569         if (boot_cpu_has(X86_FEATURE_CPB)) {
1570
1571                 cpb_capable = true;
1572
1573                 msrs = msrs_alloc();
1574                 if (!msrs) {
1575                         printk(KERN_ERR "%s: Error allocating msrs!\n", __func__);
1576                         return -ENOMEM;
1577                 }
1578
1579                 register_cpu_notifier(&cpb_nb);
1580
1581                 rdmsr_on_cpus(cpu_online_mask, MSR_K7_HWCR, msrs);
1582
1583                 for_each_cpu(cpu, cpu_online_mask) {
1584                         struct msr *reg = per_cpu_ptr(msrs, cpu);
1585                         cpb_enabled |= !(!!(reg->l & BIT(25)));
1586                 }
1587
1588                 printk(KERN_INFO PFX "Core Performance Boosting: %s.\n",
1589                         (cpb_enabled ? "on" : "off"));
1590         }
1591
1592         rv = cpufreq_register_driver(&cpufreq_amd64_driver);
1593         if (rv < 0 && boot_cpu_has(X86_FEATURE_CPB)) {
1594                 unregister_cpu_notifier(&cpb_nb);
1595                 msrs_free(msrs);
1596                 msrs = NULL;
1597         }
1598         return rv;
1599 }
1600
1601 /* driver entry point for term */
1602 static void __exit powernowk8_exit(void)
1603 {
1604         pr_debug("exit\n");
1605
1606         if (boot_cpu_has(X86_FEATURE_CPB)) {
1607                 msrs_free(msrs);
1608                 msrs = NULL;
1609
1610                 unregister_cpu_notifier(&cpb_nb);
1611         }
1612
1613         cpufreq_unregister_driver(&cpufreq_amd64_driver);
1614 }
1615
1616 MODULE_AUTHOR("Paul Devriendt <paul.devriendt@amd.com> and "
1617                 "Mark Langsdorf <mark.langsdorf@amd.com>");
1618 MODULE_DESCRIPTION("AMD Athlon 64 and Opteron processor frequency driver.");
1619 MODULE_LICENSE("GPL");
1620
1621 late_initcall(powernowk8_init);
1622 module_exit(powernowk8_exit);